Joachim R.

Daduna
Gernot Liedtke
Xiaoning Shi
Stefan Voß (Eds.)
LNCS 14239

Computational
Logistics
14th International Conference, ICCL 2023
Berlin, Germany, September 6–8, 2023
Proceedings
Lecture Notes in Computer Science 14239
Founding Editors
Gerhard Goos
Juris Hartmanis

Editorial Board Members

Elisa Bertino, Purdue University, West Lafayette, IN, USA
Wen Gao, Peking University, Beijing, China
Bernhard Steffen , TU Dortmund University, Dortmund, Germany
Moti Yung , Columbia University, New York, NY, USA
The series Lecture Notes in Computer Science (LNCS), including its subseries Lecture
Notes in Artificial Intelligence (LNAI) and Lecture Notes in Bioinformatics (LNBI),
has established itself as a medium for the publication of new developments in computer
science and information technology research, teaching, and education.
LNCS enjoys close cooperation with the computer science R & D community, the
series counts many renowned academics among its volume editors and paper authors, and
collaborates with prestigious societies. Its mission is to serve this international commu-
nity by providing an invaluable service, mainly focused on the publication of conference
and workshop proceedings and postproceedings. LNCS commenced publication in 1973.
Joachim R. Daduna · Gernot Liedtke ·
Xiaoning Shi · Stefan Voß
Editors

Computational
Logistics
14th International Conference, ICCL 2023
Berlin, Germany, September 6–8, 2023
Proceedings
Editors
Joachim R. Daduna Gernot Liedtke
Hochschule für Wirtschaft und Recht Berlin Deutsches Zentrum für Luft- und Raumfahrt
Berlin, Germany (DLR)
Berlin, Germany
Xiaoning Shi
Deutsches Zentrum für Luft- und Raumfahrt Stefan Voß
(DLR) University of Hamburg
Berlin, Germany Hamburg, Germany

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-031-43611-6 ISBN 978-3-031-43612-3 (eBook)
https://doi.org/10.1007/978-3-031-43612-3

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2023

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors
or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Paper in this product is recyclable.

 Preface

For decades, logistics tasks were of minor importance in economics as well as in busi-
ness practice; the basic functions transportation, transshipment and warehousing were
regarded as necessary factors, but not essential from a market-related as well as a compet-
itive point of view. Discussions about logistics were mainly related to the military sector,
from which, however, the key impact for a fundamental rethinking came at the end of
the 1940s. The employment of qualified former military personnel with leadership and
management experience in the civilian economy resulted in a transfer of essential logis-
tical principles into operational planning and control, the importance of which increased
steadily with the internationalization and, in particular, the globalization of the economy.
In the context of further developments, it became clear that logistics is not every-
thing, but without logistics, everything is nothing. It became increasingly obvious that
the mobility of goods is and will continue to be an essential prerequisite for ensuring sat-
isfactory efficiency in service provision processes organized on the basis of the division
of labor within the framework of spatially decentralized structures, which increasingly
determined the industrial sector as well as retail trade. Linked to this was the neces-
sity to install suitable external and internal information management for the continuous
monitoring and control of a globally connected cooperation.
The essential importance of reliable and efficient flows of goods and information
has become apparent to the general public. This holds especially as a result of the
consequences of the Covid-19 pandemic and the effects of the Russian war against
Ukraine. But even singular events, such as a well-observed accident in the Suez Canal
not too long ago, can lead to serious disruptions in logistical processes, combined with
significant negative economic consequences. Even despite increasing de-globalization
and backshoring, efficiency and costs must not be the only key objectives; ensuring
resilient logistics structures must also be included. Ultimately, this also means that
efficient logistics must be seen as a necessary core competence for the economic success
of companies, especially in the manufacturing and retail trade sectors.
In recent years, far-reaching progress has been made in logistics as well as in informa-
tion and communication technology, for example in the areas of theoretical foundations,
the available technical systems and also in operational realizations. In the foreground
were the development and improvement of algorithms, the increased automation in trans-
port and warehousing up to autonomous systems, the use of processes from the field of
artificial intelligence as well as the development of new materials and manufacturing
processes and also advances in communication technology. This results in the mandatory
necessity of a problem-adequate and object-oriented combination of logistical functions,
quantitative solution procedures and methods for decision support with elements of effi-
cient information management in theory and practice. To push these developments is the
central aim of the International Conferences on Computational Logistics (ICCL), which
have been held worldwide for 13 years now.
Starting as a joint endeavor between Hamburg and Shanghai the ICCL was held in
Shanghai (China) in 2010 (as well as in 2012), followed in 2011 by Hamburg (Germany),
2013 Copenhagen (Denmark), 2014 Valparaiso (Chile), 2015 Delft (The Netherlands),
vi Preface

2016 Lisbon (Portugal), 2017 Southampton (UK), 2018 Salerno (Italy), 2019 Barran-
quilla (Colombia), 2020 as well as 2021 Enschede (The Netherlands (online)) and 2022
Barcelona (Spain). The 14th ICCL conference, which took place in Berlin (Germany)
in September 2023, also aimed to bring together representatives from fields of research
and development as well as from practice with the objective of an intensive exchange
of knowledge and experiences. Presentations and discussions of current research results
as well as developments in various areas of logistics were at the foreground. The main
topics of this conference were grouped into five streams as follows:
1. Computational Logistics
Starting from observing the recent trend of using generative artificial intelligence
tools, the first paper provides an insight into this regarding logistics. Moreover, among
others, this group of papers considers issues of recommendation systems and logistics
platforms, internal hospital logistics, cybersecurity aspects for synchromodal freight
operations, as well as an emprical study regarding cross-border e-commerce.
2. Maritime Shipping
The largest group of papers, extending a major focus of the ICCL conferences right
from the start, concerns maritime shipping, including considerations of inland water-
ways. Specifically, the included papers span a wide area from developing digital
twins in seaports, tramp ship routing, liquified natural gas (LNG) delivery programs,
conventional as well as roll-on roll-off (RORO) stowage planning up to problems
regarding the allocation of shore-side electricity for grouped container terminals,
among others.
3. Vehicle Routing
Vehicle routing problems make up the core of many logistics considerations. Beyond
the inclusion of advances in classical vehicle routing, we encounter quite a few prac-
tical applications using a variety of modeling approaches. In more detail, we have
works dealing with the vehicle routing problem with time windows, a truck-drone
routing problem, risk and uncertainty considerations, e.g., in humanitarian logistics
as well as an e-waste collection problem. An interesting twist are novel applications
including the snow grooming routing problem, which is an extension of an arc rout-
ing problem with profits. The problem arises in cross-country skiing facilities when
scheduling the grooming of track networks.
4. Traffic and Transport
This volume extends the interest in public transport works, compared to earlier ones.
This includes papers on rolling stock circulation planning as well as robust bus driver
rostering with uncertainty, but also the prediction and analysis of travel times for
transit ferries and ridesharing issues in rural areas with autonomous electric vehicles.
The usability of public transport regarding health care services in rural areas as well
as questions regarding the combination of public transport and freight transport are
considered, too.
5. Combinatorial Optimization
While many of the above problems might be considered as combinatorial optimization
problems, there is more. Applications relate to location planning, job shop scheduling,
but also various path planning problems and rich lot-sizing problems.
 Preface vii

The editors thank all the authors for their submissions as well as the program commit-
tee and the reviewers for their helpful support and important feedback which have made
a significant contribution to the realization of this book. Reviewing was single blind with
33 full papers accepted out of 71 submissions. Each paper was assigned to at least two
reviewers and the average number of reviews received per submission was 2.42. Finally,
we would like to express our thanks to Julia Bachale for her comprehensive support and
assistance during the preparation of the conference and also the manuscript. We trust that
the present volume supports the more important advances within computational logistics
and inspires all participants and readers to further research activities.

September 2023 Joachim R. Daduna

Gernot Liedtke
Xiaoning Shi
Stefan Voß
 Organization

Program Committee Chairs

Daduna, Joachim R. Berlin School of Economics and Law, Germany

Liedtke, Gernot German Aerospace Center (DLR) Berlin,
Germany
Shi, Xiaoning German Aerospace Center (DLR) Berlin,
Germany
Voß, Stefan University of Hamburg, Germany

Program Committee and Reviewers

Amirghasemi, Mehrdad University of Wollongong, Australia

Armas, Jesica de Pompeu Fabra University, Spain
Bearzotti, Lorena Pontifical Catholic University of Valparaiso, Chile
Bell, Michael University of Sydney, Australia
Boysen, Nils University of Jena, Germany
Bubalo, Branko KPMG Berlin, Germany
Cerulli, Raffaele University of Salerno, Italy
Colak, Merve University of Hamburg, Germany
Cuerebal, Ahmet University of Hamburg, Germany
Daduna, Joachim R. Berlin School of Economics and Law, Germany
Festa, Paola University of Naples, Italy
Heilig, Leonard University of Hamburg, Germany
Gonzalez, Rosa University of the Andes, Chile
Hartl, Richard University of Vienna, Austria
Hill, Alessandro California Polytechnic State University, San Luis
Obispo, USA
Ji, Zhe University of Hamburg, Germany
Jovanovic, Raka Hamad bin Khalifa University, Qatar
Lalla-Ruiz, Eduardo University of Twente, The Netherlands
Lau, Hoong Chuin Singapore Management University, Singapore
Mattfeld, Dirk TU Braunschweig, Germany
Maturana-Ross, Javier Pontifical Catholic University of Valparaiso, Chile
Mazza, Rina Mary University of Calabria, Italy
Mazzulla, Gabriella University of Calabria, Italy
Mes, Martijn University of Twente, The Netherlands
x Organization

Møller Jensen, Rune IT University of Copenhagen, Denmark

Negenborn, Rudi Technical University of Delft, The Netherlands
Neugebauer, Julian University of Hamburg, Germany
Nourmohammadzadeh, Abtin University of Hamburg, Germany
Pacino, Dario Technical University of Denmark, Denmark
Pahl, Julia University of Southern Denmark, Denmark
Paias, Ana University of Lisbon, Portugal
Paternina-Arboleda, Carlos San Diego State University, USA
Radojicic, Nina University of Hamburg, Germany
Ramalhinho Lourenço, Helena Pompeu Fabra University, Spain
Rössler, David Free University of Berlin, Germany
Salazar Gonzalez, Juan Jose University of La Laguna, Spain
Sarhani, Malek Al Akhawayn University, Morocco
Schulte, Frederik Delft University of Technology, The Netherlands
Schwartz, Frank University of Hamburg, Germany
Shi, Xiaoning German Aerospace Center (DLR) Berlin,
Germany
Tierney, Kevin University of Bielefeld, Germany
Ulmer, Marlin University of Magdeburg, Germany
Voß, Stefan University of Hamburg, Germany
Xie, Lin University of Twente, The Netherlands
Yu, Jingjing University of Hamburg, Germany
 Contents

Computational Logistics

Successfully Using ChatGPT in Logistics: Are We There Yet? . . . . . . . . . . . . . . . 3

Stefan Voß

When Routing Meets Recommendation: Solving Dynamic Order

Recommendations Problem in Peer-to-Peer Logistics Platforms . . . . . . . . . . . . . . 18
Zhiqin Zhang, Waldy Joe, Yuyang Er, and Hoong Chuin Lau

A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics . . . . . . 36

Ebrahim Ehsanfar, Farzaneh Karami, and Tim Kerkenhoff

Cybersecurity Considerations for the Design of an AI-Driven Distributed

Optimization of Container Carbon Emissions Reduction for Freight
Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Carlos Paternina-Arboleda, Alexander Nestler, Nicholas Kascak,
and Morteza Safaei Pour

Customer’s Choice in the Context of Cross-Border E-Commerce:

An Application of Structural Equation Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Yijia Liu and Xiaoning Shi

Maritime Shipping

Towards a Deep Reinforcement Learning Model of Master Bay Stowage

Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Jaike van Twiller, Djordje Grbic, and Rune Møller Jensen

The Dynamic RORO Stowage Planning Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Alastair Ronald Main, Filipe Rodrigues, and Dario Pacino

Allocation of Shore Side Electricity: The Case of the Port of Hamburg . . . . . . . . 139
Jingjing Yu, Philip Cammin, and Stefan Voß

Stockyard Storage Space Allocation in Dry Bulk Terminals Considering

Mist Cannons and Energy Expenditure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Xinyu Tang and Frederik Schulte
xii Contents

Planning LNG Annual Delivery Programs with Speed Optimization

and Multiple Loading Ports . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
Helle V. Haug, Sigrid H. Solum, Sanna B. Warholm, Kjetil Fagerholt,
Mingyu Li, and Inge Norstad

Tramp Ship Routing with Bunker Optimization and Flexible Cargo

Quantities: Case from Dry Bulk Shipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Simen Omholt-Jensen, Kjetil Fagerholt, and Frank Meisel

Digital Twins in Seaports: Current and Future Applications . . . . . . . . . . . . . . . . . . 202

Julian Neugebauer, Leonard Heilig, and Stefan Voß

A Neural Network Approach for ETA Prediction in Inland Waterway

Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Peter Wenzel, Raka Jovanovic, and Frederik Schulte

Vehicle Routing

A Regret Policy for the Dynamic Vehicle Routing Problem with Time
Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Peter Dieter

A Tabu Search Algorithm for the Traveling Purchaser Problem

with Transportation Time Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Ilker Kucukoglu, Pieter Vansteenwegen, and Dirk Cattrysse

GRASP Solution Approach for the E-Waste Collection Problem . . . . . . . . . . . . . . 260

Aldy Gunawan, Dang Viet Anh Nguyen, Pham Kien Minh Nguyen,
and Pieter Vansteenwegen

The Heterogeneous Fleet Risk-Constrained Vehicle Routing Problem

in Humanitarian Logistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
Robert M van Steenbergen, Eduardo Lalla-Ruiz, Wouter van Heeswijk,
and Martijn Mes

The Snow Grooming Routing Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292

Henrik Andersson, Sondre Høyland, Jesper Anker Krogstad,
Truls Flatberg, and Anders N. Gullhav

A Constraint Programming Model for the Vehicle Routing Problem

with Multiple Time Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Florian Linß
 Contents xiii

A Variable Neighborhood Search Algorithm for the Truck-Drone Routing

Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
Batool Madani and Malick Ndiaye

Traffic and Transport

Prediction and Analysis of Transit Ferries Travel Time: An Open

Data-Based Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Malek Sarhani, Mohammed El Amrani, and Abdelhak El Achhab

A Bi-Objective Column Generation Approach for Real-World Rolling

Stock Circulation Planning Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Paul Päprer, Janis S. Neufeld, and Udo Buscher

An Effective Matheuristic Approach for Robust Bus Driver Rostering

with Uncertain Daily Working Hours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Abtin Nourmohammadzadeh and Stefan Voß

Beyond Cargo Hitching: Combined People and Freight Transport Using

Dynamically Configurable Autonomous Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Joris J. A. Kortekaas, Breno A. Beirigo, and Frederik Schulte

Impact of Public Transport Development on Health Care Services in Rural

Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
Joachim R. Daduna and Daniel Philipp

Ridesharing in Rural Areas with Autonomous Electric Vehicles

and Interrelated Trips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
Marvin Soth, Lennart C. Johnsen, Sebastian Scholz, and Frank Meisel

Combinatorial Optimization

Operational Integration of Supply Chain Activities with Earliness

and Tardiness Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Ece Yağmur and Saadettin Erhan Kesen

Constrained Multi-agent Path Planning Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 450

Ali Maktabifard, Dávid Földes, and Bendegúz Dezső Bak

UAV Path Planning for Area Coverage and Energy Consumption in Oil
and Gas Exploration Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Salim Sulaiman Maaji and Dario Landa-Silva
xiv Contents

Minimizing Peak Energy Demand in Flexible Job Shops . . . . . . . . . . . . . . . . . . . . 482

Michael Eley

Carbon-Aware Mine Planning with a Novel Multi-objective Framework . . . . . . . 500

Nurul Asyikeen Binte Azhar, Aldy Gunawan, Shih-Fen Cheng,
and Erwin Leonardi

Multi-product Lot-Sizing Problem with Remanufacturing, Lost Sales

and Sequence-Dependent Changeover Cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
Lucas Gana, Sebastián Dávila-Gálvez, and Franco Quezada

A Radius-Based Approach for the Bi-Objective p-Center and p-Dispersion

Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
Niels De Walsche, Carlo S. Sartori, and Hatice Çalık

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551

Computational Logistics
 Successfully Using ChatGPT in Logistics:
Are We There Yet?

Stefan Voß(B)

Institute of Information Systems, University of Hamburg,

Von-Melle-Park 5, 20146 Hamburg, Germany
[email protected]

Abstract. ChatGPT is among the recent most-commonly discussed

artificial intelligence systems. While many success stories as well as mis-
uses are exemplified in different fields like, e.g., education, the usefulness
in various academic disciplines with very close ties to practical applica-
tions still needs to be proven. In this paper we explore the use of this
artificial intelligence (AI) tool within the logistics domain. As a lesson
learned, one has to be careful. That is, answers cannot always be granted
as being correct. Beyond diving into related literature, we explore the use
of ChatGPT regarding an as yet underexplored (even without consult-
ing generative AI tools) logistics problem, that is, the stochastic vehicle
routing problem with uncertainty in the number of available vehicles.

Keywords: Generative artificial intelligence · ChatGPT ·

Computational logistics · Artificial intelligence · Stochastic vehicle
routing problem with uncertain number of vehicles · Right-hand-side
uncertainty

1 Introduction

After more than 40 years [48], artificial intelligence (AI) is gaining rejuvenated
popularity based on powerful computational environments as well as advances in
information systems research. Acronyms and phrases like deep learning, machine
learning, or generative AI currently make it into the popular press implying hopes
and fears all around the globe. Generative AI can be a system using AI techniques
being able to generate text, software, images, or other types of files/documents
in response to prompts, questions or requests; see, e.g., [41]. At the core of these
systems we see comprehensive language models being able to produce data based
on appropriately trained systems using possibly large data sets (making this a
big data application).
Popular generative AI systems include ChatGPT, a chatbot (a chatbot may
be seen as a computer program that simulates and processes human conversa-
tion) built by OpenAI using a large language model like GPT-3 or GPT-4.1

1
See, e.g., https://openai.com/blog/chatgpt; last access 30 April 2023.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 3–17, 2023.
https://doi.org/10.1007/978-3-031-43612-3_1
4 S. Voß

Other systems include Bard, which is related to the company Google2 , Perplex-
ity3 , Bing4 , as well as the Chinese counterpart Ernie available on Baidu [42,50].
A gentle introduction can be found at different places, e.g., in [19].
Generative AI has potential applications across a wide range of industries
and domains, including software development, marketing, and fashion [4]. Most
remarkable in the open public seems to be the discussion regarding its use in
education (see, e.g., [25]).
Notable examples regarding the use of ChatGPT in logistics include [47] with
a focus on Industry 5.0 and [52] concentrating on potential transportation appli-
cations. Other examples reaching out towards logistics in specific domains as
well as supply chain management are [6,17]. In a subsequent companion paper,
we focus on issues of bus bunching and bus bridging (see, e.g., [9]) in public
transport and also compare answers from ChatGPT with those from Bing [45].
In general, however, the logistics domain has not yet been comprehensively cov-
ered regarding the description of using generative AI systems. Moreover, while
attempting to showcase the opportunities and future applications, the limita-
tions regarding content also need to be exemplified. In this paper, we aim to
provide some insights regarding the use of ChatGPT in the logistics domain and
exemplify possible limitations by investigating a specific problem in stochastic
vehicle routing which has not yet been comprehensively covered in the academic
literature, indicating that even large language models might face some difficulties
for handling it, possibly due to the missing academic considerations in related
works.
The remainder of the paper is structured as follows. First, we provide a
literature review and problem description regarding the specific vehicle routing
problem (VRP) to be considered. Moreover, we provide some entries towards
using ChatGPT and generative AI to conduct academic studies and establish
related reports. Then, in Sect. 3, we utilize ChatGPT in an attempt to gain
insights regarding the advancements of a specific problem, namely the stochastic
VRP with an uncertain number of vehicles. Parts of the paper are using results
from ChatGPT where practical.5 Finally conclusions are drawn and selected
future research needs are exemplified.

2 Problem Exposition and Literature Review

We distinguish two important parts regarding this paper, i.e., vehicle routing
and generative AI. Note that some application-oriented works are subsumed in
the second part.
2
See, e.g., https://blog.google/technology/ai/bard-google-ai-search-updates/; last
access 30 April 2023.
3
https://www.perplexity.ai/; last access 22 June 2023.
4
https://www.bing.com/new?setlang=en&sid=081BEF962AF260422C13FCAE2BC
C615F; last access 22 June 2023.
5
All entries have been conducted or reconfirmed on 30 April 2023 using the release
from 23 March 2023; see https://help.openai.com/en/articles/6825453-chatgpt-
release-notes.
 Using ChatGPT in Logistics 5

2.1 Vehicle Routing

In this paper, we refer to problem settings leading towards a specific VRP as
well as some vehicle scheduling issues. Therefore, we start gently by introducing
the general problem settings before becoming specific.

“The vehicle routing problem (VRP) is a combinatorial optimization prob-

lem that involves finding the best set of routes for a fleet of vehicles to
service a set of customers or locations. In the VRP, each customer or loca-
tion has a demand for goods or services, and the goal is to determine the
most efficient set of routes for the vehicles to deliver the goods or provide
the services to the customers. The VRP is a complex problem that can be
formulated in many different ways, depending on the specific constraints
and objectives of the problem.” (ChatGPT)

Asking ChatGPT about important references on the VRP gives five mean-
ingful entries including the following (plus one reference that could not be found
in popular databases): [11,43].
VRPs are among the most studied problems in operations research and dis-
crete applied mathematics with quite a few textbooks and well-known compendi-
ums devoted to the problem with a most recent one being [12]. Also surveys seem
to be a dime a dozen including those focusing, e.g., on specific issues such as
logistics distribution [18], reverse logistics [36] or stochastic VRPs [32].
Related to the VRP (and our research focus; see below) is also the topic of
vehicle scheduling where the basic idea is to cover a given set of trips by means of
a set of vehicles. To be more specific and choosing a specialized problem setting,
given a set of timetabled trips and a fleet of vehicles the integrated vehicle and
duty scheduling problem seeks to find minimum-cost vehicle blocks (the schedule
of a vehicle) and valid driver duties such that each active trip is covered by
one block, each active trip segment is covered by one duty, and each deadhead,
pull-in, and pull-out trip used in the vehicle schedule is also covered by one duty.
That is, a vehicle performs a feasible sequence of trips from the time it leaves a
depot until it returns to the depot. Example references include [7,8,21].
Using different words, this can be described as follows:
“The integrated vehicle and duty scheduling problem (IVDSP) is a com-
binatorial optimization problem that involves scheduling a set of vehicles
and their drivers over a planning horizon to fulfill a given set of trips or
duties. The objective of the IVDSP is to minimize the total cost, which
typically includes the cost of vehicle operations, driver labor, and penalties
for violations of operational constraints.
The IVDSP is a challenging problem due to its complexity and the large
number of decision variables involved. The problem involves simultane-
ously scheduling vehicles and drivers, considering factors such as vehicle
capacity, driver skills, rest and work regulations, and travel times. Addi-
tionally, the problem often includes uncertain demand, which further com-
plicates the planning process.” (ChatGPT)
6 S. Voß

A specific area of research in the VRP domain is related to stochastic VRPs

where it seems important to first distinguish between stochastic and dynamic
VRPs. In the dynamic VRP (DVRP), also referred to as real-time or online VRP,
some input data are (only) revealed during the execution of the plan (aka a solu-
tion). Examples for such input data can be knowledge about specific customers
and/or their demands. More specifically, this can refer to new customer requests,
real-time traffic information, knowledge about disturbances like vehicle break-
downs, weather conditions, capacity changes or even customer cancellations on
a short notice.
The stochastic VRP (SVRP) is any VRP where at least one parameter is
stochastic, i.e., some future events are random variables with a known probabil-
ity distribution. Examples include the time required to serve customers where
different customers may require different service times due to the specific nature
of the service provided or varying order complexities. There might also be a
probability distribution for vehicle breakdowns or failures during the routing
process which can affect the availability of vehicles and require adjustments to
the routes.
After an older survey by Gendreau et al. [10], there are quite a few more
recent surveys on stochastic vehicle routing including [3,31,32,35,39]. While
these surveys seem to be quite comprehensive, a wealth of specific additional
works is available.
When emphasizing the SVRP, among the issues receiving general attention
in recent scientific literature on logistics and transportation are the topics of
robustness and uncertainty. This becomes especially important in the context of
disaster relief problems like the January 2023 earthquake in Turkey and Syria
or other “prominent” disasters of the last decades like the hurricane Katrina in
New Orleans, USA (2005) or the Fukushima disaster in Japan (2011). Related
VRP research in this respect might be concerned about an uncertain number
of available vehicles, though, this seems to be limited to very few papers; see,
e.g., [38]. Of course, not necessarily focusing on disaster relief issues in logistics,
specific types of problems reveal the mentioned uncertainty per se, such as the
investigation of the concept of crowdsourced delivery in [37]. Here, the crowd-
sourced couriers are not employed by the platform and this has fundamentally
changed the planning and execution of the delivery of goods as the number of
available drivers and vehicles is not known a priori. A related body of references
exists on agent-based models and transport simulations, where the number of
vehicles can vary like in taxi or ride-hailing systems; see, e.g., [49]. Actually,
issues regarding a shortage of drivers are commonly discussed in logistics lead-
ing to quite some uncertainty in the spirit of what was mentioned above (see,
e.g., [27]).
Interestingly enough, there seems to be no consideration in the above-
mentioned survey papers regarding an uncertain number of vehicles available.
Usually, the number of vehicles can be a given constant or a decision variable
[10]. However, an explicit consideration of an uncertain number of vehicles can
 Using ChatGPT in Logistics 7

hardly be found with [38] being an exception, as pointed out in the previous
paragraph.
The above considerations open up the avenue towards formulating and inves-
tigating an SVRP or a vehicle scheduling problem with an uncertain number of
vehicles. Regarding mathematical formulations, this relates to issues of right-
hand-side (RHS) uncertainty. Note that in mathematical programming, RHS
uncertainty refers to the uncertainty in the values of the right-hand side coeffi-
cients of a linear programming or mixed-integer programming model. To exem-
plify, we may consider RHS uncertainty in the context of SVRPs as well as inte-
grated vehicle and crew/duty scheduling problems and show that these problem
domains may actually be treated in a way that allows for deriving reasonable
solution approaches as well as opening up for future research efforts [8]. The
consideration of an uncertain number of vehicles may not only be motivated by
means of disaster relief issues but also be motivated by means of electric vehicles
having a restricted usable time of the battery, possibly making them unavailable
after some time.

2.2 Generative AI

As mentioned above, generative AI can be a system using AI techniques being

able to generate text, software, or other types of documents in response to ques-
tions or requests; see, e.g., [41]. ChatGPT has been widely discussed in the open
public and, according to popular search engines, there are already thousand
of articles with ChatGPT in their titles or mentioning it in their abstracts as
emphasized in the review from [51]. This also includes quite a few more or less
critical commentaries judging upon challenges and opportunities of ChatGPT;
see, e.g., [16]. In Europe, a High-Level Expert Group on Artificial Intelligence
had presented some Ethics Guidelines for what they call Trustworthy AI [1]. A
collection of possible guidelines at German universities, if available, can be found
in [40].
It seems important to both understand the technology as well as their lim-
itations and even imagine possible misuses. A comprehensive study comparing
human answers to ChatGPT answers can be found in [13]. That is, linguistic
analyses of ChatGPT-generated content compared with that of humans can be
found and is well documented in that source.6 Various areas of research received
commentaries and initial studies regarding the use of ChatGPT including, e.g.,
environmental research [53] and history [20]. The latter study is interesting as
the author claims to see some political bias in what is presented by the system.
While we do not comment on the specific result, we echo the need to be careful
regarding possible misuses of ChatGPT.
The discussion of using generative AI is closely related to cognitive intelli-
gence which is referred to as human mental ability and understanding developed

6
The dataset, code, and models are publicly available at https://github.com/Hello-
SimpleAI/chatgpt-comparison-detection; last access 30 April 2023.
8 S. Voß

through thinking, experiences and senses [29]. It is the ability to generate knowl-
edge by using existing information (as is also a common and well-known issue in
information management; see, e.g., [44]). We do not claim that ChatGPT and
generative AI tools are able towards successfully pursuing this generation, but
the fear of related claims is there.
Finally, we should note that comparative studies of different chatbots may
also be conducted but are still rare; see, e.g., [28]. This is seen in front of what
happened during the last 20 years regarding the use of chatbots especially in
supply chain management where simple tasks such as tracking shipments or
managing inventory are accompanied by related tools. Marrying natural lan-
guage processing and optimization is also on the horizon with first efforts being
on the way [34].

3 Example Explorations
First of all, we can put forward the core focus of our paper to ChatGPT in
asking “Successfully Using ChatGPT in Logistics: Are We There Yet?” leading
to the following answer.

“Some of the potential use cases of ChatGPT in logistics include:

– Customer Support: ChatGPT could assist with addressing customer
queries, tracking shipments, providing order status updates, and
resolving basic support issues.
– Route Optimization: By analyzing data and understanding logistics
constraints, ChatGPT could help in optimizing delivery routes to
improve efficiency and reduce costs.
– Inventory Management: ChatGPT could aid in managing inventory
levels and predicting demand, leading to more effective stock manage-
ment.
– Supply Chain Visibility: By processing vast amounts of data, Chat-
GPT could provide insights into the supply chain, identifying potential
bottlenecks or issues.
– Natural Language Data Entry: ChatGPT could enable users to input
data through natural language, simplifying data entry processes.
– Forecasting and Decision Making: By analyzing historical data, Chat-
GPT could assist in making predictions and suggesting optimal deci-
sions for logistics operations.”
(ChatGPT)

Current studies on ChatGPT are mostly reporting on experiences by asking

questions and providing as well as discussing the answers. This will be surveyed
in Sect. 3.1. Based on this, we perform a somewhat difficult selftest in Sect. 3.2.
 Using ChatGPT in Logistics 9

3.1 Experiences from Literature

Examples from the available literature focusing on logistics issues performing
related tests are [17] and [6], as mentioned above. Note that the latter reference
also includes links to quite a few blog entries and related sources providing
additional insights and discussions. The area of intelligent vehicles may have
a major influence regarding autonomous vehicles in the future. In [5] this area
is considered. Technically relevant questions might be answered in a superficial
way firstly, but once quite a bit of knowledge has been disseminated towards the
system, answers may become more meaningful.
Other examples touching the logistics domain include [46]. Also logistics ser-
vice providers discuss the use of generative AI tools in logistics with [24] being an
example. In [45] the concentration is on specific areas within the public transport
domain.

3.2 A Selftest on a Stochastic Vehicle Routing Problem

Earlier, we already used ChatGPT for providing answers to commonly known
aspects of the VRP. That is, ChatGPT provided a meaningful and practical
characterization of what the VRP is about. Next, we go a few steps further and
elaborate on the SVRP considering an uncertain number of vehicles (with other
input data like demands kept constant). One might assume a lower and an upper
bound on the overall number of available vehicles within a system, where the
specific number of them is uncertain. In this specific problem area the literature
(according to our understanding) should be much more comprehensive than it is.
This seems important regarding the use of a generative AI tool for its treatment
as there are very few references to copy from.
Various ideas for handling robustness with respect to uncertainties on some
specified coefficients in mathematical programming models have been proposed.
The different approaches can roughly be divided into two distinct categories,
depending on whether the underlying uncertainty model refers to possible fluc-
tuations on the row vectors of the constraint matrix (“rowwise uncertainty”), or
on column vectors (we call this “columnwise uncertainty”). RHS uncertainty may
refer to a special case of columnwise uncertainty namely the subclass of robust
models with uncertainty limited to the RHS only. In the settings mentioned
above, and they are also adaptable to the SVRP with an uncertain number of
vehicles, we assume a constraint where the number of vehicles nv is an uncertain
value on the RHS. That is, the number of used vehicles according to the problem
settings is constrained in a way that an uncertainty set is used for nv . A first
attempt can be seen in [8].
Thus, the following prompt is input into the ChatGPT interface: “How to
solve stochastic vehicle routing problems with an uncertain number of vehicles?”
“Solving stochastic vehicle routing problems with an uncertain number of
vehicles is a challenging task, as it involves dealing with both stochastic-
ity and combinatorial optimization. Here are some possible approaches to
tackle this problem:
10 S. Voß

Stochastic Integer Programming: One way to address this problem is to use

stochastic integer programming (SIP) techniques, which combine stochas-
tic optimization with integer programming. SIP models explicitly consider
the stochasticity in the decision-making process and can incorporate differ-
ent sources of uncertainty, including demand, travel times, and the number
of vehicles. To model the uncertain number of vehicles, one can use integer
variables to represent the number of vehicles used in each scenario, and
then solve the problem by minimizing the expected value of the objec-
tive function over all scenarios. SIP models can be solved using various
methods, including scenario generation, sample average approximation,
and stochastic dynamic programming.
Robust Optimization: Another approach is to use robust optimization
techniques to handle the uncertain number of vehicles. Robust optimiza-
tion aims to find solutions that are optimal under a set of possible sce-
narios, by minimizing the worst-case objective function value over all sce-
narios. To model the uncertain number of vehicles, one can use uncertain
parameters that represent the upper and lower bounds on the number of
vehicles, and then solve the problem using robust optimization methods.
Robust optimization can be applied to both deterministic and stochastic
programming problems and has been shown to be effective in dealing with
right-hand-side uncertainty.
Simulation Optimization: A third approach is to use simulation optimiza-
tion techniques, which involve simulating different scenarios and then opti-
mizing the objective function based on the simulated results. To model the
uncertain number of vehicles, one can use a range of possible values for the
number of vehicles and then simulate different scenarios for each value. The
objective function can then be optimized based on the simulated results,
using techniques such as response surface methodology, genetic algorithms,
or simulated annealing.
Some references on solving stochastic vehicle routing problems with an
uncertain number of vehicles include:
Marianov, V., Perez, J. A., & Gutierrez, J. M. (2007). A stochastic vehicle
routing problem with a varying number of vehicles. European Journal of
Operational Research, 183(1), 492–508.
Wang, H., & Ceder, A. (2013). Stochastic programming with a varying
number of vehicles for vehicle routing problems. INFORMS Journal on
Computing, 25(3), 542–557.
Gajpal, Y., & Bansal, A. (2019). An integrated approach for stochastic
vehicle routing problem with uncertain number of vehicles. Computers &
Operations Research, 109, 77–87.
Wu, X., Ouyang, Y., & Yang, Y. (2019). A robust optimization model for
the stochastic vehicle routing problem with uncertain number of vehicles.
Journal of Advanced Transportation, 2019.
These references provide an overview of different approaches to solving
stochastic vehicle routing problems with an uncertain number of vehicles
and offer insights into solution methods and applications.” (ChatGPT)
 Using ChatGPT in Logistics 11

The answer seems good despite the fact that the given references do not seem
to exist. It is an experience.
To be more specific, the mentioned authors exist and have worked on similar
problems but they have not published those specific papers. The titles of those
papers seem to be invented and they cannot be found with popular search engines
(including Google Scholar). The journals exist, but they have other works in the
mentioned issues and page numbers.
Now, let us focus on a specific issue in mathematical programming for solving
the SVRP with an uncertain number of vehicles.
Question: What is right-hand-side uncertainty?

“In mathematical programming and optimization, the right-hand side

(RHS) uncertainty refers to the uncertainty associated with the values
of the parameters or data that appear on the right-hand side of the con-
straints or objective function in a mathematical optimization problem.
In a typical optimization problem, the decision variables are chosen to
minimize or maximize an objective function subject to a set of constraints.
These constraints typically involve parameters or data that are known with
certainty, such as the number of customers, the capacity of a vehicle, or the
duration of a task. However, in many real-world optimization problems,
these parameters may be subject to uncertainty, such as demand fluctua-
tions, unexpected events, or incomplete information.” (ChatGPT)

Question: What is right-hand-side uncertainty in the vehicle routing prob-

lem and how can we solve it?

“Right-hand-side uncertainty is a type of uncertainty that affects the val-

ues of the parameters or data that appear on the right-hand side of the con-
straints or objective function. For example, in a vehicle routing problem,
the demand for a customer may be uncertain, and the capacity constraint
for a vehicle may be affected by this uncertainty.
To deal with right-hand-side uncertainty, several techniques can be used,
such as stochastic programming, robust optimization, and scenario-based
optimization. These techniques involve modeling the uncertainty using
probability distributions or a set of scenarios and incorporating them into
the optimization problem to find a robust or probabilistically optimal solu-
tion.” (ChatGPT)
RHS uncertainty is considered, e.g., in [22,23,30,33]. Asking ChatGPT about
references on this topic in different forms provides references on uncertainty and
alike, but not specifically on RHS uncertainty. It reveals the following:
“These references provide an overview of different aspects of right-hand-
side uncertainty in optimization problems, including robust optimization,
uncertainty sets, and conditional value-at-risk. They cover both theoret-
ical and practical aspects of the problem and offer insights into solution
methods and applications.” (ChatGPT)
12 S. Voß

As we found papers by Minoux [22,23] as being relevant to the topic of RHS

uncertainty, we give that name to ChatGPT asking whether Michel Minoux has
worked on RHS uncertainty.

“Yes, Michel Minoux has worked on right-hand-side uncertainty in math-

ematical programming. He is a well-known expert in the field of optimiza-
tion and has made significant contributions to the theory and applications
of mathematical programming.
In particular, Minoux has contributed to the development of robust opti-
mization methods for handling right-hand-side uncertainty in linear and
nonlinear optimization problems. He has also worked on stochastic pro-
gramming and chance-constrained optimization, which are related to right-
hand-side uncertainty.” (ChatGPT)

Then, however, we get four references that are non-existing (or can not be
found at the places/journals that they are claimed to be cited from). That is,
ChatGPT tends to generate “hallucinated” information beyond the given knowl-
edge; see, e.g., [2] for a related consideration. Hallucinations are factual inaccura-
cies, but they seem to appear as scientifically plausible (at least at first sight). Of
course this is problematic because given references may give a false impression
of existing work that does not exist.
In a yet not comprehensible form, we have worked on RHS uncertainty for
the integrated vehicle and duty/crew scheduling problem. Again, the aim is to
build minimum cost vehicle and daily crew schedules and rosters that cover a
given set of timetabled trips as well as all daily crew duties incorporating a pre-
defined days-off pattern. Details of the problem setting and a modeling approach
can be found in [7,8,21], where the first puts up the overall problem while the
latter extend in various dimensions especially including delay propagation and
uncertainty in the number of vehicles. Those works had been done without using
ChatGPT, and even attempting to let ChatGPT rejuvenate written code in
Python towards solving the problem failed. Though, when we go back to a simple
VRP or even further towards using it to provide code to input instances of
the traveling salesman problem to a commercial solver like CPLEX, we might
succeed.7

4 Conclusions
In this paper, we attempted to investigate the use of a specific generative artificial
intelligence system, widely discussed even in the open public with quite a bit of
media coverage. If generally known issues are considered, ChatGPT, the system
7
Question: Can you write code to input a problem instance of the traveling salesman
problem to CPLEX?
Answer: “Yes, I can provide a code snippet in Python to input a TSP instance
to CPLEX. Here is an example code that reads a TSP instance from a file in the
TSPLIB format, creates a CPLEX model, and solves it: ... ” (ChatGPT).
 Using ChatGPT in Logistics 13

exemplified, can provide meaningful answers. However, it needs quite a bit of

care not to be used without thinking. In addition, there are quite a few ethical
issues that have to be covered as emphasized, e.g., in [14].
Even if our conclusions deem to be close to handwaving (that is, they might be
more specific), they may be used to better understand the issues. Regarding the
specific problem settings considered in this paper, the following lessons may be
learned. Once domain knowledge is available, generative AI tools like ChatGPT
can be used to expedite the writing of a study. Nevertheless, the natural language
processing domain, generative AI and related fields still need to go a long way,
to become useful tools in practice. Once widely known terms can be utilized8 we
can envisage successful use cases. This also holds, once enabling technologies for
automatization are well understood as, e.g., in intelligent transport systems or
smart port developments (see, e.g., [15]). To simplify, in the latter case ChatGPT
may help truck drivers to enter an unknown terminal even in a language that
they usually do not utilize (e.g. supporting the avoidance of disturbances in
executable plans of the above-mentioned dynamic or stochastic vehicle routing
problems). This, however, needs additional work as ChatGPT as an AI language
model currently has no access to real-time data or specific information about
things happening in real time. However, once married with other systems, like
e.g. sensors, or digital twins as virtual representation of a physical system (like
a container terminal) plus related environment and processes the power of those
tools may be utilized. Looking at use cases of digital twins (see, e.g., [26]), this
is part of future research but one may envisage almost immediate application as
technology development seems quite fast.
In general, research issues are related to different language models and related
translation issues (also aiming to compare different tools from different providers
as well as in different hemispheres like the United States and China). Moreover,
the predictive power to enhance given knowledge needs to be improved. Most
importantly, however, the systems should be advanced so that obviously wrong
answers should be (more easily) checked and ruled out. To summarize, we most
probably have some disruptive technology at hand open for further innovation
and research. Beyond the mere of generative AI applications, we may also ask
how combinations with the metaverse may support further advances of knowl-
edge dissemination.
Generative AI tools are becoming popular and one of the future issues relates
to the development of a process model regarding the use of such tools. To accom-
modate such a model, quite some preliminaries need to be clarified to make it
meaningful. This not only relates to the focus of a study but also the reception
level of the respective user. A “handwaving” notion could incorporate a learning
curve which incorporates gaining experience based on issues that are well known
to the user. Given references could be checked at the journals mentioned.

8
Examples might be the use of INCOTERMS in logistics; see, e.g., https://iccwbo.
org/business-solutions/incoterms-rules/incoterms-2020/; last access 30 April 2023.
14 S. Voß

A specific task that might be conducted as future work could be an attempt

to provide ChatGPT with some detailed examples as inputs (including tables
and code).
The general issue of right-hand-side uncertainty needs more research not only
regarding the stochastic vehicle routing problem with an uncertain number of
vehicles. This holds for generic versions of the stochastic vehicle routing problem
as well as specific problems like integrated vehicle and duty or crew scheduling
problems.

References
1. AI HLEG: Ethics Guidelines for Trustworthy AI. European Commis-
sion, Brussels (2019). https://digital-strategy.ec.europa.eu/en/library/ethics-
guidelines-trustworthy-ai. High-Level Expert Group on Artificial Intelligence.
Accessed 22 June 2023
2. Bang, Y., et al.: A multitask, multilingual, multimodal evaluation of ChatGPT on
reasoning, hallucination, and interactivity (2023). https://doi.org/10.48550/arXiv.
2302.04023. Available on Arxiv
3. Berhan, E., Beshah, B., Kitaw, D., Abraham, A.: Stochastic vehicle routing prob-
lem: a literature survey. J. Inf. Knowl. Manag. 13(3), 1450022 (2014). https://doi.
org/10.1142/S0219649214500221
4. Castelli, M., Manzoni, L.: Editorial: generative models in artificial intelligence
and their applications. Appl. Sci. 12(9), 4127 (2022). https://doi.org/10.3390/
app12094127
5. Du, H., et al.: Chat with ChatGPT on intelligent vehicles: an IEEE TIV perspec-
tive. IEEE Trans. Intell. Veh. 8(3), 2020–2026 (2023). https://doi.org/10.1109/
TIV.2023.3253281
6. Frederico, G.F.: ChatGPT in supply chains: initial evidence of applications and
potential research agenda. Logistics 7(2), 26 (2023). https://doi.org/10.3390/
logistics7020026
7. Ge, L., Kliewer, N., Nourmohammadzadeh, A., Voß, S., Xie, L.: Revisiting the
richness of integrated vehicle and crew scheduling. Public Transp. (2022). https://
doi.org/10.1007/s12469-022-00292-6
8. Ge, L., Nourmohammadzadeh, A., Voß, S., Xie, L.: Robust optimization for inte-
grated vehicle and crew scheduling based on uncertainty in the main inputs. In: The
Fifth Data Science Meets Optimisation Workshop at IJCAI-2022, Vienna (2022).
https://sites.google.com/view/ijcai2022dso/. Accessed 30 Apr 2023
9. Ge, L., Voß, S., Xie, L.: Robustness and disturbances in public transport. Public
Transp. 14, 191–261 (2022). https://doi.org/10.1007/s12469-022-00301-8
10. Gendreau, M., Laporte, G., Séguin, R.: Stochastic vehicle routing. Eur. J. Oper.
Res. 88, 3–12 (1996). https://doi.org/10.1016/0377-2217(95)00050-X
11. Golden, B., Raghavan, S., Wasil, E. (eds.): The Vehicle Routing Problem: Latest
Advances and New Challenges. Springer, New York (2008). https://doi.org/10.
1007/978-0-387-77778-8
12. Golden, B., Wang, X., Wasil, E.: The Evolution of the Vehicle Routing Problem.
Springer, Cham (2023). https://doi.org/10.1007/978-3-031-18716-2
13. Guo, B., et al.: How close is ChatGPT to human experts? Comparison corpus, eval-
uation, and detection (2023). https://doi.org/10.48550/arXiv.2301.07597. Avail-
able on Arxiv
 Using ChatGPT in Logistics 15

14. Haman, M., Školnı́k, M.: Using ChatGPT to conduct a literature review. Account.
Res. 1–3 (2023). https://doi.org/10.1080/08989621.2023.2185514
15. Heilig, L., Lalla-Ruiz, E., Voß, S.: Digital transformation in maritime ports: anal-
ysis and a game theoretic framework. NETNOMICS 18, 227–254 (2017). https://
doi.org/10.1007/s11066-017-9122-x
16. Kasneci, E., et al.: ChatGPT for good? On opportunities and challenges of large
language models for education. Learn. Individ. Differ. 103, 102274 (2023). https://
doi.org/10.1016/j.lindif.2023.102274
17. Kim, J., Lee, J.: How does ChatGPT introduce transport problems and solutions
in North America? Findings (2023). https://doi.org/10.32866/001c.72634. Entry
72634
18. Konstantakopoulos, G., Gayialis, S., Kechagias, E.: Vehicle routing problem and
related algorithms for logistics distribution: a literature review and classifica-
tion. Oper. Res. Int. J. 22, 2033–2062 (2022). https://doi.org/10.1007/s12351-
020-00600-7
19. Lin, H.Y.: Large-scale artificial intelligence models. Computer 55(5), 76–80 (2022).
https://doi.org/10.1109/MC.2022.3151419
20. McGee, R.W.: How would American history be different if LBJ had lost the 1948
election? A ChatGPT essay (2023). https://doi.org/10.2139/ssrn.4413418. Avail-
able at SSRN
21. Mesquita, M., Moz, M., Paias, A., Pato, M.: A decomposition approach for the
integrated vehicle-crew-roster problem with days-off pattern. Eur. J. Oper. Res.
229, 318–331 (2013). https://doi.org/10.1016/j.ejor.2013.02.055
22. Minoux, M.: Robust linear programming with right-hand-side uncertainty, duality
and applications. In: Floudas, C., Pardalos, P. (eds.) Encyclopedia of Optimization,
pp. 3317–3327. Springer, Boston (2008). https://doi.org/10.1007/978-0-387-74759-
0 569
23. Minoux, M.: Two-stage robust LP with ellipsoidal right-hand side uncertainty is
NP-hard. Optim. Lett. 6(7), 1463–1475 (2012). https://doi.org/10.1007/s11590-
011-0341-z
24. Möller, P. (ed.): ChatGPT and the Like: Artificial Intelligence in Logis-
tics. DHL (2023). https://dhl-freight-connections.com/en/trends/chatgpt-and-
the-like-artificial-intelligence-in-logistics/. Accessed 10 May 2023
25. Mollick, E.R., Mollick, L.: Using AI to implement effective teaching strategies in
classrooms: five strategies, including prompts (2023). https://doi.org/10.2139/ssrn.
4391243. Available at SSRN
26. Neugebauer, J., Heilig, L., Voß, S.: Digital twins in seaports: current and future
applications. In: Daduna, J., et al. (eds.) ICCL 2023. LNCS, vol. 14239, pp. 202–
218. Springer, Cham (2023)
27. Nourmohammadzadeh, A., Voß, S.: An effective matheuristic approach for robust
bus driver rostering with uncertain daily working hours. In: Daduna, J., et al. (eds.)
ICCL 2023. LNCS, vol. 14239, pp. xx–yy. Springer, Cham (2023)
28. O’Leary, D.E.: An analysis of three chatbots: BlenderBot, ChatGPT and LaMDA.
Intell. Syst. Account. Finance Manag. 30(1), 41–54 (2023). https://doi.org/10.
1002/isaf.1531
29. Otero, I., Salgado, J.F., Moscoso, S.: Cognitive reflection, cognitive intelligence,
and cognitive abilities: a meta-analysis. Intelligence 90, 101614 (2022). https://
doi.org/10.1016/j.intell.2021.101614
30. Ouorou, A.: Robust models for linear programming with uncertain right hand side.
Networks 68(3), 200–211 (2016). https://doi.org/10.1002/net.21693
16 S. Voß

31. Oyola, J., Arntzen, H., Woodruff, D.L.: The stochastic vehicle routing problem, a
literature review, part II: solution methods. EURO J. Transp. Logist. 6(4), 349–388
(2017). https://doi.org/10.1007/s13676-016-0099-7
32. Oyola, J., Arntzen, H., Woodruff, D.L.: The stochastic vehicle routing problem, a
literature review, part I: models. EURO J. Transp. Logist. 7(3), 193–221 (2018).
https://doi.org/10.1007/s13676-016-0100-5
33. Özaltın, O.Y., Prokopyev, O.A., Schaefer, A.J.: The bilevel knapsack problem with
stochastic right-hand sides. Oper. Res. Lett. 38(4), 328–333 (2010). https://doi.
org/10.1016/j.orl.2010.04.005
34. Ramamonjison, R., et al.: NL4Opt competition: formulating optimization problems
based on their natural language descriptions (2023). https://doi.org/10.48550/
arXiv.2303.08233. Available on Arxiv
35. Ritzinger, U., Puchinger, J., Hartl, R.F.: A survey on dynamic and stochastic
vehicle routing problems. Int. J. Prod. Res. 54, 215–231 (2016). https://doi.org/
10.1080/00207543.2015.1043403
36. Sar, K., Ghadimi, P.: A systematic literature review of the vehicle routing problem
in reverse logistics operations. Comput. Ind. Eng. 177, 109011 (2023). https://doi.
org/10.1016/j.cie.2023.109011
37. Savelsbergh, M., Ulmer, M.: Challenges and opportunities in crowdsourced delivery
planning and operations. 4OR 20, 1–21 (2022). https://doi.org/10.1007/s10288-
021-00500-2
38. Shi, Q., Yang, W., Yang, Z.M., Zhao, Q.C.: Relief materials vehicles planning in
natural disasters. IEEE/CAA J. Automatica Sin. 5(2), 595–601 (2018). https://
doi.org/10.1109/JAS.2017.7510850
39. Soeffker, N., Ulmer, M.W., Mattfeld, D.C.: Stochastic dynamic vehicle routing in
the light of prescriptive analytics: a review. Eur. J. Oper. Res. 298, 801–820 (2022).
https://doi.org/10.1016/j.ejor.2021.07.014
40. Solis, T.: Die ChatGPT-Richtlinien der 100 größten deutschen Univer-
sitäten (2023). https://www.scribbr.de/ki-tools-nutzen/chatgpt-universitaere-
richtlinien/. Accessed 22 June 2023
41. Stokel-Walker, C., Van Noorden, R.: The promise and peril of generative AI. Nature
614, 214–216 (2023). https://doi.org/10.1038/d41586-023-00340-6
42. Sun, Y., et al.: ERNIE: enhanced representation through knowledge integration
(2019). https://doi.org/10.48550/arXiv.1904.09223. Available on Arxiv
43. Toth, P., Vigo, D.: Vehicle Routing: Problems, Methods, and Applications, 2nd
edn. SIAM, Philadelphia (2014)
44. Voß, S., Gutenschwager, K.: Informationsmanagement. Springer, Berlin (2001).
https://doi.org/10.1007/978-3-642-56878-7
45. Voß, S.: Bus bunching and bus bridging: what can we learn from generative AI
tools like ChatGPT? Sustainability 15(12, article #9625) (2023). https://doi.org/
10.3390/su15129625
46. Wang, D., Lu, C.T., Fu, Y.: Towards automated urban planning: when genera-
tive and ChatGPT-like AI meets urban planning (2023). https://doi.org/10.48550/
arXiv.2304.03892. Available on Arxiv
47. Wang, F.Y., Yang, J., Wang, X., Li, J., Han, Q.L.: Chat with ChatGPT on Industry
5.0: learning and decision-making for intelligent industries. IEEE/CAA J. Auto-
matica Sin. 10(4), 831–834 (2023). https://doi.org/10.1109/JAS.2023.123552
48. Winston, P.H.: Artificial Intelligence. Addison-Wesley, Boston (1992). (1st edn. in
1977)
 Using ChatGPT in Logistics 17

49. Xiong, C., Chen, X., He, X., Lin, X., Zhang, L.: Agent-based en-route diversion:
dynamic behavioral responses and network performance represented by macro-
scopic fundamental diagrams. Transp. Res. Part C Emerg. Technol. 64, 148–163
(2016). https://doi.org/10.1016/j.trc.2015.04.008
50. Yang, Z.: The bearable mediocrity of Baidu’s ChatGPT competitor. MIT Tech-
nology Review (2023). https://www.technologyreview.com/2023/03/22/1070154/
baidu-ernie-bot-chatgpt-reputation/. Accessed 10 May 2023
51. Zhang, C., et al.: One small step for generative AI, one giant leap for AGI: a com-
plete survey on ChatGPT in AIGC era (2023). https://doi.org/10.48550/arXiv.
2304.06488. Available on Arxiv
52. Zheng, O., Abdel-Aty, M., Wang, D., Wang, Z., Ding, S.: ChatGPT is on the
horizon: could a large language model be all we need for intelligent transportation?
(2023). https://doi.org/10.48550/arXiv.2303.05382. Available on Arxiv
53. Zhu, J.J., Jiang, J., Yang, M., Ren, Z.J.: ChatGPT and environmental research.
Environ. Sci. Technol. (2023). https://doi.org/10.1021/acs.est.3c01818
 When Routing Meets Recommendation:
Solving Dynamic Order
Recommendations Problem
in Peer-to-Peer Logistics Platforms

Zhiqin Zhang1 , Waldy Joe1 , Yuyang Er2 , and Hoong Chuin Lau1(B)
1
School of Computing and Information Systems, Singapore Management University,
Singapore, Singapore
{zqzhang.2020,waldy.joe.2018}@phdcs.smu.edu.sg, [email protected]
2
AI Singapore, Singapore, Singapore
[email protected]

Abstract. Peer-to-Peer (P2P) logistics platforms, unlike traditional

last-mile logistics providers, do not have dedicated delivery resources
(both vehicles and drivers). Thus, the efficiency of such operating model
lies in the successful matching of demand and supply, i.e., how to match
the delivery tasks with suitable drivers that will result in successful
assignment and completion of the tasks. We consider a Same-Day Deliv-
ery Problem (SDDP) involving a P2P logistics platform where new orders
arrive dynamically and the platform operator needs to generate a list
of recommended orders to the crowdsourced drivers. We formulate this
problem as a Dynamic Order Recommendations Problem (DORP). This
problem is essentially a combination of a user recommendation prob-
lem and a Dynamic Pickup and Delivery Problem (DPDP) where the
order recommendations need to take into account both the drivers’ pref-
erence and platform’s profitability which is traditionally measured by
how good the delivery routes are. To solve this problem, we propose
an adaptive recommendation heuristic that incorporates Reinforcement
Learning (RL) to learn the parameter selection policy within the heuris-
tic and eXtreme Deep Factorization Machine (xDeepFM) to predict the
order-driver interactions. Using real-world datasets, we conduct a series
of ablation studies to ascertain the effectiveness of our adaptive approach
and evaluate our approach against three baselines - a heuristic based on
routing cost, a dispatching algorithm solely based on the recommenda-
tion model and one based on a non-adaptive version of our proposed
recommendation heuristic - and show experimentally that our approach
outperforms all of them.

Keywords: Crowdsourced delivery · Data-driven optimization ·

Recommendations system

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 18–35, 2023.
https://doi.org/10.1007/978-3-031-43612-3_2
 When Routing Meets Recommendation 19

1 Introduction
Peer-to-Peer (P2P) transportation platforms, such as Uber, Lyft, and Food-
Panda, have been gaining popularity in recent years with greater adoption of
internet technology (such as e-commerce and mobile apps) and the growing
demand for more efficient urban mobility. These platforms include ride-hailing
or ride-sharing and crowdsourced delivery service providers where the former
provide transportation services for people while the latter for goods. The key
feature of such platforms is that the platform owners do not own the physical
assets (fleet of vehicles) nor employ the transport operators (drivers or riders,
used interchangeably) unlike the traditional taxi companies or logistics service
providers. Thus, the efficiency of such P2P transportation model lies in the suc-
cessful matching of demand and supply, i.e., how to match the delivery tasks
(or orders or jobs, used interchangeably) with the appropriate drivers that will
result in successful assignment and completion of the jobs. This is an interesting
research challenge as unlike in the traditional routing problems, the drivers have
a choice to accept or reject the jobs.
The work presented in this paper is motivated by a real-world problem sce-
nario involving a P2P logistics platform uParcel1 , where new delivery tasks are
being recommended throughout the day in the form of a ranked list (also known
as menu) to individual crowdsourced drivers who have already existing pre-
scheduled delivery tasks. This problem is essentially a combination of a user rec-
ommendation problem [23] and a Same-Day Delivery Problem (SDDP) which is
a variant of a Dynamic Vehicle Routing Problem (DVRP) [5]. For a successful
matching of delivery tasks to drivers, the recommendations need to take into
account not only the platform’s profitability (which in classical SDDP is deter-
mined by how good the delivery routes are), but also the drivers’ preferences for
the tasks recommended.
Both, the user recommendation problem and the SDDP have been widely
studied and researched in their respective fields. However, there is a great poten-
tial for synergy between the two research domains in addressing the problem pre-
sented in this paper. The traditional Recommender Systems (RS) are only able
to capture users’ general personal preferences [29] while the Sequential Recom-
mender Systems (SRS) go one step further by taking into account the sequential
dependencies of user-item interactions for more accurate recommendation [7].
Although an SRS takes into account the existing state of a user, it is only able
to recommend the user’s next action, i.e., in the context of e-commerce, the item
that he or she will purchase next. This is insufficient in addressing the problem
in this paper which is akin to recommending a new product given that users
have purchased a list of items (pre-scheduled orders) and that they are planning
to purchase another list of items in the future (new orders that arrive dynami-
cally). Meanwhile, current works in solving the SDDP assume that the drivers

1
uParcel is a Singapore start-up company which offers on-demand delivery and courier
services for business and consumers. See https://www.uparcel.sg/ (last access date
02 July 2023).
20 Z. Zhang et al.

will always accept the assigned delivery tasks [27,28], which is not the case in a
crowdsourced setting.
To address this gap, we bring together two key ideas from Reinforcement
Learning and Recommender Systems. More precisely, we propose an adaptive
recommendation heuristic (ARH) that incorporates reinforcement learning (RL)
to learn the parameter selection policy within the heuristic and a RS model called
xDeepFM [15] to predict the driver’s probability of accepting an order based on
their preference. ARH relies on a linear scalarization function that takes into
account both routing information and order-driver interaction in generating the
recommendation, and the RL-trained policy is utilized to adaptively determine
the exact weightages of this linear function at a given pre-decision state.

2 Related Works

Same-Day Delivery Problem. There are two broad categories of approaches for
solving SDDP or DVRP in general namely offline or pre-processsed decision
support and online decision [24]. In offline approaches, policies or values for
decision-making are computed prior to the execution of a plan. Here, the problem
is usually formulated and solved as an Markov Decision Process (MDP). Unfor-
tunately, MDP-based approaches (specifically tabular-based ones) fall into the
curse of dimensionality and hence are not suitable for most real-world problems
[19]. Approximate Dynamic Programming (ADP) approaches [20] are commonly
used to tackle the scalability issue, and one such ADP method for the DVRP is
Approximate Value Iteration (AVI) [1]. Increasingly, there also have been many
recent works that addressed to solve the DVRP using RL [6,8,13,14]. Mean-
while, lookahead approaches have been applied successfully to solve dynamic
and stochastic routing problems. These are commonly termed as rollout algo-
rithms, e.g., [25], the pilot method, e.g., [17] and Multiple Scenario Approach
(MSA), e.g., [4,28]. The approaches proposed in these works are not directly
applicable to our problem since we are dealing with crowdsourced drivers; how-
ever, like many current works that use RL to solve the DVRP, we utilize RL to
learn a policy to guide our recommendation heuristic.

Recommender System. There are two broad approaches to recommender sys-

tems, namely collaborative filtering methods and content-based methods. Col-
laborative filtering methods rely on past-recorded interactions between users and
predict through similar user patterns to produce new recommendations. One
such example is Factorization Machines (FM) [22]. Meanwhile, content-based
methods use item features to recommend other items similar to what the user
likes and based on their previous action or explicit feedback. One such example
is linear models with Follow-the-Regularized-Leader (FTRL) [16]. Lately, Deep
Neural Networks have been utilised to learn high-order interactions [15]. To com-
bine the strength of a wide linear model (i.e., Generalized Linear Models) and
a deep neural network for recommender systems, Google proposed a framework
called Wide & Deep [9]. Subsequently a wide and deep architecture of DeepFM
 When Routing Meets Recommendation 21

that integrates the Factorization Machine (the wide component) and the Multi-
Layer Perceptron (the deep component) was proposed in [11]. xDeepFM, an
extension of the DeepFM that can jointly model explicit and implicit feature
interactions was proposed in [15]. In this paper, we select xDeepFM to predict
drivers’ probability of selecting an order as it combines both content-based and
content filtering (hybrid approach) and mitigates the cold start and data sparsity
issues by relying on the factorization of the sparse user-item matrix.

Dynamic Matching Problem. [3] introduced the concept of dynamic order match-
ing in a Peer-to-Peer (P2P) logistics platform, which shares similarities with the
problem addressed in this paper. The authors framed the problem as a two-
stage decision problem and proposed a multiple scenario approach that involved
sampling various driver selection scenarios and solving an integer program for
each scenario to generate the final menu. However, it should be noted that the
authors of [3] considered a simplified version of the DORP where drivers are
only allowed to select one order, and no time windows or capacity constraints
are imposed. Furthermore, the authors modeled the drivers’ preferences using a
pre-determined utility function. In contrast, the problem addressed in this paper
takes into account real operational constraints and incorporates drivers’ prefer-
ences based on historical data, making it a more realistic and comprehensive
representation of the DORP.

3 Problem Description and Model

3.1 Problem Description

We assume that dynamic orders are generated throughout the planning horizon,
and at a predetermined frequency, the platform consolidates the newly arrived
dynamic orders and any unassigned orders and generates a personalized order
recommendation list for the crowdsourced drivers. The dynamic orders consist
of pickup and delivery tasks, order size, time window requirements for both
pickup and delivery, and an expiration time. If a driver arrives at the pickup
or delivery location earlier than the specified time window, a waiting time will
be incurred. In the event that no driver accepts an order and arrives at the
pickup location after the specified time limit, the order will be automatically
canceled. A simple illustration of the proposed problem can be found in the
supplementary materials2 .
The objective of the problem is to maximize the number of orders fulfilled
within the given planning horizon, specifically, a full working day. This objec-
tive can be achieved by ensuring that the order recommendation list considers
both the driver’s preference and routing considerations. The driver’s preference
ensures that they have a high probability of selecting at least one of the recom-
mended orders, while the routing considerations ensure that drivers with optimal
delivery routes can fulfill more orders while minimizing delivery costs.
2
https://anonymous.4open.science/r/iccl2023-7ADC/SupplementaryMaterials.pdf.
22 Z. Zhang et al.

Fig. 1. Sequential decision process of a DORP.

3.2 Model Formulation

This problem can be modeled as a sequential decision process with stochastic
information [20]. A visual depiction of the entire decision process is presented in
Fig. 1. Here, we present an MDP formulation of the problem.
Decision Epoch. The decision epoch in our model is time-based. We discretize
the planning horizon (e.g., a day) into multiple time periods. Here, we use k =
1, 2. . . , i, i + 1. . . , n to represent each decision point.
State. As shown in Fig. 1, a state consists of two parts, pre-decision state Sk and
post-decision state Skxk . Both states are represented by the same tuple dk , ok 
where dk denotes the list of drivers and ok denotes the list of unassigned orders.

The list of drivers can be further broken down as follows: dk = d1k , . . . , dm k
k
,
where mk denotes the number of the active drivers at decision epoch k. We
assume the drivers are active when they arrive at pickup or delivery nodes within
this decision time slot. Each element in dk consists of the following features: the
last reported location, vehicle capacity and the route of the remaining  orders.
Similarly, the list of unassigned orders is denoted as ok = o1k , . . . , onk k , where nk
denotes the number of the unsigned orders at decision epoch k. Each element in
ok consists of the following features: pickup and delivery locations, time windows
for pickup and delivery, and size of the order.
Action/Decision. The decision in this model refers to the action of generating
personalized order recommendation lists for each driver. Although in reality,
the platform needs to execute an assignment action once drivers make their
selections (see Fig. 1), we assume that this assignment is governed by a pre-
determined rule and we treatthis step as part of the environment. We denote
our decision variable as xk = x1k , . . . , xm
k
k
, where mk still denotes the number
of the riders available in this time slot. Each element xik in xk contains a subset
of unassigned orders a sorted in descending order of suitability for driver i. Note
that each order may be recommended to more than one driver and each driver
can have more than one recommended order.
Transition. There are two main transitions in this model. Firstly, the transi-
xk−1
tion occurs from the previous post-decision state, denoted as Sk−1 , to the pre-
decision state, represented as Sk . This transition is triggered by the realization
of new dynamic orders, denoted as ωk . The second transition takes place from
 When Routing Meets Recommendation 23

the pre-decision state to the post-decision state, which occurs after executing
the action xk . However, this second transition introduces additional complexity
due to multiple uncertainties that influence the outcome of the action taken.
Specifically, there are two intermediate states that arise from two sources of
uncertainty, namely the selection behaviors of the drivers and the assignment
rule.
Reward/Objective Function. The reward R(Sk , xk ) of an action xk is defined
as the total number of orders fulfilled for the decision epoch k. The solution to
this Markov Decision Process (MDP) is a policy that generates the menus at
each state. The optimal solution is the policy that maximizes the current reward
and the total expected rewards for future states.

4 Solution Approach
Generating the personalized order recommendation lists directly based on the
model formulation presented in the previous section will involve an enumeration
of all possible permutations of unassigned orders of variable length for each
driver. Even if we limit the length of the lists, the action space will still be
combinatorial. Thus, to address the curse of dimensionality, we design a heuristic
approach to generate the recommendation lists.
The remaining part of this section is organised as follows. We first introduce
how our proposed ARH generates the recommendation lists. We then explain
the two key components within the ARH, namely how RL is used to train the
weightage selection policy for this heuristic; and how the drivers’ preference is
modelled and learnt from historical data by using xDeepFM.

4.1 Adaptive Recommendation Heuristic

In DORP, the state space and action space are too large to enumerate for large-
scale problems. Compared with most dynamic same-day delivery problems such
as [10,12] where a decision involves the assignment of one dynamic order to
one rider, our problem involves a multi-order multi-driver matching, which leads
to a very large state space. In addition, the action space also faces the “curse
of dimensionality”. Unlike the action space in [3], the recommendation lists for
riders in DORP need to be sorted by the relevance (rider-order relevance) which
contributes to a larger action space. This task is even made more challenging
as decisions need to be done in real-time. Thus, we propose a heuristic solution
approach to aggregate the action space.
Heuristic Scoring Function. In our heuristic, the key idea is to compute a
driver-order pair score based on the routing cost and riders’ personal preference.
For a given rider-order pair score, the higher this score is, the more suitable
for this rider to serve this order by our consideration. The recommendation list
for each rider is then generated by these scores. The function to calculate the
driver-order score is given as follows:
24 Z. Zhang et al.

score(di , oj ) = αH1 + (1 − α)H2 (1)

where α and 1 − α are the weights on H1 and H2 , respectively. The value H1
is the estimated cost for inserting this order to this rider’s route, calculated
by the distance of the pickup location of this order to the nearest location of
the remaining orders carrying by this rider (referred to as Dn ) and normalized
between 0 and 1:
Dmax − Dn
H1 = (2)
Dmax
where Dmax is the largest distance value for any two locations in a given problem
instance.
The value of H2 is the probability of the prediction for this rider’s preference
for this order, and computed using our proposed xDeepFM model where the
details will be presented in Sect. 4.3.
Balancing Platform Objective and Driver Preferences. As a recom-
mender system, the challenge is in calibrating the values of α adaptively over
decision epochs, as the environment changes, in order to ensure good profitability
(jobs served) on one hand, and good acceptance rate by drivers on the recom-
mended jobs on the other.
From the point of view of the logistics platform, the efficiency of the dynamic
routing plan hinges on timeliness of deliveries which in turn is measured by the
physical measures such as the routing cost, as proposed in the SDDP and DVRP
literature. On the other hand, from the drivers’ point of view, they would like to
serve orders of their preferences. For instance, the drivers may be more willing to
serve the orders in the region which they are more familiar with, even though the
locations may be a little far away from their current location and/or the locations
of their pre-scheduled orders. If the heuristic generates the menu based on the
preferences of the riders, it will not only improve the acceptance of these orders
but also the user experience for drivers.
To achieve a good balance of these two potentially conflicting measures, we
propose our ARH. The key idea is that we dynamically adjust the weights for
each selection criterion based on the current state when making the decision.
Intuitively, if the delivery resources are insufficient in some time period, the
heuristic should assign orders to focus on the operational efficiency. Otherwise,
we want to recommend preferred options to riders, hoping that the orders rec-
ommended are more likely accepted by riders (according to their personal prefer-
ence). In the following, we present a Reinforcement Learning approach to obtain
the parameter values (weights) for balancing these two selection criteria.
Finally, we present our heuristic approach, ARH, to calculate driver-order
scores that take into account both driver preferences and routing costs at every
decision epoch.
ARH consists of the following components:
– Driver’s predicted preference score to represent the drivers’ selection
behaviour based on historical data.
 When Routing Meets Recommendation 25

Fig. 2. Overview of the proposed Adaptive Recommendation Heuristic approach.

– A routing heuristic to calculate the cost for a given driver to serve a given
order.
– RL-trained policy to determine the weights attached to each objective at a
given pre-decision state.

Our proposed heuristic is based on a Policy Function Approximation (PFA)

approach [21]. Traditional PFAs are analytical functions by which information
in the state variable is mapped directly to a decision. The information of the
state in this problem includes two perspectives: platform and drivers. This PFA
applies Eq. 1 to calculate the pair scores to generate the final decision, namely
the recommendation lists.
The overview of ARH can be found in Fig. 2. The two hollow arrows represent
the input of all orders and drivers’ data into the Deep Q-Network (DQN), while
the other solid arrows signify the transfer of a single value. More details about
the algorithm showing how the ARH generates the recommendation lists in a
given decision epoch can be found in supplementary materials.

4.2 RL for Selection in ARH

In order to select a proper heuristic for a given decision epoch, we propose an RL

approach. Before we elaborate this approach, we describe the MDP formulation
for the new decision problem after introducing our proposed heuristic.
MDP Reformulation with State and Action Space Aggregation. The
state of this new MDP is an aggregated state. Then we use handcrafted features
to calculate the rider-order score. To condense the new state to a set of features,
we select the following state information for our new MDP:

– Current time t for this decision epoch k. As the authors of [6,26] select the
current point of time of this decision epoch, we also include this feature in
our representation of state.
26 Z. Zhang et al.

– Total number of unassigned orders nk and active drivers mk in this time

slot (or decision epoch). We select these two features because it can help to
capture how dense the supply/demand in the current time slot is. Combined
with the current time, these two features can provide the information to learn
the policy for making decisions not only on the current decision epoch but
also looking further in terms of supply/demand ratio.
– Information about riders. We do not need all the rider’s information in the
original MDP. Here we consider to use the remaining orders (or carrying
orders) of riders. Based on the remaining orders, we can calculate the distance
and time-related information of a given new order to those remaining orders.
These are main features to capture the power of supply in the near future
time slots. For example, for driver i, the committed delivery time of carrying
orders is denoted as vector toi . Then the average bottleneck time for all
drivers tb can be calculated by
mk
1 
tb = (toi −∞ − t) (3)
mk i=1
where the negative infinite norm can give the earliest committed delivery time
of all the on-going orders of a given driver.
– Information about orders. Apart from the total number of the unassigned
orders to calculate the ratio of supply/demand, we also consider average
remaining time to obtain the information of the demand. To be more specific,
the average remaining time tr is the average of the differences between the
expired time tej for a given order j and the current time t.
nk
1 
tr = (tej − t) (4)
nk j=1

The action is to select one heuristic to calculate the rider-order pair score for
this current decision epoch. To be more specific, we use five parameters in our
proposed heuristic. These five values of α are 0, 0,25, 0.5, 0.75 and 1, respectively.
The transition and reward function remain the same as the original MDP.
Then, we formulate this problem and use a RL approach to train a policy to
select a proper heuristic to calculate the rider-order score in each decision epoch.
By these scores, we can directly generate the personalized recommendation lists
to riders.
Deep Q-Learning for Weight Selection. For a given state Sk , the action xk is
to select the most proper heuristic to generate the personalized recommendation
lists for riders. Thus, the solution of the DORP is to learn a policy π ∈ Π that
selects the action for each state. The optimal solution π∗ can maximize the total
reward. Q-learning [30] learns a value Q(Sk , x) for each state and each action
pair. This Q-value can estimate the immediate reward plus the expected future
rewards if the action x is taken to the state Sk . Definitely, we cannot calculate
and learn all these state-action pairs due to the size of the state and action space
(even after aggregation). Thus, we use the DQN approach to estimate this value.
 When Routing Meets Recommendation 27

We applied the DQN with experience replay approach in [18]. To calculate

the loss function to train this Q-network, the reward function is defined as the
proportion of accepted jobs in a specific decision epoch (time slot). The hyper-
parameters we used can be seen in the supplementary materials.

4.3 xDeepFM for Preference Prediction

In this subsection, we introduce xDeepFM to predict the job-driver preference
score which will be used by the ARH presented above.
Similar to a typical RS problem, the task is to recommend delivery jobs to
drivers. The role of the xDeepFM model is to generate the value of H2 , a score
for each driver-job pair that indicates the probability that driver dm k
k will accept
nk
job ok . This probability will subsequently be used by the ARH to take into
account the driver’s preference in selecting a job.
Architecture. The architecture of xDeepFM can be broken down into four
parts: (1) embedding layer, (2) linear/FM using the raw input features, (3) Deep
Neural Network (DNN) using the dense feature embeddings, and (4) Compressed
Interaction Network (CIN) using the dense feature embeddings.
The output of parts (2) to (4) jointly contribute to a shared sigmoid out-
put. The embedding layer takes in the FM inputs and represents them as
lower-dimensional vectors. Linear/FM uses matrix factorization to learn the low-
dimensional representations and then learns the linear regression weights for the
FM layer. For higher-order interactions, these are captured by DNN and CIN
where DNN learns implicit high-order feature interactions at the bit-wise level
while CIN learns explicit high-order feature interactions fashion at the vector-
wise manner. CIN and plain DNNs can complement each other to make the model
stronger by combining these two structures. In addition, xDeepFM is customiz-
able as it can be configured as a classical FM model by enabling the linear part
and the FM part only if the dataset does not require high-order feature inter-
actions. The architecture of xDeepFM can be seen in supplementary materials.
See [15] for more details.
Feature Engineering. The features used for training the xDeepRM model are
presented in supplementary materials.
In DORP we face two key challenges involved in the learning task, namely,
data sparsity and highly dynamic input features. The data is sparse because the
platform only records positive samples, i.e., the accepted orders but not orders
that are rejected. In addition, some important input features such as the current
location of the drivers and the current capacity of the vehicle are highly dynamic
and play an important role in determining whether a driver will accept or reject
an order. We address these challenges as follows.
Distance-Based Feature. The current location of a driver plays an important
role in determining whether he or she will accept an order. Thus, inclusion of
distance-based features would improve the model’s performance. These features
capture the distance from the driver’s current location to the pickup point and
28 Z. Zhang et al.

from the pickup point to the dropoff point. The challenge is the former. Unlike in
typical RS, this feature is highly dynamic. The main challenge lies in calculating
this value in real time as there could be some error in calculating expected versus
actual values since such highly dynamic data may not be available or accurate.
If it is not handled properly, these errors could accumulate over time, resulting
in incorrect predictions. To overcome this, we propose to use a proxy feature,
i.e., last known location to represent the current location of the driver in order
to calculate the distance of the current location to the pickup location. During
training, the last known location of a driver is defined as the location at the time
point when jobs have last been accepted by this driver.
Negative Sampling. The given dataset consists of historical jobs accepted by
individual drivers (item features), as well as a list of drivers (user features). For
training the model, the accepted jobs form the positive training samples. As the
platform does not allow drivers to reject jobs, all non-accepted jobs are auto-
matically considered implicit negative samples. Hence most samples recorded
are positive use cases. Thus, implicit negative samples are generated to ensure
a more balanced data during model training. However, not all negative samples
are useful and they may in turn introduce noise, hence negative sampling was
used to decrease the amount of negative use cases. The chosen approach is by
filtering within a time window and computing distance relative to the driver’s
most recent job/known location. We made the assumption that an active driver
who accepted order(s) or performing a job at a given time window is likely to
have seen and “rejected” the non-accepted jobs at the same time window.

5 Experiments
We evaluate our proposed approach using a real-world dataset that we collected
from a local logistics platform. We ensure that the problem settings and input
data closely resemble the real-world scenario.

5.1 Benchmark Algorithms

We present three baseline algorithms as benchmarks: one based solely on rout-
ing cost, one based solely on the driver’s preference, and one that is a linear
combination of these two aspects with equal weightage.
– RH. The first baseline algorithm is a routing heuristic (RH), which is simi-
lar to the order dispatching algorithm which uParcel currently adopts. This
algorithm only considers the distance between the driver’s current location
and the new order location. This can be done by setting the α = 1 in Eq. 1.
– DP. The second baseline algorithm is one which solely relies on the driver’s
preference (DP). This is similar to a typical recommender system which learns
driver-order score to represent the driver’s preference over a certain order.
Section 4.3 describes how this preference score is computed. Then, we set the
α = 0 in Eq. 1 to ensure that the order dispatching is solely based on the
drivers’ preference.
 When Routing Meets Recommendation 29

– FWH. The third baseline algorithm is considering both aspects in a fixed

weight. This is done by setting α in Eq. 1 as fixed value, (e.g. setting it as 0.5
as the third baseline algorithm). Thus, this model is a fixed weight heuristic
(FWH) model where we assume a ’balanced’ consideration.

5.2 Experiment Design

Data. In this experiment, we use a total of 894,794 rows of pickup-delivery tasks
collected from 2021 to 2022 provided by uParcel. To train the xDeepFM model
and the DQN, we generate two datasets from the raw data. More details about
the dataset can be found in the supplementary materials.

xDeepFM Module. As mentioned in Sect. 4.3, we use the xDeepFM model

to predict preference of the drivers. Here, we describe the detailed of our imple-
mentation of this model.
xDeepFM Training. The challenge of training the xDeepFM in this problem con-
text lies in the sparsity of data and the dynamicity of the problem. Unlike in the
typical RS, recommendation or preference scores need to be computed quickly and
sometimes the input features change dynamically and may not be available at the
moment of inference. This is because xDeepFM is customizable depending on the
nature of the data used. We conducted an ablation study to ascertain which cus-
tomized architectures of xDeepFM are effective in predicting drivers’ preference.
In addition, given that the current location of a driver plays an important role in
determining how likely an order will be accepted, we include a distance feature to
act as a proxy to determine how far the driver is from the order. Here, we use a
proxy of the driver’s real location because in practice, the exact location of a driver
may not be available or be updated in the system due to the highly dynamic nature
of this feature. See Table 1 for the list of customized models evaluated.
Model and Training Setup. The xDeepFM model that we implemented includes
a DNN with 4 layers and a CIN with 3 layers. We use optuna [2] to tune the
number of nodes in each layer. Apart from these hyperparameters, the training
epochs is set as 500, the batch size is 128 and the learning rate is 0.001. To train
the model for binary classifications, we use the following loss function:
N
1 
L=− yi log yˆi + (1 − yi )log(1 − yˆi ) (5)
N i=1

where N is the number of total training instances, yi and yˆi is predicted value 0
or 1 (accept or not) and the related label.
Metrics. To evaluate the performance of the various baseline models, we use
Precision@k and Recall@k where k refers to the top-k order recommendations
given to a driver. To be more specific, the precision@k is defined as the number
of recommended orders at top-k that are accepted by a given driver divided by
k which is the number of top-k orders the model recommended to this driver.
# of recommended orders @k that are accepted
P recision@k = (6)
# of recommended orders @k
30 Z. Zhang et al.

The recall@k is the proportion of accepted orders found in the top-k recom-
mendations.
# of recommended orders @k that are acceptd
Recall@k = (7)
total # of accepted orders
We choose k = 3 and 5 as these are realistic lengths of a recommendation list
that are both not too long (difficult to make decision) but at the same time give
drivers enough options to choose.

RL Module. As mentioned in Sect. 4.2, we train a DQN to select the weights

for our proposed solution method, Adaptive Recommendation Heuristic. Here,
we show the details related to this module.
RL Environment. To conduct the experiments, we build a simulator for the
DORP environment which simulates the selection behavior of riders. Here, we
use a multi-attribute utility model similar to the one proposed in [3]. There are
two attributes in this multi-attribute utility model namely detoured distance and
drivers’ preference, which are two considerations in decisions for our proposed
DORP. In addition, we also introduce a discount factor to reflect the position of
the recommended order in the list:
vi,j = γp (δ1 Di,j + δ2 Pi,j ) (8)
where vi,j is the utility value of order j to driver i. γp is the position discount
factor. This number from large to small represents the position of the order
appearing in the menu from top to bottom. δ1 and δ2 are two weights of the two
attributes mentioned above and they sum up to one. To capture the difference
of the driver selecting behavior, we uniformly randomize these two figures from
driver to driver. Di,j is the estimated detoured distance for driver i to serve
order j. Here we use the H1 value to estimate the routing cost value. Pi,j is the
predicted probability that driver i will accept order j, which can be estimated
by our trained xDFM model. A high value for vi.j indicates high willingness of
driver i to serve the order j. To conduct the experiments, we also borrow the
idea from [3] to set a threshold (0.5 in this environment to get a reasonable
results by the similar current algorithm implemented in the logistics platform).
Other experiment setups can be seen in the following paragraph.
Experiment Setup. The hyperparameters of the DQN can be found in
supplementary materials. We train the DQN for ARH on the two-month dataset
and test the performance of the ARH on the two-week dataset. More details
related to the dataset can be found in supplementary materials. We define each
day (or each instance) as an “episode” in the training and testing because the
problem we consider is same-day delivery. We split each episode into 144 time
steps (or decision epochs), where each step is 10 min.
Metrics. We evaluate our proposed ARH against the three baseline algorithms
based on the test dataset and we use the order fulfill rate to compare the per-
formance. The order fulfill rate can be calculated by:
# canceled order
order f ulf ill rate = 1 − (9)
# total order
 When Routing Meets Recommendation 31

Table 1. xDeepFM model with additional distance features outperforms the other
baseline models in both precision and recall measures.

Model Precision@3 Precision@5 Recall@3 Recall@5

FM only 0.431 0.436 0.109 0.110
FM + DNN 0.469 0.489 0.335 0.336
FM + CIN 0.468 0.478 0.095 0.095
FM + CIN + DNN 0.491 0.506 0.336 0.324
FM + CIN + DNN
w/distance 0.557 0.578 0.352 0.349

5.3 Experiment Results and Discussion

Firstly, we show the evaluation results of the prediction model. Then, we compare
the results of our ARH method with three baseline models described in Sect. 5.1.
xDeepFM. Figure 3 visualizes the convergence of the models with different
components. In each graph, the horizontal axis is the epochs, while the vertical
axis is the log loss. The blue and orange lines show the loss in the training data
set and test data set, respectively. From these graphs, we can find that the FM
model cannot predict a good result as the loss on test data set converge to about
1.25 to 1.5 after training. If introducing the DNN, the test loss of the prediction
model will reduce to about 0.6 but it is not stable, comparing with the FM and
the CIN model. The final prediction model, xDeepFM with all components can
get the best results as the test loss converge to the smallest value.
Table 1 also shows the results that xDeepFM with the full suite of components
outperforms the classical FM or FM with the subsets of the components. In
addition, inclusion of distance-based feature improves the prediction.

Fig. 3. Convergence of different predic- Fig. 4. Summary of the performance of

tion models. (Color figure online) different algorithms.
32 Z. Zhang et al.

Table 2. Experiments results of test data set.

Instance RH DP FWH ARH

Instance 1 (912) 93.97% 84.87% 86.51% 93.86%
Instance 2 (383) 97.91% 98.17% 98.43% 98.69%
Instance 3 (779) 89.47% 83.44% 85.49% 95.25%
Instance 4 (467) 79.87% 72.38% 76.45% 90.15%
Instance 5 (453) 100% 100% 100% 100%
Instance 6 (435) 98.39% 94.48% 96.55% 98.62%
Instance 7 (1065) 85.45% 78.87% 80.01% 91.27%
Instance 8 (920) 94.02% 86.09% 87.61% 97.83%
Instance 9 (887) 93.46% 86.70% 87.15% 94.59%
Instance 10 (752) 99.20% 89.89% 90.69% 99.34%
Instance 11 (782 93.61% 86.57% 90.41% 97.31%
Instance 12 (384) 97.14% 97.14% 97.14% 96.89%
Instance 13 (440) 90.68% 87.73% 90.23% 90.23%
Instance 14 (424) 83.19% 86.56% 86.08% 84.67%
Avg. 92.59% 88.06 89.49% 94.91%
Var. 0.39% 0.57% 0.47% 0.20%

ARH. After having trained 200 episodes (one episode includes 144 epochs) on
the training dataset, our proposed ARH with the trained DQN policy can achieve
better results on most test datasets. Table 2 shows the results of the performance
of our proposed ARH and three baseline algorithms. The entries are the order
fulfilled rate defined as Eq. 9, so larger numbers represent better performance.
The first column is instances in the two weeks data set. The number in the brack-
ets represents the total number of the dynamic orders in one instance. Columns
two to four are the results of three baseline models described in Sect. 5.1. The
fifth column shows the best results among these three baselines. The last column
shows the performance of our proposed ARH. From this table, we can observe
that our proposed ARH can get the best results in 10 instances. ARH is the
worst only in instance 12 with a slight difference. For the remaining instances,
the ARH can outperform some of the baselines. As the statistical results shows,
the average order fulfill rate of the ARH on the test dataset is the highest and
the its variance is the smallest.
Apart from the order fulfill rate, we present a summary of the comparison
on two other metrics (namely, the total routing cost and the driver preference
measured by the probability of the driver accepting the order recommended to
him/her based on the driver’s historical data). Figure 4 shows that for rout-
ing cost, RH (i.e., solely considering routing cost) achieves the best perfor-
mance compared with the other three algorithms (which is expected). Similarly,
DP performs best in terms of driver preference compared with the other three
 When Routing Meets Recommendation 33

approaches. Interestingly, our approach ARH achieves the best performance in

terms of number of orders fulfilled, while not compromising by a large margin
in terms of the other two conflicting metrics.
All in all, the results demonstrate that our proposed framework effectively
dispatches more dynamic orders for P2P platforms, thereby enhancing customer
satisfaction for both riders and customers. This positive impact translates into
increased profitability for the platform in both the short and long term.

Acknowledgements. This research project is supported by the National Research

Foundation, Singapore under its AI Singapore Programme (Award No: AISG2-100E-
2021-089). We like to thank uParcel and AI Singapore for data, domain and comments,
the ICCL PC chairs and reviewers, with special mention of Stefan Voss, for suggestions
and meticulous copy-editing during the review process.

References
1. Agussurja, L., Cheng, S.F., Lau, H.C.: A state aggregation approach for stochastic
multiperiod last-mile ride-sharing problems. Transp. Sci. 53(1), 148–166 (2019).
https://doi.org/10.1287/trsc.2018.0840
2. Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.: Optuna: a next-
generation hyperparameter optimization framework. In: Proceedings of the 25th
ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,
pp. 2623–2631 (2019). https://doi.org/10.1145/3292500.3330701
3. Ausseil, R., Pazour, J.A., Ulmer, M.W.: Supplier menus for dynamic matching
in peer-to-peer transportation platforms. Transp. Sci. 56(5), 1304–1326 (2022).
https://doi.org/10.1287/trsc.2022.1133
4. Bent, R.W., Van Hentenryck, P.: Scenario-based planning for partially dynamic
vehicle routing with stochastic customers. Oper. Res. 52(6), 977–987 (2004).
https://doi.org/10.1287/opre.1040.0124
5. Berbeglia, G., Cordeau, J.F., Laporte, G.: Dynamic pickup and delivery problems.
Eur. J. Oper. Res. 202(1), 8–15 (2010). https://doi.org/10.1016/j.ejor.2009.04.024
6. Chen, X., Ulmer, M.W., Thomas, B.W.: Deep Q-learning for same-day delivery
with vehicles and drones. Eur. J. Oper. Res. 298(3), 939–952 (2022). https://doi.
org/10.1016/j.ejor.2021.06.021
7. Chen, X., et al.: Sequential recommendation with user memory networks. In: Pro-
ceedings of the Eleventh ACM International Conference on Web Search and Data
Mining, pp. 108–116 (2018). https://doi.org/10.1145/3159652.3159668
8. Chen, Y., et al.: Can sophisticated dispatching strategy acquired by reinforcement
learning? In: Proceedings of the 18th International Conference on Autonomous
Agents and MultiAgent Systems, pp. 1395–1403 (2019). https://doi.org/10.48550/
arXiv.1903.02716
9. Cheng, H.T., et al.: Wide & deep learning for recommender systems. In: Proceed-
ings of the 1st Workshop on Deep Learning for Recommender Systems, pp. 7–10
(2016). https://doi.org/10.1145/2988450.2988454
10. Dayarian, I., Savelsbergh, M.: Crowdshipping and same-day delivery: employing
in-store customers to deliver online orders. Prod. Oper. Manag. 29(9), 2153–2174
(2020). https://doi.org/10.1111/poms.13219
34 Z. Zhang et al.

11. Guo, H., Tang, R., Ye, Y., Li, Z., He, X.: DeepFM: a factorization-machine
based neural network for CTR prediction. arXiv preprint arXiv:1703.04247 (2017).
https://doi.org/10.48550/arXiv.1703.04247
12. Hou, S., Gao, J., Wang, C.: Optimization framework for crowd-sourced delivery
services with the consideration of shippers’ acceptance uncertainties. IEEE Trans.
Intell. Transp. Syst. 24(1), 684–693 (2022). https://doi.org/10.1109/TITS.2022.
3215512
13. Joe, W., Lau, H.C.: Deep reinforcement learning approach to solve dynamic vehicle
routing problem with stochastic customers. In: Proceedings of the International
Conference on Automated Planning and Scheduling, vol. 30, pp. 394–402 (2020).
https://doi.org/10.1609/icaps.v30i1.6685
14. Li, X., et al.: Learning to optimize industry-scale dynamic pickup and deliv-
ery problems. In: 2021 IEEE 37th International Conference on Data Engineering
(ICDE), pp. 2511–2522. IEEE (2021). https://doi.org/10.1109/ICDE51399.2021.
00283
15. Lian, J., Zhou, X., Zhang, F., Chen, Z., Xie, X., Sun, G.: xDeepFM: combining
explicit and implicit feature interactions for recommender systems. In: Proceedings
of the 24th ACM SIGKDD International Conference on Knowledge Discovery &
Data Mining, pp. 1754–1763 (2018). https://doi.org/10.1145/3219819.3220023
16. McMahan, H.B., et al.: Ad click prediction: a view from the trenches. In: Proceed-
ings of the 19th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pp. 1222–1230 (2013). https://doi.org/10.1145/2487575.2488200
17. Mendoza, J.E., Castanier, B., Guéret, C., Medaglia, A.L., Velasco, N.: Construc-
tive heuristics for the multicompartment vehicle routing problem with stochastic
demands. Transp. Sci. 45(3), 346–363 (2011). https://doi.org/10.1287/trsc.1100.
0353
18. Mnih, V., et al.: Human-level control through deep reinforcement learning. Nature
518(7540), 529–533 (2015). https://doi.org/10.1038/nature14236
19. Pillac, V., Gendreau, M., Guéret, C., Medaglia, A.L.: A review of dynamic vehicle
routing problems. Eur. J. Oper. Res. 225(1), 1–11 (2013). https://doi.org/10.1016/
j.ejor.2012.08.015
20. Powell, W.B.: Approximate Dynamic Programming: Solving the Curses of Dimen-
sionality, vol. 842. Wiley, Hoboken (2011)
21. Powell, W.B.: Designing lookahead policies for sequential decision problems in
transportation and logistics. IEEE Open J. Intell. Transp. Syst. 3, 313–327 (2022).
https://doi.org/10.1109/OJITS.2022.3148574
22. Rendle, S.: Factorization machines. In: 2010 IEEE International Conference on
Data Mining, pp. 995–1000. IEEE (2010). https://doi.org/10.1109/ICDM.2010.
127
23. Resnick, P., Varian, H.R.: Recommender systems. Commun. ACM 40(3), 56–58
(1997)
24. Ritzinger, U., Puchinger, J., Hartl, R.F.: A survey on dynamic and stochastic
vehicle routing problems. Int. J. Prod. Res. 54(1), 215–231 (2016). https://doi.
org/10.1080/00207543.2015.1043403
25. Secomandi, N.: A rollout policy for the vehicle routing problem with stochastic
demands. Oper. Res. 49(5), 796–802 (2001). https://doi.org/10.1287/opre.49.5.
796.10608
26. Ulmer, M.W., Thomas, B.W.: Meso-parametric value function approximation for
dynamic customer acceptances in delivery routing. Eur. J. Oper. Res. 285(1), 183–
195 (2020). https://doi.org/10.1016/j.ejor.2019.04.029
 When Routing Meets Recommendation 35

27. Ulmer, M.W., Thomas, B.W., Mattfeld, D.C.: Preemptive depot returns for
dynamic same-day delivery. EURO J. Transp. Logist. 8(4), 327–361 (2019).
https://doi.org/10.1007/s13676-018-0124-0
28. Voccia, S.A., Campbell, A.M., Thomas, B.W.: The same-day delivery problem for
online purchases. Transp. Sci. 53(1), 167–184 (2017). https://doi.org/10.1287/trsc.
2016.0732
29. Wang, S., Hu, L., Wang, Y., Cao, L., Sheng, Q.Z., Orgun, M.: Sequential recom-
mender systems: challenges, progress and prospects. In: 28th International Joint
Conference on Artificial Intelligence, IJCAI 2019, pp. 6332–6338 (2019). https://
doi.org/10.24963/ijcai.2019/883
30. Watkins, C.J., Dayan, P.: Q-learning. Mach. Learn. 8(3), 279–292 (1992). https://
doi.org/10.1007/BF00992698
 A Reactive-Periodic Hybrid Optimization
for Internal Hospital Logistics

Ebrahim Ehsanfar1,2(B) , Farzaneh Karami3,4 , and Tim Kerkenhoff1

1
Fraunhofer Institute for Material Flow and Logistics (IML), Dortmund, Germany
{ebrahim.ehsanfar,tim.kerkenhoff}@iml.fraunhofer.de
2
Faculty of Mechanical Engineering, TU Dortmund, Dortmund, Germany
3
Department of Computer Science, KU Leuven, Ghent, Belgium
4
DeltaQ, Brussels, Belgium
[email protected]

Abstract. Internal hospital logistics (IHL) involves the scheduling of

materials and patient transportations employing a fleet of transporters.
The problem of collecting and delivering these items within a hospital
can be modeled as a Pickup and Delivery Problem with Time Win-
dows (PDPTW). This paper proposes a hybrid dynamic optimization to
address the IHL problem based on a two-step heuristic. This algorithm
combines reactive and periodic optimizations to assign logistic’s trans-
ports to the most suitable transporters while considering the urgency of
each transport. To conserve resources, this algorithm addresses logistics
transports with higher urgency reactively and handles less urgent trans-
ports periodically. The initial assignment is constructed using the earliest
due date first (EDDF) assignment method. To further improve the effi-
ciency of the procedure, a ruin and recreate heuristic is developed and
tested. Computational experiments have been conducted utilizing hospi-
tal data from a large hospital with approximately 2100 beds located in
Germany to evaluate the performance of the proposed dynamic hybrid
optimization. Results show that the hybrid policy outperforms the base-
line reactive policy used in the hospital in terms of service quality and
cost efficiency.

Keywords: Internal Hospital Logistics · Dynamic Optimization ·

Dynamic Pickup and Delivery Problems · Ruin & Recreate Heuristics

1 Introduction

Recently, many developed countries have experienced a significant rise in their

healthcare costs. This is due to higher demands to strengthen the foundations of

This research was conducted at the Fraunhofer Institute for Material Flow and Logistics
as part of the KIK-Dispo project, funded by the Federal Ministry of Education and
Research of Germany (Grant No. 01IS19041B). The responsibility for the content lies
with the authors.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 36–55, 2023.
https://doi.org/10.1007/978-3-031-43612-3_3
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 37

health systems, which results in guaranteeing higher levels of health services for
a major part of the population. According to data from 1980 onwards, in most
OECD1 countries the percentage of GDP spent on health has increased over
time [9]. Where, for example, Germany in 2020 allocated 12.5% of its GDP for
healthcare, with approximately 30% of that expenditure being directed towards
hospital services [18]. Logistics operation among different hospital activities is
responsible for 20% to 40% of total hospital cost [11]. Therefore one of the
interesting topics for researchers is investigating efficient algorithms for reducing
the costs of logistics services.
In order to enhance the efficiency of hospital logistics, a crucial part of the
cost, Internal Hospital Logistics (IHL) optimization must be prioritized. Effective
management of IHL has been a longstanding challenge. Recently, the COVID-19
pandemic has added additional challenges to IHL operations by placing a heavy
burden on their processes. Developing efficient IHL operations frees up valuable
transport resources, particularly when hospitals face strict distancing guidelines
and limited transport resources. In such circumstances, hospitals may need to
quickly expand intensive care capacities to ensure continued care in the event of
increased patient numbers. Additionally, there has been a significant increase in
demand for protective equipment and hygiene products. Logistics management
must therefore implement fast, pragmatic solutions to prevent such bottlenecks
within hospitals.
IHL’s timely transport services are typically planned between different loca-
tions within the hospital with an intention of maintaining continuous care. As
such, the problem of collecting and delivering patients and materials (e.g., medi-
cal equipment and consumables) within hospitals can be considered as the Pickup
and Delivery Problem with Time Windows (PDPTW) [8]. However, in real-world
IHL problems, uncertainty in the transport data (e.g. unknown arrival times or
time windows) makes it challenging to solve such a dynamic problem, efficiently.
To effectively handle the dynamic PDPTW challenges, several steps need to
be taken. One crucial step is collecting and arranging transports for optimization.
This involves determining how to collect data on new transports, how to integrate
them with existing ones, and how to prioritize them. Another step is selecting
an appropriate algorithm that can handle the dynamic nature of the PDPTW.
The algorithm should rapidly and efficiently re-optimize schedules when new
transports arrive or conditions change. Finally, it is essential to implement the
schedules produced by the algorithm. This can be achieved using an automated
transport management system that assigns transports to transporters, tracks
their progress, and updates schedules in real-time. Addressing all these steps is
crucial for providing high-quality services at a minimal cost.
Like many other optimization problems, the PDPTW can be solved using
exact and heuristic methods. These methods are mostly differentiated by their

1
The OECD (Organisation for Economic Co-operation and Development) is an inter-
national organization made up of 38 member countries. These member countries are
generally considered to be developed, industrialized economies with relatively high
levels of economic growth, stability, and well-being.
38 E. Ehsanfar et al.

ability to generate optimal solutions and their computational demands. Exact

algorithms produce optimal solutions; however, they are computationally expen-
sive and memory-demanding. According to Rais et al. [13], it takes three hours
to calculate routing and scheduling in a midsize hospital, which is not practical
given the limited time available to conduct their logistics services, especially for
vital and urgent transports. Heuristics on the other side can produce high-quality
solutions within a reasonable computation time for large-scale problems.
The majority of research on the PDPTW employs heuristic solutions [12,19].
There are two types of heuristics as single-phase-based and two-phase-based
heuristics. Single-phase-based heuristics, such as insertion heuristics, construct
a feasible solution by adding pickup or delivery services to the transporter sched-
ule one at a time. On the other hand, two-phase-based heuristics [5,7,17] gen-
erate the transports’ schedule by first constructing a solution (e.g. by insertion)
and then improving this solution iteratively. A few examples of two-phase-based
heuristics introduced as route-first-cluster-second heuristic [4], the XOR-based
heuristic [14], and two-phase simulated annealing [1]. Recently, a ruin and recre-
ate (R&R) based heuristic [3] has been introduced as a state-of-the-art two-phase
heuristic.
A common policy known as reactive optimization assigns services to a trans-
porter as soon as they are received. Employing this policy allows immediate
reactions without losing decision time and ensures that the quality of service is
not compromised. In IHL operation, particularly for emergency transports, it
is important to guarantee service quality. Another policy is the periodic opti-
mization [7] which makes decisions only after a set of predefined criteria have
been met. The periodic optimization policy schedules the transports in periods
and re-optimizes the schedule after each period, enabling adjustments and reas-
signments of transports to transporters as necessary. It is especially beneficial
when managing limited resources, such as transporters and vehicles, in an effi-
cient manner. In [6], they have studied and compared these two approaches and
investigated the circumstances under which one may be more effective than the
other. However, this paper identifies a method to achieve the advantages of both
policies. Therefore, a hybrid policy is developed that uses a ruin and recreate
heuristic algorithm to provide high-quality services while minimizing route costs
consistently. The term optimization is used for the pursuit of efficient solutions
within the inherent complexities of the IHL. It does not imply the achievement
of globally optimal solutions.

2 Definitions and Preliminaries

To enhance comprehension of the hybrid scheduling algorithm, the following

definitions and preliminaries are presented:

Transport: Movement of a specific item from its pickup location to

its delivery location.
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 39

Transport type: Mode of transportation used for each transport, e.g.,

wheelchairs or hospital beds for patients, or trolleys for
medicines and equipment.
Transport locations: Specific locations where a transport begins or ends.
Time window: Each transport must start and be completed within a
specified time window. Transports cannot be scheduled
before the start of their time window but can be sched-
uled after the end of the time window.
Priority: Categorizes each transport by its level of importance
or urgency.
Transporter: Logistics personnel responsible for transporting items
within the hospital locations.
Shift: The period of time during which a transporter is work-
ing continuously.
Tour: Collection of all transports to be completed by a single
transporter during their shift.

According to a study [2], a hospital with 1000 beds typically requires approxi-
mately 750 transportations daily. Transports are requested by clinical and execu-
tive staff and conducted by transporters. In this paper, data from a pilot hospital
in Germany with approximately 2100 beds is utilized. This data is collected from
specialized software designed for logistics and services management. The soft-
ware facilitates the hospital in acquiring data pertaining to new transportation
requests, assimilating them with the already existing ones, and classifying them
based on critical factors such as urgency and proximity. The average statistics
for transports and tours in this hospital are presented in Table 1.
In order to enhance the efficacy of IHL transport operations, medical institu-
tions often implement management software systems, which facilitate the record-
ing, processing, and handling of all transportation operations while also furnish-
ing electronic assistance for manual dispatch and automatic scheduling. Auto-
mated scheduling software allows hospitals to distribute transport services based
on current demand and available resources. In the daily operations of hospi-
tals, medical staff members enter transport orders into the logistic management
software. These orders contain information about each transport’s pickup and
delivery locations, time windows, transport type, and priority. Logistic managers
then assign these transports to available transporters based on their skills and
availability.

2.1 Hospital Internal Graph and Travel Time Calculation

A hospital’s internal layout can be represented by a graph of several locations,

where each vertex corresponds to a location within the hospital, such as patient
rooms, labs, and pharmacies, and each edge represents the connections between
two vertices. Even for small hospitals, the resulting graph can be large and
complex, making scheduling and logistics management challenging.
40 E. Ehsanfar et al.

Table 1. The average daily statistics for transports and tours in the pilot hospital.

Transports Tours
Number or Transports 1900 Number of Tours 122
Completed Transports 1500 Tours at Night Shifts 26
Emergency Transports 40 Tours at Day Shifts 96
Patient Transports 1100 Average Duration 550 min
Cito Transports 220 Average Number of Transports 19
Material Transports 570

To accurately represent the transport times between locations within the

hospital, this paper employs a data-driven approach that uses historical data
based on the time required to travel between different locations. While Euclidean
distance is one of the most commonly used methods to calculate these distances,
it does not always reflect the actual travel time between locations, as factors
such as stairways and the layout of the hallways can significantly influence the
travel time. Thus, using time as an index, rather than distance, helps to obtain
a more accurate representation of transport times between different locations.
The travel time matrix is the set of the travel times between all locations of the
hospital.
To create a more realistic travel time matrix for internal hospital logistics, it
is important to consider the varying speeds of different transport types, such as
wheelchairs and beds. The travel time for each transport type can vary greatly
based on the nature of the task, such as the transportation of a patient on a bed
versus a wheelchair. We also calculate the travel time for empty paths between
locations, where the transporters travel without carrying any patient or object.
Additionally, the accuracy of the travel time matrix can be improved by consid-
ering the time of day. Hospital corridors may be less crowded at night, leading to
shorter travel times between locations compared to the day. This variability in
transport times is factored into the travel time matrix, providing a more accurate
representation of the actual travel times. The travel time matrix provides a valu-
able tool for optimizing total travel time in internal hospital logistics operations.
Each element of the matrix is calculated as the 60th percentile of all travel time
records between similar pickup and delivery locations for each hour and trans-
port type. While the median is a commonly used measure of central tendency
in such situations, we chose the 60th percentile to account for calculation errors
and ensure that our data is closer to realistic. Prior to this calculation, the data
was subjected to a cleaning procedure to remove any anomalous values that may
have adversely affected the validity of the results.
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 41

Fig. 1. Example of a transport tour based on real data collected from a hospital. The
tour consists of 22 transports that were performed between 14 locations. (Color figure
online)

Figure 1 shows an example of a transport tour performed by a single trans-

porter, based on real data collected from a hospital. The blue boxes indicate
hospital locations, the blue arrows indicate completed transports, and the red
arrows indicate empty transports when the transporter is traveling without any
equipment. The tour illustrates that when either material or a patient is assigned
to a transporter, transporters sometimes need to walk to reach the pickup loca-
tion. Some of these empty travels may take a considerable amount of time.

2.2 Problem Definition

From the perspective of hospitals, reducing logistic costs and improving trans-
port service quality are two critical objectives. Providing high-quality service
means ensuring that each transport service is provided within its predefined
time window. Longer travel times result in higher costs for the hospital, partic-
ularly for skilled transporters who are paid more per hour of work. As a result,
hospitals aim to efficiently utilize these valuable resources. Minimizing the total
travel time of transporters helps to meet the hospitals’ demand for cost reduc-
tion and efficient resource utilization, while also ensuring that transport services
are provided within the designated time windows.
42 E. Ehsanfar et al.

The problem involves assigning each transport to a transporter with the nec-
essary skill and determining the sequence of transports to minimize the total
travel time for all transporters. The objective of the current paper is to mini-
mize the sum of the lateness of all transports and the total travel time for all
transporters while also considering the various constraints and characteristics of
the transport and transporters. For example, each transport should be assigned
to a transporter with the necessary skill, and it is necessary to ensure that each
transport is conducted within its given time window.

3 Algorithm

The dynamic and complex nature of hospital logistics has resulted in an increased
demand for efficient algorithms that can quickly adapt to unexpected changes
to maintain high-quality patient care services. In response, a novel algorithm is
presented in this section, which consists of a pre-processing phase followed by a
two-step heuristic solver.

3.1 Pre-processing

The pre-processing phase is a critical stage in the algorithm that prepares

input data by combining suitable transports. Combining two or more transports
implies that they can be accomplished by the same transporter during a single
tour. Specifically, in this stage, transports that share the same time windows
and have identical pickup or delivery locations are combined, provided that a
combination is possible and allowed. The combined transports are then assigned
to the same transporter in the subsequent stage. However, certain combinations
may be forbidden under specific circumstances. For example, it may not be pos-
sible to transport a blood sample and a patient on a bed simultaneously due to
health regulations. However, this depends on the specific regulations in place.
Another example is that transporting more than one bed at the same time may
not be possible for a single transporter.
To determine the best transport combination, subsets of transports are gen-
erated that can be combined based on their time windows and transport type.
The objective is to create the largest possible subset that covers all transports,
and then calculate the permutations of elements of each subset to find the best
tour with the least travel time. The pre-processing step is given by Algorithm 1,
which reduces the search space for the heuristic, leading to improved efficiency
and transport quality.
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 43

Algorithm 1. Combine Transports

1: Input: A set T of n transports and a lateness threshold 
2: Output: A covering set CT ⊆ P(T ) of T with minimum cardinality such that each
transport in T is in exactly one element of CT
3: function CombineTransports(T )
4: f ound ← false
5: for k = 1 to NT do  Initialize the covering set as CT = {T }
6: for Any covering set CT with cardinality k do
7: for each subset S of CT do
8: for each permutation PS of S do
9: if Any Forbidden Combinations in PS then
10: continue
11: end if
12: if Lateness (PS ) >  then  Lateness(PS ) = Sum of lateness for
all consecutive t ∈ PS
13: continue
14: end if
15: f ound ← true
16: end for
17: end for
18: end for
19: if f ound then
20: return CT
21: end if
22: end for
23: return t : t ∈ T  No valid covering set found, return set of singletons (no
possible combination)
24: end function

3.2 Heuristics Algorithm

To address the IHL problem, the scheduling starts with creating two separate
service pools based on priority: urgent and non-urgent transports. Transport that
exceeds the priority threshold is categorized as urgent and placed in the corre-
sponding urgent transports pool, while non-urgent transports are stored sepa-
rately. The proposed heuristic algorithm assigns transports stored in each pool
individually. To illustrate, the paper begins by discussing urgent pool scheduling,
followed by an explanation of how the non-urgent pool is scheduled. Finally, an
overview of the entire heuristic algorithm is provided.

Urgent Pool: To schedule transports from the urgent pool, the algorithm fol-
lows a reactive scheduling policy, which focuses on urgent transports. We clas-
sify the urgent pool based on their priority level as either emergencies or highly
urgent transports. Emergencies have the highest priority and must be handled
immediately without any combination with other transports. Examples of emer-
gency transports include emergency medical supplies for the operating room,
moving patients who have suffered accidents, or transporting critical patients
44 E. Ehsanfar et al.

from the emergency department to the intensive care unit. As soon as the sys-
tem identifies an emergency, it assigns the best available transporter, i.e., the
available skilled transporter closest to the pickup location. The process of assign-
ing the best transporters to each transport is shown in Algorithm 2.

Algorithm 2. Find Best Transporter

1: Input: A transport t and a set of available transporters P
2: Output: The best transporter p in P to carry out transport t
3: function FindBestTransporter(t, P )
4: Tt ← transport time of t
5: dmin ← very large number
6: pbest ← None
7: for each transporter p in P do
8: Lp ← last seen location of p
9: Ap ← next availability time of p
10: T ravelT imep ← empty travel time between Lp and pick-up location of t
11: if Ap + T ravelT imep + Tt < dmin and p has the required skills for t then
12: dmin ← Ap + T ravelT imep + Tt
13: pbest ← p
14: end if
15: end for
16: return pbest
17: end function

Algorithm 3. Reactive Assignment Algorithm

1: Input: A set of transports T and a set of transporters P .
Output: A set of pairs Re with elements (ti , pj ), where ti is in T and pj is in P .
2: 
ti ∈T ti = T .
3: procedure ReactiveAssignment(T, P )
4: Re ← ∅
5: Sort T in ascending order based on the delivery time.
6: for ti ∈ T do
7: pj ← FindBestTransporter(ti , P )  Find the best transporter for ti
(Algorithm 2)
8: Re ← Re ∪ (ti , pj )  Add the pair (ti , pj ) to Re
9: end for
10: return Re
11: end procedure

After assigning the emergency transports employing Algorithm 2, there may

still be other highly urgent transports in the pool, such as pharmacy transporta-
tion to patient rooms. For these transports, first, their possible combination is
examined. If it is determined that the combination will result in lateness, they
are scheduled separately with no combination. If it is possible to combine mul-
tiple transports without incurring any lateness, the system will roll them out
as a single transport, using the same transporter. For each transport after the
pre-processing phase, the system will assign a transporter based on the earliest
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 45

Algorithm 4. Urgent Pool Assignment

1: Input: A set of transports T and transporters P .
Output: A set of pairs Re with elements (ti , pj ), where ti is in T and pj is in P .
2: 
ti ∈T ti = T .
3: procedure UrgentPoolAssignment(T, P )
4: T hr ← threshold urgency factor
5: Tnot ← {ti ∈ T | urgency factor (ti ) ≤ T hr}  Emergency transports
6: Turg ← T \ Tnot  Urgent transports
7: Pnot ← P  Initialize Emergency transporter set
8: Renot ← ReactiveAssignment(Tnot , Pnot )  Reactive Assignment for
Emergency transports
9: for (ti , pj ) ∈ Renot do
10: Update Pnot based on the availability and last location of pj
11: end for
12: Turg ← Combiningtransports(Turg )  Combine transports for urgent
transports (Algorithm 1)
13: Purg ← P \ Pnot  Urgent transporter set
14: Reurg ← ReactiveAssignment(Turg , Purg )  Reactive Assignment for
urgent transports
15: Re ← Renot ∪ Reurg  Combine emergency and urgent assignments
16: return Re
17: end procedure

delivery time and proximity. If no transporter is available to perform a trans-

port without incurring lateness, the algorithm will assign the transporter with
the least amount of lateness to each transport. The whole process is shown in
Algorithm 3.
Algorithm 4 outlines the process of scheduling the urgent pool and select-
ing the efficient transporters for handling emergencies and highly urgent trans-
ports using a reactive scheduling policy. Each transport that is scheduled will
be removed from the pool. Additionally, each transporter that is assigned to a
transport will receive a new availability time, which will be after the end of the
delivery of the assigned transport. The transporter’s last available location will
be the delivery location of the assigned transport. The transporter’s availability
at this time and location for later assignments should be taken into account.

Non-urgent Pool: The proposed algorithm adopts a periodic scheduling pol-

icy, whereby a fixed duration of time (r minutes) is observed before scheduling
transports from the non-urgent pool. Subsequently, each scheduling is executed
after every r minutes. This policy is designed to optimize the tours by minimiz-
ing the total transporters’ travel time while guaranteeing that they receive their
due services within the stipulated period. At each scheduling step, the algorithm
first checks if it is possible to combine any transports. If so, the pool is updated
accordingly. The transports in the pool are then sorted based on their due date
in ascending order, which represents the deadline for their delivery. The periodic
assignment Algorithm 5 has two main steps as follows:
46 E. Ehsanfar et al.

Step 1 - Earliest due date first (EDDF) Algorithm 6: The earliest due date
first algorithm is employed to optimize tours considering the service due date
such that transports with earlier due dates are assigned first. As a result of
step 1, a set of tours is constructed as the initial solution.

Step 2 - R&R algorithm: The R&R algorithm is applied to improve the initial
solution. This algorithm is repeated several times to optimize tours while
minimizing total service lateness. The number of transports to be reassigned
and the number of iterations at each iteration can be adjusted based on the
transports pool size and the desired solution quality. In the next part, we will
discuss R&R in general and the algorithm in more detail.

Algorithm 5. Periodic Assignment Algorithm

1: Input: Size of optimization periods r, set of transports T and set of transporters
P in each loop
2: Output: Assignments, updated set of transports and updated set of transporters
in each loop
3: procedure PeriodicAssignment(r)  r: size of optimization periods in minutes
4: while true do  Loop until stopped. e.g. the end of working day
5: Get (T, P )  T : set of transports, P : set of transporters
6: initialAssignments ← EDDF(T, P )
7: ref inedAssignments ← RuinRecreate(initialAssignments)
8: Execute refined assignments
9: Update (T )  the assigned transports are removed from T
10: Update (P )  the next availability and location of the transporters get
updated
11: Wait for r minutes
12: end while
13: end procedure

Algorithm 6. Earliest Due Date First Assignment (EDDF)

1: Input: A set of transports T and transporters P .
2: Output: The list of assignments in the initial step
3: procedure Edd(T , P )
4: Assignments ← Empty DataFrame
5: Sorted transports ← Sort(T, key = λt : t.earliest due date)  Sort transports
by earliest due date
6: for i in 1 to |Sorted transports| do
7: t ← Sorted transports[i]
8: p ← FindBestTransporter(t, P )
9: Assignments.append([t, p])
10: end for
11: return Assignments
12: end procedure
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 47

Algorithm 7. Ruin & recreation algorithm (R&R)

1: Input: Transport pool T , List of tours P
2: Output: List of updated tours
3: procedure Ruin(T, P )
4: T ourEmptyF actors ← Calculate empty travel factor for each tour in P
5: SelectedT our ← Tour with the highest empty travel factor from
T ourEmptyF actors
6: RemovedT ransports ← All transports from the selected tour in P
7: T ransportP ool ← Add the removed transports back to the transport pool T
8: return RemovedT ransports, T ransportP ool  Removed transports and
updated transport pool
9: end procedure
10: procedure Recreate(T, P, RemovedT ransports)  Transport pool T , List of
tours P , Removed transports
11: LatenessBef ore ← Calculate the overall lateness of P
12: for all transport in RemovedT ransports do
13: LatenessBest ← ∞
14: InsertionT our ← None
15: for all tour in P do
16: LatenessInsert ← Calculate the lateness of inserting transport into
tour
17: if LatenessInsert == 0 then
18: Insert transport into tour and return
19: end if
20: if LatenessInsert < LatenessBest then
21: LatenessBest ← LatenessInsert
22: InsertionT our ← tour
23: end if
24: end for
25: Insert transport into InsertionT our
26: end for
27: LatenessAf ter ← Calculate the overall lateness of P
28: if LatenessAf ter < LatenessBef ore then
29: return P  Recreated tours
30: else
31: return None  Recreate failed, return None
32: end if
33: end procedure

3.3 Ruin & Recreation (R&R) Algorithm

Many studies show that R&R is an effective algorithm for solving complex opti-
mization problems such as PDPTW. The basic principle of R&R is to first “ruin”
a large part of the initial solution by randomly selecting a certain number of
transports and removing them from their assigned routes, and then “recreate”
a new solution by reassigning the removed transports to routes in a way that
improves the overall solution quality. For the first time, the R&R algorithm was
proposed by Schrimpf et al., [16] for solving traveling salesman problems and
48 E. Ehsanfar et al.

has since been extensively studied and improved by researchers in the field. For
example, Ropke and Pisinger [15] developed various ruin methods such as ran-
dom, worst, cluster, or history-based, and then a greedy algorithm was used to
recreate the solution. Later, Pisinger and Ropke [10] proposed an R&R algorithm
that employs adaptive large neighborhood search. More recently, the state-of-
the-art R&R algorithm has been proposed for the Vehicle Routing Problem with
Time Windows (VRPTW) [3].
The R&R algorithm is particularly effective in dealing with complex and
“discontinuous” problems, where small changes in the solution can result in large
differences in quality. This is often the case in the VRPTW, where solutions must
meet a variety of constraints and a small change in the scheduled transportation
can cause a solution to become infeasible [3]. Using a large number of transports
to remove from the solution can significantly enhance the performance of the
R&R algorithm, allowing it to explore a larger solution space and increase the
likelihood of finding high-quality solutions [3].
This paper optimizes scheduled transportation by employing the R&R algo-
rithm (Algorithm 7). The algorithm begins by calculating an empty travel factor
for each tour, which represents the total empty travel duration of each trans-
porter divided by the total tour duration. This factor indicates the amount of
time each transporter spends traveling between locations without performing
any services. Tours with higher empty travel factors are more likely to benefit
from improvement.
The R&R algorithm sorts tours based on their empty travel factor and
attempts to reduce the empty travel duration by reassigning transports to other
transporters.
Ruin: The iterative R&R algorithm initiates by identifying the tours with the
highest empty travel time factor. At each iteration, the tours are sorted based on
this factor, and the tour with the highest score is chosen to be ruined. Ruining
the chosen tour involves removing all transports from the tour and reintroducing
them back into the transport pool.
Recreation: The recreate step has two sub-steps: (1) inserting the removed
transports into another existing tour or (2) creating a new tour with an empty
transporter. When deciding on the insertion location for a removed transport,
a simple criterion of minimizing the overall lateness caused by the insertion can
be taken into account.
When determining the insertion location for a removed transport, the inser-
tion is accepted into the existing tour if it adds zero lateness. If there is a late-
ness, the tour is continued to be built, and the overall lateness is compared to
the previous overall lateness. If the new tour has less lateness, the new insertion
is accepted. From an operational perspective, it is crucial to acknowledge that
transporters have limited knowledge of their future transports, usually, they are
only aware of a small number of upcoming ones. Therefore, the creation of new
tours during rescheduling does not present significant challenges. This is mainly
because the rescheduling process and the generation of new tours occur seam-
lessly in the background, ensuring a streamlined and effective workflow for the
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 49

transporters. Therefore, the system can adapt to changes in real-world situations,

making single modifications when needed while preserving efficient operational
performance.

3.4 Overview of the Hybrid Algorithm in Practice

The hybrid algorithm proposed in this paper is illustrated in Fig. 2. The imple-
mentation of the hybrid algorithm starts with creating a list of transporters that
includes their names, availability, and the last seen location of each. If a trans-
porter has conducted any transport previously, the last location is considered
as the delivery location of the last transport. Otherwise, it is assumed that the
transporters begin their shift from the logistics center of the hospital.

Fig. 2. Flow chart of hybrid algorithm

Upon receiving a transport order, the algorithm initially evaluates whether

it corresponds to an urgent or non-urgent transport. For urgent transports, the
algorithm checks again if the transport is an emergency. If it is, then the algo-
rithm assigns the best available transporter to handle the transport with minimal
delay. This is accomplished by identifying the transporter with the lowest travel
time to the pick-up location of the transport. For highly urgent transports that
are not emergencies, the process is the same, with the addition of a combination
step (in case there is no lateness caused by the combination) before assigning the
best transporter to the transport. This combination step is not usually necessary,
unless there are multiple urgent transports announced at the same time. If there
is only one urgent transport, then the pool contains only one at the same time
as the assignment process is very fast, and the assignment will happen before
the next transports are received by the algorithm.
After scheduling the transport, the table of transporters is updated to reflect
their new availability and locations. If the transport request is non-urgent, the
50 E. Ehsanfar et al.

periodic policy is implemented, which includes the pre-processing step to com-

bine transports and the two-step procedure that uses the EDDF and R&R algo-
rithms. The transports are then assigned, and the list of transporters is updated
accordingly to reflect the changes.

4 Computational Experiment
A set of computational experiments was conducted based on real-world hospital
data. The hospital in question has an average of 42 transporters per working
shift, with facilities located on different campuses, buildings, levels, and rooms,
totaling 364 locations. The computational experiments involved simulating 6
months of logistic planning, from January 1st, 2019 to June 30th, 2019, based on
actual historical data. All transport orders and schedules were stored in the hos-
pital’s logistic management systems. Transporter route schedules were assigned
manually by logistic operators or automatically by a greedy algorithm. The
transport orders were sorted based on their arrival time to ensure chronological
processing. The list of available transporters was obtained from the historical
data, and this information was used to create shifts for each transporter. Table 2
summarizes the primary features of the ILH data. The optimization period was
set to 10 min. Therefore, the system waits for 10 min for re-optimization and
then optimizes the non-urgent transports. We conducted the experimental results
using a machine with 40 CPU cores and 187 GB of RAM. In all cases, the process
of reactive assignments occurred within a matter of milliseconds, while periodic
assignments, including the ruin & recreate phase, took a maximum of 1.43 s per
period.

Table 2. The primary features of the historical data and their respective types

Feature Name Feature Type Description

Entry Time Timestamp The first time the transport is known in the database
TransportID Real The unique ID for each transport
Time Window Time Interval Including the earliest scheduled time for pickup and the latest
scheduled time for delivery
PriorityID Real A number from set {1, 500, 1000, 1010, 1100, 1200}, smaller
number indicates higher priority.
1 indicates emergencies, 500 indicates other urgent transports
Pickup String The pickup location identifier in the hospital
Delivery String The delivery location identifier in the hospital
Transporter String A transporter who has carried out the transport
Real Begin & End Time Interval the realised pick-up and delivery time in the database
Empty Path Float The time (in seconds) that the transporter was walking to
reach the pickup location
TransportType String The type of transport, e.g. transport bed, wheelchair, etc. (16
types of transport in total)

To create shifts, the first timestamp that a transporter appeared in the data
was set as the start time of the shift, and the last timestamp was set as the
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 51

end time of the shift. This allowed for the assignment of transports to specific
transporters based on their availability during their shifts. A loop was used to
feed each transport into the hybrid algorithm, and each was assigned to the most
suitable transporter. The waiting time for each transporter was considered the
same as the real data to ensure accurate simulation. Once tours were scheduled,
lateness and travel time were recorded.
To assess the performance of the hybrid algorithm, the results were com-
pared to the actual historical data. The evaluation metrics consisted of the total
lateness for both emergency and non-emergency services, the total transporters’
travel time, the percentage of emergency transports delivered on time, and the
travel time balance index (T T BI) among transporters. A higher T T BI value
indicates an uneven distribution of travel time among the transporters. The
T T BI is calculated using the following formula:

T T BI = (max travel time − min travel time)/(average travel time)

where max travel time is the total travel time taken by the transporter with
the highest workload, min travel time is the total travel time taken by the
transporter with the lowest workload, and average travel time is the average
travel time taken by all the transporters. A higher T T BI indicates an uneven
travel time distribution among the transporters.
It was observed that weekends and holidays had an average of 423 daily
transports per day, whereas workdays had an average of 2070 daily transports.
Therefore, two types of days were defined in hospital operations: busy and non-
busy days, where non-busy days were considered those with less than 500 trans-
ports per day. Table 3 presents the average daily evaluation parameters for both
busy and non-busy days, where the results of the hybrid algorithm are compared
with the hospital records. Additionally, the benchmarks for each day of the week
from Monday to Sunday were evaluated and presented in Tables 4, 5, 6, 7 and
8. The results indicate that the approach was effective in optimizing scheduled
tours for different days and transport volumes.

5 Results and Discussion

In this section, the results of the evaluation of the hybrid algorithm for emergency
and non-emergency transport operations are presented and compared to the his-
torical data. Firstly, the lateness for emergencies was evaluated, and the average
results are reported in Table 4. These results confirm that the proposed algo-
rithm offers almost the same results as the historical data. The reactive policy
followed by the hospital shows the main reason for such similarity. Therefore, the
hybrid algorithm performs as well as the existing operations for emergencies in
the hospital in most cases. However, in situations where the number of transports
is low e.g. in non-working days and weekends (resulting in a reduced availability
of transporters), some urgent transports may experience slight delays. This is
primarily due to improved resource planning for non-emergency orders, where
periodic assignments are incorporated for enhanced management. As a result, a
52 E. Ehsanfar et al.

few transporters may still be occupied when an emergency is announced, caus-

ing a minor delay in their availability. Nonetheless, the overall benefit lies in the
considerable reduction of delays observed in non-emergency cases, justifying the
tradeoff.
Furthermore, Table 5 reports the percentage of on-time delivery performance
for emergencies. This result is also very similar to the historical data. Addi-
tionally, the results reported in Table 6 show that the workload balance of the
proposed hybrid algorithm is very similar to the historical results. The main
improvement provided by the hybrid algorithm was for non-emergency trans-
ports, as reported in Table 7. The hybrid algorithm reduces the lateness of these
transports by employing a periodic policy, which allows transports to be com-
bined and assigned to transporters more efficiently. This leads to saving time
and resources, as well as more effective assignment of transporters, resulting in
fewer empty tours. Moreover, the total transporters’ travel time, as shown in
Table 8, indicates that the hybrid algorithm performs slightly better than the
historical hospital operations. The reduction of empty travel distances and the
benefits of combining these transports may also contribute to the reduction of
total travel time.

Table 3. The average daily evaluation parameters for days with less than and more
than 500 transports during the six-month period of historic hospital operation data.
The results of the hybrid algorithm are compared with the hospital records

Parameter Transports (>500) Transports (<500)

Historical Hybrid Gap (%) Historical Hybrid Gap (%)
Data Algorithm Data Algorithm
TTBI 0.25 0.21 16 0.29 0.32 −10.34
Lateness Emergencies 7.83 7.66 2.17 1.25 1.43 −14.4
On-Time Emergencies (%) 91 90 1.1 98 97 1.02
Lateness Non-Emergencies 579.87 214.67 62.97 70.04 27.55 60.66
Total Travel Time 13177 10786 18.12 2891 2002 30.75

Table 4. The average daily sum of lateness, in minutes, for emergency transports on
weekdays within the experience horizon, as obtained from the results of the hybrid
algorithm and hospital records

Parameter Mon Tue Wed Thu Fri Sat Sun

Lateness 7.83 7.70 7.65 7.87 8.05 1.27 1.20
Emergencies
Historic Data
Lateness 7.64 7.45 8.11 6.98 7.04 1.56 1.32
Emergencies Hybrid
Algorithm
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 53

Table 5. The average on-time delivery percentages of emergency transports for week-
days within the experience horizon, comparing the results of the hybrid algorithm with
hospital records.

Parameter Mon Tue Wed Thu Fri Sat Sun

On-time Emergencies Historic 91% 89% 90% 92% 96% 99% 98%
Data (%)
On-time Emergencies Hybrid 90% 90% 93% 92% 94% 97% 98%
Algorithm (%)

Table 6. The average daily TTBI parameter for weekdays within the experience hori-
zon for the results of the hybrid algorithm and the hospital records.

Parameter Mon Tue Wed Thu Fri Sat Sun

TTBI Historic Data 0.25 0.22 0.19 0.18 0.19 0.28 0.29
TTBI Hybrid Algorithm 0.25 0.24 0.19 0.20 0.20 0.26 0.26

Table 7. The average daily sum of lateness, in minutes, for non-urgent transports on
weekdays within the experience horizon, as obtained from the results of the hybrid
algorithm and hospital records

Parameter Mon Tue Wed Thu Fri Sat Sun

Lateness Non-Emergencies 567.2 545.8 543.1 512.0 487.2 222.3 204.4
Historic Data
Lateness Non-Emergencies 198.4 200.2 214.8 186.3 215.4 67.9 72.6
Hybrid Algorithm

Table 8. The average daily sum of total travel time, in minutes, on weekdays within the
experience horizon, as obtained from the results of the hybrid algorithm and hospital
records

Parameter Mon Tue Wed Thu Fri Sat Sun

Total Travel Time 13749 13876 10174 13543 13001 2812 2945
Historic Data
Total Travel Time 10653 10431 10111 10056 10399 1988 2061
Hybrid Algorithm
54 E. Ehsanfar et al.

6 Conclusion and Future Works

This paper presents a hybrid dynamic optimization approach for internal hospi-
tal logistics (IHL). This approach utilizes a two-step heuristic to assign logistic
transports to transporters in a way that considers both urgency and resource
conservation. The experiments show that the proposed method outperforms the
baseline reactive policy used in the hospital in terms of service quality, result-
ing in less lateness. Additionally, the approach offers improved cost efficiency,
resulting in less travel time.
While the proposed hybrid dynamic optimization approach has shown
promising results in addressing the IHL problem, further research can be con-
ducted to enhance its overall performance and impact. In particular, there are
several areas where the incorporation of predictive analytics could be beneficial.
First, predictive analytics could be utilized to forecast the IHL workload
based on various factors such as time of day and location within the hospital.
This would allow for better scheduling decisions that take into account potential
fluctuations in demand, ultimately leading to more efficient use of resources
and increased service quality. Additionally, predictive analytics could be used
to forecast the occurrence of urgent cases, allowing transporters to be reserved
specifically for such cases, further enhancing service quality.
Second, the use of AI can further improve the proposed approach by adapting
the size of periods and waiting strategies dynamically based on the workload and
urgency of transports. This would enable more flexible scheduling and allow for
better resource allocation to meet the demands of the hospital.
Lastly, the deployment of the proposed approach in a real-time scenario
within a hospital could provide valuable insights into its performance and effec-
tiveness. On-site evaluations that consider several other factors such as patient
satisfaction, transporter workloads, and overall hospital efficiency could provide
a more comprehensive assessment of the approach’s impact and potential for
implementation in practice.
Overall, the incorporation of predictive analytics and AI has the potential to
significantly enhance this research. By utilizing these tools to forecast workload,
adapt scheduling policies, and identify areas for improvement, the service quality
and efficiency of IHL operations can ultimately be improved, enhancing patient
care.

References
1. Bent, R., Van Hentenryck, P.: A two-stage hybrid algorithm for pickup and delivery
vehicle routing problems with time windows. Comput. Oper. Res. 33(4), 875–893
(2006). https://doi.org/10.1016/j.cor.2004.08.001
2. Bredehorn, T., Johow, L.: KI in der Krankenhaus-Transport-Disposition (2020).
https://publica.fraunhofer.de/handle/publica/265840
3. Christiaens, J., Vanden Berghe, G.: Slack induction by string removals for vehicle
routing problems. Transp. Sci. 54(2), 417–433 (2020). https://doi.org/10.1287/
trsc.2019.0914
 A Reactive-Periodic Hybrid Optimization for Internal Hospital Logistics 55

4. Garside, A.K., Laili, N.R.: A cluster-first route-second heuristic approach to solve

the multi-trip periodic vehicle routing problem. Jurnal Teknik Industri 20(2), 172–
181 (2019)
5. Gendreau, M., Guertin, F., Potvin, J.-Y., S’eguin, R.: Neighborhood search heuris-
tics for a dynamic vehicle dispatching problem with pick-ups and deliveries. Transp.
Res. Part C Emerg. Technol. 14(3), 157–174 (2006). https://doi.org/10.1016/j.trc.
2006.03.002
6. Karami, F., Vancroonenburg, W., Vanden Berghe, G.: Periodic vs. reactive schedul-
ing for internal hospital logistics. Available as TR, KU Leuven (2019)
7. Karami, F., Vancroonenburg, W., Vanden Berghe, G.: A periodic optimization
approach to dynamic pickup and delivery problems with time windows. J. Sched.
23, 711–731 (2020). https://doi.org/10.1007/s10951-020-00650-x
8. Karami, F., Vancroonenburg, W., Vanden Berghe, G.: A buffering-strategy-based
solution method for the dynamic pickup and delivery problem with time windows.
In: The 12th International Conference on the Practice and Theory of Automated
Timetabling, Vienna, Austria, 28–31 August 2018, pp. 256–272. PATAT (2018).
http://patatconference.org/patat2018/files/proceedings/paper37.pdf
9. OECD. Health expenditure and financing (2020). https://stats.oecd.org/viewhtml.
aspx?datasetcode=SHAlang=en#
10. Pisinger, D., Ropke, S.: A general heuristic for vehicle routing problems. Comput.
Oper. Res. 34(8), 2403–2435 (2007)
11. Poulin, E.: Benchmarking the hospital logistics process: a potential cure for the
ailing health care sector. CMA Manag. 77(1), 20–23 (2003). https://doi.org/10.
1016/j.cor.2005.09.012. CMA, Canada
12. Psaraftis, H.N., Wen, M., Kontovas, C.A.: Dynamic vehicle routing problems: three
decades and counting. Networks 67(1), 3–31 (2016). https://doi.org/10.1002/net.
21628
13. Rais, A., Alvelos, F., Figueiredo, J., Nobre, A.: Optimization of logistics services in
hospitals. Int. Trans. Oper. Res. 25(1), 111–132 (2018). https://doi.org/10.1111/
itor.12370
14. Renaud, J., Boctor, F.F., Ouenniche, J.: A heuristic for the pickup and delivery
traveling salesman problem. Comput. Oper. Res. 27(9), 905–916 (2000). https://
doi.org/10.1016/S0305-0548(99)00066-0
15. Ropke, S., Pisinger, D.: An adaptive large neighborhood search heuristic for the
pickup and delivery problem with time windows. Transp. Sci. 40(4), 455–472
(2006). https://doi.org/10.1287/trsc.1050.0135
16. Schrimpf, G., Schneider, J., Stamm-Wilbrandt, H., Dueck, G.: Record breaking
optimization results using the ruin and recreate principle. J. Comput. Phys. 159(2),
139–171 (2000). https://doi.org/10.1006/jcph.1999.6413
17. van Lon, R.R.S., Ferrante, E., Turgut, A.E., Wenseleers, T., Vanden Berghe, G.,
Holvoet, T.: Measures of dynamism and urgency in logistics. Eur. J. Oper. Res.
253(3), 614–624 (2016). https://doi.org/10.1016/j.ejor.2016.03.021
18. Visa Guide. Statistics about the German health care system (2020). https://
visaguide.world/international-health-insurance/germany/healthcare-statistics
19. Zhang, J., Van Woensel, T.: Dynamic vehicle routing with random requests: a
literature review. Int. J. Prod. Econ. 256, 108751 (2022). https://doi.org/10.1016/
j.ijpe.2022.108751
 Cybersecurity Considerations for the Design
of an AI-Driven Distributed Optimization
of Container Carbon Emissions Reduction
for Freight Operations

Carlos Paternina-Arboleda(B) , Alexander Nestler, Nicholas Kascak,

and Morteza Safaei Pour

Department of Management Information Systems, San Diego State University, San Diego,
CA 92182, USA
[email protected]

Abstract. The transportation industry is a vital component of the global econ-

omy, responsible for the movement of goods between different locations. The
intermodal freight transportation system involves the use of different modes of
transportation, such as trucks, trains, and ships, to move freight containers. How-
ever, this system is loaded with inefficiencies due to the poor availability of real-
time coordination and disruptions, causing delays, increased costs, and thus, higher
carbon emissions. AI has the potential to improve the intermodal freight transporta-
tion system’s efficiency by optimizing operations in real-time and self-evolving
the models to make better/faster decisions. While both policymaking and business
operations would benefit from using real-time optimization models, the implica-
tions and applications of these models are different in each context. In policymak-
ing, real-time optimization models are used to improve public services, reduce
overall network costs, and setting regulations for sustainable management of the
network. The system can consider real-time traffic conditions, weather, and other
factors to optimize the routing of the trucks, reducing transportation costs, improv-
ing delivery times, maintaining resiliency, and managing emissions. This work
aims to contribute with a better understanding on how these information systems
can be protected from cyberthreats, while performing the optimization of freight
synchromodal transportation operations in real-time in terms of efficiency, cost-
effectiveness, and carbon emissions reduction, considering the dynamic nature
and heterogeneity of the intermodal freight system.

Keywords: Cybersecurity · Digital Twins · Transportation · AI-Driven

Optimization · Environmental Impact

1 Introduction
If anything, the last three years have taught the maritime industry that the supply chain
is vulnerable and very much dependent on global events. Ports around the world have
witnessed some of the biggest cargo congestion challenges in modern history. This

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 56–84, 2023.
https://doi.org/10.1007/978-3-031-43612-3_4
 Cybersecurity Considerations for the Design 57

congestion, driven by a continuous increase in the gap between consumer demand and
logistics supply, exacerbated by supply-chain disruptions (among which, cybersecurity
breaches are a main concern), is negatively impacting the maritime footprint, despite
large investments by the industry to streamline, modernize, and reduce supply chain
carbon emissions. Consumer demand for goods shipped across the world is set to be
steady or increase. Major ports around the world are witnessing similar trends, leading
to a rise in regional congestion and emissions at bottlenecks, whether in the port, at port
gates, or in dray corridors. Saxon and Stone (2017) build on the container disruption of
the shipping business and how the industry unfolds in the next 50 years. They suggest
that by 2067, the annual growth of global trade of containers will range from 1.9% to
3.2% in the high case scenario. This equates to an increase from 182M TEUs (2016)
to 464M TEUs and 858M TEUs by 2067 respectively for low and high. Every year,
container ships plying the world’s waterways spew about 1 billion metric tons of CO2
into the air (about 3% of global greenhouse gas emissions). Transportation stakeholders
must continue to invest and develop groundbreaking solutions that monitor, track, and
reduce cargo’s carbon footprint.

Fig. 1. Projected CO2 emissions, shipping

The increase of a global cargo movement, coupled with the added complexity of
equipment and systems that are deployed to optimize the movement of goods, pose a
new challenge of measuring the impact of intermodal container travel. This comes at a
critical time as private sectors and government agencies are seeking reliable measure-
ment sources to accurately assess carbon neutrality and drive decarbonization. Despite
many efforts to measure carbon impact, there continue to be a lack of standardization
and discrepancies among the metrics deployed to measure GHG emissions and their
impact. The goal is to accurately measure GHG emissions and levelized costs of Mt-km
(LCOTKM) analyzing the multi-variation by modes of energy and goods transportation
and compare to projected values. Figure 1 shows a shipping emissions projection.
To control carbon emissions, we first need to accurately measure their impact. We
propose a Digital-Twin enabled, AI-driven RT Distributed Optimization Platform, based
upon a system level view of the intermodal network, mapping energy consumption, and
calculating emissions, and offering suggestions for route and mode switching of contain-
ers, which could greatly improve those metrics. Due to the diversity and complexity of
58 C. Paternina-Arboleda et al.

data this Platform would depend upon, success requires a massive data ingestion from
hundreds of data sources while combining extensive data modeling, AI processing,
and algorithms along with simulation. Investing in a high-fidelity simulation Platform
such as Argonne’s POLARIS simulation (Auld et al., 2016), and proven optimization
meta-modeling information systems (Velasquez-Bermudez, 2020), is more beneficial
and rewarding in the long run, as real-time Intermodal efficient mode switch is essential
for these systems to convey good results. Jubiz-Diaz et al. (2021) modeled the Colombia
Intermodal Transportation Network, and examine the relationship between infrastruc-
ture investment, freight accessibility, and GDP, with a data-driven geospatial approach,
targeting investments with greatest GDP impact.
A digital twin is a virtual model designed to accurately reflect a physical object
by using real-time data (IBM, n.d.). According to IBM, there are four types of digital
twins depending on the level of magnification. To successfully implement a digital
twin infrastructure capable of real-time coordination and reduce carbon emissions for
intermodal freight transportation systems in the United States, several sources of data
and information will need to be integrated to form various component twins.
An AI-driven real-time distributed optimization (RT-DO) system is deployed, con-
sidering the dynamic nature of the intermodal freight transportation system. The model
uses real-time data and analytics to optimize operations, reducing costs and carbon emis-
sions. The system considers real-time data on transportation routes, cargo volumes, and
costs. The modeling methodology is based on the concepts of advanced parallel, and dis-
tributed analytics, as support to “artificial brains,” the decision support systems in the era
of Industry 4.0. The system is deployed at different levels. The first level presents real-
time distributed optimization mathematical models for cargo routing which are used to
determine the most efficient/sustainable path for traffic flow in a transportation network,
considering a wide range of factors such as intermodal switching, traffic, and demand
resource limitations. Intermodal switching refers to the transfer of cargo between differ-
ent modes of transportation, such as from a ship to a truck or from a train to a waterway.
We base our RT Optimization modeling framework in Large-Scale Optimization theo-
ries, such as Benders/Lagrangean decomposition, and a novel Event-Driven Real-Time
modeling structure described in Abril et al. (2023). Paternina-Arboleda et al. (2008)
show an integrated simulation-optimization framework for logistical systems.
The main contribution of the research paper entitled “Cybersecurity Considerations
for the Design of a Digital Twin Enabled AI-Driven Real-Time Distributed Optimization
of Container Carbon Emissions Reduction for Synchromodal Freight Operations” lies
in highlighting and addressing the crucial cybersecurity considerations associated with
the design and implementation of a digital twin-enabled AI-driven real-time distributed
optimization system for synchromodal freight operations.
The paper identifies the specific cybersecurity challenges that need to be taken into
account when designing such a system, including securing the data flow, ensuring the
security of digital twin models, and safeguarding the system against cyber-attacks. It pro-
poses practical solutions to address these challenges, such as implementing encryption
mechanisms, access control policies, anomaly detection systems, and other cybersecurity
measures.
 Cybersecurity Considerations for the Design 59

From the cybersecurity perspective, Wang and Liu (2022) discusses the increasing
application of cyber-physical systems (CPSs) in the rail industry and the corresponding
rise in cyber threats that can cause significant failures and consequences. The authors
propose a risk management methodology for addressing cyber security risks in rail
CPSs, focusing on proactive identification, clear definition, and proper handling of these
risks. Beaumont (2018) presents a case study of automated maritime container terminals
(CTs). It has the aim of demonstrating that the risks derived from the use of technol-
ogy associated with the Fourth Industrial Revolution (4IR) are both real and dangerous.
In the shipping industry, the four largest shipping lines in the world have been hit by
cyber-attacks since 2017 (Song, 2021). Cybersecurity is significant in maritime trans-
portation because the costs of cyberattacks can be vast and the consequences fatal. For
instance, digital navigation systems could be manipulated so that they sheer off or run
aground, which would endanger the lives of the crew, people at sea and on land. The
financial effects on shipowners and ship operators would also be immense (Tsvetkova
et al., 2021). Woschank et al. (2020) mentions that AI may be used also in higher-level
processes to detect fraud, prevent cybersecurity threats, and generally optimize higher-
level processes for Smart Logistics in the future. De la Peña-Zarzuelo et al. (2020),
present Internet of Things and sensing solutions, cybersecurity, horizontal and vertical
system integration, cloud computing, 3D printing and additive manufacturing, big data
and business analytics, augmented reality and simulation and modeling are the pillars
of Industry 4.0. From all viewpoints, cybersecurity is a top concern for all AI-driven
digital systems of today’s logistics industry.
The paper emphasizes the critical importance of protecting the integrity, confidential-
ity, and availability of data in a digital twin-enabled AI-driven system. It underscores the
potential risks and vulnerabilities that can arise in the transportation industry, especially
when leveraging advanced technologies for real-time optimization and carbon emissions
reduction. This paper is significant as it sheds light on the importance of cybersecurity
in the context of digital twin-enabled AI-driven systems for synchromodal freight oper-
ations. It provides valuable insights for researchers, practitioners, and decision-makers
in the transportation industry, offering a foundation for the development and imple-
mentation of secure and resilient systems that can effectively reduce carbon emissions,
optimize operations, and improve sustainability.

2 Decision Support System Description

2.1 LCA and GHG Emissions
The first step is to define the system boundaries, which determine processes and activities
that will be included in the analysis. The next step is to collect data on the energy source
and the processes included in the system boundary. This involves data from suppliers,
manufacturers, and energy producers, as well as conducting on-site measurements and
calculations. Using the collected data, the GHG emissions associated with each process
and activity are calculated. This involves using emissions factors or equations that esti-
mate GHG emissions based on energy consumption, fuel use, and production processes.
The GHG emissions calculated for each process and activity are then aggregated to
determine the total GHG emissions associated with the low carbon energy source. This
60 C. Paternina-Arboleda et al.

will become either an objective function or a constraint in our optimization models. We

use a modified version to the framework proposed by Guo et al. (2022). In their article,
the authors present a methodology for estimating carbon emissions in hinterland-based
container inter-modal networks. Their study focuses on a specific intermodal network in
China to show the application of the methodology which involves an estimation model
that considers factors such as distance traveled, mode used, cargo type, and considers
energy consumption of different modes and emissions associated. The study found that
carbon emissions can be reduced by optimizing the network, including increased use of
rail, reduced distance traveled, and use of clean energy sources.
The network estimation model evaluates carbon emissions in a specific period.
Paternina-Arboleda et al. (2023), show an estimation of SO2 emissions derived from
ships when hoteling and maneuvering, and cruising in the port, showing predictive mod-
eling of port emissions inventories. The data relates to vessel’s AIS, and Port’s IoT
environmental sensor network. Cammin et al. (2023) also present a similar approach to
the one above mentioned (Fig. 2).

Fig. 2. Methodology for predicting Emissions Inventories (an example for Port Emissions),
Paternina-Arboleda et al. (2023)

We are expanding the methodology to account for geospatial emissions for the inter-
modal network to be optimized based on the predicted metrics, with valuable insights
to improve environmental practices and foster sustainable development.

2.2 Disruption Management

The system is designed to implement level metrics to assess the resiliency of a given
“sustainable transportation network.” These metrics provide a way to measure the ability
of a transportation system to withstand and recover from disruptions/stresses, while still
providing reliable and efficient transportation services to users.
A resilient transportation network should (a) have enough capacity to handle expected
traffic volumes, while also being able to adapt to changes in demand and unexpected
disruptions, (b) have high travel time reliability, to expect consistent travel times even
in the face of disruptions, (c) be adaptable, to quickly adjust to changing circumstances
 Cybersecurity Considerations for the Design 61

and provide reliable and efficient services, (d) have an appropriate level of redundancy,
to have multiple options for reaching their destination in the event of disruptions, and
(e) be environmentally sustainable (Yang et al., 2023). The global resiliency metric
of a system depends on the resiliency of each individual transportation mode within
because a transportation system is made up of multiple modes, each of which has its
own characteristics and vulnerabilities. The mode switching optimization models must
be tightly coupled with the disruption models. Some metrics proposed for our system
are:
• System capacity as a measure of the ability to handle traffic volumes during
normal/peak periods, considering number/capacity of lanes, number/frequency of
vehicles, availability of modes.
• Travel time reliability as a measure of the predictability of travel times of a network.
This metric considers i.e., congestion, incidents, and weather conditions that can
cause delays or disruptions.
• System adaptability as a measure of the ability of a transportation network to adapt to
changes in demand, technology, or other factors that can impact services. This metric
considers factors such as new technologies, flexibility to adjust based on demand,
and respond to disruptions.
• System redundancy as a measure of availability of alternative routes or modes in the
event of disruptions to the primary network. This metric considers factors such as
number and location of alternative routes, availability of transit or other modes, and
ability to switch between modes.
• Environmental impact as a measure of the sustainability, which takes into consid-
eration factors such as energy consumption, emissions, and the impact on local
ecosystems.
The resiliency of each transportation mode is influenced by a range of factors, includ-
ing quality and age of infrastructure, availability of alternative transportation modes,
the level of investment in research and development, and the degree of coordination
and communication between transportation providers and policymakers (Trucco and
Petrenj, 2023). Understanding resiliency of individual transportation modes is critical
to developing strategies for enhancing resiliency of whole transportation systems. Our
proposed solution is designed to determine the sensitivity to inputs such as routes, low
carbon infrastructure rollout, GHG reductions, and other variables that can help foster
resiliency in several ways. The solution can help identify vulnerabilities in the transporta-
tion system by simulating the effects of various input scenarios on system performance
via a digital twin. We can simulate the impact of a disruption to a key route or a delay
in the rollout of low carbon infrastructure and help identify areas where the system is
most vulnerable.
We could prioritize investments in infrastructure and other resources based on the
potential impact on system performance (identify most critical investments to enhance
resiliency of the system and reduce vulnerability to disruptions), while testing strate-
gies to enhance the resiliency of the system. It provides a visual representation of the
system’s performance under different input scenarios to help understand the potential
impact of actions and decisions on the resiliency of the system to promote more effective
collaboration. Intermodal logistics can be impacted by a range of disruptions (local and
62 C. Paternina-Arboleda et al.

global). Some examples are: (a) natural disasters, (b) labor strikes, (c) disruptions in the
supply chain, (d) global trade disruptions, (e) Pandemics, (f) Network Congestion, and
the focus of this article, (g) Cybersecurity threats. Effective management of disrup-
tions requires a comprehensive understanding of the system’s vulnerabilities, as well
as proactive planning and response strategies to mitigate the impact of disruptions and
maintain the resiliency of the transportation system.

2.3 Full Synchromodal System Optimization

Intermodal transportation systems are an essential component of modern logistics and
supply chain management. The complexity of these systems presents significant chal-
lenges in terms of optimizing routes, minimizing costs, and ensuring deliveries. Real-
time (RT) optimization is a promising approach to address these challenges by leveraging
advanced data analytics and optimization algorithms to make informed decisions in real-
time. It allows transportation planners to respond quickly to unexpected events, such as
traffic congestion or delays in shipments, and make necessary adjustments to routes
and schedules to ensure efficient and timely delivery. Zhang et al. (2018), present an
approach to RT optimization for transportation networks which discusses a method for
optimizing the dynamic stowage of cargo at highway freight stations. They propose an
algorithm that considers the cargo, the available space on the trucks, and the order in
which the cargo is loaded and unloaded using a combination of dynamic programming
and branch-and-bound techniques. They also use simulation to test the effectiveness of
the algorithm under different scenarios.
The proposed system is designed to implement a framework for Autonomous Real-
Time Distributed Optimization (ARTDO), as described in Velasquez-Bermudez et al.
(2020). It enables autonomous optimization of transportation logistics in real-time, using
distributed decision-making algorithms. The methodology improves transportation effi-
ciency while minimizing the environmental impact of transportation systems in real-
time. It involves: (1) Autonomous Distributed decision-making, with nodes equipped
with decision-making algorithms that make autonomous decisions based on local data
and feedback allowing the system to adapt to changing conditions in real-time. (2)
Real-time data collection, to inform decision-making, including information on routes,
schedules, capacities, and other factors that may impact transportation logistics. (3)
Optimization algorithms. (4) Feedback mechanisms for nodes to provide/receive feed-
back and learn/improve based on their performance. To automatically run this math
framework as required, we must embed all the models as a library of smart algorithms
for the Optimization of the Intermodal Network, independently engaged in parallel but
mathematically coordinated to achieve ARTDO across the full network. We will use the
optimization platform to integrate the framework as a set of massive distributed although
interconnected parallel routines. We must connect all the data models, and math models
with data sources, to automatically generate (and maintain) master tables for the whole
intermodal network. The system can incorporate real-time data on GHG emissions from
transportation modes to identify the most efficient transportation modes and routes for
minimizing GHG emissions, while meeting transportation needs, optimizing transporta-
tion modes and routes in RT, based on RT data on traffic congestion, weather conditions,
and other factors. The system can be programmed to switch transportation modes/routes
 Cybersecurity Considerations for the Design 63

in response to changing conditions for GHG emissions or the economic cost of avoided
carbon, while meeting efficiency needs (economic cost/freight ton-Km).

2.4 Designing the Decision Support Information System

The DSS is built in OPTEX (Velasquez-Bermudez, 2020), and it will group all math
models, and predictive/prescriptive analytics for the system. The MMIS (Math Model-
ing information system) manages elements (objects, entities) i.e., tables, fields, indexes,
sets, variables, parameters, constraints, equations, objectives, problems, models, DSS,
and applications. MMIS standardizes the management of entities and relationships cen-
tered on its database algebraic language for linear/non-linear equations. These objects
are critical to address large-scale problems by coordinating multi-problem models and
the information flow. The output of one model is used as the input to others. Due to
the complexity of the real system, the DSS consists of multiple math models integrated
through the data stream, thereby generating the information required by the decision
maker to address all levels (strategy/tactic/real-time). The different models share infor-
mation stored on a common/coherent standardized database to allow for data integration
along with the decision-making chain, so that this coordinated effort guarantees “opti-
mization” of the entire system. The latter is practically impossible to obtain with a single
model. From the platform we can generate math programs in high-level algebraic lan-
guages, like GUROBI or CPLEX. Optex (Fig. 3) is a generic meta-platform that works as
interface for multiple technologies. The modular concept is fundamental to implement
large-scale methodologies, based on partitioning and system decomposition.

Fig. 3. The Optex meta-modeling system

The main large-scale concept in our system is Real-Time Distributed Optimization

(RT-DO), where multiple agents coordinately keep the system on the optimality path
(Velasquez-Bermudez et al., 2020). It follows a standard 7-step process:
• Step 1 – Algebraic Model Load: Fill the database to the MMIS and follow next steps.
• Step 2 – Model Data Load: Generated simultaneously with the loading of math mod-
els (table structures determined by the math model), by sets, parameters, variables,
constraints.
64 C. Paternina-Arboleda et al.

• Step 3 – Generation of the visual user Interface: The system automatically generates
a user interface. The shell windows are associated with main/secondary tables.
• Step 4 – Analysis and Review of the Algebraic Model Formulation: It involves coordi-
nating two simultaneous activities: Review loaded algebraic formulation into MMIS,
and results.
• Step 5 – Run the system and store the Algebraic Model results: Performed
automatically.
• Step 6 – Set the Algebraic Model: Necessary changes to the MMIS and IDIS. This
cyclic process ends when the implemented model produces the correct results, ready
to be delivered.
• Step 7 – Data Access by the end-user: Finally, the end-user can access the model for
use, based on the data stored in the IDIS and results generated by the models.

2.5 System’s Artificial Intelligence

A decision support module will need to run in parallel to the AI process. As the AI
optimizes, updates, or forecasts routes, the schedule will need to be updated to reflect
these changes. Further, common routes and common alternate routes for each region
optimized by the system or updated due to new routes (i.e., construction) need to be
stored as the system generates them for efficient retrieval in the future. This will allow
the system to retrieve optimal routes previously generated providing operators with
options that are already optimized for CO2 emission reduction.

Fig. 4. High level view of the proposed AI process

The predictions from all three base models will be combined for the final regional
inference of the best route the container should take based on the location, traffic condi-
tions, and weather. The algorithms for the AI need to focus specifically on minimizing
 Cybersecurity Considerations for the Design 65

the CO2 emissions overall based on the ton-mileage and minimizing the time-to-arrival
when taking the route by avoiding any traffic. Since CO2 emissions are based on ton-
mileage, without exact emissions per hour data, the best the system can do is avoid
idling in traffic to reduce those idling emissions by minimizing the time-to-arrival and
ton-mileage for the route (Cammin et al., 2023; Paternina-Arboleda et al., 2023).
The AI system will continuously update this information as historical data con-
tinues to be generated and new optimized or updated routes are discovered. Figure 4
shows a high-level view of the AI Process from collecting regional historical data to
generating decisions from inferences made by the system. In previous research articles,
Paternina-Arboleda and Das (2001, 2005) demonstrate how artificial intelligence tech-
niques, specifically Multi-Agent RL, can be used to address complex logistics problems.
By leveraging the power of AI, logistics companies can optimize their operations, reduce
costs, and improve efficiency and customer satisfaction.

3 Data Types and Structures

To successfully implement a digital twin infrastructure capable of real-time coordina-

tion and reduce carbon emissions for intermodal freight transportation systems, several
sources of data and information will need to be integrated to form various component
twins. There will need to be lots of data and information inputs needed to build out an
AI-based model that will provide positive value to a transportation service business. The
AI-based model portion of the system will be built out using the information and data
inputs that are laid out below.
For this system to work and function as intended, there is a large amount of raw
data that the system collects from various sources. These sources include the Global
Positioning Systems (GPS), local cyber sensors at each fielded unit, historical trans-
portation metrics, personnel information, real-time traffic, and weather data, vessel AIS
data, among others. There will be more added as the project continues to grow in scope.
The more data that can be cleaned, indexed, and inserted into the system, the more AI can
learn and adapt to optimize the transportation routes. Dzemydienė et al. (2023) present
a system architecture for monitoring and managing freight intermodal transportation
using the IoT and WSNs.
Before any data or information is collected, there needs to be a shared data repository
to collect the data. This data repository, also known as a data lake, will be primarily in the
cloud with incremental, local hard drive backups. The repository will be internal to the
company’s private network and protected through various cybersecurity controls. After
the shared data repository is established, data can start getting collected. The system will
have a centralized data collection point at the Hub as well as local storage of data on the
system data sensor. The sensor will also monitor data as it transits through the network
and into the model to make sure the data is not compromised.
Historical transportation production data will be used to have a baseline of the his-
torically best routes to take. The historical data should be gathered and broken up in a
logical manner for ease of analysis. This includes recent data on the actual paths the
truck, train or boat took to get from point A to point B. The inputted data can be broken
up by month, year, season, quarters, and year over year changes. Different months and
66 C. Paternina-Arboleda et al.

seasons have different weather patterns that must be kept in mind. Comparing yearly
data could prove to show an abundance of statistical insights. By gathering this historical
data, the AI model will have some data points and ability to make rudimentary decisions.
There is a variety of different information metrics that will assist in the training of
the model. To start off, priority levels need to be set for each method of transportation,
cost, schedule, weather, payload type and traffic patterns. These set priority levels will
lead the model to make more accurate predictions. These priority levels will also be a
large part of the COA selection list that the model outputs.
The mode of transportation and associated personnel are required to physically move
the goods from point A to point B. Trucks are, and will continue to be, the primary
method of travel. However, the transportation method could also be a boat or train and
the associated personnel for each method. Driver information for each situation will have
to be inputted into the system and protected. Driver will be used as a colloquial term in
this paper to represent a truck driver, ship driver and train driver into one central category.
This information varies from driver to driver and method to method. For example, a truck
driver will need to have all their medical information, personal information such as social
and next of kin, time on the road, consecutive days traveled, average speed, follow on
trip and GPS positioning data.
The system will need to keep payload information with set priorities for different
manifests. Each item or good will include the name, Radio Frequency ID (RFID), weight
and priority level. Each item in the manifest will combine to bring the overall payload
to a quantifiable level of priority. This will be inserted into the model for further course
of action training of data.
For the data to reach the Hub and central data repository securely, a data transportation
method needs to be established. The endpoint devices and users will be connected to
the Hub via a private Virtual Private Network (VPN). A VPN is used to “mitigate the
security risks inherent to providing remote network access by offering strong encryption
to provide data security and strong authentication to limit access to applications based
on defined security policies” (Loshin, 2019). Aggarwal, (2022), and Aggarwal (2023)
define data encryption as a security tool that transforms plaintext data into encoded data
called ciphertext that can only be read with a unique key that is given at the time of
the encryption, which can be used for both in transit and at rest data. The two different
types of VPN that the system can implement are TLS VPN or IPsec VPN in tunnel
mode. Both modes reach the same conclusion of secure data traffic and communication
through slightly different means. A TLS VPN lives in the application layer, layer 4.
IPSec VPNs live in the network layer, layer 3. In other words, “The major difference
between an IPsec VPN and an SSL VPN comes down to the network layers at which
encryption and authentication are performed” (Loshin, 2019). Again, both VPNs provide
a private, secure network for communication, data storage, data transmission and network
operations. The project manager can implement whichever one his engineering team
recommends.
The above criteria, data, and information can be combined for model use to output the
Courses of Action (COAs). The model will use the data as inputs and produce COAs that
show cost-reduction, environmental benefit, scheduling patterns, transportation method,
 Cybersecurity Considerations for the Design 67

destinations, and departure suggestions. These COAs will be the optimal transportation
strategy and give management options to pursue.

3.1 Identifying and Structuring the Data

The goal of our system is to provide transportation efficiency through automation, costs-
reducing methods, and increase sustainability of the transportation ecosystem. Artificial
intelligence (AI) has the potential to improve the intermodal freight transportation sys-
tem’s efficiency by optimizing operations in real-time and self-evolving the models to
make better/faster decisions. The system accomplishes this goal. For the system to give
accurate predictive and reactive solutions, there needs to be data structure, software sys-
tem, and hardware system requirements of the system. When implemented correctly, the
system will output a corrective course of action that is an optimal solution.
Identifying and establishing a data structure is the beginnings of building out the
system. The data structure needs to be set up in a way that optimizes the dashboard’s
utilities. The structure of the data will be in a data frame form. What this will do is
allow for data to be structured cleanly and in a format that can be queried. A data frame
is defined as “a data structure that organizes data into a 2-dimensional table of rows
and columns, much like a spreadsheet” (“Data Frames,” 2022). This allows for easy
sorting and indexing of the raw data. The data structure will be set up very similarly
to an excel spreadsheet with numbered rows and columns. Data frames are “one of the
most common data structures used in modern data analytics because they are a flexible
and intuitive way of storing and working with data” (“Data Frames,” 2022). This ease of
implementation and flexibility will pay dividends once queries, analysis and data start
to flow through. As an AI-based model that will be learning from the data provided, the
flexibility of data frames allows for results to be optimized.
The proposed decision support system follows as an extension of Port Community
Systems (PCS), Freight optimization systems, and Freight brokering systems. A PCS is
a platform that connects the various stakeholders in a port community, such as shipping
companies, freight forwarders, port authorities, and customs officials, to facilitate com-
munication, coordination, and collaboration. Data is a crucial component of a PCS, and
various data types are involved in the system (Moros-Daza et al., 20167, 2018, 2020).
Vehicle routing optimization (VRO) is the process of finding the most efficient routes for
vehicles to transport goods or people. Data is a crucial component of VRO, and various
data types are involved in the optimization process (De la Cruz et a., 2013, Amador et al.,
2014, Palma et al., 2019). A freight brokering system is a platform that connects ship-
pers with carriers to facilitate the movement of goods. Data is a crucial component of a
freight brokering system, and various data types are involved in the system (Moros-Daza
et al., 2019). Sarabia et al. (2006), present simulation-based decision support models as
computer-based tools used to optimize river cargo transportation. These models simu-
late the river system and cargo flows to help decision-makers identify the most efficient
transportation routes and schedules. Various data types are involved in SDSM for river
cargo transportation. The typical data types that need to be consider in such systems
follow (Table 1).
68 C. Paternina-Arboleda et al.

3.2 System Data Flow Model

This section focuses on the design of the data flow system that enables real-time synchro-
modal transportation optimization using digital twin technology and AI. The proposed
system aims to provide a comprehensive view of the transportation network, including
infrastructure, assets, and operations, to optimize the transportation flow and minimize
carbon emissions. The system’s design includes the integration of various data sources,
such as real-time sensor data, historical transportation data, and external data sources,
to enable the AI algorithms to make informed decisions in real-time. Figure 4 shows the
context diagram for the system data flow.
The primary data input into the system at regular intervals will be the GPS technology
tracking the containers. The GPS tracking will be used by the system to create the content
twin of the containers as they are moved between modes of transport from starting point
to destination. Supplemental information such as the current engine being used to move
the container, current starting point, and destination will also need to be gathered each
time the container changes mode of transport until the container reaches its destination.
This information will allow the system to track CO2 emissions based on the engine
being used for transport and calculate anticipated & actual CO2 emissions based on the
starting point, destination, and actual miles traveled.
To create the content twins of weather and traffic conditions, real-time weather,
and traffic data APIs such as Weather.gov, WeatherAPI.com and Google Traffic will be
utilized to access the available data from the cloud. With this information collected, the
system will be able to monitor the weather conditions and local traffic conditions where
transport is taking place. Adjustments to routes will be made possible by analyzing the
weather and traffic data allowing the system to make routing decisions in a timely and
emission friendly manner. Figure 5 shows the context diagram of the system data flow.

Fig. 5. Context Diagram of System Data Flow

 Cybersecurity Considerations for the Design 69

The core component twins of the container, weather, and traffic will work together
to form three distinct asset twins within the digital twin system. Container and weather
component twins will form a single asset twin to study the interactions of the container
routes and weather. Similarly, the container and traffic component twins will form a
single asset twin to study the interactions between container routes and traffic.
Finally, the weather and traffic component twins will form a single asset twin to study
the interactions between weather and traffic where intermodal transport occurs. The data
collected from these three asset twins will be processed and analyzed by the system
for optimization of routes, schedules, and CO2 emission reduction during intermodal
container transport, forming the complete digital twin data structure (Fig. 6).

Fig. 6. High-Level Data Structure of the Digital twin system.

3.3 Dynamic Data Acquisition

Real-time data can be obtained from a variety of sources. Advanced analytics and ML
algorithms can be used to process and analyze real-time data, providing insights and sug-
gestions to stakeholders in real-time. Table 1 specifies the update frequency, latency, and
spatial granularity of dynamic data. The goal is for system level optimizations to be com-
puted within a day and dynamic adjustments in under 1 h. Real-time data can be critical
for optimizing intermodal transportation networks by providing up-to-date information
on infrastructure, traffic patterns, weather conditions, and other factors that can impact
transportation operations. Some common types of RT data used in intermodal transporta-
tion optimization include Vessel AIS, GPS data, Rail freight traffic, road Traffic, Port
traffic, Weather, Inventory and shipment data, and data from Customer Experience (CX)
and Customer Relationship Management (CRM) platforms along with feedback chan-
nels (for which Natural Language Processing models are needed). All the above sources
can provide RT or near-RT information. Effective use of RT data improves network
optimization, leading to increased efficiency, reduced costs, and improved customer
satisfaction.
70 C. Paternina-Arboleda et al.

Table 1. Sample Data Sources

Requirement Historic data source Dynamic data source: freq., latency,

spatial dimension
Maritime freight traffic Same as dynamic, 60 days active, 4+ Customs and Border Protection,
years in cold store, Port of Los Manifest for the Port of Los Angeles
Angeles 5+ years Wharfinger, Spire This applies to the main ocean carriers
except for COSCO, Wan Hai shipping
lines
Rail freight traffic Rail data provider such as Railinc Rail data provider such as Railinc
Road freight traffic 5+ years in-gate/out-gate for POLA POLA In/out-gate (Excludes Road
traffic), GeoStamp
Maritime routes Reports from ocean carriers’ web Dynamic routes accessible from
carriers’ websites
Rail routes Class 1 and short lines, Wabtec Class 1 and short lines, Wabtec
Inland Waterways routes Inland waterways Carriers Information on the river system, such
as its topology, current, and water
levels. This data is used to simulate the
river system and the flow of cargo on it
Truck routes Dray providers Providers (area of focus). Data can be
provided daily
Port traffic 2+ Y historical of Marine Exchange, Marine Exchange, GeoStamp Vessel to
Spire Global API, Live data for ALL Port ETA functionality. Ability to call
vessels within USA EEZ on all historical data Spire has (back to
2011)
Security Data Own source Security threats, risk assessments, and
incident reporting. This data is used to
ensure the safety and security of the
port community
Shipment Data On demand Freight data O/D, type of goods being transported,
and desired delivery date. Used to
match shippers with carriers to meet
needs
Weather Data Several sources The latency and frequency of weather
data can vary depending on the source
and the type of data. Data collected
from ground-based sensors may have
lower latency and higher frequency
than weather data collected from
satellites. Similarly, certain types of
weather data, such as temperature and
precipitation, may be collected and
updated more frequently than other
types of weather data, such as cloud
cover or wind speed

4 Cybersecurity Considerations

The CIA triad, which stands for Confidentiality, Integrity, and Availability, is a fun-
damental concept in cybersecurity that applies to various systems and technologies,
including digital twins. Let’s discuss the CIA triad in the context of ports, container, and
shipping digital twin systems:
 Cybersecurity Considerations for the Design 71

Confidentiality: Confidentiality in the context of digital twin systems for ports, con-
tainers, and shipping refers to protecting sensitive information from unauthorized access.
It involves ensuring that only authorized individuals or entities can access and view the
data related to the digital twin system. Confidentiality measures can include strong
access controls, encryption of sensitive data, and secure communication channels to
prevent eavesdropping or data interception. For example, in a shipping digital twin, con-
fidential information may include cargo details, shipping schedules, or trade secrets of
the involved parties. Safeguarding this information is critical to prevent unauthorized
disclosure or exploitation by malicious actors.
Integrity: Integrity refers to the assurance that the data and information within the
digital twin system remain accurate, unaltered, and trustworthy. It involves protecting
the data from unauthorized modifications, ensuring that it reflects the true state of the
physical system being represented. In the context of ports, containers, and shipping
digital twins, maintaining data integrity is crucial to avoid manipulation or tampering that
could result in incorrect decisions, compromised safety, or financial losses. Implementing
integrity controls such as data validation, digital signatures, and audit trails can help
ensure the reliability and authenticity of the digital twin’s data.
Availability: Availability refers to ensuring that the digital twin system is accessible
and operational whenever it is needed. It involves safeguarding the system against disrup-
tions, outages, or denial-of-service attacks that could render it inaccessible or unusable.
For ports, containers, and shipping digital twins, availability is vital for maintaining
smooth operations and efficient logistics. Downtime or unavailability of the digital twin
system can lead to delays, inefficiencies, and potential financial losses. Implementing
robust backup and disaster recovery mechanisms, redundancy, and proactive monitoring
can help ensure high availability of the digital twin system.
In brief, the CIA triad provides a comprehensive framework for addressing cyber-
security concerns in ports, container, and shipping digital twin systems. By prioritizing
confidentiality, integrity, and availability, organizations can establish a strong security
foundation to protect sensitive data, maintain the accuracy of information, and ensure
the continuous and reliable operation of their digital twin systems.

4.1 System’s Cybersecurity Factors

The proposed optimization system will have cyber-attack resiliency through address-
ing key security factors. These key security factors aim to address the confidentiality,
integrity, and availability (CIA) cybersecurity triad. Although the following factors are
not fully encompassing of all security measures, they are the foundational factors for
basic cyber hygiene.
Access controls is the first step to ensure data confidentiality. Limiting access to
the system data lakes in the cloud, locally on premises, or off premises is essential to
denying cyber-attacks. Access controls also play the role of giving the system’s end user
the proper level of access needed to do their jobs. For example, a system administrator,
security analyst and truck driver will all need different levels of access. The later needing
access to their specific data that is relevant to their logistic routes. Along with access
control, authentication measures that match users with their digital identity is critical. The
foundational authentication measures include using FIDO-compliant technology, a PKI
72 C. Paternina-Arboleda et al.

methodology, and asymmetric encryption methods. A more in-depth description on these

authentication measures will be covered in the next section regarding protecting privacy
of sensitive data. The system will need to establish an asset management program to track,
monitor and update assets as needed to complete orders. By having a comprehensive list
and inventory of assets allows the security team to manage the security requirements
of each asset and their associated components. The system assets fall into four general
categories that need to be managed. These categories are central hub assets, digital assets,
transportation assets, and local driver assets. Each category will have their own asset
inventory list with full detail of the assets themselves. Knowing what assets and data
goes along with those assets is critical to being able to protect them from cyber-attacks
and implement a change management program. The change management program is
another key cybersecurity measure to consider when designing the system. The goal
is to track and document all proposed engineering changes to individual subsystems
and components. These changes will be analyzed with a wholistic view. This validates
that the changes to each subsystem or component will not degrade or eliminate the
functionality of the system.
Another key security factor is being able to protect the data and system in case of
catastrophic loss through an attack. The cybersecurity measure of Backup and Restore
(BAR) is vital to success. The copying of physical or virtual files or databases to a sec-
ondary location for preservation in case of equipment failure or catastrophe, and backups
provide a way of restoring deleted files or recovering a file when it is accidentally over-
written or becomes corrupted or may support recovery from a cyber-incident. To ensure
successful backup and restore functionality, a full periodic full backup of the digital
twin will be required at the current baseline. A full backup is the most complete type
of backup where you duplicate all the selected data of the digital twin’s configuration.
This includes files, folders, SaaS applications, hard drives and more; while a full backup
requires minimal to restore data, it takes longer to backup compared to other types of
backups and places a burden on required storage space. When you back up data, all the
information on your disk is saved as a single file called an image. If your digital twin
crashes or gets corrupted, you may lose some or all your data. You can use this file to
restore your digital twin exactly back to how it was before data loss.
Security monitoring is an important measure to implement to mitigate potential
security violations and cyber-attacks. The Security Orchestration, Automated Response
(SOAR) method security alert subsystem will have to be implemented to respond to the
security alerts and potential attacks. This automated response method, SOAR, builds off
the previously mentioned SIEM (security information and event management) alerting
structure. Although the Security Incident/Event Management Subsystem is sufficient to
accomplish security monitoring, SOAR takes things one step further by automating an
actual response to help mitigate the risk as quickly as possible. Although SOAR is a
more recent method to accomplishing security alert system along with the follow-on
response playbooks. This is a cost-saving and efficiency optimization measure. Security
alerts responses will largely be automated except for egregious violations.
SOAR improves the security operations when the subsystem is allowed to “automate
investigation path workflows can significantly cut down on the amount of time required
to handle alerts. They also provide lessons about the security admin skill set required
 Cybersecurity Considerations for the Design 73

to complete an investigation path” (Froehlich, 2023). This analysis by former Cisco

Enterprise-level IT security professional, Andrew, points out two key factors of SOAR
that provide an added security layer of responses to security alerts.
To effectively protect an AI-driven real-time optimization information system for
synchromodal freight operations, it is essential to utilize industry-standard cybersecurity
frameworks and guidelines such as the Security Technical Implementation Guideline
(STIG), ISO 27002, and NIST standards.
As with the Security Technical Implementation Guideline (STIG), these guidelines
are specifically designed for the technology stack and components used in the AI-
driven system. Regularly update and patch software and hardware components as per
STIG guidelines to address vulnerabilities and protect against emerging threats. Conduct
periodic STIG compliance assessments and audits to ensure ongoing adherence to the
guidelines.
ISO 27002 could also be adopted as a comprehensive framework for establishing
and maintaining an information security management system (ISMS). The IT managers
could conduct a risk assessment to identify potential security risks and vulnerabilities in
the AI-driven system. Develop and implement security policies, procedures, and controls
based on ISO 27002 to mitigate identified risks. Regularly monitor, review, and update
the security controls to align with evolving threats and organizational requirements.
Ensure compliance with ISO 27002 through internal audits and periodic assessments.
Last, follow the NIST Cybersecurity Framework, which provides a flexible and cus-
tomizable approach to managing and improving cybersecurity. Identify and categorize
the AI-driven system’s assets, assess risks, and develop risk mitigation strategies follow-
ing NIST guidelines. Implement appropriate security controls, such as access controls,
encryption, intrusion detection systems, and incident response plans, as recommended
by NIST. Continuously monitor, detect, respond to, and recover from security incidents
by aligning with NIST’s incident response and recovery best practices.
By leveraging the Security Technical Implementation Guideline (STIG), ISO 27002,
and NIST standards, organizations can establish a robust security posture for their AI-
driven real-time optimization information systems in synchromodal freight operations.
These frameworks provide comprehensive guidance and best practices to mitigate risks,
protect sensitive data, and ensure the system’s resilience against cyber threats.

4.2 System’s Cybersecurity Risks

A digital twin-enabled, AI-driven optimization system for intermodal freight operations
is a semi-autonomous system that leverages digital twins and Artificial Intelligence (AI)
to optimize logistical operations. This is accomplished by collecting raw data from active
operations, analyzing real-time goods movement data, and generating insights on how
to improve the system.
The first cybersecurity risk that arises is faulty code that is at fault due to malicious
actors. A large part of cybersecurity checks and evaluations has historically been solely
at the end of the development process. This is dangerous for a variety of reasons and
allows these malicious actors to start their cyber-attack early in the process and hide
a trojan-type virus into the system’s code. To field a testable digital twin that mimics
systems in production, the code cannot have this potential vulnerability.
74 C. Paternina-Arboleda et al.

The first code type that logically should be secured is the source code. The source
code that is used as a baseline configuration for all development could potentially be
compromised and go undetected until deployment. This type of source code data poi-
soning and embedded viruses can hide if the malicious actor enters them smartly in
the beginning of the development process. These data vulnerabilities can be part of
the implementation process and deployed to production units without the developer’s
knowledge. Drivers and cyber operations dashboards would not be able to track security
alerts and be utilized as they were intended. To mitigate this risk, a best practice known
as “Shift-Left” can be implemented into the systems’ security posture.
The Development, Security, Operations (DevSecOps) process incorporates secu-
rity into the early stages of the software development life cycle. According to IBM,
DevSecOps “integrates application and infrastructure security seamlessly into Agile
and DevOps processes and tools. It addresses security issues as they emerge when they
are easier, faster, and less expensive to fix. Additionally, DevSecOps makes applica-
tion and infrastructure security a shared responsibility of development, security, and IT
operations teams, rather than the sole responsibility of a security silo” (“DevSecOps,”
2023). DevSecOps gives developers access to security tools throughout the process so
that there are no integration problems at the end. The security vulnerabilities and flaws
would be shifted left and found sooner rather than later.
“Shift left”, promotes the integration of security measures into the early stages of
the DevOps process (“DevSecOps”, 2023). By moving security from the right (end)
to the left (beginning) of the development cycle, software engineers are encouraged to
prioritize security from the very beginning of the project. In a DevSecOps environment,
cybersecurity architecture and development engineering is incorporated as part of the
development team to guarantee that every phase has security measures. Furthermore,
the configuration items in the stack are securely patched, configured, and documented.
This will benefit change management and version control measures. By adopting a shift
left approach, DevSecOps teams can detect security risks and vulnerabilities early and
take immediate action to address them. This ensures that not only is the product being
built efficiently, but that security is also being implemented simultaneously.
The next cybersecurity risk that will be relevant to the system is the risk of lateral
movement, network device compromises and privilege escalation after bypassing the
initial access controls and authentication stages. This risk is because of the increased
attack surface of traditional cybersecurity measures that focus on this initial access
barrier (North/South traffic) and allow freedom to move around within the network
(East/West traffic). To mitigate this risk, using zero trust cybersecurity practices would
be a sufficient mitigation path. Zero Trust Architecture (ZTA) takes an orthogonal view
of cybersecurity by accounting for internal movement and east/west traffic within a
network. This lets ZTA be viewed as “data-centered” rather than “user-centered”. ZTA
secures the data that flows all through the network while traditional measures attempted
to secure the users in entering the network. The United States Government Accountabil-
ity Office (GAO), states “Zero trust architecture (ZTA) is a cybersecurity approach that
authenticates and authorizes every interaction between a network and a user or device”
(Zero Trust Architecture, 2022). In a normal cybersecurity paradigm, a network user is
allowed to roam freely east and west through the network once access is granted. An
 Cybersecurity Considerations for the Design 75

attack surface is much smaller using a zero-trust mindset because there are consistent
authorization measures that are implemented into the system design. This will be imple-
mented as Policy-as-Code and will be automated through firewall parameters, router
rules and ACLs. ZTA will be very important for the system because of the data struc-
ture and hybrid-cloud environment. The GAO describes the importance of ZTA with
some background. “Given the increasingly complex nature of IT networks, including
cloud and hybrid environments, ZTA’s goals are to reduce opportunities for attackers by
restricting access and to detect attacks by monitoring user behavior and other network
activity” (Zero Trust Architecture, 2022). Given the hybrid data storage and data move-
ment paths through secure VPNS, this major reduction of attack surface will directly
lower system’s security risk.
The last major cybersecurity risks the system will be faced with is being in the dark
about security threats and attacks. The system will need the ability to receive notifications
of cyber security attacks. Whenever there is data out there that can be compromised, the
system vulnerabilities will be attacked by cyber criminals. The criminals hope to gain
some sort of gain. Whether that gain be a financial, competitive advantage, competitor
insights or some sort of notoriety. When these cyber-attacks occur, the system will have
a prompt, security alert and security management subsystem to alert security analysts of
the breach. As mentioned, the Security Incident/Event Management (SIEM) subsystem
is sufficient to mitigate this cybersecurity risk and accomplish security monitoring. This
subsystem will be expanded on in the next section to automate the responses. However,
to automate a response to a security breach, a type of SIEM subsystem is the foundation
for mitigating security alerts that will arise. The SIEM tools are “a way to centrally
collect pertinent log and event data from various security, network, server, application,
and database sources. SIEMs then detect and alert on security events” (Froehlich, 2023).
This allows security analysts to review the log data, identify the threat, isolate, and
mitigate as appropriate. As mentioned, this will be automated by another subsystem in
later phases. Another benefit of the SIEM subsystem is that “Aggregated data is analyzed
by the SIEM in real time to spot potential security issues. Because multiple data sources
are analyzed, the SIEM identifies threats by correlating information from more than one
source” (Froehlich, 2023). The system will leverage this feature to collect and store data
at each local host and aggregated host data at the hub. Additionally, “The SIEM then
intelligently ranks the events in order of criticality” (Froehlich, 2023). The SIEM AI
feature that stacks and ranks the security alerts will be very beneficial to cost savings
and efficiency. Buckbee (2022) states some of the top SIEM tools used including Splunk,
IBM QRadar, and Logrhythm.
Table 2 shows a summary of the cybersecurity considerations and the measures to
counteract.
76 C. Paternina-Arboleda et al.

Table 2. Cybersecurity threats and measures for system design

Cybersecurity threat Measures suggested for system design

Malware and Ransomware Attacks Implement Strong Endpoint Protection, Patch and Update
Software, Conduct Employee Awareness Training, Enable
Email Filtering and Spam Detection, Implement Network
Segmentation, Regularly Backup and Encrypt Data,
Enforce Strong Password Policies, Maintain Incident
Response and Recovery Plans, Monitor Network Traffic
and Behavior, Regularly Test and Update Security
Measures
Phishing and Social Engineering Employee Education and Awareness, Implement Email
Filtering and Spam Detection, Multi-Factor Authentication
(MFA), Encourage Reporting of Suspicious Activities,
Strong Password Policies, Incident Response and Mock
Exercises, Implement Web Filtering
Insider Threats User Access Control, User Behavior Monitoring,
Employee Awareness and Training, Confidentiality
Agreements and Policies, Regular Security Awareness
Programs, Segregation of Duties, Employee Offboarding
Processes, Cultivate a Positive Work Environment
Distributed Denial of Service (DDoS) Network Monitoring and Traffic Analysis, DDoS
Attacks Mitigation Services, ensure a Scalable Network
Infrastructure, Traffic Filtering and Rate Limiting, Content
Delivery Network (CDN) Services, Deploy Intrusion
Detection and Prevention Systems (IDPS), Incident
Response Planning, Deploy Redundancy and Failover
Systems, Traffic Scrubbing and Blackholing, Regular
Testing and Preparedness
Data Breaches and Unauthorized Access Data Encryption, Access Controls and Authentication,
Secure Network Infrastructure, Vulnerability Assessments
and Penetration Testing, Employee Education and
Awareness, Data Backup and Disaster Recovery (Business
Continuity), Intrusion Detection and Monitoring, Incident
Response Planning, Data Privacy and Compliance, Regular
Security Audits and Compliance Assessments
Supply Chain Attacks Vendor Risk Management, Secure Development Lifecycle,
Continuous Monitoring, Multi-Factor Authentication
(MFA), Secure Software and Firmware Updates, Secure
Configuration Management, Incident Response and
Recovery Planning, Threat Intelligence Sharing, Employee
Awareness and Training, Supply Chain Resilience, Use
secure protocols, such as HTTPS or SFTP, for transmitting
origin/destination data over networks, Minimize the
collection, storage, and retention of origin/destination data
to reduce the potential impact of a data breach, Third-Party
Risk Management
IoT and Operational Technology (OT) Implement strong security controls for IoT and OT devices,
Vulnerabilities Network Segmentation, Access Control and Privileged
Access Management, Regular Patching and Updates,
Secure Configuration Management, Network Monitoring
and Intrusion Detection, Security Testing and Penetration
Testing, Vendor Management and Supply Chain Security,
Employee Training and Awareness, Incident Response and
Business Continuity Planning
 Cybersecurity Considerations for the Design 77

4.3 Preserving Security and Privacy of Sensitive Data

There is a large need for the system to have overarching security protections on sensitive
data to protect customer, supplier, and employee information. Data Privacy rulesets are
baked into the source code that the system uses. Both data at rest and data in transit
should have cybersecurity measures built into the system design.
The primary method to store and secure the data will be in a cloud instance. This
allows the cloud storage infrastructure to be managed at the Hub and treats the production
units as Thin Clients. Thin clients are defined as “a virtual desktop computing model
that runs on the resources stored on a central server instead of a computer’s resources”
(Gillis, 2021). Both cloud and hard drive based local backup methods have advantages
and disadvantages. To take advantage of both and fully address any possible data loss,
a hybrid backup strategy is recommended, where backups are created locally and in the
cloud. The 3-2-1 backup rule industry standard describes that the data is kept in three
places, across two media, with one backup stored offsite, such as in the cloud.
Local hard drive data storage will be utilized on an incremental, scheduled basis. This
local, hard drive backup refers to the process of backing up the system, applications, and
data to a device located internally or as a connected component, such as tape, disk, hard
disk, flash drive, CD, external hard drive, or other media. This type of backup will be
located and kept solely by data custodians located on site at the Hub. Due to the growing
storage capacity restraints for the massive amounts of data that gets produced, this will
not be a full data storage strategy. However, a very important data storage strategy if the
cloud was ever compromised.
On-Prem means the data is stored on the Hub’s premises but in a cloud storage
location. Cloud based data storage on-premises backup refers to the process of backing
up the system, applications, and data to a public or private cloud that is located on the
Hub premises. The cloud will be configured as a secure database that collects the data
and stores it into the centralized servers. These servers are located on site at the Hub.
Off-premises backup refers to the process of backing up the system, applications, and
data to a public or private cloud that is located off the premises. (i.e., on a production
unit). This will be the primary method of collecting and storing data for the different
transportation methods and personnel involved. Off Prem is when the data is not part of
the centralized cloud storage infrastructure. Rather, a means to store local data of driver
information and movement in a separate, secure cloud storage that acts as a thin client.
Thin clients are very useful because it reduces the cyber-attack surface and protects the
data that is generated by each unit. However, there is a drawback. This drawback being
“Normally thin clients take the form of low-cost computing devices that heavily rely on
a server for computation” (Gillis, 2021). This puts the computational power back onto
the server side located at the Hub instead of the clients themselves on the truck, train, or
boat.
Protecting data at rest within the Hub database and storage locations can be accom-
plished by enforcing three key security procedures of database encryption, access and
authentication (“Protecting data at rest,” 2023). Enforcing access controls and proper
authentication using MFA points protects the core cybersecurity triad. Confidentiality,
data integrity and availability are protected because there is no unauthorized access to the
data being stored. All data that is meant to be seen by authorized personnel is available.
78 C. Paternina-Arboleda et al.

If there is unauthorized access, the database encryption is the fallback security measure
that ensures the sensitive data is protected. The method to protect sensitive data in transit
is to set up an enterprise level Virtual Private Network (VPN) and direct, secure con-
nection between the Hub and fielded units. This method of protecting sensitive data in
transit allows the system to mitigate security risks inherent to providing remote network
access by offering strong encryption to provide data security and strong authentication,
to limit access to applications based on security policies (Loshin, 2019).
Having the ability to hide the identity of information is another way to protect
private data, as it would make it difficult to pinpoint whose data it is. QueryPie (2022)
goes into detail about the difference between anonymization and pseudonymization, but
how both can be beneficial to protecting personal data. Anonymization is the process of
eliminating all identifiers from the data, whereas pseudonymization uses a pseudonym
identifier takes place of the actual identifier of the data.
Cybersecurity and data privacy are prioritized steps to ensuring safe use of the pro-
posed optimization system. To reach the appropriate level of security and privacy for
the system, there are several legal and regulatory issues that must be addressed. Intellec-
tual Property (IP) rights are one of the leading legal obstacles that must be considered
when using AI-driven models and algorithms. Medeiros and Sanft (2018) state that IP
is associated with inventions, brands, new technologies, source code, and artistic work
that have patents, trademarks, and copyright ownership.

5 Key Challenges in Integrating Cybersecurity and Data Privacy

Considerations into the Proposed Optimization System
There are many cybersecurity and data privacy concerns that will need to be addressed
before design and implementation occurs. This section will address and dive into three
key challenges that the system will face integrating cybersecurity and data privacy.
After the key challenge is identified and defined, the follow-on framework to address
the privacy concern will be laid out.
Key challenge 1 (KC1) touches on both cybersecurity and data privacy considera-
tions of the system through policy implementation. This key challenge is a cybersecurity
policy implementation measure while also striving to protect data privacy. The second
key challenge, KC2, will be to keep the system up to date with security and vulnera-
bility guidance on the system components and subsystems. This mainly addresses the
cybersecurity consideration portion. The third key challenge, KC3, is a pure data privacy
consideration that involves securing personal data. The framework for this key challenge
builds off the FIDO-based security technology that includes a public key infrastructure
(PKI) and asymmetric encryption.
The first key challenge that the system will be faced with is policy implementa-
tion and policy enforcement. Cybersecurity policies implementation and enforcement
is important because these policies “defines which conditions must be met in order for
a code to pass a security control and be deployed” (Carroll, 2023). The system will
be leveraging the DevSecOps process with built in policy implementation that act as
guardrails. If these policies are violated, the code is considered to be not secure and
should be flagged as a security alert. To address this policy implementation challenge,
 Cybersecurity Considerations for the Design 79

Policy-as-Code (PaC) will be embedded to automate the process. Policy-as-Code (PaC)

is defined as “an approach to policy management in which policies are defined, updated,
shared, and enforced using code. By leveraging code-based automation instead of rely-
ing on manual processes to manage policies, policy-as-code allows teams to move more
quickly and reduce the potential for mistakes due to human error” (Carroll, 2023). PaC
allows policy enforcement and implementation to be automated, cheaper to audit, and
allows for more efficiencies throughout the system. Implementing the PaC framework
to put up guardrails during initial design and fielding puts security in a better posture as
well. PaC is set up to be a collaboration tool as well. The system can use “a policy-as-
code approach to domains like security makes it possible to define and manage policies
in ways that different types of stakeholders – such as developers and security engineers –
can understand” (Carroll, 2023). This improves the developer and manager collaboration
and project communication. Regardless of technical skill or policy expertise, the PaC
explicitly outlines the policies and sets up enforcement guidelines at the same time.
The next key challenge to cybersecurity and data privacy is to address security
technical implantation guidance (STIGs) updates and procedures. STIGs are used “for
verifying that the product has been configured properly, and/or for identifying unautho-
rized configuration changes to the product” (Quinn et al., 2018). Historically, this was
very much a manual process of reviewing the STIGs and finding out how they differ than
what is currently fielded. STIGS can be automatically delta checked and implemented
through an automated checklist program. The checklist will be designed like how the
U.S. government uses the National Checklist Program (NCP). The NCP is “the U.S. gov-
ernment repository of publicly available security checklists (or benchmarks)” (National
Checklist Program, 2023). These components and subsystems that will need constant
security implementation include operating systems, AI-model, the SIEM subsystem,
database, and cloud services (i.e., AWS), firewalls, routers, switches, network logging
subsystem, vulnerability subsystem, audit collection subsystem and the virtual hosting
environment. The NCP is beneficial because “Applying checklists to operating systems
and applications can reduce the number of vulnerabilities that attackers can attempt to
exploit and lessen the impact of successful attack” (Quinn et al., 2018). Essentially, the
checklist is a proactive security process that shores up the system for known vulnerabil-
ities. The NCP is a checklist “that provides detailed low-level guidance on setting the
security configuration of operating systems and applications” (Quinn et al. 2018). NIST
identifies the main security benefit of using the NCP is to “minimize the attack surface,
reduce vulnerabilities, lessen the impact of successful attacks, and identify changes that
might otherwise go undetected” (Quinn et al., 2018). Lower impact, reduced vulnera-
bilities and less attack surface leads to lower project cost, heightened security posture
and faster project implementation.
The last key challenge that will be addressed that poses a risk to the system operations
is data privacy and governance of a data privacy program. The governance portion really
looks at the business operations and aligns the data privacy measures accordingly. Using
the appropriate level of data privacy measures and establishing a framework for the
system is how this key challenge can be addressed.
80 C. Paternina-Arboleda et al.

6 Managerial Insights
This article highlights the critical managerial insights for designing an AI-driven dis-
tributed optimization system with a strong focus on cybersecurity considerations. By
adopting a cybersecurity-first approach, investing in robust cybersecurity measures, fos-
tering awareness, collaborating with industry partners, conducting regular risk assess-
ments, and planning for incident response and recovery, managers can enhance the secu-
rity and resilience of the system, protecting valuable data and ensuring the successful
implementation of container carbon emissions reduction in freight operations.
It highlights the importance of adopting a cybersecurity-first approach when design-
ing an AI-driven distributed optimization system for container carbon emissions reduc-
tion in freight operations. Managers should prioritize cybersecurity considerations right
from the system’s design phase to mitigate potential risks and safeguard sensitive data.
It highlights the importance of investing in robust cybersecurity measures. Given
the increasing connectivity and interdependence of systems in the freight industry,
it is imperative to invest in robust cybersecurity measures. Managers should allocate
resources for implementing encryption mechanisms, access controls, anomaly detection
systems, and other cybersecurity solutions to protect the system against cyber-attacks
and unauthorized access.
The article also highlights the importance of fostering a culture of cybersecurity
awareness. It is essential to conduct regular risk assessments and audits to identify
vulnerabilities, evaluate the effectiveness of implemented cybersecurity measures, and
ensure compliance with relevant industry standards and guidelines. Managers should
allocate resources for periodic assessments, penetration testing, and audits to detect and
address any weaknesses or gaps in the system’s security.
Last, the article shows the importance of planning for incident response and recovery.
Despite preventive measures, it is important to have a well-defined incident response
plan in place. Managers should establish protocols for detecting, responding to, and
recovering from cybersecurity incidents. This includes defining roles and responsibili-
ties, establishing communication channels, and regularly testing incident response plans
through tabletop exercises and simulations.

7 Conclusions
The design of a digital twin-enabled AI-driven real-time distributed optimization system
for synchromodal freight operations has the potential to significantly reduce carbon
emissions and increase efficiency in the transportation industry. However, with the use
of advanced technologies also comes the potential for cybersecurity risks.
This article highlights the key cybersecurity considerations that need to be addressed
in the design of such a system. These include securing the data flow, ensuring the security
of the digital twin models, preserving the privacy for all stakeholders, and protecting the
system against cyber-attacks. The proposed solutions to these challenges include the use
of encryption, access control, and anomaly detection systems, and mitigation strategies
when an attack happens, among others.
The success of the proposed system depends on the implementation of these cyber-
security measures. The transportation industry must prioritize cybersecurity to ensure
 Cybersecurity Considerations for the Design 81

the safety and integrity of the transportation network and its stakeholders. By adopting a
cybersecurity-first approach, the users of this system can mitigate risks and ensure that
the benefits of the system are fully realized.
Future research should focus on further refining the proposed cybersecurity solutions
and addressing new cybersecurity challenges that may arise as the system is deployed.
One area that requires attention is the potential impact of emerging technologies, such as
quantum computing, on the security of the system. Quantum computing has the potential
to break the encryption algorithms currently used to secure data, and thus, there is a need
to explore new encryption methods that are resistant to quantum attacks.
Another area of interest is the development of improved, more automated machine
learning algorithms that can detect and prevent cyber-attacks in real-time. Traditional
cybersecurity measures, such as firewalls and intrusion detection systems, may not be
sufficient in detecting sophisticated cyber-attacks that can exploit system vulnerabilities.
Machine learning algorithms can learn from past cyber-attacks and detect new and
unknown cyber-attacks, leading to better and more effective cybersecurity measures.
Furthermore, future research should also explore the potential of blockchain technol-
ogy in securing the system. Blockchain technology provides a decentralized and tamper-
proof data storage and transaction control mechanism that can enhance data security and
integrity, making it a promising technology for securing digital twin-enabled AI-driven
real-time distributed optimization systems.
Overall, this paper describes certain topics as to how a cybersecurity-aware design
approach can enable the successful implementation of a digital twin-enabled AI-driven
real-time distributed optimization system for synchromodal freight operations, leading
to significant benefits for the transportation industry and the environment.

References
Cost of a Data Breach Report 2022. IBM (2022). https://www.ibm.com/reports/data-breach?
utm_medium=OSocial&utm_source=Blog&utm_content=SSSWW&utm_id=Security-Intell
igence-Blog-Banners%20&_ga=2.169531361.1119445732.1681783239-1093745095.168
1783239. Accessed 14 Apr 2023
What is DevSecOps? IBM (2023). https://www.ibm.com/topics/devsecops#Benefits%20of%20D
evSecOps. Accessed 17 Apr 2023
Zero Trust Architecture. GAO (2022). https://www.gao.gov/products/gao-23-106065. Accessed
15 Apr 2023
Protecting data at rest. AWS Inc. (2023). https://docs.aws.amazon.com/wellarchitected/latest/sec
urity-pillar/protecting-data-at-rest.html. Accessed 17 Apr 2023
US Department of Commerce, National Oceanic and Atmospheric Administration: API web
service (2023). https://www.weather.gov/documentation/services-web-api. Accessed 10 Apr
2023
NIST. Post-quantum cryptography: CSRC (2017). https://csrc.nist.gov/projects/post-quantum-cry
ptography. Accessed 18 Apr 2023
NIST. Round 3 submissions - post-quantum cryptography: Computer Security Resource Cen-
ter (CSRC) (2023). https://csrc.nist.gov/Projects/post-quantum-cryptography/post-quantum-
cryptography-standardization/round-3-submissions. Accessed 17 Apr 2023
NVIDIA. NVIDIA H100 Tensor Core GPU Datasheet. NVIDIA (2022). https://resources.nvidia.
com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet. Accessed 13 Apr 2023
82 C. Paternina-Arboleda et al.

NVIDIA. Triton Inference Server. NVIDIA Developer (2023). https://developer.nvidia.com/nvi

dia-triton-inference-server#ecosystem. Accessed 14 Apr 2023
NVIDIA. Nvidia TENSORRT. NVIDIA Developer (2023). https://developer.nvidia.com/tensorrt.
Accessed 14 Apr 2023
JSON and XML weather API and Geolocation Developer API. Free Weather API - Weather-
API.com (n.d.). https://www.weatherapi.com/. Accessed 10 Apr 2023
Abril, D., Velasquez-Bermudez, J., Paternina-Arboleda, C.D., Cantillo, V.: Integrated tactical-
operational event-driven real-time optimization framework for smart ports operations planning.
Marit. Econ. Logist. (2023, Submitted)
Aggarwal, B.K., Gupta, A., Goyal, D., Gupta, P., Bansal, B., Barak, D.D.: A review on investigating
the role of block-chain in cyber security. Mater. Today Proc. 56(Part 6), 3312–3316 (2022).
ISSN: 2214-7853. https://doi.org/10.1016/j.matpr.2021.10.124
Aggarwal, A., Dhurkari, R.K.: Association between stress and information security policy non-
compliance behavior: a meta-analysis. Comput. Secur. 124, 102991 (2023). ISSN: 0167-4048.
https://doi.org/10.1016/j.cose.2022.102991
Amador-Fontalvo, J.A., Paternina-Arboleda, C.D., Montoya-Torres, J.R.: Solving the hetero-
geneous vehicle routing problem with time windows and multiple products via a bacterial
meta-heuristic. Int. J. Adv. Oper. Manag. 6(1), 81–100 (2014)
Auld, J., Hope, M., Ley, H., Sokolov, V., Xu, B., Zhang, K.: POLARIS: agent-based modeling
framework development and implementation for integrated travel demand and network and
operations simulations. Transp. Res. Part C Emerg. Technol. 64, 101–116 (2016)
Beaumont, P.: Cybersecurity risks and automated maritime container terminals in the age of 4IR.
In: Information and Cybersecurity in the Fourth Industrial Revolution. IGI Global (2018).
https://doi.org/10.4018/978-1-5225-4763-1.ch017
Cammin, P., Yu, J., Voß, S.: Tiered prediction models for port vessel emissions inventories. Flex
Serv. Manuf. J. 35, 142–169 (2023). https://doi.org/10.1007/s10696-022-09468-5
Cobb, M.: What is the RSA algorithm? Definition from search security. Security (2021). https://
www.techtarget.com/searchsecurity/definition/RSA. Accessed 17 Apr 2023
De la Cruz, J.J., Paternina-Arboleda, C.D., Cantillo, V., Montoya-Torres, J.R.: A two-pheromone
trail ant colony system—tabu search approach for the heterogeneous vehicle routing problem
with time windows and multiple products. J. Heuristics 19, 233–252 (2013). https://doi.org/
10.1007/s10732-011-9184-0
De la Peña-Zarzuelo, I., Soeane, M.J.F., Bermúdez, B.L.: Industry 4.0 in the port and maritime
industry: a literature review. J. Ind. Inf. Integr. 20, 100173 (2020). ISSN: 2452-414X. https://
doi.org/10.1016/j.jii.2020.100173
Dzemydienė, D., Burinskienė, A., Čižiūnienė, K., Miliauskas, A.: Development of E-service pro-
vision system architecture based on IoT and WSNs for monitoring and management of freight
intermodal transportation. Sensors 23(5), 2831 (2023). https://doi.org/10.3390/s23052831
Froehlich, A.: SOAR vs. SIEM: What’s the difference? Tech Target (2023). https://www.techtarget.
com/searchsecurity/answer/SOAR-vs-SIEM-Whats-the-difference. Accessed 15 Apr 2023
Gillis, A.: Thin client (lean client). Tech Target (2021). https://www.techtarget.com/searchnetwor
king/definition/thin-client. Accessed 16 Apr 2023
Guo, X., He, J., Lan, M., Yu, H., Yan, W.: Modeling carbon emission estimation for hinterland-
based container intermodal network. J. Clean. Prod. 378, 134593 (2022). ISSN: 0959-6526.
https://doi.org/10.1016/j.jclepro.2022.134593
Jubiz-Diaz, M.A., Saltarin-Molino, M.A., Arellana, J., Paternina-Arboleda, C.D., Yie-Pinedo, R.:
Effect of infrastructure investment and freight accessibility on gross domestic product: a data-
driven geographical approach. J. Adv. Transp. 2021 (2021). Article ID: 5530114. https://doi.
org/10.1155/2021/5530114
 Cybersecurity Considerations for the Design 83

Lipner, S.B., et al.: Security kernels. In: Proceedings of the May 6–10, 1974, National Computer
Conference and Exposition on - AFIPS 1974, vol. 43, pp. 977–978 (1974). https://doi.org/10.
1145/1500175.1500361
Loshin, P.: IPsec vs. SSL VPN: comparing speed, security risks and technology. Tech Target
(2019). https://www.techtarget.com/searchsecurity/tip/IPSec-VPN-vs-SSL-VPN-Comparing-
respective-VPN-security-risks
Moros-Daza, A., Solano, N.C., Amaya, R., Paternina, C.: A multivariate analysis for the creation
of Port Community System approaches. Transp. Res. Procedia 30, 127–136 (2018). ISSN:
2352-1465. https://doi.org/10.1016/j.trpro.2018.09.015
Moros-Daza, A., Hoz, D.-D., Jaller-Martelo, M., Paternina-Arboleda, C.D.: Using advanced infor-
mation systems to improve freight efficiency: results from a pilot program in Colombia. In:
Paternina-Arboleda, C., Voß, S. (eds.) ICCL 2019. LNCS, vol. 11756, pp. 22–38. Springer,
Cham (2019). https://doi.org/10.1007/978-3-030-31140-7_2
Moros-Daza, A., Amaya-Mier, R., Paternina-Arboleda, C.: Port Community Systems: a structured
literature review. Transp. Res. Part A Policy Pract. 133, 27–46 (2020). ISSN: 0965-8564. https://
doi.org/10.1016/j.tra.2019.12.021
Möller, M., Vuik, C.: On the impact of quantum computing technology on future developments
in high-performance scientific computing. Ethics Inf. Technol. 19(4), 253–269 (2017). https://
doi.org/10.1007/s10676-017-9438-0
Nichols, B.: The Current State of DevSecOps Metrics. CMU (2021). https://insights.sei.cmu.edu/
blog/the-current-state-of-devsecops-metrics/
Palma-Blanco, A., González, E.R., Paternina-Arboleda, C.D.: A two-pheromone trail ant colony
system approach for the heterogeneous vehicle routing problem with time windows, multiple
products and product incompatibility. In: Paternina-Arboleda, C., Voß, S. (eds.) ICCL 2019.
LNCS, vol. 11756, pp. 248–264. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-
31140-7_16
Paternina-Arboleda, C.D., Das, T.K.: A multi-agent reinforcement learning approach to obtaining
dynamic control policies for stochastic lot scheduling problem. Simul. Model. Pract. Theory
13(5), 389–406 (2005)
Paternina-Arboleda, C.D., Das, T.K.: Intelligent dynamic control policies for serial production
lines. IIE Trans. 33(1), 65–77 (2001). https://doi.org/10.1023/A:1007641824604
Paternina-Arboleda, C.D., Montoya-Torres, J.R., Fabregas-Ariza, A.: Simulation-optimization
using a reinforcement learning approach. In: 2008 Winter Simulation Conference, Miami, FL,
USA, pp. 1376–1383 (2008). https://doi.org/10.1109/WSC2008.4736213
Paternina-Arboleda, C.D., Agudelo-Castañeda, D., Voß, S., Das, S.: Prediction of SO2 ports
emissions inventories. Sustainability (2023, Submitted)
Rakos, D.: Understanding GPU caches. [web log] (2021). https://www.rastergrid.com/blog/gpu-
tech/2021/01/understanding-gpu-caches/. Accessed 12 Apr 2023
Sarabia, Carolina, M., John, H., Rios-Griego, Carlos, D., Paternina-Arboleda: Simulation-based
decision support models for river cargo transportation. In: 2006 IEEE Systems and Information
Engineering Design Symposium, pp. 142–145 (2006)
Saxon, S., Stone, M.: Container shipping: the next 50 years. A McKinsey & Company report
(2017)
Saxena, A.: What are AI accelerators and how does it work? BISinfotech (2022). https://www.bis
infotech.com/what-are-ai-accelerators-and-how-does-it-work/. Accessed 14 Apr 2023
Shankland, S.: Quantum computers could crack today’s encrypted messages. That’s a prob-
lem. CNET (2021). https://www.cnet.com/tech/computing/quantum-computers-could-crack-
todays-encrypted-messages-thats-a-problem/. Accessed 17 Apr 2023
Song, D.: A literature review, container shipping supply chain: planning problems and research
opportunities. Logistics 5, 41 (2021). https://doi.org/10.3390/logistics5020041
84 C. Paternina-Arboleda et al.

Trucco, P., Petrenj, B.: Characterisation of resilience metrics in full-scale applications to interde-
pendent infrastructure systems. Reliab. Eng. Syst. Saf. 235, 109200 (2023). ISSN: 0951-8320.
https://doi.org/10.1016/j.ress.2023.109200
Tsvetkova, A., Gustafsson, M., Wikström, K.: Digitalizing maritime transport: digital innovation
as a catalyzer of sustainable transformation. In: Montero, J., Finger, M. (eds.) A Modern Guide
to the Digitalization of Infrastructure. Elgar Modern Guides, pp. 123–148. Edward Elgar (2021)
Velasquez-Bermudez, J.M., Abril, D., Paternina-Arboleda, C.D.: Optimizing port operations to
cope with shipping congestion in South American countries. OR/MS Today 49(2) (2022)
Velásquez-Bermúdez, J.M., Khakifirooz, M., Fathi, M. (eds.): Large Scale Optimization in Supply
Chains and Smart Manufacturing: Theory and Applications. Springer, Cham (2020). ISBN:
978-3-030-22790-6. https://doi.org/10.1007/978-3-030-22788-3
Wang, Z., Liu, X.: Cyber security of railway cyber-physical system (CPS) – a risk management
methodology. Commun. Transp. Res. 2, 100078 (2022). ISSN: 2772-4247. https://doi.org/10.
1016/j.commtr.2022.100078
Woschank, M., Rauch, E., Zsifkovits, H.: A review of further directions for artificial intelligence,
machine learning, and deep learning in smart logistics. Sustainability 12, 3760 (2020). https://
doi.org/10.3390/su12093760
Yang, Z., et al.: Indicator-based resilience assessment for critical infrastructures – a review. Saf.
Sci. 160, 106049 (2023). ISSN: 0925-7535. https://doi.org/10.1016/j.ssci.2022.106049
Zemp, B.: The Intersection Between AI and Blockchain Technology – Industries of Tomor-
row. Forbes (2023). https://www.forbes.com/sites/forbesbooksauthors/2023/02/28/the-inters
ection-between-ai-and-blockchain-technology--industries-of-tomorrow/?sh=1731ee7d4de7.
Accessed 14 Apr 2023
Zhang, W., Guhathakurta, S., Khalil, E.B.: The impact of private autonomous vehicles on vehicle
ownership and unoccupied VMT generation. Transp. Res. Part C Emerg. Technol. 90, 156–165
(2018)
 Customer’s Choice in the Context
of Cross-Border E-Commerce: An
Application of Structural Equation
Modelling

Yijia Liu1(B) and Xiaoning Shi2

1
Faculty of Mathematics, Informatics and Natural Sciences, University of Hamburg,
Hamburg, Germany
[email protected]
2
Institute of Transport Research, DLR (Deutsches Zentrum für Luft- und
Raumfahrt, German Aerospace Center), Berlin, Germany
[email protected]

Abstract. Emerging E-Commerce enables the purchasing and delivery

activities during the pandemic. Along with the eased situation of the
post-pandemic era, however, the E-Commerce business model is facing an
increasing number of competitive service providers and demanding cus-
tomers. One of the discriminating issues between service providers seems
to relate to the Customer Repurchase Intention (CRI). Especially in the
textile industry in the context of Cross-Border E-Commerce (CBEC),
CRI might be influential. In this context, we investigate the customer’s
choice to provide commercial insights to E-Commerce service providers.
To achieve this, we apply Structural Equation Modeling (SEM). More
specifically, the combination of ESEM (Exploratory Structural Equation
Modelling) and CFA (Confirmatory Factor Analysis) is used as an impor-
tant SEM approach to explore which factors most significantly impact
the CRI of textile products in the context of CBEC. The study is con-
ducted based on the 187 feedback of distributed questionnaires. In our
results, Satisfaction and MIS (Management Information System) Tech-
nical Factor have significant impacts on CRI.

Keywords: E-Commerce · Structural Equation Modeling (SEM) ·

Customer Repurchase Intention · Cross-Border E-Commerce

1 Introduction
In the post-pandemic era, all business models, i.e., traditional business models,
E-Commerce and Cross-Border E-Commerce (CBEC) have some gradual devel-
opments. With the popularity of smart phones and other devices, the devel-
opment of the Internet of Things (IoT) technologies, the establishment of 5G
Supported by the Project ‘FuturePorts’, DLR (Deutsches Zentrum für Luft- und Raum-
fahrt, German Aerospace Center).
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 85–102, 2023.
https://doi.org/10.1007/978-3-031-43612-3_5
86 Y. Liu and X. Shi

stations, logistics industry services and the improvement of information security

collectively provide the foundation for the prosperous development of CBEC.
According to [43], the E-Commerce revenue of the fashion industry in Ger-
many will rise to around 45.33 billion U.S. dollars in 2025. Some statistics in
[42] explicates the T-shirt and hoodie as the two most purchased product cate-
gories. Additionally, from the perspective of the supply side of logistics services,
e.g., rapid development of logistics network connectivity and Asia-Europe rail-
way infrastructures and facilities, CBEC between Europe and Asia is making
rapid progress. However, considering the demand side, a challenging economic
situation in the post-pandemic era occurs. With a three-year observation on the
freight rates of the containerized rail freight provided by Transfracht (TFG) [46],
compared to 2022 and 2021 the freight rates in 2023 increase by approximately
12.94% and 24.63%, respectively.
No matter of the traditional business process or the E-Commerce model,
merchants intend to increase the Return of Investment (ROI) through providing
high quality goods or services. Merchants study consumer’s behavior in order
to build a long-term relationship with customers [22]. They often carry out a
series of strategies to amplify a Customer Repurchase Intention (CRI) and to
attract potential customers through improvement of service quality according to
customer demand (Fig. 8). CRI is a customer decision behavior [16]; therefore,
it is meaningful to analyze which factors can affect, trigger and amplify the
repurchase intention. However, it is noted that there exists not only one factor
affecting the CRI. Instead, there may be a series of unique factors, such as
perceived benefits, perceived risk [38], and so on.
This paper aims to explore the causality among several factors and CRI of
textile products in the context of CBEC. Although T-shirt and hoodie are the
particularly most purchased products in the CBEC business [42], T-shirts seem
to be more widely and commonly used. Hence, during the process of conduct-
ing this research we focus on the customer’s choice of T-shirt products. At the
outset, SPSS is utilized to determine the factor number based on rotated analy-
sis. Following that, the ESEM+CFA1 framework is to be constructed on Mplus
for analyzing the significant impacts derived from different factors. The process
of redefining factors is conducted based on the results of the factor analysis
provided by Mplus. A questionnaire is designed and distributed digitally for the
purpose of data collection. We employ SmartPLS 4 to evaluate the measurement
model.2 In case a company or platform plans to proactively increase its revenue
within a limited budget, the outcome of our research would provide commercial
insights as well as indicate some prioritized strategies.
Nevertheless, some challenges exist in this research. Firstly, ESEM+CFA are
not yet widely applied in the CBEC field. The available references are there-
fore limited. Secondly, in empirical research, the sample size plays a crucial
role. However, it is difficult to reach a large sample size by collecting huge

1
ESEM+CFA is a combination of ESEM and CFA. It exhibits the similarity with
ESEM within CFA, which is introduced in Mplus User’s Guide [29].
2
SmartPLS 4: the current Version is SmartPLS 4.0.9.4, https://www.smartpls.com/.
 Customer’s Choice in Cross-Border E-Commerce 87

amounts of feedback within the limited time. Thirdly, instead of naming the
factors in ESEM in advance, significant factors are to be redefined by running
ESEM+CFA. Fourthly, in ESEM indicators have cross-loading, the assignment
of indicators may differ from our original design. In other words, it is challenging
for the process of factor redefinition. Moreover, while refining the model fit, the
modification indices (MI) are not unified. Therefore, two distinctive processes
are discussed and applied based on a better result. Lastly, some evaluation cri-
teria of PLS-SEM are applied for assessing the measurement models with some
concerns on their validity. The factor redefining process provided by this research
is going to overcome the above-mentioned challenges.

2 Literature Review
CBEC is a busi-
ness model based
on Internet and E-
Commerce websites
or platform, where
two parties involved
in the transaction
come from different
countries or terri-
tories [15]. Hu and
Luo [17] state that
CBEC is efficient
and low-cost, and it
is an application of Fig. 1. Process of CBEC
internet technology
on the field of for-
eign trade. Different roles in CBEC are characterized by different functions
and responsibilities (Fig. 1). For instance, customs is regulating and facilitating
international trade [6] and the customs also aims to maintain a certain degree
of fairness and transparency in global trade. It is vital for all parties involved
in CBEC to know customs declarations and regulations. The customs releases
declared items comprehensively, such as, the calculation of customs duties includ-
ing over-size or over-weight products’ duties and taxes. When the sellers launch
their products, the CBEC platform should automatically inform them about the
relevant regulations. Generally speaking, the owner of products and its represen-
tative are responsible to submit declarations [45]. For the purpose of reducing the
operational time, such a declaration should be submitted with considerations of
accuracy, authenticity and completeness. It is noted for cross-border trade that
the payment of customs duties complies with Incoterms3 such as Delivered Duty
Paid (DDP), Delivered Duty Unpaid (DDU), etc., where DDP is relative widely
used due to efficiency and safety concerns in E-Commerce [9].
3
Incoterm: International Commercial Term.
88 Y. Liu and X. Shi

Our research is empirical research based on quantitative methods such as sur-

veys [7], though this type of approach might be criticized; see, e.g., [50] indicating
the possible use of mystery shopping instead. It often lacks a theoretical foun-
dation [11]. Kawa and Zdrenka [19] point out that the biggest barrier in CBEC
is delivery-service-related cost with required time windows, which becomes an
important factor affecting customer satisfaction. In [36] customer satisfaction,
enjoyment and trust are found to exert positive impacts on repurchase inten-
tions. Obviously, service and system quality have positive impacts on repurchase
intentions [23]; see also [51]. Consumer’s choice behavior is actually influenced
by various values including functional, social, emotional, and other values [39].
Customer’s repurchase intention is actually a judgement and decision behavior
[16] affected by several factors. Repurchase behavior and repurchase intention
can be regarded as consumer’s choice influenced by various factors including
satisfaction, trust, perceived value, etc.
In [3], the authors consider the perspectives of service quality and commercial
and technical safety, and depict some keywords, namely, “privacy”, “online pay-
ment security”, “website design”, “website speed”, etc. Valdez-Juárez et al. [49]
also consider “website security” in their study. The user interface (UI) including
its usability and attractiveness affects the repurchase intention [26]. According
to [31], customer’s satisfaction is categorized as a level of fulfillment associated
with a consumption perspective. However, perceived value is different from sat-
isfaction [44]. Perceived value [24], online service quality, and logistics service
[32] have been investigated in E-Commerce research in the past. Product variety
has been regarded as an important factor for customer satisfaction [18]. Trust
can be deemed as customer’s confidence on the involved business partner [5,30].
Service is usually intangible, heterogeneous, inseparable and perishable [20]. In
E-Commerce, service is often divided into online and offline categories. Online
service usually refers to handling customer complaints through the platform,
while offline service usually refers to the logistics service, product replacement
and re-delivery [51]. Thus, these two categories of service are considered in our
research, too. In addition, Mou et al. [28] explore the perceived risk of “customs
clearance” in CBEC.
In our research, such as customs duties declaration, chat-bot service, etc. are
also considered. Essentially, IT and Management Information Systems (MIS) are
preconditions for business intelligence, digitization and digital transformation. A
qualified MIS and infrastructure can facilitate the performance of E-Commerce.
For instance, a well-designed E-Commerce website interface gains customer’s
curiosity. In fact, a tracking Application Programming Interface (API) is a very
practical tool acknowledged by many. In a policy of Amazon, the Valid Tracking
Rate (VTR) is analyzed as one key indicator to evaluate the performance of a
seller [1]. Hence, the implementation of an API makes sense, and we deploy an
API-based indicator. Besides, some technologies can be utilized for the identity
authentication [33], which provides us insights to a technological perspective.
CRI is deemed as a willingness leading a customer repeatedly towards purchasing
goods or services from the same seller [51]. In CBEC, it also stands for repeated
 Customer’s Choice in Cross-Border E-Commerce 89

purchases on the same platform. A series of studies have explored the repurchase
intention or behaviour with SEM.
Methodologies, principles, concepts, the criteria of metrics, and theories of
SEM are listed in this section as prerequisite for the next sections. RMSEA
(Root Mean Square Error of Approximation), CFI (Comparative Fit Index), TLI
(Tucker Lewis Index), df (Degree of Freedom), and the probability of RMSEA ≤
0.05 as criteria are comprehensively introduced in the references [21,37,41]. For
evaluation of the measurement model in SEM, an Average Variance Extracted
(AVE) ≥ 0.5 is acceptable [13]. However, Shrestha [41] states if composite reality
> 0.6 but AVE < 0.5, the convergent validity remains acceptable. The theory
of ESEM, guideline of Mplus, and examples of Mplus support our research [2,
27,29]. Modification benefits a better model fit. With regard to investigating
the exploration of the ESEM framework, at the outset, a pair with highest
modification index is modified [48]. Even though many articles have explored
the CRI prior to the pandemic, it remains valuable to investigate it in the post
pandemic era.

3 Methodology

3.1 Analysis Process

Model specification is conducted as the first

step. Since EFA+CFA has both exploratory
and confirmatory characteristics, ESEM+CFA
is applied (EFA: Exploratory Factor Analy-
sis). Following that, we design a question-
naire with 7 Likert scale for data collec-
tion. Our questionnaire covers 16 questions,
and it is distributed through different soft-
ware and platforms (Table 2). Next, data is
implemented on SPSS for EFA and Mplus for
ESEM+CFA, respectively. The factor num-
bers are to be determined based on the result
of EFA in advance, so that we can use it in
ESEM+CFA. After that, before the model is
going to be evaluated according to the criteria.
Modification based on different assumptions
for refining model fit is conducted. Next, the Fig. 2. Analysis Process
model is extracted for analysis forward. We
implement it on SmartPLS for assessing the
measurement model. However, some criteria are derived from PLS-SEM. Hence,
the p-value remains a crucial metric. After that, we attempt to discuss the exper-
imental results based on different sample size, refined measurement model, and
a different structural model. Besides, a comparison with related literature is also
to be discussed in the final stage. Figure 2 displays the our analysis process.
90 Y. Liu and X. Shi

3.2 ESEM+CFA

In the early stage, EFA is applied to observe the extracted factor numbers, which
will be utilized in ESEM+CFA. ESEM within CFA is introduced in Mplus User’s
Guide [29] and Mai and Wen’s study [25]. Instead of ESEM within CFA, we
prefer to call it ESEM+CFA, as we think that this model is constructed by
two independent models. Hence, this model is an integration, not an inclusion.
Accordingly, we suppose two different modification processes in Sect. 3.3.
Beyond their original struc-
tural models they also share the Table 1. Criteria for Model Evaluation
same structural model through Metrics Cut-off value Sources
RMSEA ≤0.08 [8, 12, 13, 21, 29, 35,
connection with each other. Being 37, 41, 48]
SRMR ≤0.08
different from typical SEM, the CFI ≥0.8
hypothesis is not to be established TLI ≥0.8
Probability of ≥ 0.05
in advance. After estimation, the RMSEA≤0.05
Chi Square –
factors are named based on their Chi Square ≤3
df
indicators. In our view, the CFA Cronbach’s alpha ≥0.6
AVE ≥0.5
or EFA is an analysis method for R Square ≥0.5
≥0.9
factor analysis. While factor anal- NFI (Normed Fit
Index)
ysis do not place emphasize on Indicators loading ≥0.6
Composite ≥0.6
the analysis of factor relationship, reliability
HTMT ≤0.9
SEM concentrates on it. ESEM is (Heterotrait-
Monotrait Ratio)
a SEM based on EFA, highlight-
ing the analysis of significant relationships among factors. Such an exploratory
process is not only for the relationship among factors, but also for indicators.
The definition of ESEM also provides us a new understanding of SEM. For
instance, how to understand the Confirmatory Structural Equation Modelling
(CSEM). In other words, if researchers plan to verify an old model, whether
this process can be called CSEM. Furthermore, there are two kinds of SEM,
i.e., PLS-SEM (Partial Least Squares SEM) and CB-SEM (Covariance Based
SEM). They can be distinguished by some differences including distinct criteria
for model assessment. PLS-SEM is for the extension of a verified model, theory
testing and comparison is a feature of CB-SEM [13,14]. To conduct an analysis
of ESEM+CFA based on CB-SEM, Mplus is to be applied. ESEM, Bi-factor,
ESEM+CFA are highlights of Mplus [29]. At present Mplus 8.9 is available [29],
and in our study Mplus 8.3 is deployed for data implementation. To observe
the measurement model, SmartPLS 4 is also applied. The applied criteria are
succinctly depicted in Table 1, providing a overview.

3.3 Modification

As described, two different modification processes are applied according to two

assumptions.

Assumption 1. Considering the restriction of model, modification is based on

a priority principle. Firstly, crossing structural model is not to be allowed. The
indicator are initially modified through deletion, as they are pointed to by some
 Customer’s Choice in Cross-Border E-Commerce 91

factors residing in different structure model. Secondly, the indicators are to be

modified by correlation within a same structural model. Next, the indicators
residing in different measurement model are going to be correlated in last steps.
Assumption 2. Despite structural limitations, the modification is started from
the highest MI. The principle of “each step, one pair” are adhered to by both
two assumptions. Additionally, uniform utilization of modification is guided by
a better result of model fit.

3.4 Questionnaire

This research applies a

questionnaire with a 1–7
Likert scale. Because of the
exploratory nature of our
research, we apply a ques-
tionnaire by self-defining
for collecting feedback. A
sample of 187 responses
is collected. Respondents
come from China (circa
64%), Germany (circa 25%),
the USA (circa 4%), Canada
(circa 2%) and other coun- Fig. 3. The most visited CBEC website
tries4 (circa 5%). Addition-
ally, around 103 respondents have purchased in Amazon before (Figure 3). 75.4%
of the respondents are between 21 to 30 years old, while 43.9% are students. To
filter the valid response, respondents are asked by a question, namely “Have you
ever purchased in one of CBEC before?”. Then 162 valid data are implemented.
28 indicators are extracted and summarized from different perspectives, namely

Table 2. Questionnaire content

Questionnaires Content
Collection Period Feb 20. 2023 to March 17. 2023
Numbers of question 16
Total participants 187
Response Rate around 36%
Languages English, Chinese, German
Female, Male 46%, 54%
Spread medium Linkedin, WeChat, Email and
What’sApp, Survey Circle, SMS

4
UK, Australia, Japan, Panama, Norway, Ireland, and Switzerland.
92 Y. Liu and X. Shi

satisfaction, online service quality, offline service quality, trust, and technical
factors. 28 indicators are analyzed using SPSS to study EFA. and six indica-
tors from CRI are employed for CFA. Their indicators are defined from different
literature. The keywords of indicators are succinctly described in Table 3.
Table 3. Keywords of indicators

Perspectives Description of Items (Keywords)

Satisfaction (y1–y4) 1: service quality, 2: logistics service,
3: product quality, 4: product variety
Trust (y5–y10) 5: website privacy, 9: product description,
6 and 10: authenticity of evaluation,
7: website security, 8: payment method,
Offline service 11: pick-up and delivery, 12: fee rate,
quality (y11–y16) 13: responsibility of logistics,
14: customs declaration and clearance,
15: customs duty, 16: go through Customs
Online service 17: urgent service, 18: 24-hours service,
quality (y17–y22) 19: chat bot service, 20: return service,
21: problem resolving,
22: human service accessibility
Technical factor 23: attractive UI, 24: shipping API,
(y23–y28) 25: multilingual UI, 26: promotion newsletter,
27: log-in authentication,
28: payment authentication
CRI (cri1–cri6) 1: repurchase other T-shirts in the same vendor
2: repurchase same products in the same vendor
3: repurchase other T-shirts in the same web
4: repurchase other products in the same web
5: preference for future shopping
6: recommend to friends and family

4 Results and Discussions

4.1 Results

(1) 28 indicators are estimated on SPSS with rotated Varimax under Maxi-
mum Likelihood. According to the estimation performed by SPSS, the KMO
(Kaiser-Meyer-Olkin) is 0.878, and the p-value is approximate to 0. Hence,
in accordance with [41] the result is acceptable. Therefore, 28 indicators are
extracted into six factors to be applied for the subsequent steps.
(2) The integration of ESEM+CFA is implemented on the Mplus. Based on
the first-round estimation, RMSEA is 0.076, CFI is 0.878, the probability of
RMSEA is 0.000, and SRMR is 0.050. So the model does not fit. Afterwards,
it is to be modified using the MI until no pair exceeds 10. According to the
first modification method, 12 steps are carried out. After being modified
by the second modification process, its model fit is also acceptable. Despite
this, it does not outperform the first one (Table 4b).
(3) After being modified, 3 factors have significant impacts on the CRI. Accord-
ing to the result and their indicators, they are redefined as Satisfaction (SA),
MIS Technical factor (MIS), and Facilitation (FA), respectively (Fig. 5a).
 Customer’s Choice in Cross-Border E-Commerce 93

Table 4. Result of ESEM+CFA

Criteria Result Criteria Result

RMSEA 0.051 Acceptable RMSEA 0.056 Acceptable
CFI 0.950 Acceptable CFI 0.938 Acceptable
TLI 0.925 Acceptable TLI 0.907 Acceptable
SRMR 0.041 Acceptable SRMR 0.040 Acceptable
Probability of Probability of
0.423 Acceptable 0.144 Acceptable
RMSEA ≤ 0.05 RMSEA ≤ 0.05

(a) With the first modification process (b) With the second modification process

Fig. 4. Diagram on SmartPLS

Other 3 factors are not going to be defined any more. However, the coefficient
of Facilitation is negative, i.e. −0.216, it is inconsistent with common sense.
Following that, in order to proceed with further exploration these 3 factors
with CRI are extracted for establishing a new model further (Fig. 5a).
(4) As estimated, Facilitation has no significant impact on CRI anymore. This
provides a new vision which is different from current policies and com-
mon sense which is previously accepted [10,40,47]. In contrast to that,
the impacts of MIS Technical factor and Satisfaction persist significantly.
Figure 5a → Fig. 5b demonstrates the transformation of model. In the
extracted model, correlations that cross structural model have existed, there-
fore, model-crossing is allowed. This step is implemented on Mplus. Follow-
ing that, the further evaluation of measurement model and structural model
are required.

Since the current SmartPLS 4 supports CB-SEM, we implement it on the

SmartPLS 4 for measurement model analysis (Fig. 4). Table 6 displays the esti-
94 Y. Liu and X. Shi

Table 5. Result on SmartPLS 4 and Mplus 8

Criteria Result on SmartPLS 4 Result on Mplus 8

p-value 0.001 0.000 Acceptable
RMSEA 0.051 0.050 Acceptable
CFI 0.953 0.955 Acceptable
TLI 0.942 0.943 Acceptable
SRMR 0.058 0.055 Acceptable
NFI 0.859 – Inadequate
Probability of – 0.423 Acceptable
RMSEA ≤ 0.05
Chi Square
df
1.414 1.405 Acceptable
R2 0.645 0.641 Acceptable

mation of reliable indicator loading based on SmartPLS. Firstly, there are 8

indicators’ loading higher than 0.708. According to [8] and [12], in CB-SEM an
AVE needs to be higher than 0.5, and an acceptable loading need to exceed
0.6. Thereof, 13 indicators are noted. Hence, although the model fits good, but
convergent validity is insufficient. Cheung et al. [4] support this points too. How-
ever, it is uncertain to apply them for testing the validity in our ESEM+CFA
model. Hence, we are going to use it to evaluate measurement model. In Table 6,
the results of Cronbach’s alpha are acceptable, while all composite reliability are
greater than 0.6. Although results in AVE are less than 0.5, but the convergent
validity remains adequate [41]. Discriminant validity is evaluated through HTMT
(Heterotrait-Monotrait Ratio). If HTMT < 0.85, it indicates conceptually differ-
ent factors (Table 7). Considering the result of measurement model is not very
sufficient; therefore, the model is optimized for further refinement. According to
[13], researchers can to enhance measurement model through indicator deletion.
Furthermore, we still concentrate on p-value.

Table 6. Result of measurement model based on SmartPLS 4

Factors Indicators Coefficient Cronbach’s Composite Convergent

alpha reliability (rho c) validity (AVE)
(standardized)
CRI cri1 0.754 0.834 0.796 0.487
cri3 0.731
cri4 0.693
cri5 0.742
Facilitation y6 0.853 0.745 0.632 0.314
y10 0.845
MIS TechnicalFactor y25 0.625 0.788 0.722 0.310
y27 0.818
y28 0.766
Satisfaction y1 0.919 0.829 0.826 0.412
y2 0.691
y3 0.794
y4 0.606
y24 0.615
 Customer’s Choice in Cross-Border E-Commerce 95

The minimum path loading ωmin is inversely derived in accordance with our
sample size and applied significance level 5% [13]. Based on observation, the path
loading of Satisfaction and MIS Technical Factor exceed 0.167, but Facilitation
not. Apart from that, it is discussed whether the factor loading needs to be
higher than 0.4, 0.5 or 0.7 [4]. CRI is an endogenous variable in the model,
its R Square on SmartPLS is 0.645, while on Mplus is 0.641. Therefore, the
power of prediction of CRI is moderate [13]. Table 5 explicates the result of
the extracted model estimated on Mplus and SmartPLS, but Mplus can output
the probability of RMSEA ≤ 0.05 individually. Regardless of the differences of
outputs, the results are almost the same. Hence, the structural model of the
extracted model fits well.
With the result of the extracted model, a proactive business strategy jour-
ney is designed (Fig. 8). In terms of that, some items are hoped to applied for
improvement. Even though not all of the indicators loading are acceptable. For
the purpose of further observation and verification on this extracted model and
those of criteria forward this diagram is derived from the valid p-value.

Fig. 5. Model Extraction

4.2 Discussions
(1) The measurement models are going to be optimized by removing indica-
tors, namely the indicator with lowest loading is deleted initially. In line
with modification only one pair is deleted in each step. In Fig. 6, Facilita-
tion remains insignificantly. There is no cross-loading among different mea-
surement model. Additionally, all of 13 indicators are reliable theoretically.
Table 8a and 8b depict the results of the variation in the measurement model
and the model fit. All of the metrics are definitely improved, especially AVE.
The results of convergent reliability of CRI, Facilitation and MIS Technical
Factor meet the required criteria and are deemed as acceptable, and AVE
of Satisfaction is approximate to 0.5.
96 Y. Liu and X. Shi

Through this operation some metrics of Table 7. Result on SmartPLS 4

model fit are enhanced. For instances, CRI FA MIS
it increases CFI, TLI, RMSEA, SRMR,
Chi Square FA 0.734 – –
df , NFI, respectively, as well as
MIS 0.807 0.848 –
HTMT. Moreover, the p-value of model
is also increased to 0.005. In the begin- SA 0.776 0.841 0.761
ning, for a better model fit performed
on Mplus, we modify our model. Per-
haps due to that, the measurement model in our ESEM+CFA is not so
good. However, it lacks of rich evidence to be verified so far.
Table 8. Result of refined model on SmartPLS 4

Criteria Result
Composite
Cronbach’s RMSEA 0.057 Acceptable
reliability AVE
alpha
(rho c) CFI 0.967 Acceptable
CRI 0.842 0.792 0.539 TLI 0.955 Acceptable
FA 0.845 0.856 0.744 SRMR 0.050 Acceptable
MIS 0.771 0.769 0.632 NFI 0.912 Acceptable
SA 0.826 0.829 0.499 Chi Square
1.535 Acceptable
df
(a) Result of measurement model
(b) Result of model fit

Fig. 6. Refined extracted model on SmartPLS 4

 Customer’s Choice in Cross-Border E-Commerce 97

(2) Variation of the sample size within ESEM+CFA. To investigate the effect of
the sample size, the different sample sizes are applied for additional compar-
ison on Mplus; therefore, experiments with three different sample sizes are
conducted. They are distinguished by the response time, namely the first
week (75 available responses), the second week (124 available responses)
and overall, respectively. The result of 75 sample sizes shows the model fit is
not acceptable. After modification, almost all metrics dissatisfy with cut-off
value. With the increased sample sizes the model fit is definitive improved.
The Fig. 7 displays that the probability of RMSEA ≥ 0.05 is influenced by
sample size most significantly. Beyond that, all of metrics are refined with
the increased sample sizes. Accordingly, it is valuable to study on it with a
larger sample size in the future.
(3) Comparison with previous research. On one hand, our experiment explicates
the significant recursive relationship of “Satisfaction → CRI” and “MIS
Technical → CRI”. As described in literature, Satisfaction has significant
impacts on CRI. Perhaps the definition of factors and the design of indicators
jointly affect the result. On the other hand, the loading reliability of “as
a preference for future shopping” in our case is regarded as acceptable.
However, in [28], “as a preference for future shopping” and “recommend
it to others” are both acceptable. Maybe the different research objectives,
sample sizes, different backgrounds, different models or respondents (target
groups) lead to the above-mentioned different results.

Fig. 7. Comparison of different sample sizes within ESEM+CFA

98 Y. Liu and X. Shi

Fig. 8. Strategy journey with output of extracted model

5 Conclusions and Future Research

With an application of ESEM+CFA the CRI in the context of the CBEC is
investigated and explored. In addition, 28 indicators are simulated within ESEM
on Mplus, while six indicators are used for CRI to study CFA. Satisfaction,
Facilitation, and MIS Technical Factor have significant impacts on CRI. After
that, the model is extracted into a new independent cluster model to explore
further. Facilitation has no significant impact on CRI any more. In fact, our
result demonstrates a process from the model specification, generation, extrac-
tion, and improvement. In addition, we also analyze the measurement model
and structural model with the new version of SmartPLS. We design a strategy
journey based on the significant indicators and factors. In terms of that, with
the increased sample size the model fit of ESEM+CFA is improved. But it can-
not be concluded that, as long as the sample size is increased, the result can be
improved constantly, because questionnaire’s design plays a role in this regard.
Besides, the process of refining model is performed by applying SmartPLS. The
process ends when all metrics are beyond threshold.
In future research, some experiments focusing on the increased sample sizes
will be considered initially, because 187 respondents is indeed insufficient for a
conclusive result. In addition, some investigations on the changes and upgrades
of current insignificant factors would be of value in the new norm of global trade.
It is also of interest to conduct a Confirmatory Tetrad Analysis (CTA) and com-
pare it with the formative measurement model [34]. The above-mentioned future
directions would contribute and push forward modification processes. Some fac-
tors are going to be selected for observing moderation and mediation [13], such
as perceived and individual factors in the questionnaire. In summary, there is
 Customer’s Choice in Cross-Border E-Commerce 99

still space to refine and leverage our understanding on the CBEC with focus on
customer intention, choice and derived cargo volumes.

Acknowledgements. Special thanks go to the 187 respondents of the questionnaire,

and the support of FuturePorts Project, DLR (Deutsches Zentrum für Luft- und Raum-
fahrt, German Aerospace Center). In addition, the constructive comments provided by
the anonymous referees are greatly appreciated.

References
1. Amazon: Valid Tracking Rate (VTR) (2023). https://sellercentral.amazon.co.uk/
help/hub/reference/external/G201817070?locale=en-GB. Accessed 11 Apr 2023
2. Asparouhov, T., Muthén, B.: Exploratory structural equation modeling. Struct.
Equ. Model. 16(3), 397–438 (2009). https://doi.org/10.1080/10705510903008204
3. Cao, X., Shen, B., Liu, Y., Wang, M.: Research on evaluation of B to C e-commerce
website based on AHP and grey evaluation. In: Second International Symposium
on Electronic Commerce and Security, vol. 2, pp. 405–408 (2009). https://doi.org/
10.1109/ISECS.2009.242
4. Cheung, G., Cooper-Thomas, H., Lau, R., Wang, L.: Reporting reliability, con-
vergent and discriminant validity with structural equation modeling: a review and
best-practice recommendations. Asia Pac. J. Manage. (2023). https://doi.org/10.
1007/s10490-023-09871-y
5. Corbitt, B.J., Thanasankit, T., Yi, H.: Trust and e-commerce: a study of consumer
perceptions. Electron. Commer. Res. Appl. 2(3), 203–215 (2003). https://doi.org/
10.1016/S1567-4223(03)00024-3
6. Customs & Excise Division Dominica: Role of Customs (2022). http://www.
customs.gov.dm/index.php/about-customs/role-of-customs. Accessed 20 Mar
2023
7. Dan, V.: Empirical and Nonempirical Methods, pp. 1–3. Wiley (2017). https://
doi.org/10.1002/9781118901731.iecrm0083
8. Dash, G., Paul, J.: CB-SEM vs PLS-SEM methods for research in social sciences
and technology forecasting. Technol. Forecast. Soc. Chang. 173, 121092 (2021).
https://doi.org/10.1016/j.techfore.2021.121092
9. Deutsche Post Group DHL: Customs, duties & taxes (2023). https://www.
dhl.com/discover/en-my/e-commerce/e-commerce-advice/ddu-vs-ddp-for-
ecommerce-business. Accessed 17 Mar 2023
10. Fang, L., Shi, X., Hu, H.: Analysis of influencing factors and location prediction
of dry port based on logit model. J. Chongqing Jiaotong Univ. 5, 31 (2012)
11. Gaskell, T.: The process of empirical research: a learning experience?
Res. Post-Compulsory Educ. 5(3), 349–360 (2000). https://doi.org/10.1080/
13596740000200084
12. Hair, J., Black, W., Babin, B.: Multivariate Data Analysis: A Global Perspective.
Global Edition, Pearson Education (2010)
13. Hair, J., Hult, G.T.M., Ringle, C., Sarstedt, M., Danks, N., Ray, S.: Partial Least
Squares Structural Equation Modeling (PLS-SEM) Using R: AWorkbook. Springer,
Heidelberg (2021). https://doi.org/10.1007/978-3-030-80519-7
14. Hair Jr., J.F., Matthews, L.M., Matthews, R.L., Sarstedt, M.: PLS-SEM or CB-
SEM: updated guidelines on which method to use. Int. J. Multivariate Data Anal.
1, 107–123 (2017). https://doi.org/10.1504/IJMDA.2017.087624
100 Y. Liu and X. Shi

15. Hazarika, B.B., Mousavi, R.: Review of cross-border e-commerce and directions for
future research. J. Glob. Inf. Manag. 30(2), 1–18 (2021). https://doi.org/10.4018/
JGIM.20220301.oa1
16. Hellier, P., Geursen, G., Carr, R., Rickard, J.: Customer repurchase intention: a
general structural equation model. Eur. J. Mark. 37, 1762–1800 (2003). https://
doi.org/10.1108/03090560310495456
17. Hu, B., Luo, Q.: Cross-border e-commerce mode based on Internet +. In: IOP
Conference Series: Materials Science and Engineering, vol. 394, no. 5, p. 052014
(2018). https://doi.org/10.1088/1757-899X/394/5/052014
18. Kaiser, U., Kesler, R., Reisinger, M.: The interplay between product variety and
customer satisfaction: theory and evidence (2023). https://doi.org/10.2139/ssrn.
4022124. Accessed 09 Apr 2023
19. Kawa, A., Zdrenka, W.: Conception of integrator in cross-border E-commerce.
LogForum 12, 63–73 (2016). https://doi.org/10.17270/J.LOG.2016.1.6
20. Khan, M.K., Nawaz, M., Ishaq, M.I., Tariq, M.I.: Product versus service: old myths
versus new realities. J. Basic Appl. Sci. Res. 4(1), 15–20 (2014)
21. Kline, R.: Principles and Practice of Structural Equation Modeling. Methodology
in the Social Sciences, 4th edn. Guilford, London (2015)
22. Lei, F., Lei, S., Liu, J., Wu, L.: Studying customer behavior of shopping on e-
commerce platforms from the perspective of college students. In: 2020 International
Conference on E-Commerce and Internet Technology (ECIT), pp. 41–44 (2020).
https://doi.org/10.1109/ECIT50008.2020.00017
23. Luo, H., Liang, Y., Wu, Z., Liu, Z., Shu, B., Luo, Y.: Research on the impact
of cross-border e-commerce platform quality on customer cross-buying intention.
In: 16th International Conference on Service Systems and Service Management
(ICSSSM), pp. 1–6 (2019). https://doi.org/10.1109/ICSSSM.2019.8887725
24. Luo, H., Zhang, X., Huang, J., Su, F., Huang, W., Luo, Y.: An empirical study on
customer repurchase willingness of fresh e-commerce products based on customer
perceived value. In: 16th International Conference on Service Systems and Service
Management (ICSSSM), pp. 1–6 (2019). https://doi.org/10.1109/ICSSSM.2019.
8887610
25. Mai, Y., Wen, Z.: Exploratory structural equation modeling (ESEM): an integra-
tion of EFA and CFA. Adv. Psychol. Sci. 21(5), 934 (2013). https://doi.org/10.
3724/SP.J.1042.2013.00934
26. Matthew, D., Hellianto, G.R., Putra, N.S., Sundjaja, A.M.: The effect of monthly
promotion, gamification, user interface usability & attractiveness on the market-
place repurchase intention. In: International Conference on Informatics, Multime-
dia, Cyber and Information System, pp. 193–199 (2021). https://doi.org/10.1109/
ICIMCIS53775.2021.9699114
27. Morin, A.J.S., Marsh, H.W., Nagengas, B.: Supplementary materials for: Chapter
10. Exploratory structural equation modeling (2013). https://www.statmodel.
com/download/Chapter%2010%20-%20ESEM-%20supplements%2020131.pdf.
Accessed 11 Feb 2023
28. Mou, J., Cui, Y., Kurcz, K.: Trust, risk and alternative website quality in B-Buyer
acceptance of cross-border e-commerce. J. Glob. Inf. Manag. 28, 167–188 (2020).
https://doi.org/10.4018/JGIM.2020010109
29. Muthén, L.K., Muthén, B.O.: MPlus version 8 user’s guide. Muthén & Muthén
(2017)
30. Nurdani, Y., Sandhyaduhita, P.I.: Impact of express delivery service quality
towards repurchase intention by B2C and C2C: a case of Indonesia. In: Inter-
 Customer’s Choice in Cross-Border E-Commerce 101

national Conference on Advanced Computer Science and Information Systems

(ICACSIS), pp. 221–227 (2016). https://doi.org/10.1109/ICACSIS.2016.7872779
31. Oliver, R.: Satisfaction: A Behavioral Perspective on the Consumer. McGraw Hill
(1997). https://doi.org/10.4324/9781315700892
32. Rachmawati, D.Z., Agus, A.A.: E-service and logistics service quality in e-
commerce, study case: shopee Indonesia. In: 2020 3rd International Conference
on Computer and Informatics Engineering (IC2IE), pp. 218–223 (2020). https://
doi.org/10.1109/IC2IE50715.2020.9274597
33. Rajesh, H.M., Song, I.: E-commerce websites/blogs - hassles and predicaments. In:
Proceedings of the 6th International Conference on Communication and Network
Security, ICCNS 2016, pp. 56–61. Association for Computing Machinery, New York
(2016). https://doi.org/10.1145/3017971.3017990
34. Ringle, C.M., Wende, S., Becker, J.M.: Confirmatory tetrad analysis in PLS
(CTA-PLS) (2022). https://www.smartpls.com/documentation/algorithms-and-
techniques/cta-pls/. Accessed 20 Mar 2023
35. Ringle, C.M., Wende, S., Becker, J.M.: SmartPLS 4 (2022). https://www.smartpls.
com/documentation/algorithms-and-techniques/model-fit/. Accessed 20 Mar 2023
36. Rini, D.P., Josephat, E., Ririmasse, Y.R., Sundjaja, A.M.: The effect of trust,
satisfaction and enjoyment toward repurchase intention in E-commerce. In: 2021
3rd International Conference on Cybernetics and Intelligent System (ICORIS), pp.
1–6 (2021). https://doi.org/10.1109/ICORIS52787.2021.9649504
37. Schumacker, R., Lomax, R.: A Beginner’s Guide to Structural Equation Modeling,
4th edn. Taylor & Francis (2015). https://doi.org/10.4324/9780203851319
38. Shan, S., Hua, F., Zeng, Q.: B2C e-commerce consumer decision-making model
based on perceived benefit and perceived risk. In: 2010 International Conference
on E-Business and E-Government, pp. 2222–2225 (2010). https://doi.org/10.1109/
ICEE.2010.561
39. Sheth, J.N., Newman, B.I., Gross, B.L.: Why we buy what we buy: a theory of
consumption values. J. Bus. Res. 22(2), 159–170 (1991). https://doi.org/10.1016/
0148-2963(91)90050-8
40. Shi, X., Ma, L., Voß, S.: On designing a slot sharing E-platform for liner shipping
services. In: Stephanidis, C., Marcus, A., Rosenzweig, E., Rau, P.-L.P., Moallem,
A., Rauterberg, M. (eds.) HCII 2020. LNCS, vol. 12423, pp. 499–513. Springer,
Cham (2020). https://doi.org/10.1007/978-3-030-60114-0 34
41. Shrestha, N.: Factor analysis as a tool for survey analysis. Am. J. Appl. Math.
Stat. 9, 4–11 (2021). https://doi.org/10.12691/ajams-9-1-2
42. Statista: What is your best selling streetwear product? (2019). https://www.
statista.com/statistics/1092487/best-selling-products-among-streetwear-retailers-
worldwide/?locale=en. Accessed 10 Jan 2023
43. Statista: E-commerce revenue forecast in Germany from 2017 to 2025,
by segment. https://www.statista.com/forecasts/480260/e-commerce-revenue-in-
germany-fashion-by-segments. Accessed 10 Apr 2023
44. Sweeney, J.C., Soutar, G.N.: Consumer perceived value: the development of a mul-
tiple item scale. J. Retail. 77(2), 203–220 (2001). https://doi.org/10.1016/S0022-
4359(01)00041-0
45. Taxation and Customs Union: Customs declaration (2021). https://taxation-
customs.ec.europa.eu/customs-4/customs-procedures-import-and-export-0/
customs-procedures/customs-declaration en. Accessed 10 Apr 2023
46. TFG Transfracht GmbH: Transfracht (2023). https://www.transfracht.com/
transfracht. Accessed 20 Mar 2023
102 Y. Liu and X. Shi

47. Tousypanah, H., Shi, X., Viohl, B., Hoffmann, J.: Analysis of influencing factors
and location prediction of dry port based on logit model. UNCTAD Transp. Newsl.
43, 6–10 (2009)
48. Tóth-Király, I., Bõthe, B., Rigó, A., Orosz, G.: An illustration of the exploratory
structural equation modeling (ESEM) framework on the passion scale. Front. Psy-
chol. 8 (2017). https://doi.org/10.3389/fpsyg.2017.01968
49. Valdez-Juárez, L.E., Gallardo-Vázquez, D., Ramos-Escobar, E.A.: Online buyers
and open innovation: security, experience, and satisfaction. J. Open Innov.: Tech-
nol. Mark. Complexity 7(1), 37 (2021). https://doi.org/10.3390/joitmc7010037
50. Voß, S., Mejia, G., Voß, A.: Mystery shopping in public transport: the case of
bus station design. In: Stephanidis, C., Marcus, A., Rosenzweig, E., Rau, P.-L.P.,
Moallem, A., Rauterberg, M. (eds.) HCII 2020. LNCS, vol. 12423, pp. 527–542.
Springer, Cham (2020). https://doi.org/10.1007/978-3-030-60114-0 36
51. Yu, Y.W., Chiu, S.C., Chumpoothepa, S., Chen, J.K.: Factors influencing online
repurchase intention: the case of clothing customers in Thailand. In: Proceedings
of PICMET 2014 Conference: Portland International Center for Management of
Engineering and Technology; Infrastructure and Service Integration, pp. 658–669
(2014)
Maritime Shipping
 Towards a Deep Reinforcement Learning
Model of Master Bay Stowage Planning

Jaike van Twiller(B) , Djordje Grbic, and Rune Møller Jensen

IT University of Copenhagen, Rued Langgaards Vej 7, 2300 Copenhagen, Denmark

{jaiv,djgr,rmj}@itu.dk

Abstract. Major liner shipping companies aim to solve the stowage

planning problem by optimally allocating containers to vessel locations
during a multi-port voyage. Due to a large variety of combinatorial
aspects, a scalable algorithm to solve a representative problem is yet
to be found. This paper will show that deep reinforcement learning can
optimize a non-trivial master bay planning problem. Our experiments
show that proximal policy optimization efficiently finds reasonable solu-
tions, serving as preliminary evidence of the potential value of deep rein-
forcement learning in stowage planning. In future work, we will extend
our architecture to address a full-featured master bay planning problem.

Keywords: Maritime logistics · Liner shipping · Stowage planning ·

Deep reinforcement learning · Markov decision processes

1 Introduction
In the past century, maritime transport has become the backbone of global trade
and modern consumerism. Many transported goods are shipped by container ves-
sels of liner shipping companies. To ensure timely arrivals and resource-efficient
operations, these liner companies use stowage planning to allocate containers to
vessel slots at each port of the voyage. The goal is to maximize vessel utilization
and minimize operational costs by creating robust stowage plans. This is an NP-
hard task due to (hatch-)overstowage [18,35], which is further complicated by
the problem size (20,000 Twenty Foot Equivalent Unit (TEU) vessels visiting at
least 10 ports) and combinatorial aspects as container dimensions, seaworthiness
and stowage regulations, demand uncertainty and planning best practices.
Several contributions have been unable to directly solve their problem for-
mulation (e.g., [8,23]). Consequently, it is suggested to hierarchically decompose
the stowage problem into master bay and slot planning (e.g., [28,37]), which is
further explained in Sect. 3. Even though plenty of contributions tried, a scalable
algorithm for a representative decomposed problem is yet to be found.
This paper will provide a proof of concept for a novel application of deep rein-
forcement learning (DRL) to solve the master bay planning problem (MBPP)
This work is partially sponsored by the Danish Maritime Fund under grant 2021-069.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 105–121, 2023.
https://doi.org/10.1007/978-3-031-43612-3_6
106 J. van Twiller et al.

as a Markov decision process (MDP). To the best of our knowledge, the only
stowage contributions involving reinforcement learning (RL) relax many of the
complex combinatorial aspects [33,38]. Furthermore, DRL is used to solve opti-
mization problems similar to MBPP [14,17,22,24]. Hence, we believe that DRL
implementations can contribute to solving the stowage planning problem.
In our work, we model an episodic MDP that maximizes vessel utilization
and minimizes hatch-overstowage for equal-sized cargo with two weight classes,
while satisfying demand and location capacity, as well as ensuring longitudinal
and vertical stability. We refer to Sect. 4 for details on the problem. This is not a
full-featured MBPP, but rather a non-trivial problem to provide us with a proof
of concept. Each episode generates a Gaussian equivalent to the Mixed instances
by [4]. We have implemented a proximal policy optimization (PPO) architecture
that learns an actor and critic network to find a policy that can efficiently solve
the MDP. This architecture is compared against an equivalent mixed integer
program (MIP) to evaluate performance.
Our experiments show that PPO can learn a policy that optimizes the objec-
tive function on a limited training budget. Subsequently, the policy can be gener-
alized to efficiently find reasonable solutions for a non-trivial MBPP in a fraction
of the MIP runtime on limited hardware. Thus, we have provided preliminary
evidence for the potential of DRL in stowage planning.
The remainder of this paper is structured as follows: Sect. 2 introduces the
domain of container vessel stowage planning. Section 3 outlines related work,
and Sect. 4 defines the MBPP with a MIP formulation. In Sect. 5, some RL
preliminaries are provided, while our MDP and PPO architecture are described
in Sect. 6. Finally, Sect. 7 compares the results of our PPO architecture with a
MIP solver, and Sect. 8 concludes the main findings of this study.

2 Domain of Container Vessel Stowage Planning

We will first give a brief introduction to container vessel stowage planning. For a
comprehensive description of the domain, we refer to [18]. Liner shipping employs
a fleet of container vessels to serve ports with fixed schedules on a closed-loop
route. An intuitive analogy is a maritime bus line that transports cargo instead
of people between ports. Visiting all ports once is referred to as a voyage.
To transport cargo efficiently, container vessels have a cellular design as
shown in Fig. 1. The vessel is divided into bays (02–38) with stacks and rows
that contain cells capable of holding one 40’ container or two 20’ containers. To
secure the below deck stacks (02–12), the vessel incorporates cell guides. In addi-
tion, the below-deck hold is sealed with hatch covers. Any stacks on deck (82–92)
either rest on hatch covers or the ship deck. Twist locks are used to bind stacked
containers together, lashing rods connect container corners to the deck or lashing
bridges that increase stability. Each stack has weight limits, while below-deck
stacks also have height limits. Figure 1 shows cells (82–84) with power plugs for
refrigerated containers (reefers) indicated by the symbol ∗.
The cargo is typically 20 ft., 40 ft. and 45 ft. long, 8 ft. wide, and 8 ft. 6 in. high
for dry cargo or 9 ft. 6 in. high for highcubes. Container weight ranges between 4
 Towards Deep Reinforcement Learning for Master Bay Planning 107

and 30 tons, depending on the load. Note that empty containers also need to be
redistributed from time to time. Any container has an origin named port of load
(POL), and a destination called the port of discharge (POD). Furthermore, cargo
types can also differ, as Specials include reefers, IMDG that carry dangerous
goods, and OOG containers that deviate from the typical cargo dimensions. To
ensure safety, IMDGs are often separated from other cargo types.

Fig. 1. Vessel side, top and front view [36].

Above and below-deck locations are separated by hatch covers, and thus the
cover must be removed to reach below-deck cargo in any bay. If cargo destined for
future ports is still placed on the cover, cranes must remove and reload (restow)
the on-deck containers that overstow the below-deck cargo. This is referred to
as hatch overstowage. Note that minimizing hatch overstowage is an NP-hard
task, which prevents us from finding an optimal solution in polynomial time [35].
Equivalently, stack overstowage can also occur if stacks are not in ascending order
of PODs from top to bottom. Furthermore, vessels must be seaworthy according
to safety regulations. In general, the cargo must be distributed along the vessel
to ensure acceptable levels of stability and stress forces [18].

3 Related Work
In essence, container vessel stowage planning is a multi-port problem, which
aims to balance a myriad of combinatorial aspects [18]. The field can roughly be
subdivided into single-port work to create light-weight operational stowage plans
(e.g., [2,3,10,23]), and multi-port work to generate realistic stowage plans (e.g.,
[4,8,28,30,31,37]). As demonstrated by [8,23], an exact and optimal solution to
the multi-port problem is yet to be found. Consequently, [37] suggested a hierar-
chical decomposition into master bay planning to allocate groups of containers to
general locations on the vessel (e.g., [7,9,26]), and slot planning to subsequently
allocate containers to slots in locations (e.g., [19,21,29]). For a comprehensive
description of hierarchical decomposition, we refer to [18]. Recently, heuristic
frameworks have also gained traction as an alternative to hierarchical decompo-
sition (e.g., [23,27]).
108 J. van Twiller et al.

Regardless, several solution methods have been proposed to solve different

stowage planning problems, for instance, exact methods (e.g., [31,39]), greedy
heuristics (e.g., [4,11]), population-based (e.g., [12,16]) or neighborhood-based
metaheuristics (e.g., [2,27]), matheuristics (e.g., [21,30]), tree-based methods
(e.g., [5]), or hybrid frameworks (e.g., [7,37]). To the best of our knowledge,
the number of contributions related to RL in stowage planning is limited, which
addressed single-port problems without key combinatorial aspects by deep Q-
learning [33] and Monte Carlo tree search [38]. Despite this variety, we are yet
to find scalable algorithms that solve representative stowage planning problems.
DRL has been rather successful for various combinatorial optimization prob-
lems (e.g., 0–1 knapsack [22], capacitated vehicle routing [14,17,22], and job shop
scheduling [17]). Similar to stowage planning with the master bay subproblem,
chip design is accelerated significantly by optimizing the chip floorplanning sub-
problem with DRL [24]. Furthermore, actor-critic methods can mitigate known
drawbacks of learning from experience and policy gradients by combining both
techniques [34]. In general, actor-critic methods are relatively sample-efficient
with reliable performance, from which PPO performs best in several continuous
control problems [32]. Thus, we believe that PPO merits further investigation.

4 Problem Formulation of Master Bay Planning Problem

Given the previous sections, we can introduce the MBPP. During a voyage, the
MBPP assigns groups of containers on port loadlists to vessel locations. By doing
so, we effectively abstract away individual containers and slots.
The unidirectional voyage is represented by an ordered set P = {1, 2, ...}
of port calls, of which a set of all possible transport pairs T = {τ = (i, j) ∈
P 2 | i < j} is constructed to specify the POL and POD of cargo. At the departure
from an arbitrary port p, the onboard cargo can be characterized by the set
T OB (p) = {(i, j) ∈ P 2 | i ≤ p, j > p}. At the next port, this onboard cargo
is either discharged or remains on board (ROB) as defined by set T ROB (p) =
{(i, j) ∈ P 2 | i < p, j > p}. Each group of containers has the same POL and
POD (i, j) ∈ T as well as class k ∈ K that defines the dimensions, weight, and
type of the containers in the group. Each vessel location is defined by bay b ∈ B
and deck d ∈ D = {DO , DH }, where DO represents on-deck locations and DH
represents below-deck locations in the hold.
The capacity expressed in TEU of a location in bay b ∈ B and deck d ∈ D
is given by cb,d . The weight in tons per container in a group with class k ∈ K
is given by wk . Notice that our cargo is equal-sized and specials are not taken
into account. The cargo demand in TEU from port i (POL) to port j (POD)
of class k is given by qτ,k , where τ = (i, j). Different instances are generated by
sampling qτ,k from a Gaussian distribution for each class k and transport τ as
shown in Eq. 1. The expected value μ is a single random instance of the Mixed
instances by [4] and σ represents the standard deviation to introduce variability
around μ. Hence, Gaussian equivalents of Mixed instances are generated.
qτ,k ∼ Qτ,k = N (μ, σ) ∀τ ∈ T, k ∈ K (1)
 Towards Deep Reinforcement Learning for Master Bay Planning 109

The primary objective is to maximize vessel utilization by loading cargo that

satisfies port demand, while the secondary goal is to minimize hatch overstowage
and arises from efficiency best practices. To prevent safety hazards (e.g., capsiz-
ing or falling containers), vessels must have an even keel and sufficient transverse
stability, measured by trim and metacentric height (GM). Which in turn are
determined by longitudinal (LCG) and vertical centre of gravity (VCG) [18].

4.1 MIP Model of the MBPP

The following model of the MBPP is inspired by the MIP formulation of [28].
⎛ ⎞
   
max ⎝ f1 xb,d
τ,k − f2 yb,p ⎠ (2)
p∈P b∈B d∈D k∈K τ ∈T OB (p) b∈B


s.t. xb,d
τ,k ≤ qτ,k ∀p ∈ P, τ ∈ T OB (p), k ∈ K (3)
b∈B d∈D
 
xb,d
τ,k ≤ cb,d ∀p ∈ P, b ∈ B, d ∈ D (4)
k∈K τ ∈T OB (p)
 
xb,d
(p,j),k ≤ M zb,p ∀p ∈ P, b ∈ B, d ∈ DH (5)
k∈K j∈P :j>p
 
xb,d
(i,p),k ≤ M zb,p ∀p ∈ P, b ∈ B, d ∈ DH (6)
k∈K i∈P :i<p
 
xb,d
τ,k − M (1 − zb,p ) ≤ yb,p ∀p ∈ P, b ∈ B, d ∈ DO (7)
k∈K τ ∈T ROB (p)

LM (p) − TW (p)LCG ∗ ≤ υ · TW (p) ∀p ∈ P (8)

LM (p) − TW (p)LCG ∗ ≤ −υ · TW (p) ∀p ∈ P (9)
VM (p) − TW (p)VCG ∗ ≤ υ · TW (p) ∀p ∈ P (10)
VM (p) − TW (p)VCG ∗ ≤ −υ · TW (p) ∀p ∈ P (11)
    b,d
TW (p) = wk xτ,k ∀p ∈ P (12)
k∈K τ ∈T OB (p) d∈D b∈B
   
LM (p) = LP b wk xb,d
τ,k ∀p ∈ P (13)
b∈B k∈K τ ∈T OB (p) d∈D
   
VM (p) = VP d wk xb,d
τ,k ∀p ∈ P (14)
d∈D k∈K τ ∈T OB (p) b∈B

The primary decision variable of our MIP model is xb,d (i,j),k ∈ R≥0 , which is
the number of TEU from port i to j of class k planned to be stowed in bay b
and deck d. The secondary decision variable yb,p ∈ R≥0 is the number of TEU
that causes hatch overstowage in bay b at port p, while the last decision variable
zb,p ∈ {0, 1} indicates whether the hatch is opened in bay b at port p. Equation 2
shows that the MIP aims to maximize vessel utilization and minimize hatch
overstowage, where f1 is the gain per loaded TEU and f2 is the cost per hatch
overstow in TEU. Equation 3 limits the total vessel utilization by demand qτ,k ,
110 J. van Twiller et al.

while Eq. 4 limits the onboard cargo by the capacity volume cb,d . Equation 5
and 6 connects stowed containers xb,d(i,j),k to hatch moves zb,p using the big M
notation, whereas Eq. 7 computes hatch restows based on whether the cargo is
placed on the to-be-opened hatch. Let us define LCG(p) = LM (p)/TW (p) and
VCG(p) = VM (p)/TW (p)
 at port p, and their optimal equivalents LCG ∗ =
∗
b LP b /|B| and V CG = d VP d /|D|. Equations 8 and 9 limit the difference
between the actual and optimal LCG with parameter υ. Similarly, Eqs. 10 and
11 limit the difference between the actual and optimal VCG with parameter
υ. Note that Eqs. 8–11 are linear transformations of |LCG(p) − LCG ∗ | < υ
and |VCG(p) − VCG ∗ | < υ. Equations 12–14 define the total weight TW (p),
longitudinal moment LM (p), and vertical moment VM (p). Notice that LP b and
VP d refer to the longitudinal and vertical position of locations respectively.

5 Preliminaries of Reinforcement Learning

RL is a set of algorithms inspired by neurological reward processes, where agents
learn how to take sequential decisions in environments [34]. The idea might be
inspired by trial-and-error learning, in which experiments indicated that animals
tend to strengthen or weaken connections to situations based on positive or nega-
tive outcomes respectively [34]. Those environments are often modeled by MDPs.
The agent’s goal is to find an optimal policy π ∗ that maximizes discounted future
rewards, thereby solving the MDP. Since RL is relatively unknown in stowage
research, we will introduce some fundamentals and refer to additional citations.
We define an episodic MDP with tuple (S, A, P, R), where S is a set of
states and A is a set of actions. The stochastic function P(s |s, a) provides the
likelihood of reaching state s given state s and action a. If the probability
function is deterministic as in our case, T : S × A → S is used as a transition
function. The reward function is defined as R : S × A → R. We also define
our policy π(a|s) as the probability that action a is executed in state s. In the
deterministic case, we can define a policy function π : S → A.
Let us assume a stochastic transition and policy function. An episode starts
with the agent in the initial state s0 and obtaining action a from policy π(a|s).
This is referred to as an on-policy agent. Subsequently, the environment obtains
the next state s from P(s |s, a) but also reward r from reward function R(s, a),
then the agent observes s and r. This closed-loop interaction continues until
a terminal state sterm is reached, which ends the episode and returns the sum
of accumulated rewards. The next episode resets the agent to s0 and continues
until the total number of steps Ts is reached. The performance of a trajectory
under π is defined as the expected return in Eq. 15, which is extended to Eqs. 16
and 17 as the expected returns dependent on state st (i.e., value function), and
state-action pair (st , at ) (i.e., Q-function) respectively. Note that γ ∈ [0, 1) is a
discount factor that enables asymptotic convergence.
 Towards Deep Reinforcement Learning for Master Bay Planning 111

 Ts  Ts−t 
 t

J = Eπ γ rt (15) V (st ) = Eπ k
γ rt+k |s = st (16)
t=0 k=0
Ts−t 

Q(st , at ) = Eπ k
γ rt+k |s = st , a = at (17)
k=0

Solving MDPs is analogous to finding an optimal policy π ∗ that prescribes

actions to maximize future discounted rewards, thereby causing J to converge.
Several RL algorithms can be used to do so. For instance, dynamic programming
(DP) consists of model-based methods using knowledge of P(s |s, a) and R(s, a)
to estimate solution V (s) by planning ahead as shown in Eq. 18. Even though
convergence to π ∗ (a|s) is guaranteed, DP is intractable for large state and action
spaces and assumes perfect information which is often unrealistic [34].


V (s) = max R(s, a) + γ P(s |s, a)V (s ) (18)
a
s

Instead, one could consider learning from experience using model-free meth-
ods, often suffering from bias. For example, Monte Carlo methods perform trial-
and-error simulations, whereas temporal difference learning obtains approximate
functions with bootstrapping [34]. Another way is through DRL, which approx-
imates good solutions with parameterized functions in large state and action
spaces (e.g., policy πθ ). The policy gradient theorem states that the derivative
of the expected return is proportional to the terms in Eq. 19, where ψπ (st ) is the
on-policy probability of state st . Doing so, πθ can be learned by updating θ with
respect to expected return J [34]. Due to large policy updates, gradient-based
methods can suffer from variance. Anybody unfamiliar with RL is suggested to
read more on the topic in [34].

 
 
∇J(θ) ∝ Eπ ψπ (st ) ln ∇πθ (at |st )Qπ (st , at ) (19)
st at

6 Solving MBPP with Reinforcement Learning

The state st ∈ S at time step t is defined with pair st = (u(t), q) and is fully
observable. It consists of the vessel utilization u(t) ∈ R|B|×|D|×|T |×|K| and a
constant voyage demand quantity q ∈ R|T |×|K| . As an example ub,d (i,j),k (t) = 0.05
means that 5% of the vessel capacity is cargo stowed in bay b at deck d with
POL i, POD j and class k at step t. Similarly, q(i,j),k = 0.13 means that there is
a cargo demand of 13% of the vessel capacity with POL i, POD j and class k.
At port p, the agent takes an action at ∈ A for every (i, j) ∈ T OB (p) : i =
p, j = j  . Each action is a combination of port p and future port j  as shown in
Fig. 2. The action is defined by a pair at = (l(t), τt ), where l(t) ∈ R|B|×|D|×|K| is
112 J. van Twiller et al.

the fraction of vessel capacity to load in bay b at deck d of class k on transport

τt = (p, j  ). In Eq. 20, we define the transition function to step t + 1 with an
input of τt . If cargo remains on board, then ub,d
τ,k (t) is unchanged. If cargo should
be loaded (i.e., τ = τt ), then the utilization will become lkb,d (t). Otherwise, cargo
that should not be on board will be set to zero.

⎧ b,d
⎪
⎨uτ,k (t) if τ ∈ T OB (p) \ {τt }
ub,d (t + 1) = lkb,d (t) if τ = τt ∀b ∈ B, d ∈ D, τ ∈ T, k ∈ K (20)
τ,k
⎪
⎩
0 otherwise

Since agents struggle with sparsity [13], a constant reward signal is neces-
sary for each action. Equation 21 defines the reward of satisfying demand with
reward coefficient f1 and input τt = (p, j  ) for current port p and future port
j  . A penalty is incurred with coefficient f3 if the agent ships more cargo than
demanded. Note that f3 > f1 holds to enforce feasibility.
  
f1 b,d,k τ ∈T OB (p) ub,d b,d uτt ,k ≤ qτt ,k
b,d
τ,k , if 
rτDS
t
=   ∀k ∈ K (21)
−f3 b,d,k τ ∈T OB (p) ub,d
τ,k , otherwise

In addition, the following terms are evaluated at (p, j  ) ∈ T : j  = |P |.

Equations 22 and 23 enforce that the actual longitudinal or vertical centre of
gravity (i.e., LCG t or VCG t ) can at most deviate υ from the optimal longitudinal
or vertical centre of gravity (i.e., LCG ∗ or VCG ∗ ) with infeasibility penalty.
Given the definitions of Eqs. 12 until 14, let us substitute xb,d b,d
τ,k by uτ,k and define
LM (p) VM (p)
LCG t = TW (p) and VCG t = TW (p) . Equation 24 incurs an infeasibility penalty
for violating capacity. While hatch overstowage is evaluated in Eq. 25, where we
penalize hatch restows with coefficient f2 . The cost of restowing a container is
lower than its revenue, and thus f1 > f2 holds.

LS −f3 |LCG t − LCG ∗ |, if j  = |P | ∧ |LCG t − LCG ∗ | > υ

r(p,j ) = (22)
0, otherwise

VS −f3 |VCG t − VCG ∗ |, if j  = |P | ∧ |VCG t − VCG ∗ | > υ

r(p,j ) = (23)
0, otherwise
⎧  
⎪
⎪ −f3 ub,d
τ,k ,
if j  = |P | ∧ ∃b ∈ B, d ∈ D.
⎪
⎨    
CS
r(p,j ) =
b,d,k τ ∈T OB (p)
ubτ,k,d > cb ,d (24)
⎪
⎪ k τ ∈T OB (p)
⎪
⎩
0, otherwise
 Towards Deep Reinforcement Learning for Master Bay Planning 113
⎧  
⎪
⎪ −f2 ub,d
τ,k ,
if j  = |P | ∧ ∃b ∈ B, d ∈ DH .
⎪
⎪  
⎪
⎪ O τ ∈T ROB (p)  
⎪
⎪
b,k,d∈D
ubi,p,k
,d
>0
⎪
⎪
⎨   k i∈P :i<p
HO
r(p,j ) = −f2 ub,d
τ,k ,
if j  = |P | ∧ ∃b ∈ B, d ∈ DH . (25)
⎪
⎪  
⎪
⎪ b,k,d∈D O τ ∈T ROB (p)
 
ubp,j,k
,d
⎪
⎪ >0
⎪
⎪
⎪
⎪ k j∈P :j>p
⎩
0, otherwise

Equation 26 defines a reward function R(st , at ) for each step t, where the
above-mentioned terms are summed. However, a final term penalizes unsatisfied
demand by including the upper bound of rτDSt
with parameter f1 . Hence, we
maximize a reward function with an upper bound of 0.

r(st , at ) = rτDS
t
+ rτLS
t
+ rτVS
t
+ rτCS
t
+ rτHO
t
− f1 qτ,k (26)
k,τ ∈T OB (p)

Figure 2 illustrates an example with 4 ports of our episodic MDP. Each

episode consists of |T | = (|P |2 − |P |)/2 actions during a unidirectional voy-
age, corresponding to the number of above-diagonal elements in a matrix P 2 .
An episode is initialized with an empty vessel in state s0 (i.e., u0 is set to 0),
whereas a sample qτ,k is drawn from the demand distribution Qτ,k . During the
episode, the agent traverses all ports p ∈ P and takes an action according to
policy πθ (at |st ) for each τt = (p, j  ) with j  ∈ P : j  > p. Figure 2 differenti-
ates between rewards for each transport pair rτt and the last action of each port
r(p,|P |) . An episode is terminated at arbitrary step t that satisfies mod t/|T | = 0
as all actions are performed. Finally, the episodic reward is provided as output.

Fig. 2. Overview of master bay planning MDP with colours corresponding to PODs.

6.1 Proximal Policy Optimization Architecture

This on-policy architecture consists of an actor neural network with weights θ to
create policy πθ (at |st ) and a critic neural network with weights ω to evaluate the
performance of the policy by approximating the value function Vω (st ). The value
114 J. van Twiller et al.

function estimation Vω (st ) is used to express the relative advantage of taking an

action. Equation 27 defines the general advantage estimate Ât of (st , at )-pair at
step t, where λ is the exponential decay rate and δt is the temporal difference
residual found by stepwise bootstrapping in Eq. 28. Both πθ (at |st ) and Vω (st )
are multilayer perceptrons that approximate non-linear functions by learning the
mean and standard deviation of continuous Gaussian distributions.

H−(t+1)
Ât = (γλ)l δt+l (27) δt = r(st , at ) + γVω (st+1 ) − Vω (st ) (28)
l=0

Let us define the loss function with respect to ω as the mean squared error
between the estimated value function and the sum of discounted rewards in
Eq. 29. Given the estimated advantage Ât , Eq. 30 defines a clipped loss func-
tion with respect to θ, where pt (θ) = ππθ θ (a(at |s t)
t |st )
is the importance sampling
old
ratio between the current and previous policy distribution. The clipping limit
prevents disruptive changes to θ. In Eq. 31, both functions are combined with
entropy regularization, where cf 1 and cf 2 are coefficients and S[πθ ](st ) is the
entropy term that promotes exploration based on πθ and st . Hence, Eq. 31 is
maximized to update θ and thereby policy πθ .
 2
VF

t
L (ω) = Êt Vω (st ) − t
γ r(st , at ) (29)
i=0
 
LCLIP (θ) = Êt min(pt (θ)Ât , clip(pt (θ), 1 − , 1 + )Ât ) (30)
 
LCLIP +V F (θ) = Êt LCLIP (θ) − cf 1 LV F (ω) + cf 2 S[πθ ](st ) (31)

A description of PPO is shown in Algorithm 1. For each step t, let N parallel

actors run policy πθold based on parameter θold for a time horizon of H timesteps.
The estimated advantage Ât is computed for every step t ∈ {1, ..., H}. Using a
minibatch of M steps, we optimize the actor and critic loss w.r.t. θ and ω using
the Adam solver [20] for E epochs. If θ and ω are trained for a sufficient number
of steps, then we can obtain an actor with πθ ≈ π ∗ and a critic with Vω ≈ V ∗ .

Algorithm 1: Proximal Policy Optimization

Initialize ω = ω0 , θ = θ0 ;
for t = 1, 2, ..., Ts do
for actor = 1, 2, ..., N do
Run policy πθold in environment for horizon with H timesteps;
Compute general advantage estimates Â1 , ..., ÂH ;
Optimize LV F (ω) w.r.t. ω with E epochs and minibatch M ≤ N H;
Optimize LCLIP +V F (θ) w.r.t. θ with E epochs and minibatch M ≤ N H;
ωold ← ω; θold ← θ;
return πθ
 Towards Deep Reinforcement Learning for Master Bay Planning 115

6.2 Hyperparameter Tuning

There are many hyperparameters to be determined that impact the performance

of PPO. We will, however, limit ourselves to the most impactful ones mentioned
below. The time horizon with H steps is defined as the number of samples
propagated through the network for each actor, while the number of epochs E is
the number of passes through the experience buffer. The total buffer size amounts
to N H and the minibatch size equals the buffer M = N H. The learning rate
α ∈ (0, 1] defines the update size of network weights (i.e., θ, ω), which decreases
over time to reduce the impact of updates. The actor and critic networks have the
same number of hidden layers (depth) and neurons in each hidden layer (width).
The layer activation function is ReLu to deal with vanishing gradients [15]. Both
πθ (at |st ) and Vω (st ) are learned by their mean and log standard deviation that
are initialized at 0 and init log std respectively.
We implement a tree-structured Parzen estimator using an optimization
framework [1], which is a Bayesian method that can efficiently sample hyper-
parameters for various optimization use-cases [6]. The goal is to maximize the
episodic return of trials in 106 steps based on samples. To improve efficiency,
a median pruner stops trials if its best intermediate result is worse than the
median of earlier trials at the same time step [1]. In total, we run 100 trials with
different hyperparameters, of which 95 trials use the pruner.
The trials are run with the following MBPP parameters: ports |P | = 4,
classes |K| = 2, bays |B| = 4, decks |D| = 2, location capacity in TEU cb,d =
50, weight class wk = {1, 2}, absolute centre of gravity tolerance υ = 0.05,
longitudinal position of bays LP b = {0.25, 0.75, 1.25, 1.75}, vertical position of
decks VP d = {0.5, 1.5}, gain per loaded TEU f1 = 1, cost per hatch overstowed
TEU f2 = 1/3, and penalty to violate constraints f3 = 3. Given N = 1, the
highest episodic return is obtained with H = 512, E = 10, α = 3e−4, 3 hidden
layers, 1024 neurons, and init log std = 6.75. Other hyperparameters are set as
follows: Ts = 4 · 107 , M = 512, γ = 0.99, λ = 0.95, cf 1 = 0.5 and cf 2 = 1e−6.

7 Results

In Subsect. 6.2, the respective environment and PPO (hyper)parameters are

defined. The usefulness of PPO is demonstrated by addressing a small-scale
MBPP. Even though the MBPP is not representative of real-life stowage plans,
it does capture fundamental combinatorial aspects to form a non-trivial prob-
lem. The results analyze whether the policy learns to optimize the objective,
but also whether the policy generalizes to new test instances. To evaluate test
performance, we compare the objective value and runtime of PPO with a MIP
solver on two sets of instances. The first set of 100 instances Gaus-MBPP con-
tains Gaussian equivalents to the Mixed instances of [4]. The second set of 100
instances Unif-MBPP consists of uniform versions of the Mixed instances. Due
to reward shaping, the MIP objective and the reward function are proportional
but not equal, and therefore we transform the objective before comparison.
116 J. van Twiller et al.

The experiments are run on a Windows machine with an NVIDIA RTX

A2000 Laptop GPU with 12.0 GB memory, and an Intel Core i7-11800H pro-
cessor with 8 cores and 32.0 GB memory, running at 2.3/4.6 GHz. The work is
implemented in Python 3.9 supported by libraries such as Gym 0.21 to model
the environment, PyTorch 1.11 and Stable Baselines 1.6.2 to implement PPO,
Optuna 3.0.3 to tune hyperparameters, and CPLEX 22.1 to solve the MIP model.
In Fig. 3, several training metrics are included to analyze training. Subfigure
3(a) plots the episodic return against episodes to evaluate the training perfor-
mance of PPO on Gaus-MBPP. In general, the training curve seems to con-
verge towards an increasingly stable return around 70,000 episodes. Due to the
stochasticity of Gaus-MBPP, performance will vary over time. In early training,
PPO learns to avoid large negative rewards associated with infeasible solutions,
after which the episodic return follows an upward trend without converging defi-
nitely. In addition, Subfigure 3(b) shows the loss function optimized by the actor,
which has yet to converge but starts to reduce in volatility. To evaluate the dif-
ference between distributions, one can use the approximate KL divergence [13].
Subfigure 3(c) shows that the approximate KL divergence between πθ (at |st ) and
πθold (at |st ) is stabilizing. Similarly, Subfigure 3(d) shows that the change to pt (θ)
is lower than the clipping range , indicating small actor changes. In Subfigure
3(e), we observe that log std starts to converge, meaning that the variability of
πθ (at |st ) reduces. Thus, we have found an increasingly stable policy with one
actor on a limited training budget, suggesting there is room for improvement.
Nonetheless, we have found a policy that optimizes the objective function.
In Table 1, we evaluate the performance of PPO and MIP on Gaus-MBPP
and Unif-MBPP. Since the instance size is relatively small, the benchmark MIP
will solve all instances. In comparison, PPO finds 98 feasible solutions with an
average gap of 12.5% on Gaus-MBPP with low levels of variability. Even though
the results are not near-optimal, a larger training budget is likely to improve
performance as our policy approaches the global optimum [24,25]. Moreover, if
MIP solvers become intractable for large-scale problems, the significantly shorter
runtime of PPO will be advantageous. Hence, we find that PPO generalizes its
policy to the Gaus-MBPP instances with reasonable performance.
Furthermore, the policy struggles to generalize to the Unif-MBPP instances,
which is not the case for the MIP. Even if feasible solutions are found, then
the average optimality gap is 81.3% with large variability in the objective value.
Our function approximator struggles to accurately predict uniform instances
with a different ratio than Gaus-MBPP. DRL methods generally learn to adapt
to the underlying distribution, and thus it is likely that sampling outside of
this distribution leads to performance loss. Though this is a clear drawback of
PPO, it is our understanding that demand often follows predictable patterns.
Therefore, it is important to accurately model demand based on real data.
Although we have found encouraging results, we realize that DRL demands
considerably more computational power than provided by our machine appropri-
ate for MIP optimization. Usually, laptop GPUs have fewer cores, less VRAM,
and slower clock speeds than their desktop counterparts. This causes long train-
Towards Deep Reinforcement Learning for Master Bay Planning 117

Fig. 3. Training metrics on Gaus-MBPP instances.

118 J. van Twiller et al.

Table 1. Evaluation of PPO and MIP on two sets of 100 test instances generated from
Gaus-MBPP and Unif-MBPP. Results are expressed in the mean average and standard
deviation metrics for the objective value (Obj.) and runtime in seconds (Time). The
number of feasible solutions (#) is given without std. This also holds for the optimality
gap (Gap) that is computed relative to the optimal MIP objective.

Methods Metric Gaus-MBPP Unif-MBPP

# Obj. Gap Time # Obj. Gap Time
PPO Mean 98 2.092 12.5% 0.013 13 0.444 81.3% 0.006
PPO Std 0.135 0.002 0.359 0.000
MIP Mean 100 2.366 0.0% 0.058 100 1.930 0.0% 0.035
MIP Std 0.261 0.022 0.387 0.016

ing times, sub-optimal convergence, and reduced test performance. To train a

DRL model efficiently, it is therefore recommended to use specialized hardware
such as high-end GPUs, TPUs, or cloud-based computing resources. Since this
is not easily accessible, we will invest in comparable hardware to a workstation
with 4x NVIDIA RTX A6000 with 48 GB memory and a Threadripper Pro 3955
WX with 16 cores and 256 GB memory, running at 3.9/4.3 GHz.
In conclusion, PPO can learn a policy that optimizes the objective function
during training on Gaus-MBPP instances. Afterward, this policy can be gener-
alized to efficiently achieve reasonable performance on unseen Gaus-MBPP test
instances. When confronted with the Unif-MBPP instances, PPO mostly obtains
infeasible or weak solutions. Using this proof of concept, we argue in favor of
using PPO to solve the MBPP.

8 Conclusion

This paper presents a DRL approach towards solving the MBPP in container
vessel stowage planning. In particular, we introduce a MDP equivalent to a MIP
model with NP-hard combinatorial aspects, as well as suggest PPO to solve the
MDP. These preliminary experiments show that PPO learns to optimize the
objective value during training on limited hardware, after which its policy can
be generalized to efficiently find reasonable solutions for test instances. Hence,
we have provided a proof of concept for applying PPO to the MBPP.
In future work, we will extend the environment to become a full-featured
MBPP with a representative demand simulator. The algorithm will also be
improved to increase performance and deal with more complex problems. Since
voyages are inherently sequential, we could implement temporal dynamic poli-
cies using recurrent neural networks (e.g., long short-term memory networks).
Considering the cellular shape of container vessels, we might also leverage graph
representation learning in our architecture.
 Towards Deep Reinforcement Learning for Master Bay Planning 119

References
1. Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.: Optuna: a next-generation
hyperparameter optimization framework (2019). http://arxiv.org/abs/1907.10902
2. Ambrosino, D., Anghinolfi, D., Paolucci, M., Sciomachen, A.: An experimental
comparison of different heuristics for the master bay plan problem. In: Festa,
P. (ed.) SEA 2010. LNCS, vol. 6049, pp. 314–325. Springer, Heidelberg (2010).
https://doi.org/10.1007/978-3-642-13193-6 27
3. Ambrosino, D., Sciomachen, A., Tanfani, E.: Stowing a containership: the master
bay plan problem. Transp. Res. Part A 38(2), 81–99 (2004). https://doi.org/10.
1016/j.tra.2003.09.002, http://linkinghub.elsevier.com/retrieve/pii/S09658564030
00892, http://www.sciencedirect.com/science/article/pii/S0965856403000892
4. Avriel, M., Penn, M., Shpirer, N., Witteboon, S.: Stowage planning for container
ships to reduce the number of shifts. Ann. Oper. Res. 76(1-4), 55–71 (1998).
http://search.ebscohost.com/login.aspx?direct=true&db=bth&AN=18925445&
site=ehost-live, http://link.springer.com/article/10.1023/A:1018956823693
5. Azevedo, A., Cassilda Maria, R., de Sena, G.J., Chaves, A.A., Neto, L.L.S., Moretti,
A.C.: Solving the 3D container ship loading planning problem by representation by
rules and meta-heuristics. Int. J. Data Anal. Tech. Strategies 6(3), 228–260 (2014).
https://doi.org/10.1504/IJDATS.2014.063060
6. Bergstra, J., Bardenet, R., Bengio, Y., Kégl, B.: Algorithms for hyper-parameter
optimization. In: NIPS 2011: Proceedings of the 24th International Con-
ference on Neural Information Processing Systems, pp. 2546–2554 (2011).
https://proceedings.neurips.cc/paper files/paper/2011/file/86e8f7ab32cfd12577bc
2619bc635690-Paper.pdf
7. Bilican, M.S., Evren, R., Karatas, M.: A mathematical model and two-stage heuris-
tic for the container stowage planning problem with stability parameters. IEEE
Access 8, 113392–113413 (2020). https://doi.org/10.1109/ACCESS.2020.3003557
8. Botter, R., Brinati, M.: Stowage container planning: a model for getting an optimal
solution. Comput. Appl. Autom. Shipyard Oper. Ship Design VII(C), 217–228
(1992)
9. Chao, S.L., Lin, P.H.: Minimizing overstowage in master bay plans of large con-
tainer ships. Maritime Econ. Logist. 23(1), 71–93 (2021). https://doi.org/10.1057/
s41278-019-00126-6
10. Delgado, A., Jensen, R.M., Guilbert, N.: A placement heuristic for a commercial
decision support system for container vessel stowage. In: 38th Latin America Con-
ference on Informatics, CLEI 2012 - Conference Proceedings (2012). https://doi.
org/10.1109/CLEI.2012.6427181
11. Ding, D., Chou, M.C.: Stowage planning for container ships: a heuristic algo-
rithm to reduce the number of shifts. Eur. J. Oper. Res. (2015). https://
doi.org/10.1016/j.ejor.2015.03.044, http://linkinghub.elsevier.com/retrieve/pii/
S0377221715002660
12. Dubrovsky, O., Levitin, G., Penn, M.: A genetic algorithm with a compact solu-
tion encoding for the container ship stowage problem. J. Heuristics 8(6), 585–
599 (2002). https://doi.org/10.1023/A:1020373709350, http://link.springer.com/
article/10.1023/A:1020373709350
13. Engstrom, L., et al.: Implementation matters in deep policy gradients: a case study
on PPO and TRPO (2020). http://arxiv.org/abs/2005.12729
14. Foa, S., Coppola, C., Grani, G., Palagi, L.: Solving the vehicle routing problem
with deep reinforcement learning (2022). http://arxiv.org/abs/2208.00202
120 J. van Twiller et al.

15. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016). http://
www.deeplearningbook.org
16. Hamedi, M.: Containership load planning with crane operations. Ph.D. thesis, Uni-
versity of Maryland (2011)
17. Hottung, A., Kwon, Y.D., Tierney, K.: Efficient active search for combinatorial
optimization problems. In: International Conference on Learning Representations
2022 (2021). http://arxiv.org/abs/2106.05126
18. Jensen, R.M., Pacino, D., Ajspur, M.L., Vesterdal, C.: Container vessel stowage
planning. Weilbach (2018)
19. Kebedow, K.G., Oppen, J.: Including containers with dangerous goods in the slot
planning problem. In: 2019 Proceedings of the International Conference on Indus-
trial Engineering and Operations Management, pp. 225–232 (2019)
20. Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. In: 2015 Inter-
national Conference on Learning Representations (2014). http://arxiv.org/abs/
1412.6980
21. Korach, A., Brouer, B.D., Jensen, R.M.: Matheuristics for slot planning of con-
tainer vessel bays. Eur. J. Oper. Res. 282(3), 873–885 (2020). https://doi.org/10.
1016/j.ejor.2019.09.042
22. Kwon, Y.D., Choo, J., Kim, B., Yoon, I., Gwon, Y., Min, S.: POMO: policy opti-
mization with multiple optima for reinforcement learning (2020). https://doi.org/
10.48550/ARXIV.2010.16011, https://arxiv.org/abs/2010.16011
23. Larsen, R., Pacino, D.: A heuristic and a benchmark for the stowage planning
problem. Maritime Econ. Logist. 23(1), 94–122 (2021). https://doi.org/10.1057/
s41278-020-00172-5
24. Mirhoseini, A., et al.: A graph placement methodology for fast chip design. Nature
594(7862), 207–212 (2021). https://doi.org/10.1038/s41586-021-03544-w
25. Mnih, V., et al.: Playing Atari with Deep Reinforcement Learning (2013)
26. Pacino, D.: An LNS approach for container stowage multi-port master planning.
In: Computational Logistics, pp. 35–44 (2013). https://doi.org/10.1007/978-3-642-
41019-2 3, http://link.springer.com/10.1007/978-3-642-41019-2 3
27. Pacino, D.: Crane intensity and block stowage strategies in stowage planning. In:
Cerulli, R., Raiconi, A., Voß, S. (eds.) ICCL 2018. LNCS, vol. 11184, pp. 191–206.
Springer, Cham (2018). https://doi.org/10.1007/978-3-030-00898-7 12
28. Pacino, D., Delgado, A., Jensen, R.M., Bebbington, T.: Fast generation of near-
optimal plans for eco-efficient stowage of large container vessels. In: Böse, J.W., Hu,
H., Jahn, C., Shi, X., Stahlbock, R., Voß, S. (eds.) ICCL 2011. LNCS, vol. 6971, pp.
286–301. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-24264-
9 22
29. Pacino, D., Jensen, R.M.: A 3-phase randomized constraint based local
search algorithm for stowing under deck locations of container ves-
sel bays (2010). https://en.itu.dk/-/media/EN/Research/About-ITU-Research/
Technical-Reports/2010/ITU-TR-2010-123-pdf
30. Parreño-Torres, C., Çalık, H., Alvarez-Valdes, R., Ruiz, R.: Solving the generalized
multi-port container stowage planning problem by a matheuristic algorithm. Com-
put. Oper. Res. 133, 105383 (2021). https://doi.org/10.1016/j.cor.2021.105383
31. Roberti, R., Pacino, D.: A decomposition method for finding optimal container
stowage plans. Transp. Sci. 52(6), 1444–1462 (2018). https://doi.org/10.1287/trsc.
2017.0795
32. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy
optimization algorithms (2017). http://arxiv.org/abs/1707.06347
 Towards Deep Reinforcement Learning for Master Bay Planning 121

33. Shen, Y., Zhao, N., Xia, M., Du, X.: A deep Q-learning network for ship stowage
planning problem. Pol. Maritime Res. 24(S3), 102–109 (2017). https://doi.org/10.
1515/pomr-2017-0111
34. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press,
Cambridge (2018)
35. Tierney, K., Pacino, D., Jensen, R.M.: On the complexity of container stowage
planning problems. Discrete Appl. Math. 169, 225–230 (2014). https://doi.org/10.
1016/j.dam.2014.01.005, http://dx.doi.org/10.1016/j.dam.2014.01.005
36. van Twiller, J., Sivertsen, A., Pacino, D., Jensen, R.M.: Literature survey on the
container stowage planning problem. Under review (2023)
37. Wilson, I.D., Roach, P.A.: Container stowage planning: a methodology for generat-
ing computerised solutions. J. Oper. Res. Soc. 51(11), 1248–1255 (2000). https://
doi.org/10.1057/palgrave.jors.2601022, https://www.tandfonline.com/doi/full/10.
1057/palgrave.jors.2601022
38. Zhao, N., Guo, Y., Xiang, T., Xia, M., Shen, Y., Mi, C.: Container ship stowage
based on Monte Carlo tree search. J. Coastal Res. 83, 540–547 (2018). https://
doi.org/10.2112/SI83-090.1
39. Zhu, H., Ji, M., Guo, W.: Integer linear programming models for the containership
stowage problem. Math. Probl. Eng. 2020 (2020). https://doi.org/10.1155/2020/
4382745
 The Dynamic RORO Stowage Planning
Problem

Alastair Ronald Main(B) , Filipe Rodrigues , and Dario Pacino

DTU Management, Technical University of Denmark,

Akademivej Building 358, 2800 Kgs. Lyngby, Denmark
{aroma,rodr,darpa}@dtu.dk
https://www.man.dtu.dk/

Abstract. Reducing emissions from the Roll-on/Roll-off (RORO) ship-

ping industry has seen an increase in focus in the past years. Reducing
the turnaround time in ports through stowage planning will increase
slow-steaming use. Stowage planning assigns cargo to positions on board
the vessel. This paper studies how the dynamic arrival of cargo affects
stowage planning by considering revenue from shipping cargo vs. fuel
costs incurred from time spent waiting and stowing cargo. A mixed-
integer program formulation and a heuristic approach are presented to
solve the problem. The computational study shows how the heuristic
can find stowage plans for the dynamic arrival of cargo. A sensitivity
analysis is conducted to investigate algorithm sensitivity in relation to
revenue, fuel costs, and cargo handling time. The results indicate a high
sensitivity in the number of units of cargo stowed when these parameters
fluctuate.

Keywords: RORO shipping · Dynamic stowage planning · Heuristic

1 Introduction
The interest in reducing Green House Gas (GHG) emissions within the ship-
ping sector has greatly increased over the past few years [17]. Governments and
organizations are adopting new targets for reducing GHG emissions. The Inter-
national Maritime Organization’s (IMO) target is to reduce GHG emissions by
50% from shipping by 2050 compared to the GHG levels in 2008 [2]. The GHG
emissions from shipping depend on various parameters, i.e., ship size, weight,
hull shape, ballast water management, and sailing speed [17]. As the shipping
industry is experiencing a general increase in demand, it is paramount to develop
new strategies and technologies for the industry [2,6].
One of these shipping industries is Roll-on/Roll-off (RORO) shipping. RORO
shipping focuses on the transportation of wheeled cargo such as semi-trailers,
cars, buses, and farm equipment. RORO ships have the advantage of being able
This work is partly funded by the Innovation Fund Denmark (IFD) under File No.
0177-00022B.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 122–138, 2023.
https://doi.org/10.1007/978-3-031-43612-3_7
 The Dynamic RORO Stowage Planning Problem 123

to transport heterogeneous cargo with respect to weights, sizes, and shapes,

allowing for more flexibility. RORO cargo is the fourth largest shipping type,
only surpassed by liquid, dry bulk, and container shipping [11].
A known way to reduce GHG emissions of ships is through slow steaming.
By reducing the sailing speed, the consumption of fuel is significantly reduced.
To increase the time available to sail between port calls, the turnaround time in
ports needs to be reduced. This can be done by optimizing the stowage planning.
Stowage planning is the process of assigning cargo to specific locations on board
the vessel. When assigning cargo on board the vessel, several constraints must be
considered. The allocation must ensure vessel stability. Hazardous and refriger-
ated cargo requires special attention due to the need for electrical connections or
safety distances [7,14]. Other cargo can have size and weight limitations. Since
RORO vessels handle cargo approximately as a Fist-In-Last-Out (FILO) queue,
great care must be taken when assigning cargo. Cargo for later ports can end
up in front of cargo with an earlier discharge port. Such situations can result in
unnecessary cargo moves, called “shifts”, and increase the turnaround time for
vessels at the port [4].
Prior research assumes that all the cargo is available for stowage when the
ship arrives at the port. In real life, the dynamic arrival time of cargo creates a
trade-off between waiting for cargo to arrive, resulting in a revenue increase, and
incurring an increase in costs due to an increase in fuel consumption resulting
from the necessary speed increase, to reach the next port on time.
The contributions of this paper are twofold. First, we propose the extension
of the RORO stowage planning problem with dynamic arrival times and emis-
sion considerations (modeled in the objective function weighted against revenue
gains). Second, we propose a novel mixed-integer programming formulation for
the problem and present a heuristic approach to solve it.
The remainder of this paper is structured as follows. Section 2 presents
the prior research regarding stowage planning for RORO shipping. Section 3
presents a mathematical model formulation and performance. Section 4 presents
the heuristic framework for optimizing the stowage plan. Section 5 presents a
computational study analysis of the heuristic performance along with a sensi-
tivity analysis of the heuristic. We conclude on the findings and present future
work in Sect. 6.

2 Related Work

Initial research regarding the optimization of stowage plans for RORO shipping
utilized a lane layout. The deck would be divided into a set of strips; whereafter
the cargo would be stowed in a FILO queue. A mathematical model was devel-
oped to maximize revenue gained by stowing cargo in FILO lanes. The model
would define the lanes’ length, width, and height and simultaneously ensure
ship stability [12]. The mathematical model could not solve large instances, so
a heuristic approach was applied to large real-life instances. The authors also
studied optimizing the routing and scheduling of RORO vessels [13].
124 A. R. Main et al.

A different approach was later proposed, where the lane layout was replaced
by a grid, which allowed for sideways movement of cargo. A mathematical model
was proposed to assign positions of cargo on board. Several objectives were
utilized for finding good positioning, including grouping constraints of cargo
to minimize shifting. The solutions structure was evaluated by calculating the
shortest path for cargo to exit the vessel. The complexity of the model, however,
would limit its use in real-life instances [5]. Later a shortest path heuristic for
evaluating the stowage plan was presented [3].
Following the grid layout’s introduction, research focused on increasing the
performance of using MIP formulations for stowage planning and maximizing
revenue. The discretization of the vessel layout was optimized to reduce the
need for a very fine grid representation. Furthermore, refrigerated cargo and
hazardous cargo were considered during the stowage process. The need for shift-
ing of cargo was removed by computing the shortest path from a position to an
exit and utilizing inequality constraints to force a non-shift policy. The math-
ematical model also includes ship stability; however, the model was shown to
solve medium-sized problems well but not large instances [14].
An Adaptive Large Neighborhood Search was presented to minimize the num-
ber of cargo shifts. The study introduced a mathematical model to optimize
cargo assignment on board and showed that it was incapable of solving large
instances. This method showed that a ship could be stowed efficiently for each
deck by minimizing the number of shifts. The model utilizes Dijkstra’s algorithm
to evaluate how many shifts are generated by a given stowage plan [4].
The grid layout increases the computational complexity. A third approach to
the deck layout is to use slots. The slots represent specific positions on the deck,
and cargo can be assigned to these positions. To reduce the GHG emission by
shipping cargo, the assignment of cargo could be used to minimize the amount
of ballast water needed to stabilize the ship, thus reducing dead weight [8] and
thereby fuel consumption. The authors also conducted studies to estimate cargo
discharge times [9] and optimize the discharging process by dual cycling [10].
Drawing on the prior research within RORO stowage planning, the study
of this paper focused on the interconnection between maximizing revenue and
GHG emissions. A grid representation was utilized to describe the shifting and
cargo movement better. Ballast tanks and stability constraints were not included.
However, dynamic cargo arrival times were added to analyze the relationship
between time spent stowing the vessel, fuel consumption, and revenue.

3 Mathematical Model Formulation

The following mathematical model is an extension of the work done by [3]. The
original model has been modified to include the arrivals of each individual piece
of cargo. Furthermore, the objective function has been modified to be a weighted
function between fuel consumption and revenue. The weight α can be set to how
important it is for shippers to reduce GHG emissions vs. transporting cargo.
The function for fuel consumption is assumed to be non-linear with respect to
 The Dynamic RORO Stowage Planning Problem 125

sailing speed. The consumption of fuel has previously been described as cubic
[15]. For the mathematical model, the function is linearized by discretizing it
into W points, where B(v) is the fuel consumption, v is the speed, v ∗ is the
design speed, and B(v ∗ ) is the design consumption.

 v 3
B(v) = B(v ∗ ) (1)
v∗
A predefined route for port calls is assumed to be available, resulting in a
set of ports P . The ports are assumed to be separated into a loading region
P L and an unloading region P U . Each piece of cargo can be picked up at its
loading port PcL and unloaded at its unloading port PcU . In the original model,
each piece of cargo c ∈ C consisted of several vehicles of the same width, length,
and weight. However, in the following model, each piece of cargo is assumed to
have a single vehicle. This allows for a better description of the individual arrival
times of each cargo. The inter-arrival times of the cargo were assumed to follow
an exponential distribution.
The deck layout of the ship follows a grid representation of I rows and J
columns resulting in (i, j) pairs of locations. Each piece of cargo is modeled as
a rectangle. The positioning of a piece of cargo follows the cargo’s lower left
corner. Due to cargo width and height, a piece of cargo will span over several
grid points. The grid layout contains restricted areas where, e.g., pillars, ramps,
and walls are located.
If a piece of cargo is shifted, it is removed from the vessel and returned to its
original position. The model guarantees that the ship makes it in time for the
next port. The ship can, therefore, in order not to experience tardiness regulate
sailing speed.

Table 1. Sets, Parameters, and Decision Variables for the mathematical model

Sets
C Set of cargoes
P Set of ports
I Set of rows
J Set of columns
N Set of squares where cargo can be placed N ⊆ I × J
Nc Set of squares where the lower left corner of cargo c ∈ C can be
placed, Nc ⊆ N
Qijc The set of squares where placing cargo c ∈ C would cover position
(i, j)
Dijcd If cargo d ∈ C is in position (i, j). The set contains all squares that
would initiate a shift of cargo d if cargo c ∈ C uses these squares
(continued)
126 A. R. Main et al.

Table 1. (continued)

Sets
L
Dijcd Same as Dijcd , however, specifically for when cargo is being loaded
individually and sequentially
Bij Set of neighboring squares for position (i, j)
PL Set of loading ports, P L ⊆ P
PU Set of unloading ports, P U ⊆ P
PcL Set of loading ports p ∈ P L for cargo c ∈ C
PcU Set of unloading ports p ∈ P U for cargo c ∈ C
P LU Set of ports where shifting can occur P LU = P \{1, |P |}
PcLU Set of ports where shifting can occur for cargo
c ∈ C, P LU = P \{1, |P |}
CpR Set of cargo to be loaded or unloaded at port p ∈ P
CpN Set of cargo not to be loaded or unloaded at port p ∈ P
CpRL Set of cargo to be loaded at port p ∈ P including p = 1
W Set of discrete points for nonlinear fuel consumption discretization
EI , EJ Set of exit squares for row and column pair (I, J).
Parameters
α Parameter for choosing the importance level between fuel
consumption and revenue
Mdp Big M
Rc Number of vehicles in cargo c ∈ C
hc Handling time of cargo c ∈ C
Acp Arrival time of cargo c ∈ C at port p ∈ P
Tpmean Average time to sail between ports p, p + 1 ∈ P
Tpslow Time to sail between ports p, p + 1 ∈ P at the slowest vessel speed
possible
Tpf ast Time to sail between ports p, p + 1 ∈ P at the fastest vessel speed
possible
C rev Revenue gain for transporting a piece of cargo
C f uel Fuel cost per ton
Cwcon
Coefficient of fuel consumption at discrete point w ∈ W
time
Cpw Coefficient of time to sail at discrete point w ∈ W at port p ∈ P
(continued)
 The Dynamic RORO Stowage Planning Problem 127

Table 1. (continued)

Sets
Decision Variables
xijc 1 if the lower left corner of a vehicle from cargo c ∈ C is placed in
position (i, j), 0 otherwise
ycdp 1 if cargo c ∈ C is stowed before cargo d ∈ C in port p ∈ P L , 0
otherwise
δijcp 1 if cargo c is shifted at position (i, j) in port p
dijcp Supply or demand variable for cargo c at position (i, j) in port p. If
dijcp > 0, then position (i, j) is a supply square. If dijcp < 0, it is a
demand square. If dijcp = 0, it is a transit square
fijklcp Flow of cargo c from position (i, j) to position (k, l) at port p.
Bp Fuel consumption per time unit between port p and p + 1
STcp Start time of loading of cargo c ∈ C at port p ∈ PcL
T T Sp Time to sail between port p and p + 1 ensuring
Tpf ast ≤ T T Sp ≤ Tpslow
λwp Used to select the discrete point w ∈ W of the nonlinear fuel
consumption function between port p and p + 1
Deptp Departure time at port p ∈ P

The notation for the mathematical model is described in Table 1, and the
formulation is as follows.
  
min α C f uel Bp Tpmean − (1 − α) C rev xijc (2)
p∈P LU c∈C (i,j)∈Nc

Subject to:

xijc ≤ Rc ∀c ∈ C (3)
(i,j)∈Nc

xi j  c ≤ 1 ∀(i, j) ∈ N (4)
(i ,j  )∈Qijc

δijcp ≤ xijc ∀p ∈ P LU , c ∈ CpN , (i, j) ∈ Nc (5)

dE I E J cp ≤ Rc ∀c ∈ C, p ∈ PcL (6)
dijcp = −xijc ∀c ∈ C, p ∈ PcL , (i, j) ∈ Nc \{E , E }
I J
(7)
dE I E J cp ≥ −Rc ∀c ∈ C, p ∈ PcU (8)
dijcp = xijc ∀c ∈ C, p ∈ ∈ Nc \{E , E }
PcU , (i, j) I J
(9)
 
fijklcp − fklijcp = dijcp ∀p ∈ P , c ∈ CpR , (i, j) ∈ Nc
LU
(10)
(k,l)∈Bij (k,l)∈Bij
128 A. R. Main et al.
  
(fijk l cp + fk l ijcp ) ≤ Mdp (1 − xkld + δkldp )
c∈CpR (k ,l )∈D klcd (i,j)∈Bk j  (11)
∀p ∈ P LU
,d ∈ CpN , (k, l) ∈ Nd
STcp +hc −(1 − ycdp )Mcdp ≤ STdp
(12)
∀p ∈ P LU , c ∈ CpRL , d ∈ CpRL , c = d
ycdp + ydcp = 1 ∀p ∈ P LU , c ∈ C RL , d ∈ C RL , c = d (13)
STcp ≥ Acp − Mcp (1 − xijc ) ∀c ∈ C, p ∈ PcL , (i, j) ∈ Nc (14)
 
(fijk l cp + fk l ijcp ) ≤ Mdp (2 − xkld − ydcp )
(k l )∈Dklcd
L (i,j)∈Bk j  (15)
∀p ∈ P LU
,c ∈ CpRL , d ∈ CpRL , (k, l) ∈ Nd , c = d
Deptp ≥ STcp + hc ∀c ∈ C, p ∈ PcL (16)

Bp = Cwcon
λwp ∀p ∈ P LU (17)
w∈W

T T Sp = time
Cpw λwp ∀p ∈ P LU (18)
w∈W

λwp = 1 ∀p ∈ P LU (19)
w∈W

Deptp = Tpslow − T T Sp − 2δijcp hc ∀p ∈ P LU (20)
c∈C,(i,j)∈Nc

Deptp ≤ TpF ast ∀p ∈ P LU (21)

xijc ∈ {0, 1} ∀(i, j) ∈ Nc , c ∈ C (22)
δijcp ∈ {0, 1} ∀p ∈ P LU
, (i, j) ∈ Nc , c ∈ C (23)
dijcp ∈ R ∀p ∈ PcLU , c ∈ C, (i, j) ∈ Nc (24)
dijcp ∈ [−1, 1] ∀p ∈ PcLU , c ∈ C, (i, j) ∈ Nc \{E , E } I J
(25)
fijklcp ∈ R +
∀p ∈ P LU
, (i, j) ∈ Nc , (k, l) ∈ Nc , c ∈ CpR (26)
STcp ∈ R +
∀p ∈ PcLU , c ∈C (27)
ycdp ∈ {0, 1} ∀p ∈ P LU
, c ∈ C, d ∈ C (28)
Bp , T T Sp , Deptp ∈ R+ ∀p ∈ P LU (29)
λwp ∈ {0, 1} ∀w ∈ W, p ∈ P LU
(30)
The objective function (2) minimizes the weighted function between the cost
of fuel consumption vs. the revenue from shipping cargo. Constraints (3) ensure
that the cargo is stowed on board the vessel if chosen. Constraints (4) ensure that
each square can only be covered by one piece of cargo. Constraints (5) ensure that
cargo c ∈ C may only be shifted from position (i, j) ∈ I × J in port p ∈ P if it
is indeed in position (i, j). Constraints (6) set the supply of cargo c equal to the
vehicle numbers in c at the loading port. Constraints (7) set the demand for cargo
 The Dynamic RORO Stowage Planning Problem 129

in position (i, j), which is the chosen spot for cargo c. Constraints (8)–(9) state
the demand and supply for cargo c in port p similar to (6)–(7). Constraints 10
ensure that the outflow minus the inflow must be equal to the demand/supply in
the position (i, j) for each cargo c in port p. Constraint (11) ensure that if a piece
of cargo that is being loaded or unloaded in port p uses a position (k, l) from a
stationary piece of cargo d, it would result in the shifting of cargo d. Constraints
(12) ensure that the start time plus handling time for stowing cargo c in port p is
less than the starting time of cargo d if we stow cargo c before cargo d. Constraints
(13) ensure a sequence of stowing one piece of cargo c after another d in port p.
Constraints (14) ensure that the start time for stowing cargo c in port p must be
greater than or equal to the arrival time of cargo c if at all chosen to be stowed.
Constraints (15) ensure that if position (i, j) is chosen for cargo c, all other pieces
of cargo stowed after cannot cross the covered positions by cargo c. Constraints
(16) ensure that the departure time must be greater than the starting time of cargo
c plus the handling time of cargo c. Constraints (17) calculate the fuel consumption
for a given λwp value chosen. Constraints (18) calculate the time to sail between
port p and p + 1 given a λwp -value chosen. Constraints 19 ensure only one point
on the discretized non-linear fuel consumption curve can be picked. Constraints
(20) calculate the time to sail between ports p and p + 1. The departure time is
given to be between the time it takes to sail at the fastest and slowest speed. The
departure time is penalized if shifting of cargo occurs. Constraints (21) ensure that
the ship leaves at the latest time possible. The remaining constraints (22)–(30) are
for variable declarations.

4 A Heuristic Approach
To represent the deck layout without having an over-representation of the spaces
taken up by cargo, a grid size of 150 × 400 is needed [4]. Since the MIP formu-
lation is not able to solve the problem even for toy-size instances, we present a
heuristic approach. The heavy computations required to evaluate a single solu-
tion indicate the need to use strongly guided solution methods that do not
require many iterations. We propose a heuristic procedure where a pool of can-
didate solutions is iteratively improved by a truncated local search.
The pseudo-code for the heuristic is presented in Algorithm 1. The algorithm
encodes the solution as a sequence of cargo loads. A deterministic procedure then
decodes the sequence by assigning specific positions to the cargo. The initial
solution is constructed by a sequence where cargo is sorted in descending order
with regard to their leg length (line 1). The cargo positions are generated by
the deterministic decoding heuristic_placement (described later). An initial
pool of solutions is created in line 2. For the first iteration of the heuristic, the
pool only contains a single solution; however, later, it will represent the set of
current solutions. Out of all the solutions in the pool of current solutions, only
the best β solutions are selected for further processing (line 4). The next step
of the algorithm is to generate a new pool of solutions (lines 13–15). For each
of the best solutions (sols), η truncated local search procedures are executed.
Each of the found solutions is added to the new pool of current solutions.
130 A. R. Main et al.

Algorithm 1: Heuristic for DRSPP

Input: Problem instance π ∈ Π, Time limit τ 1 , Time limit τ 2 , Number of
nodes to expand β, Number of swaps κ, Layer limit ξ,
Pertubation indicator , Number of expansions per node η
Output: Stowage plan s ∈ S or ∅
1 s ← heuristic placement(init sequence())
2 pool ← {s}
1
3 while time limit (τ ) is not reached do
4 sols ← best solutions(pool, β)
5 if ξ iterations or τ 2 seconds have passed then
6 Eval(sols)
7 sols ← best(sols)
8 if No improvement after iterations then
9 sols ← double bridge(sols)
10 end
11 end
12 pool ← ∅
13 for r ∈ sols do
14 pool ← {LocalSearch(r, κ) ∀ i ∈ 1..η}
15 end
16 end
17 s ←best(Eval(pool))

The local search procedure performs κ swaps in the stowage sequence. From
the heuristic placement, it is known which piece of cargo blocks most other cargo.
The most blocking cargo is swapped to a new random position within the stowage
sequence. Using the heuristic placement again, the new sequence is evaluated. If
the solution is not better, the second worst piece of cargo is tried swapped. If
the solution is better, a new, better stowage sequence has been discovered, and
the most blocking cargoes have therefore changed. The local search continues
until a total of κ swaps have been performed.
During the search, all solutions are evaluated using an approximation. Given
a cargo stowage sequence, the placement heuristic assigns cargo to a specific
position iteratively. For each piece of cargo, a set of available positions is calcu-
lated (starting from the top-right corner of the cargo). A position is not available
if any of the grid positions, potentially occupied by the cargo, are already occu-
pied. This is calculated for the entire leg the cargo will be on board. Figure 1
shows an example set of available positions in green. Notice how only positions in
direct contact with the vessel structure or other placed cargo are available. Each
available position is evaluated based on its potential impact on the objective
function. Positions towards the bow and starboard are prioritized. Calculating
shifting costs is computationally heavy; we approximate the evaluation consid-
ering cargo c ∈ C blocking cargo d ∈ C iff PdL > PcL ∨ PdU < PcU looking from
cargo c towards the bow.
 The Dynamic RORO Stowage Planning Problem 131

At each ξ iterations (or at intervals of τ 2 seconds), a full evaluation of the

best solutions is performed. The full evaluation requires the calculation of all
the shortest-paths between the cargo and the exit. This is done using an A∗
algorithm (line 6). Furthermore, should the best solutions not be improved after
iterations, a double bridge move over the cargo sequence is performed as diver-
sification (line 9). The search continues until a time limit τ 1 is reached (line 3).
After termination, the best solution is returned (line 17).

Fig. 1. Creation of available positions for the heuristic placement procedure

5 Computational Study
The instances were available online [4] and were changed to fit the new problem
formulation in relation to cargo only containing a single vehicle and adding
arrival times of cargo. The inter-arrival times of the cargo were assumed to
follow an exponential distribution with μ = 60, and it was assumed that 20%
of cargo was available to stow immediately. The distances between ports were
generated by selecting routes between ports within the European RORO market.
The fuel costs were estimated through available bunker prices [16]. The revenue
generated for a piece of cargo was selected by analyzing pricing for spot cargo in
the RORO market. Due to a lack of better data, fuel consumption was estimated
based on the smallest vessel from LinerLib [1]. The grid resolutions used range
from 75 × 200 to 150 × 400 and are described as large instances. The number of
port calls ranges from 6–10, and the number of cargo units is app. 200–300.
The study was conducted on an Intel R
Xeon R
Gold6226R processor. The
mathematical model was solved for toy instances with a grid resolution of 10 × 10
up to 35 × 35. The number of cargoes was 77–193. The MIP was given 4 h to
perform precomputation of needed sets and model generation and given 1 h to
solve the problem. The problem was solved using Gurobi 9.1.0 and four processor
cores. It was not possible to generate solutions for the instances within 1 h for
grid resolutions as low as 10 × 10 and 77 pieces of cargo. The heuristic was tested
on several large problem instances. For each of these problem instances, several
different arrival patterns were generated. Furthermore, the algorithm was tested
several times for each instance with different seeds to evaluate the algorithm’s
performance. The heuristic was implemented in Julia 1.6.1 and proved capable
of generating stowage plans.
132 A. R. Main et al.

There are several parameters for the heuristic (η = 2, κ = 5, β = 2, ξ =

15, τ 2 = 20 s, = 4). These parameters were tuned manually to find a set that
performed well for large instances. Initially, the algorithm was tested with base-
line parameters to find the general behavior with respect to changes in α - values.
The algorithm was given half an hour (τ 1 ) to optimize the stowage plan.

Fig. 2. The average objective function plotted for different α-values along with the
95% confidence interval.

The initial analysis showed that increasing the α-value will result in less
cargo being stowed. This is intuitively supported by the nature of the objective
function. Increasing the α-value will emphasize fuel consumption (and thereby
GHG emissions) more than revenue. However, the variance increases as α goes to
1, as can be seen in Fig. 2. Instances generally differ regarding the number of port
calls and the different port sequences. The timing becomes more pronounced as
the cargo stowed does not dominate the objective function. If we look at the
amount of cargo stowed in Fig. 3 and the number of shifts occurring in the
stowage plan in Fig. 4, the number decreases with increasing α-values. This also
holds true for the variance. However, this is expected as the variance must follow
when less cargo is stowed.
We aimed to investigate how the heuristic behaved if the coefficients of the
objective function changed. The objective function was affected by the cargo’s
revenue, the fuel cost, and indirectly by the handling time of the cargo. The han-
dling time impacted the vessel’s departure time and, thereby, fuel consumption.
From the initial analysis, we can conclude that the algorithm is of most interest
if 0.00 ≤ α ≤ 0.30. The most cargo was stowed in this interval, so a sensitivity
analysis is conducted within this range.
 The Dynamic RORO Stowage Planning Problem 133

Fig. 3. The average number of stowed cargo plotted for different α-values along with
the 95% confidence interval.

Fig. 4. The average number of shifts incurred in the stowage plan plotted for different
α-values along with the 95% confidence interval.

An initial baseline was established for the sensitivity analysis. The problems
were solved for a handling time of 2 min per cargo, a revenue of 270 euros
per cargo, and a fuel consumption cost of 520 euros per ton. A baseline was
generated for each α-value. For the baseline, the mean and standard deviation
were calculated across the instances for how many units of cargo were stowed
on board. The sensitivity analysis used the same setup as the one used for
the computational study above. The first analysis was conducted where the
handling time was kept fixed, whereafter the other parameters varied slightly.
The percentage change in the number of cargo stowed changed significantly when
revenue and fuel costs fluctuated, as shown in Fig. 5. For an α-value of 0.0, there
is very little change in the amount of cargo stowed. However, as α increases, the
variance becomes significantly more noticeable. Lowering the revenue results in
134 A. R. Main et al.

a negative percentage change in the amount of cargo stowed, and an increase in

fuel costs has the same effect.

Fig. 5. The percentage change in the average number of cargoes stowed and the 95%
confidence intervals for fixed handling time.

Fixing the fuel costs at the baseline level will allow the handling time to
fluctuate. The heuristic was highly susceptible to increased handling time, as
seen in Fig. 6. An increase in handling time of a single minute resulted in a
significant drop in stowed cargo irrespective of the α-value compared to the
baseline. The susceptibility of the algorithm is most likely connected to the
stowage sequence and the number of cargoes stowed. If the stowage list is badly
ordered in relation to cargo arrival, a lot of time will be spent waiting for cargo.
This is only exacerbated by potential cargo shifts. The time penalty for each
shift is double the handling time. This results in the algorithm being penalized
heavily in relation to fuel consumption incurred. This also holds true for an α-
value of 0.0. No direct cost is incurred at this alpha value; however, shifts result
in time being spent not stowing cargo. Thereby less cargo could end up being
stowed as the ship is forced to leave port.
 The Dynamic RORO Stowage Planning Problem 135

Increasing fuel costs and handling time resulted in a significant decrease in

the amount of cargo stowed. The algorithm was very sensitive to the handling
time. In contrast, when fluctuating, the fuel costs changed very little, as can
be seen in Fig. 7. The algorithm displayed a high sensitivity to an increase in
α-values. Furthermore, when allowing handling time to fluctuate, the amount of
cargo stowed decreases with respect to the baseline.

Fig. 6. The percentage change in the average number of cargoes stowed and the 95%
confidence intervals for fixed fuel costs.
136 A. R. Main et al.

Fig. 7. The percentage change in the average number of cargoes stowed and the 95%
confidence intervals for fixed revenue.

6 Conclusion

This study of the dynamic RORO stowage planning problem focused on cargo
arriving at the port over time. The cargo had to be assigned a position on
board the vessel. Sailing between ports, the stowage plan should consider the
fuel consumption incurred from waiting for arriving cargo, and the vessel could
leave early to reduce costs. Furthermore, the number of cargo re-handling could
be minimized to reduce time spent stowing the ship. A MIP formulation was
presented; however, it could not solve even toy-sized instances of the problem,
so a heuristic was applied instead. The heuristic proved effective in generating
feasible stowage plans. The heuristic showed a clear sensitivity to changes in
coefficients for revenue and fuel consumption costs. However, the algorithm was
highly sensitive to the handling time of each piece of cargo on board the vessel.
Given how sensitive the solutions are to changes in the parameters, it might
be advisable to study different cost structures and to perform a deeper analysis
of alternative objectives as seen in [18]. Furthermore, an investigation into a
different problem representation could prove advantageous in relation to the
algorithm’s performance.
 The Dynamic RORO Stowage Planning Problem 137

References
1. Brouer, B.D., Alvarez, J.F., Plum, C.E.M., Pisinger, D., Sigurd, M.M.: A base
integer programming model and benchmark suite for liner-shipping network design.
Transp. Sci. 48(2), 281–312 (2014). https://doi.org/10.1287/trsc.2013.0471
2. Faber, J., et al.: Fourth IMO GHG study 2020. Technical report, International
Maritime Organization, London (2021)
3. Hansen, J.R., Fagerholt, K., Stålhane, M.: A shortest path heuristic for evaluat-
ing the quality of stowage plans in roll-on roll-off liner shipping. In: Bektaş, T.,
Coniglio, S., Martinez-Sykora, A., Voß, S. (eds.) ICCL 2017. LNCS, vol. 10572, pp.
351–365. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-68496-3 24
4. Hansen, J.R., Fagerholt, K., Stålhane, M., Rakke, J.G.: An adaptive large neigh-
borhood search heuristic for the planar storage location assignment problem: appli-
cation to stowage planning for Roll-on Roll-off ships. J. Heuristics 26(6), 885–912
(2020). https://doi.org/10.1007/s10732-020-09451-z
5. Hansen, J.R., Hukkelberg, I., Fagerholt, K., Stålhane, M., Rakke, J.G.: 2D-packing
with an application to stowage in roll-on roll-off liner shipping. In: Paias, A., Ruth-
mair, M., Voß, S. (eds.) ICCL 2016. LNCS, vol. 9855, pp. 35–49. Springer, Cham
(2016). https://doi.org/10.1007/978-3-319-44896-1 3
6. IEA: International Shipping. Technical report, International Energy Agency, Paris
(2022). https://www.iea.org/reports/international-shipping. Accessed 29 Apr 2023
7. Jia, B., Fagerholt, K.: Step-wise stowage planning of roll-on roll-off ships trans-
porting dangerous goods. Maritime Transp. Res. 2(June), 100029 (2021). https://
doi.org/10.1016/j.martra.2021.100029
8. Jia, B., Fagerholt, K., Reinhardt, L.B., Rytter, N.G.M.: Stowage planning with
optimal ballast water. In: Lalla-Ruiz, E., Mes, M., Voß, S. (eds.) ICCL 2020. LNCS,
vol. 12433, pp. 84–100. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-
59747-4 6
9. Jia, B., Rytter, N.G.M., Reinhardt, L.B., Haulot, G., Billesø, M.B.: Estimating
discharge time of cargo units – a case of Ro-Ro shipping. In: Paternina-Arboleda,
C., Voß, S. (eds.) ICCL 2019. LNCS, vol. 11756, pp. 122–135. Springer, Cham
(2019). https://doi.org/10.1007/978-3-030-31140-7 8
10. Jia, B., Tierney, K., Reinhardt, L.B., Pahl, J.: Optimal dual cycling operations in
roll-on roll-off terminals. Transp. Res. Part E: Logist. Transp. Rev. 159(February),
102646 (2022). https://doi.org/10.1016/j.tre.2022.102646
11. Mattfield, D.C.: The Management of Transshipment Terminals, 1st edn. Springer,
New York (2006). https://doi.org/10.1007/0-387-30854-7
12. Øvstebø, B.O., Hvattum, L.M., Fagerholt, K.: Optimization of stowage plans for
RoRo ships. Comput. Oper. Res. 38(10), 1425–1434 (2011). https://doi.org/10.
1016/j.cor.2011.01.004
13. Øvstebø, B.O., Hvattum, L.M., Fagerholt, K.: Routing and scheduling of RoRo
ships with stowage constraints. Transp. Res. Part C: Emerg. Technol. 19(6), 1225–
1242 (2011). https://doi.org/10.1016/j.trc.2011.02.001
14. Puisa, R.: Optimal stowage on Ro-Ro decks for efficiency and safety. J. Marine Eng.
Technol. 20(1), 17–33 (2018). https://doi.org/10.1080/20464177.2018.1516942
15. Reinhardt, L.B., Plum, C.E., Pisinger, D., Sigurd, M.M., Vial, G.T.: The liner
shipping berth scheduling problem with transit times. Transp. Res. Part E: Logist.
Transp. Rev. 86, 116–128 (2016). https://doi.org/10.1016/j.tre.2015.12.006
16. Ship & Bunker: EMEA Average (2023). https://shipandbunker.com/prices/av/
global/av-g20-global-20-ports-average#VLSFO. Accessed 29 Apr 2023
138 A. R. Main et al.

17. UNCTAD: REVIEW OF MARITIME TRANSPORT 2020. Technical report,

UNCTAD, New York, NY (2020)
18. Voß, S., Schwarze, S.: A note on alternative objectives for the blocks relocation
problem. In: Paternina-Arboleda, C., Voß, S. (eds.) ICCL 2019. LNCS, vol. 11756,
pp. 101–121. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-31140-7 7
 Allocation of Shore Side Electricity: The
Case of the Port of Hamburg

Jingjing Yu1,2(B) , Philip Cammin1 , and Stefan Voß1

1
Institute of Information Systems, University of Hamburg,
Von-Melle-Park 5, 20146 Hamburg, Germany
{philip.cammin,stefan.voss}@uni-hamburg.de
2
State Key Laboratory of Coastal and Offshore Engineering,
Faculty of Infrastructure Engineering, Dalian University of Technology, Dalian
116024, Liaoning, China
[email protected]

Abstract. This study addresses the problem of allocating shore side

electricity (SSE) supply points. We consider a group of terminals, for
which the overall environmental benefit is to be maximized. To this end,
an optimization model is developed that yields the optimal SSE sup-
ply point allocation plan. The proposed model solves the berth alloca-
tion problem with port call data. The novelty of this model lies in the
distinction of terminals, respecting the individual target terminal per
vessel. This supports the application on a group of terminals, even span-
ning multiple ports. Previous models only support such a use case by
requiring further analysis on the results of per-terminal applications. We
examine the results for various generated scenarios based on historical
port call data of one month for a container terminal group in the Port
of Hamburg, Germany. The results show that the best found allocation
plan in this study, enabling 2.54 GWh of SSE consumption, is slightly
better than the allocation plan published by the Port of Hamburg.

Keywords: Shore side electricity · container terminal · berth

allocation

1 Introduction
Shore side electricity (SSE) has been given great attention to significantly reduce
at-berth emissions from vessels. For instance, as of 2023, the Port of Los Angeles
invested in 79 SSE supply points and requires all container, cruise, and reefer
vessels to use SSE or an approved emission control technology during berthing
(POLA, 2023). By 2021, the average coverage rate of SSE at 11 major coastal
ports in China is about 71.5%, and vessels that have berthing durations longer
than three hours are required to use SSE (CAA, 2022). The European Com-
mission aims at reducing net greenhouse gas emissions by at least 55% by 2030
compared to 1990 levels; for instance, vessels larger than 5,000 gross tonnages
berthing at European ports for more than two hours would be required to use
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 139–153, 2023.
https://doi.org/10.1007/978-3-031-43612-3_8
140 J. Yu et al.

SSE as of January 2030 (see the “Fit for 55” package (EP, 2023)). In 2021,
the ports of Rotterdam, Hamburg, Bremen, Antwerp, and Haropa agreed to a
memorandum of understanding (MOU) to provide SSE to ultra-large container
vessels, 2028 at the latest (POA, 2021).
More specifically, the Port of Hamburg contribution manifests in the provi-
sion of eight SSE supply points for container vessels (HPA, 2022). From an eco-
nomic perspective, Williamsson et al. (2022) note that one of the obstacles to the
adoption of SSE is the huge initial investment in SSE supply points. For example,
the cost of constructing one SSE supply point at the Port of Hamburg is about
nine million EUR (HPM, 2019). The correct allocation of SSE supply points to
terminals is crucial to maximize the SSE consumption, and thus, to maximize
the environmental benefit with a limited investment. Consequently, allocation
plans should be scrutinized by using suitable methods; however, we deem the
existing methods having some drawbacks. Our literature review outlines that
most methods are lacking consideration of the temporal overlap of arriving ves-
sels (neglecting the berth allocation problem (BAP)); this potentially reduces
the accuracy concerning past assessment periods. Moreover, existing methods do
not consider terminal-assigned port calls for a terminal group with the shared
objective to maximize the overall environmental benefit; this requires separate
applications of methods for each port or terminal, as well as further analysis to
find the best allocation that maximizes the overall environmental benefit.
Consequently, this paper integrates the shared investment and shared goal by
terminals to maximize the overall environmental benefit by maximizing the SSE
consumption. Different scenarios are explored, which concern a variable number
of SSE supply points and SSE-enabled vessels. The objective of this paper is
(1) to optimize the allocation of SSE supply points among container terminals
under different scenarios with respect to an uncertain future, and (2) to provide
the environmental benefit of this allocation. The aim of this paper is to adapt
the methodological support for SSE adoption to fit a grouped terminal setting
and to investigate the current plan for the Port of Hamburg. Toward this aim,
an optimization model and a solution approach using a genetic algorithm (GA)
are implemented and the obtained allocations are discussed. To the best of our
knowledge, this paper is the first to focus on the shared investment and shared
goal of a terminal group. The case study concerns the three container terminals
operated by the Hamburger Hafen und Logistik AG (HHLA) at the Port of
Hamburg.
The remainder of this paper is structured as follows. Section 2 presents a review
of relevant studies. Section 3 then describes the optimization model. In Sect. 4, the
developed optimization model is applied in a case study followed by a presentation
of the results and discussions. At last, conclusions are summarized in Sect. 5.

2 Literature Review
Table 1 summarizes the most recent literature concerning the adoption of SSE in
maritime ports. The first research branch, assessment, primarily estimates the
 Allocation of Shore Side Electricity 141

environmental benefit of using SSE for ports for a certain time horizon. Herein,
we present the building blocks that are changed for different future scenarios. For
instance, Lathwal et al. (2021) estimate the environmental benefits of using SSE
with varying electricity generation structures, such as reducing the share of coal
generation and increasing the share of renewable generation in 2030 compared
to 2017. This leads to different emission factors. Most works in this branch
assume all vessels to be SSE-enabled and SSE-served. Contrary, Dai et al. (2019)
experiment with the number of SSE supply points for ten consecutive years. The
authors calculate the SSE consumption by multiplying the increasing adoption
rate of SSE with the total energy consumption of vessels during berthing. The
adoption rate is a simple fraction between the total number of berths and the
number of SSE supply points. Nguyen et al. (2021) feature two scenarios based
on two shipping routes, each of which has a different number of SSE-enabled
vessels. Gore et al. (2023) consider different numbers of SSE-enabled vessels
based on their calling frequency and make assumptions on the subsequently
required number of SSE supply points. In summary, the building blocks that
are changed for different future scenarios comprise varying emission factors, the
number of SSE-enabled vessels, the number of SSE supply points, as well as the
annual energy demand at berths.
The second branch, optimization, in a nutshell, yields the number of SSE
supply points to be installed to abate air emissions. Unlike the works in the
assessment branch, only the emission factors are subject to change to build dif-
ferent scenarios. Peng et al. (2019, 2021) optimize which specific berths of one
terminal should be equipped with SSE supply points to minimize the cost and
minimize the CO2 emissions, providing a trade-off analysis. Similarly, Vaishnav
et al. (2016) maximize inter alia the economic benefits. In contrast to the assess-
ment works, Peng et al. (2019, 2021) do not change the number of SSE-enabled
vessels for scenarios. To address the question of which berths should be equipped
with SSE supply points, it is useful to know which berths are actually used the
most. Insights to more detailed operational data may be helpful, such as those
data obtained by solving the berth allocation problem for vessels. Berths with
the longest utilization rate, and occupied by large vessels with a higher energy
demand are to be preferred. The work by Peng et al. (2021) solves the BAP.
The authors assume all vessels to be SSE-enabled, arrival times to be based on a
probability distribution, their berthing durations to be based on averaged histor-
ical vessel size, and the berth utilization to be based on the case study port. The
use of averaged historical data is commonly observed in strategic-oriented works
as can be seen in Table 1, and applies to both the optimization and assessment
branches. The takeaway is that the works in the optimization branch do not
consider a varying number of SSE-enabled vessels for different scenarios. This
cannot reflect current or near-future scenarios. Taken together, the references
foci are either concerned only with a single terminal or a group of ports. We
deem it necessary to address the rate of SSE-enabled vessels, as well as a new
scope of application, i.e., a group of terminals. We justify this grouping with the
case of the Port of Hamburg where HHLA operates three terminals whose berths
 142

Table 1. Summary of literature regarding the investment of SSE supply points in ports

Reference Objective Input data Scenario development

Winkel et al. (2016) Assess the economic and Seven typical European ◦ Annual energy demand at berths estimated
environmental benefit of SSE seaports’ one-year throughput for 2010 and 2020
J. Yu et al.

for each port • All vessels are SSE-enabled and SSE-served

Dai et al. (2019) Assess the reduced CO2 Number and size of vessels ◦ Accumulated annual adoption rate of SSE
emission and economic benefit calling at the Port of Shanghai supply points from 2018 to 2028
of using SSE for one port in one year • Berthing duration based on vessel size
• All vessels are SSE-enabled
Lathwal et al. (2021) Assess environmental and 21,937 port calls by 5,732 ◦ Emission factors based on changes in elec-
economic benefit of SSE for vessels at 12 ports in India tricity generation structure in local grid esti-
each port mated for 2017 and 2030
• All vessels are SSE-enabled and SSE-served
Martı́nez-López et al. Assess environmental benefit of Port calls during one year of • Berthing duration at ports based on vessel
(2021) SSE for each port two vessels with specified sizes size and type
and types in three selected • All vessels are SSE-enabled and SSE-served
European shipping routes
Nguyen et al. (2021) Assess environmental benefit of 1,122 port calls during one ◦ Number of SSE-enabled vessels based on
SSE for each port year at two container terminals shipping routes, e.g., only local vessels or
at the Port of Kaohsiung international vessels are SSE-enabled
• All vessels are SSE-enabled and SSE-served
Spengler and Tovar Assess environmental benefit of Port calls during one year of • External costs of emissions based on local
(2021) SSE for each port vessels to all ports in Spain city population
• All vessels are SSE-enabled and SSE-served
Stolz et al. (2021) Assess environmental benefit of 10,710 vessels at 714 European • Berthing duration and energy demand at
SSE for each port ports during one year berths based on vessel type
• All vessels are SSE-enabled and SSE-served
Herrero et al. (2022) Assess environmental benefit of 18,071 port calls during 11 ◦ Emission factor of CO2 based on changes in
SSE for each port years at the Port of Santander electricity generation
structure in local grids from 2011 to 2021
• Auxiliary engine powers based on vessel size
and type
• All vessels are SSE-enabled and SSE-served
(continued)
 Table 1. (continued)

Reference Objective Input data Scenario development

Gore et al. (2023) Assess cost and environmental 4,664 port calls during one ◦ Number of SSE-enabled vessels based on
benefit of SSE for each port year by 21 passenger vessels at calling frequency, e.g., top 5 or top 10 vessel
Dublin and Belfast ports callers are SSE-enabled
◦ Number of SSE supply points based on
average berthing duration of SSE-enabled
vessels and the assumption of temporal
overlaps among vessels
Vaishnav et al.(2016) Optimize the number of SSE 46,000 port calls during 18 ◦ External cost of emissions based on two dif-
supply points at each port and months by 3,300 vessels at 187 ferent emission impact assessment models
SSE-enabled vessels’ number to U.S. ports • Berthing duration based on vessel size and
maximize the environmental average rate of utilization of berths of each
and net benefits port
Peng et al. (2019) Optimize the allocation of SSE Container terminal in China • Time interval and daily number of arriving
supply points for one terminal with five berths and its vessels based on empirical distribution
to minimize the cost and CO2 throughput, service level, quay • Berthing duration based on vessel size and
emissions crane working efficiency average rate of utilization of berths in one ter-
minal
• All vessels are SSE-enabled
Peng et al. (2021) Optimize the allocation of SSE Bulk terminal in China with 10 ◦ External cost of CO2 based on environmen-
supply points at berths and berths and its throughput and tal tax rate
solve the BAP for one terminal service level • Time interval and daily number of arriving
to minimize costs and vessel vessels based on a probability distribution
emissions • Berthing duration based on vessel size and
average rate of utilization of berths in one ter-
minal
• All vessels are SSE-enabled
This paper Optimize the allocation of SSE 473 port calls during one ◦ Number of SSE supply points based on bud-
supply points and solve the month by 135 vessels at three get
BAP for a group of terminals container terminals at ◦ Number of SSE-enabled vessels based on
at one port Hamburg port vessel size
Allocation of Shore Side Electricity

Empty bullet points (◦) signal varying parameters leading to multiple scenarios, whereas filled bullet points (•) signal fixed parameters.
143
144 J. Yu et al.

are planned to be equipped with SSE supply points. Moreover, the use of vessels’
actual arrival times and berthing durations (port call data) is instrumental in
increasing the accuracy of the results, at least for an a posteriori analysis. For
future scenarios with uncertain port call data, an a priori analysis can be fed
with adapted/generated port call data, e.g., increasing the number of port calls.
Based on the literature review, we consider the following aspects to build
various scenarios: The number of SSE-enabled vessels and the number of SSE
supply points. Furthermore, we solve a BAP using a varying number of SSE-
enabled vessels together with the use of vessels’ actual arrival times and berthing
durations. The aim is to identify the number of berths that should be equipped
with SSE supply points, i.e., to optimize the allocation of SSE supply points
among the three terminals of HHLA in our case study.

3 Optimization Model
In this section, we develop an optimization model to allocate SSE supply points
to different terminals, and to solve the BAP, with the objective to maximize
the total SSE consumption of all vessels. Each vessel can only berth at its indi-
vidually defined target terminal. The set of terminals can be appended, imagin-
ably even spanning multiple ports. The BAP is simplified by using the historic
berthing start and departure time as parameters. Let us introduce the assump-
tions, input sets and parameters, as well as the decision variables:
Assumptions

• There is no distinction between the characteristics of berths, such as capacity,

length and depth.
• One berth can be equipped with exactly one SSE supply point providing
power for SSE-enabled vessels at this berth.

Sets

G set of terminals, G = {1, 2, ..., g, ..., G}

Bg set of berths at terminal g, Bg = {1, 2, ..., i, ..., Bg }
V set of vessels, V = {1, 2, ..., j, ..., V }
Vg set of vessels calling at target terminal g, Vg ⊂ V

Parameters

Pjaux auxiliary engines’ power of vessel j, kW

λaux auxiliary engines’ load factor during berthing
sj berthing start time of vessel j, h
dj departure time of vessel j, h
Wj SSE consumption of vessel j during berthing, kWh
φig 1 if berth i at terminal g is equipped with an SSE supply point, 0 otherwise
M a large positive number
 Allocation of Shore Side Electricity 145

Decision variables

yijg 1 if vessel j is served at berth i at terminal g, 0 otherwise

Objective function
  
max C SSE = Wj φig yijg (1)
g∈G i∈Bg j∈Vg

Subject to

yijg = 1 ∀j ∈ Vg , ∀g ∈ G (2)
i∈Bg

dj−1 yi(j−1)g ≤ sj yijg + M (1 − yijg ) ∀i ∈ Bg , ∀j ∈ Vg , ∀g ∈ G (3)

Wj = Pjaux λaux (dj − sj ) ∀j ∈ V (4)

The objective function (1) is to maximize the SSE consumption of SSE-

enabled vessels, C SSE . Constraints (2) ensure that each SSE-enabled vessel is
served at exactly one berth at its target terminal. Constraints (3) are used to
avoid the temporal overlap of two SSE-enabled vessels that are assigned to the
same berth. Constraints (4) gauge the SSE consumption of each SSE-enabled
vessel. Subsequently, based on the allocation of SSE supply points and the SSE
consumption, the saved emissions and the total emissions of vessels are calculated
by Eq. (5) and Eq. (6), where EFkf uel and EFkSSE are emission factors of emission
type k by using fuel oil and SSE, k ∈ K ={SOx , NOx , PM2.5 , PM10 , CO2 }.

 
SavedEmissions = C SSE EFkf uel − EFkSSE ∀k ∈ K (5)

T otalEmissions = Wj EFkf uel − SavedEmissions ∀k ∈ K (6)
j∈V

4 Case Study
The case study is based on the container terminals operated by HHLA in the
Port of Hamburg. Those terminals comprise the Container Terminal Altenwerder
(CTA), the Container Terminal Burchardkai (CTB), and the Container Terminal
Tollerort (CTT) with four, ten, and four berths, respectively. In the aforemen-
tioned MOU of five large ports in Europe (including the Port of Hamburg), it
is stated that “[Article 2] The Parties agree that all berths at container termi-
nals serving, on a regular basis, Ultra Large Container vessels (ULCV) with a
length overall (LOA) above 366 m, will be subject to this MoU [...] OPS [onshore
power supply] installations, as referred to in Article 2, must be operational by
1 January 2028 at the latest” (POA, 2021). The plan of the Port of Hamburg is
to allocate five SSE supply points to HHLA’s container terminals by 2024, i.e.,
CTA (1), CTB (3), and CTT (1) (HPA, 2022).
146 J. Yu et al.

4.1 Input Data

The dataset includes the entire port calls made at those terminals during Decem-
ber 2021. Those accumulate to 473 port calls by 135 unique vessels. The dataset
has been obtained on 27 January 2022 from HHLA (2020).
For each port call, the vessel name, vessel type (seagoing vessel, feeder),
scheduled berthing start time, departure time, and target terminal are available.
Based on the vessel name, the deadweight tonnage (DWT) and vessel length
in meters are collected from vesselfinder.com. The power of a vessel’s auxiliary
engines is estimated as P aux = 0.22 × P main , where P main is the power of
the main engines estimated by the DWT regression equation, i.e., P main =
2.4351×DW T 0.8673 (Wan et al., 2020). Based on some collected sample data, we
adjust the regression equation by multiplying a correction factor of 23 . The power
of the auxiliary engines is used to estimate the energy consumption of the vessel
during berthing by Eq. (4). The load factor, λaux , is highly correlated with the
port and its determination would constitute a separate project (see, e.g., efforts
by the Port of Los Angeles to express a nonlinear relationship between the load
factor and capacity, where container vessels are binned by twenty-foot equivalent
unit (TEU) (POLA, 2022)); here, we simplify and set λaux as 0.1. Emission
factors are used to assess the environmental benefit of using SSE (UBA, 2022)
compared to using auxiliary engines with low sulphur marine gas oil (LSMGO)
(POLA, 2022) and are shown in Table 2.

Table 2. Emission factors

Factor Unit SOx NOx PM2.5 PM10 CO2

Emission factors by LSMGO g/kWh 0.42 2.0 0.17 0.19 696
Emission factors by SSE g/kWh 0.178 0.355 0.001 0.008 375

4.2 Scenario Development

We use two aspects to develop a series of scenarios: the number of SSE sup-
ply points and the number of SSE-enabled vessels. Increasing the number of
SSE supply points enables the exploration of the trade-offs between investments
and environmental benefits. With a maximum of nine SSE supply points to be
allocated among the three terminals, 150 allocations exist (φig ≥ φ(i+1)g , ∀i ∈
Bg , ∀g ∈ G). This is sufficient to provide some solution space under and above
the planned allocation of five SSE supply points by the Port of Hamburg.
The number of SSE-enabled vessels is uncertain and beyond the decision
scope of the port. It is reasonable to suspect that larger container vessels are
prioritized for being equipped with SSE in the initial build or for retrofitting.
Hence, DWT classes are suitable to increase the number of SSE-enabled vessels
for different scenarios. The inspection of the extended port call data shows that
 Allocation of Shore Side Electricity 147

all vessels having a length larger than 366 m fall in the class of vessels ≥140,000
DWT. This class represents the type of vessels that the Port of Hamburg aims
at providing SSE first. In total, three DWT classes are defined to provide some
solution space for future developments, i.e., ≥200,000 DWT, ≥140,000 DWT,
and ≥100,000 DWT. The vessels in the three DWT classes are subsequently
treated as SSE-enabled vessels. Figure 1 provides the distribution of port calls
at each terminal in three DWT-intervals: [100,000, 140,000) DWT, [140,000,
200,000) DWT, and [200,000, 250,000) DWT.

Fig. 1. Frequency of port call DWT-intervals by terminal

Table 3. Parameter settings of GA

Parameter Values
Maximum number of 300
generations
Population 100, 120
Crossover probability 0.8
Mutation probability 0.01
Crossover function scattered, intermediate
Mutation function adaptive feasible
148 J. Yu et al.

4.3 Application

The presented optimization model, presented in Sect. 3, is solved using the MAT-
LAB GA optimization toolbox. The performance of the GA is largely influ-
enced by its parameters. The parameter settings of the GA are listed in Table 3,
which are based on initial experiments with respect to the scenarios developed
in Sect. 4.2.
Based on the 150 allocations and the three DWT classes (i.e., ≥200,000 DWT,
≥140,000 DWT, and ≥100,000 DWT) described in Sect. 4.2, two population
values, two crossover functions, and three GA execution repetitions to increase
the probability to find better solutions, there are 5,400 solutions obtained.

Fig. 2. SSE consumption of best solutions under different scenarios

4.4 Results and Discussion

For each scenario, the best solution is selected and shown in Fig. 2. The objective
value represents the SSE consumption of SSE-enabled vessels. Table 4 presents
the two best solutions for each scenario in more detail. Moreover, the best solu-
tion that adheres to the MOU-based constraint is shown. The MOU text provides
some room for interpretation concerning when a berth regularly serves vessels in
class ≥140,000 DWT: “The Parties agree that all berths at container terminals
serving, on a regular basis, Ultra Large Container vessels (ULCV) with a length
overall (LOA) above 366 m [140,000 DWT], will be subject to this MoU” (POA,
2021). The MOU could imply that each terminal should be allocated at least one
SSE supply point (MOU-based constraint), to be able to serve vessels of class
≥140,000 DWT at all terminals.
For the class ≥200,000 DWT, three supply points are sufficient to meet the
energy demand of all SSE-enabled vessels. The best solution is to allocate two
 Table 4. Allocation of SSE supply points for different scenarios

Scenario SSE allocation SSE SOx (t) NOx (t) PM2.5 (t) PM10 (t) CO2 (kt)
Class No. CTA CTB CTT (GWh) T S % T S % T S % T S % T S %
≥0 0 0 0 0 n/a 1.97 n/a n/a 9.39 n/a n/a 0.80 n/a n/a 0.90 n/a n/a 3.27 n/a n/a
≥200 1 0 1 0 0.50 1.85 0.12 6.13 8.57 0.82 8.74 0.72 0.08 10.50 0.81 0.09 10.13 3.11 0.16 4.90
≥200 1 0 0 1 0.33 1.89 0.08 4.11 8.84 0.55 5.86 0.75 0.06 7.05 0.83 0.06 6.80 3.16 0.11 3.29
≥200 2 0 1 1 0.83 1.77 0.20 10.24 8.02 1.37 14.60 0.66 0.14 17.55 0.74 0.15 16.92 3.00 0.27 8.19
≥200 2 0 2 0 0.62 1.82 0.15 7.65 8.36 1.02 10.91 0.70 0.11 13.12 0.78 0.11 12.65 3.07 0.20 6.12
≥200 3 0 2 1 0.96 1.74 0.23 11.76 7.81 1.57 16.77 0.64 0.16 20.16 0.72 0.17 19.44 2.96 0.31 9.41
≥200 3 0 1 2 0.83 1.77 0.20 10.24 8.02 1.37 14.60 0.66 0.14 17.55 0.74 0.15 16.92 3.00 0.27 8.19
≥200 3 1 1 1 0.83 1.77 0.20 10.24 8.02 1.37 14.60 0.66 0.14 17.55 0.74 0.15 16.92 3.00 0.27 8.19
≥140 1 0 1 0 0.85 1.76 0.21 10.42 7.99 1.39 14.85 0.66 0.14 17.85 0.74 0.15 17.21 2.99 0.27 8.33
≥140 1 0 0 1 0.71 1.80 0.17 8.75 8.21 1.17 12.48 0.68 0.12 15.00 0.77 0.13 14.47 3.04 0.23 7.00
≥140 2 0 1 1 1.56 1.59 0.38 19.17 6.82 2.57 27.33 0.54 0.26 32.86 0.61 0.28 31.68 2.77 0.50 15.32
≥140 2 0 2 0 1.54 1.60 0.37 18.89 6.86 2.53 26.93 0.54 0.26 32.38 0.62 0.28 31.22 2.77 0.49 15.10
≥140 3 0 2 1 2.26 1.42 0.55 27.77 5.67 3.72 39.60 0.42 0.38 47.60 0.48 0.41 45.90 2.54 0.73 22.20
≥140 3 0 3 0 1.74 1.55 0.42 21.38 6.52 2.86 30.49 0.51 0.29 36.65 0.58 0.32 35.34 2.71 0.56 17.09
≥140 3 1 1 1 1.58 1.59 0.38 19.37 6.79 2.59 27.61 0.54 0.27 33.20 0.61 0.29 32.01 2.76 0.51 15.48
≥140 4 0 3 1 2.52 1.36 0.61 30.94 5.25 4.14 44.11 0.38 0.43 53.03 0.44 0.46 51.13 2.46 0.81 24.73
≥140 4 0 2 2 2.28 1.42 0.55 27.99 5.64 3.75 39.91 0.42 0.38 47.98 0.48 0.41 46.26 2.54 0.73 22.38
≥140 4 1 2 1 2.28 1.42 0.55 27.97 5.64 3.74 39.88 0.42 0.38 47.94 0.48 0.41 46.23 2.54 0.73 22.36
≥140 5 0 4 1 2.54 1.35 0.61 31.24 5.21 4.18 44.54 0.37 0.43 53.55 0.43 0.46 51.63 2.45 0.82 24.97
≥140 5 1 3 1 2.53 1.36 0.61 31.11 5.22 4.16 44.36 0.37 0.43 53.33 0.44 0.46 51.42 2.45 0.81 24.87
≥140 6 0 4 2 2.56 1.35 0.62 31.46 5.18 4.21 44.85 0.37 0.43 53.92 0.43 0.47 51.99 2.44 0.82 25.15
≥140 6 1 4 1 2.56 1.35 0.62 31.44 5.18 4.21 44.82 0.37 0.43 53.89 0.43 0.47 51.96 2.45 0.82 25.13
≥140 7 1 4 2 2.58 1.35 0.62 31.65 5.15 4.24 45.13 0.37 0.44 54.26 0.43 0.47 52.32 2.44 0.83 25.31
≥140 7 0 4 3 2.56 1.35 0.62 31.46 5.18 4.21 44.85 0.37 0.43 53.92 0.43 0.47 51.99 2.44 0.82 25.15
≥100 1 0 1 0 0.85 1.76 0.21 10.48 7.98 1.40 14.94 0.66 0.14 17.96 0.74 0.16 17.32 2.99 0.27 8.38
≥100 1 0 0 1 0.71 1.80 0.17 8.75 8.21 1.17 12.48 0.68 0.12 15.00 0.77 0.13 14.47 3.04 0.23 7.00
≥100 2 0 1 1 1.56 1.59 0.38 19.17 6.82 2.57 27.33 0.54 0.26 32.86 0.61 0.28 31.68 2.77 0.50 15.32
≥100 2 0 2 0 1.55 1.59 0.37 19.02 6.84 2.55 27.12 0.54 0.26 32.60 0.61 0.28 31.43 2.77 0.50 15.21
≥100 3 0 2 1 2.20 1.44 0.53 27.03 5.77 3.62 38.54 0.43 0.37 46.33 0.50 0.40 44.68 2.56 0.71 21.61
≥100 3 1 2 0 1.85 1.52 0.45 22.78 6.34 3.05 32.49 0.49 0.31 39.05 0.56 0.34 37.65 2.67 0.59 18.21
≥100 4 0 3 1 2.55 1.35 0.62 31.33 5.19 4.19 44.68 0.37 0.43 53.71 0.43 0.46 51.79 2.45 0.82 25.05
≥100 4 1 2 1 2.37 1.39 0.57 29.14 5.49 3.90 41.56 0.40 0.40 49.96 0.46 0.43 48.17 2.51 0.76 23.30
≥100 5 1 3 1 2.78 1.30 0.67 34.22 4.81 4.58 48.79 0.33 0.47 58.65 0.39 0.51 56.55 2.37 0.89 27.36
≥100 5 2 2 1 2.66 1.32 0.64 32.74 5.00 4.38 46.69 0.35 0.45 56.12 0.41 0.48 54.11 2.41 0.86 26.18
≥100 6 2 3 1 3.00 1.24 0.73 36.91 4.45 4.94 52.64 0.29 0.51 63.28 0.35 0.55 61.01 2.30 0.96 29.51
Allocation of Shore Side Electricity

≥100 6 1 4 1 2.90 1.27 0.70 35.59 4.62 4.76 50.75 0.31 0.49 61.01 0.37 0.53 58.83 2.34 0.93 28.46
≥100 7 3 3 1 3.20 1.20 0.77 39.28 4.13 5.26 56.01 0.26 0.54 67.33 0.31 0.58 64.93 2.24 1.03 31.41
≥100 7 2 4 1 3.13 1.21 0.76 38.52 4.23 5.15 54.92 0.27 0.53 66.03 0.33 0.57 63.66 2.26 1.01 30.80
• Total amount of emissions in the current scenario (T), i.e., emissions by all port calls ≥0k DWT. Saved emissions in the current scenario, absolute (S), percent (%)
compared to the baseline scenario.
149

• The baseline scenario requires vessels to use LSMGO only (first row). • The SSE consumption is the objective value (SSE).
150 J. Yu et al.

SSE supply points to CTB and one to CTT, i.e., CTA (0), CTB (2), CTT (1),
leading to 0.96 GWh of SSE consumption. Generally, this allocation correlates
to the distribution of port call DWT-classes at each terminal as shown in Fig. 1.
As mentioned before, the class ≥140,000 DWT covers the prioritized vessels
based on the MOU. Based on Fig. 2, four supply points appear to be sufficient,
and Table 4 confirms a marginal performance difference between the best allo-
cations using four and five SSE supply points. The best allocation using four is
CTA (0), CTB (3), CTT (1), 2.52 GWh, as well as CTA (0), CTB (4), CTT (1),
2.54 GWh, using five. This best allocation using five supply points is slightly bet-
ter compared to the allocation planned by the Port of Hamburg, i.e., CTA (1),
CTB (3), CTT (1) with 2.53 GWh of SSE consumption.
Concerning the class ≥100,000 DWT, up to seven SSE supply points can be
exhausted, leading to 3.20 GWh of SSE consumption. Notably, in this class, the
allocation CTA (1), CTB (3), CTT (1), leading to 2.78 GWh, as planned by the
Port of Hamburg, is the best for allocating five supply points.
Negative implications for the MOU-based constraint are exhibited, only if
less than five supply points are allocated. More concretely, for the class ≥200,000
DWT and allocating three supply points, the only feasible allocation is CTA (1),
CTB (1), CTT (1) that decreases SSE consumption by 0.13 GWh compared to
the best allocation, i.e., CTA (0), CTB (2), CTT (1). For the classes ≥140,000
DWT and ≥100,000 DWT, more significant implications are observed with three
and four supply points. Regarding ≥140,000 DWT, for instance, the SSE con-
sumption is reduced by 0.68 GWh and 0.24 GWh.
In summary, Fig. 2 shows SSE supply points and SSE-enabled vessels to cor-
relate positively with the consumed SSE until the energy demand is fully met.
The energy demand by port calls in class ≥140,000 DWT is almost satisfied
by four supply points. Allocating more than four supply points yields a much
lower cost-effectiveness, e.g., an increase from four to five supply points leads
to only 0.02 GWh more SSE consumption, while an increase from three to four
yields 0.26 GWh. However, in the advent that class ≥100,000 DWT vessels are
SSE-enabled, further increasing the supply points to seven is beneficial.
Let us describe the limitations of this study. First, the results are limited by
the data set of one month. Future studies could include annual data sets. Sec-
ond, this study assumes vessels to use SSE during the entire berthing duration.
However, only a partial consumption may take place with respect to time of
usage energy pricing (Yu et al., 2022), or insufficient power supply due to energy
production or infrastructure limitations. The third limitation is the heuristic
approach that cannot guarantee to find the best solution for a BAP that could
further increase the SSE consumption and therefore change the ranking of the
presented allocation plans. Fourth, the BAP in this paper is simplified by using
the historic scheduled berthing start and departure time of our case study termi-
nals that have no SSE supply points yet. If the port had SSE supply points, the
times might have been adjusted in favor of SSE consumption. Fifth, the auxiliary
engines’ power and load factor are based on simplified methods where obtaining
precise values requires a case study per port. Moreover, the emission factors could
 Allocation of Shore Side Electricity 151

be implemented to change over the assessed time period. At last, the potential
use of scrubber and heavy fuel oil is neglected, leading to uncertainty of the
presented SOx emissions. At the end of 2020, about 14% of the container vessel
world fleet had been equipped with scrubbers (ICCT, 2021). The SSE imple-
mentation performance depends on the SSE allocation plan and other factors
such as the BAP plan, electricity pricing, incentives, and the number of SSE-
enabled vessels. An emissions inventory, such as presented in Table 4, thus can
only be regarded a composite index, a proxy measure for the SSE implementa-
tion performance. However, emissions inventory methodologies and key figures
can be standardized (Cammin and Voß, 2021), and subsequently be utilized to
establish sensible goals, i.e., specific, measurable, achievable, relevant, and time-
bound (SMART) goals (Drucker, 1954). A component of a SMART goal could
be the saved emissions relative to the baseline emissions, as presented in Table 4.
So far, SMART goals incorporating such quantifications of emissions have not
been declared in the MOU. Thus, rethinking the design of goals could benefit
future evaluations of SSE implementations.

5 Conclusions

In this study, we addressed the SSE supply point allocation problem. More
specifically, we implemented an optimization model for a terminal group with
the shared objective to maximize the environmental benefit by using SSE while
solving the BAP. We applied our model to a case study concerning the largest
container terminal group in the Port of Hamburg under a series of scenarios with
varying numbers of SSE supply points (from one to nine) and SSE-enabled ves-
sels (classes ≥200,000 DWT, ≥140,000 DWT, and ≥100,000 DWT). Based on
the results, we discussed the trade-offs between the environmental benefit and
the allocation of SSE supply points. For instance, the energy demand of vessels
in classes ≥200,000 DWT, ≥140,000 DWT, and ≥100,000 DWT are almost sat-
isfied by three, four, and seven supply points, respectively. With a total number
of five supply points, as planned by the Port of Hamburg, the best plan enables
0.96 GWh, 2.54 GWh, and 2.78 GWh of SSE consumption, for vessels in class
≥200,000 DWT, ≥140,000 DWT, and ≥100,000 DWT, respectively. We believe
our work can spark discussions revolving the intelligent use of funds, concerning
terminal groups, and future port cooperations. Future studies could extend this
study by considering the entirety of terminals of the ports of Rotterdam, Ham-
burg, Bremen, Antwerp, and Haropa that are subject to the MOU to promote
SSE, and to expose the implications to local air quality.

References
CAA. Evaluation of Air and Climate Synergies of Typical Chinese Ports. Clean Air Asia
(2022). http://allaboutair.cn/uploads/soft/230308/1-23030Q34007.pdf. Accessed 15
Apr 2023
152 J. Yu et al.

Cammin, P., Voß, S.: Towards smart maritime port emissions monitoring: a platform
for enhanced transparency. In: Chandra Kruse, L., Seidel, S., Hausvik, G.I. (eds.)
DESRIST 2021. LNCS, vol. 12807, pp. 71–76. Springer, Cham (2021). https://doi.
org/10.1007/978-3-030-82405-1 9
Dai, L., Hao, H., Wang, Z., Shi, Y., Ding, W.: An environmental and techno-economic
analysis of shore side electricity. Transp. Res. Part D: Transp. Environ. 75, 223–235
(2019). https://doi.org/10.1016/j.trd.2019.09.002
Drucker, P.: The Practice of Management. Harper, New York (1954)
EP: Sustainable maritime fuels ‘Fit for 55’ package: the FuelEU Maritime
proposal (2023). https://www.europarl.europa.eu/RegData/etudes/BRIE/2021/
698808/EPRS BRI(2021)698808 EN.pdf
Gore, K., Rigot-Müller, P., Coughlan, J.: Cost-benefit assessment of shore side elec-
tricity: an Irish perspective. J. Environ. Manage. 326, 116755 (2023). https://doi.
org/10.1016/j.jenvman.2022.116755
Herrero, A., Piris, A.O., Diaz-Ruiz-Navamuel, E., Gutierrez, M.A., Lopez-Diaz, A.-I.:
Influence of the implantation of the onshore power supply (OPS) system in Spanish
medium-sized ports on the reduction in CO2 emissions: the case of the Port of
Santander (Spain). J. Marine Sci. Eng. 10(10), 1446 (2022). https://doi.org/10.3390/
jmse10101446
HHLA. Report vessel operations: Sailing list. Hamburger Hafen und Logistik AG
(2022). https://coast.hhla.de/report?id=Standard-Report-Segelliste. Accessed 10
Feb 2023
HPA. Plugging in to eliminate emissions. Hamburg Port Authority (2022). https://
www.hafen-hamburg.de/en/port-of-hamburg-magazine/hafen-und-klima/stecker-
rein-emissionen-aus/. Accessed 10 Feb 2023
HPM. Cleaner air and climate protection in Port of Hamburg - New shore-based power
units for cruise liners and containerships. Hamburg Port Marketing (2019). https://
www.hafen-hamburg.de/en/press/news/cleaner-air-and-climateprotection-in-port-
of-hamburg-new-shore-based-power-units-for-cruiseliners-and-containerships-
36499/. Accessed 10 Feb 2023
ICCT. Global scrubber washwater discharges under IMO’s 2020 fuel sulfur limit. Inter-
national Council on Clean Transportation (2021). https://theicct.org/sites/default/
files/publications/scrubber-discharges-Apr2021.pdf. Accessed 10 Mar 2023
Lathwal, P., Vaishnav, P., Morgan, M.G.: Environmental and health consequences of
shore power for vessels calling at major ports in India. Environ. Res. Lett. 16(6),
064042 (2021). https://doi.org/10.1088/1748-9326/abfd5b
Martı́nez-López, A., Romero, A., Orosa, J.A.: Assessment of cold ironing and LNG as
mitigation tools of short sea shipping emissions in port: a Spanish case study. Appl.
Sci. 11(5), 2050 (2021). https://doi.org/10.3390/app11052050
Nguyen, D.-H., et al.: Reduction of NOx and SO2 emissions by shore power adoption.
Aerosol Air Qual. Res. 21(7), 210100 (2021). https://doi.org/10.4209/aaqr.210100.
Peng, Y., Dong, M., Li, X., Liu, H., Wang, W.: Cooperative optimization of shore
power allocation and berth allocation: a balance between cost and environmental
benefit. J. Clean. Prod. 279, 123816 (2021). https://doi.org/10.1016/j.jclepro.2020.
123816
Peng, Y., Li, X., Wang, W., Wei, Z., Bing, X., Song, X.: A method for determining
the allocation strategy of on-shore power supply from a green container terminal
perspective. Ocean Coast. Manag. 167, 158–175 (2019). https://doi.org/10.1016/j.
ocecoaman.2018.10.007
 Allocation of Shore Side Electricity 153

POA. Memorandum of understanding onshore power ambitions for container terminals

in port. Port of Antwerp, Port of Bremerhaven, Port of Hamburg, Haropa Port, Port
of Rotterdam (2021). https://www.portofrotterdam.com/sites/default/files/2021-
07/mou ops ports of antwerp-bremerhaven-hamburgharopa port-rotterdam -
final 20210618 003.pdf. Accessed 01 Mar 2023
POLA. Port of Los Angeles inventory of air emissions 2021. Port of Los
Angeles (2022). https://kentico.portoflosangeles.org/getmedia/f26839cd-54cd-4da9-
92b7-a34094ee75a8/2021 Air Emissions Inventory. Accessed 15 Feb 2023
POLA. Alternative maritime power (AMP). Port of Los Angeles (2023). https://www.
portoflosangeles.org/environment/air-quality/alternative-maritimepower-(amp).
Accessed 15 Apr 2023
Spengler, T., Tovar, B.: Potential of cold-ironing for the reduction of externalities from
in-port shipping emissions: the state-owned Spanish port system case. J. Environ.
Manage. 279, 111807 (2021). https://doi.org/10.1016/j.jenvman.2020.111807
Stolz, B., Held, M., Georges, G., Boulouchos, K.: The CO2 reduction potential of
shore-side electricity in Europe. Appl. Energy 285, 116425 (2021). https://doi.org/
10.1016/j.apenergy.2020.116425
UBA. Specific emission factors for electricity mix in Germany. Umweltbun-
desamt (2022). https://www.umweltbundesamt.de/themen/luft/emissionen-
vonluftschadstoffen/spezifische-emissionsfaktoren-fuer-den-deutschen. Accessed 01
Mar 2023
Vaishnav, P., Fischbeck, P.S., Morgan, M.G., Corbett, J.J.: Shore power for vessels
calling at U.S. ports: benefits and costs. Environ. Sci. Technol. 50(3), 1102–1110
(2016). https://doi.org/10.1021/acs.est.5b04860.
Wan, Z., Ji, S., Liu, Y., Zhang, Q., Chen, J., Wang, Q.: Shipping emission inventories
in china’s Bohai bay, Yangtze river delta, and pearl river delta in 2018. Mar. Pollut.
Bull. 151, 110882 (2020). https://doi.org/10.1016/j.marpolbul.2019.110882
Williamsson, J., Costa, N., Santén, V., Rogerson, S.: Barriers and drivers to the imple-
mentation of onshore power supply–a literature review. Sustainability 14(10), 6072
(2022). https://doi.org/10.3390/su14106072
Winkel, R., Weddige, U., Johnsen, D., Hoen, V., Papaefthimiou, S.: Shore side elec-
tricity in Europe: potential and environmental benefits. Energy Policy 88, 584–593
(2016). https://doi.org/10.1016/j.enpol.2015.07.013
Yu, J., Voß, S., Song, X.: Multi-objective optimization of daily use of shore side electric-
ity integrated with quayside operation. J. Clean. Prod. 351, 131406 (2022). https://
doi.org/10.1016/j.jclepro.2022.131406
 Stockyard Storage Space Allocation
in Dry Bulk Terminals Considering Mist
Cannons and Energy Expenditure

Xinyu Tang(B) and Frederik Schulte

Delft University of Technology, 2600AA Delft, The Netherlands

{X.Tang-3,F.Schulte}@tudelft.nl

Abstract. Storage space management in bulk terminals has become an

important focus for research and practical operation due to the increasing
demand for bulk cargo and limited storage space in stockyards. The study
of storage space management in dry bulk terminals is less thorough and
comprehensive, and the existing research investigates the storage space
allocation problem with other operational problems like berth allocation
problems, but little environmental consideration has been incorporated.
We investigate the storage space allocation problem with the considera-
tion of stacker-reclaimer assignment and mist cannon operation to deal
with the dust generated during material stacking. A mixed integer pro-
gramming model has been established with the aim of minimizing energy
consumption to reflect the pursuit of the growing emphasis on climate-
neutral operations and sustainability. We test the effectiveness of the
model by conducting computational experiments. We use the commer-
cial solver CPLEX to obtain the optimal solutions for most of the test
instances. Useful managerial insights extracted from the computational
results may serve as a reference for storage space management in dry
bulk terminals.

Keywords: Storage space allocation · Dry bulk terminal · Mist

cannon operations · Stacker-reclaimer operations · Energy consumption

1 Introduction

Stockyard in dry bulk terminal acts as a centralized storage area for bulk mate-
rials, and also functions as an essential buffer for differences between incoming
and outgoing streams of bulk materials [16]. Under the influence of the increas-
ing demand for bulk cargo and the trend towards larger vessels, storage space
in the stockyard tends to become a scarce resource, leading terminal operators
to focus on stockyard storage space management. Utilizing the storage space
reasonably can help improve the operational efficiency of the cargo transport
chain. Therefore, it is necessary to study the storage space allocation problem
in dry bulk terminals.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 154–169, 2023.
https://doi.org/10.1007/978-3-031-43612-3_9
 Stockyard Space Allocation with Mist Cannons and Energy Expenditure 155

Apart from the operation efficiency, the port now needs to take responsibility
for environmental concerns driven by the growing emphasis on a cleaner environ-
ment, neutral climate, and sustainability. A large amount of dust is generated
during the operation in the dry bulk stockyard, such as stacking materials into
the empty fields of stock pads. Mist cannon (also known as dust cannon, or
dust suppression cannon) is a widely used facility to control dust. The fogging
process involves the action of fog nozzles which nebulize water into very small
micro-droplets of water under pressure. The fog drives airborne dust particles
to the ground and wets the surface to prevent fugitive dust particles. When
water is combined with dust in the air, due to the adhesion of the surface of
water molecules, it will be combined with the dust, and the effect of gravity will
drop after condensing, to achieve the purpose of dust suppression. Materials are
stacked into stock pads accompanied by the water spray of the mist cannon. We
consider mist cannon operation in the space allocation process, which is rarely
addressed in the existing literature.
The stockyard is a complex logistics system that consists of components
of different operation machines like stacker-reclaimers and mist cannons. We
study Storage Space Allocation Problem (SSAP) in the stockyard of dry bulk
terminals, which together considers the assignment of stacker-reclaimers and
mist cannons for incoming materials. In this study, we consider import dry bulk
terminals in which cargo flows from the berth area to the stockyard area. We
aim to determine the specific storage locations for incoming materials with the
objective of minimizing energy consumption. The contributions of the paper
include the following: (1) We first model the operations of mist cannons to control
the dust and study the storage space allocation, stacker-reclaimer assignment
and mist cannon assignment in an integrated manner. (2) We incorporate energy
consumption into the optimisation objective, corresponding to the requirements
of sustainable operations.
The rest of the paper is organized as follows. Section 2 reviews relevant papers
in the literature. Section 3 describes the storage space allocation problem fol-
lowed by mathematical formulation in Sect. 4. Section 5 shows the preliminary
results from computational experiments to test the effectiveness of the mathe-
matical model. Finally, conclusions are drawn in Sect. 6.

2 Literature Review

Storage space management in dry bulk terminal stockyards has received far
less attention than that in container terminal yards. Storage space allocation
problem (SSAP) is usually investigated with other operational problems (such as
berth allocation problem (BAP), train scheduling, etc.) in an integrated manner.
Table 1 lists the related work on the SSAP problem in dry bulk terminals.
Ouhaman et al. [13] studied the SSAP in an export dry bulk terminal. The
authors formulated the problem as a mixed-integer linear programming (MILP)
problem and proposed a heuristic method to solve large-scale data sets. They
limited the problem to the material flow from the production plant to storage
156 X. Tang and F. Schulte

Table 1. Related literature on storage space allocation problem in dry bulk terminals

Literature Publication year Problem studied Solution

Ouhaman et al. [13] 2020 SSAP heuristic method
Tang et al. [17] 2016 SSAP, BAP Benders decomposition
Sun et al. [16] 2020 SSAP logic-based Benders
decomposition
Tang et al. [18] 2022 SSAP genetic algorithm
Ago et al. [1] 2007 SSAP, conveyor belt Lagrangian
routing decomposition
Menezes et al. [12] 2017 scheduling and planning branch and price
of product flows
De Andrade et al. [3, 4] 2021,2022 BAP, SSAP, scheduling column generation,
and planning of product diving heuristic
flows
Boland et al. [7] 2012 SSAP, BAP heuristic method
Robenek et al. [14] 2014 SSAP, BAP branch and price
Al-Hammadi and Diabat [2] 2017 SSAP, BAP branch and price
Babu et al. [5] 2015 SSAP, ship scheduling, logic-based Benders
train scheduling decomposition
Hu and Yao [9] 2012 SRSP genetic algorithm
Belassiria et al. [6] 2019 SRSP branch and bound
method, tabu search
algorithm
Hanoun et al. [8] 2013 operational scheduling bi-objective
of the continuous coal Optimization
handling
van Vianen et al. [20] 2015 SRSP simulation
Unsal and Oguz [19] 2019 SSAP, BAP, SRSP logic-based Benders
decomposition
Our work 2023 SSAP with MIP
environmental
consideration
SSAP: Storage space allocation problem, BAP: Berth allocation problem, SRSP:
Stacker-reclaimer scheduling problem, MIP: Mixed integer programming.

hangars. Tang et al. [17] addressed the integrated berth scheduling and SSAP in
the context of bulk raw material ports of large iron and steel companies. They
developed an integer programming model to minimize total costs and solved the
model using a Benders decomposition-based approach. Sun et al. [16] followed
up on the work of Tang et al. [17]. They first developed a mixed-integer pro-
gramming (MIP) formulation that could avoid generating scattered small fields
and schedule the unloading, stacking and reclaiming operations. A logic-based
Benders approach was proposed to solve the problem optimally. SSAP in large
iron ore terminal stockyards was studied in the work of Tang et al. [18]. They
used continuous variables to describe the specific locations of the empty fields
to highlight the continuous cargo flow characteristic of dry bulk stockyard oper-
ation. A MIP model was developed and a heuristic algorithm based on genetic
 Stockyard Space Allocation with Mist Cannons and Energy Expenditure 157

algorithm was proposed to solve it. Ago et al. [1] considered both the SSAP and
conveyor belt routing problems, using a Lagrangian decomposition algorithm.
Menezes et al. [12] studied a production planning and scheduling problem in
bulk cargo terminals. The problem considered planning and scheduling the flow
of products between supply, storage, and demand nodes and a branch and price
algorithm was applied to solve it. De Andrade et al. [3,4] extended this problem
by integrating berth allocation and yard assignment decisions.
Boland et al. [7] studied stockyard management in coal export terminals,
considering the rail system transporting coal from mines to the terminal. The
decisions in this study included berth allocation, stockpile location and the
start times of stockpile assembly and reclaiming decisions. Robenek et al. [14]
solved the integrated berth allocation and yard assignment problem in import
and export bulk ports as a single large-scale optimization problem. They con-
structed a MIP model and applied an exact algorithm based on a branch and
price framework. Al-Hammadi and Diabat [2] added constraints on the stacking
space capacity to the model of Robenek et al. [14]. Babu et al. [5] minimized the
delay of ships in bulk terminals by simultaneously considering ship scheduling,
stockyard planning, and train scheduling.
As an important component in the stockyard, the stacker-reclaimer is also
a factor that could influence the operation in the stockyard, which is consid-
ered in the decision-making of stockyard operation. Hu and Yao [9] discussed
the stacker-reclaimer scheduling problem (SRSP) with the objective of mini-
mizing the maximum completion time. The scheduling problem is formulated
as a MIP model, and a genetic algorithm is developed to solve it. Belassiria et
al. [6] also studied SRSP and developed two different heuristic methods, a branch
and bound method, and a tabu search algorithm. Hanoun et al. [8] addressed
the operational scheduling of the continuous coal handling problem. A model
of stockyard operations within a coal mine was described and the problem was
formulated as a Bi-Objective Optimization Problem (BOOP). In van Vianen et
al.’s work [20], simulation was applied to reschedule the stacker-reclaimers oper-
ation to increase the dry bulk terminal’s performance by reducing the waiting
time of cargo trains being loaded at the terminal. Unsal and Oguz [19] integrated
berth allocation, yard space allocation, and SRSP and modelled the master and
subproblems with MIP and constraint programming, respectively, solving the
problem with a logic-based Benders decomposition algorithm.
While the climate and environmental sustainability objectives are increas-
ingly recognized in port (operations) research [10,11,15], the existing research
related to the considered problem mostly focuses on the operation efficiency of
the stockyard with little consideration of environmental aspects. Dust suppres-
sion in the stockyard is paid little attention to. Our study is an attempt to
incorporate environmental protection considerations and neutral climate targets
into stockyard space management.
158 X. Tang and F. Schulte

3 Problem Description
The whole stockyard consists of some long stock pads, which may be pre-occupied
by existing stockpiles, leaving empty fields available for incoming materials.
Figure 1 shows the layout of a typical dry bulk terminal stockyard. One stacker-
reclaimer (abbreviated as “SR” below) travels on a track between two pads, and
the two pads are served by the same SR. Between two stock pads is a space
interval. Three intervals can be found in Fig. 2, and we label them as Int 1, 2,
and 3. Like SR, mist cannon (abbreviated as “MC” below) is also mobile. The
difference is that MC can move among intervals instead of moving on a fixed
track. They usually move from one interval to another through the end area of
the stock pad. When an SR stacks materials into the stockyard, the MC sprays
water mist to control the dust generated. There are two moving routes for the
SR in Fig. 2. After completion of handling material 1, if the MC continues to
serve material 2, it needs to move from Int 1 to Int 2, which brings vertical
transport distance. Otherwise, if the MC continues to serve material 3 planned
to be allocated in pad 2, it only needs to move horizontally.

Fig. 1. Typical stockyard layout in dry bulk terminals.

We investigate the Stockyard Storage Space Allocation Problem with the

MC operation to control the dust generated during stacking. We aim to find the
 Stockyard Space Allocation with Mist Cannons and Energy Expenditure 159

Fig. 2. Representation of mist cannon operation.

specific storage position for each incoming material with the objective of mini-
mizing the energy consumption of operations in the stockyard. The main energy
consumers in the stockyard are conveyor belts, SRs, and MCs. The energy con-
sumption of SR and MC are further divided into two parts: energy consumption
when handling materials and moving from the position of handled material to
the position of material to be next served. To calculate energy consumption, we
multiply the power of operating machines by the time consumed during handling
or moving. Time can be obtained by dividing distance by moving speed.
When calculating the travel distance, we need to mark the position of each
material and the working machine. We, therefore, introduce a virtual material 0
for each SR and MC (each SR and MC will at least serve material 0) and then
set the position of the virtual material as the initial position of the machines.
Hence, the moving distance can be normally calculated.

4 Mathematical Model
4.1 Assumptions
The following assumptions are made in this problem:
(1) There are plenty of empty fields to accommodate the incoming materials
over the whole planning period, which means there is always enough space
for stacking.
(2) One material can be stacked next to another material if a safe distance is
maintained to avoid mixing. We consider a dedicated terminal for a partic-
ular cargo type. For example, in an iron ore terminal, the different types of
iron ore are usually distinguished from each other by the size of the diameter
of the particles and there is no strict restriction that they cannot be stacked
next to each other, only that a safe distance is maintained to avoid mixing.
(3) Once the goods are stacked in the yard, the position will not change during
the planning period, which is consistent with the situations in practice.
(4) The operating capability of machines in the stockyard is constant during
the handling process.
(5) The moving distance is calculated from the end coordinate of the material
that has just been completed handling to the start coordinate of the material
to be next served.
160 X. Tang and F. Schulte

4.2 Model Formulation

In this section, we present the mathematical model of the stockyard storage

space allocation problem.
Sets and Parameters

P set of stock pads

C set of incoming materials
S set of stacker-reclaimers
W set of mist cannons
N set of intervals
L the horizontal length of a pad
dis the distance between two intervals in stockyards
li length of the space needed to stack the material i, i ∈ C
mi mass of the incoming material i, i ∈ C
ly p the vertical distance from the entrance of the stockyard to
stock pad p, p ∈ P
st
αnw binary parameter which indicates whether MC w is in the
st
interval n before planning starts, n ∈ N , w ∈ W ; if αnw = 1,
MC w is in the interval n
asp binary parameter which indicates whether SR s can cover
stock pad p, s ∈ S; if as p = 1, SR can cover stock pad p
bpn binary parameter which indicates whether pad p is adjacent
to interval n, n ∈ N ; if bpn = 1, pad p is adjacent to interval n
xss0 the start position of SR s when the stacking operation starts
xsw0 the start position of MC w when the stacking operation starts
capsrh handling capability of SR
capc handling capability of conveyor belt
spesrm moving speed of SR
spewcm moving speed of MC
specm moving speed of conveyor belt
Psrh power of SR when handling materials
Psrm power of SR when moving
Pwch power of MC when spraying water mist
Pwcm power of MC when moving
Pc power of conveyor belt
sd safe distance between two adjacent materials
M a positive number that serves as infinity for the problem
 Stockyard Space Allocation with Mist Cannons and Energy Expenditure 161

Decision Variables

kip 0–1 variable, equal to 1 if and only if the material i is

allocated to stock pad p
xsi start coordinate of material i
zijp 0–1 variable, equal to 1 if and only if material i and j are
allocated to stock pad p and material i is allocated to the left
of material j satisfying safe distance constraint
xis 0–1 variable, equal to 1 if material i is handled by SR s
yiw 0–1 variable, equal to 1 if material i is handled by MC w
αin 0–1 variable, equal to 1 if the MC serving material i is
located in the interval n
θijs 0–1 variable, equal to 1 if and only if the material i and j are
both handled by SR s, and material j is handled by SR s
right after material i
μijw 0–1 variable, equal to 1 if and only if the material i and j are
both handled by MC w, and material j is handled by MC w
right after material i
ρijw 0–1 variable, equal to 1 if and only if MC w moves in the
same interval when complete serving material i and then
continues to serve material j
πijw 0–1 variable, equal to 1 if and only if MC w move from one
interval to another when completing serving material i and
then continues to serve material j
ui , vi random number related to material i
δ, τ 0–1 variable

The mathematical model can be formulated as follows:

  
 mi  mi  mi  Pc xsi + p ly p kip
min Psrh + Pwch + Pc +
i
capsrh i
capsrh i
capc i
specm
 
 θ0js |xsj − xss0 |    θijs xsj − xsi − li 

+ Psrm + Psrm
s j
spesrm s i j
spesrm
  
 μ0jw dis  n nαjn − n nαnw  st
+ Pwcm
w j
spewcm
  
 μijw dis  n nαjn − n nαin 
+ Pwcm (1)
w i j
spewcm
 ρ0jw |xsj − xsw0 |  π0jw (L − xsw0 + L − xsj )
+ Pwcm + Pwcm
w j
spewcm w j
spewcm
 
 ρijw xsj − xsi − li 
+ Pwcm
w i j
spewcm
 πijw (L − xsi − li + L − xsj )
+ Pwcm
w i j
spewcm
162 X. Tang and F. Schulte

s.t. 
kip = 1, ∀ i ∈ C \ {0} (2)
p∈P

x0 s = 1, ∀s∈S (3)

xis = 1, ∀ i ∈ C \ {0} (4)
s∈S

kip ≥ 1 + M (1 − xis ) , ∀ i ∈ C \ {0} , s ∈ S (5)
{p∈P |asp =1}

kip ≤ 1 − M (1 − xis ) , ∀ i ∈ C \ {0} , s ∈ S (6)
{p∈P |asp =1}

θijs ≤ 1 + M (1 − xjs ) , ∀ j ∈ C \ {0} , s ∈ S (7)
i

θijs ≥ 1 − M (1 − xjs ) , ∀ j ∈ C \ {0} , s ∈ S (8)
i

θijs ≤ 1 + M (1 − xis ) , ∀ i ∈ C \ {0} , s ∈ S (9)
j∈C\{0}

xis > 1 − M (1 − δ) , ∀s∈S (10)
i

θ0js ≥ 1 − M (1 − δ) , ∀s∈S (11)
j∈C\{0}

θ0js ≤ 1 + M (1 − δ) , ∀s∈S (12)
j∈C\{0}

vi < vj + M (1 − θijs ) , ∀i, j ∈ C \ {0} , s ∈ S (13)

θijs + θjis ≤ 1, ∀ i, j ∈ C, s ∈ S (14)
θijs ≤ xis , ∀i, j ∈ C, i = j, s ∈ S (15)
θijs ≤ xjs , ∀ i, j ∈ C, i = j, s ∈ S (16)
xsj − (xsi + li ) − sd > M (zijp − 1) , ∀ i, j ∈ C \ {0} , i = j, p ∈ P (17)
xsj − (xsi + li ) − sd ≤ M zijp , ∀ i, j ∈ C \ {0} , i = j, p ∈ P (18)
zijp + zjip ≥ 1 − M (2 − kip − kjp ) , ∀ i, j ∈ C \ {0} , i = j, p ∈ P (19)
zijp ≤ kip , ∀ i, j ∈ C \ {0} , p ∈ P (20)
zijp ≤ kjp , ∀ i, j ∈ C \ {0} , p ∈ P (21)

αin = 1, ∀ i ∈ C \ {0} (22)
n∈N

αin ≥ 1 − M (1 − kip ) , ∀i ∈ C \ {0} , p ∈ P (23)
{n∈N |bpn =1}
Stockyard Space Allocation with Mist Cannons and Energy Expenditure 163

αin ≤ 1 + M (1 − kip ) , ∀ i ∈ C \ {0} , p ∈ P (24)
{n∈N |bpn =1}

y0w = 1, ∀w∈ W (25)


yiw = 1, ∀ i ∈ C \ {0} (26)
w

μijw ≤ 1 + M (1 − yjw ) , ∀ j ∈ C \ {0} , w ∈ W (27)
i

μijw ≥ 1 − M (1 − yjw ) , ∀ j ∈ C \ {0} , w ∈ W (28)
i

μijw ≤ 1 + M (1 − yiw ) , ∀ i ∈ C \ {0} , w ∈ W (29)
j∈C\{0}

yiw > 1 − M (1 − τ ) , ∀w∈W (30)
i

μ0jw ≥ 1 − M (1 − τ ) , ∀w∈W (31)
j∈C\{0}

μ0jw ≤ 1 + M (1 − τ ) , ∀w∈W (32)
j∈C\{0}

ui < uj + M (1 − μijw ) , ∀ i, j ∈ C \ {0} , w ∈ W (33)

μijw + μjiw ≤ 1, ∀ i, j ∈ C, w ∈ W (34)
μijw ≤ yiw , ∀ i, j ∈ C, w ∈ W (35)
μijw ≤ yjw , ∀i, j ∈ C, w ∈ W (36)
 
nαjn ≤ nαin + M (1 − ρijw ) , ∀i, j ∈ C \ {0} , w ∈ W (37)
n n
 
nαjn ≥ nαin − M (1 − ρijw ) , ∀ i, j ∈ C \ {0} , w ∈ W (38)
n n
 
nαjn ≤ st
nαnw + M (1 − ρ0jw ) , ∀ j ∈ C \ {0} , w ∈ W (39)
n n
 
nαjn ≥ st
nαnw − M (1 − ρ0jw ) , ∀ j ∈ C \ {0} , w ∈ W (40)
n n

ρijw ≤ μijw , ∀ i, j ∈ C, w ∈ W (41)

πijw ≤ μijw , ∀ i, j ∈ C, w ∈ W (42)
ρijw + πijw ≤ 1 + M (1 − μijw ) , ∀ i, j ∈ C, w ∈ W (43)
ρijw + πijw ≥ 1 − M (1 − μijw ) , ∀ i, j ∈ C, w ∈ W (44)
0 ≤ xsi ≤ L, ∀i∈C (45)
0 ≤ xsi + li ≤ L, ∀i∈C (46)
164 X. Tang and F. Schulte

kip , xis , yiw , θijs , μijw , zijp , αin ∈ {0, 1} ,

(47)
∀ i, j ∈ C, p ∈ P, s ∈ S, w ∈ W, n ∈ N
The objective (1) is to minimize the total energy consumption of operation
in the stockyard. The first three terms are the energy consumption of SRs, MCs
and conveyor belts, respectively of handling (SRs and MCs) and transporting
(conveyor belts) materials. The fourth term is the energy consumption of the
conveyor belt when moving from the stockyard entrance to the stacking posi-
tion of materials in the stockyard. The next two terms represent the energy
consumption of SRs when moving from the position of the material they have
just handled to the position of the material they will next serve. The energy
consumption of MC when moving from one material to the next one is calcu-
lated by the last six terms. The transport distance is calculated vertically and
horizontally respectively. The terms with variable μ0jw , ρ0jw , π0jw calculate the
energy consumption when moving from the initial position of each MC to the
position of material they first serve.
Constraint (2) guarantees that each material can be allocated to only one
stock pad. Constraint (3) allocates a virtual starting material for each SR. Con-
straint (4) ensures that each material is served by only one SR. Constraint
(5) and (6) define the relationship between variable xis and kip . If material i
is served by SR s, the material i will be allocated to the stock pad p which
stacker-reclaimer s can cover. Constraints (7)–(13) ensure that for each SR, if it
is assigned to serve material in the planning period, the served material num-
ber will form a sequence beginning with the number 0, with each number of
served material appearing only once. Constraint (14) states that two materi-
als served by one SR can not be handled simultaneously. Constraints (15) and
(16) are the premise of the non-overlapping constraints in time for materials
served by the same SR. Constrains (17), (18), and (19) are the non-overlapping
restrictions for any two materials stacked in the same pad and they must satisfy
the safe distance restriction. Constraints (20) and (21) are the premise of the
non-overlapping restrictions in space.
Constraint (22) guarantees that the MC serving each material can be located
to only one interval. Constraint (23) and (24) defines the relationship between
variable αin and kip . If material i is located in stock pad p, the MC serving
material i must be located in the interval to which the stock pad p is adjacent
to. Constraint (25) allocates a virtual starting material for each MC. Constraint
(26) states that each material is served by only one MC. Constraints (27)–(33)
ensure that for each MC, if it is assigned to serve material in the planning period,
the served material number will form a sequence beginning with the number
0, with each number of served material appearing only once. Constraint (34)
states that two materials served by one MC can not be handled simultaneously.
Constraints (35) and (36) are the premise of the non-overlapping constraints in
time for materials served by the same MC. The meaning of ρijw is defined in
constraints (37) and (38). If MC w remains in the same interval when moving
from the position of material i to that of material j, the MC serving material i
and j are in the same interval. Constraints (39) and (40) have the same meaning
 Stockyard Space Allocation with Mist Cannons and Energy Expenditure 165

as the above two constraints, but they describe the situation of an MC moving
from the initial position to its first serving material. Constraints (41) and (42)
are the premise of variable ρijw and πijw . Constraint (43) and (44) state that
only one of the two variables ρijw and πijw can be 1 if there is pair i→j in
the material handling sequence of MC w. Constraint (45) and (46) show the
value range of start and end coordinates. Lastly, constraint (47) determines the
domains of 0–1 variables.

5 Computational Experiments

In this section, we conducted computational experiments to test the effective-

ness of the proposed mathematical model using solver CPLEX 22.1 under C++
environment with a time limit of 3600 s.

5.1 Instance Generation and Computational Results

We generated 12 instances and divided them into 3 sets according to the number
of stock pads, SRs, and MCs. We consider |C| materials, |P | stock pads, |S| SRs
and |W | MCs in computational experiments. The configuration of each instance
can be found in Table 1. Set 1 has 4 stock pads with 2 SRs and 1 MC. Set 2 has
the same configuration as set 1, except for 2 MCs. Set 3 has 6 stock pads with
3 SRs and 3 MCs. Each set consists of computational instances with different
material numbers. The material information (mass and length) in instances with
the same material numbers is the same. (Instances 1, 5, and 9 have the same
information, which is the same situation for instances 2, 6, and 10; instances 3,
7, and 11 and instances 4, 8, and 12.) The material information is generated
randomly.
Table 2 shows the results obtained within the time limit. Since the first three
terms in the objective (1) are fixed terms, we use “Objective value of variable
parts” in the table to explicitly show the objective value dependent on decision
variables, that is, the value of objective (1) excluding the first three terms. The
“objective value” in the following text refers to the value of objective excluding
the fixed terms. Optimal solutions can be found for 9 of the 12 instances. The
increase in the number of stock pads, SRs, MCs and materials expands the scale
of the model, thus making the computation time increase dramatically. It can
be noticed that the results of instances 1, 5, and 9 are the same (Instances 2, 6,
and 10 are in the same situation.), which means the change of some parameters
like the number of stock pads, SRs and MCs have little influence on optimal
solutions to some extent.
We compare the composition of objective value (of variable parts) of instance
5–8 in Fig. 3. The energy consumption of the conveyor belt moving to the spec-
ified stacking position for materials takes a larger proportion compared to the
other two components (energy consumption of SR and MC) during moving. It
is observed that there is a clear absolute variation of the energy consumption of
the belt conveyor moving among different instances, while the change of energy
166 X. Tang and F. Schulte

Table 2. Computational results for the test instances

Set Ins No. |C| |P | |S| |W | Objective value of time(s)

variable parts(kW·h)
1 5 4 2 1 15.25 4.64
2 5 4 2 1 13.25 5.04
1
3 6 4 2 1 19.76 68.61
4 7 4 2 1 24.17 2480.27

5 5 4 2 2 15.25 9.91
6 5 4 2 2 13.25 41.12
2
7 6 4 2 2 19.76 104.54
8 7 4 2 2 24.17 3600

9 5 6 3 3 15.25 226.45
10 5 6 3 3 13.25 216.92
3
11 6 6 3 3 19.76 3600
12 7 6 3 3 23.36 3600

consumption of SR and MC moving is rather less obvious. The possible reason

is that each SR and MC moves between two stock pads (in one interval), the
transport distance of SR and MC is therefore relatively shorter than the distance
the conveyor belt travels.

5.2 Managerial Insights

With the calculation results in Sect. 5.1, we can extract some managerial insights
which may be helpful for storage space management in the stockyard.

(1) The number of stock pads open can be determined according to the number
of incoming materials. Stockyards do not always need to open all the stock
pads. We can find that the objective values of instances 1,5,9 are the same,
which means the results obtained in the background of 4 stock pads and 6
stock pads are the same. Opening stock yards depending on the number of
materials may help save operational costs of the stockyard.
(2) Materials are stacked into the adjacent stock pads which one stacker-
reclaimer and mist cannon can cover, which could shorten the travel dis-
tance of SR and MC, thus saving the energy consumed. As shown in Fig. 4
and 5, all the materials are stacked in the adjacent pad 1 and 2.
(3) Materials tend to be stacked in the stock pads close to the stockyard entrance
so that the conveyor belts do not need to move a long distance, which
could reduce the energy consumption of the conveyor belt. Figure 4 and 5
show that materials are stacked in the pad 1 and 2 which are closest to the
entrance of the stockyard.
Stockyard Space Allocation with Mist Cannons and Energy Expenditure 167

Fig. 3. Comparison of the objective value of instance 5–8.

Fig. 4. Representation of solution for instance 5.

Fig. 5. Representation of solution for instance 7.

168 X. Tang and F. Schulte

6 Conclusion
We study the Storage Space Allocation Problem with the consideration of oper-
ation machines in the stockyard. Existing studies investigate the storage space
allocation problem with other operation problems such as berth allocation and
stacker reclaimer scheduling. However, they pay little attention to environmen-
tal sustainability. We take the operation of mist cannon to control dust during
stacking into consideration and aim to minimize the energy consumption during
stacking operation. A mixed integer programming model is established, and we
show the effectiveness of the proposed model by conducting computational exper-
iments on some small-scale instances. Some managerial insights are extracted,
which may help in storage space management as a reference.
Our future work will focus on the improvement of the formulation of the
model and the development of an efficient algorithm to solve the model for large-
scale instances. And we need to collect operational data from real-world practice
as the input of computational experiments to validate the results considering
practical implications.

References
1. Ago, M., Nishi, T., Konishi, M.: Simultaneous optimization of storage allocation
and routing problems for belt-conveyor transportation. J. Adv. Mech. Design Syst.
Manuf. 1(2), 250–261 (2007)
2. Al-Hammadi, J., Diabat, A.: An integrated berth allocation and yard assignment
problem for bulk ports: formulation and case study. RAIRO - Oper. Res. 51(1),
267–284 (2017). https://doi.org/10.1051/ro/2015048. Identifier: ro151165
3. de Andrade, J.L.M., Menezes, G.C.: An integrated planning, scheduling, yard allo-
cation and berth allocation problem in bulk ports: model and heuristics. In: Mes,
M., Lalla-Ruiz, E., Voß, S. (eds.) ICCL 2021. LNCS, vol. 13004, pp. 3–20. Springer,
Cham (2021). https://doi.org/10.1007/978-3-030-87672-2 1
4. de Andrade, J.L.M., Menezes, G.C.: A column generation-based heuristic to solve
the integrated planning, scheduling, yard allocation and berth allocation problem
in bulk ports. J. Heuristics 29(1), 39–76 (2022). https://doi.org/10.1007/s10732-
022-09506-3
5. Babu, S.A.K.I., et al.: Minimizing delay of ships in bulk terminals by simultaneous
ship scheduling, stockyard planning and train scheduling. Maritime Econ. Logist.
17(4), 464–492 (2015). https://doi.org/10.1057/mel.2014.20
6. Belassiria, I., Mazouzi, M., Elfezazi, S., Torbi, I.: A heuristic methods for stacker
and reclaimer scheduling problem in coal storage area. In: 2019 International Col-
loquium on Logistics And Supply Chain Management (LOGISTIQUA) (2019)
7. Boland, N., Gulczynski, D., Savelsbergh, M.: A stockyard planning problem. EURO
J. Transp. Logist. 1(3), 197–236 (2012). https://doi.org/10.1007/s13676-012-0011-
z
8. Hanoun, S., Khan, B., Johnstone, M., Nahavandi, S., Creighton, D.: An effec-
tive heuristic for stockyard planning and machinery scheduling at a coal handling
facility. In: 2013 11th IEEE International Conference On Industrial Informatics
(INDIN), pp. 206–211 (2013)
 Stockyard Space Allocation with Mist Cannons and Energy Expenditure 169

9. Hu, D., Yao, Z.: Stacker-reclaimer scheduling in a dry bulk terminal. Int. J. Com-
put. Integr. Manuf. 25(11), 1047–1058 (2012)
10. Iris, Ç., Lam, J.S.L.: A review of energy efficiency in ports: operational strategies,
technologies and energy management systems. Renew. Sustain. Energy Rev. 112,
170–182 (2019)
11. Lam, J.S.L., Notteboom, T.: The greening of ports: a comparison of port manage-
ment tools used by leading ports in Asia and Europe. Transp. Rev. 34(2), 169–189
(2014)
12. Menezes, G.C., Mateus, G.R., Ravetti, M.G.: A branch and price algorithm to
solve the integrated production planning and scheduling in bulk ports. Eur. J.
Oper. Res. 258(3), 926–937 (2017). https://doi.org/10.1016/j.ejor.2016.08.073
13. Ouhaman, A.A., Benjelloun, K., Kenne, J.P., Najid, N.: The storage space allo-
cation problem in a dry bulk terminal: a heuristic solution. In: 21st IFAC World
Congress on Automatic Control - Meeting Societal Challenges, vol. 53, pp. 10822–
10827 (2020). https://doi.org/10.1016/j.ifacol.2020.12.2868
14. Robenek, T., Umang, N., Bierlaire, M., Ropke, S.: A branch-and-price algorithm to
solve the integrated berth allocation and yard assignment problem in bulk ports.
Eur. J. Oper. Res. 235(2), 399–411 (2014). https://doi.org/10.1016/j.ejor.2013.08.
015
15. Schulte, F., Voß, S., Ascencio, L.M., González-Ramı́rez, R.G.: Directions for sus-
tainable ports in Latin America and the Caribbean. Int. J. Transp. Econ. 43(3),
315–337 (2016)
16. Sun, D., Meng, Y., Tang, L., Liu, J., Huang, B., Yang, J.: Storage space allocation
problem at inland bulk material stockyard. Transp. Res. Part E: Logist. Transp.
Rev. 134 (2020). https://doi.org/10.1016/j.tre.2020.101856
17. Tang, L., Sun, D., Liu, J.: Integrated storage space allocation and ship scheduling
problem in bulk cargo terminals. IIE Trans. 48(5), 428–439 (2016). https://doi.
org/10.1080/0740817x.2015.1063791
18. Tang, X., Jin, J.G., Shi, X.: Stockyard storage space allocation in large iron ore
terminals. Comput. Industr. Eng. 164, 107911 (2022). https://doi.org/10.1016/j.
cie.2021.107911
19. Unsal, O., Oguz, C.: An exact algorithm for integrated planning of operations in
dry bulk terminals. Transp. Res. Part E: Logist. Transp. Rev. 126, 103–121 (2019).
https://doi.org/10.1016/j.tre.2019.03.018
20. van Vianen, T., Ottjes, J., Lodewijks, G.: Simulation-based rescheduling of the
stacker-reclaimer operation. J. Comput. Sci. 10, 149–154 (2015). https://doi.org/
10.1016/j.jocs.2014.06.004
 Planning LNG Annual Delivery Programs
with Speed Optimization and Multiple
Loading Ports

Helle V. Haug1 , Sigrid H. Solum1 , Sanna B. Warholm1 , Kjetil Fagerholt1(B) ,

Mingyu Li2 , and Inge Norstad2
1
Department of Industrial Economics and Technology,
Norwegian University of Science and Technology, Trondheim, Norway
[email protected]
2
Quorum Software, Houston, TX, USA
[email protected]

Abstract. We study the planning problem of creating Annual Deliv-

ery Programs (ADPs) in the Liquefied Natural Gas Industry (LNG), in
which an LNG producer must fulfill a series of long-term contracts for
LNG deliveries with customers all over the world with a given fleet of
LNG vessels. We expand existing models in the literature by also con-
sidering speed optimization and multiple loading ports. We denote this
planning problem as the LNG-ADP with Speed Optimization and Mul-
tiple Load Ports (LNG-ADP-SO-MLP). We propose a novel and efficient
mixed integer programming (MIP) model for the LNG-ADP-SO-MLP.
The model is solved using a commercial MIP-solver on a number of realis-
tic instances for two different LNG producers. It is shown that the solver
obtains solutions with less than 1% optimality gaps within one hour for
instances with a planning horizon of up to 180 days. This indicates that
the model can efficiently be embedded within a rolling horizon heuristic
to solve instances with longer planning horizons.

Keywords: Liquefied Natural Gas · Annual Delivery Program · Ship

Routing · Inventory Management · Speed Optimization

1 Introduction
The earth contains enormous amounts of natural gas, which are often located far
from where the gas is needed. For long distances the gas must be transported by
sea. To make transportation volume-efficient and safe, the natural gas is cooled
down and condensed to its liquid form, i.e., into Liquefied Natural Gas (LNG),
at liquefaction plants, before being loaded onto specially designed LNG vessels
for transportation to its destination where the LNG is boiled to natural gas at
a regasification port.
The demand for LNG has steadily been increasing and today there are liqui-
dation and regasification ports organized for shipping of LNG in large parts of
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 170–184, 2023.
https://doi.org/10.1007/978-3-031-43612-3_10
 Planning LNG Annual Delivery Programs 171

the world. Countries in some areas, especially the Middle East, USA, and Aus-
tralia, are net exporters of LNG, while countries in areas including China, Japan,
and Europe are net importers. Large geographical distances between suppliers
and customers create a high demand for shipping LNG by sea.
The average charter rate (rental price) for a 160 000 cubic meter LNG vessel
in 2021 was around 89 000 USD/day, up from around 59 300 USD/day in 2020 [4].
However, the spot charter rate reached as high as 400 000 USD/day during 2022,
mainly due to political tensions related to Russia being at war and cutting off
gas supplies to Europe, as well as the sabotage of two pipelines in the Baltic Sea
[3]. Although this is an unusual situation, it highlights the current importance
of optimizing LNG transportation to maximize the utilization of the vessel fleet.
In this paper, we study a special case of the maritime inventory routing
problem for an LNG producer, which is the planning of so-called Annual Delivery
Programs (ADPs). We denote this planning problem as the LNG-ADP, in which
the producer must fulfill a series of long-term contracts with customers all over
the world while also having the option to sell LNG in the spot market. The
producer is in charge of the inventory of LNG at its liquefaction plant(s), the
loading port with a limited number of berths, and the routing and scheduling
of a heterogeneous fleet of LNG vessels. The LNG-ADP describes the customer
deliveries of LNG cargoes for the next 12 months. In the LNG-ADP, the LNG
producer aims at maximizing the profits, which consists of the revenue from the
sales of LNG less the sailing and vessel chartering costs.
The LNG-ADP is, due to its importance, well studied. An important pioneer-
ing study was published by Rakke et al. [10], who proposed a Rolling Horizon
Heuristic (RHH) for this planning problem. Later, Stålhane et al. [11] used a
construction and improvement heuristic that makes a set of solutions with a
greedy insertion procedure, improves them with a first-descent neighborhood
search and/or branch-and-bound on a mathematical formulation. Al Haidous et
al. [1] base their model on the same assumptions as [10], but consider the spe-
cific case of using a homogeneous fleet of LNG-delivery vessels, with the goal of
minimizing fleet size. Andersson et al. [2] also model the LNG-ADP problem in
the same way as [10], but add four families of valid inequalities to improve the
lower bounds on the problem’s optimal value.
The study by Mutlu et al. [8] stands out by including two extensions: 1) the
production of LNG is a decision variable and not an input parameter, and 2)
allowing split deliveries, where a vessel can unload partial cargoes in more than
one unloading port. Li et al. [6] were the first to include transshipment in the
LNG-ADP problem, motivated by the Yamal LNG case where a transshipment
port helps the producer to avoid longer-than-necessary voyages with ice-breaking
LNG vessels, as normal LNG vessels can take over the LNG and deliver it to the
customers. Li et al. [6] introduce a continuous-time formulation for this problem,
while Li et al. [5] later proposes a time-discrete model for the same problem. Both
models are embedded into and solved with an RHH.
In this paper, we expand previous models of the LNG-ADP by including opti-
mization of the vessels’ sailing speeds. In contrast to currently existing models
172 H. V. Haug et al.

from the literature, our models also allows for multiple loading ports, which is
becoming relevant for some producers as the industry develops. Hence, we refer
to our version of the LNG-ADP with Speed Optimization and Multiple Load-
ing Ports (LNG-ADP-SO-MLP). Speed optimization in ship routing has been
widely studied, e.g., [9,12], although not in the context of LNG-ADP planning.
We propose a novel Mixed Integer Programming (MIP) model for the LNG-ADP-
SO-MLP. We generate a number of realistic instances based on data from two
LNG producers and show that our MIP model can be efficiently solved with low
optimality gaps for instances with a planning horizon of up to 180 days using
a commercial MIP-solver. Furthermore, the results show that including speed
optimization in the planning can improve the profits and make the planning
more flexible.
The remaining of this paper is organized as follows. Section 2 gives a formal
definition of the LNG-ADP-SO-MLP, while our MIP model for this problem is
introduced in Sect. 3. The computational study is presented in Sect. 4, before we
conclude in Sect. 5.

2 Problem Definition
The LNG-ADP-SO-MLP considers an LNG producer’s planning of the supply
of LNG to customers over a planning horizon with a given vessel fleet. The
producer controls one or more loading (liquefaction) ports where the LNG is
produced (liquefied). Each loading port has a given daily production, which
may vary throughout the year. Furthermore, each loading port has a given set
of berths from which the vessels load LNG. The number of available berths may
also vary over time, e.g., due to maintenance. In addition, each loading port has
storage tanks with a given maximum and minimum storage capacity.
There is a given set of customers, which either require LNG Delivered Ex-Ship
(DES), where the LNG producer transports the cargo to the customer with its
own vessels, or Free-On-Board (FOB), where the customers pick up the cargo
at an agreed loading port with their own vessels. Each DES contract has an
associated unloading port, where the vessels unload the LNG cargoes. Since the
customer decides when it wants the LNG delivered, we assume there is always
available inventory capacity and berth capacity at the unloading ports. Both
DES and FOB customers define their demand in advance for specific subsets
of the planning horizon, hereby referred to as partitions. A customer typically
has annual, quarterly, and/or monthly partitions with an associated delivery
requirement. The customers’ required contract volumes are satisfied by one or
more deliveries in each partition. For a DES customer, the minimum demand
within each partition must be satisfied, and a revenue is associated with deliveries
above this minimum up to a specified maximum demand within the partition.
For FOB contracts, each delivery must be the size of the vessel picking it up. In
addition to these contract deliveries, it is also possible to have optional one-time
deliveries of spot cargoes with associated customer destinations and revenues.
These have to be delivered either using the LNG producer’s vessels (like DES)
or by using the customer’s own vessel (like FOB).
 Planning LNG Annual Delivery Programs 173

Each vessel in the LNG producer’s fleet has a given fuel consumption depend-
ing (non-linearly) on the sailing speed, as well as a maximum and minimum
sailing speed. The distances between ports are known in advance. If the LNG
producer is short on vessels, additional vessels can be chartered at a given daily
rate. Each vessel has a specific loading capacity and is always loaded up to
capacity due to physical limitations. It is also required to unload the entire load
at the unloading port, except for a minimum tank volume share that is needed
to keep the tanks cool due to the boil-off effect, which means that a fraction
of the LNG that is transported will evaporate during a voyage, typically at a
rate of 0.10–0.12% per day [7]. The actual delivery volume of each delivery is
therefore defined by the capacity of the vessel and the boil-off, which depends
on the sailing distance. Each vessel has a given start position at the beginning
of the planning horizon. A vessel becomes available for a new voyage when it
returns to a loading port, and may therefore become available some days into
the beginning of the planning horizon.
Some vessels might be scheduled for maintenance during the planning hori-
zon. This must take place at a pre-defined maintenance port and start within a
given time window. After maintenance, the vessel has to go through a purge and
a cool-down procedure, which must be done at a loading port and is assumed to
take a given amount of time. All vessels are assumed to have the same specified
time associated with loading and unloading (i.e., one day).
The main decisions of the LNG-ADP-SO-MLP are where to send each vessel
at what times and at what speeds, i.e., the vessel routes and schedules, while
making sure that all customer delivery requirements are satisfied and that the
inventory levels at the loading ports are kept within its limits at all times. Not
all vessels can visit all ports, and some vessels cannot serve all customers due to
their properties not being compatible with the customer’s unloading port. Also,
some vessels are restricted to serving specific customer contracts. Furthermore,
no more vessels than there are available berths can load on a given day at a
given production port. Additional decisions include how much spot cargo to sell
within certain demand and revenue specifications.
The planning objective of the LNG-ADP-SO-MLP is to maximize the gross
margin less the transportation costs (hereby referred to as the “profit”). The
revenue comes from spot cargoes transported DES and FOB, from long-term
contracted LNG deliveries (DES and FOB), and lastly from the value of the
remaining LNG at the end of the planning period. The costs are the (speed-
dependent) sailing costs for the producer’s fleet of vessels, as well as the char-
tering costs. The resulting plan is an Annual Delivery Program, describing the
routes and schedules for the vessels in the fleet over the given planning horizon.

3 Mathematical Model

In the following, we explain the modeling assumptions and generation of the

underlying network in Sect. 3.1. We introduce the notation in Sect. 3.2, before
the model is presented in Sect. 3.3.
174 H. V. Haug et al.

3.1 Modeling Assumptions and Network Generation

The proposed MIP model is time-discrete where an arc represents a one-way

sailing between two ports in a time-space network. Each time period is one day,
which is a reasonable resolution for the LNG-ADP-SO-MLP. Before running the
MIP model, an arc generation procedure generates feasible arcs, which are given
as input to the MIP model. The arcs are generated so that we allow that the LNG
vessels can wait at a port before loading/unloading for up to a specified number
of days. This makes the model more flexible when it comes to handling both
inventory constraints at the loading ports and contract delivery requirements
for the customers.
A node in the time-space network is a combination of a port and a time
period, denoted (i, t) or (j, t ), where i and j denote ports, and t and t denote
time periods. Each port represents one of five port types: 1) a loading port,
where LNG is stored before being loaded onto the vessels before transportation;
2) an unloading (customer) port, where LNG is unloaded from the vessel and
regasified; 3) a maintenance port, which the vessels that require maintenance
must visit within the specified time winodws; 4) a spot unloading port, which is
similar to an unloading port, except that only optional cargoes can be unloaded
here; 5) the artificial initialization and destination ports for the vessels.
An arc in the network ((i, t), (j, t )) for a vessel links two nodes. The start
time of an arc is defined by the point in time when the vessel starts sailing from
the start port. The end time of an arc is the start time of the arc, plus the
time it takes to complete four vessel activities: 1) sailing; 2) potential waiting;
3) potential purge and cool-down procedure; 4) the port operations, i.e., the
time for loading/unloading. An example of an arc in the time-space network is
shown in Fig. 1. The blue line represents the arc ((i, t), (j, t )), and the pink lines
represent the actual processes included in the arc. Waiting is relevant when the
sailing distance and times of an arc imply a speed lower than the minimum speed
of the vessel. The vessel will then sail at its minimum speed and wait outside the
port in order to fulfill the time requirement. The purge and cool-down process
is only relevant when a vessel comes from a maintenance port.
As shown in Fig. 1, the arcs are defined so that the sailing precedes the other
activities. This modeling choice is made to simplify the berth constraints at the
loading port and to reduce symmetry. In the case of purge and cool-down, a
vessel requires more time in the loading port after maintenance. An arc provides
information about what type of node a vessel came from (the i in ((i, t), (j, t ))).
Therefore, we know if a purge-and-cool down is necessary, and thus how long a
berth is busy before the next vessel can arrive.
Artificial nodes and their associated arcs are included for two purposes. First,
this makes sure only one journey is started for each vessel. This is done by having
only one electable arc from the artificial origin node to the time and location
where the vessel first becomes available. The artificial origin node is located in
(i, t) = (0, 0). The arcs from the artificial origin node have 0 associated costs.
Second, it easily allows for not using a vessel at all, represented by an arc going
from the artificial origin node in (i, t) = (0, 0) to the artificial destination node
 Planning LNG Annual Delivery Programs 175

Fig. 1. Illustration of an arc ((i, t), (j, t )) and the vessel activities it represents

in (i, t) = (0,|T | + 1), where |T | is the length of the planning horizon. These arcs
also have 0 associated costs.
Since there is always an option to either shut down production or accept
potential additional spot contracts for the LNG producer throughout the year, an
assumption about defining an artificial spot FOB contract is made. This artificial
spot FOB contract has zero revenue and allows surplus LNG to potentially be
sold throughout the planning period. Each cargo loaded by the artificial spot
FOB contract in the final ADP represents a possibility for the LNG producer to
find a corresponding real-life spot contract for the given cargo’s pick-up interval.
Following this, we generate the network that is used as input to the MIP
model. Note that each arc has a calculated cost and also represents a given
associated sailing speed, and hence fuel cost. As an example, consider the two
arcs ((i, t), (j, t )) and ((i, t), (j, t + 1)), the latter arc allows for spending an
additional day than the former on the sailing between the ports i and j. This
allows for sailing at a lower speed and sailing cost.

3.2 Notation

Sets

V Set of producer-operated vessels that are available during the planning horizon
VM Set of vessels that require maintenance during the planning horizon, V M ⊂ V
Vi Set of vessels that can serve node i, Vi ∈ V
N Set of ports
NL Set of loading ports, N L ⊂ N
NU Set of unloading ports, N U ⊂ N
NS Set of spot unloading ports, N S ⊂ N
NM Set of maintenance ports, N M ⊂ N
176 H. V. Haug et al.

Av Set of feasible arcs vessel v can sail

AMv Set of feasible maintenance arcs vessel v can sail to start maintenance, AMv ⊂ Av
AUv Set of feasible arcs ship v can sail to deliver a DES long-term contracted cargo, AU
v ⊂ Av
ASv Set of feasible arcs ship v can sail to deliver a DES spot cargo, AS
v ⊂ Av
Fi Set of FOB cargoes of LNG that want to be picked up at loading port i
FiU Set of long-term contracted FOB cargoes of LNG that must be picked up at
loading port i, FiU ⊂ Fi
FiS Set of Spot FOB cargoes who’s load can be picked up by a FOB vessel at loading
port i, also including the artificial FOB-pickup (f = 1), FiS ⊂ F
Pj Set of time partitions where customer j has DES delivery requirements
T Set of time periods (days) in the planning horizon
TL Set of time periods in loading days, where the vessels can lift LNGÂ from a loading port,
TL ⊂ T
TU Set of time periods in unloading days, where vessels can deliver LNGÂ to a customer,
TU ⊂ T
Tjp Set of time periods within partition p for customer j, Tjp ⊂ T
TfF OB Set of time periods where FOB cargo f can be picked up, TfF OB ⊂ T
Tv Set of time periods where vessel v is available to be scheduled, Tv ⊂ T
TvM The time period where maintenance of vessel v is scheduled to start, TvM ⊂ Tv

Parameters
S
Cvitjt Sailing cost of each feasible arc ((i, t), (j, t )) for vessel v
C
Citj Costs of using a charter vessel to deliver a cargo at port j, loading the cargo at
loading port i at time t
NvST ART Start port of vessel v
TvST ART First time period where vessel v is available to be scheduled
O
Tvij Operational time associated with sailing from port location i to port location j for
vessel v
C
Tij Sailing time for a charter vessel sailing from loading port i to unloading port j
TfOF
i
OB
Operational time associated to port location j for FOB cargo f
Lv Capacity of vessel v
C
L Upper limit for capacity of a charter vessel
LC Lower limit for capacity of a charter vessel
LFf
OB
Loading quantity of FOB cargo f
D jp Maximum demand of unloading port j in partition p
D jp Minimum demand of unloading port j in partition p
RfSF OB Revenue per volume unit of LNG loaded for FOB spot contract f
RfU F OB Revenue per volume unit of LNG loaded for long-term FOB contract f
DES
Rjt  Revenue per volume unit of LNG for delivering DES contract to customer j at time t
RiEN D Unit value of LNG left in storage tanks at loading port i at the end of the planning
horizon
Bjt Berth capacity at port j at time t
QP
it Produced quantity of LNG in loading port i in time period t

Si Maximum storage level of LNG at loading port i

Si Minimum storage level of LNG at loading port i
Si Initial storage level of LNG at the start of the planning horizon at loading port i
E Boil-off rate in percent of total vessel capacity
 Planning LNG Annual Delivery Programs 177

Variables
xvitjt 1 if vessel v sail arc ((i, t), (j, t )), and 0 otherwise
zf t 1 if a cargo from FOB contract f is done loading in time period t , and 0 otherwise
witj 1 if a charter vessel starts sailing from loading port i at time t to deliver a cargo at j,
and 0 otherwise
gitj Amount loaded by a charter vessel in loading port i at time t to deliver in
unloading port j
sit Remaining storage at loading port i at the end of time period t

3.3 Mathematical Model

Objective Function. The Objective function (1) maximizes profit, which is

represented by revenue less vessel sailing costs. The first seven terms represent
1) the revenue from the long-term DES contracts served by own vessels, 2) the
long-term DES contract shipments shipped by charter vessels, 3) the DES spot
contracts served by own vessels, 4) the DES spot contract served by charter
vessels, 5) the long-term FOB contracts, 6) the revenue from the spot FOB
contracts, and 7) the value of the remaining LNG at the end of the planning
horizon, respectively. Note that the boil-off effect is accounted for by including
the boil-off rate E in terms 1 to 4. The last two terms are the cost terms. The
first cost term is the cost of sailing the producer-operated vessels, while the last
term represents the cost of using charter vessels.

  DES 
maxz = Rjt Lv (1 − (t − t)E)xvitjt
v∈V ((i,t),(j,t ))∈AU
v
   DES C
  DES 
+ Rj,t+T C (1 − Tij E)gitj + Rjt Lv (1 − (t − t)E)xvitjt
ij
i∈N L t∈T L j∈N U v∈V ((i,t),(j,t ))∈AS
v
   DES C
   U F OB F OB
+ Rj,t+T C (1 − Tij E)gitj + Rf Lf zf t
ij
i∈N L t∈T L j∈N S i∈N L f ∈F U t ∈T
i
   SF OB F OB
 EN D
+ Rf Lf zf t + Ri si,|T |
i∈N L f ∈F S t ∈T i∈N L
i
  S
   C
− Cvitjt xvitjt − Citj witj
v∈V ((i,t),(j,t ))∈Av i∈N L t∈T j∈N U

(1)

Inventory Constraints. Constraints (2) to (4) handle the inventory at the

loading ports. Constraints (2) initialize the constraint for each loading node,
while Constraints (3) make sure that the inventory requirement is fulfilled for
each day afterwards. Constraints (4) ensure that the inventory level is always
within its limits.
   
si1 = Si + QP
i1 − Lv xvi1jt − gi1j
j∈N U t ∈T v∈Vi j∈N U
  (2)
− LF
f
OB
zf t , i ∈ NL
f ∈Fi t ∈TfF OB
178 H. V. Haug et al.

   
sit = si,t−1 + QP
it − Lv xvitjt − gitj
j∈N U t ∈T v∈Vi j∈N U
  (3)
− LF
f
OB
zf t , i ∈ N L , t ∈ T L \{1}
f ∈Fi t ∈TfF OB

Si ≤ sit ≤ Si , i ∈ N L, t ∈ T L (4)

Maintenance and Vessel Flow Constraints. Constraints (5) state that the
vessels in need of maintenance perform this exactly one time. Constraints (6)
ensure that each vessel either starts at its start position at its first available day
or is not used at all. Constraints (7) are vessel flow conservation constraints.

xvitjt = 1, v ∈ VM
(5)
((i,t),(j,t ))∈AM
v

xv,0,0,NvST ART ,TvST ART + xv,0,0,0,|T |+1 = 1, v∈V (6)

 |T |
 t
−1  
xvitjt = xvjt it , v ∈ V, j ∈ N , t ∈ T (7)
i∈N t=0 i∈N t=t +1

Demand Constraints. Constraints (8) are demand constraints ensuring that

demand is satisfied for each DES long-term contract in each time partition.
The constraints account for the difference in loading and unloading volumes by
subtracting the boil-off.

   
Djp ≤ Lv (1 − (t − t)E)xvitjt
v∈Vi i∈N L t∈Tv t ∈Tjp
 
+ gitj (1 − TijC E) ≤ Djp , j ∈ N U ∪ N S, p ∈ Pj
i∈N L t∈(Tjp −Tij
C)

(8)
Constraints (9) and (10) are the FOB contract constraints. Constraints (9)
make sure that each long-term FOB cargo contract is fulfilled once within its time
window, while Constraints (10) make sure each that spot FOB cargo contract is
fulfilled at most once in the relevant time window, except for the artificial spot
FOB contracts (contract 1), which can be fulfilled as many times as wanted.

zf t = 1, j ∈ N L , f ∈ FjU
F OB
(9)
t ∈Tf

zf t ≤ 1, j ∈ N L , f ∈ FjS \{1}
(10)
t ∈TfF OB
 Planning LNG Annual Delivery Programs 179

Berth and Charter Constraints. Constraints (11) are the berth capacity
constraints. Constraints (12) ensure that if a charter vessel loads a cargo, the
amount loaded is bounded within an upper and lower limit. This allows for some
flexibility in the capacities of the charter vessels.

t +T O
   vij 
xvitjτ + wjt j 
v∈V P i∈N t∈T τ =t +1 j  ∈N U
(11)
t +TfOF OB
  j

+ zf τ ≤ Bjt , j ∈ N L, t ∈ T U
f ∈Fj τ =t +1

C
LC witj ≤ gitj ≤ L witj , i ∈ N L, t ∈ T L, j ∈ N U ∪ N S (12)

Variable Domains. Constraints (13) to (17) specify the domains of our three
binary variables xvitjt , zf t and wvitjt as well as the continuous variables gitj and
sit . The binary variable zf t is only defined for loading ports and the relevant
time periods. The binary decision variable for using a charter vessel, variable
witj , is only defined for arcs that go from a loading node to an unloading node.

xvitjt ∈ {0, 1}, v ∈ V, ((i, t), (j, t )) ∈ Av (13)

zf t ∈ {0, 1}, j ∈ N L , f ∈ Fj , t ∈ TfF OB (14)

witj ∈ {0, 1}, i ∈ N L, t ∈ T L, j ∈ N U ∪ N S (15)

gitj ≥ 0, i ∈ N L, t ∈ T L, j ∈ N U ∪ N S (16)

sit ≥ 0, i ∈ N,t ∈ T L (17)

4 Computational Study
The MIP model (including the network generation) was implemented in Python
and solved by Gurobi (version 10.0.0) on a Dell Optiplex 7780 computer with an
Intel Core i7-10700 @ 2.90 GHz. In the following, we describe the test instances
used in Sect. 4.1, before we present the computational results in Sect. 4.2 and
provide some managerial insights in Sect. 4.3.
180 H. V. Haug et al.

4.1 Input Data and Test Instances

A number of realistic test instances are generated based on data for two differ-
ent LNG producers, one with only one liquefaction (loading) port, and one with
two. Table 1 presents the different configurations of test instances, varying in the
number of loading- and unloading ports |N L | and |N U |, long-term FOB con-
tracts |F U |, vessels |V|, and time periods |T |. Three different test instances were
generated for each configuration, slightly differing in the number of unloading
ports (customer ports) and long-term FOB contracts.
The black horizontal line in Table 1 splits the configurations into two sets,
where the upper set of configurations (with a prefix ‘n’) is based on data for the
LNG producer with one loading port, while the set at the bottom (with a prefix
’a’) is based on data from the producer with two loading ports. The leftmost
column represents the configuration ID, where each of the numbers indicates the
corresponding size of the set on the corresponding position in the instance ID.
For example, configuration n-2-6-12-11-60 has two loading ports, six unloading
ports, 12 FOB long-term contracts, 11 vessels, and 60 time periods of one day.

4.2 Computational Results

A summary of the configurations’ optimality gaps and objective function values

are presented in Table 2 (averaged over the three instances for each configura-
tion). Columns 2–4 show the optimality gaps after 300, 3600, and the maximum
run time of 10 800 s, respectively. A gap of 0.0% means that a guaranteed optimal
solution was found. The next column presents the instances’ average objective
function values after running the commercial solver to the maximum time limit
of 10 800 s. The rightmost column, TTFFS, indicates how much time it takes for
the solver to find the first feasible solution. Note that we were not able to find
any feasible solution within the time limit for instance n-1-17-75-23-365.
We see from the results in Table 2 that once the solver finds a feasible solution
to an instance, the corresponding optimality gap gets reasonably small after a
relatively short amount of time. Furthermore, we see that the solver is able to
find good feasible solutions with small optimality gaps to instances with planning
horizons up to 180 days within minutes, which is very promising. To make some
comparison, test instance B4 in [10] (16 vessels, 121 days) is close in size to
instances in our configuration n-1-7-27-23-120. [10] obtains a gap of 10.63%
for instance B4 after 24 h. It should be noted though that [10] minimizes the
sum of costs minus revenues from the sale of spot, leading to a relatively lower
objective value than our model for same-sized problems. The optimality gaps
thereby seem larger measured in percent. Due to the time-varying LNG prices
in our model, we were not able to remove revenues from long-term DES and FOB
contracts in our objective function to better compare the gaps with ones in [10].
Furthermore, it is worth mentioning that the both MIP solvers and computers
have improved over the last 12 years since [10], which also can explain some of
the differences in gaps.
 Planning LNG Annual Delivery Programs 181

Table 1. Overview of the different test instances’ problem sizes, differing in the number
of loading ports |N L |, average number of unloading ports |N U |, average number of
long-term FOB contracts |F U |, number of vessels |V| and number of time periods |T |.

Configuration ID |N L | |N U | |F U | |V| |T |
n-1-2-2-18-30 1 2 2 18 30
n-1-6-16-18-45 1 6 16 18 45
n-1-6-16-18-60 1 6 16 18 60
n-1-6-10-18-75 1 6 20 18 75
n-1-7-21-18-90 1 7 21 18 90
n-1-7-27-23-120 1 7 23 23 120
n-1-7-40-23-180 1 7 40 23 180
n-1-17-75-23-365 1 17 75 23 365
a-1-6-12-11-60 1 6 12 11 60
a-2-6-12-11-60 2 6 12 11 60
a-2-6-38-15-180 2 6 38 15 180

Table 2. Summary of the configurations’ optimality gaps and objective function values
averaged over three instances for each configuration.

Configuration ID Optimality Gap [%] Objective value TTFFS

300 s 3600 s 10 800 s [mill. USD] [s]
n-1-2-2-18-30 0.0 0.0 0.0 73.1 0
n-1-6-16-18-45 0.40 0.30 0.27 426.3 3
n-1-6-16-18-60 0.39 0.23 0.23 668.1 12
n-1-6-10-18-75 0.37 0.26 0.19 675.7 12
n-1-7-21-18-90 0.34 0.21 0.17 867.7 23
n-1-7-27-23-120 0.71 0.51 0.48 1 168 58
n-1-7-40-23-180 1.55 0.42 0.22 1 615 240
n-1-17-75-23-365 - - - - -
a-1-6-12-11-60 1.63 1.32 1.30 263.1 33
a-2-6-12-11-60 2.27 1.70 1.48 294.4 26
a-2-6-38-15-180 1.11 1.10 0.42 1 400 154

Even though one should be careful with comparisons like this, the low gaps
from running our model indicates that our modeling approach is efficient and
can be used to find good solutions for all practical purposes. It should be noted
though that we were not able to find feasible solutions for the largest test instance
with 365 days planning horizon. However, the results show that our model could
be embedded in a rolling horizon heuristic, e.g., like in [5,10], for obtaining
practical solutions also for instances with longer planning horizons.
182 H. V. Haug et al.

4.3 Managerial Insights: Effect of Speed Optimization

To test the effect of speed optimization, we compare the solutions obtained when
including speed optimization and when using only the vessels’ service speeds.
When running these tests, we noticed that the number of binary variables more
than doubled when including speed optimization since it generates more arcs in
the network. As a representative example, Table 3 shows the breakdown of the
objective function for two cases of an instance from configuration n-1-7-27-23-
120, i.e., without (1) and with (2) speed optimization. The objective function is
broken down into six terms. Our objective function 1 has nine terms but since
there were no spot contracts in this particular instance, three terms have been
omitted from further study in this section.

Table 3. Objective function breakdown for an instance of configuration n-1-7-27-23-

120 (run time 1800 s, optimality gap 0.24% for case 1 and 0.19% for case 2).

Configuration ID 1: With default speed 2: With speed optimization Change [%]

Revenue, DES long-term contracts, own vessels 747.7 815.8 +9.1
Revenue, DES long-term contracts, chartered 181.8 116.8 -35.7
Revenue, FOB long-term contracts 304.3 304.3 0
Revenue, tank left-over LNG 1.6 2.1 +23.1
Costs, own vessels 79.9 90.1 +12.9
Costs, chartering vessels 38.9 22.9 -41.2
Total profit 1 117 1 126 +0.83

We know that each of the two cases in Table 3 has the same lower demand
limits and LNG prices regardless of speed configuration. Therefore, the increase
in DES revenue comes from delivering amounts of LNG that are closer to the
upper demand limits of the customers. Overall, the DES revenue (the first two
terms) increase by 3.1 mill. USD compared to the configuration without speed
optimization, which is an increase of 0.34%. This is possible including speed
optimization allows for sailing at higher speeds so that the producer can deliver
somewhat larger volumes of LNG to the customers. The higher sailing speeds
also results in higher fuel costs for the own vessels (increase of 12.9%). Charter
costs decrease by 35.7% in the case of speed optimization since the producer’s
own fleet is able to transport more due to the higher sailing speeds. This indicates
that if the charter rates were to increase, speed optimization could turn out even
more valuable than it does in the current test instances.
It should be noted that our test instances do not include spot contracts, nor
does our model place a value on artificial FOB. The effect of speed optimization
is expected to become even stronger if these elements are included since there
would be more potential to exploit additional revenue sources. The additional
cost that any extra revenue would imply is often relatively low, so exploiting
more revenue opportunities is likely to be beneficial.
Figure 2 shows the distribution of the speeds of the vessels in the optimal
solution of instance n-1-7-27-23-120-a. The left violin plot is the case with
 Planning LNG Annual Delivery Programs 183

Fig. 2. Distribution of vessel speeds with- and without speed optimization for instance
n-1-7-27-23-120-a. The central horizontal line denotes the average speed of the ves-
sels.

speed optimization, while the right plot is the case with default speeds. The plots
show that the speed optimization model uses the whole spectrum of allowed
vessel speeds, with a slight peak at around 14 and 15 knots. The average is
slightly lower than the default service speeds, which are concentrated around
16 and 17 knots. This would imply a lower fuel cost per distance traveled on
average.

5 Concluding Remarks
We have considered the planning Annual Delivery Programs (ADPs) in the Liq-
uefied Natural Gas (LNG) industry, in which an LNG producer must fulfill a
series of long-term contracts with customers all over the world with a given
fleet of LNG vessels. We have expanded previous definitions and models for this
problem by including speed optimization and the possibility of having multiple
loading ports, which is becoming relevant for some producers as the industry
develops. We denote this planning problem as the LNG-ADP with Speed Opti-
mization and Multiple Load Ports (LNG-ADP-SO-MLP).
Our contribution to the literature is a rich formulation of the LNG-ADP
planning problem, modeled as a mixed-integer programming (MIP) model. The
model is designed to handle each leg of a voyage separately (because of the
multiple loading ports), rather than a round trip (as in the previous models
in the literature). This also allows for efficient modeling of vessel maintenance,
184 H. V. Haug et al.

as the vessel can sail directly to a maintenance port after unloading without
returning to a loading port.
We used a commercial MIP-solver to test our model on a number of realistic
instances for two different LNG producers (one with two loading ports). The
computational study showed that the commercial solver was able to find feasible
solutions with less than 1% optimality gaps within one hour for instances with a
planning horizon of up to 180 days. It was also shown that including speed opti-
mization brings some extra flexibility and increases the profit. The commercial
solver did not find a feasible solution for the largest instance with a full planning
horizon of a year. However, the promising results for the instances with shorter
planning horizons show that our model can be embedded into a Rolling Hori-
zon Heuristic (RHH) in order provide practical solutions also for instances with
longer planning horizons.

References
1. Al-Haidous, S., Msakni, M.K., Haouari, M.: Optimal planning of liquefied natural
gas deliveries. Transp. Res. Part C: Emerg. Technol. 69, 79–90 (2016)
2. Andersson, H., Christiansen, M., Desaulniers, G., Rakke, J.G.: Creating annual
delivery programs of liquefied natural gas. Optim. Eng. 18(1), 299–316 (2017)
3. FreightWaves: LNG shipping rates ‘shooting for the stars’ at $500,000 per day.
https://www.freightwaves.com/news/lng-shipping-rates-shooting-for-the-stars-
at-500000-per-day
4. GIIGNL: GIIGNL Annual Report 2022. The International Group of Liquefed Natu-
ral Gas Importers, Neuilly-sur-Seine, France. https://giignl.org/resources2/ (2022)
5. Li, M., Fagerholt, K., Schütz, P.: Maritime inventory routing with transshipment:
the case of Yamal LNG. Flex. Serv. Manuf. J. (2022). https://doi.org/10.1007/
s10696-022-09476-5
6. Li, M., Schütz, P.: Planning annual LNG deliveries with transshipment. Energies
13(6), 1490 (2020)
7. Msakni, M.K., Haouari, M.: Short-term planning of liquefied natural gas deliveries.
Transp. Res. Part C: Emerg. Technol. 90, 393–410 (2018)
8. Mutlu, F., Msakni, M.K., Yildiz, H., Sönmez, E., Pokharel, S.: A comprehensive
annual delivery program for upstream liquefied natural gas supply chain. Eur. J.
Oper. Res. 250(1), 120–130 (2016)
9. Norstad, I., Fagerholt, K., Laporte, G.: Tramp ship routing and scheduling with
speed optimization. Transp. Res. Part C: Emerg. Technol. 19(5), 853–865 (2011)
10. Rakke, J., et al.: A rolling horizon heuristic for creating a liquefied natural gas
annual delivery program. Transp. Res. Part C: Emerg. Technol. 19(5), 896–911
(2011)
11. Stålhane, M., Rakke, J.G., Moe, C.R., Andersson, H., Christiansen, M., Fagerholt,
K.: A construction and improvement heuristic for a liquefied natural gas inventory
routing problem. Comput. Ind. Eng. 62(1), 245–255 (2012)
12. Wen, M., Ropke, S., Petersen, H.L., Larsen, R., Madsen, O.B.G.: Full-shipload
tramp ship routing and scheduling with variable speeds. Comput. Oper. Res. 70,
1–8 (2016)
 Tramp Ship Routing with Bunker
Optimization and Flexible Cargo
Quantities: Case from Dry Bulk Shipping

Simen Omholt-Jensen1 , Kjetil Fagerholt1 , and Frank Meisel2(B)

1
Department of Industrial Economics and Technology,
Norwegian University of Science and Technology, Trondheim, Norway
[email protected], [email protected]
2
Faculty of Business, Economics and Social Science, Kiel University, Kiel, Germany
[email protected]

Abstract. We study the Tramp Ship Routing and Scheduling Prob-

lem with Bunker Optimization (TSRSPBO). The TSRSPBO includes,
in addition to the routing and scheduling of the vessels in the given
fleet, decisions about how much to bunker (refuel) in which ports. Fur-
thermore, the cargo quantities to be transported are given in an inter-
val, which means that determining the optimal transport quantities also
becomes a decision. We propose an arc flow model and a solution method
based on a reformulation into a path flow model that uses all feasible ves-
sel routes (columns) as input. For each route, we determine the optimal
bunker amounts and cargo quantities along the route. The path flow
solution method is tested on a set of instances generated based on data
from our case company. The computational results show that the path
flow solution method outperforms the arc flow model solved by a com-
mercial solver. We also present how chartering managers may use this
as decision support in the negotiation of freight contracts and bunker
prices.

Keywords: Ship routing and scheduling · Bunkering optimization ·

Dry bulk shipping

1 Introduction
Maritime transportation is the backbone of international trade, with more than
80% of the total volume traded internationally [12]. One important segment is
dry bulk shipping, which concerns transporting raw materials shipped in bulk,
such as iron ore, coal, and grain. In this paper, we study a real-life tramp ship
routing and scheduling problem in which a dry bulk shipping operator is to
deploy its vessel fleet to maximize the profit for the entire fleet. To do so, the
operator has to select from sets of mandatory contracted cargoes and optional
spot cargoes, potentially paying a cost for not servicing a contracted cargo. In dry
bulk shipping, an operator can often choose the amount of cargo to transport
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 185–201, 2023.
https://doi.org/10.1007/978-3-031-43612-3_11
186 S. Omholt-Jensen et al.

within a predetermined interval. Thus, the operator also needs to decide how
much of each cargo to transport, which can be challenging given that the vessels
may carry multiple cargoes simultaneously. Furthermore, as the ships eventually
run out of bunker (fuel), the operator must choose where to bunker and the
quantity of bunker to purchase. We denote this planning problem as the Tramp
Ship Routing and Scheduling Problem with Bunker Optimization (TSRSPBO).
There have been significant contributions to the academic field of tramp
ship routing and scheduling over recent years, summarized in the review papers
[6,9,11]. The aspect of flexible cargo quantities in ship routing is considered by
[2,7,8], while bunker optimization was included by [10,13]. However, none of
the previous studies integrates these two important aspects. Furthermore, the
latter two studies do only consider full shiploads, meaning that a vessel can have
at most cargo on board at the same time, in contrast to the problem we con-
sider in this paper. Hence, to our knowledge, the TSRSPBO studied here is the
first to integrate both flexible cargo quantities and bunker optimization. Fur-
thermore, by allowing vessels to detour to refuel bunker and the simultaneous
service of multiple cargoes, the TSRSPBO realistically represents our case com-
pany Western Bulk’s operational environment. We propose an arc flow model
for the TSRSPBO, and to solve instances of larger sizes, we propose a solution
method based on a path flow model with routes (columns) generated a priori.
For each of these routes, we have to determine the optimal cargo and bunker
quantities along the route, which we do by solving a linear programming model.
By solving test instances based on real-life data, we show how our models and
proposed solution method may act as an aid to chartering managers in dry bulk
shipping companies.
The remaining of this paper is organized as follows. Section 2 provides a
definition of the TSRSPBO, while Sect. 3 presents its arc flow model. Then, in
Sect. 4, we propose the path flow solution method. The computational study is
provided in Sect. 5, while concluding remarks are given in Sect. 6.

2 Problem Definition
This section formally defines the TSRSPBO, which is relevant to many shipping
companies operators, e.g., in the dry bulk shipping industry. The TSRSPBO
considers a ship operator controlling a given heterogeneous fleet of vessels with
differing cargo and bunker carrying capacities, as well as initial positions and
bunker levels. Each vessel has a given sailing speed and fuel consumption.
The vessel fleet is set up to transport a set of cargoes for the following plan-
ning horizon, e.g., the next two months. Each cargo is specified by its cargo
type, pickup and delivery ports and time windows, and freight rate. A cargo
can either be a contractual mandatory cargo or a spot cargo. A contractual
cargo usually stems from a long term Contract of Affreightment (CoA), and is
therefore denoted a CoA cargo. A spot cargo is an optional cargo for which the
ship operator not yet has signed a contract. In dry bulk shipping it is common
that the cargo contracts offer some flexibility to the ship operator regarding the
 Tramp Ship Routing with Bunker Optimization 187

quantities to be transported through a contractual term called More or Less

(ship-)Owners Option (MoLOO). A MoLOO term states that the cargo quan-
tity can be selected from within a given range. For example, the quantity range
may be defined as 66,000 metric tonnes with 10% MoLOO. This means that the
operator can choose to transport any quantity between 59,400 to 72,600 metric
tonnes. The times spent in port for loading and unloading then depend on the
chosen cargo quantity. A cargo’s pickup (delivery) time windows define the ear-
liest and latest time for when the loading (unloading) of the cargo can start in
the pickup (delivery) port.
In addition to the ship operator’s fleet, spot ships can be chartered on voyage
charter contracts to service the CoA cargoes. In addition to the cargoes’ pickup
and delivery ports, there is a predefined set of available bunker ports, from which
the ships in the fleet can bunker fuel. The bunker ports have different fuel prices,
which are assumed to be known and constant over the planning horizon.
A vessel will start its route from a given initial position. A vessel route
consists of a sequence of ports, which may be pickup, delivery, and bunker ports.
Depending on the cargo quantities and vessel capacities, a vessel might have the
possibility to transport multiple cargoes simultaneously. For each sailing leg
between two ports, there is an associated vessel-dependent sailing time and cost.
Ships traveling through a canal will have to pay an additional canal cost.
The main decisions of the TSRSPBO are the routes and schedules of each
vessel in the operator’s fleet. This implicitly incorporates the following addi-
tional decisions: 1) which spot cargoes to service, 2) which vessels should service
which cargoes, 3) which CoA cargoes should be serviced by spot vessels, and
4) which port the vessels should stop for bunkering. Furthermore, we need to
determine the quantity of each cargo along the vessels’ routes (within its limits),
as well as the bunker amounts. All these decisions impact each other and must
be determined simultaneously.
All transported cargoes must be picked up within their pickup time windows
and delivered within their delivery time windows. The cargo capacity limit of
each vessel must be respected along its route, and its bunker level must be above
a safety limit and not exceed the bunker capacity limit.
The overall objective of the TSRSPBO is to maximize the vessel fleet’s profit
over the planning period. Profit is the generated revenue of the transported
cargoes less the sum of all variable costs. The variable costs consists of the fuel,
port, and canal costs, as well as voyage chartering costs for spot vessels.

3 Arc Flow Model

This section presents the arc flow (AF) model for the TSRSPBO. We call this
model an AF model because we here use arc flow variables specifying which arcs
the vessels traverse within a network representation of the problem. Section 3.1
describes the modeling assumptions and notation, while Sect. 3.2 presents the
AF model.
188 S. Omholt-Jensen et al.

3.1 Modeling Assumptions and Notation

In Sect. 2, it was stated that there is a delivery time window for each cargo. In
reality, it is often only a pickup time time window, while it is specified that deliv-
ery should occur at the ”earliest convenience”. Thus, to exclude the possibility
of a vessel carrying a cargo for too long time, artificial delivery time windows are
constructed by multiplying the direct sailing time between the cargo’s pickup
and delivery ports and port time with a factor, e.g., 1.5. This design choice
allows detours to purchase bunker and pick up other cargoes while ensuring the
delivery of all cargoes within a reasonable time.
Furthermore, we assume that the time it takes to load bunker is not depen-
dent on the amount. This assumption is in line with [13]. A fixed bunker time
is therefore added to the sailing time for trips arriving at a bunker port. At
the beginning of the planning period, vessels might be in port or at sea as they
operate 24/7. We assume that each vessel becomes available at a given port, rep-
resented by an origin node, corresponding to the discharge port of their ongoing
voyage. After completing the last port call in the planning horizon, a vessel ends
its route by traveling to an artificial destination node at zero cost and time.
The vessel’s bunker load at this destination node is considered a resource, so the
difference in bunker load between the origin and destination node is priced at
an average price for all the bunker ports. The value of this additional bunker is
reflected in the model’s objective.
The notation and formulation presented here is based on the arc flow formu-
lation presented in [4,5,13]. Let the fleet of vessels be represented by the set V.
There are N cargoes available for transport in the following planning horizon,
where the sets of pickup and delivery nodes are given by N P = {0, . . . , N } and
N D = {N + 1, . . . , 2N }, respectively. These sets allow each cargo i to be repre-
sented by a pickup node i ∈ N P and a delivery node (N + i) ∈ N D . As such, the
set of all cargo-related nodes is represented as N = N P ∪ N D . The set of pickup
nodes N P is partitioned into N P = N C ∪ N O , where N C and N O denote the
mandatory contracted and optional spot cargoes, respectively. A vessel v has a
node of origin o(v), while it is assigned an artificial destination node d(v). Due to
cargo capacity, depth, and cargo handling equipment restrictions, some vessels
in V and nodes in N may be incompatible. Thus, the set of nodes a vessel v may
visit is represented as Nv ⊆ N ∪ {o(v), d(v)}. (Nv , Av ) defines a base network
for each vessel v, where the set of arcs Av ⊆ {(i, j)|i ∈ Nv , j ∈ Nv } represents
all arcs traversable by vessel v with respect to time, capacity and bunker con-
sumption. The pickup and delivery nodes compatible with vessel v are defined
as NvP = N P ∩ Nv and NvD = N D ∩ Nv .
The next step is extending the base network (Nv , Av ) to allow vessel visits
at bunker nodes (ports). Let Bv be the set of bunker nodes that vessel v may
visit. The set of base nodes Nv for vessel v is extended to include the nodes in
Bv such that N̂v = Nv ∪ Bv . Furthermore, the set of base arcs Av is extended
by adding all feasible arcs connecting nodes in Nv \d(v) with nodes in Bv . As
such, the extended set of arcs is defined as Âv = Av ∪ AB v , where Av is the set
B
 Tramp Ship Routing with Bunker Optimization 189

of arcs for vessel v connecting bunker option nodes Bv to the nodes in Nv . The
extended cargo-bunker network is thereby (N̂v , Âv ).
S
Let Tijv be the sailing time between nodes i and j for vessel v. If node j is a
S
bunker node, a fixed time is added to Tijv , signifying the time spent bunkering.
The time required to load or discharge one unit of cargo i with vessel v is
Q
defined as Tiv . The variable voyage cost Cijv accounts for the costs of visiting
node i and the sailing from node i to node j for vessel v. Note that the cost
of the purchased bunker is not included in this parameter as the purchase of
bunker is modeled separately. Instead, let Pi denote the cost of purchasing one
unit of bunker available in bunker node i. The bonus bunker remaining at the
destination node d(v) is rewarded by P per unit (calculated as the mean of all
available bunker node prices). The voyage charter cost of servicing a CoA cargo
i by a spot ship is denoted CiS . Furthermore, the bunker consumption for vessel
v traversing arc (i, j) ∈ Âv is denoted BijvS
. The bunker consumed per time unit
P
in port i by vessel v is given by Biv .
 The quantity of cargo that can be transported is flexible within interval
Qi , Qi . There is a per unit revenue Ri for transporting cargo i ∈ N P and
 
a time window T iv , T iv for when service must start at node i (strengthened
for each vessel v). Additionally, the bunker level on board each vessel v must
be within a lower limit B v and an upper limit B v to ensure realistic and safe
operation. Finally, the cargo carrying capacity of vessel v is denoted Kv .
The binary arc flow variable xijv is assigned the value 1 if vessel v traverses
arc (i, j) ∈ Âv , and 0 otherwise. The variable tiv denotes the time vessel v begins
service at node i. The quantity of cargo i transported by vessel v is represented
by qiv , and the quantity of bunker purchased by vessel v at bunker node i is
given by the variable biv . The total cargo load on board a vessel v when it leaves
C B
node i is represented by the variable liv . Similarly, the variable liv represents the
total bunker load on board vessel v just after completing service at node i. To
represent whether a CoA cargo is serviced by a spot ship, the binary variable zi
is assigned the value 1, or 0 otherwise. Finally, the binary variable yi is assigned
a value of 1 if the optional spot cargo i is serviced, and 0 otherwise.

3.2 Arc Flow Formulation

Using the above notation, we can now formulate the AF model as follows.

Objective Function
    
max Ri qiv − Cijv xijv − CiS zi
v∈V i∈N P v∈V (i,j)∈Âv i∈N C
  (1)
− Pi biv + P · (ld(v)v
B
− Bv0 )
v∈V i∈Bv v∈V

The objective function (1) maximizes the total profit. The first term specifies
the revenue generated for both contracted and optional cargoes, while the second
190 S. Omholt-Jensen et al.

term specifies the variable voyage costs. The third term represents the voyage
charter costs associated with servicing CoA cargoes with spot ships. The fourth
term describes the cost associated with purchasing bunker, while the last term
calculates the value of the bonus bunker remaining at the destination node. Note
that if the bunker level at the destination is higher than at the origin, this term
will be positive.

Network Flow Constraints

 
xijv + zi = 1 i ∈ NC (2)
v∈V j∈N̂v
 
xijv − yi = 0 i ∈ NO (3)
v∈V j∈N̂v

xo(v)jv = 1 v∈V (4)
j∈N̂v
 
xijv − xjiv = 0 v ∈ V, i ∈ N̂v \{o(v), d(v)} (5)
j∈N̂v j∈N̂v

xid(v)v = 1 v∈V (6)
i∈N̂v
 
xijv − xN +i,jv = 0 v ∈ V, i ∈ NvP (7)
j∈N̂v j∈N̂v

Constraints (2) ensure that all mandatory CoA cargoes are transported by a
vessel in the predefined fleet or by a spot ship. In contrast, Constraints (3) state
that all spot cargoes may be transported. Constraints (4) ensure that all vessels
begin their routes from their origin node and travel to a pickup node, a bunker
option node, or directly to their destination node. Constraints (5) denote the
flow conservation, while Constraints (6) force each vessel’s route to end at its
artificial destination node. Constraints (7) ensure that the same vessel services
both the pickup and the cargo’s corresponding delivery node.

Temporal Constraints
 
Q
xijv tiv + Tiv S
qiv + Tijv − tjv ≤ 0 v ∈ V, (i, j) ∈ Âv
|i ∈ NvP (8)
 
Q
xijv tiv + Tiv S
qi−N,v + Tijv − tjv ≤ 0 v ∈ V, (i, j) ∈ Âv
|i ∈ NvD (9)
xijv tiv + S
Tijv − tjv ≤ 0 v ∈ V, (i, j) ∈ Âv
|i ∈
/ NvP ∧ i ∈
/ NvD (10)
Q
+i,v − tN +i,v ≤ 0 v ∈ V, i ∈ NvP
S
tiv + Tiv qiv + Ti,N (11)
 Tramp Ship Routing with Bunker Optimization 191

T iv ≤ tiv ≤ T iv v ∈ V, i ∈ N̂v (12)

Constraints (8)–(9) specify the time progression for the pickup and delivery
nodes, respectively. Note, however, that the cargo quantity variable qiv is only
defined for pickup nodes. Constraints (10) describe a similar time flow for origin,
bunker, and destination nodes as there is no cargo loading or unloading at these
nodes. Constraints (11) are precedence constraints forcing the pickup node to
be serviced before its corresponding delivery node. Finally, Constraints (12) are
the time window constraints.

Cargo Constraints
C
xijv liv + qjv − ljv
C
=0 v ∈ V, (i, j) ∈ Âv
|j ∈ NvP (13)
C
xi,N +j,v liv − qjv − lN
C
+j,v = 0 v ∈ V, (i, N + j) ∈ Âv
|j ∈ NvP (14)
C
xijv liv − C
ljv =0 v ∈ V, (i, j) ∈ Âv
|j ∈ Bv (15)
C
lo(v),v =0 v∈V (16)

qiv ≤ liv
C
≤ Kv xijv v ∈ V, i ∈ NvP (17)
j∈N̂v

0 ≤ lN
C
+i,v ≤ (Kv − qiv ) xN +i,jv v ∈ V, i ∈ NvP (18)
j∈N̂v
 
Qi xijv ≤ qiv ≤ Qi xijv v ∈ V, i ∈ NvP (19)
j∈N̂v j∈N̂v

Constraints (13)–(15) represent the cargo load balance constraints for pickup,
delivery and bunker nodes, respectively. Constraints (16) initialize the cargo load
on board the vessels to zero. Constraints (17) and (18) describe the cargo load
capacity restrictions at pickup nodes and delivery nodes, respectively. Finally,
Constraints (19) ensure that the cargo loaded at pickup port i is within the
minimum and maximum cargo load limits. The summations in Constraints (17)–
(19) are introduced to strengthen the formulation.

Bunker Constraints
 
P Q
B
xijv liv − Bijv
S
− Bjv Tjv qjv − ljv
B
=0 v ∈ V, (i, j) ∈ Âv
|j ∈ NvP (20)
 
P Q
B
xijv liv − Bijv
S
− Biv Tjv qj−N,v − ljv
B
=0 v ∈ V, (i, j) ∈ Âv
|j ∈ NvD (21)
192 S. Omholt-Jensen et al.

B
xijv liv − Bijv
S
+ bjv − ljv
B
= 0 v ∈ V, (i, j) ∈ Âv
|j ∈ Bv (22)
B
xijv liv − Bijv
S
− ljv
B
= 0 v ∈ V, (i, j) ∈ Âv
|j ∈ {d(v)} (23)
B
=
lo(v)v Bv0 v ∈ V (24)
  
P Q
S
B v + Bijv + Bjv Tjv qjv xijv ≤ liv
B
≤ Bv xijv v ∈ V, i ∈ N̂v (25)
j∈N̂v j∈N̂v

biv ≤ B v − B v xijv v ∈ V, i ∈ Bv (26)
j∈N̂v

Constraints (20)–(23) describe the bunker balance constraints for the pickup,
delivery, bunker and destination nodes, respectively. Constraints (24) specify
the bunker level on board vessel v at the beginning of the planning period. Con-
straints (25) make sure that the bunker loads on board the vessels are kept within
its limits, while Constraints (26) ensure that the amount bunker purchased is
feasible. The summations in (25) and (26) are again included to strengthen the
formulation.

Variable Domains

xijv ∈ {0, 1} v ∈ V, (i, j) ∈ Âv (27)

yi ∈ {0, 1} i∈N O
(28)
zi ∈ {0, 1} i∈N C
(29)
C B
tiv , liv , liv ∈R
+
v ∈ V, i ∈ N̂v (30)
qiv ∈ R+ v ∈ V, i ∈ NvP (31)
biv ∈ R +
v ∈ V, i ∈ Bv (32)

In the above arc flow model, the balance constraints for time, cargo, and
bunker are non-linear. The same is the case for Constraints (18), handling the
discharge load capacities of the vessels, as well as for Constraints (25), controlling
the bunker load capacities of the vessels. These constraints are all linearized,
using standard linearization techniques with big-M constraints (e.g., [4]), to make
the model solvable by a commercial MIP-solver.
 Tramp Ship Routing with Bunker Optimization 193

4 Path Flow Solution Method

Since a commercial MIP-solver can only solve very small instances of the arc
flow model presented in Sect. 3.2, we propose a path flow solution method. By
inspecting the arc flow model, it is clear that only Constraints (2) and (3) are
fleet-specific, while the remaining constraints, as well as the objective function,
are vessel-specific with no interaction among the vessels. As such, the struc-
ture of the problem lends itself to Dantzig-Wolfe decomposition, in which the
vessel-specific constraints concerning the routing and scheduling can be handled
separately in a subproblem for each vessel. The fleet-specific constraints can
thus be placed in a path flow formulation (or a master problem) with a reduced
number of constraints but with a large number of columns. Here, we propose
a solution method based on such a path flow model, presented in Sect. 4.1, for
which the columns and parameters are generated through an a priori generation
of all feasible columns, i.e., vessel routes, as described in Sect. 4.2.

4.1 Path Flow Formulation

Let Rv be a feasible route for vessel v. In this setting, feasible means a route
adhering to Constraints (4)–(26). Rv denotes the set of all such feasible routes
for vessel v. For each feasible route Rv for vessel v, the optimal amounts of
the cargoes to be transported and bunker to be purchased are found as later
described in Sect. 4.2. We refer to such an optimized route as a schedule s since
it also includes information on the timing of the activities along the route. Each
such schedule is associated with an optimal vessel-specific profit, denoted Psv .
Let Sv be the set of all optimal schedules for vessel v. The new parameter, Aisv ,
is defined to be equal to 1 if vessel v carries cargo i in schedule s, and 0 otherwise.
The binary variable xsv takes the value of 1 if vessel v is chosen to sail schedule
s, and 0 otherwise. Reusing some of the notation defined in Sect. 3.1, we can now
define the path flow model as follows:

  
max Psv xsv − CiS zi (33)
v∈V s∈Sv i∈N C
 
Aisv xsv + zi = 1 i ∈ NC (34)
v∈V s∈Sv
 
Aisv xsv − yi = 0 i ∈ NO (35)
v∈V s∈Sv

xsv = 1 v∈V (36)
s∈Sv

xsv ∈ {0, 1} v ∈ V, s ∈ Sv (37)

Objective function (33), corresponding to objective function (1) in the arc

flow model, maximizes the fleet profit, which consists of the vessel-specific profit
of the selected schedules less the cost of hiring a spot ship to service a CoA
194 S. Omholt-Jensen et al.

cargo. Constraints (34) state that each contracted cargo is either transported by
a vessel in the fleet or by a spot ship, similar to Constraints (2) in the arc flow
model. Constraints (35) specify that optional cargoes may be serviced by a vessel
in the fleet, similar to Constraints (3). Constraints (36) ensure that each vessel
is assigned exactly one schedule. Finally, Constraints (37) define the domains of
the binary schedule assignment variables.

4.2 A Priori Column Generation with Cargo and Bunker

Optimization

First, all routes are generated using a depth-first search for each vessel. Here,
we use the minimum cargo quantities Qi to ensure that all feasible routes are
generated, and we also make sure that they adhere to the vessel-specific con-
straints in the arc flow model. Given the set of all feasible routes, we solve a
Vessel Scheduling Problem (VSP) to determine the optimal amounts of bunker
and the cargo quantities along the route (together with timing of the activities).
It should be noted that, since the route is now given, the VSP does not
have any binary variables and is a Linear Programming (LP) model, which is
considerably easier to solve than the whole AF model. Solving the VSP for each
feasible route for each vessel yields the set of optimal schedules Sv for each vessel
v. The objective function value associated with each optimal schedule s is the
vessel-specific optimal profit Psv for route s used in the path flow formulation.

5 Computational Study
The arc flow and the path flow models were implemented in C++ and solved by
Gurobi (version 10.0.0). All of the computational experiments were performed on
a MacBook Pro with an Apple M1 Pro processor. In the following, we describe
the test instances used in Sect. 5.1, before we present the computational results
in Sect. 5.2 and provide some managerial insights in Sect. 5.3.

5.1 Input Data and Test Instances

This section explains how real-life cargo and voyage data from the industry
partner, Western Bulk1 , was compiled into suitable test instances. In addition
to Western Bulk’s operational data, routing and port information were pro-
vided by Maritime Optima2 , a maritime data analytics company. The remaining
input data needed to generate realistic test instances was either gathered or con-
structed from open sources, estimated from available data, or randomly sampled
from reasonable uniform intervals.
We generated a set of test instances of different size classes, as summarized
in Table 1. Within each class, we generated five instances. For each test instance,

1
www.westernbulk.com.
2
www.maritimeoptima.com.
 Tramp Ship Routing with Bunker Optimization 195

Table 1. Classes of generated test instances

Name Size # Cargoes # Vessels # Bunker nodes

C6V2B4 Small 6 2 4
C9V2B4 Small 9 2 4
C9V3B4 Small 9 3 4
C15V5B10 Medium 15 5 10
C30V5B10 Medium 30 5 10
C30V10B10 Large 30 10 10
C45V10B10 Large 45 10 10

the number of spot cargoes was fixed to be twice the number of CoA cargoes.
As such, for a test instance with 15 cargoes, there are five CoA cargoes and 10
spot cargoes. This is representative for Western Bulk’s operational environment.
The cargoes’ MoLOO limits were estimated by creating a ±10% interval
from the actual quantities of historical transported cargoes. The origin ports of
the vessels were randomly drawn from a set of the most frequent delivery ports
visited by Western Bulk’s vessels. A final group of parameters was randomly
sampled from sensible intervals. These include the time when vessels first become
available to transport cargoes, T o(v)v , the beginning of the time windows for
pickup ports, T iv , i ∈ NvP , and the initial bunker level on board the vessels, Bv0 .
For pickup ports, the upper time windows were constructed by adding ten days
to the start, which corresponds to what is typical in Western Bulk’s operation.

5.2 Computational Results and Comparison of Solution Methods

The results for the arc flow model are summarized in Table 2. The column Solved
shows the number of test instances within each class that were solved to opti-
mality for the AF model within the maximum allowed running time, which is set
to one hour. The Time column shows the average solution times in seconds. The
BestBoundAF column displays the best dual bound found when solving the AF
model, averaged over the test instances in each class. Similarly, the ObjValsAF
column states the best primal bounds, i.e., the average profit of the obtained
solutions. The Gap column shows the average percentage difference between
the best primal and dual bounds.
Table 2 shows that only the instances of the first two classes were solved to
optimality within one hour. For the C9V3B4 class, two out of five test instances
were solved to optimality. For the larger instances, the AF model could not
provide an optimal solution within one hour, and the gaps are high. It should
be noted that introducing bunker optimization to the tramp ship routing and
scheduling problem significantly increases the problem complexity. To illustrate
this, the arcs going to and from the bunker nodes constitute as much as 62.5%
of the total number of arcs (on average across all test instances).
196 S. Omholt-Jensen et al.

Table 2. Arc flow (AF) model results averaged over instances in each instance class

Instance Class Size Solved Time BestBoundAF ObjValAF Gap (%)

(seconds) (mill. USD) (mill. USD)
C6V2B4 Small 5/5 86 3.76 3.76 0.00
C9V2B4 Small 5/5 153 5.22 5.22 0.00
C9V3B4 Small 2/5 2,890 9.24 6.75 34.26
C15V5B10 Medium 0/5 3,600 22.3 11.1 104.39
C30V5B10 Medium 0/5 3,600 44.3 12.8 250.77
C30V10B10 Large 0/5 3,600 49.9 27.6 81.56
C45V10B10 Large 0/5 3,600 69.9 26.2 169.93

The results of the path flow solution method are summarized in Table 3,
where the Solved column shows the number of test instances within each class
that were solved to optimality within one hour. The NR column displays the
average number of routes generated for instances within each class. The column
Time shows the solution times in seconds, wheremost of the time is spent in the
generation of the routes and solving the LPs to determine the optimal bunker
amounts and cargo quantities. The NSS and NSC columns state the average
number of spot ships used and the number of spot cargoes transported in the
optimal solutions, respectively. Furthermore, the ObjValTOT column states the
optimal objective values, i.e., the profit, averaged over the test instances in each
class. Finally, the Δ column reports the average percentage improvements in the
objective value found when solving the path flow model compared to the best
objective value found when solving the AF model (after one hour).

Table 3. Path flow model results averaged over the instances in each instance class

Instance Class Size Solved NR Time NSS NSC ObjValTOT Δ

(secs) (mill. USD) (%)
C6V2B4 Small 5/5 816 0.3 1 2 3.76 0.00
C9V2B4 Small 5/5 1370 0.6 1 2 5.22 0.00
C9V3B4 Small 5/5 1581 0.6 1 3 6.76 0.15
C15V5B10 Medium 5/5 90,194 56.3 1 6 12.6 13.51
C30V5B10 Medium 5/5 406,435 416 5 7 15.5 21.09
C30V10B10 Large 5/5 1,312,606 1,516 0 14 31.4 13.77
C45V10B10 Large 2/5 2,095,022 3,708 3 13 34.7 32.44

In general, the path flow solution method clearly outperforms the AF model,
both with respect to solution time and quality. As we can see from Table 3, the
path flow solution method is able to solve all except for three of the largest
instances. Nevertheless, the run time grows quickly along with the number of
feasible routes/columns when the instance size increases. Eventually, this also
results in memory problems, and the largest instances solved here represent more
 Tramp Ship Routing with Bunker Optimization 197

or less the limit of the path flow solution method. For even larger test instances,
either heuristics or more advanced exact methods will be needed.

5.3 Managerial Insights

Our model offers potential for a wide range of analyses that can help managers
in making the addressed decisions. In this section, we concentrate on exploring
how the negotiation of flexible cargo quantities and bunker purchase discounts
may impact profits. There is surely a large range of further factors that could
be analysed, such as fleet diversity, fleet age, or repositioning flexibility (see [1])
but we leave this to future research also as some of these issues may require
extensions of our proposed model.

Impact of Flexible Cargo Quantities: The MoLOO limits are subject to

negotiation between the ship operator and the shippers. As such, ship operators
would benefit from understanding the impact of increased flexibility in cargo
quantity flexibility. Figure 1 shows the effect on profit for different levels flexi-
bility cargo quantities for a test instance in the C15V5B10 class.

Fig. 1. The effect of utilizing flexible cargo quantities for a test instance in C15V5B10

For large parts of Fig. 1, there is a (close to) linear profit increase from
increased flexibility. This comes simply from the possibility of transporting larger
quantities of the same cargoes. However, when the MoLOO limits are increased
from ±3% to ±4% and from ±24% to ±25%, we notice two jumps in the profit.
This is because the additional flexibility makes it possible to reduce the quantities
for some cargoes, and hence be able to transport additional spot cargoes. This
shows that increased contract flexibility in cargo quantities may have a huge
impact on profitability. These results are in accordance with [8].
198 S. Omholt-Jensen et al.

Impact of Bunker Price Negotiations: The bunker prices used in these

tests were retrieved from Maritime Optima’s [3] interface. As such, the prices
are considered market prices and available to anyone who might want to pur-
chase bunker at a given port. In reality, shipping companies devote resources to
acquire favorable bunker purchase deals ahead of time. For example, Western
Bulk has an entire department responsible for negotiating bunker purchase deals
with different ports. Here, we show how we can provide decision support in this
process.
Figure 2 shows the bunker prices and the total number of visits to bunker
ports over all test instances in class C15V5B10. It can be seen that the ports
of Hong Kong, Busan, Lianyungang, Port Elizabeth, and Gladstone are never
visited, even though some of them are offering bunker at very low prices. For
Gladstone and Port Elizabeth this observation is explained by their positioning
in Australia and South Africa, respectively, which is too far for vessels operating
in the Indo-Pacific region as is the case in our instances. Furthermore, neither
of these countries is a major oil producer, which explains their relatively high
prices. However, the other mentioned ports are actually closer located to the
operation areas of the ships, yet they do not play a role in bunkering either. For
those ports, it is less clear what there disadvantages are. They are probably just
marginally worse compared to those ports that are actually visited frequently,
i.e., Al Fujayrah and Singapore.

Fig. 2. Bunker prices and total number of bunker port visits in test instances C15V5B10
 Tramp Ship Routing with Bunker Optimization 199

Fig. 3. Averaged effects of procuring discounted bunker for test instances C15V5B10

Figure 3 shows the increase in profit averaged over the five test instances
in C15V5B10 at different bunker discount rates. The bunker discount rates are
applied one at a time to each of the bunker ports. As expected, increasing the
discount rate leads to increased profit. A bunker price discount at the ports of
Singapore, Al Fujayrah and/or Kandia has the highest impact on the profits.
Their impact on profits is in line with Fig. 2, as these ports are among the most
frequently visited ports. As such, if Western Bulk has to choose which port they
should focus their efforts on procuring a discounted bunker purchase contract,
Fig. 3 suggests either one of these. For the previously discussed unattractive
ports Gladstone and Port Elizabeth, even a bunker price discount of 20% has
(almost) no impact on the solutions. Interestingly, Hong Kong, a port that was
never visited in Fig. 2, displays a greater impact on the average profit for the
test instances than Zhoushan Port, which was visited five times even without a
bunker discount. The explanation is that Hong Kong is more centrally located in
the Indo-Pacific region than Zhoushan Port. Whereas Zhoushan Port is located
further north, closer to Shanghai, Hong Kong lies further south, closer to Sin-
gapore. As such, decreasing the price of bunker in Hong Kong can lead to more
port visits than lowering the bunker price in Zhoushan. Bunker purchase man-
agers in Western Bulk may therefore leverage the information provided in Fig. 3
during bunker purchase negotiations.

6 Concluding Remarks

We have considered a real rich ship routing problem arising in the dry bulk ship-
ping segment, a problem we have denoted Tramp Ship Routing and Scheduling
Problem with Bunker Optimization (TSRSPBO). The TSRSPBO includes, in
200 S. Omholt-Jensen et al.

addition to the routing and scheduling of the given vessels in the fleet, deci-
sions about how much to bunker (refuel) in which ports. Furthermore, the cargo
quantities to be transported in the TSRSPBO are given in an interval (i.e.,
MoLOO limits), which means that determining the optimal transport quantities
also becomes a decision. We presented an arc flow model for the TSRSPBO, but
the commercial MIP solver is only able to solve tiny instances of the problem,
partly because the size of the network increases drastically with the inclusion of
the bunker ports. Hence, we have proposed a path flow solution method based
on a Dantzig-Wolfe decomposition of the problem. This solution method consists
of a path flow model using all feasible vessel routes (columns) as input. For each
such route, we determine the optimal amounts of both the bunker and cargoes
along the route by solving an LP problem.
The solution method was tested on a set of realistic instances generated
based on data from Western Bulk, a dry bulk shipping operator. The computa-
tional study suggests that the path flow solution method outperforms solving the
arc flow model by a commercial solver and that it is able to consistently solve
instances of up to 30 cargoes, 10 vessels, and 10 bunker ports within 30 min.
Furthermore, additional tests show how different MoLOO limits and bunker dis-
counts in different ports may affect the ship operator’s profit. Hence, the ship
operator may use these insights in the negotiation of freight contracts and to
determine for which ports it is most beneficial to negotiate bunker price dis-
counts.

References
1. Bai, X., Cheng, L., Çağatay, I.: Data-driven financial and operational risk man-
agement: empirical evidence from the global tramp shipping industry. Transp. Res.
Part E: Logistics Transp. Rev. 158, 102617 (2022). https://doi.org/10.1016/j.tre.
2022.102617
2. Brønmo, G., Christiansen, M., Nygreen, B.: Ship routing and scheduling with flex-
ible cargo sizes. J. Oper. Res. Soc. 58(9), 1167–1177 (2007). https://doi.org/10.
1057/palgrave.jors.2602263
3. BunkerEx: BunkerEx — Bunker Prices — Bunker Broker United Kingdom (2022).
https://www.bunker-ex.com. (accessed 2022-11-11)
4. Christiansen, M., Fagerholt, K.: Chapter 13: ship routing and scheduling in indus-
trial and tramp shipping. In: Toth, P., Vigo, D. (eds.) Vehicle Routing, pp. 381–408.
Society for Industrial and Applied Mathematics, Philadelphia, PA (2014). https://
doi.org/10.1137/1.9781611973594.ch13
5. Christiansen, M., Fagerholt, K., Nygreen, B., Ronen, D.: Chapter 4 maritime trans-
portation. In: Barnhart, C., Laporte, G. (eds.) Handbooks in Operations Research
and Management Science, Transportation, vol. 14, pp. 189–284. Elsevier (2007).
https://doi.org/10.1016/S0927-0507(06)14004-9
6. Christiansen, M., Fagerholt, K., Nygreen, B., Ronen, D.: Ship routing and schedul-
ing in the new millennium. Eur. J. Oper. Res. 228(3), 467–483 (2013). https://
doi.org/10.1016/j.ejor.2012.12.002
7. dos Santos, P.T.G., Kretschmann, E., Borenstein, D., Guedes, P.C.: Cargo routing
and scheduling problem in deep-sea transportation: case study from a fertilizer
 Tramp Ship Routing with Bunker Optimization 201

company. Comput. Oper. Res. 119, 104934 (2020). https://doi.org/10.1016/j.cor.

2020.104934
8. Korsvik, J.E., Fagerholt, K.: A tabu search heuristic for ship routing and scheduling
with flexible cargo quantities. J. Heuristics 16(2), 117–137 (2010). https://doi.org/
10.1007/s10732-008-9092-0
9. Ksciuk, J., Kuhlemann, S., Tierney, K., Koberstein, A.: Uncertainty in maritime
ship routing and scheduling: a Literature review. Eur. J. Oper. Res. 308, 499–524
(2023). https://doi.org/10.1016/j.ejor.2022.08.006
10. Meng, Q., Wang, S., Lee, C.Y.: A tailored branch-and-price approach for a joint
tramp ship routing and bunkering problem. Transp. Res. Part B: Methodol. 72,
1–19 (2015). https://doi.org/10.1016/j.trb.2014.11.008
11. Pache, H., Kastner, M., Jahn, C.: Current state and trends in tramp ship routing
and scheduling. In: Hamburg International Conference of Logistics (HICL), vol.
28, pp. 369–394. epubli GmbH (2019). https://doi.org/10.15480/882.2504
12. UNCTAD: Review of maritime transport 2021. United Nations, S.l. (2022).
https://unctad.org/publication/review-maritime-transport-2021, Accessed 19
June 2023
13. Vilhelmsen, C., Lusby, R., Larsen, J.: Tramp ship routing and scheduling with
integrated bunker optimization. EURO J. Transp. Logistics 3(2), 143–175 (2013).
https://doi.org/10.1007/s13676-013-0039-8
 Digital Twins in Seaports: Current
and Future Applications

Julian Neugebauer(B) , Leonard Heilig , and Stefan Voß

Institute of Information Systems, Universität Hamburg,

Von-Melle-Park 5, 20146 Hamburg, Germany
[email protected]
https://www.bwl.uni-hamburg.de/iwi.html

Abstract. Information systems in major seaports have evolved to digi-

tal hubs where state-of-the-art technologies are used to gather and ana-
lyze (near) real-time data from infrastructure and superstructure. Digi-
tal twins are seen as key enabler of Industry 4.0 applications and digi-
tal transformation in seaports. This paper presents case studies of digi-
tal twins in global seaports and investigates implementation layers and
decision support. Based on a literature review and interviews with port
stakeholders from different seaports, we identify potentials, challenges
and requirements for integrating digital twins in seaports. Moreover, we
present results and insights from a major project building a digital twin
for the EUROGATE container terminal in Hamburg, Germany. As such,
the paper provides an overview on the maturity of digital twin applica-
tions in the port sector and discusses important aspects to be considered
during implementation.

Keywords: Digital Twins · Seaports · Container Terminals · Case

Studies · Maritime Logistics

1 Introduction
Digital twins have evolved as one of the main drivers for digital transformation
in major seaports and are seen as key technology of Industry 4.0 applications. In
industry, digital twin technologies are recognized as one of the most important
areas of investment for terminal operators in the coming years [27]. The main
reason is that they can be implemented in various forms and support many use
cases to enable automation and improve productivity, efficiency, resilience and
sustainability in terminals and seaports.
According to Vanderhorn et al. [39] a digital twin is a “virtual representa-
tion of a physical system (and its associated environment and processes) that is
updated through the exchange of information between the physical and virtual
system.” As such, it is comprised of a set of technologies to gather (near) real-
time data from infrastructure and superstructure and its environment, associate
it with related port processes, visualize and analyze it to make decisions having
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 202–218, 2023.
https://doi.org/10.1007/978-3-031-43612-3_12
 Digital Twins in Seaports: Current and Future Applications 203

an impact on port operations by controlling infrastructure and superstructure

automatically. Moreover, they can be viewed as a tool for applying a variety of
methods for enabling data-driven decision making, such as by using simulations,
machine learning (ML) or advanced optimization algorithms utilizing the vast
amounts of data being collected within a digital twin.
Different forms of digital twins have been discussed and established in
research and practice in recent years [20]. As the term is used as another buzz-
word in industry, one has to be careful regarding applications being coined as
digital twin. Using established definitions from literature, it is important to ana-
lyze present applications in research and practice to analyze the current state-
of-the-art and to explore potentials and challenges in seaports. In this paper,
we present case studies of major implementations of digital twins in maritime
seaports and terminals. Based on a set of aspects and criteria, we provide an
overview and lessons learned from different forms of implementation. In this con-
text, we analyze which requirements or conditions must be met to successfully
implement or integrate digital twins in seaports and how these affect the inter-
action or coordination between different port stakeholders. For that purpose, we
analyze implementations on different layers and present insights from interviews
with industry experts involved in digital twin projects. Finally, we showcase
the implementation of an advanced digital twin being developed for the EURO-
GATE Container Terminal Hamburg (CTH) and provide insights about the key
challenges and developments in the project.
The remainder of the paper is structured as follows. Section 2 describes
related literature to get an overview on the current state of research in this area.
In Sect. 3, we present a framework to categorize a case study for the identified
digital twin applications in industry. Using this framework, we investigate the
details of each case and carve out the specific features, challenges, shortcomings
and potentials. This includes findings from interviews with various port experts
being involved in the respective projects. We further analyze the application of
advanced methods for decision support and explore a digital twin being realized
in the Port of Hamburg in Sect. 4. The overall results and lessons learned from
the case studies are discussed in Sect. 5. Finally, we provide some conclusions
and an outlook to future research.

2 Related Literature

Each digital twin, as defined by Vanderhorn et al. [39], consists of a model of a

physical system or operation with varying levels of complexity [11]. Complexity
in this context refers to the level of detail, including the number of variables and
equations used in a model. For a digital twin to be realized, it must be updated
in (near) real time with changes in the underlying physical system or operation,
and it must learn and adapt automatically. Seaports, facing high demand, are
undergoing digital transformation [13], utilizing technologies such as the Internet
of Things (IoT), big data, simulation, and ML applications. There are countless
other technologies and concepts available, leading to numerous use cases.
204 J. Neugebauer et al.

One example is the implementation of renewable energy sources in Italian

port areas, which involves modeling wind and solar energy production for deci-
sion making and then incorporating IoT monitoring [2]. Others are described
by Pavlic Skender et al. [26], whether they are autonomous drones for emissions
tracking used at the Port of Antwerp (Belgium), automated gates for dispatch-
ing in Kaohsiung (Taiwan), or the application of 5G for various other purposes
as described by the Port Authority of Hamburg (Germany).
A more advanced utilization of digital twins in use cases is decision support.
Hofmann and Branding [16] describe this for container terminal operations plan-
ning, discuss dispatching support and continuous monitoring and evaluation.
This is made possible by full modeling of key container terminal processes and
python-based simulation connected to IoT enabled devices. To do this, the col-
lected data is used to give the digital twin a picture of the current state, which
must be adjusted according to the information needed to meet the respective
task [45]. If the goal is to create a more resilient port, for example, analyzing
the implications of power supply shortages, different scenarios can be simulated
using the O2 DES.Net open source framework for allocating power to different
vehicles in a container terminal using automated guided vehicles (AGV) based
on power and container terminal simulations [46]. The simulation as part of the
digital twin allows decisions for future challenges to be analyzed in (near) real
time and an optimal strategy for vehicles and the entire container terminal to
be defined. Similar analyses are helpful for ongoing operations. These take over
the automated storage yard scheduling and work in the form of simulation-based
optimization [12].
To develop a digital twin with decision support in the domain of container
terminals, the framework described by Wang et al. [40] can be applied. The
development is in five steps starting with the “data aquisition” and creating
a “georeferenced design” which is not only used for visualization but also for
mapping data. Thirdly the integration of (near) real-time data to a model is
described, followed by an implementation of some form of “simulation, analysis
and optimization” for progressing the model and utilizing it to give new insights
from the digitally twinned perspective. Last but not least, the sharing and deliv-
ery of insights is described to ensure that the potential lessons and outcomes
of the digital twin are understood and executed. Yang et al. [43] describe and
apply a similar framework at the Port of Qingdao (China), albeit focusing on the
different layers of application. It is comparable to the framework described by
Zheng et al. [45]. The following areas are categorized as part of the framework:
– Physical resource layer (system-wide resources and their description)
– Virtual resource layer (model of, and mapping to real components)
– Data connection layer (network, interfaces and data exchange)
– Twin data layer (providing data of analysis, simulation and virtual scenes)
– Application service layer (information for a dynamic knowledge base)
Redelinghuys et al. [28] present a digital twin system from a manufacturing per-
spective and use similar layers with an additional configuration layer that takes
into account the self-adaptation and self-improvement of the digital twin system.
 Digital Twins in Seaports: Current and Future Applications 205

However, a suitable procedure and selection of the use cases to be implemented

must be made according to the area of application and the application goals.
The large number of possible use cases, technologies that can be used, and
ways to implement them effectively demonstrate the complexity of implementing
a digital twin. In addition, suitable criteria are needed to describe use cases in a
precise way so that a structured decision can be made in the selection process.

3 Digital Twin Enabled Use Cases

3.1 Digital Twins in Global Ports

To get an overview on current digital twin implementations, we categorize a

number of digital twin applications in the port domain and analyze them from
different angles. We classify them based on their application area, namely port
infrastructure (PI), container terminal (CT), multi-purpose terminal (MPT),
and bulk terminal (BT). For each digital twin, we identify the application pur-
pose, data sources, data processing techniques, type of user interface, integration
approach and a general level of detail based on a comparison. The latter gives
a rough estimate for the amount and granularity of data being mapped in the
respective digital twin, the general complexity of the implementation, and the
capability to support decision making. The results are gathered from a case study
and interviews with port experts involved in respective digital twin projects, as
indicated in Table 1.1 For further insights from the conducted interviews and
additional references the reader is referred to the authors upcoming paper “Dig-
ital Twins in the Context of Seaports and Terminal Facilities” [23].
The results in Table 1 show that digital twins have diverse application areas
in the port domain, ranging from infrastructure management to cargo handling.
They are also used to monitor and manage port infrastructure, improve oper-
ational efficiency, predict maintenance, and improve the ecological footprint.
While the applications have similarities regarding the use of information tech-
nology, they differ in terms of the utilization of data for advanced analytics
and decision support. For port infrastructure, often managed by a port author-
ity, the digital twin in the Port of Antwerp-Bruges, Belgium, can be seen as a
good example. One of many use cases is the identification of pollution sources,
mostly caused by ships in the ports or on the river sections between the ports.
To enable proper identification of emission sources, a large number of sensors
was installed in the ports, but for accurate localization, autonomous drones are
also sent to the locations where emissions are measured outside the norm. These
drones also take measurements, which are then sent in (near) real time to the
digital twin and used as a basis for visualizing and simulating certain scenarios.
The results provide decision support for the port authority and help lowering
greenhouse gas (GHG) emissions. Another example is the Port of Hamburg, Ger-
many, where sensor technologies and BIM are applied to visualize conditions of
1
BIM: building information modeling; VR: virtual reality; AR: augmented reality;
GNSS: global navigation satellite systems; TOS: terminal operating system.
206 J. Neugebauer et al.

Table 1. Categorization of Digital Twins in Global Seaports

Location of the Use Case Application Area Application Purpose Data Sources Data Processing User Interface Twin Integration Level of Detail Source
Antwerp- PI General port IoT, BIM, Big data, 3D (Near) ····· [26] Interview
Bruges, and emissions cameras, simulation, visualization, Real-time data
Belgium management power- and ML VR/AR integration
weather data, and feedback
autonomous loop
drones
Anzio, Italy PI Renewable IoT, BIM, Simulation Dashboards (Near) ·· [1]
energy geological data Real-time data
utilization
Bremen, BT Position GNSS, IoT, Big data, 3D (Near) Real-time ······ Interview
Germany monitoring cameras, simulation, visualization, data integration,
and move weather data ML integrated into automated decision
optimization control units making and (near)
real-time control
Dalian, China CT Equipment IoT, BIM, Big data, 3D visualization (Near) Real-time ······ [21] Interview
and terminal cameras, simulation, data integration,
management GNSS, TOS ML automated decision
and overall making and (near)
optimization real-time control
Genoa, Italy Multiple Renewable IoT, TOS Simulation Dashboards Data ··· [4, 9]
energy integration
utilization and and process
process model for
monitoring predictions
Gothenburg, Multiple Infrastructure IoT Big data, ML n.a. n.a. ··· [8]
Sweden and traffic
management
Hamburg, PI Infrastructure and IoT, cameras, Big data, 3D (Near) Real-time ····· [3, 29, 37] Interview
Germany traffic management GNSS, simulation, visualization, data integration,
weather data ML VR/AR, automated decision
dashboards, making and (near)
integrated into real-time control
traffic displays
Livorno, Italy PI, BT Infrastructure IoT, cameras, Big data, 3D (Near) ····· [7, 25, 33] Interview
and traffic GNSS, simulation, visualization, Real-time data
management weather data ML VR/AR integration
and feedback
loop
Mawan, China BT, CT Fleet and IoT, cameras, Big data, 3D visualization (Near) Real-time ····· [17]
terminal GNSS, TOS simulation, data integration,
management ML automated decision
making and (near)
real-time control
Multiple, BT, CT Terminal IoT, TOS Big data, Dashboards Data ···· [26, 41]
Thailand management simulation integration
and process
model for
predictions
Oulu, Finland MPT General port IoT, BIM, Big data, 3D visualization, dashboards (Near) Real-time ··· [5]
and process cameras simulation data integration and
management feedback loop
Qingdao, China CT Equipment IoT, BIM, Big data, simulation, ML 3D visualization (Near) Real-time ······ [6, 42, 43]
and terminal cameras, data integration,
management GNSS, TOS, automated decision
and overall weather data making and (near)
optimization real-time control
Rotterdam, Netherlands PI Infrastructure IoT, cameras, Big data, n.a. (Near) Real-time ····· [15, 24, 26],
and traffic GNSS, simulation, data integration, [17, 32, 34]
management weather data ML automated decision
making and (near)
real-time
optimization
Shanghai, CT Process GNSS, Big data, 3D (Near) Real-time ····· [10, 19, 26, 44]
China monitoring cameras, BIM, simulation, visualization data integration, Interview
and general TOS ML automated decision
optimization making, predictions
and (near) real-time
control
Singapore, CT, PI Infrastructure n.a. Simulation 3D visualization Data integration ···· [17, 18, 21, 26]
Singapore and terminal and process model Interview
management for predictions
Valencia, CT Terminal IoT, GNSS, Big data, 3D (Near) Real-time ·· [30, 31]
Spain management BIM, TOS simulation, visualization data integration and Interview
ML feedback loop

port infrastructure (e.g., streets, bridges, locks), which supports infrastructure

planning. One example is the project smartBridge to optimize maintenance by
monitoring (near) real-time conditions of the Köhlbrand Bridge connecting the
port with a main highway.2
Applications in bulk terminals include the precise positioning of steel coils
using various movement sensors attached to the forklifts, as done in Bremen,
Germany. Such precise positioning of goods and vehicles avoids seek times and
enables a better warehouse management in their enterprise resource planning
(ERP) system. With accurate positioning and transparent start and end times
for all handling operations, the bulk terminal is able to predict movement times
and control the number of vehicles needed for a given workload. In the Port of
Mawan, China, a similar system was implemented, but with the additional use
2
See, e.g., https://www.homeport.hamburg/portfolio/smartbridge.
 Digital Twins in Seaports: Current and Future Applications 207

case of fleet management. In this case, the focus is on historical data, which
serves as a basis for analyzing bottlenecks and technical problems in order to
avoid them during planning [17]. Another example in this context can be found
in the Port of Livorno, Italy, where advanced technologies, such as 3D LIDAR
sensors and Wide Dynamic Range (WDR) cameras, are used for the localization
of cargo boxes and forklifts. Based on the planned tasks, the forklift driver is
guided within the terminal by wearing AR glasses.3 An example of a digital twin
in multi-purpose terminals is the Port of Oulu, Finland, where both vessel- and
terminal related information are represented in a geographic information system
(GIS) in order to increase situational awareness for improving the efficiency and
safety in the terminals. The digital twin provides a number of tools enabling
planning of port infrastructure elements, such as industrial lighting or cameras
of the monitoring system [5].
Looking at the application area of container terminals, the digital twin at the
Port of Qingdao realizes a wide variety of use cases. The status of vehicles along
with respective processes are represented in a complex 3D visualization. Not
only the positions and possible technical problems are displayed based on (near)
real-time GNSS and IoT data, but also simulations and optimization methods
for path planning and forecasts of the future handling rate based on ML are inte-
grated. As required for digital twins, decisions of the digital twin are propagated
to the control units of the AGVs [43]. Another focus is on various visualizations
for the operating status or the system behavior, as well as the structured imple-
mentation [42]. Similar applications can be found in the Port of Shanghai and
Dalian, China, where a very detailed digital twin supporting decision making for
automated and semi-automated terminals has been developed [10,21]. Several
digital twins apply big data technologies, such as Hadoop, to process various
sources of data. In the Port of Valencia, for instance, key performance indicators
are calculated based on operational data, AIS (automatic identification system),
and weather data. Another advanced digital twin for a container terminal in the
Port of Hamburg, Germany, is detailed in Sect. 4.

3.2 Maturity and Applied Layers

The analysis of the digital twins in Table 1 reveals that data sources, process-
ing techniques, user interfaces and twin integration vary across the different
application areas. For example, some applications rely solely on IoT data, while
others use many data sources and establish advanced decision support. Simi-
larly, some use cases require a high level of detail, while others can operate at
a more abstract level. Although the described applications give a good idea of
the use cases, the following section will take a generalized look at the different
layers of implementations and their maturity. The following analysis focuses on
the technical layers described in Sect. 2 which were developed by Yang et al. [43]
and extended by a sixth layer of self-adaption in Redelinghuys et al. [28].
3
See, e.g., https://www.ericsson.com/en/blog/2020/12/digital-twins-port-
operations.
208 J. Neugebauer et al.

The physical resource layer is present in all digital twins listed in Table 1,
albeit to very different degrees. In the Port of Anzio, Italy, only certain aspects
regarding energy production and consumption are represented while the digital
twin in Shanghai includes models of vehicles, infrastructure, weather and other
factors influencing terminal operations. Interacting mechanisms as well as the
current state of each object are taken into account and thus show the physical
resources in more detail. Across the layer, it is important to note that the level
of detail and the number of objects being represented are almost unlimited.
The trade off between the complexity and utility of displaying detailed physical
resources and the associated cost of implementing them must be considered.
This layer describes only the resource and its general attributes, excluding more
advanced process definitions or rules and more extensive environmental and
business-related information.
To model additional aspects besides the physical objects, especially environ-
mental aspects and business processes (as required by digital twins by Van Der
Horn [39]), the virtual resource layer is used. Similar to the previous level, the
more detailed and extensive the contextual information, the better and more
accurate the digital twin. During the interviews, the person in charge at the
ports of Antwerp-Bruges mentioned their goal to build the port’s nervous sys-
tem, including things like smell, which was an example for air quality sensors
and sight for example, realized with cameras and various visualizations. Other
environmental factors are derived from IoT and external data, which included,
for example, infrastructure conditions and vessel tracking. Examples for com-
bining physical objects with business processes in which they interact are given
in the case study presented in Sect. 4.
Data connectivity as a layer of digital twins is described in detail in the Port
of Qingdao. In order to use all the information and connect the representation
of physical resources and their behavior as well as 3D models, the network,
protocols, interfaces and general data exchange must be defined and managed.
This is critical for establishing an integration and is the focus of many digital twin
projects. For the Port of Singapore, for example, an open source framework was
used to model the digital twin in order to ensure future ease of integration and
extensibility, and all of the ports interviewed cited the specification of interfaces
as one of the biggest challenges in development. The data connection layer is
directly linked to the twin data layer, which includes the storage and processing
of data, but also advanced data processing techniques such as big data analytics
and other data mining techniques. For example, in the interview with the Port of
Hamburg, it was mentioned that traffic information is combined with simulation
data to be used in various applications. Another example is a BIM infrastructure
model, used for the before-mentioned Köhlbrand bridge, which was then used in
the application service layer to predict damage and schedule maintenance.
The application service layer includes decision support, using advanced meth-
ods, such as simulation modeling, optimization approaches and ML, but also
monitoring and business intelligence dashboards for supporting human decision
making. Examples from the digital twins include the actual traffic management
 Digital Twins in Seaports: Current and Future Applications 209

service and its visualization implemented at the Port of Livorno but also the
analysis and optimization of operations for the container terminal at the Port
of Mawan. Note that although all projects listed in Table 1 incorporate this
layer, albeit with varying degrees of detail, not all feed these insights back into
the physical domain in an automated or semi-automated and (near) real-time
manner.
The self-adaptation layer can be realized when data from the lower layers are
considered on the application layer to adapt to certain situations and propagate
decisions back to the physical resources so that they can adapt accordingly. If
the self-adaptation layer is implemented in an architecture for cross-layer infor-
mation and the various layers are implemented with a high level of detail, it
is assumed that there is a huge potential for automation. In addition, a large
number of use cases can then be supported. This is also the goal in the presented
case study in the following section.

4 Digital Twin at the EUROGATE Container Terminal

Hamburg

The research-industry project TwinSim, currently carried out by EUROGATE

Holding, EUROGATE Technical Services, Akquinet port consulting and the
Institute of Information Systems (IWI), University of Hamburg, is pursuing the
development of a digital twin for the visualization and simulation-based opti-
mization of terminal operations at the EUROGATE Container Terminal Ham-
burg (CTH). The research project has a project volume of 3.65 million Euros
and is funded by the German Federal Ministry of Digital Affairs and Trans-
port (BMDV) within the frame of the Innovative Port Technologies (IHATEC)
program over a period of three years [38]. The scope is on container terminal
processes and maintenance for straddle carrier (SC) and gantry crane operations
in order to strengthen the efficiency, eco-awareness, and resilience of the termi-
nal. In a comprehensive requirements analysis a multitude of use cases utilizing
a digital twin were identified in several workshops (for details, the interested
reader is referred to by Neugebauer [22]). The final result of the requirements
analysis was a selection of 14 use cases to be implemented within the project
and moreover serve to develop a best practice implementation of a digital twin
for the application area of container terminals considering all levels discussed in
Sect. 2.
Starting with the representation of the physical resources of the terminal,
the digital twin firstly considers containers, infrastructure and superstructure
involved in container transport and storage processes for both sea- and landside
operations. The following list represents physical objects and potential dynamic
attributes to be mapped:
210 J. Neugebauer et al.

– Container (position, reefer status)

– Straddle carrier (position, technical status, operational status)
– Gantry cranes (position, technical status, operational status)
– Special handling equipment (e.g., overheight frames) (position)
– Vessel (position, estimated time of arrival)
– Truck and trailers (position, operational status)
– Light (operational status)
– Reefer station (position, technical status)
– Yard (operational status)
– Truck gate (operational status)
– Buildings

These objects are modeled as 3D representations in a simulation model, which

is later also used for (near) real-time visualizations. Buildings are included for
reference purposes only, as they have no influence on the operation of the con-
tainer terminal. The container yard, albeit represented with different roads as
well as one-way driving restrictions, is also only modeled as a reference and pro-
vided with information on transshipment areas, hazardous materials and reefer
storage locations, and other higher level information. Surface conditions, for
example, are not included. A preview of the visualization can be seen in Fig. 1.

Fig. 1. 3D Visualization of CTH

However, only a part of the data comes directly from the physical devices. In
the first phase, the project focuses on collecting data from SCs and gantry cranes.
This is achieved by connecting the equipment’s programmable logic controller
(PLC), sensors, and GNSS though an edge processor to an on-premises message
broker, which is communicating with a cloud message broker using the MQTT
(Message Queuing Telemetry Transport) protocol, as shown in Fig. 2.
 Digital Twins in Seaports: Current and Future Applications 211

Fig. 2. IoT-Edge Architecture

The virtual resource layer focuses mainly on building cycles of SCs and gantry
cranes. A cycle contains all equipment activities for moving a container from
one position (e.g., in the yard) to another position (e.g., at the quay). In some
situations also dual cycles need to be considered, meaning that two cycles are
considered together, such as when loading and unloading a full container from a
vessel in two subsequent moves. Note that one cycle can also contain more than
one container move (e.g., twin moves). For building cycles, from the edge device
containing PLC, GNSS and sensor data are combined with TOS data in order
to relate the IoT data with the business process context. The contextual data is
given by a job instruction from the TOS, representing all information to conduct
the movement of a container, such as the start and end positions, container
information, planned schedules, etc., resulting from the first planning phases
(e.g., vessel planning). The combination enables situational awareness about the
behavior and conditions of terminal equipment in relation to certain operations,
time, and other environmental impacts, such as weather if provided. Moreover, it
supports multiple use cases and has implications for all areas of operations since
it allows to analyze operations in detail and use extracted insights to improve the
planning and enhance the transparency for terminal operators and customers.
Added to this is ecologically relevant information such as fuel consumption or
tire wear, which can also be used later to monitor the total GHG emissions,
not only in general, but also the consumption of individual containers during
their lifetime at the terminal. In the project, the data from the equipment is
further used to enable predictive maintenance. As shown in Fig. 3, the IoT and
contextual data is shown when selecting physical objects in the digital twin. Note
that this is an interim result; later, extensive data dashboards and predictions
are included to better monitor conditions and key performance indicators.
Data connectivity, especially in a project of this size, with a multitude of
stakeholders, with data sources from various types of equipment from diffe-
rent suppliers and different information systems (e.g., TOS, ERP) in numer-
ous formats, becomes a complex task and thus requires standardized interfaces.
Currently one major problem with respect to the equipment is that suppliers
are not using the same definitions for measuring certain aspects, e.g., the opera-
tional time of a SC. In one case the engine-on time is used, in the other case, the
212 J. Neugebauer et al.

Fig. 3. Contextual information for SC

working hours of the driver are used. At the EUROGATE container terminals,
SCs of different suppliers are in use, making it difficult to compare metrics.
The Terminal Industry Committee 4.0 (TIC 4.0) is working on data stan-
dards for container terminals and includes crane manufacturers, TOS suppliers
and more (see [36]). A number of publications has been done for standardizing
the syntax and semantics for different equipment data sources [35]. In the project,
TIC 4.0 standards are used to assemble position data and technical information
from SCs (e.g., tire pressure, oil temperature, etc.) and gantry cranes, send them
in (near) real-time to message brokers and combine them with standardized job
instructions filled with information from the TOS to build respective cycles.
One major advantage of the proposed digital twins is that the standardization
allows to connect information systems (e.g., TOS) and equipment from differ-
ent suppliers/providers without changing the core implementation by talking
or translating into the same language. Therefore, all higher level applications,
including advanced decision support functionality, can read and write data in
this predefined format and solutions can be developed without spending time
on interface definition and in an agile manner. As equipment suppliers have just
started with implementing TIC 4.0 compliant data for newer equipment, it is
still necessary to transform the equipment data in TIC messages. As shown in
Fig. 2, a special edge processor has been implemented in the project to trans-
form all data records before forwarding them to the cloud. The (near) real-time
data connectivity is enabled by a scalable message broker system that connects
edge devices from physical objects in on-premises networks with the cloud-based
streaming processors and the twin data layer.
The gathered data is processed in several data processing stages for multiple
purposes in the twin data layer, which is build with AWS cloud services. This
layer includes a cloud-based data lake and big data technologies to prepare the
 Digital Twins in Seaports: Current and Future Applications 213

data for the higher level applications. At both the data connectivity and twin
data layers, some data streams need to be processed in (near) real time (e.g.,
objects for (near) real-time visualization). This affected the cloud services and
technologies chosen. A list of all applications realized for the digital twin at the
CTH is almost impossible to compile and is also not the goal of this work. Rather,
those that show the decision area of the digital twin are discussed. Currently the
terminal planning is realized with rule-based approaches using the TOS. For
dispatching the SCs, for example, a score for each container is calculated based
on several weighted factors, such as the shortest distance from a theoretical
position of an SC to its planned position. This is a rather greedy approach that
does not consider the combination of different moves.
Therefore, the goal is to improve this planning with more advanced algo-
rithms based on the vast availability of granular data, which will be discussed in
future works. Starting from the integration of accurate positions based on GNSS
data, the distance-based calculations can be replaced by those that include, for
example, the estimated actual distances traveled by the SCs. In the project, a
ML model for predicting travel times will be implemented. In order for these pre-
dictions to be as accurate as possible and thus represent the physical resources
most accurately, weather information is collected by sensors on the container
terminal. This provides information on precipitation, wind speed and direction,
current visibility and much more. Weather conditions also have an impact on
planned operations and maintenance activities, which may need to be postponed
due to several weather conditions. Thus, weather predictions also serve for early
warnings so that plannings can be adjusted accordingly. In addition, water levels
and more as well as arrival time predictions of vessels from external sources are
integrated.
As a second stage a simulation will provide a feedback loop for optimiza-
tion methods and will allow for a more in-depth analysis of manual decisions.
Through simulating past shifts with new conditions (e.g., more SCs being used)
lessons learned can be taken for improving the current situation. By simulating
the next shift, the future performance of the terminal can be evaluated without
interrupting ongoing operations. In the final phase of realizing the described dig-
ital twin, a simulation-based optimization is implemented for the dispatching of
SCs and container storage allocation, using all available information, integrating
ML, and applying operations research methods to find an optimized plan and
evaluate it through simulation. By taking into account granular data from the
past, the simulation helps to validate the plan in a more accurate and complex
representation of the container terminal. The results are subsequently reported
to the TOS and passed on, e.g., to SC drivers, who then take an optimized route
better utilizing available resources, avoiding unproductive moves, waiting times,
and minimizing the environmental impact. Moreover, the solution can lead to a
greater operational resilience as unforeseen events can be simulated and emer-
gency plans or more resilient plans can be assessed in advance. Since this system
controls for the movement of the vehicles, it is also easier to use automated
equipment in the future.
214 J. Neugebauer et al.

5 Discussion and Conclusion

Digital twins offer many benefits and opportunities for seaports and container
terminals, such as improving efficiency, safety, sustainability, and resilience. How-
ever, they also pose some challenges and limitations. From the analyzed cases,
interviews, and from the on-going project, we could identify the following recur-
ring challenges, which may require more research:

– Choosing and integrating suitable IoT devices and data processing technolo-
gies, also regarding implementation and operation costs
– Ensuring data quality, standardization, and security
– Scaling and adapting to changing needs
– Interacting with other systems and stakeholders (e.g., port community sys-
tems)
– Complying with ethical and legal regulations (e.g., the General Data Protec-
tion Regulation (GDPR) in Europe)
– Estimating and managing implementation costs and time
– Avoiding model bias and prejudice
– Expanding the scope to other parts of the supply chain
– Anticipating long-term impacts and dependencies on the digital twin

This paper has examined how digital twins can be used in seaports and
container terminals to optimize and improve terminal planning and operations
as well as maintenance. To this end, a review of existing publications on the
topic was conducted and a number of port cases were discussed to assess the
current state of research and practice. The interviewed experts expressed a need
for more standardization, further collaboration, and more empirical evidence
to support the adoption and evaluation of digital twins in this domain. Some
possible solutions or recommendations to address the challenges and limitations
of digital twins are:

– Developing standards and guidelines for data collection, processing, sharing,

and protection
– Enhancing collaboration and communication among different stakeholders,
such as port authorities, terminal operators, shipping companies, logistics
providers, regulators, and researchers
– Conducting more empirical studies and evaluations to measure the impact
and value of digital twins for seaports and container terminals

Regarding the latter aspect, we see that in industry the question whether
digital twins are just costly toys or valuable support tools still arises.4 In this
context, this work has shown the possibilities of a digital twin on different levels
and with different degrees of complexity. Furthermore, it has been shown that an
in-depth requirements analysis in terms of valuable use cases for the individual
port or terminal is decisive to gain value (as further discussed in [22]).

4
See, e.g., http://www.bpoports.com/BPC/agenda-3.pdf.
 Digital Twins in Seaports: Current and Future Applications 215

As digital twins become more mature and widely adopted, we expect to see
more benefits and opportunities for seaports and container terminals in two main
areas. The first area is expanding the scope and impact of digital twins to other
parts of the supply chain. Digital twins can enable a seamless integration and
coordination of port operations with other actors and processes in the supply
chain, such as shippers, carriers, logistics providers, customs authorities, and
end customers. This can improve the visibility, resilience, and efficiency of the
entire supply chain, as well as reduce costs, risks, and environmental impacts.
Digital twins can also support the development of new supply chain concepts
and solutions, such as synchromodality or the circular economy. The second area
is creating new business models and value propositions for port services and
products. Digital twins can enable ports and terminals to offer new services and
products that leverage their data and insights to create value for their customers
and partners. For example, ports and terminals may provide data-driven services
such as (near) real-time monitoring, forecasting, optimization, or simulations of
port operations and processes to their customers. They can also integrate data-
enabled objects such as “smart” containers that can communicate and interact
with the digital twin (see, e.g., [14]). These services and products can create new
revenue streams and competitive advantages for worldwide ports and container
terminals.

References
1. Agostinelli, S., Cumo, F., Nezhad, M.M., Orsini, G., Piras, G.: Renewable energy
system controlled by open-source tools and digital twin model: zero energy port
area in Italy. Energies 15(5), 1817 (2022). https://doi.org/10.3390/en15051817
2. Agostinelli, S., Neshat, M., Majidi Nezhad, M., Piras, G., Astiaso Garcia, D.:
Integrating renewable energy sources in Italian Port areas towards renewable
energy communities. Sustainability 14(21), 13720 (2022). https://doi.org/10.3390/
su142113720
3. Ali-Tolppa, J., Kajo, M.: Mobility and QoS prediction for dynamic coverage opti-
mization. In: Proceedings of the IEEE/IFIP Network Operations and Management
Symposium (NOMS), pp. 1–2. IEEE, Budapest, Hungary (2020). https://doi.org/
10.1109/NOMS47738.2020.9110396
4. Battilani, C., Galli, G., Arecco, S., Casarino, B., Granero, A., Lavagna, K., Varna,
R., Ventura, M., Revetria, R., Damiani, L.: Business process re-engineering in pub-
lic administration: the case study of Western Ligurian sea port authority. Sustain.
Futures 4, 100065 (2022). https://doi.org/10.1016/j.sftr.2022.100065
5. Blare.Tech: Increasing port operations efficiency and safety thanks to the full situ-
ational awareness through the use of digital twin (2023). https://www.blare.tech/
project/port-of-oulu-digital-twin
6. Cai, J., Li, G., Zhang, L.: Study on the construction of digital twin-oriented oper-
ation ecosystem model of large-scale comprehensive port. In: Yang, M., Lanzinha,
J.C., Samui, P., Bao, X., Hu, J. (eds.) Advances in Transdisciplinary Engineering,
pp. 1280–1285. IOS Press (2022). https://doi.org/10.3233/ATDE220998
7. Cavalli, L., et al.: Addressing efficiency and sustainability in the port of the future
with 5G: the experience of the Livorno port. A methodological insight to measure
216 J. Neugebauer et al.

innovation technologies’ benefits on port operations. Sustainability 13(21), 12146

(2021). https://doi.org/10.3390/su132112146
8. Dalaklis, D., Christodoulou, A., Ölcer, A., Ballini, F., Dalaklis, A., Lagdami, K.:
The port of Gothenburg under the influence of the fourth stage of the industrial
revolution: implementing a wide portfolio of digital tools to optimize the conduct
of operations. Marit. Technol. Res. 4(3), 1–18 (2021). https://doi.org/10.33175/
mtr.2022.253844
9. Damiani, L., Revetria, R., Morra, E.: A digital twin for supporting energy manage-
ment in complex maritime terminals. In: Proceedings of the International Maritime
Transport and Logistics Conference, pp. 1–19. Rome, Italy (2019). https://marlog.
aast.edu/attachments/cms/solid/0920ac36890175d746314dd28d79a004.pdf
10. Ding, Y., et al.: Real-time monitoring and optimal resource allocation for auto-
mated container terminals: a digital twin application at the Yangshan Port. J.
Adv. Transp. 2023, 6909801 (2023). https://doi.org/10.1155/2023/6909801
11. Fuller, A., Fan, Z., Day, C., Barlow, C.: Digital twin: enabling technologies, chal-
lenges and open research. IEEE Access 8, 108952–108971 (2020). https://doi.org/
10.1109/ACCESS.2020.2998358
12. Gao, Y., Chang, D., Chen, C.H., Xu, Z.: Design of digital twin applications in
automated storage yard scheduling. Adv. Eng. Inform. 51, 101477 (2022). https://
doi.org/10.1016/j.aei.2021.101477
13. Heilig, L., Lalla-Ruiz, E., Voß, S.: Digital transformation in maritime ports: anal-
ysis and a game theoretic framework. Netnomics 18, 227–254 (2017). https://doi.
org/10.1007/s11066-017-9122-x
14. Heilig, L., Voß, S.: A cloud-based SOA for enhancing information exchange and
decision support in ITT operations. In: González-Ramı́rez, R.G., Schulte, F., Voß,
S., Ceroni Dı́az, J.A. (eds.) ICCL 2014. LNCS, vol. 8760, pp. 112–131. Springer,
Cham (2014). https://doi.org/10.1007/978-3-319-11421-7 8
15. Henrı́quez, R., Martı́nez de Osés, F.X., Martı́nez Marı́n, J.E.: Technological drivers
of seaports’ business model innovation: an exploratory case study on the port
of Barcelona. Res. Transp. Busi. Manage. 43, 100803 (2022). https://doi.org/10.
1016/j.rtbm.2022.100803
16. Hofmann, W., Branding, F.: Implementation of an IoT- and cloud-based digital
twin for real-time decision support in port operations. IFAC-PapersOnLine 52(13),
2104–2109 (2019). https://doi.org/10.1016/j.ifacol.2019.11.516
17. Klar, R., Fredriksson, A., Angelakis, V.: Assessing the maturity of digital twinning
solutions for ports (2022). https://doi.org/10.48550/ARXIV.2212.07722
18. Li, H., Cao, X., Sharma, P., Lee, L.H., Chew, E.P.: Framework of O2 DES.NET:
digital twins for next generation ports and warehouse solutions. In: Proceedings of
the Winter Simulation Conference (WSC), pp. 3188–3199. IEEE, Orlando, USA
(2020). https://doi.org/10.1109/WSC48552.2020.9384111
19. Li, Y., Chang, D., Gao, Y., Zou, Y., Bao, C.: Automated container terminal pro-
duction operation and optimization via an AdaBoost-based digital twin framework.
J. Adv. Transp. 2021, 1–16 (2021). https://doi.org/10.1155/2021/1936764
20. Madusanka, N.S., Fan, Y., Yang, S., Xiang, X.: Digital twin in the maritime
domain: a review and emerging trends. J. Mar. Sci. Eng. 11(5), 1021 (2023).
https://doi.org/10.3390/jmse11051021
21. Mi, W., Liu, Y.: Smart port and digital monitoring and diagnosis. In: Mi, W., Liu,
Y. (eds.) Smart Ports, pp. 171–188. Springer, Singapore (2022). https://doi.org/
10.1007/978-981-16-9889-7 10
 Digital Twins in Seaports: Current and Future Applications 217

22. Neugebauer, J., Heilig, L., Voß, S.: Anforderungsanalyse zur Umsetzung eines
digitalen Zwillings im Containerterminal. HMD Praxis der Wirtschaftsinformatik
60(1), 110–131 (2023). https://doi.org/10.1365/s40702-022-00941-1
23. Neugebauer, J., Heilig, L., Voß, S.: Digital twins in the context of seaports and
terminal facilities. In: Flexibil Services and Manufacturing, pp. 1–57 (2023)
24. Özkanli, A., DeniZhan, B.: Digitalization roadmap for Turkish seaports. Eur. J.
Sci. Technol. 358–363 (2020). https://doi.org/10.31590/ejosat.araconf46
25. Pagano, P., Antonelli, S., Tardo, A.: C-ports: a proposal for a comprehensive stan-
dardization and implementation plan of digital services offered by the “ort of the
future” Comput. Ind. 134, 103556 (2022). https://doi.org/10.48550/arXiv.2104.
13175
26. Pavlić Skender, H., Ribarić, E., Jović, M.: An overview of modern technologies in
leading global seaports. Pomorski Zbornik 59(1), 35–49 (2020). https://doi.org/
10.18048/2020.59.02. https://hrcak.srce.hr/249260
27. Port Technology: PSA highlights digital twin as biggest emerging technology
trend (2021). https://www.porttechnology.org/news/psa-highlights-digital-twin-
as-biggest-emerging-technology-trend/
28. Redelinghuys, A.J.H., Basson, A.H., Kruger, K.: A six-layer architecture for the
digital twin: a manufacturing case study implementation. J. Intell. Manuf. 31(6),
1383–1402 (2019). https://doi.org/10.1007/s10845-019-01516-6
29. Rost, P., et al.: Customized industrial networks: network slicing trial at Ham-
burg seaport. IEEE Wirel. Commun. 25(5), 48–55 (2018). https://doi.org/10.1109/
MWC.2018.1800045
30. Sarabia-Jacome, D., Lacalle, I., Palau, C.E., Esteve, M.: Enabling industrial data
space architecture for seaport scenario. In: Proceedings of the IEEE 5th World
Forum on Internet of Things (WF-IoT), p. 101–106. IEEE, Limerick, Ireland
(2019). https://doi.org/10.1109/ACCESS.2019.2963283
31. Sarabia-Jacome, D., Palau, C.E., Esteve, M., Boronat, F.: Seaport data space for
improving logistic maritime operations. IEEE Access 8, 4372–4382 (2020). https://
doi.org/10.1109/ACCESS.2019.2963283
32. Simoni, M., Schiavone, F., Risitano, M., Leone, D., Chen, J.: Group-specific busi-
ness process improvements via a port community system: the case of Rotterdam.
Prod. Plan. Control 33(4), 371–385 (2022). https://doi.org/10.1080/09537287.
2020.1824029
33. Tardo, A., Pagano, P., Antonelli, S., Rao, S.: Addressing digitalization though out
a prototyping framework for agile smart services development: the case of Livorno
Port. J. Phys: Conf. Ser. 2311(1), 012007 (2022). https://doi.org/10.1088/1742-
6596/2311/1/012007
34. Taylor, S.J.E., Johansson, B., Jeon, S., Lee, L.H., Lendermann, P., Shao, G.: Using
simulation and digital twins to innovate: are we getting smarter? In: Proceedings of
the Winter Simulation Conference (WSC), pp. 1–13. IEEE, Phoenix, USA (2021).
https://doi.org/10.1109/WSC52266.2021.9715535
35. Terminal Industry Committee 4.0: Publications of the Terminal Industry Commit-
tee 4.0 (2023). https://tic40.org/publications/
36. Terminal Industry Committee 4.0: TIC4.0 Members (2023). https://tic40.org/tic4-
0-members/
37. Tesse, J., Baldauf, U., Schirmer, I., Drews, P., Saxe, S.: Extending internet of things
enterprise architectures by digital twins exemplified in the context of the Hamburg
port authority. In: Proceedings of the 27th Americas Conference on Information
Systems (AMCIS), pp. 1–11. Montreal, Canada (2021)
218 J. Neugebauer et al.

38. TÜV Rheinland: Verbundprojekt TwinSim zum 01.10.2021 gestartet IHATEC

(2021). https://www.innovativehafentechnologien.de/verbundprojekt-twinsim-
zum-01-10-2021-gestartet/
39. Van Der Horn, E., Mahadevan, S.: Digital twin: generalization, characterization
and implementation. Decis. Support Syst. 145, 113524 (2021). https://doi.org/10.
1016/j.dss.2021.113524
40. Wang, K., Hu, Q., Zhou, M., Zun, Z., Qian, X.: Multi-aspect applications and
development challenges of digital twin-driven management in global smart ports.
Case Stud. Transp. Policy 9(3), 1298–1312 (2021). https://doi.org/10.1016/j.cstp.
2021.06.014
41. Wattanakul, S., Henry, S., Napaporn, R., Ouzrout, Y.: A port digital twin model
for operational uncertainty management. J. Modern Project Manage. 9(3), 154–167
(2022). https://hal.archives-ouvertes.fr/hal-03579344
42. Yang, A., Liu, Y., Xin, C., Chen, Q., Wang, L.: Towards intuitive visualisation
goals for the operation optimisation of automated container terminal based on
digital twin technology. Marit. Policy Manage. 1–22 (2023). https://doi.org/10.
1080/03088839.2023.2166686
43. Yang, W., et al.: A digital twin framework for large comprehensive ports and a
case study of Qingdao Port. Int. J. Adv. Manuf. Technol. (2022). https://doi.org/
10.1007/s00170-022-10625-1
44. Yao, H., Wang, D., Su, M., Qi, Y.: Application of digital twins in port system.
J. Phys. Conf. Ser. 1846(1), 012008 (2021). https://doi.org/10.1088/1742-6596/
1846/1/012008
45. Zheng, Yu., Yang, S., Cheng, H.: An application framework of digital twin and
its case study. J. Ambient. Intell. Humaniz. Comput. 10(3), 1141–1153 (2018).
https://doi.org/10.1007/s12652-018-0911-3
46. Zhou, C., et al.: Analytics with digital-twinning: a decision support system for
maintaining a resilient port. Decis. Support Syst. 143, 113496 (2021). https://doi.
org/10.1016/j.dss.2021.113496
 A Neural Network Approach for ETA
Prediction in Inland Waterway Transport

Peter Wenzel1 , Raka Jovanovic2 , and Frederik Schulte1(B)

1
Technical University of Delft, Mekelweg 5, 2628 Delft, CD, Netherlands
{p.a.wenzel,f.schulte}@tudelft.nl
2
Qatar Environment and Qatar Environment and Energy Research Institute,
Hamad bin Khalifa University, PO Box 5825, Doha, Qatar
[email protected]

Abstract. Ensuring the accuracy of the estimated time of arrival (ETA)

information for ships approaching ports and inland terminals is increas-
ingly critical today. Waterway transportation plays a vital role in freight
transportation and has a significant ecological impact. Improving the
accuracy of ETA predictions can enhance the reliability of inland water-
way shipping, increasing the acceptance of this eco-friendly mode of
transportation. This study compares the industry-standard approach for
predicting the ETA based on average travel times with a neural network
(NN) trained using real-world historical data. This study generates and
trains two NN models using historical ship position data. These models
are then assessed and contrasted with the conventional method of cal-
culating average travel times for two specific areas in the Netherlands
and Germany. The results indicate by using specific input features, the
quality of ETA predictions can improve by an average of 20.6% for short
trips, 4.8% for medium-length trips, and 13.4% for long-haul journeys
when compared to the average calculation.

Keywords: Neural Networks · Machine Learning · Inland Waterway

Transport · Estimated Time of Arrival Prediction

1 Introduction

In 2021, the total volume of goods transported on European inland waterways

was 524 million tonnes, increasing by 3.9 % compared with the previous year.
The complete transport reached 136 billion tonne-kilometres, up 3.3 % of the
prior year [9]. Inland Waterway Transportation (IWT) is part of the critical
infrastructure that provides essential services that are substantial to the safety
and the economic and social welfare of society [20]. Therefore inland waterway
infrastructure maintenance and management is critical but not ideal. In 2015,
for example, around 85% of locks, 73% of weirs, and 87% of pumping stations
were in an inadequate state of repair [20]. Navigation hazards and safety mea-
sures in inland waterways play an important role. For 2010, 2011, and 2013, the
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 219–232, 2023.
https://doi.org/10.1007/978-3-031-43612-3_13
220 P. Wenzel et al.

most frequent accident was the collision with infrastructure and bridges. This
type accounted for 38–40% of all accidents. The second most frequent type of
accident was the collision between ships (18–19%). Due to increasing low water
periods caused by climate change, low water levels will lead to smaller vessels
and, therefore, an increase in transportation costs [13]. In order to make accu-
rate predictions, it is critical to consider and account for exceptions in the inland
waterway network.
Intermodal connectivity and transport efficiency in inland waterways are cru-
cial to model cost competitiveness compared with other transportation modes
such as road and rail [21]. Waterway transportation has an irreplaceable key
position in the entire transportation development process. It is also one of the
important modes of transportation to have a significant ecological impact [12].
The maritime operations in a port involve many parties, including pilots,
tug boats, boatmen, agents, supervision agencies such as customs and police,
stevedores and others. The arrival of a ship triggers activities at all these parties,
who then determine the performance of the port as a whole by each contributing
their specialist activity [19]. The need for reliable tools to verify and ensure
the accuracy of the estimated time of arrival (ETA) information provided by
ships as they approach ports has never been more critical than it is today [3,
6]. This paper establishes a groundwork for future research by demonstrating
the advantages of using neural networks (NN) with simple input parameters
for ETA prediction in inland waterway transportation. The development and
testing of various NN architectures make it possible to consider external factors
such as weather forecasts, river depth forecasts, and infrastructural factors when
predicting ETA.
In Sect. 2, this paper presents the related research. The methodology used is
delineated, and the process of data preparation is provided in Sect. 3. Moving on
to Sect. 4, the training of our neural network model is detailed, along with the
accompanying results. Finally, the conclusion of the paper offers the limitations
and potential future research in Sect. 5.

2 Related Research
The utilisation of Automatic Identification System (AIS) data for ETA predic-
tion in seagoing vessels is a widely researched topic. It has been shown in various
research work [11,15,23] that data-driven algorithms achieve higher accuracy in
terms of the time of ETA error. The main goal of this research work is to improve
ETA time calculation to improve efficiencies in port operations. Valero [18] pre-
dicts ETA times to improve short sea shipping. El Mekkaoui [8] focuses more
on the improvements of predictions for bulk ports, and Pani [16], and Yu [23]
focus on container and transhipment terminals. The applied methods in overseas
shipping research used vary from machine learning algorithms such as NNs [2,7],
random forest [23], reinforcement learning [17], bayesian learning [15] and deep
learning [8,11] to pathfinding algorithms [1].
There are significant differences in IWT when estimating the ETA for seago-
ing vessels. Water levels and weather conditions are crucial factors that can
 Neural Network ETA Prediction in Inland Waterway Transport 221

affect the vessel’s speed and route. Also, ports are only sometimes fixed, and
routes may need to be adjusted accordingly, making ETA prediction more com-
plex. Limited research is carried out to address ETA predictions in the context
of IWT specifically. Zhong et al. [25] proposed a deep learning method based
on bi-directional long short-term memory recurrent neural networks (BLSTM-
RNNs) for restoring AIS trajectory data and applied it to inland ship trajec-
tory restoration. The paper focused on ETA prediction in IWT. Noman et al.
[14] developed a data-driven approach for ETA prediction using gradient boost-
ing decision trees (GBDT), multi-layer perceptron neural networks (MLP), and
gated recurrent unit NNs (GRU) algorithms trained on past inland waterway
AIS data. The approach was tested for both natural and artificial waterways,
and the results showed that the GRU model outperformed the other models in
accuracy and efficiency. Overall, both papers focused on ETA prediction in the
context of IWT and proposed different machine-learning algorithms to address
this problem. Zhong et al. [25] focused on restoring AIS trajectory data, while
Noman et al. [14] focused on developing a data-driven approach for ETA predic-
tion. Two closely related papers have been identified. The paper of Xie and Liu
[22] proposed a deep learning model based on long short-term memory networks
(LSTMs) for vessel traffic flow prediction in inland waterways. The model was
designed to predict a wide range of traffic flow aspects, including short-term,
long-term, and the influence of water level factors. Yu et al. [24] explored deep
learning approaches for AIS data association in the context of maritime domain
awareness. The paper presented two methods for inferring ship association prob-
ability. One predicts the ship’s position before computing association probability,
while the other compute association probability directly using only longitude,
latitude, and time.

The literature review highlights the significance of accurate ETA predictions

in facilitating port operations. Specifically, the research emphasises sea-going
vessels, mainly in deep sea ports, with relatively less attention paid to inland
waterway shipping. However, predicting precise ETA times in inland shipping
for smaller terminals, river terminals, and transhipment terminals is crucial.
Achieving reliable IWT, comparable to rail and road transportation, is vital to
encourage a modal shift towards IWT.
Unlike a road network with numerous crossings and decision points, a river
network is relatively straightforward. Derrow-Pinion [5] put forth a technique in
their publication that involves dividing a road network into segments and trans-
forming it into a graph. They subsequently employ a Graph Neural Network
(GNN) to analyse the network. Our approach uses real-world AIS data to seg-
ment the river network data, thus simplifying the network. This simplification
allows us to use a NN for travel time prediction.
In line with the works of Fancello [10] and Yu [24], a NN is employed for ETA
time prediction. Yu [24] partitions the Baltic and North Seas close to Copenhagen
into segments and utilises track projection and ETA prediction on the dataset.
This work also creates segments for the inland waterway. Meanwhile, Xie [22]
222 P. Wenzel et al.

incorporates data from the Wuhan Yangtze River, including water levels that
fluctuate during flood periods throughout the year. This model does not consider
water levels. The author optimises the model instead of comparing the prediction
outputs to average calculations.

3 Methodology

This work’s objective is to predict the duration it takes for ships to travel from
the initial location A to the final location B, with a route passing through several
segments s. An A* algorithm is implemented to determine the segments travelled
from A to B, followed by the typical industry practice of computing the average
travel time. Afterwards, a NN is designed and trained to compare the model’s
predictions with the computed average. Our methodology and data preparation
process are outlined in Sect. 3, while the training of our NN model and the
corresponding results are presented in Sect. 4.

3.1 Data Description and Data Preparation

The maritime domain utilises the AIS to enhance ship traffic safety through real-
time broadcasting of vessel information, including identity, speed, location, and
course. For this study, prefiltered AIS data was downloaded from a cooperation
partner via an API call. The received data contains the position data of the ships,
including the following information, MMSI (Maritime Mobile Service Identities,
the unique ship number), Status of the ship (underway, moored, etc.), speed over
ground, course over ground, true heading, time of last update, ETA, destination
(manual input from the skipper), data source, maximum draught of the ship, new
streaming update and location (latitude and longitude). The data were filtered
based on the “Moored” status and vessels, not in motion, requiring a speed over
ground greater than zero.
Segments with a radius of 2.5 km are generated to cover the entire area.
The position data, consisting of start latitude, longitude, and end latitude and
longitude, is mapped to start node ID and end node ID. The position data is then
looped through to eliminate multiple position data for each segment, keeping
only the data closest to the midpoint of the segment. The resulting table is
referred to as segment crossings. The crossing duration is computed from the
start and end times of each segment crossing. For each segment, the number
of ships crossing it is calculated. Segments with less than 50 appearances and
trip durations below 50 s are excluded. From the resulting ship-crossing data,
the stop time and destination were dropped from the dataset. The dataset was
further enriched by adding this additional data, such as compass bearing. Any
rows with missing values were removed to ensure data integrity and accuracy in
the analysis. A NN model is created to predict the duration to travel from one
segment to another.
To improve the reliability of predictions for IWT, it is advisable to make
predictions for individual segments and obtain accurate predictions for complete
 Neural Network ETA Prediction in Inland Waterway Transport 223

ship trips that traverse multiple sections. Therefore, developing and incorporat-
ing trip-level prediction models in addition to segment-level models is recom-
mended. Consequently, an algorithm was created to identify the travels of a ship
and add a trip ID to the dataset. The code iterates over all segment’s crossings
and adds a new value, “trip id”. For each row, it assigns a trip ID based on the
MMSI (an identifier for a vessel), the start time, and the start and end nodes. If
the MMSI changes or the time difference between consecutive rows is more than
1000 s seconds, or the start node of the current row is not the same as the end
node of the previous row, then it increments the trip ID counter. The following
Fig. 1 shows a sample trip in the dataset.

Fig. 1. River network with segment ids

3.2 Average Travel Time Calculation

Similar to Derrow-Pinion [5], the Average Travel Times (ATT) between each
pair of cells are calculated based on the actual times ds for a ship-crossing s.
The ATT is calculated by the sum of the duration of the ship-crossing ds of the
number of start and end node pairs N and then divided by N .
N
i=1 ti
AT T = (1)
N

3.3 Neural Network

Transposing and describing train features in Python is an essential step in

preparing the data for machine learning, as it ensures that the data is organ-
ised in a format suitable for training models and provides valuable information
224 P. Wenzel et al.

about the characteristics of the features. To ensure that the features are organ-
ised in the correct format for training the model, train features are transposed
and described. Using z-score normalisation in the previous approach could have
been more effective in producing accurate results. This paper utilised min-max
normalisation to scale the data between 0 and 1, preserving the original range.
The equation for min-max normalisation is as follows:

x − min(x)
x = (2)
max(x) − min(x)
x is the original value of the data, min(x) is the minimum value of the data,
max(x) is the maximum value of the data, and x is the normalized value of x
between 0 and 1. The numerator, (x − min(x)), subtracts the minimum value
from the original value to measure the distance between the original value and the
minimum value of the data. The denominator, (max(x)−min(x)), calculates the
range of the data by subtracting the minimum value from the maximum value.
By dividing the distance between the original value and the minimum value by
the range of the data, the resulting value is normalised between 0 and 1.
A validation set that is not used for training but to evaluate the model’s
performance on unseen data is created. 80 % of the data are used for training,
20 % are used for validating the model. The actual travel duration was excluded
from both the train and validation datasets.
The used input parameters are the following: the id of the start segment,
the id of the end segment, the Status of the ship (underway, moored, reserved),
Course over ground, True heading, Time of last update, ETA, Destination (man-
ual input from the skipper), Data Source, Maximum draught of the ship, new
streaming update, location (latitude and longitude), compass bearing and the
actual travel time from origin segment to destination segment.
Several correlations can be observed among the provided input parameters.
The course over ground and true heading is typically closely related, representing
the direction of ship movement. The time of last update and the new streaming
update are correlated, with the new streaming update expected to have a more
recent timestamp. The maximum draught of the ship and moored status might
be correlated, as the draught becomes less critical when the ship is stationary.
Furthermore, a correlation exists between the specified destination and the esti-
mated ETA, as the ETA reflects the projected time of reaching the destination.
The created model design is adapted from Chondrodima [4] using similar
parameters for performance evaluation and hyper-parameter selection. This is
a sequential NN model, an artificial deep-learning architecture. The sequential
model is a linear stack of layers, where the output of one layer serves as the
input for the next layer sequentially. The model has three layers, the first dense
layer with 30 output units and 510 trainable parameters. The activation func-
tion is a ReLu function-the second dense layer (hidden layer) with 30 output
units. The activation function is a ReLu function. The third and final dense
layer, with one output unit, represents the model’s output prediction. It has 31
trainable parameters. The activation function is a ReLu function. The model
 Neural Network ETA Prediction in Inland Waterway Transport 225

has 1,471 parameters, all of which are trainable. The parameters of the model
are adaptively adjusted during training based on the patterns and information
contained in the training data. This adaptive learning process allows the model
to learn the appropriate representations and relationships in the data, enabling
it to make better predictions on new, unseen dataThese parameters are adjusted
during training to optimise the model’s performance on the given task. The
model has no non-trainable parameters, meaning all parameters are updated
during training. Figure 2 shows the design of the NN.

Fig. 2. NN with three layers, an input layer, two hidden layers and one output layer

Similar to the work of Xiu [22], the root two widely used criteria are adopted
to measure the error of the predicted data, they are the Root Mean Squared Error
(RMSE) and the Mean Relative Error (MRE). These criteria are commonly used
to evaluate the accuracy of predictive models.

4 Results
A case study was conducted to demonstrate the higher accuracy of neural
network-based ETA prediction for IWT. The code was implemented in Python
within a Google Colab environment, utilising an Intel Xeon CPU @2.20 GHz, 25
GB RAM, a Tesla K80 accelerator, and 12 GB GDDR5 VRAM.
The following section provides an overview of the case study data and an eval-
uation of the model training results. Subsequently, the predictions are compared
with the average results of the entire network and two specific areas.

4.1 Case Study Data

This study’s position data covers January to April 2022 and encompasses a 50km
radius around Rees. Figure 3 shows the segments.
226 P. Wenzel et al.

Fig. 3. Filtered river network segments with applied filters

The resulting area contains the busy area in Duisburg, a crossing close to
Doornenburg, the border crossing in Bimmen, and a straight line without any
stops between Emmerich and Wesel. After data preparation, 82 segments and
152,650 rows of ship-crossing data were obtained. Figure 3 illustrates the remain-
ing segments post-data filtering.

4.2 Model Evaluation

For comparison, two different approaches were employed. The first approach, the
Overall NN (ONN), involved using the entire network to train the model. The
second approach Individual NN (INN), utilised individually trained models for
each start and end node pair. The training and validation are made for the same
in the dataset, allowing for a comprehensive assessment of the performance of
each method.
The model was trained using data from January to March, and the training
process involved 100 epochs. The history epoch loss graph shows the model’s
MAE change over successive epochs and is illustrated in the following Fig. 4.
Similarly, the validation loss graph shows the change in loss on a separate
validation set during training. The validation graph provides insights into how
well the model generalises to new data. Since the validation loss remains consis-
tently low with the training loss, the model performs well on both the training
and validation data. After the 20th epoch, the ONN model’s learning rate is not
improving significantly. For the different INN models, the learning drops at the
80th epoch.
Xie [22] conducted predictions on a more granular level. The evaluation met-
rics for the models are crucial, and the Root Mean Square Error (RMSE) for
the ONN model is 5.8%, and the Mean Relative Error (MRE) is 12.32 %. For
 Neural Network ETA Prediction in Inland Waterway Transport 227

(a) Loss function for ONN (b) 10 sample loss functions for INNs

Fig. 4. Loss and Validation Loss Function shows the differences between ONN and
INN

the INN model, the RMSE is 25%, and the MRE is 35.8%. However, the model
trained in this study produced similar results for a one-day prediction.

4.3 Comparison of Averages to NN

NNs are commonly evaluated by randomly partitioning data into training, val-
idation, and testing sets. However, this section uses historical data to predict
future data. Although the predicted data is also historical, it represents the
future concerning the training data used. A future dataset containing 59,000
entries of start and end segment information for April was obtained and prepro-
cessed for analysis.
t = (s1 , s2 , . . . , sn ) (3)
A trip t contains multiple segment-crossings si , for i = 1 . . . n. Using the struc-
ture of the trip t, the total duration is calculated using the following equation.
Ns

dt = ds (4)
s=1

In (4), Ns is defined as the number of segments per trip. Each segment-crossing

has a duration of ds . Therefore the sum of all segments crossing is the duration
of the trip dt .
Ns

AT Tt = AT Ts (5)
s=1

The averages AT Tt are calculated for all the segment-crossings s in trip t.

Nt
t=1AT Ts
AT T = (6)
Nt
228 P. Wenzel et al.

For all trips Nt in the test dataset, the average AT T is calculated. The following
Eq. 7 shows the prediction error P E for the Average Calculation (AC):
Nt

P EAC = AT T − dt (7)
t=1

P EAC is calculated by substracting the sum of all trip durations dt from the
average travel times.
To create trip duration predictions from the NN, the algorithm iterates over
each segment in the route and predicts the duration of the segment-crossings.
The predicted duration is added to the current timestamp, which serves as an
input to the NN, and this updated timestamp is used to make the next prediction
for the next segment crossings. Predictions are made for ONN, and INN uses the
same methodology. Both the ONN and INN models are employed to generate
the predictions.
Ns

P N Nt = P N Ni (8)
i=1

The prediction duration of the trip for the NN (PNN) is the sum of all
segment crossing predictions.
Nt
i=1 P N Ni
PNN = (9)
Nt
For all trips Nt in the test dataset, the average of PNN is calculated.
Nt

P EN N = P N N − di (10)
i=1

The prediction error for the NN is calculated by subtracting the sum of all
trip durations from the PNN.
The dataset was analysed and segmented into short trips (3 segment cross-
ings), medium trips (9 segment crossings), and long trips (18 segment crossings).
To compare different trip lengths, long trips with more than 18 segment cross-
ings were truncated to 18 segment crossings, ensuring that at least 100 trips
were available for each trip length category.
The results of ETA prediction error using the three different models - AC,
ONN, and INN - on different data sets with varying numbers of segments and
durations. The first column in table 1 indicates the number of crossed segments
Ns per trip t, while the second column represents the duration dt in seconds.
The following three columns show the average predictions made by each model
for the respective data sets. Finally, the last three columns indicate the relative
standard deviation in percentage for each model’s predictions.
 Neural Network ETA Prediction in Inland Waterway Transport 229

Table 1. Results of the comparison of prediction methods

Ns d[s] Prediction error Relative. Std.[%]

AC NN INN AC NN INN
18 6898 1743 820 1220 25.3 11.9 17.7
9 5944 358 343 539 12.8 5.8 9.1
3 3687 143 34 268 2.4 0.9 2.1

The results indicate that the NN methods have smaller prediction errors than
the AC method for all three trip length sets. The relative standard deviation is
the smallest for the ONN method for the 3-segment and 9-segment crossings,
while the INN method has the lowest relative standard deviation for the 18-
segment crossing. Overall, the results suggest that the ONN and INN method
are slightly more accurate in predicting segment-crossing duration than the AC
method.
Computational experiments have been performed on two selected regions to
provide an additional comparison, as shown in Fig. 5.

Fig. 5. Area of Kleve (green) and Duisburg (red) for comparison. (Color figure online)

The region surrounding Duisburg is assigned the abbreviation “DUI,” while

the red region near Kleve is designated as “KLE.” Given the smaller geographical
scope and absence of long-distance port journeys, emphasis is placed on trips
with three and 9-segment crossings. After segmenting the data, 100 random
samples were selected for each 3-segment and 9-segment crossing trip.
Similar to Table 1, the prediction errors are shown in Table 2 for the three
methods: AC, NN and INN.
230 P. Wenzel et al.

Table 2. Results of the comparison of prediction methods for DUI and KLE

Ns d[s] Prediction error Relative. Std.[%]

AC NN INN AC NN INN
9 (DUI) 5500 745 652 742 13.5 11.9 13.5
3 (DUI) 3105 102 38 130 3.2 1.2 4.1
9 (KLE) 6085 774 10 193 12.7 0.1 3.1
3 (KLE) 3578 704 37 111 19.7 1.0 3.1

Trip durations are, in general, a bit shorter in the DUI area. The largest
inland port in Europe is located in Duisburg, so the dataset includes shorter
trips that involve docking at specific locations within the port. In the case of
the Klave region, noisy data may be attributed to crossing the border between
Germany and the Netherlands.

5 Conclusion

Previous studies have emphasised the significance of enhancing ETA prediction

accuracy for efficient terminal and inland port operations [3,6]. Xie [22] has
showcased the effectiveness of using LSTM models for ETA prediction. However,
related studies mainly focused on fine-tuning predictions for a single segment
crossing rather than predicting entire trips.
The paper discusses the advantages of precise ETA prediction in IWT and
uses real-world data to train a dedicated NN. The results demonstrate a con-
siderable enhancement in ETA prediction compared to conventional travel time
averages. Additionally, this work showcases how predictions can be utilised for
complete ship trips crossing multiple segments using an A* path algorithm.
There are certain limitations that should be considered. Historical data’s
availability and representativeness could impact the predictions’ generalizabil-
ity. Additionally, the generalizability of the results may be limited if the exper-
imentation is conducted only with specific segment sizes. Including additional
input parameters such as weather data and river, depths could improve predic-
tion accuracy. While effective in improving neural network model performance,
hyperparameter selection techniques have certain limitations. These include com-
putational complexity, sensitivity to initial conditions, and the need for careful
consideration of the search space.
Future work should aim at enhancing the accuracy of results by download-
ing more historical data. In addition, further experimentation with smaller and
larger segments will be conducted. More input parameters, including weather
data and river depths, will be added to improve the accuracy of predictions.
Future work will further incorporate infrastructural information, such as closed
bridges and locks at certain timestamps, to provide more precise predictions.
Finally, the results of this study will be compared across different river areas to
 Neural Network ETA Prediction in Inland Waterway Transport 231

validate the model’s effectiveness. It is anticipated that Graph Neural Networks

(GNNs) will be utilised for ETA prediction in IWT. The advantage of GNNs lies
in their ability to learn from graph-structured data and model the dependencies
among nodes in a graph. In the case of IWT, the river network can be repre-
sented as a graph where the nodes represent the various segments and the edges
represent the waterways connecting them. By incorporating this graph structure
into the GNN, the accuracy of ETA predictions is expected to be significantly
improved.

References
1. Alessandrini, A., Mazzarella, F., Vespe, M.: Estimated time of arrival using his-
torical vessel tracking data. IEEE Trans. Intell. Transp. Syst. 20(1), 7–15 (2019).
https://doi.org/10.3390/app11052429
2. Bautista-Sánchez, R., Barbosa-Santillan, L.I., Sánchez-Escobar, J.J.: Method for
Select Best AIS Data in Prediction Vessel Movements and Route Estimation. Appl.
Sci. 11(5), 2429 (2021). https://doi.org/10.3390/app11052429
3. Bouchard, P., Moros-Daza, A., Voß, S.: Simulation of an AIS system for the port of
Hamburg. In: Mes, M., Lalla-Ruiz, E., Voß, S. (eds.) ICCL 2021. LNCS, vol. 13004,
pp. 21–35. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-87672-2 2
4. Chondrodima, E., Georgiou, H., Pelekis, N., Theodoridis, Y.: Particle swarm opti-
mization and RBF neural networks for public transport arrival time prediction
using GTFS data. Int. J. Inf. Manage. Data Insights 2(2), 100086 (2022). https://
doi.org/10.1016/j.jjimei.2022.100086
5. Derrow-Pinion, A., et al.: ETA prediction with graph neural networks in google
maps. In: Proceedings of the 30th ACM International Conference on Information &
Knowledge Management, pp. 3767–3776 (2021). https://doi.org/10.1145/3459637.
3481916
6. El Mekkaoui, S., Benabbou, L., Berrado, A.: Predicting ships estimated time of
arrival based on AIS data. In: Proceedings of the 13th International Conference on
Intelligent Systems: Theories and Applications, pp. 1–6 (2020)
7. El Mekkaoui, S., Benabbou, L., Berrado, A.: Machine learning models for effi-
cient port terminal operations: case of vessels’ arrival times prediction. IFAC-
PapersOnLine 55(10), 3172–3177 (2022)
8. El Mekkaoui, S., Benabbou, L., Caron, S., Berrado, A.: Deep learning-based ship
speed prediction for intelligent maritime traffic management. J. Mar. Sci. Eng.
11(1), 191 (2023). https://doi.org/10.3390/jmse1101019
9. Eurostat: eurostat Statistics explained. https://ec.europa.eu/eurostat/statistics-ex
plained/index.php?title=Inland waterway transport statistics#Eurostat#Statisti
csExplained
10. Fancello, G., Pani, C., Pisano, M., Serra, P., Zuddas, P., Fadda, P.: Prediction
of arrival times and human resources allocation for container terminal. Maritime
Economics & Logistics 13(2), 142–173 (2011). https://doi.org/10.1057/mel.2011.3
11. Huang, C., Huang, Y., Yu, Y., Xiao, B.: Predicting liner arrival time based on deep
learning. In: 2021 IEEE 3rd International Conference on Civil Aviation Safety and
Information Technology (ICCASIT), pp. 1127–1131. IEEE (2021)
12. Jing, M., Zheng, W.: Economic development advantages of low-carbon economy
under waterway transportation. IOP Conf. Ser. Mater. Sci. Eng. 780(6), 062029
(2020). https://doi.org/10.1088/1757-899X/780/6/062029
232 P. Wenzel et al.

13. Jonkeren, O., Rietveld, P., van Ommeren, J., te Linde, A.: Climate change and
economic consequences for inland waterway transport in Europe. Reg. Environ.
Change 14(3), 953–965 (2013). https://doi.org/10.1007/s10113-013-0441-7
14. Noman, A., Heuermann, A., Wiesner, S., Thoben, K.D.: Towards data-driven GRU
based ETA prediction approach for vessels on both inland natural and artificial
waterways. vol. 2021, pp. 2286–2291 (2021). https://doi.org/10.1109/ITSC48978.
2021.956488
15. Ogura, T., Inoue, T., Uchihira, N.: Prediction of arrival time of vessels considering
future weather conditions. Appl. Sci. 11(10), 4410 (2021). https://doi.org/10.3390/
app1110441
16. Pani, C., Fadda, P., Fancello, G., Frigau, L., Mola, F.: A data mining approach
to forecast late arrivals in a transhipment container terminal. Transport 29(2),
175–184 (2014). https://doi.org/10.3846/16484142.2014.930714
17. Park, K., Sim, S., Bae, H.: Vessel estimated time of arrival prediction system based
on a path-finding algorithm. Marit. Transp. Res. 2, 100012 (2021). https://doi.org/
10.1016/j.martra.2021.100012
18. Valero, C.I., Martı́nez, n., Oltra-Badenes, R., Gil, H., Boronat, F., Palau, C.E.:
Prediction of the estimated time of arrival of container ships on short-sea shipping:
a pragmatical analysis. IEEE Latin Am. Trans. 20(11), 2354–2362 (2022)
19. Veenstra, A., Harmelink, R.: On the quality of ship arrival predictions. Marit. Econ.
Logistics 23(4), 655–673 (2021). https://doi.org/10.1057/s41278-021-00187-6
20. Wehrle, R., Wiens, M., Schultmann, F.: A framework to evaluate systemic risks
of inland waterway infrastructure. Prog. Disaster Sci. 16, 100258 (2022). https://
doi.org/10.1016/j.pdisas.2022.100258
21. Wiegmans, B., Konings, R.: Intermodal Inland waterway transport: modelling con-
ditions influencing its cost competitiveness. Asian J. Shipping Logistics 31(2), 273–
294 (2015). https://doi.org/10.1016/j.ajsl.2015.06.006
22. Xie, Z., Liu, Q.: LSTM networks for vessel traffic flow prediction in inland water-
way. In: 2018 IEEE International Conference on Big Data and Smart Computing
(BigComp), pp. 418–425 (2018). https://doi.org/10.1109/BigComp.2018.00068
23. Yu, J., et al.: Ship arrival prediction and its value on daily container terminal oper-
ation. Ocean Eng. 157, 73–86 (2018). https://doi.org/10.1016/j.oceaneng.2018.03.
038
24. Yu, J.Y., Sghaier, M.O., Grabowiecka, Z.: Deep learning approaches for AIS data
association in the context of maritime domain awareness. In: 2020 IEEE 23rd Inter-
national Conference on Information Fusion (FUSION), pp. 1–8 (2020). https://doi.
org/10.23919/FUSION45008.2020.9190283
25. Zhong, C., Jiang, Z., Chu, X., Liu, L.: Inland ship trajectory restoration by recur-
rent neural network. J. Navig. 72(6), 1359–1377 (2019). https://doi.org/10.1017/
S0373463319000316
Vehicle Routing
 A Regret Policy for the Dynamic Vehicle
Routing Problem with Time Windows

Peter Dieter(B)

Paderborn University, Warburger Str. 100, 33098 Paderborn, Germany

[email protected]

Abstract. In this work, we present a regret policy for the dynamic vehi-
cle routing problem with time windows (DVRPTW) in which customer
order arrivals are revealed over a day and which has been the considered
problem in the EURO Meets NeurIPS 2022 Vehicle Routing Competi-
tion. The problem requires two types of decisions: Dispatching orders
and planning the respective vehicle routes. The objective is to minimize
the travel time while dispatching all orders over the day and adhering to
the respective time windows. The regret policy we present in this paper
focuses on the former decision, namely the dispatching of orders, as the
routing can then be performed by a regular VRPTW solver. The basic
idea of the policy is to calculate a regret value for each order which rep-
resents possible cost savings that we would miss when dispatching the
order immediately. To attain this regret value, we make use of the cus-
tomer distribution from which orders are sampled. If the value is below
a predefined threshold, the order is dispatched and routes are planned
with a state-of-the-art VRPTW solver. Otherwise, the order needs to
be dispatched at a later stage. The proposed policy outperforms several
benchmark policies and ranked 6th place among approximately 50 teams.
Furthermore, the proposed regret policy has a high level of explainability,
is simple to implement, and can be generalized to other problems in the
domain of dynamic vehicle routing.

Keywords: Stochastic dynamic vehicle routing · Anticipatory vehicle

routing · Vehicle dispatching

1 Introduction

The logistics industry is disrupted by customers’ demand for services such as

same-day delivery (SDD). For example, Amazon has introduced “Prime Now”,
a service that promises customers the delivery of goods in one hour [8]. Accord-
ing to a survey in the United States, more than half of the respondents say that
they are interested in using same-day delivery but have a limited willingness
to pay extra for this service [4]. Due to these developments, dynamic vehicle
routing is becoming increasingly important [12] and efficient same-day delivery
planning can give last-mile delivery companies a competitive edge. However,
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 235–247, 2023.
https://doi.org/10.1007/978-3-031-43612-3_14
236 P. Dieter

efficient planning is difficult since customer orders arrive in the course of a day
and before some decisions are already made and implemented. Next to plan-
ning vehicle tours, it is important to decide which customer orders to dispatch,
especially when orders have time windows. Too early dispatching might lead to
inefficiencies since it limits the possibility of consolidation. Too-late dispatch-
ing on the other hand might lead to limited flexibility in planning tours since
time-windows need to be adhered to. This paper describes a regret policy for
the dynamic vehicle routing problem with time windows (DVRPTW), which
has been the posed problem in the EURO Meets NeurIPS 2022 Vehicle Routing
Competition [9]. In the considered problem, orders are sampled from a known
customer distribution over a day. Once an hour, it needs to be decided which
orders need to be dispatched and which ones to postpone to consolidate with
future requests that may arrive later. After it has been decided which orders
are dispatched, feasible routes for these orders need to be created. Customer
time windows are hard and an unlimited vehicle fleet is assumed. The overall
goal is to minimize the time traveled during all decision epochs. The policy pre-
sented in this paper only focuses on the dispatching decision, as the routing task
is then a standard VRPTW, for which there exists an abundance of literature
[1]. This is done by calculating a regret value for each undue order, i.e., orders
which can be postponed, considering the total set of customers from which orders
are randomly sampled. Furthermore, we describe methods to extend the regret
value with information on time windows and other undue orders. The result-
ing VRPTW instance is then solved by a state-of-the-art hybrid genetic search
(HGS) algorithm developed by [10], which is an extension of the HGS solver
which has been proven successful for other VRP variants [16,17]. We show that
the proposed policy outperforms two benchmark policies. Due to its high level of
explainability, easiness of implementation, and generalizability to other problem
settings, we believe that the proposed policy can serve as a benchmark policy for
other studies on dynamic vehicle routing, in which anticipating future customer
orders is crucial. Furthermore, several possible policy extensions exist and could
be considered in future work. The further structure of the paper is as follows:
Sect. 2 includes related work. The problem is described in Sect. 3. In Sect. 4, we
present the developed regret policy. In Sect. 5, we evaluate the policy by com-
paring it to the two benchmark policies. The paper concludes with a summary
and an outlook in Sect. 6.

2 Related Work

The body of literature on dynamic vehicle routing is vast and covers differ-
ent problem settings as well as various algorithmic contributions. For a general
overview of stochastic dynamic vehicle routing, we refer to [13] and [5]. In this
section, we first examine related work on the concept of regret in dynamic vehi-
cle routing. Second, we present the winning approaches of the EURO Meets
NeurIPS 2022 Vehicle Routing Competition challenge.
 A Regret Policy for the DVRPTW 237

To the best of our knowledge, the study of [3] is the first one on dynamic
vehicle routing that used the concept of regret. The authors consider a prob-
lem with the objective to maximize the number of orders served. The problem,
therefore, involves deciding which orders to serve and which ones not to serve,
considering time windows and a limited vehicle fleet. The authors suggest a
multi-scenario approach in which scenarios are created by sampling potential
future orders from a known customer distribution. A regret value for each order
already in the system o is then calculated which is defined as the difference in
objective value between the optimal solution considering all orders (i.e., already
in the system or drawn in the scenario) and the best solution serving request o.
This idea is further developed in [14].
[6] apply a multi-scenario approach for a problem in which the main objective
is to maximize the number of orders served and is more similar to the problem
considered in this paper, as orders can be postponed to later epochs. Rather than
solving a problem for each scenario once and then choosing one of the solutions,
they suggest a branch-and-regret heuristic which identifies structural decisions
(e.g., postponing an order) for each scenario and branches on them until all sce-
nario problems have solutions where all known orders are served according to
the same plan. The concept of regret represents the possible improvement that
might be gained from postponing an order to be dispatched in a later decision
epoch. This is similar to our proposed regret policy. However, our suggested
approach does not require sampling orders to create scenarios but directly uses
the known customer distribution, which allows for faster computations. Further-
more, in contrast to [6], we consider a problem in which all orders need to be
served and the total travel time over all decision epochs must be minimized.
The two most successful approaches to the challenge both applied machine
learning to the problem. In the winning approach, [7] reformulated the problem
into a prize-collecting VRPTW, i.e., a prize is assigned to each order during an
epoch and this prize is collected if an order is visited in the respective epoch.
The authors then apply imitation learning, by learning the prizes for each order
based on solutions with complete information. Instead of learning from solutions
with full information, [11] apply a reinforcement learning approach that learns
how to modify the costs of traversing arcs to an order in a respective epoch.
While the two approaches outperformed our proposed policy (by 4.95% and
2.4%, respectively), the regret policy is easier to implement and offers a higher
degree of explainability.

3 Problem Description
The problem at hand can be defined as a multi-stage VRPTW. We consider a
discrete-time horizon H = {1, ..., n} such that each decision occurs at a discrete
time (decision epoch) h ∈ H. Further, let Gh = (Vh , Ah ) be a complete graph,
where Vh = {0, Nh } is the set of nodes, Nh is a set of order nodes in epoch h
and 0 identifies the depot. The set of arcs is denoted by Ah = {(i, j) : i, j ∈ Vh }.
The time of traversing arc (i, j) is given by cij . Each order i possesses a time-
window {ei , li } where ei denotes the earliest arrival time and li the latest arrival
238 P. Dieter

time. Moreover, service times and demands for order i are given by si and qid ,
respectively. A vehicle from an unlimited fleet is denoted by k ∈ Kh , where Kh
is a set of vehicles in epoch h. The fleet is homogeneous and a vehicle possesses
a capacity of q v . In each decision stage h, a set of new orders Oh arrives. Orders
which arrived before epoch h but have not been dispatched yet are denoted by
Fh and the set of all available orders in epoch h is then given by Nh = Oh ∪ Fh .
Out of these orders, some might be due as dispatching them in a later epoch,
would lead to non-adherence to its time window. We define the set of these due
orders as Dh ⊆ Nh . In the last decision epoch n, all orders are due, so Nn = Dn .
The decision variable xijk denotes if arc (i, j) : i, j ∈ Vh is traversed by vehicle
k and tik denotes the arrival time of vehicle k at order i. A decision vector in
epoch h defined as xh = {xijk : i, j ∈ Vh ; k ∈ Kh }. As xh indicates which orders
are dispatched in h and which are not, the set of not dispatched orders at epoch
h + 1, Fh+1 , is dependent on xh . For example, if we would dispatch all orders in
epoch h, Fh+1 would be an empty set. The problem in epoch h can be formulated
as follows:
  
minimize cij · xijk + E[Q(xh )] (1)
i∈Vh j∈Vh k∈K

subject to
 
xjik = 1 ∀i ∈ Dh (2)
j∈Vh k∈Kh
j=i
 
xijk − xjik = 0 ∀k ∈ Kh ; j ∈ Vh (3)
i∈Vh i∈Vh
 
qid · xjik ≤ q v ∀k ∈ Kh (4)
j∈Vh i∈Nh

tik + cij + si − M · (1 − xijk ) ≤ tjk ∀i ∈ Vh ; j ∈ Nh ; ∀k ∈ Kh (5)

ei ≤ tik ≤ li ∀i ∈ Nh ; k ∈ Kh (6)
tik ≥ 0 ∀i ∈ Nh ; k ∈ Kh (7)
xijk ∈ {0, 1} ∀i, j ∈ Vh ; k ∈ Kh (8)

where Q(xh ) is the future costs of implementing decision xh in epoch h. The

objective is to minimize the time the vehicle travels and waiting times are not
considered. Constraint (2) assures that each due order is visited. Constraint
(3) assures that if a vehicle visits a node it must also leave this node. Not
exceeding the vehicle capacity is assured by Constraint 4. Constraints (5) and
(6) assure that each order is visited in its time window and also serve as sub-
tour elimination constraints, where M is a large enough number. Constraints
(7) and (8) are non-negativity and binary constraints, respectively. The sets,
parameters, and decision variables are summarized in Table 3 which is to be
found in the Appendix. Further, we note that the problem is a sequential decision
process and an alternative problem formulation would be to model it as a Markov
 A Regret Policy for the DVRPTW 239

Decision Process (MDP). In such a formulation, a state would contain all orders
revealed but not yet dispatched and the action space would contain all feasible
vehicle tours. A transition between two decision epochs would occur by removing
all dispatched orders from the state and adding all orders that newly entered
the system. The objective is to find a policy that minimizes the expected cost
of serving all orders. For a detailed problem formulation as an MDP, we refer
to [2].

4 Regret Policy

In this section, we first describe the general algorithmic procedure of the regret
policy. We then describe the regret function and explain two extensions of the
basic regret function, namely the incorporation of information about time win-
dows and undue orders. An open-source Python implementation of the policy is
available on GitHub1 .

4.1 General Procedure

The general procedure of the algorithm is described in Algorithm 1. The pro-

cedure requires a set of due orders Dh ⊆ Nh , which are orders that need to
be dispatched immediately in epoch h as postponement would lead to a time
window violation. Moreover, the following is required as input: a set of undue
orders Uh ⊆ Nh , a set of all customers from which orders are drawn, C, and a
threshold value t which can further be tuned. To enter the while loop, we set the
best (lowest) regret value r to t. While the best regret value is equal or below
the threshold, we do the following: For each undue order u ∈ Uh we calculate its
regret value with the GetRegret function, which takes the order u, Dh , and C
as input. This function is further explained in Subsect. 4.2. If the lowest regret
value is below threshold value t, the respective order denoted by c (the one with
the lowest regret value) is marked as dispatched, added to the set of due orders
Dh , and removed from the set of undue orders Uh . This procedure is repeated
until the regret values of all orders are above t, or Uh is an empty set. Therefore,
the final set of Dh includes all orders which are dispatched, and consequently,
feasible routes for these orders are determined by the HGS solver [10,15].

4.2 Regret Function

The motivation behind the proposed regret function is the following: If an undue
order u ∈ Uh is dispatched in epoch h, we can expect that u will be served con-
secutively before or after its closest order d ∈ Dh which is marked as dispatched.
Closest is here defined by the lowest average arc costs from u to d and vice versa.
Let cud be the time (costs) needed to travel from the customer of order u to the
customer of order d. We then approximate that dispatching order u will result
1
https://github.com/PeterDieter/RegretPolicy DVRPTW.
240 P. Dieter

Algorithm 1. Regret policy procedure

Input: Set of due orders Dh , set of undue orders Uh , set of all customers C, threshold
value t
1: r ← t
2: while r ≤ t do
3: R ← {}
4: for u ∈ Uh do
5: R ← R ∪ GetRegret(u, Dh , S)
6: end for
7: r, c ← min(R), argmin(R)
8: if r ≤ t then
9: Dh ← Dh ∪ c
10: Uh ← Uh \ c
11: end if
12: end while
Output: Set of orders to dispatch Dh

in additional costs of cud +c

2
du
, assuming that cud does not deviate much from
cdu and no further orders being dispatched. We note that this is only a simple
approximation, as order u might be on the route to a dispatched order, which
might result in lower costs. As order u is not due yet, better (closer) orders might
arrive in the future. Let B ⊆ C be those closer customers. The improvement of
routing u consecutively after/before i ∈ B compared to d is given by:
cud + cdu cui + ciu
− (9)
2 2
These improvements can be seen as a regret, i.e., possible improvements which
we would miss when dispatching u immediately. As the regret also depends on
the probability that a better customer will be drawn until order u is due, we
adjust the regret by this probability: Let no be the number of orders that are
expected to be drawn until order u is due. The probability that a customer is
1
drawn in a single draw is |C| and its counter probability is, therefore, 1 − |C|
1
.
The probability that a customer is not drawn in no drawings is then given by
1 no
(1− |C| ) . The probability that a customer is drawn in no drawings is therefore,
1 − (1 − |C|
1 no
) . The adjusted regret value Ru of order u is therefore given by:

 cud + cdu cui + ciu cui + ciu cud + cdu 1 no

Ru = ( − [ < ]) · (1 − (1 − ) ) (10)
2 2 2 2 |C|
i∈C

where square brackets are Iverson brackets, i.e., its inner value is 1 if the condi-
tion ( cui +c
2
iu
< cud +c
2
du
) holds and 0 otherwise.
To provide further explanation on the regret function, it is visualized in Fig. 1,
where the regret value for order i is determined. There are two orders which are
marked as due. The distance cid is an approximation of the additional routing
costs that would occur when order i is dispatched together with the two due
 A Regret Policy for the DVRPTW 241

orders. However, orders from two closer customers (1 and 2) might be drawn in
the future and 99 orders are expected to be drawn until order i is due. If the
calculated regret value of 1.2 is below the threshold, the order will be dispatched.

Fig. 1. Illustration of the regret function.

4.3 Incorporation of Time Windows

So far, the regret policy only takes time windows into account for determining
when orders become due. However, neglecting time windows can lead to inef-
ficient route plannings, in which vehicles need to wait before an order can be
served. Therefore, we multiply the regret value with a term that penalizes the
time a vehicle would need to wait if serving order u directly before/after order d.
We note that also an additive term could be added, but pretests have shown that
a multiplicative term leads to better solutions. Let [ed , ld ] be the time window of
order d and [eu , lu ] of order u, respectively. The minimal waiting time for serv-
ing order u directly before order d is given by wdu = max(0, eu − ld + cdu ) and
respectively, wud is given by max(0, ed − lu + cud ). The minimum waiting time
of serving u directly before d or vice versa is then min(wdu , wud ). Therefore, we
update the regret value as follows:

Ru ← Ru · (1 + min(wdu , wud )) · h2 (11)

where h2 is a hyperparameter that determines how much waiting time is penal-
ized. High values of h2 lead to higher penalizations and its value should be
carefully tuned.

4.4 Incorporation of Undue Orders

In the basic formulation of the regret function, other orders which are not due
Uh are not taken into account. However, these known undue orders provide
242 P. Dieter

valuable information and should therefore be incorporated. Let us assume the

following example: Two orders u and v are very close to each other and have
compatible time windows. Moreover, u is due in one epoch while v is due in
5 epochs. Therefore, the regret value for v is likely to be high and the order
might not be dispatched immediately. On the contrary, as order u is due soon,
its regret value will likely be low and the order might be dispatched immediately.
Orders u and v would therefore not be dispatched together, which might result
in inefficient routes. The procedure of incorporating Uh into the regret function
is similar to the incorporation of the set of all customers C described in the
previous subsection. However, there are two important differences: Instead of
taking into account all customers C, we now only consider customers of orders
which are in the system and not dispatched yet. Let d ∈ Dh be again the closest
due order to u as defined by the lowest average arc costs from u to d and vice
versa. The subset of better orders which are not due is then defined as B ⊆ Uh
where:
cuv + cvu cud + cdu
< ∀v ∈ Uh . (12)
2 2
Furthermore, as orders B are already in the system, we do not correct the value
with a probability term. Therefore, we update the regret value as follows:
 cud + cdu cui + ciu
Ru ← Ru + ( − ) · h1 (13)
2 2
i∈B

where h1 is a hyperparameter that determines the weight given to possible

improvements by using the information of undue orders.

5 Experiments and Discussion

5.1 Experimental Setup

In this section, we evaluate the regret policy by comparing it to two benchmark

policies. These benchmark policies are a greedy dispatching rule which dispatches
orders as soon as possible and a lazy dispatching rule which dispatches only
orders which are due. To tune hyperparameters for our proposed regret policy,
we applied a grid search. Suitable ranges used in the grid search have been
determined in pretests. The grid search ranges are presented in Table 1). This
search resulted in the following best combination: t = 0.08, h1 = 0.01, and
h2 = 0.002.
We test the policies on 50 randomly chosen instances which derive from real-
world instances and are publicly available [9]. The problem instances derive from
data of ORTEC, with an explicit duration matrix giving (non-euclidean) realistic
road driving times between orders. The number of customers from which orders
are sampled is between 200 and 1000. Further, 100 orders per epoch are randomly
drawn from the given customer distribution and a typical instance has between
5 and 7 epochs. Moreover, we apply a random seed value of 1 (see [9] for further
 A Regret Policy for the DVRPTW 243

Table 1. Grid search ranges

Parameter Values
t {0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10}
h1 {0.005, 0.01, 0.02, 0.03, 0.04}
h2 {0.01, 0.004, 0.002, 0.001, 0.0005}

information on the systems’ environment). The time limit for the HGS solver
is set to 10 s for all policies (regret and benchmark policies). Experiments are
performed on an Intel Core i7 with 2.6GHz and 12 GB RAM.

5.2 Results

The results are summarized in Table 2, which shows the improvements of the
policy in the row, compared to the policy in the column. Additionally, the per-
centage of instances in which one policy led to better results is shown in brackets.
Detailed results over all tested instances can be found in the Appendix (Table 4).
On average, the proposed policy leads to improvements of around 8.45% com-
pared to the greedy benchmark and 40.88% compared to the lazy benchmark.
Further, the regret policy always leads to better results in all cases compared
to the greedy benchmark and in 47 out of 50 instances (94%) compared to the
greedy policy. Moreover, we can deduce that a greedy policy performs better
than a lazy policy (improvement of 35.14%).

5.3 Discussion

Firstly, we can see from the results that the regret policy outperforms the two
benchmarks to a large extent. Further, we can see from the results that a greedy
policy performs better than a lazy policy. Therefore, it might be better to attain
some slack in the routing decision compared to trying to maximize the con-
solidation of orders by dispatching them as late as possible. The regret policy
seems to achieve a better balance between those two extremes. In fact, our pro-
posed method can be seen as a combination of these two benchmark policies.
If the threshold value is 0, only due orders will be dispatched (lazy policy), if
the threshold value is infinitely big, all orders will be dispatched (greedy policy).
This perspective contributes to the explainability of the proposed method since
the greediness/laziness of our method is determined by the threshold value and
can be tuned accordingly. In terms of ease of implementation, the regret policy
remains fairly simple compared to the two benchmarks, since standard opera-
tions can be used to calculate the regret. However, this simplicity also comes
with a price, since it is difficult to capture all relevant information in a (simple)
function. In the following Section (6), we mention how the regret function could
be extended and further propose hybridizations of our proposed method with
other methods.
244 P. Dieter

Table 2. Policy Comparison

Greedy benchmark Regret policy

Lazy benchmark 35.14% (100%) 40.88% (100%)
Greedy benchmark – 8.45% (94%)

6 Conclusion

In this paper, we present a regret policy for the dynamic vehicle routing problem
with time windows (DVRPTW). In the studied problem, orders are sampled from
a known customer distribution over a day. Once an hour, it needs to be decided
which orders need to be dispatched and which ones to postpone. After it has been
decided which orders are dispatched, feasible routes for these orders need to be
created. The policy presented in this paper considers the first decision stage, i.e.,
determining which orders to dispatch, by anticipating orders that might arrive
in the future.
We first introduce the general procedure of the policy and the basic regret
function, which is applied to determine a regret value for each undue order, i.e.,
orders that can be postponed. This regret value represents possible improvements
that might be gained by postponing the order and consolidating it with future
orders. If the regret value is below a certain threshold, the order is dispatched.
Furthermore, we present the methods applied to incorporate information about
other undue orders and time windows in the regret function. The developed pol-
icy consistently outperforms benchmark policies such as a greedy dispatching
rule which dispatches all orders as soon as possible or a lazy dispatching rule
which dispatches only orders which are due. Due to its high level of explainabil-
ity, ease of implementation, and generalizability to other variants of dynamic
vehicle routing, the proposed regret policy can serve as a benchmark policy for
other work on anticipatory vehicle routing. Moreover, multiple extensions of the
proposed policy could be investigated. Currently, only the closest due order is
regarded to determine the regret value. Future work could look for ways to incor-
porate the entire set of due (or already dispatched) orders to determine regret
values. Furthermore, it could be tried to better account for time windows, e.g., by
immediately constructing an initial route plan in the proposed policy. Another
possible extension is the development of dynamic thresholds. Even though the
time until an order is due is implicitly accounted for in the regret policy, dynamic
thresholds could further improve the policy. For example, it might be beneficial
to release orders more easily when the number of epochs the orders are in the
system is low, as this results in wider time windows for the routing decision,
and consequently, fewer vehicles could be used. Methods such as Value Func-
tion Approximation (VFA) could be used to determine these state-dependent
thresholds. Another possible hybridization of methods would be to reformulate
the problem as a price collection VRPTW and use the proposed regret function
to determine prizes for undue orders.
 A Regret Policy for the DVRPTW 245

Appendix
Table 3. Sets, Parameters and Decision Variables

Notation Definition
Sets
H Set of decision epochs, H := {1, ..., n}
Nh Set of all known orders in epoch h
Oh Set of orders that arrived in epoch h
Fh Set of orders that arrived before epoch h but have not been dispatched yet
Dh Set of all orders that are due in epoch h, Dh ⊆ Nh
Vh Set of all nodes in epoch h, where 0 is the Depot, Vh := {0} ∪ Nh
Kh Set of vehicles in epoch h
Parameters
cij Travel time from node i ∈ Vh to node j ∈ Vh
qv Vehicle capacity (same for all vehicles)
qid Demand of order i ∈ Nh
si Service time for serving order i ∈ Nh
[ei , li ] Time window of order i ∈ Nh
Decision Variables
xijk = 1 if vehicle k ∈ Kh travels from node i ∈ Vh to node j ∈ Vh
(0 otherwise)
tik arrival time of vehicle k ∈ Kh at node i ∈ Vh

Table 4. Objective values over all tested instances.

Instance Lazy benchmark Greedy benchmark Regret policy

n204-k12 708992 292545 260257
n206-k12 552684 350988 326561
n214-k12 443799 309934 272562
n214-k15 942720 308814 317443
n225-k14 606387 376851 359847
n225-k16 630183 442811 418218
n230-k15 1399494 728339 717362
n237-k16 509679 348427 313138
n237-k20 501440 327922 303989
n238-k30 583763 276100 281693
n241-k18 584001 452830 373796
n250-k15 654395 425656 386436
n254-k18 858338 328418 312425
n254-k20 555495 363851 344465
n270-k18 349255 255878 226454
(continued)
246 P. Dieter

Table 4. (continued)

Instance Lazy benchmark Greedy benchmark Regret policy

n290-k18 545620 379116 396671
n291-k16 641405 474666 422344
n299-k21 555080 331866 292910
n301-k16 821085 552070 519155
n302-k20 898116 435965 376439
n302-k25 753785 551246 481033
n308-k35 966468 468322 436023
n309-k19 605191 351529 324527
n311-k25 565430 357539 341638
n313-k20 531487 329314 315648
n316-k19 655239 437355 380248
n318-k20 366332 254295 223473
n319-k25 645225 483209 454389
n320-k20 833531 464472 426446
n321-k35 662557 536213 472861
n325-k20 394283 300151 255569
n326-k25 1032144 618376 572630
n327-k20 449911 283930 255615
n329-k25 696282 485038 410575
n332-k25 362024 254561 233209
n335-k22 624164 425939 402006
n338-k21 470721 371466 344624
n338-k24 704357 506885 437091
n339-k20 553098 402506 368421
n349-k30 601643 478866 404215
n353-k21 556843 368884 335157
n363-k22 631352 493885 431556
n413-k35 624703 390816 358661
n418-k28 989808 521679 495567
n420-k26 644202 462920 397164
n430-k35 628483 401273 366917
n455-k45 593949 372825 341316
n457-k30 597019 394398 373642
n491-k36 517331 431485 374174
n504-k38 636083 386395 370372
 A Regret Policy for the DVRPTW 247

References
1. Baldacci, R., Mingozzi, A., Roberti, R.: Recent exact algorithms for solving the
vehicle routing problem under capacity and time window constraints. Eur. J. Oper.
Res. 218(1), 1–6 (2012). https://doi.org/10.1016/j.ejor.2011.07.037, https://www.
sciencedirect.com/science/article/pii/S0377221711006692
2. Baty, L., Jungel, K., Klein, P.S., Parmentier, A., Schiffer, M.: Combinatorial opti-
mization enriched machine learning to solve the dynamic vehicle routing problem
with time windows (2023)
3. Bent, R., Van Hentenryck, P.: Regrets only! Online stochastic optimization under
time constraints. In: Proceedings of the 19th National Conference on Artifical
Intelligence, pp. 501–506. AAAI’04, AAAI Press (2004)
4. Ecker, T., Hans, M., Neuhaus, F., Spielvogel, J.: Same-day delivery: ready for
takeoff, January 2020. https://www.mckinsey.com/industries/retail/our-insights/
same-day-delivery-ready-for-takeoff
5. Hildebrandt, F.D., Thomas, B.W., Ulmer, M.W.: Opportunities for reinforcement
learning in stochastic dynamic vehicle routing. Comput. Oper. Res. 150, 106071
(2023). https://doi.org/10.1016/j.cor.2022.106071
6. Hvattum, L.M., Løkketangen, A., Laporte, G.: A branch-and-regret heuristic for
stochastic and dynamic vehicle routing problems. Networks 49(4), 330–340 (2007).
https://doi.org/10.1002/net.20182
7. Jungel, K., Baty, L., Klein, P.: Structured Learning for the EURO Meets NeurIPS
2022 Vehicle Routing Competition. Technical report (2022)
8. Klapp, M.A., Erera, A.L., Toriello, A.: The dynamic dispatch waves problem for
same-day delivery. Eur. J. Oper. Res. 271(2), 519–534 (2018). https://doi.org/10.
1016/j.ejor.2018.05.032
9. Kool, W., et al.: EURO Meets NeurIPS 2022 Vehicle Routing Competi-
tion (2022). https://euro-neurips-vrp-2022.challenges.ortec.com/assets/pdf/euro-
neurips-vrp-2022-rules-2-aug.pdf. Accessed 15 Nov 2022
10. Kool, W., et al.: Hybrid Genetic Search for the Vehicle Routing Problem with Time
Windows: a High-Performance Implementation. Technical report, 12th DIMACS
Challenge, April 2022
11. Kwon, Y.D., Choo, J., Bae, D., Kim, J., Hottung, A.: Cost Shaping via Reinforce-
ment Learning for Vehicle Routing Problems. Technical report (2022)
12. Ulmer, M.W., Mattfeld, D.C., Köster, F.: Budgeting time for dynamic vehicle rout-
ing with stochastic customer requests. Transp. Sci. 52(1), 20–37 (2018). https://
doi.org/10.1287/trsc.2016.0719
13. Approximate Dynamic Programming for Dynamic Vehicle Routing. ORSIS, vol.
61. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-55511-9
14. Van Hentenryck, P., Bent, R., Upfal, E.: Online stochastic optimization under
time constraints. Ann. Oper. Res. 177, 151–183 (2010). https://doi.org/10.1007/
s10479-009-0605-5
15. Vidal, T.: Hybrid genetic search for the CVRP: open-source implementation and
swap* neighborhood. Comput. Oper. Res. 140, 105643 (2022). https://doi.org/10.
1016/j.cor.2021.105643
16. Vidal, T., Crainic, T.G., Gendreau, M., Lahrichi, N., Rei, W.: A hybrid genetic
algorithm for multidepot and periodic vehicle routing problems. Oper. Res. 60(3),
611–624 (2012). https://doi.org/10.1287/opre.1120.1048
17. Vidal, T., Crainic, T.G., Gendreau, M., Prins, C.: A unified solution framework
for multi-attribute vehicle routing problems. Eur. J. Oper. Res. 234(3), 658–673
(2014). https://doi.org/10.1016/j.ejor.2013.09.045
 A Tabu Search Algorithm for the Traveling
Purchaser Problem with Transportation Time
Limit

Ilker Kucukoglu1(B) , Pieter Vansteenwegen2 , and Dirk Cattrysse2

1 Industrial Engineering Department, Bursa Uludag University, Bursa, Turkey
[email protected]
2 KU Leuven, KU Leuven Institute for Mobility – CIB, Leuven, Belgium

{pieter.vansteenwegen,dirk.cattrysse}@kuleuven.be

Abstract. This study extends the well-known traveling purchaser problem (TPP)
by considering a transportation time limit of perishable food in cold-chain logistics.
The problem is called the traveling purchaser problem with transportation time
limit (TPP-TTL). The objective of the TPP-TTL is to find a route and procurement
plan for the purchaser to satisfy the demand of a number of product types with
minimum cost. To satisfy the product demand, the purchaser visits a number
of capacitated markets, in which the available amount of products is limited.
Furthermore, since the travel times cause deterioration on the perishable products,
a transportation time limit is taken into account in the TPP-TTL for each product
type. The problem is formulated as a mixed-integer programming model and
solved by using a tabu search (TS) algorithm. In the computational experiments, TS
is carried out for a number of different-sized instances and the results are compared
to the results obtained by GUROBI solver to determine the performance of the
proposed algorithm. The results of the experiments show that the TS is capable to
find many optimal results with less computational time than the GUROBI solver.

Keywords: Traveling Purchaser Problem · Combinatorial Optimization · Tabu

Search Algorithm

1 Introduction
The traveling purchaser problem (TPP) is a well-known combinatorial optimization
problem that was first introduced by Ramesh [17] in 1981. The TPP is a variant of the
traveling salesperson problem (TSP), in which a set of product demands is satisfied from
a number of markets. Distinct from the TSP, the TPP takes into account the decisions
about the selection of the markets to be visited, the number of products purchased from
each market, and the route of the purchaser. The aim of the problem is to satisfy product
demand with minimum total traveling and procurement costs [13].
The original version of the TPP has been extended in many different ways in the
literature. Boctor et al. [5] introduced the uncapacitated TPP, in which the amount of a
product is sufficient to satisfy demand if it is available in a market. Almeida et al. [1]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 248–259, 2023.
https://doi.org/10.1007/978-3-031-43612-3_15
 A Tabu Search Algorithm for the Traveling Purchaser Problem 249

proposed a bi-objective TPP by considering different goals in the objective function. A

budget limit for the purchaser is taken into account by Choi and Lee [7], where the total
cost of the procurement cannot exceed the specified budget limit. Batista-Galvan et al.
[2] extended the TPP by considering the pick-up and delivery operations for multiple
stacks. Bianchessi et al. [4] introduced a multi-vehicle TPP, in which a distance limit is
defined for each vehicle. Gouveia et al. [11] limited the number of markets to be visited
in the route for the purchaser. Cheaitou et al. [6] included the environmental effect of
transportation in the TPP with variable speeds. Kucukoglu [12] extended the TPP with
a duration time limit for the purchaser and a release time for each product at markets.
Bianchessi et al. [3] introduced a capacitated multiple vehicles TPP for incompatible
products, in which two incompatible products cannot be loaded into the same vehicle.
In almost all studies, the TPP has been extended by considering the route-based
(i.e., pick-up and delivery, multiple-vehicle, vehicle speed), purchaser-based (i.e., budget
limit, duration time limit, maximum market visit) or market-based (i.e., market capacity,
the release time of products) limitations. However, the products to be purchased from
the markets is another factor that may affect the route and procurement plan of the
purchaser. In this context, perishable food procurement activity in cold chain logistics
is a specific application of the TPP since maintaining the quality of perishable foods is
critical to reduce the cost of quality loss due to the deterioration of products while they
are in transportation. Therefore, in most real-life applications, the transportation time
of perishable products is limited by product-specific service-level agreements. Hence,
the time the products spend in the vehicle is an important factor in terms of product
quality. Particularly, when traditional vehicles are used in logistics operations (instead
of temperature-controlled vehicles), transportation of different types of foods having
their own different storage requirements has to be well planned to ensure acceptable
product quality [19].
Based on the aforementioned motivation, this paper introduces a new variant of the
TPP based on the procurement of perishable food in cold chain logistics. Since the
transportation time of the perishable products is limited to maintain product quality, the
TPP is extended with a transportation time limit for each product type. The problem is
called the traveling purchaser problem with transportation time limit (TPP-TTL). The
aim of the problem is to minimize the total transportation and procurement cost of the
purchaser while satisfying the market capacity and transportation time limit restrictions.
In accordance with this purpose, a mixed-integer mathematical model formulation is
introduced for the TPP-TTL. To solve the given model, a tabu search (TS) algorithm is
proposed. Computational experiments are used to evaluate the performance of the TS
on a number of different-sized TPP-TTL instances.
This study contributes to the literature in two ways: First, a new variant of the TPP
is introduced regarding the procurement of perishable food in cold chain operations.
To the best of our knowledge, the TPP has not been addressed in the literature with
a transportation time limit for each product type. Distinct from the existing studies
related to TPP, the mathematical model introduced for the TPP-TTL takes into account
the real-life restrictions of cold chain activities by including the transportation time
limits of perishable foods. The second contribution of this study is the proposed solution
250 I. Kucukoglu et al.

methodology, which employs an advanced local search procedure to explore the solution
space.
The remainder of the paper is organized as follows. Section 2 introduces the details
of the TPP-TTL and its mathematical formulation. The proposed TS is presented in
Sect. 3. The computational results and performance analysis of the TS are given in
Sect. 4. Finally, the conclusions are given in Sect. 5.

2 Problem Definition and Mathematical Model

The TPP-TTL is an extension of the well-known capacitated version of the TPP, in

which the demand of a set of products is provided from a set of capacitated markets.
In each market, the available quantity of a product can be less than the demand for that
product. Therefore, the demand of a product type can be purchased from more than one
market. In addition to the capacity restriction of the markets, the time each product can
be transported is limited. The purchaser starts its route, visits a number of markets, and
returns to the depot after all required products are purchased. The aim of the TPP-TTL
is to minimize the total transportation and procurement cost of the purchaser. Based on
the assumptions given above, the mathematical model of the TPP-TTL is formulated as
follows.
Parameters
K Set of products
M Set of markets
Mk Set of markets in which the product k is available (Mk ⊆ M ), k ∈ K.
{0} Depot node
V Set of markets and depot (M ∪ {0})
dk Demand amount of product k, k ∈ K.
qik Available amount of product k at market i, k ∈ K, i ∈ Mk .
pik Price of product k at market i, k ∈ K, i ∈ Mk .
cij Traveling cost from node i to node j, i, j ∈ V .
tij Traveling time from node i to node j, i, j ∈ V .
hi Unit service time to purchase any product at market i; i ∈ M .
tlk Maximum allowed transportation time for product k, k ∈ K.
γ Large number.
Decision Variables
xij Binary variable: 1 if the purchaser travels from node i to node j, otherwise 0; i, j ∈ V ,
i = j.
yi Binary variable: 1 if market i is visited by the purchaser, otherwise 0; i ∈ M .
zik Amount of product k purchased from market i; k ∈ K, i ∈ Mk .
oik Binary variable: 1 if product k is purchased at market i, otherwise 0; k ∈ K, i ∈ Mk .
ri Arrival time of purchaser to node i; i ∈ V .
si Time spent at market i; i ∈ M .
 A Tabu Search Algorithm for the Traveling Purchaser Problem 251

Model
 
Min cij xij + pik zik (1)
i∈V j∈V k∈K i∈Mk

S.t.

zik = dk k∈K (2)
i∈Mk

zik ≤ qik yi k ∈ K, i ∈ Mk (3)

oik ≤ zik ≤ qik oik k ∈ K, i ∈ Mk (4)


xij = yi i∈M (5)
j∈V
i = j

xij = yj j∈M (6)
i∈V
i = j

xi0 = 1 (7)
i∈M

x0j = 1 (8)
j∈M

t0i ≤ ri + γ (1 − x0i ) i∈M (9)

t0i ≥ ri − γ (1 − x0i ) i∈M (10)

 
ri + si + tij ≤ rj + γ 1 − xij i ∈ M , j ∈ V, i = j (11)

 
ri + si + tij ≥ rj − γ 1 − xij i ∈ M , j ∈ V, i = j (12)

si = hi zik i∈M (13)
k∈K

r0 − ri − si ≤ tlk + γ (1 − oik ) k ∈ K, i ∈ Mk (14)

xij ∈ {0, 1} i, j ∈ V , i = j (15)

yi ∈ {0, 1} i∈M (16)

252 I. Kucukoglu et al.

oik ∈ {0, 1} k ∈ K, i ∈ Mk (17)

zik ≥ 0 k ∈ K, i ∈ Mk (18)

ri ≥ 0 i∈V (19)

si ≥ 0 i∈M (20)

Equation (1) in the proposed model describes the objective function to minimize the
total transportation and procurement cost of the purchaser. Constraints (2) provide that
each product demand is satisfied from the markets. Constraints (3) and (4) ensure that
the purchased amount of a product type at a specific market cannot exceed the available
stock. Constraints (5) and (6) guarantee that the purchaser enters and exits from a market
if a procurement is made from the related market. Constraints (7) and (8) set the initial
and last node of the route as the depot node. Constraints (9)–(12) determine the arrival
time of the purchaser to the visited nodes considering the traveling time between the
locations and service time at markets. These constraints also eliminate the sub-tours by
tracking the arrival time between the visited nodes. Constraints (13) compute the service
time at the visited markets. Constraints (14) guarantee that the travel time and service
time of each product in the vehicle cannot exceed the corresponding time limit. Finally,
constraints (15)–(20) identify the domain of the decision variables.

3 Solution Methodology
Tabu search, which was first introduced by Glover in 1986, is a well-known meta-
heuristic to solve combinatorial optimization problems [9, 10]. The main feature of TS
is the memory mechanism that prevents the search from cycling back to previously
found solutions. The memory structure of TS, called tabu list, guides the search of the
algorithm. The tabu list allows to find unexplored solutions by prohibiting the usage of
recently utilized moves in the tabu list. With the help of this memory structure, TS has
proved to be effective to solve various routing problems [8, 16, 18]. Moreover, TS is
used in [20] to solve TPP, in which two dynamic tabu search strategies are operated.
In this context, TS is employed in this study to find effective results for the TPP-TTL.
Figure 1 presents the pseudo code of the proposed algorithm. The details of the main TS
procedures are described in the following sub-sections.

3.1 Initial Solution

The proposed algorithm searches the solution space for a feasible solution considering
the capacity of the markets and the transportation time limit of each product. Therefore,
the initial solution generation procedure (Line 2 in Fig. 1) is designed to produce a
feasible solution for the TPP-TTL. In this context, the procedure iteratively selects an
unmet product that has the smallest transportation time limit and fulfills the demand of
the product. First, the visited markets are taken into account to satisfy the demand. In
 A Tabu Search Algorithm for the Traveling Purchaser Problem 253

Fig. 1. Pseudo code of the proposed TS

case the demand is not completely met from the visited markets, an unvisited market with
minimum cost for the selected product is inserted into the first position in the route. This
product selection and adding procedure repeats until all demand is satisfied. Although
the capacity restriction of the markets can be completely satisfied by this procedure, the
transportation time limit of any product may be violated while inserting another product
into the solution. In such a case, a feasible solution is provided by randomly selecting a
product that violates the time restriction and purchasing it from a proper market. Here,
in addition to the visited markets, unvisited markets are also taken into account. Then,
a market is randomly selected from the considered list of markets. If a proper market is
not found, the initial solution generation procedure is restarted in the algorithm. Since
the markets are randomly selected in this procedure, different results can be obtained at
each restart. In this way, at the end of the procedure, a feasible initial solution (X 0 ) is
generated to be improved in the main loop of the TS.

3.2 Neighborhood Generation

In each iteration of the TS, the algorithm employs two independent local search proce-
dures to improve the existing solutionX. In the first local search procedure, a number
of neighborhoods are generated by applying a procurement-based move operator (Lines
5–7 in Fig. 1). The procurement-based move operators are used to improve the existing
solution by making changes in the procurement plan. To change the procurement plan,
three moves are considered: unvisited market insertion, visited market removal, and
product exchange between the visited markets. For each application of the procurement-
based search, one of the move operators is randomly selected. Based on the selected
operator, all feasible neighbor solutions are generated. Regarding the tabu list criteria of
the
 TS, the best alternative
 solution X ∗ is selected from a set of candidate neighborhoods
  
( X 1 , X 2 , X 3 , . . . ).
254 I. Kucukoglu et al.
 
In the second local search procedure, a set of candidate solutions ( X 1 , X 2 , X 3 , . . . )
are generated based on X ∗  by using a route-change-based move operator (Lines 8–
10 in Fig. 1). Similar to the procurement-based search approach, four different move
operators are used to generate neighborhoods. Four move operators are considered:
market exchange, a market movement to a further position, a market movement to a
backward position, and a sub-path reverse in the route. After all feasible candidate
solutions are generated, the best solution X ∗ , which is not tabu or leads to a new best
solution, is selected to be assigned as the current solution X.

3.3 Tabu List Structure

The tabu list of the proposed algorithm includes the move information for both the
procurement-based and route-change-based search procedures. In this context, two inde-
pendent tabu lists (TL and TL ) are employed in the algorithm, where TL is used for
the procurement-based search procedure while TL is used for the route-change-based
search procedure. In each iteration of the TS, the information corresponding to the
selected neighborhood in the applied move operator is added to the related tabu list. A
capacitated tabu list strategy is taken into account with first-in-first-out logic, in which
move information is kept in the tabu list for a specified number of iterations (L). A move,
which is in the tabu list, is prohibited for L iterations. However, regarding the aspiration
criterion, a move is accepted if it leads to a new best solution for the algorithm. At the
end of each iteration of the TS, the oldest move information in the list is removed.

3.4 Termination Criterion

The main loop of the proposed algorithm consists of neighborhood generation, fitness
evaluation, tabu list update, and best solution update procedures. The main loop of the
TS is repeated until the termination criterion is met. For the termination criterion, a
maximum iteration number (Iter Max ) is used in the algorithm.

4 Computational Results
In the computational studies, first, the difference between the TPP and TPP-TTL based
on the total cost is analyzed by using the GUROBI solver. Following this analysis, the
performance of the proposed TS is demonstrated by comparing the proposed algorithm
with the GUROBI solver. For these studies, a well-known capacitated TPP benchmark
problem set introduced by Laporte et al. [14] is used. The TPP instances are adapted to
the TPP-TTL by randomly assigning a transportation time limit for each product type.
To generate a random value for the transportation time limits, first, a reference duration
time (RDT ) for the problem is calculated regarding the transportation and service time
of the purchaser. The transportation time of the purchaser is calculated by using the
nearest neighborhood algorithm ignoring the transportation time limit of the products.
On the other hand, the service time is determined regarding the lowest service time for
each product type. After the RDT is obtained for the problem, a transportation time limit
is randomly specified as tlk = 0.50 × RDT + rnd () × RDT where rnd () is a uniformly
 A Tabu Search Algorithm for the Traveling Purchaser Problem 255

distributed random number in the interval [0, 1]. In case the generated time limits leads
an infeasible route for the purchaser, the procedure is restarted. Based on this procedure,
27 different-sized TPP-TTL instances are generated where the problem size is specified
by the pair of two problem data: {|V |, |K|} where the value of each parameter is specified
as 10, 20, or 30 for TPP-TTL instance generation.
The proposed TS is implemented for each TPP-TTL instance by setting the Iter Max =
1000 and L = 50. With these parameter settings, ten independent runs are carried out
for each problem. The results of the TS are investigated through the total cost (f ) and
solution time (t) for both the best solution obtained in the runs and the average of the
runs. Here, it should be noted that the f and t are used to represent the GUROBI results
to indicate optimum solution and solution time, respectively. Each TS and GUROBI
computation is carried out on a workstation with Intel® Xeon® CPU E5-2643 v3 CPU
and 64 GB of memory.
To indicate the difference between the TPP and TPP-TTL results based on the total
cost, the model formulation proposed for the TPP-TLL is simply adapted to the TPP by
ignoring the constraints (9)–(14). However, the Miller-Tucker-Zemlin sub-tour elimina-
tion constraints [15] are included in the model to eliminate sub-tours. To find the optimal
solution for TPP and TPP-TTL each instance, the GUROBI is executed without setting
any time limit. Table 1 shows the results of the GUROBI for both the TPP and TPP-TTL
instances, where the percentage gap between the two solutions is computed by using
Eq. (21), where fTPP and fTPP−TTL are the optimum solutions obtained for the TPP and
TPP-TTL instances, respectively.

fTPP−TTL − fTPP
Gap1 % = × 100% (21)
fTPP

The Gap1 % values presented in Table 1 show that up to 15.02% increase on total
cost is observed due to the transportation time limits of the products. Particularly, the gap
between the solution is high when the number of product types is high (i.e., {|V |, |K|} =
{(10, 30), (20, 30), (30, 30)}). On the other hand, the same results are obtained for four
instances. In addition to the comparison based on the total cost, it should be seen from
Table 1 that the solution time of GUROBI for the TPP-TLL is quite high according to the
TPP model computation. Therefore, it should be stated that finding an optimal solution
for the TPP-TTL requires more computational effort with respect to the TPP.
256 I. Kucukoglu et al.

Table 1. GUROBI results for the TPP and TPP-TLL instances

Problem # |V | |K| TPP TPP-TTL

f t f Gap1 % t
1 10 10 2652 0.42 2700 1.81 6.41
2 10 10 2608 0.62 2710 3.91 4.57
3 10 10 3292 0.39 3393 3.07 5.60
4 10 20 3432 0.48 3486 1.57 20.79
5 10 20 3499 0.42 3638 3.97 11.43
6 10 20 3818 0.44 3882 1.68 14.74
7 10 30 3895 0.42 4181 7.34 95.95
8 10 30 3885 0.39 4303 10.76 81.55
9 10 30 3521 0.44 4050 15.02 97.71
10 20 10 3048 2.03 3048 0.00 138.24
11 20 10 3530 3.34 3530 0.00 180.33
12 20 10 3983 2.21 3991 0.20 157.71
13 20 20 4074 0.97 4344 6.63 470.27
14 20 20 4109 1.36 4245 3.31 444.65
15 20 20 3688 4.65 3760 1.95 350.35
16 20 30 3975 1.05 4274 7.52 678.94
17 20 30 6483 1.06 6855 5.74 340.95
18 20 30 5120 2.20 5597 9.32 906.16
19 30 10 3467 8.52 3467 0.00 451.66
20 30 10 3684 8.02 3684 0.00 440.06
21 30 10 4865 5.82 4871 0.12 545.06
22 30 20 4748 8.70 4788 0.84 881.62
23 30 20 6938 10.11 7063 1.80 956.97
24 30 20 5836 7.74 5867 0.53 1118.01
25 30 30 5078 3.51 5250 3.39 955.05
26 30 30 6342 2.40 6572 3.63 884.27
27 30 30 7369 11.50 7507 1.87 858.87
Average 4331.1 3.30 4484 3.56 411.03

In addition to the comparison of TPP and TPP-TTL, the second part of the compu-
tational experiments is carried out to demonstrate the efficiency of the proposed TS. In
this context, the TS results are compared to the optimal TPP-TTL solutions obtained by
GUROBI. The results of GUROBI and TS, and the comparison are given in Table 2, in
which Gap2 % indicates the percentage gap between the best/average result of the TS
 A Tabu Search Algorithm for the Traveling Purchaser Problem 257

and GUROBI result. Gap2 % is calculated by using Eq. (22) where fTS is the best/average
result of the TS and fGUROBI is the GUROBI result.

Table 2. TS and GUROBI comparison for the TPP-TTL instances

Problem # GUROBI TS
Best Solution Average Solution
f t f Gap2 % t f Gap2 % t
1 2700 6.41 2700 0.00 0.45 2700.0 0.00 0.53
2 2710 4.57 2710 0.00 0.94 2710.0 0.00 1.12
3 3393 5.60 3393 0.00 1.04 3393.0 0.00 1.32
4 3486 20.79 3486 0.00 0.79 3486.0 0.00 0.72
5 3638 11.43 3638 0.00 0.99 3638.0 0.00 0.92
6 3882 14.74 3882 0.00 0.73 3887.8 0.15 0.86
7 4181 95.95 4181 0.00 0.85 4181.0 0.00 0.93
8 4303 81.55 4303 0.00 1.06 4303.0 0.00 1.15
9 4050 97.71 4050 0.00 1.11 4052.6 0.06 1.09
10 3048 138.24 3048 0.00 1.15 3048.0 0.00 1.12
11 3530 180.33 3530 0.00 1.23 3530.0 0.00 1.29
12 3991 157.71 3991 0.00 1.32 3991.0 0.00 1.31
13 4344 470.27 4344 0.00 1.22 4348.9 0.11 1.19
14 4245 444.65 4245 0.00 1.41 4245.0 0.00 1.38
15 3760 350.35 3760 0.00 1.38 3771.3 0.30 1.41
16 4274 678.94 4274 0.00 1.44 4276.5 0.06 1.35
17 6855 340.95 6855 0.00 1.51 6855.0 0.00 1.49
18 5597 906.16 5597 0.00 1.47 5601.2 0.08 1.53
19 3467 451.66 3467 0.00 1.62 3467.0 0.00 1.55
20 3684 440.06 3684 0.00 1.58 3684.0 0.00 1.49
21 4871 545.06 4871 0.00 1.69 4872.1 0.02 1.71
22 4788 881.62 4788 0.00 1.55 4795.6 0.16 1.43
23 7063 956.97 7063 0.00 1.78 7077.9 0.21 1.74
24 5867 1118.01 5867 0.00 1.71 5869.7 0.05 1.73
25 5250 955.05 5250 0.00 1.83 5275.3 0.48 1.79
26 6572 884.27 6572 0.00 1.81 6593.3 0.32 1.83
27 7507 858.87 7507 0.00 1.93 7542.1 0.47 1.88
Average 4484 411.03 4484 0.00 1.32 4488.7 0.09 1.33
258 I. Kucukoglu et al.

fTS − fGUROBI
Gap2 % = × 100% (22)
fGUROBI
Based on the best-found result of the TS given in Table 2, the optimal solution is
provided for each instance by the proposed algorithm. Furthermore, the average result
of the TS is the same with the optimal solution for 14 of 27 instances. For the remaining
instances, the highest percentage gap is 0.48%. For the overall performance of the TS, the
average percentage gap of the proposed algorithm is 0.09% with respect to the optimal
solution. Furthermore, the average computational time of the TS is 1.33 s, while the
average computational time of the GUROBI is 411.03 s. Regarding the solution quality
and computational effort of the proposed algorithm, it can be stated that the TS is capable
to find the optimal or near-optimal solution for the TPP-TTL in much less computational
time.

5 Conclusions
In this study, the traveling purchaser problem with transportation time limit, which is an
extension of the well-known traveling purchaser problem, is proposed by considering
the transportation restrictions of perishable food in cold-chain logistics. The TPP-TTL
extends the capacitated version of the TPP by including a transportation time limit for
each product type in the vehicle. In this context, any product cannot be kept in the vehicle
for more than a specified time limit. Regarding both the market capacity and product
transportation restrictions, the TPP-TTL is formulated as a mixed-integer programming
model, in which the objective is to minimize the total transportation and procurement
cost of the purchaser. To solve the problem, a tabu search algorithm is introduced. The
proposed algorithm employs a number of procurement-based and route-change-based
local search moves to generate neighborhoods in each iteration. The performance of the
proposed algorithm is tested on different-sized TPP-TTL problems generated by using
a well-known TPP benchmark problem set. The TS is compared to the GUROBI solver
to identify the solution quality of the proposed algorithm. Results of the computations
show that the proposed TS outperforms the GUROBI solver by finding optimal results
in much less computational time.
For future research, this study can be extended with considering different real-life
restrictions of cold chain logistics. First, the deterioration cost of perishable products due
to the traveling in a vehicle can be integrated. On the other hand, the procurement plan can
consider multiple trips by the purchaser, where each trip can be done with a different type
of vehicle (i.e., temperature controlled vehicle, traditional vehicle). With this assumption,
the quality level of some products that are more sensitive to temperature change may be
transported with a temperature-controlled vehicles with additional cooling cost. Another
practical assumption for the TPP-TTL may be the proportional transportation time limit
for the product at different markets. This assumption may be valid when a product
is produced in different time periods in different markets. Considering the quality of
a product at each market depends on the production date, the purchaser may need to
include this assumption for its procurement and route plan.
 A Tabu Search Algorithm for the Traveling Purchaser Problem 259

Acknowledgement. This work is supported by the Commission of Scientific Research Projects

of Bursa Uludag University, Project Number FUİ-2022-1042.

References
1. Almeida, C.P., Gonçalves, R.A., Goldbarg, E.F., Goldbarg, M.C., Delgado, M.R.: An exper-
imental analysis of evolutionary heuristics for the biobjective traveling purchaser problem.
Ann. Oper. Res. 199(1), 305–341 (2012)
2. Batista-Galvan, M., Riera-Ledesma, J., Salazar-Gonzalez, J.J.: The traveling purchaser prob-
lem, with multiple stacks and deliveries: a branch-and-cut approach. Comput. Oper. Res. 40,
2103–2115 (2013)
3. Bianchessi, N., Irnich, S., Tilk, C.: A branch-price-and-cut algorithm for the capacitated
multiple vehicle traveling purchaser problem with unitary demand. Discret. Appl. Math. 288,
152–170 (2021)
4. Bianchessi, N., Mansini, R., Speranza, M.G.: The distance constrained multiple vehicle
traveling purchaser problem. Eur. J. Oper. Res. 235, 73–87 (2014)
5. Boctor, F.F., Laporte, G., Renaud, J.: Heuristics for the traveling purchaser problem. Comput.
Oper. Res. 30(4), 491–504 (2003)
6. Cheaitou, A., Hamdan, S., Larbi, R., Alsyouf, I.: Sustainable traveling purchaser problem
with speed optimization. Int. J. Sustain. Transp. 15(8), 621–640 (2021)
7. Choi, M.J., Lee, S.H.: The multiple traveling purchaser problem for maximizing system’s
reliability with budget constraints. Expert Syst. Appl. 38(8), 9848–9853 (2011)
8. Gendreau, M., Hertz, A., Laporte, G.: A Tabu search heuristic for the vehicle routing problem.
Manag. Sci. 40(10), 1276–1290 (1994)
9. Glover, F.: Tabu search – Part I. ORSA J. Comput. 1(3), 190–206 (1989)
10. Glover, F.: Tabu search – Part II. ORSA J. Comput. 2(1), 4–32 (1990)
11. Gouveia, L., Paias, A., Voß, S.: Models for a traveling purchaser problem with additional
side-constraints. Comput. Oper. Res. 38, 550–558 (2011)
12. Kucukoglu, I.: The traveling purchaser problem with fast service option. Comput. Oper. Res.
141, 105700 (2022)
13. Manerba, M., Mansini, R., Riera-Ledesma, J.: The traveling purchaser problem and its
variants. Eur. J. Oper. Res. 259(1), 1–18 (2017)
14. Laporte, G., Riera-Ledesma, J., Salazar-González, J.-J.: A branch-and-cut algorithm for the
undirected traveling purchaser problem. Oper. Res. 51(6), 940–951 (2003)
15. Miller, C.E., Tucker, A.W., Zemlin, R.A.: Integer programming formulation of traveling
salesman problems. J. ACM 7(4), 326–329 (1960)
16. Qiu, M., Zhuo, F., Eglese, R., Qiong, T.: A Tabu search algorithm for the vehicle routing
problem with discrete split deliveries and pickups. Comput. Oper. Res. 100, 102–116 (2018)
17. Ramesh, T.: Traveling purchaser problem. Opsearch 18(1–3), 78–91 (1981)
18. Taillard, E., Badeau, P., Gendreau, M., Guertin, F., Potvin, J.-Y.: A Tabu search heuristic for
the vehicle routing problem with soft time windows. Transp. Sci. 31(2), 170–186 (1997)
19. Tsang, Y.P., Choy, K.L., Wu, C.H., Ho, G.T.S., Lam, H.Y., Tang, V.: An intelligent model for
assuring food quality in managing a multi-temperature food distribution centre. Food Control
90, 81–97 (2018)
20. Voß, S.: Dynamic tabu search strategies for the traveling purchaser problem. Ann. Oper. Res.
63, 253–275 (1996)
 GRASP Solution Approach
for the E-Waste Collection Problem

Aldy Gunawan1(B) , Dang Viet Anh Nguyen1 ,

Pham Kien Minh Nguyen2 , and Pieter Vansteenwegen3
1 School of Computing and Information Systems, Singapore Management University,
80 Stamford Road, Singapore 178902, Singapore
{aldygunawan,dvanguyen}@smu.edu.sg
2 Department of Industrial Management, National Taiwan University of Science

and Technology, No. 43, Keelung Rd., Taipei 106335, Taiwan

[email protected]
3 KU Leuven Institute for Mobility - CIB, University of Leuven, Celestijnenlaan 300,
Leuven 3001, Belgium
[email protected]

Abstract. The digital economy has brought significant advancements

in electronic devices, increasing convenience and comfort in people’s
lives. However, this progress has also led to a shorter life cycle for these
devices due to rapid advancements in hardware and software technol-
ogy. As a result, e-waste collection and recycling have become vital for
protecting the environment and people’s health. From the operations
research perspective, the e-waste collection problem can be modeled as
the Heterogeneous Vehicle Routing Problem with Multiple Time Win-
dows (HVRP-MTW). This study proposes a metaheuristic based on the
Greedy Randomized Adaptive Search Procedure complemented by Path
Relinking (GRASP-PR) to solve the HVRP-MTW problem. The exper-
iment demonstrates that the proposed algorithm can efficiently handle
HVRP-MTW instances, even of large-scale. Moreover, the comparison
with CPLEX indicates that our approach can achieve optimal solutions
for small instances and outperform the commercial solver in large-scale
instances.

Keywords: E-waste · Vehicle routing problem · time windows ·

GRASP · Path-relinking

1 Introduction
With the rapid advancement of technology and the continuous development of
production and industrial design, electronic and electrical devices have become
ubiquitous in lives, providing convenience and comfort on a daily basis. However,
this progress has also resulted in a significant drawback-the ever-shortening life
cycle of electrical and electronic devices. Consequently, there has been a marked
increase in the annual amount of e-waste - discarded electrical and electronic
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 260–275, 2023.
https://doi.org/10.1007/978-3-031-43612-3_16
 GRASP Solution Approach for the E-Waste Collection Problem 261

devices and their components which are not reused. According to UNITAR,
in 2019, the world generated a staggering 53.6 Megatons of e-waste, of which
only 17.35% was properly managed in an environmentally sound manner, while
the fate of the remaining e-waste was unknown [1]. The collection of e-waste
plays a pivotal role in its overall management and treatment cycle, typically
involving the use of heavy vehicles and differing from conventional waste man-
agement practices. If not adequately planned, the e-waste collection process can
create numerous challenges for urban areas, such as straining the urban trans-
port system, contributing to increased air pollution, and burdening the e-waste
management companies. Unfortunately, the literature on optimizing the e-waste
collection process remains relatively limited. In the field of operations research,
the e-waste collection problem can be effectively modeled as a variant of the Vehi-
cle Routing Problem (VRP), which is extensively studied and widely applied in
the realm of logistics and supply chain management. To gain a comprehensive
understanding of the development of VRP models, methods, and applications,
interested readers are encouraged to refer to the work by Braekers et al. [2].
We provide a summary and highlight relevant literature on various algorithms
proposed to address different variants of the VRP in the context of e-waste man-
agement. Yao et al. [13] focused on two subproblems: optimizing the location of
Waste of Electric and Electronic Equipment (WEEE) transit sites and the VRP.
They proposed an exact method using the quadratic optimization and an ant
colony algorithm to solve instances in the context of Shanghai (China). Mar-
Ortiz et al. [11] formulated the WEEE collection problem as the VRP with
Split Loads and Time Windows and proposed a Greedy Randomized Adaptive
Searching Procedure (GRASP) metaheuristic to solve a large dataset in the con-
text of Galicia (Spain). Król et al. [10] presented a model for e-waste collection
at both doorsteps and collection points within a single day. The problem is
formulated as the VRP with intermediate facilities, considering heterogeneous
vehicles, breaks between shifts, and regional cost differences. A multiple neigh-
borhood search algorithm was proposed to tackle this problem. Pourhejazy et
al. [12] formulated the actual waste-collecting scheme as a Capacitated General
Routing Problem with Time Windows and adopted Tabu Search to handle it.
They considered population density to separate requests from on-call customers
and also accounted for a scheme of heterogeneous vehicles.
To the best of our knowledge, none addresses the waste collecting with dis-
tinct periods within distinct days in the schedule. In the most recent research
[7], a Heterogeneous VRP with Multiple Time Windows (HVRP-MTW) model
is introduced to optimize e-waste collection vehicle routing. CPLEX is employed
to solve the Mixed Integer Linear Programming formulation of the problem.
However, CPLEX is inefficient in solving practical-scale instances due to a sig-
nificant amount of time to handle the problem. In this work, we propose an
efficient approach based on the GRASP metaheuristic to obtain solutions for
larger instances within reasonable computational times. The remainder of this
article is organized as follows. Section 2 provides details of the problem descrip-
tion and the proposed algorithm. In Sect. 3, we discuss our experimental setup,
262 A. Gunawan et al.

benchmark instances, and the computational results. Finally, we conclude this

article and provide potential research directions in Sect. 4.

2 Methodology
2.1 Problem Description
The problem addressed arises from the e-waste collection process in the context
of Singapore, which was modeled as a Mixed Integer Linear Program (MILP)
in [7]. E-waste public drop-off bins (e-bins) have been set up across Singapore.
Unlike conventional waste, which is collected daily, the e-bins only require col-
lection when the amount of e-waste reaches a specified level. Every day, the
system updates information about the e-bins that need to be collected, allow-
ing staff to carry out the collection task. Additionally, companies also cater to
on-demand e-waste collection requests from households and businesses. For each
on-demand request, customers are charged a fee. Moreover, the on-demand col-
lection requests need to be completed within a specific time frame specified by
the customer when making the request through a mobile application, creating
time windows.
Each customer’s request includes the preferred collection days and time win-
dows, resulting in a multiple time window problem. If the vehicle arrives before
the lower bound of the time window, an idle cost will be incurred. Conversely,
a penalty cost will be imposed if the vehicle arrives after the upper bound of
the time window, with the penalty cost typically being greater than the idle cost
due to its impact on the customer satisfaction. The company typically employs a
heterogeneous fleet of vehicles with two different types. These vehicles have vary-
ing capacities and operating costs per unit of time. Different collection points
have different amounts of e-waste that need to be collected and require a certain
amount of time for processing. Figure 1 illustrates a solution of the described
HVRP-MTW. The problem involves planning the routes for the existing vehi-
cles within a specific collection period, with the objective of maximizing profits
from the e-waste collection process.

2.2 Proposed Algorithm

GRASP is a multi-start and memoryless metaheuristic that has been successfully
applied to solve various combinatorial optimization problems [4]. The original
GRASP metaheuristic consists of two main phases: construction and improve-
ment. In each iteration, an initial solution is constructed by a greedy randomized
heuristic, and local search procedures are applied to improve the initial solution.
The pseudocode for the basic GRASP algorithm for maximization problems is
presented in Algorithm 1 with 𝑆 and 𝑆 ∗ representing the initial solution and
the best-found solution so far, respectively. A two-phase construction heuristic
is proposed to generate an initial solution. The first phase focuses on addressing
the vehicle capacity constraint and heterogeneous fleet characteristics, while the
second phase aims to ensure the feasibility of the collection period constraint.
 GRASP Solution Approach for the E-Waste Collection Problem 263

Fig. 1. A solution of the HVRP-MTW with four collection days

Algorithm 1. GRASP general structure for maximization problem

1: Input: Number of iterations 𝑖𝑚𝑎𝑥
2: 𝑓 (𝑆 ∗ ) ← −∞; //Set the objective function of 𝑆 ∗ to a negative infinite value
3: for 𝑖 ← 1 to 𝑖𝑚𝑎𝑥 do
4: 𝑆 ← ConstructionHeuristic(); //Build the initial solution
5: 𝑆 ← LocalSearch(𝑆 );//Implement the local search
6: if 𝑓 (𝑆) > 𝑓 (𝑆 ∗ ) then
7: 𝑓 (𝑆 ∗ ) ← 𝑓 (𝑆) ;
8: 𝑆∗ ← 𝑆;
9: end if
10: end for
11: Return: 𝑆 ∗

To establish initial routes that satisfy the capacity constraint, the Clarke
and Wright savings heuristic is utilized [3]. However, the traditional Clarke and
Wright procedure is ineffective when dealing with a heterogeneous vehicle fleet.
In the merge step, the heuristic tends to combine all subtours until the total
demand reaches the capacity of the largest vehicle in the fleet. However, relying
solely on the use of large vehicles is not an optimal approach for solving the
heterogeneous fleet problem. To address this issue, we have adopted the Realistic
Opportunity Savings - 𝛾 (𝑅𝑂𝑆 − 𝛾) [6] to calculate the savings of the Clarke
 
and Wright heuristic in our solution construction phase: 𝑠𝑖 𝑗 = 𝑠𝑖 𝑗 + (1 − 𝛾)𝑡 𝑖 𝑗 .
Where 𝛾 is a route-shape parameter with the value varying from 0.0 to 3.0 [6].
By varying the value of 𝛾, the composition of the mixed fleet can be altered.
For instance, when 𝛾 = 0, the solution tends to favor larger vehicles, while
increasing 𝛾 to 2.5 results in the utilization of more small vehicles in the initial
solution. Therefore, 𝛾 serves as a diversification factor in creating a mixed fleet

initial solution. 𝑠𝑖 𝑗 are the savings calculated in Realistic Opportunity Savings
 
(ROS): 𝑠𝑖 𝑗 = 𝑠¯𝑖 𝑗 + 𝛿(𝜃)𝐹 (𝑃(𝑧 𝑖 + 𝑧 𝑗 ) − 𝑧 𝑖 − 𝑧 𝑗 ). The parameter 𝛿(𝜃) is a binary
indicator of whether the vehicle threshold 𝜃 has been exceeded. This threshold
is calculated by the formula 𝜃 = 𝑃(𝑧𝑖 + 𝑧 𝑗 ) − 𝑃(max(𝑧 𝑖 , 𝑧 𝑗 )), where 𝑃(𝑧) denotes
the capacity of the smallest vehicle that can serve a demand of 𝑧. If the vehicle
threshold is not exceeded, then 𝛿(𝜃) is set to 0. Conversely, when 𝜃 > 0, indicating

that the threshold has been exceeded, 𝛿(𝜃) is set to 1. 𝐹 (𝑧) in our study is
the operation cost rate of the largest vehicle that has a capacity less than or
264 A. Gunawan et al.

equal to 𝑧. 𝑠¯𝑖 𝑗 are savings computed in a Combined Savings (CS) approach:

𝑠¯𝑖 𝑗 = 𝑠𝑖 𝑗 + 𝐹 (𝑧 𝑖 ) + 𝐹 (𝑧 𝑗 ) − 𝐹 (𝑧 𝑖 + 𝑧 𝑗 ). Here, 𝐹 (𝑧) stand for the operation cost of the
smallest vehicle that can service a demand of z. The maximum vehicle capacity
in Clarke and Wright with the ROS-𝛾 approach is set as the capacity of the
largest vehicle minus 50 to relax the repair procedure.

Algorithm 2. Clarke and Wright savings heuristic with 𝑅𝑂𝑆 − 𝛾

1: Input: Set of nodes N
2: Initialize individual route for each node in N

3: Calculate the savings based on the 𝑅𝑂𝑆 − 𝛾 approach, 𝑠 for each pair of nodes
4: Sort the savings list in descending order
5: while the savings list is not exhausted do

6: Select pair of nodes (𝑖, 𝑗) |𝑖 ∈ 𝑣𝑖 , 𝑗 ∈ 𝑣 𝑗 with highest savings 𝑠𝑖 𝑗
7: if two distinct routes 𝑣𝑖 and 𝑣 𝑗 are possible to merge then
8: Merge two routes 𝑣𝑖 and 𝑣 𝑗 to a single route 𝑣𝑚
9: Remove pair (𝑖, 𝑗) from the savings list
10: end if
11: end while
12: Return: Merged routes

Algorithm 2 presents the Clarke and Wright savings heuristic with the
𝑅𝑂𝑆 − 𝛾 approach. First, the algorithm creates individual routes for each node
(e-bins and households) (Line 2). Then, it calculates the savings for each pair of
nodes using the 𝑅𝑂𝑆 − 𝛾 approach and sorts the savings list in the descending
order (Lines 3–4). Starting with the top of the sorted savings list, the algorithm

examines the savings 𝑠𝑖 𝑗 for two distinct routes, 𝑣 𝑖 and 𝑣 𝑗 . If it is possible to
merge these two routes (i.e., 𝑖 and 𝑗 are endpoints of their respective subtours,
and the merged subtour containing both 𝑖 and 𝑗 has a total demand not greater
than the capacity of the largest vehicle), the algorithm combines both by remov-
ing the arcs (𝑖, 0) and ( 𝑗, 0) and adding a new arc (𝑖, 𝑗). The merging process
continues until no further pairs of routes is feasible to merge, and the merged
routes are returned at the end of the algorithm (Lines 5–12).

Repair Procedure. The initial solution generated by ROS-𝛾 satisfies the vehi-
cle capacity constraint but it does not take the collection period into account.
To address this issue, Algorithm 3 proposes a repair procedure. It is initiated
with an initial solution, denoted as 𝑆. To ensure that all routes in 𝑆 comply with
the collection period constraint, the algorithm checks the routes violating the
collection day and stores them in the infeasible routes set, denoted as Ω (Line
2). To identify the violated nodes in each infeasible route, we select a collection
day for this route and then remove the nodes which are not available in the
selected collection day. This step is done by the following procedures. The algo-
rithm chooses the day with the highest customer availability as collection day
𝑘 𝜔 of an infeasible route 𝜔. If there are more than one day with same number of
highest customer availability, we choose collection day 𝑘 𝜔 randomly from these
days 𝐾 𝜔 (Lines 3 - 12).
 GRASP Solution Approach for the E-Waste Collection Problem 265

The repair procedure then examines each node 𝑗 in the infeasible route 𝜔
to determine its availability on day 𝑘 𝜔 . Nodes that are unavailable on day 𝑘 𝜔
are recorded in a set, denoted as 𝜋 𝑤 ∈ Π, where Π is the set that stores all
nodes that violate the collection day of all infeasible routes. To repair the infea-
sible routes in 𝑆, the algorithm employs five repair operators (Lines 13–29). The
repair operators have a primary focus on enhancing the solution feasibility while
minimizing modifications to the initial solution’s route structure. Additionally,
these operators are designed to address varying levels of complexity in handling
infeasible solutions. Ideally, the repair procedure aims to rectify an infeasible
solution using a single swap operator, thus reducing the algorithm’s time com-
plexity. Below, we provide a breakdown of the five repair operators and their
respective details:
– R1: This operator is implemented by swapping a violated node 𝑖 from the
set of violated nodes 𝜋 𝜔 in route 𝜔 with another violated node 𝑗 from the
set of violated nodes 𝜋 𝜔 in another infeasible route 𝜔  ∈ Ω. However, for
this operator to be applied, it is necessary for the collection day 𝑘 𝜔 of the
infeasible route 𝜔 to be compatible with the available period 𝐻 𝑗 of node 𝑗. If
the node 𝑗 does not exist, meaning that R1 can only deal with several violated
nodes, the operator will return False, and we need to employ additional repair
operators to resolve the remaining violated nodes in 𝜔 (Line 15).
– R2: The R2 operator is implemented by selecting a violated node 𝑖 from the
set of violated nodes 𝜋 𝜔 in an infeasible route 𝜔 and swapping it with another
node 𝑗 ∉ 𝜋 𝜔 of another infeasible route 𝜔 . Similar to R1, the R2 operator
requires the available period 𝐻 𝑗 of node 𝑗 to be compatible with the collection
day 𝑘 𝜔 of infeasible route 𝜔. If the node 𝑗 does not exist, the operator returns
False, indicating that additional repair operators are required to address the
remaining violated nodes (Line 17).
– R3: This operator operates on the same principles as R1 and R2, but instead
of selecting a swap node 𝑗 from an infeasible route 𝜔 , it selects 𝑗 from a
feasible route 𝑣 ∉ Ω. Once 𝑗 has been selected, R3 swaps the violated node
𝑖 ∈ 𝜋 𝜔 with 𝑗, provided that the available period 𝐻 𝑗 is compatible with the
collection day 𝑘 𝜔 of route 𝜔. If there is no node 𝑗 in any feasible route 𝑣
available on 𝑘 𝜔 , the operator returns False, indicating that additional repair
operators are required to address the remaining violated nodes (Line 19).
– R4: The R4 operator inserts the violated node 𝑖 ∈ 𝜋 𝜔 into a feasible route
𝑣 ∉ Ω, provided that there is at least one day where the available day of
node 𝑖, 𝐻𝑖 , matches the collection period 𝐾 𝑣 of feasible route 𝑣. If there is no
feasible route 𝑣 with a suitable collection period to insert the violated node
𝑖, the operator returns False (Line 21).
– R5: The R5 operator repairs the current infeasible solution by creating a new
single trip with the violated node 𝑖 ∈ 𝜋 𝜔 and then deleting 𝑖 from the current
infeasible route 𝜔 (Line 23).
The initial solution then satisfies both capacity and collection period con-
straints. To diversify the initial solution, we randomly select a collection day in
feasible collection day for each route. Moreover, to determine the departure time
266 A. Gunawan et al.

𝑑 𝑣 of the vehicle at each route 𝑣, we use the equation 𝑑 𝑣 = max(𝐷 1𝑣 , . . . , 𝐷 𝑣|𝐷 | ).

Here, 𝐷 𝑖𝑣 = max(𝑙𝑖 − 𝑎 𝑖 , 0) is the idle time if the vehicle arrival time 𝑎 𝑖 is earlier
than the lower bound 𝑙𝑖 of the collection time window at node 𝑖. To determine the
cost and revenue of a solution, we first assign vehicles in the fleet to their respec-
tive routes. The assignment process follows this principle: if the total demand of
the considered route is less than the capacity of a small vehicle, we assign a small
vehicle to that route. On the other hand, if the demand exceeds the capacity of
a small vehicle or if no small vehicles are available in the fleet, we assign a large
vehicle to the route.

Algorithm 3. Repair procedure

1: Input: An initial solution 𝑆
2: Ω ← infeasible routes in 𝑆 ;
3: for each route 𝜔 in Ω do
4: 𝐾 𝜔 ← period with highest number of available nodes;
5: Select a random day 𝑘 𝜔 ∈ 𝐾 𝜔 as the collection period for infeasible route 𝜔 ;
6: for each node 𝑗 in 𝑤 do
7: 𝐻 𝑗 ← available period of node 𝑗 ;
8: if 𝑘 𝜔 not in 𝐻 𝑗 then
9: set of nodes violating period constraint in route 𝜔 , 𝜋 𝜔 ← 𝑗 ;
10: end if
11: end for
12: end for
13: while Ω ≠ ∅ do
14: while 𝜋𝑤 ≠ ∅ do
15: 𝜔 ← R1( 𝜔 )
16: if 𝑅 1 ( 𝜔) returned False then
17: 𝜔 ← R2( 𝜔 )
18: if 𝑅 2 ( 𝜔) returned False then
19: 𝜔 ← R3( 𝜔 )
20: if 𝑅 3 ( 𝜔) returned False then
21: 𝜔 ← R4( 𝜔 )
22: if 𝑅 4 ( 𝜔) returned False then
23: 𝜔 ← R5( 𝜔 )
24: end if
25: end if
26: end if
27: end if
28: end while
29: end while
30: Return: 𝑆 ;

Local Search. Once the initial solution has been obtained, the following local
search operators are applied to intensify the search for the optimal solution:
– N1 - Intra-relocation: The N1 operator explores the neighborhood solu-
tion by removing customer 𝑖 from route 𝑣 𝑖 and then reinserting 𝑖 at another
position in the same route 𝑣 𝑖 .
– N2 - Inter-relocation: The N2 operator removes customer 𝑖 from its current
route 𝑣 𝑖 and reinserts 𝑖 to a new route 𝑣 𝑗 provided that there is at least one
day in the available period of 𝑖 that coincides with the collection day of 𝑣 𝑗 .
Additionally, the total capacity of 𝑣 𝑗 after the reinsertion must not exceed
the capacity of the largest vehicle used.
 GRASP Solution Approach for the E-Waste Collection Problem 267

– N3 - Intra-swap: The N3 operator selects node 𝑖 in route 𝑣 𝑖 and swaps it

with another node 𝑗 in the same route.
– N4 - Inter-swap: The N4 operator selects node 𝑖 from route 𝑣 𝑖 and another
node 𝑗 from a different route 𝑣 𝑗 . The selected nodes are swapped provided
that their available periods are compatible with their target routes’ 𝑣 𝑖 and
𝑣 𝑗 collection day. Furthermore, the total demand of the two routes 𝑣 𝑖 and 𝑣 𝑗
after swapping must not exceed the maximum capacity of the largest vehicle
in the fleet.
– N5- 2-opt (intra-route): The N5 operator creates a new solution from an
initial solution 𝑆 by removing two consecutive arcs 𝐸 (𝑖, 𝑖 + 1) and 𝐸 ( 𝑗, 𝑗 + 1)
from route 𝑣 𝑖 . Then, two other arcs 𝐸 (𝑖, 𝑗) and 𝐸 (𝑖 + 1, 𝑗 + 1) are added to
reconnect the selected routes. Additionally, the subpath excluding the depot
between nodes 𝑖 + 1 and 𝑗 is inverted.
– N6 - 2-opt* (inter-route): The principle of generating a new solution
using the N6 operator is similar to that of the N5 operator. However, the
N6 operator selects and removes two arcs 𝐸 (𝑖, 𝑖 + 1) and 𝐸 ( 𝑗, 𝑗 + 1) from
two distinct routes 𝑣 𝑖 and 𝑣 𝑗 , respectively. Two new arcs are then added to
reconnect the selected routes, and both the inverted and non-inverted subpath
versions of the new routes are considered. The N6 operator selects the higher-
profit neighborhood between the two versions of new solutions. The condition
of this operator is that two selected routes 𝑣 𝑖 and 𝑣 𝑗 have the same collection
day and the two new routes after reconnecting have to satisfy the capacity
constraint.
– N7 - or-opt (inter-route): The solution in the or-opt neighborhood is
defined by relocating a subpath (𝑖 + 1, . . . , 𝑗) from route 𝑣 𝑖 to another route
𝑣 𝑗 . The conditions for applying the N7 operator are the same as those for
the N6 operator, which require that two selected routes 𝑣 𝑖 and 𝑣 𝑗 have the
same collection day and the total demand of the route 𝑣 𝑗 after inserting the
subpath is less than the capacity of the largest vehicle in the fleet.
– N8 - 3-opt (intra-route): The N8 operator deletes three arcs 𝐸 (𝑖, 𝑖 + 1),
𝐸 ( 𝑗, 𝑗 + 1), and 𝐸 (𝑙, 𝑙 + 1) in a route 𝑣 𝑖 . It then considers all ways to reconnect
these subpaths, including both inverted and non-inverted versions. In total,
eight neighborhoods are explored, and the operator selects the one with the
highest profit.
– N9 - new route: The N9 operator randomly selects a node 𝑖 from route
𝑣 𝑖 and removes it. Then, it creates a new single tour 𝑣 𝑖 from node 𝑖. The
collection day 𝑘 𝑣 for tour 𝑣 𝑖 is selected by comparing the profits of solution
if route 𝑣 𝑖 carries out in different days within the collection period 𝐾 𝑣 . Then
the solution with highest profit is selected.
– N10 - split to single: The N10 operator deletes two randomly selected
nodes 𝑖 and 𝑗 from two distinct routes 𝑣 𝑖 and 𝑣 𝑗 , respectively. It then creates
a new tour by connecting the two nodes. There are two neighborhood created
by N10 and the one with higher profit is created. The condition for this move
is that the two selected nodes 𝑖 and 𝑗 must have at least one available day in
common.
268 A. Gunawan et al.

– N11 - combine tours: The N11 operator selects two random tours 𝑣 𝑖 and 𝑣 𝑗
that have a combined total capacity that is less than the maximum capacity
of the largest vehicle in the fleet. It then combines these tours to create a new
tour. Two neighborhoods are created by the N11 operator, and the one with
the higher profit is selected.
All local search operators are applied in the first-improving strategy, as shown
in Algorithm 4.

Algorithm 4. First-improving local search procedure

1: Input: A feasible solution, 𝑆
2: improvement ← TRUE
3: while improvement = TRUE do
4: improvement ← FALSE
5: for 𝑆  ∈ 𝑁 (𝑆) do
6: if 𝑓 (𝑆  ) > 𝑓 (𝑆) then
7: 𝑆 ← 𝑆
8: improvement ← TRUE
9: end if
10: end for
11: end while
12: Return: 𝑆

Algorithm 5. Path-relinking
1: Input: 𝑆𝑖 , 𝑆𝑔
2: 𝑆 ← 𝑆𝑖 ;
3: 𝑓 (𝑆 ∗ ) ← max { 𝑓 (𝑆𝑖 ), 𝑓 (𝑆𝑔 ) };
4: 𝑆 ∗ ← arg max { 𝑓 (𝑆𝑖 ), 𝑓 (𝑆𝑔 ) };
5: Symmetric difference, Δ(𝑆, 𝑆𝑔 ) ← 𝑖 : 𝑆 [𝑖 ] ≠ 𝑆𝑔 [𝑖 ] ;
6: while Δ(𝑆, 𝑆𝑔 ) ≠ ∅ do
7: An index 𝑡 ∈ Δ(𝑆, 𝑆𝑔 ) where 𝑆 [𝑡 ] ≠ 𝑆𝑔 [𝑡 ] ;
8: Select a node 𝑖 = 𝑆 [𝑡 ] and 𝑗 = 𝑆 [𝑆𝑔 [𝑡 ] ] ;

9: 𝑆 ←Swap two selected nodes, swap(𝑆𝑖 , 𝑖, 𝑗 );

10: if valid(𝑆 ) = True then
11: if 𝑓 (𝑆 ) > 𝑓 (𝑆 ∗ ) then

12: 𝑆∗ ← 𝑆 ;

13: 𝑓 (𝑆 ∗ ) ← 𝑓 (𝑆 ) ;
14: end if
15: end if
16: end while
17: Return: 𝑆 ∗ , 𝑓 (𝑆 ∗ ) ;

Path-Relinking (PR). PR is an intensification search strategy [5] to create

a series of restricted neighborhood solutions by exploring the trajectory that
connects two elite solutions from the solution pool. In our study, PR is employed
to find the solution that can effectively deal with the penalty and idle time caused
by the time windows soft constraint, as shown in Algorithm 5. Firstly, the PR
procedure takes two solutions 𝑆𝑖 and 𝑆 𝑔 and sets the solution with higher profit
 GRASP Solution Approach for the E-Waste Collection Problem 269

as the current best solution 𝑆 ∗ (Lines 3–4). Then, Line 5 computes the symmetric
difference Δ between two solutions 𝑆 and 𝑆 𝑔 , which stores the index where nodes
in 𝑆 differ from 𝑆 𝑔 . The procedure performs the swap operator to relink the
solution 𝑆 to 𝑆 𝑔 (Lines 6–9). An example for a solution with eight nodes is

illustrated in Fig. 2. If the explored solution 𝑆 in the relinking process has a
higher profit than the current best solution 𝑆 ∗ , it becomes the best solution
in the following relinking path. There is a validity step in this procedure to
avoid partial solutions created in the relinking process (Line 10). The procedure
terminates when 𝑆 reaches 𝑆 𝑔 (i.e., Δ(𝑆, 𝑆 𝑔 ) = ∅), and it returns the best solution
in the relinking process 𝑆 ∗ .

Fig. 2. The path-relinking process for solutions with eight nodes

Integration of PR in GRASP. Algorithm 6 presents the main procedure

of GRASP complemented by PR. The original GRASP comprises two phases,
executed from Lines 4 to 12. In each iteration, GRASP starts by creating an
initial solution using the construction heuristic, followed by applying the repair
operator to ensure feasibility with the available period constraint (Lines 5–6).
Next, the collection day 𝑘 of each tour in the initial solution 𝑆 is randomly
selected within the available period of that tour 𝐾 (Line 8). Eleven local search
operators are then applied to improve the quality of the initial solution 𝑆 (Lines
10–12). If the local optimal solution 𝑆 satisfies the condition 𝑓 (𝑆) > (1− 𝜇) 𝑓 (𝑆 ∗ ),
where 𝜇 is the diversification coefficient and 𝑆 has not already been added to
pool 𝑃, it is added to 𝑃 (Line 13). If the pool is full, 𝑆 replaces the oldest elite
solution in the pool. If there is at least one elite solution in 𝑃, the algorithm
stores the elite solution 𝑝 ∈ 𝑃 having the same number of routes with current

solution 𝑆 to a set 𝑃 . After that, a random elite solution 𝑆 ∈ 𝑃  is selected, and
forward and backward PR are performed between the current solution 𝑆 and
270 A. Gunawan et al.


elite solution 𝑆 (Lines 21–22). The better solution achieved from forward and
backward PR becomes the current solution 𝑆, and if it is not in 𝑃, it is added
to the pool (Lines 23–30). Next, the current solution 𝑆 is compared with the
best solution 𝑆 ∗ . If 𝑆 is better than 𝑆 ∗ , it becomes the best solution 𝑆 ∗ for the
following iteration (Lines 32–34). After 𝑖 𝑚𝑎𝑥 iterations, the algorithm returns
the best found solution 𝑆 ∗ (Line 37).

Algorithm 6. GRASP reinforced by path-relinking

1: Input: Set of nodes, N
2: Number of iteration, 𝑖𝑚𝑎𝑥
3: Initial best profit 𝑓 (𝑆 ∗ ) ← −∞
4: for 𝑖 = 1 to 𝑖𝑚𝑎𝑥 do
5: 𝑆 ← ConstructionHeuristic(N )
6: 𝑆 ←RepairProcedure(𝑆 )
7: for each route 𝑠 in 𝑆 do
8: Randomly select a collection day 𝑘 in available period 𝐾
9: end for
10: for each operator 𝑁 ∈ {𝑁6 , 𝑁7 , 𝑁8 , 𝑁5 , 𝑁1 , 𝑁2 , 𝑁4 , 𝑁9 , 𝑁10 , 𝑁11 , 𝑁3 } do
11: 𝑆 ← 𝑁 (𝑆)
12: end for
13: if 𝑓 (𝑆) > ( 1 − 𝜇) 𝑓 (𝑆 ∗ ) then
14: if 𝑆 ∉ 𝑃 then
15: Add 𝑆 to 𝑃
16: end if
17: end if
18: if |𝑃 | > 0 then
19: Set of elite solution which have the same number of routes with 𝑆 , 𝑃

20: Randomly select an elite solution 𝑆 ∈ 𝑃
𝑓 
21: 𝑆 ← PathRelinking(𝑆 , 𝑆 )

22: 𝑆 𝑏 ← PathRelinking(𝑆 , 𝑆 )
23: if 𝑓 (𝑆 𝑓 ) > 𝑓 (𝑆 𝑏 ) then
24: 𝑆 ← 𝑆𝑓
25: else
26: 𝑆 ← 𝑆𝑏
27: end if
28: if 𝑆 ∉ 𝑃 then
29: Add 𝑆 to 𝑃
30: end if
31: end if
32: if 𝑓 (𝑆) > 𝑓 (𝑆 ∗ ) then
33: 𝑓 (𝑆 ∗ ) ← 𝑓 (𝑆)
34: 𝑆∗ ← 𝑆
35: end if
36: end for
37: Return: 𝑆 ∗

3 Computational Results

We proposed a new set of instances based on the well-known Gehring &

Homberger instances [7], by adding various attributes, as summarized in Table 1.
N1 and N2 denote the number of households and e-bins, while 𝑉1 and 𝑉2 rep-
resent the numbers of vehicle types I and II, respectively. For vehicle type I,
the instance specifications include a capacity of 500, an operating cost of 2, a
penalty cost of 7, and an idle cost of 3. On the other hand, for vehicle type II, the
 GRASP Solution Approach for the E-Waste Collection Problem 271

corresponding values are 100, 1, 2, and 5, respectively. It is important to note

that all instances consider a 4-day delivery period. Instances in Group C fea-
ture clustered customers, whereas instances in group R have customer locations
generated randomly over a square.

Table 1. Summary of benchmark instances

No Name N1 N2 𝑉1 𝑉2 No Name N1 N2 𝑉1 𝑉2
1 C1_6_1_01 4 4 4 4 19 R1_6_1_01 15 15 8 8
2 C1_6_1_02 4 8 6 6 20 R1_6_1_02 20 20 10 10
3 C1_6_1_03 4 12 8 8 21 R1_6_1_03 25 25 13 13
4 C1_6_1_04 8 4 3 3 22 R1_6_1_04 30 30 15 15
5 C1_6_1_05 12 4 4 4 23 R1_6_1_05 35 35 18 18
6 C1_6_1_06 6 6 3 3 24 R1_6_1_06 40 40 20 20
7 C1_6_1_07 8 8 4 4 25 R1_6_1_07 45 45 23 23
8 C1_6_1_08 10 10 5 5 26 R1_6_1_08 50 50 25 25
9 C1_6_1_09 12 12 6 6 27 R1_6_1_09 55 55 28 28
10 C2_6_1_01 4 4 2 2 28 R2_6_1_01 15 15 8 8
11 C2_6_1_02 4 8 3 3 29 R2_6_1_02 20 20 10 10
12 C2_6_1_03 4 12 4 4 30 R2_6_1_03 25 25 13 13
13 C2_6_1_04 8 4 3 3 31 R2_6_1_04 30 30 15 15
14 C2_6_1_05 12 4 4 4 32 R2_6_1_05 35 35 18 18
15 C2_6_1_06 6 6 3 3 33 R2_6_1_06 40 40 20 20
16 C2_6_1_07 8 8 4 4 34 R2_6_1_07 45 45 23 23
17 C2_6_1_08 10 10 5 5 35 R2_6_1_08 50 50 25 25
18 C2_6_1_09 12 12 6 6 36 R2_6_1_09 55 55 28 28

The exact solutions are obtained by IBM ILOG CPLEX Optimization Studio
- Academic version 12.10.0.0 using branch and bound methods on a personal
computer with Intel Core i7-12700 2.10 GHz, 16 GB RAM, 1000 GB HDD +
500 GB SSD, Windows 10 Education version 22H2 64-bit. The time limits are
set to 4 h per instance with the mathematical model adopted from [7]. Following
a series of tuning experiments, we have determined the maximum number of
iterations to 𝑖 𝑚𝑎𝑥 = 5000. For the PR procedure, we set a pool size of 𝜆 = 20
and a diversification coefficient of 𝜇 = 0.007. The GRASP-PR algorithm was
implemented by using Python 3.11.
Table 2 provides the results obtained by both GRASP-PR and CPLEX in
terms of solution quality and computational time. The bold numbers are optimal
solutions obtained by CPLEX, only 4 out of 36 instances. The remaining ones are
the best found so far or unsolved. Even for small instances like C1_6_1_01 and
C2_6_1_01, CPLEX requires 1200 and 2160 s seconds, respectively, to obtain
results. On the other hand, GRASP-PR can generate solutions within signifi-
cantly shorter computational times. For example, GRASP-PR can achieve an
optimal solution for instance C1_6_1_01 in 0.54 s. The maximum computa-
272 A. Gunawan et al.

Table 2. Comparison performance between MILP and GRASP-PR

No Instance CPLEX GRASP-PR No Instance CPLEX GRASP-PR

Obj CPU(s) Best Avg. 10 CPU (s) Obj CPU(s) Best Avg. 10 CPU (s)
1 C1_6_1_01 377 1200 377 377 0.54 19 R1_6_1_01 – 14400 867 323.3 185.36
2 C1_6_1_02 –222 14400 –222 –222 51.17 20 R1_6_1_02 – 14400 546 244.2 290.29
3 C1_6_1_03 –1059 14400 –749 –823.7 83.3 21 R1_6_1_03 – 14400 284 –278.1 388.25
4 C1_6_1_04 1161 14400 1191 1190.5 47.97 22 R1_6_1_04 – 14400 –796 –1201 442.99
5 C1_6_1_05 1997 14400 1884 1815.7 77.84 23 R1_6_1_05 – 14400 –619 –1464.9 578.9
6 C1_6_1_06 –77 14400 –77 –102.7 52.17 24 R1_6_1_06 – 14400 –1386 –2171 721.2
7 C1_6_1_07 730 14400 799 629.1 79.53 25 R1_6_1_07 – 14400 –1676 –3391.6 761.32
8 C1_6_1_08 572 14400 614 528 115.76 26 R1_6_1_08 – 14400 –3420 –5128.4 877.44
9 C1_6_1_09 – 14400 1963 1790.4 147.68 27 R1_6_1_09 – 14400 –3994 –6677.4 982.15
10 C2_6_1_01 –306 2160 –306 –306 29.64 28 R2_6_1_01 –1574 14400 –380 –901.4 184.09
11 C2_6_1_02 –595 14400 –595 –595 40.25 29 R2_6_1_02 – 14400 –1597 –2768.8 254.2
12 C2_6_1_03 –931 14400 –939 –1020.5 68.36 30 R2_6_1_03 – 14400 –3278 –5106 308.97
13 C2_6_1_04 639 14400 670 646.9 59 31 R2_6_1_04 – 14400 –9532 –13630.8 308.81
14 C2_6_1_05 852 14400 861 415.6 103.51 32 R2_6_1_05 – 14400 –6070 –11593.8 429.289
15 C2_6_1_06 346 14400 246 181.5 41.52 33 R2_6_1_06 – 14400 –15829 –20584.9 496.659
16 C2_6_1_07 494 14400 428 281.3 97.14 34 R2_6_1_07 – 14400 –23385 –37733.6 563.798
17 C2_6_1_08 320 14400 60 –163.4 111.4 35 R2_6_1_08 – 14400 –19285 –30553.5 730.22
18 C2_6_1_09 – 14400 371 93.9 146.04 36 R2_6_1_09 – 14400 –29888 –50331.5 819.195

tional time required by GRASP-PR for all instances is 982.15 s, with an average
running time of 296.55 s per instance.
In terms of solution quality, unlike CPLEX, GRASP-PR can successfully
solve all test instances and reach optimal solutions for instances solved by
CPLEX optimally. Moreover, for the remaining instances where CPLEX fails
to achieve optimality, the best found solution by GRASP-PR in 10 runs outper-
forms CPLEX in 7 out of 13 instances. For instance C1_6_1_03, the solution
found by GRASP-PR can improve the profit in CPLEX’s solution by 29.27%,
while for a larger instance R2_6_1_01 with 30 customers, the improvement is
as high as 75.85%. However, GRASP-PR shows limitations in its ability to scale
up to larger problem instances. The solutions obtained by GRASP-PR for the
same instances in different runs become inconsistent when the size of the problem
increases. Specifically, for the two largest instances in our test set, R2_6_1_08
and R2_6_1_09, with 100 and 110 customers, respectively, the average objec-
tive value of ten runs is almost 1.5 times worse than the objective value of the
best-found solution.
We also compare the 2-opt and 3-opt local search operators in terms of solu-
tion quality. We created two alternative scenarios: (i) the GRASP-PR meta-
heuristic without the 2-opt operator, and (ii) the GRASP-PR metaheuristic
without the 3-opt operator. Results are summarized in Table 3, together with
those of the original algorithm in terms of solution quality and computational
times.
We observe that both scenarios produce worse solutions than those of the
original GRASP-PR metaheuristic. This outcome is expected since the HVRP-
MTW represents a tightly constrained problem, and reducing the local search
operator leads to a decrease in the search space, thereby making it more chal-
 GRASP Solution Approach for the E-Waste Collection Problem 273

Table 3. Comparing the impact of 2-opt and 3-opt on algorithm performance

No Instance GRASP-PR i ii
Best Avg. 10 CPU (s) Best Avg. 10 CPU (s) Best Avg. 10 CPU (s)
1 C1_6_1_01 377 377 0.54 377 373.3 0.54 377 377 0.49
2 C1_6_1_02 –222 –222 51.17 –222 –228.4 51.65 –222 –222 48.35
3 C1_6_1_03 –749 –823.7 83.30 –1045 –1086.9 83.40 –1052 –1101.2 74.60
4 C1_6_1_04 1191 1190.5 47.97 1191 1187.9 49.41 1191 1187.8 42.77
5 C1_6_1_05 1884 1815.7 77.84 1850 1786.6 73.53 1884 1776.9 60.26
6 C1_6_1_06 –77 –102.7 52.17 –77 –104.1 49.84 –77 –99.8 45.63
7 C1_6_1_07 799 629.1 79.53 799 669.3 80.41 751 648.5 67.08
8 C1_6_1_08 614 528 115.77 596 507.5 118.63 592 513.6 101.48
9 C1_6_1_09 1963 1790.4 147.68 1932 1732.9 141.90 1863 1690.2 125.71
10 C2_6_1_01 –306 –306 29.64 –306 –306 30.71 –306 –306 25.65
11 C2_6_1_02 –595 –595 40.25 –595 –595 42.10 –595 –595 32.43
12 C2_6_1_03 –939 –1020.5 68.36 –957 –1034.5 65.50 –1019 –1083.1 55.09
13 C2_6_1_04 670 646.9 59.00 669 630.9 59.26 670 592.6 54.39
14 C2_6_1_05 861 415.6 103.51 516 367.4 104.50 595 377.5 86.38
15 C2_6_1_06 246 181.5 41.52 189 161.9 41.05 258 152.8 35.58
16 C2_6_1_07 428 281.3 97.14 303 229.7 93.32 363 286.8 82.50
17 C2_6_1_08 60 –163.4 111.40 5 –190.8 105.90 –24 –269.9 91.48
18 C2_6_1_09 371 93.9 146.04 328 44.6 141.26 196 22.2 119.51

lenging to achieve better solutions. Comparing both scenarios, alternative (i)

exhibited better performance in most instances, confirming that the 3-opt oper-
ator outperforms the 2-opt operator in terms of solution quality. The reason for
this disparity lies in the fact that the 3-opt operator enables the algorithm to
explore much larger neighborhoods compared to the 2-opt operator. The better
solution quality by the original algorithm, in comparison to the two scenar-
ios, suggests that the 2-opt and 3-opt operators lead to different local optima.
Consequently, combining both the 2-opt and 3-opt operators can enhance the
algorithm’s performance.
Regarding the computational time, the reduction in the 2-opt operator had
a negligible impact, approximately 1%. Conversely, the GRASP-PR without the
3-opt operator exhibited an average acceleration of approximately 15%. This
discrepancy can be attributed to the time complexity of the 3-opt operator being
𝑂 (𝑛3 ), whereas the 2-opt operator has a time complexity of 𝑂 (𝑛2 ).

4 Conclusions
In this study, we propose a GRASP with Path-Relinking to address the e-waste
collection route planning problem modelled as Heterogeneous Vehicle Routing
Problem with Multiple Time Windows (HVRP-MTW). To evaluate the pro-
posed algorithm, named GRASP-PR, we conducted comparative experiments
274 A. Gunawan et al.

against a commercial solver. The results show that GRASP-PR can effectively
handle HVRP-MTW instances with up to 110 customers within reasonable com-
putational times. The comparison revealed that GRASP-PR outperforms the
commercial solver in large-scale instances and can generate optimal solutions
for small instances. However, the metaheuristic approach of GRASP-PR has its
limitation in convergence when the problem size increases. For future research,
the integration of advanced extensions with a learning process such as fixed set
search [8,9] is considered to enhance the convergence and performance of the
current metaheuristic.

Acknowledgement. This research was supported by the Singapore Ministry of Edu-

cation (MOE) Academic Research Fund (AcRF) Tier 1 grant.

References
1. Baldé, C., D’Angelo, E., Deubzer, V.L.O., Kuehr, R.: Global transboundary e-
waste flows monitor - 2022, United Nations Institute for Training and Research
(UNITAR), December 2022. https://ewastemonitor.info/gtf-2022/. Accessed 26
June 2023
2. Braekers, K., Ramaekers, K., Van Nieuwenhuyse, I.: The vehicle routing problem:
state of the art classification and review. Comput. Ind. Eng. 99, 300–313 (2016)
3. Clarke, G., Wright, J.W.: Scheduling of vehicles from a central depot to a number
of delivery points. Oper. Res. 12(4), 568–581 (1964)
4. Feo, T.A., Resende, M.G.: A probabilistic heuristic for a computationally difficult
set covering problem. Oper. Res. Lett. 8(2), 67–71 (1989)
5. Glover, F.: Tabu search and adaptive memory programming-advances, applications
and challenges. In: Barr, R.S., Helgason, R.V., Kennington, J.L. (eds.) Interfaces
in Computer Science and Operations Research. Operations Research/Computer
Science Interfaces Series, vol. 7, pp. 1–75. Springer, Boston, MA (1997). https://
doi.org/10.1007/978-1-4615-4102-8_1
6. Golden, B., Assad, A., Levy, L., Gheysens, F.: The fleet size and mix vehicle routing
problem. Comput. Oper. Res. 11(1), 49–66 (1984)
7. Gunawan, A., Nguyen, M.P.K., F. Yu, V., Nguyen, D.V.A.: The heterogeneous
vehicle routing problem with multiple time windows for the e-waste collection
problem. In: submitted to IEEE CASE 2023 (accepted) (2023)
8. Jovanovic, R., Sanfilippo, A.P., Voß, S.: Fixed set search applied to the clique
partitioning problem. Eur. J. Oper. Res. 309(1), 65–81 (2023)
9. Jovanovic, R., Tuba, M., Voß, S.: Fixed set search applied to the traveling sales-
man problem. In: Blesa Aguilera, M.J., Blum, C., Gambini Santos, H., Pinacho-
Davidson, P., Godoy del Campo, J. (eds.) HM 2019. LNCS, vol. 11299, pp. 63–77.
Springer, Cham (2019). https://doi.org/10.1007/978-3-030-05983-5_5
10. Król, A., Nowakowski, P., Mrówczyńska, B.: How to improve WEEE management?
Novel approach in mobile collection with application of artificial intelligence. Waste
Manag. 50, 222–233 (2016)
11. Mar-Ortiz, J., González-Velarde, J.L., Adenso-Díaz, B.: Designing routes for
WEEE collection: the vehicle routing problem with split loads and date windows.
J. Heuristics 19, 103–127 (2013)
 GRASP Solution Approach for the E-Waste Collection Problem 275

12. Pourhejazy, P., Zhang, D., Zhu, Q., Wei, F., Song, S.: Integrated e-waste trans-
portation using capacitated general routing problem with time-window. Transp.
Res. Part E Logist. Transp. Rev. 145, 102169 (2021)
13. Yao, L., He, W., Li, G., Huang, J.: The integrated design and optimization of a
WEEE collection network in Shanghai, China. Waste Manag. Res. 31(9), 910–919
(2013)
 The Heterogeneous Fleet
Risk-Constrained Vehicle Routing
Problem in Humanitarian Logistics

Robert M van Steenbergen(B) , Eduardo Lalla-Ruiz ,

Wouter van Heeswijk , and Martijn Mes

Department of High-Tech Business and Entrepreneurship, University of Twente,

Enschede, The Netherlands
[email protected]

Abstract. While distributing essential supplies in volatile environ-

ments, humanitarian transport is often exposed to threats such as
attacks. To mitigate the negative consequences of attacks, we intro-
duce the heterogeneous fleet risk-constrained vehicle routing problem
(HFRCVRP), in which we aim to minimize transportation costs and
the expected loss of getting robbed. An Adaptive Large Neighborhood
Search (ALNS) heuristic is presented to solve the problem. The trade-off
between transportation costs and expected loss of attacks is analyzed
with a real-world case in South Sudan. Results show that the trade-off
is especially relevant in the heterogeneous variant, in which Unmanned
Aerial Vehicles (UAVs) can effectively mitigate risks of truck transport,
providing a 7.1% improvement of the objective value compared to the
same instance with only trucks. Risks can be completely eliminated by
increasing transportation costs by a factor of five. Additionally, the risk
variant decreases the objective value by 14.8% compared to considering
only transportation costs and ignoring risks.

Keywords: Risk-constrained Routing · Humanitarian Logistics ·

UAVs · Heterogeneous Fleet · Metaheuristic

1 Introduction
During crises and disasters, humanitarian assistance is essential for alleviating
the suffering of those in need. This involves transporting goods such as food,
medicine, and in increasing numbers cash-based interventions [9]. However, these
movements are not without risks. People may resort to violent behavior in des-
perate times with a lack of resources such as food, water, or medication. For
instance, one week after the 2023 Turkey-Syria earthquake, multiple incidents of
violence and armed looting were reported. Furthermore, in conflict zones, rebel-
lious groups can also be interested in relief goods. Facing the attacks in multiple
cities in Sudan in the spring of 2023, with explosions, gunfire, and looting in the
streets, the World Food Programme had to suspend one of its largest food dis-
tribution missions. Moreover, during the civil war in South Sudan (2013–2020),
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 276–291, 2023.
https://doi.org/10.1007/978-3-031-43612-3_17
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 277

combined with hunger and malnutrition in the country, dozens of incidents were
reported each month, including murders of aid workers. To avoid risks on the
roads, costly airlifts by helicopters and aircraft are often performed, which are
mainly suitable for large shipments. Unmanned Aerial Vehicles (UAVs) can pro-
vide a flexible and efficient alternative for smaller-scale distributions of aid orga-
nizations. Humanitarian cargo UAVs are not necessarily bound to fly between
airports, have the ability to deliver supplies to numerous locations spread over
an area, and can visit hard-to-reach places quickly [17].
Literature offers various secure routing approaches for high-value goods, often
referred to as cash-in-transit (CIT) problems. The transportation of valuable
relief goods in disaster areas can face comparable risks as cash transport. In most
CIT problems, the risk of vehicle attacks is mitigated by diversifying routing
plans through time inconsistency or path inconsistency [7]. However, random-
ization of routing plans is less suitable for humanitarian operations, as these
often deal with urgent, variable or one-off demands, sparse road networks, and
opponents that tend to operate opportunistically. An alternative approach is the
risk-constrained vehicle routing problem (RCVRP) introduced by Talarico et al.
[22], which limits the risk exposure in routes. Despite these approaches in CIT
literature, security issues are seldom considered in humanitarian research, yet
being a top concern for humanitarian organizations [3].
This paper investigates the problem of transporting high-value goods with a
heterogeneous fleet (i.e., trucks and UAVs) in high-risk disaster environments.
The following contributions are made in this paper. (i) We formulate the hetero-
geneous fleet risk-constrained vehicle routing problem (HFRCVRP) including
indirect losses of an attack. (ii) We propose an adaptive large neighborhood
search (ALNS) algorithm with specific operators to solve the HFRCVRP in a
reasonable time, as the HFRCVRP may need to be solved daily or suddenly after
a disaster. (iii) The methods are tested and results are analyzed using real-world
data from South Sudan.
The remainder of this paper is organized as follows. Related literature on the
HFRCVRP is reviewed in Sect. 2. In Sect. 3, we describe the problem and formu-
late the model. Section 4 introduces our ALNS heuristic. Section 5 describes the
real-world case study of South Sudan, the experimental design, and the results.
Finally, the conclusions and future research directions are stated in Sect. 6.

2 Related Work
In this section, we first describe the risk-constrained vehicle routing problem
(RCVRP) and then review literature on variants of the RCVRP that are pro-
posed in recent years. Talarico et al. [22] formally introduce the RCVRP. The
expected loss along edge (i, j) is the product of the probability of being robbed
pij (i.e., risk probability) and the loss Di of valuables transported from node i.
In the case of collecting cash, Di will increase every stop, in the case of deliveries,
Di will decrease at a stop. As the probability of being robbed along an edge is
low, Talarico et al. [22] simplify the expected loss associated with route r by
taking the sum of expected losses of robbery on each arc (i, j):
278 R.M. van Steenbergen et al.


pij · Dir (1)
(i,j)∈r

The authors assume the probability of an attack (pij ) is proportional to

the length of the edge traversed. They aim to minimize the total length of the
routes while constraining the expected loss in each route. As the RCVRP is
more complex than the traditional VRP, which is already known to be NP-
hard, the authors propose four construction heuristics in a greedy randomized
adaptive search procedure (GRASP) and an iterative local search (ILS) to solve
the problem in reasonable time. Talarico et al. [24] develop a combination in
which ant colony optimization is used for generating an initial solution that is
later improved by a large neighborhood search. Moreover, for the same problem,
a fuzzy GRASP with path relinking is presented by Radojičić et al. [16]. Their
method proves to be more effective and further improves the results of solving
the RCVRP regarding solution objectives and computation times.
In recent years, several variants of the RCVRP have been proposed. The
multi-objective variant of the RCVRP is proposed by Talarico et al. [23] and
evaluated on real-world instances with capacity constraints. They minimize (i)
the transportation costs and (ii) the maximum risk in a trip, instead of includ-
ing risk as a routing constraint. A weighted sum of the objective is proposed to
adjust the two terms according to the decision-maker’s preference. A progres-
sive multi-objective optimization with ILS is suggested to solve the problem.
Ghannadpour & Zandiyeh [8] build upon the multi-objective variant and define
a zero-sum game with two players (the vehicle and the robber) to periodically
update the risk probabilities pij . For generating route plans each period and
minimizing travel distance and risk, a multi-objective hybrid genetic algorithm
is developed and its efficiency and effectiveness are examined through several
instances. Soeanu et al. [21] introduce a multi-depot variant to minimize rout-
ing costs, vehicle breakdowns (i.e., risk based on vehicle value), and delivery
failure costs (i.e., risk based on transported goods). A learning-based heuris-
tic technique is proposed, combining a guided stochastic cost-insertion gradient
descent technique with iterative solution improvement, and evaluated in a com-
parative study. Mazdarani et al. [15] analyze a multi-objective RCVRP problem
with overlapping links (i.e., arcs) to minimize total costs and expected loss. The
authors consider an increased risk when routes overlap and develop a hybrid
genetic algorithm to solve the problem. Tikani et al. [25] extend the RCVRP
with time-dependent travel times, where increasing travel times (e.g., due to traf-
fic jams) increase the exposure to danger. A hybrid method based on dynamic
programming and genetic algorithms is proposed and the results of this heuris-
tic approximate the values of optimal solutions. Allahyari et al. [2] combine the
time-dependent travel times and multi-objective RCVRP with pickup and deliv-
ery operations and time windows. An efficient approach combining GRASP and
ILS, including a robust objective, is developed to minimize travel costs and a
specific risk index. In this index, they are not only considering the probability
and loss of a robbery (risk exposure) but also minimizing the use of overlap-
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 279

ping arcs and the spread arrival times (i.e., route predictability). Allahyari et
al. [1] apply the same risk index in a secure pickup and delivery problem with
time windows and include facility location decisions as well. To solve realistic
instances, they propose an adaptive large neighborhood search (ALNS) algo-
rithm, which outperforms an exact method considering a computational time
limit, and minimizes the risk index, depot costs, vehicle costs, and travel costs.
A two-echelon variant of the RCVRP is presented by Fallahtafti et al. [6], which
includes facility location decisions for multiple depots. A multi-objective simu-
lated annealing approach provides the most promising results in a computational
study to minimize depot costs, transportation costs, and risk exposure. Kian et
al. [11] consider an RCVRP for cash-based interventions in the Syrian refugee
crisis. They extend the works on facility location decisions and multi-echelon
problems by introducing temporary facilities and mobile depots. The goal is to
maximize the number of reached beneficiaries within a limited time period while
minimizing logistics costs. A hierarchical metaheuristic is developed to obtain
efficient solutions. An overview of all the works and their characteristics is pre-
sented in Table 1.
Table 1. Overview of literature regarding the RCVRP. (MO = multi-objective, MD
= multi-depot, ME = multi-echelon, LR = location-routing, TD = time-dependent
risk, PD = pickup and delivery, TW = time windows, OL = overlapping links, MP =
multi-period, HF = heterogeneous fleet)

Work Problem feature Solution approach Instances

MO MD ME LR TD PD TW OL MP HF
Talarico et al. (2015) [22] GRASP/ILS Stylized
Talarico et al. (2017a) [23]  PMOO-ILS Stylized
Talarico et al. (2017b) [24] ACO/LNS Stylized
Radojičić et al. (2018) [16] Fuzzy GRASP Stylized
Ghannadpour & Zandiyeh (2020) [8]   Hybrid GA Case study
Soeanu et al. (2020) [21]   SGD-ILS Stylized
Fallahtafti et al. (2021) [6]     Archived SA Case study
Tikani et al. (2021) [25]  Hybrid DP-GA Case study
Allahyari et al. (2021a) [1]     ALNS Case study
Allahyari et al. (2021b) [2]     GRASP × ILS Stylized
Kian et al. (2022) [11]      Hierarchical LS Case study
Mazdarani et al. (2023) [15]   Hybrid GA Case study
This paper   ALNS Case study

The trend of the RCVRP has shifted towards multi-objective problems,

which minimize (weighted) risks and transportation costs. This creates a trade-
off between the two often conflicting objectives, which can be adapted to the
decision maker’s preferences. Variants such as multi-depot, multi-echelon, time
windows, and location-routing problems have been proposed. The heterogeneous
fleet variant, where certain vehicle types may be exposed to different risks, has
not received much attention. A mixed fleet of vehicles could benefit from risk
mitigation compared to single-vehicle solutions. The application of the RCVRP
to a humanitarian setting is limited as of yet (only [11]), despite daily threats
of violence and the increasing importance of cash-based assistance [10]. Indirect
or unforeseeable losses from attacks, such as human injury, vehicle damage, and
280 R.M. van Steenbergen et al.

road damage, are often ignored in the literature. These losses are considered
to not be under the control of the transportation company [6,22]. Nonetheless,
they can significantly impact operations and could still be mitigated by adapting
decisions (e.g., avoiding them by using UAVs). The risk of vehicle failure is con-
sidered by [21], showing similarities with indirect losses, as vehicle failure is also
unrelated to the amount of cargo. We address this gap in literature by including
indirect losses by estimating the value of the indirect costs of an attack. Further-
more, we formulate the heterogeneous fleet RCVRP, propose an adaptive large
neighborhood search heuristic with specific operators to solve the problem, and
analyze the problem in a humanitarian context in South Sudan.

3 Heterogeneous Fleet Risk-Constrained Vehicle Routing

Problem
In this work, we address the heterogeneous fleet risk-constrained vehicle routing
problem (HFRCVRP), a new variant of the RCVRP. Following the research
trend in this area, we propose a bi-objective variant, in which the objective
is to minimize both the expected loss (risk exposure) and the transportation
costs. The HFRCVRP is defined on a directed graph G = (V, A), where V =
{0} ∪ N = {0, 1, ..., n} corresponds to the depot (vertex 0) and the customers
N = {1, ..., n}. Each customer i ∈ N has a demand di . The set of arcs is defined as
A = {(i, j) ∈ V × V, i ≠ j}. The transportation costs cijk and probability of being
robbed pijk are associated with each arc (i, j) ∈ A and vehicle type k ∈ K. Each
vehicle type k ∈ K has an associated vehicle capacity Qk and multiple routes
r can be planned for each type k ∈ K. Determining the number of routes for
each vehicle type is part of the optimization problem, and is at most equal to
the number of customers n in a solution in which one vehicle type visits each
customer in a separate route. Hence, we define route index r on customer set N .
The decision variables are as follows:

– xijkr , is equal to 1 if arc (i, j) ∈ A is traversed along route r ∈ N of vehicle

type k ∈ K and 0 otherwise.
– Dijkr is the loss in case a robbery happens on arc (i, j) ∈ A along route r ∈ N
of vehicle type k ∈ K.

Besides the direct consequences of a robbery (i.e., the loss of valuable goods),
indirect losses may also occur, such as required vehicle repairs due to damages,
emotional/physical abuse of employees, or administrative work (e.g., insurance).
We include these indirect losses that occur irrespective of the amount of lost
valuables in Dijr . The objective function of the HFRCVRP is to find routes
that minimize the total transportation costs and the total expected loss:
 
min (cijk xijkr + pijk Dijkr ) (2)
k∈K r∈N (i,j)∈A

In the HFRCVRP, the following constraints are considered:

 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 281

– The demand of each customer has to be fully fulfilled.

– All vehicle routes start and end at the depot.
– The sum of deliveries between depot visits cannot exceed the vehicle capacity.
– The loss in case of a robbery is determined by the value of the goods on
the vehicle and a fixed amount of indirect losses. The number of goods on a
vehicle is determined by the sum of planned deliveries before returning to the
depot.
The RCVRP is an extension of the traveling salesman problem, both known
as NP-hard. Hence, the HFRCVRP as an extension of the RCVRP is also NP-
hard and even more challenging to solve. In the next section, we propose a
heuristic approach to solve realistic-sized instances quickly.

4 Adaptive Large Neighborhood Search for the

HFRCVRP
In this section, we adapt the adaptive large neighborhood search (ALNS) meta-
heuristic, introduced by Ropke and Pisinger [18], to solve the HFRCVRP. The
ALNS framework can be modified to various problems and has been successfully
applied to many different VRP variants in the past, including heterogeneous
fleet VRPs (e.g., [14,26]), VRPs with drones (e.g., [13,19]), and the RCVRP [1].
ALNS progressively improves an initial solution by repeatedly destroying and
repairing the current solution. According to their performance achieved during
the search process, the destroy- and repair operators are randomly selected fol-
lowing an adaptive mechanism. A new solution is accepted according to a pre-
determined criterion. We subsequently describe the construction of the initial
solution, the destroy and repair operators, the adaptive selection of operators,
and the acceptance and stopping criteria.

4.1 Initial Solution

We develop a modified cheapest insertion heuristic with randomized selection.
We define a route as a sequence of nodes, starting and ending at the depot
(e.g., route (0, 5, 3, 0)). The construction heuristic starts with empty routes (only
including the depot) and then determines for each customer the increase in the
objective value when including it into a route of each vehicle type k ∈ K. The
combination of customer and vehicle type that results in the lowest increase of
the objective function (i.e., the cheapest insertion) is added to the (empty) route
of the associated vehicle type. For each following iteration, the heuristic considers
all possible insertion positions in existing and empty routes of each vehicle type
and determines for each customer the insertion costs in these positions. The
customer and insertion position with the cheapest insertion is included in this
insertion position at the respective route of the associated vehicle type. Before
the insertion, it is checked if the sum of deliveries in the route does not exceed
the vehicle capacity. To diversify the constructive heuristic, the insertion costs
are multiplied by a random number before selecting the cheapest insertion. The
process is repeated until all customers are allocated.
282 R.M. van Steenbergen et al.

4.2 Destroy Operators

We design five destroy operators that each remove q locations from the solution,
where q is a random variable determined by the parameter settings. The removed
locations are added to the removal list Φ.

– Random removal. Randomly selects q locations that are removed from the
routes and added to Φ.
– Distance removal. Selects a random location and removes it together with
its q − 1 nearest neighbors based on Euclidean distance.
– String removal. This operator proposed in [4] removes adjacent sequences of
locations (i.e., strings) in the routes. First, a seed location is selected, followed
by the removal of a string that encompasses the seed location. Thereafter, a
string is removed from another route that contains a location adjacent to the
seed location.
– Route removal. Randomly selects an existing route of one vehicle type
and completely destroys it. Specifically useful for heterogeneous fleets, as this
allows to mix up the locations of one specific vehicle type.
– Worst (cost/risk) removal. These operators remove the location that
results in the largest decrease in the objective value. To include some ran-
domness in the worst removal, e.g., to avoid the same location removed mul-
tiple times and to avoid local optima, the removal value is randomized, by
multiplying the value by a random number. This process is repeated until q
locations are removed. Two additional variants are developed for the HFR-
CVRP; the worst cost removal only considers the transportation costs and
the worst risk removal only considers the expected loss of the objective value.

4.3 Repair Operators

Once we have partially destroyed the solution by removing locations, we have to
repair the solution by inserting the locations until the removal list Φ is empty.
We have defined four repair operators, related to, e.g., [1,18].

– (Random) greedy repair. The greedy repair operators insert the removed
locations iteratively. In each iteration, the location with the smallest increase
of the objective is inserted in the best possible position, which is either in an
existing route or in a new route. For the random variant, the greedy repair is
randomized by multiplying the insertion value with a random number, aiming
to overcome the myopic behavior of the greedy repair [18].
– (Random) regret repair. The regret repair operators insert locations with
the highest regret. Regret is the difference in the insertion value between
inserting the location at its best position and at another position. When
the other position is the second-best position, it is denoted as 2-regret. In
each iteration of the regret operator, the location with the highest 2-regret is
inserted at its best position. Hence, we insert the location that we would most
regret if it later could not be inserted at its best position. For the random
variant, the regret is multiplied by a random number.
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 283

4.4 Operator Selection and Adaptive Weight Adjustment

The choice of destroy and repair operators in each iteration of ALNS is based
on the roulette-wheel selection principle (e.g., [12,18]). To select the different
operators, each operator k has a weight wi , i ∈ {1, 2, .., k} and we select the
operator j with probability:
wj
k (3)
i=1 wi
The weights of the operators are all initialized with a weight of 1 and updated
throughout the search. The entire search is divided into a number of segments.
Throughout each segment, we keep track of the performance of the operators
by assigning scores σ1 , σ2 , and σ3 . At the start of each segment, the score is set
to zero for each operator. If the operators find a new best solution, the score
is increased by σ1 . If operators find a solution that is better than the current
solution, the score is increased by σ2 . If the solution is worse than the current
solution but is accepted, the score is increased by σ3 . The latter is included to
encourage operators that diversify the search and explore new sections of the
solution space [18]. At the end of a segment, weights are updated according to
their scores. Let wij be the weight of operator i in segment j, πi be the score
of operator i, and θi be the number of times heuristic i is used in segment j. A
reaction factor γ controls the weight update:
πi
wi,j+1 = wij (1 − γ) + γ (4)
θi

4.5 Acceptance and Stopping Criteria

The acceptance criterion we apply is the Record-to-Record Travel (RRT), pro-

posed by [5], and according to a comparative study of [20], the preferred accep-
tance criterion for ALNS. When minimizing, this algorithm accepts a new solu-
tion s if the gap between the value of the new solution f (s ) and the value of the
best-known solution f ∗ is smaller than the positive parameter T . The threshold
parameter T decreases every iteration moving towards its end value, such that
near the end of the search process, only higher-quality solutions are accepted
[20]. We incorporate a linear decay of T : T start /I, where I is the number of
iterations.

5 Case and Numerical Experiments

In this section, we describe the case, the instances, and the experimental settings
which we use to solve the HFRCVRP, and clarify the tuning of the parameters
of the ALNS.
284 R.M. van Steenbergen et al.

5.1 South Sudan Case

A civil war, a widespread scarcity of food, and multiple floods have plagued
South Sudan for years, especially in 2017. According to the humanitarian snap-
shots of the United Nations Office for the Coordination of Humanitarian Affairs
(OCHA) in 2017, Unity was the state with severe food scarcity, the most inter-
nally displaced people, and the highest violence rate in South Sudan. Inci-
dents in Unity included violence against humanitarian personnel (also murders),
ambushes and looting, active hostilities, violence against humanitarian assets,
and operational interference and restrictions of movement. In 2017, the OCHA
reported 182 incidents in Unity, with 6 murders of aid workers among them.
Tonnes of nutritional supplies, essential medicines, and utilities were stolen and
in November 2017 food assistance was suspended in the south of Unity due to
ongoing fights in November 2017. In this study, we focus on this southern part,
containing the districts of Koch, Leer, and Mayiendit. Based on assessments from
the International Organization of Migration (IOM), we obtain a set of 125 loca-
tions where internally displaced people rely on food assistance or health support
(see Fig. 1). The depot is situated in the village of Leer, where a humanitarian
hub and an airstrip are located. The locations contain between 19 and 2,579
individuals, with a total number of 42,654 individuals in the area.

Fig. 1. Map of the operating area; the colors give a sense of the three districts and the
size of the locations scales according to the number of individuals.
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 285

5.2 Experimental Design and Instances

To provide a safer alternative for trucks to distribute goods in the area, we
compare trucks with the deployment of unmanned aerial vehicles (UAVs). Cargo
UAVs can fly over high-risk environments, deliver goods without requiring an
airstrip by means of airdrop or horizontal landing, and are generally cheaper than
manned aircraft when transporting relatively small amounts of goods. The UAV
specifications are based on the humanitarian MiniFreighter from Wings For Aid,
and comparable to the conceptual Nuuva V300 from Pipistrel. We determine
the UAV travel time by using the Euclidean distance between locations and a
cruising speed of 120km/h. To obtain realistic estimates for truck driving times
between all locations, we use ArcGIS 10.7 to obtain the rural driving times. To
obtain the cost parameters cijk between node i and j for vehicle type k, we
multiply the obtained travel times by hourly vehicle costs. Risk probabilities
pijk can depend on various factors, such as the road type, travel time, time of
the day, vehicle type, surroundings, and other information (e.g., insights that
rebel soldiers are located near a specific road). As this detailed information on
attacks is not available, we assume, for trucks, that the probability of an attack is
proportional to the travel time of an arc. The travel times are weighted by a fixed
risk factor to obtain the risk probabilities for trucks (pij0 ). In our experiments,
we use a risk factor of 0.01, based on the information found on the Unity state
in the humanitarian snapshots of the OCHA and the monthly figures of the
Logistics Cluster in 2017. The value of indirect losses of a robbery, irrespective
of the value of goods on the truck, is estimated to be 10,000 euro. For the UAVs,
we omit the risk of a robbery, as they will fly over the risky environment, so
the risk probability of being robbed pij1 = 0 between each node i and node j for
UAVs. The vehicle specifications are summarized in Table 2.

Table 2. Vehicle specifications used in the experiments.

Speed (km/h) Costs (e/h) Capacity Qk Indirect loss (e) Risk factor
Truck 37.5a 70 6,000 10,000 0.01
UAV 120 250 200 0 0.00
a
On average, based on the obtained distances and travel times from ArcGIS

To evaluate the ALNS framework for the HFRCVRP, we perform experi-

ments on seven subsets of the case instance. These subsets are summarized in
Table 3 and are based on the three districts in the south of the Unity state: Koch,
Mayiendit, and Leer. We solve the RCVRP with only trucks and we solve the
HFRCVRP with trucks and UAVs. In the experiments, we sum the transporta-
tion costs and the expected loss to a single objective. Furthermore, we solve the
instance with a focus on only costs, ignoring the expected loss in the objective.
Lastly, as the results are dependent on the level of risk in an area, we perform a
sensitivity analysis with the largest instance where we adapt the weights of the
transportation costs and expected loss (risk exposure) in the objective function.
286 R.M. van Steenbergen et al.

For all experiments, we consider that each individual requires 0.5kg of goods,
with a value of 2 euro per kg.

Table 3. Instances with combinations of the three districts.

Subset District #locations #individuals

1 Koch 31 9,756
2 Mayiendit 40 13,642
3 Leer 54 19,256
4 Koch-Mayiendit 71 23,398
5 Koch-Leer 85 29,012
6 Mayiendit-Leer 94 32,898
7 All 125 42,654

5.3 ALNS Parameter Tuning

The ALNS framework contains several parameters that can be tuned according
to the problem setting, but once they are tuned on one instance, results tend to
be robust for a wide range of instances [18]. We tune our parameters on instance
7, the largest and most complex instance. To find suitable parameters, we varied
the parameters and chose the setting that showed the best performance. The
number of locations to remove each iteration q (i.e., the degree of destruction)
is picked from the uniform interval [0.1n, 0.2n], where n is the total number of
locations. The level of randomization of the costs for the initial construction, the
worst removal operators, and the randomized repair operators is set to [0.9, 1.1].
Regarding the scoring and weight updating of the operators, we divide the search
into segments of 50 iterations. In each iteration, weights get a score based on
their performance and the scores σ1 , σ2 , and σ3 , which are set to 30, 10, and
5, respectively. The reaction factor γ for weight updating is set to 0.1. The
acceptance threshold T of RRT is a function of the best solution f ∗ , with initial
threshold T start = 0.05f ∗ . The threshold decays linearly throughout the search
until T = 0.

5.4 Experimental Results

The results obtained by the ALNS heuristic for all variants on the instances
are presented in Table 4, and were obtained within minutes on a 4-core 1.90
GHz computer with 16GB RAM. Without considering risks, UAVs are only
applied in three subsets and allocated to a total of five locations. About 1.1%
in transportation costs is saved in these instances compared to truck-only sce-
narios. When including the expected loss in the objective, UAVs are deployed
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 287

in all instances visiting on average 19% of the locations, improving the objec-
tive of transportation costs and expected loss by 7.1% on average compared to
only trucks. Additionally, transportation costs per location decrease by 30% on
average for trucks when UAVs are deployed, showing that UAVs visit locations
further away from the depot. These relatively remote locations are also those
that result in a high exposure to risk for trucks, due to their longer travel times.
When comparing the bi-objective results with the results of the single cost
objective, we observe that for the truck-only instances, the transportation costs
increase by 11.9% and the expected loss decreases by 13.4% on average. For the
heterogeneous fleet instances, the effect is about four times as large, with an
average increase of 53.0% in costs, and a 47.7% average decrease of expected
loss, resulting in an overall decrease of 14.8% in the objective value.

Table 4. Results for all experiments, where each instance stands for the case subset
(1–7), truck-only (T) or heterogeneous fleet with trucks and UAVs (HF), and a single
objective based on costs (C) or bi-objective based on costs and risks (CR). The lowest
total costs and expected losses per subset are marked in bold.

Instance Truck costs UAV costs Expected loss Total # truck # UAV
locations locations
1-T-C 1824.2 0 4198.1 6022.2 31 0
1-T-CR 1850.8 0 3168.2 5019.0 31 0
1-HF-C 1824.2 0 4198.1 6022.2 31 0
1-HF-CR 622.9 2927.4 1076.8 4627.2 20 11
2-T-C 413.2 0 722.2 1135.4 54 0
2-T-CR 417.6 0 706.6 1124.2 54 0
2-HF-C 343.9 67.7 648.4 1059.9 51 3
2-HF-CR 269.9 204.8 484.7 959.4 42 12
3-T-C 837.1 0 1523.3 2360.4 40 0
3-T-CR 843.9 0 1453.2 2297.2 40 0
3-HF-C 762.7 54.9 1678.8 2496.5 39 1
3-HF-CR 445.7 946.2 733.8 2125.8 32 8
4-T-C 2186.7 0 4440.9 6627.6 85 0
4-T-CR 2224.2 0 3832.8 6057.0 85 0
4-HF-C 2186.7 0 4440.9 6627.6 85 0
4-HF-CR 896.8 3109.0 1591.3 5597.0 65 20
5-T-C 2283.9 0 4403.7 6687.6 71 0
5-T-CR 2294.9 0 3933.1 6228.0 71 0
5-HF-C 2214.1 54.9 4220.1 6489.1 70 1
5-HF-CR 1315.0 2078.8 2402.3 5796.2 63 8
6-T-C 1052.8 0 2474.0 3526.8 94 0
6-T-CR 1064.3 0 1910.1 2974.4 94 0
6-HF-C 1052.8 0 2474.4 3527.2 94 0
6-HF-CR 1034.5 72.4 1848.3 2955.2 92 2
7-T-C 2481.7 0 5194.7 7676.4 125 0
7-T-CR 2524.1 0 4398.7 6922.8 125 0
7-HF-C 2481.7 0 5194.7 7676.4 125 0
7-HF-CR 1481.3 2311.2 2831.8 6624.3 105 20
288 R.M. van Steenbergen et al.

For the largest instance with all districts, we perform a sensitivity analysis
to compare the separate objectives of transportation costs and the expected loss
with different weights, see Fig. 2. We weigh the transportation costs by ξ and
the expected loss by 1 − ξ in the objective, with different values of ξ ∈ [0, 1]. When
the transportation costs weigh 0.6 or higher, UAVs are not deployed, and trans-
portation costs and expected loss only slightly change. When risk becomes more
important, UAVs are increasingly deployed. Expected loss can be eliminated by
increasing transportation costs by a factor of five (i.e., from  = 0.6 for the case
without UAVs to  = 0.1 where UAVs fully diminish the risk exposure). The
changes in truck routes are minimal when risk becomes more important, which
can be explained by the indirect loss. Multiple shorter routes that decrease the
expected direct loss causes not only higher transportation costs but also increase
the expected indirect loss due to more time on the road. This allows for fewer
possibilities in reducing risks with only trucks.

Fig. 2. Results for case subset 7 for both trucks only and trucks & UAVs scenarios
with different weights for transportation costs and expected loss.

6 Conclusions

Risks in disaster areas can have a significant impact on the operations of human-
itarian organizations. In this work, we presented a heterogeneous fleet RCVRP
with the consideration of indirect losses (loss irrespective of the amount of load
carried, such as material- or emotional damage) to mitigate these impacts. We
developed an ALNS metaheuristic with specific destroy and repair operators
to solve this problem variant effectively. We tested this method on real-world
instances from high-risk areas in South Sudan.
The results show that both the minimization of expected losses and the use
of UAVs increase transportation costs, but simultaneously decrease the exposure
to risks significantly. This enables safer operations for aid workers of humani-
tarian organizations. Hence, UAVs provide a valuable alternative in high-risk
environments, with an average decrease of the objective value of 7.1% in the
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 289

case of equal consideration of transportation costs and expected loss. UAVs are
especially valuable in situations in which locations are scattered and relatively
far away from humanitarian hubs. These locations cause long travel times and
high exposure to risks for road transport. Hence, UAVs can take over the most
challenging locations. In the largest instance, 16% of the locations were visited
by UAVs.
The inclusion of indirect losses for a more accurate consideration of risks
resulted in less flexibility in mitigating risks with trucks. Only minimal changes in
routes and objective value were found for different weights of the transportation
costs and the expected loss. For the heterogeneous fleet variant, risks could be
eliminated entirely in the most extreme variants by increasing transportation
costs by a factor of five. However, when the focus was mainly on transportation
costs, UAVs were not deployed.
In future work, we aim to test the ALNS framework on benchmark instances
and compare it with exact approaches to evaluate the performance of each
method. Furthermore, we suggest introducing prioritization based on the vulner-
abilities of specific beneficiaries, as the delivery of goods to certain groups might
be more important than to others. This could be done by explicitly considering
deprivation costs that can occur when beneficiaries have to wait when deliveries
cannot be performed due to attacks. Facility location decisions could also be
included, including setup costs for an operating base of UAVs.

References
1. Allahyari, S., Yaghoubi, S., Van Woensel, T.: A novel risk perspective on location-
routing planning: an application in cash transportation. Transp. Res. Part E Logist.
Transp. Rev. 150, 102356 (2021). https://doi.org/10.1016/j.tre.2021.102356
2. Allahyari, S., Yaghoubi, S., Van Woensel, T.: The secure time-dependent vehicle
routing problem with uncertain demands. Comput. Oper. Res. 131, 105253 (2021).
https://doi.org/10.1016/j.cor.2021.105253
3. Besiou, M., Van Wassenhove, L.N.: Humanitarian operations: a world of opportu-
nity for relevant and impactful research. Manuf. Serv. Oper. Manag. 22(1), 135–145
(2020). https://doi.org/10.1287/msom.2019.0799
4. Christiaens, J., Vanden Berghe, G.: Slack induction by string removals for vehicle
routing problems. Transp. Sci. 54(2), 417–433 (2020). https://doi.org/10.1287/
trsc.2019.0914
5. Dueck, G.: New optimization heuristics: the great deluge algorithm and the record-
to-record travel. J. Comput. Phys. 104(1), 86–92 (1993). https://doi.org/10.1006/
jcph.1993.1010
6. Fallahtafti, A., Ardjmand, E., Young Ii, W.A., Weckman, G.R.: A multi-objective
two-echelon location-routing problem for cash logistics: a metaheuristic app-
roach. Appl. Soft Comput. 111, 107685 (2021). https://doi.org/10.1016/j.asoc.
2021.107685
7. Fröhlich, G.E., Gansterer, M., Doerner, K.F.: Safe and secure vehicle routing: a
survey on minimization of risk exposure. Int. Trans. Oper. Res. (2022). https://
doi.org/10.1111/itor.13130
290 R.M. van Steenbergen et al.

8. Ghannadpour, S.F., Zandiyeh, F.: A new game-theoretical multi-objective evolu-

tionary approach for cash-in-transit vehicle routing problem with time windows (a
real life case). Appl. Soft Comput. 93, 106378 (2020). https://doi.org/10.1016/j.
asoc.2020.106378
9. Hagen-Zanker, J., Ulrichs, M., Holmes, R.: What are the effects of cash transfers
for refugees in the context of protracted displacement? findings from Jordan. Int.
Soc. Secur. Rev. 71(2), 57–77 (2018). https://doi.org/10.1111/issr.12166
10. Holm-Nielsen, P.V., Raju, E., Furu, P.: The transformative effect of cash and
voucher assistance experienced by humanitarian organizations. Int. J. Disaster Risk
Reduct. 80, 103238 (2022). https://doi.org/10.1016/j.ijdrr.2022.103238
11. Kian, R., et al.: Logistics planning of cash transfer to Syrian refugees in Turkey.
Eur. J. Oper. Res. 296(3), 1007–1024 (2022). https://doi.org/10.1016/j.ejor.2021.
04.054
12. Laporte, G., Musmanno, R., Vocaturo, F.: An adaptive large neighbourhood search
heuristic for the capacitated arc-routing problem with stochastic demands. Transp.
Sci. 44(1), 125–135 (2010). https://doi.org/10.1287/trsc.1090.0290
13. Li, H., Wang, H., Chen, J., Bai, M.: Two-echelon vehicle routing problem with
time windows and mobile satellites. Transp. Res. Part B Methodol. 138, 179–201
(2020). https://doi.org/10.1016/j.trb.2020.05.010
14. Mancini, S.: A real-life multi depot multi period vehicle routing problem with
a heterogeneous fleet: formulation and adaptive large neighborhood search based
matheuristic. Transp. Res. Part C Emerg. Technol. 70, 100–112 (2016). https://
doi.org/10.1016/j.trc.2015.06.016
15. Mazdarani, F., Ghannadpour, S.F., Zandieh, F.: Bi-objective overlapped links vehi-
cle routing problem for risk minimizing valuables transportation. Comput. Oper.
Res. 106177 (2023). https://doi.org/10.1016/j.cor.2023.106177
16. Radojičić, N., Djenić, A., Marić, M.: Fuzzy grasp with path relinking for the risk-
constrained cash-in-transit vehicle routing problem. Appl. Soft Comput. 72, 486–
497 (2018). https://doi.org/10.1016/j.asoc.2018.05.022
17. Rejeb, A., Rejeb, K., Simske, S., Treiblmaier, H.: Humanitarian drones: a review
and research agenda. Internet Things 16, 100434 (2021). https://doi.org/10.1016/
j.iot.2021.100434
18. Ropke, S., Pisinger, D.: An adaptive large neighborhood search heuristic for the
pickup and delivery problem with time windows. Transp. Sci. 40(4), 455–472
(2006). https://doi.org/10.1287/trsc.1050.0135
19. Sacramento, D., Pisinger, D., Ropke, S.: An adaptive large neighborhood search
metaheuristic for the vehicle routing problem with drones. Transp. Res. Part C
Emerg. Technol. 102, 289–315 (2019). https://doi.org/10.1016/j.trc.2019.02.018
20. Santini, A., Ropke, S., Hvattum, L.M.: A comparison of acceptance criteria for the
adaptive large neighbourhood search metaheuristic. J. Heuristics 24(5), 783–815
(2018). https://doi.org/10.1007/s10732-018-9377-x
21. Soeanu, A., Ray, S., Berger, J., Boukhtouta, A., Debbabi, M.: Multi-depot vehicle
routing problem with risk mitigation: model and solution algorithm. Expert Syst.
Appl. 145, 113099 (2020). https://doi.org/10.1016/j.eswa.2019.113099
22. Talarico, L., Sörensen, K., Springael, J.: Metaheuristics for the risk-constrained
cash-in-transit vehicle routing problem. Eur. J. Oper. Res. 244(2), 457–470 (2015).
https://doi.org/10.1016/j.ejor.2015.01.040
23. Talarico, L., Sörensen, K., Springael, J.: A biobjective decision model to increase
security and reduce travel costs in the cash-in-transit sector. Int. Trans. Oper. Res.
24(1–2), 59–76 (2017). https://doi.org/10.1111/itor.12214
 The Heterogeneous Fleet Risk-Constrained VRP in Humanitarian Logistics 291

24. Talarico, L., Springael, J., Sörensen, K., Talarico, F.: A large neighbourhood meta-
heuristic for the risk-constrained cash-in-transit vehicle routing problem. Comput.
Oper. Res. 78, 547–556 (2017). https://doi.org/10.1016/j.cor.2016.04.003
25. Tikani, H., Setak, M., Demir, E.: A risk-constrained time-dependent cash-in-transit
routing problem in multigraph under uncertainty. Eur. J. Oper. Res. 293(2), 703–
730 (2021). https://doi.org/10.1016/j.ejor.2020.12.020
26. Vincent, F.Y., Jewpanya, P., Redi, A.P., Tsao, Y.C.: Adaptive neighborhood sim-
ulated annealing for the heterogeneous fleet vehicle routing problem with multiple
cross-docks. Comput. Oper. Res. 129, 105205 (2021). https://doi.org/10.1016/j.
cor.2020.105205
 The Snow Grooming Routing Problem

Henrik Andersson1(B) , Sondre Høyland1 , Jesper Anker Krogstad1 ,

Truls Flatberg2 , and Anders N. Gullhav1
1
Department of Industrial Economics and Technology Management,
Norwegian University of Science and Technology,
Alfred Getz veg 3, Trondheim, Norway
[email protected]
2
SINTEF Industry/Sustainable Energy Technology,
P.O. Box 4760, Torgarden, 7465 Trondheim, Norway

Abstract. There are more than 30,000 km of cross-country skiing tracks

in Norway, and to maintain these track networks municipalities and local
ski clubs spend more than 250 million NOK every year. The primary cost
driver is the daily grooming operations which are manually planned based
on the experience of the snowcat operators. Large networks with many
vehicles starting at different depots complicate the problem of finding
effective routes. The result is unnecessary high costs due to suboptimal
route choices, yielding a benefit of solving the route planning problem.
Employees of the municipal enterprise responsible for the cross-
country facilities in Trondheim, Trondheim Bydrift, explain that today’s
planning of grooming activities is based on experience and old habits. As
Trondheim is an area known for unstable weather conditions, long-term
planning lacks robustness. Meetings are therefore conducted every morn-
ing to handle the variations. The multifaceted Snow Grooming Routing
Problem (SGRP) is an arc routing problem with profits that involves
multiple depots, time windows, a heterogeneous fleet of vehicles and track
segments having numerous attributes.
We formally define the SGRP and present a mathematical formu-
lation of the problem. The formulation is then tested on a number of
different networks taken from Bymarka, the largest cross-country area
around Trondheim, Norway.

Keywords: Snow Grooming · Arc routing

1 Introduction

There are 5360 km of cross-country skiing tracks in Norway’s top ten most pop-
ular ski facilities. Municipalities, ski clubs and commercial actors in these areas
spend more than 50 million NOK in total each season to maintain their track
networks [3].
According to Statistics Norway [13], 34266 km of tracks were planned to be
groomed in 2022. Extrapolating the costs from the popular facilities results in a
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 292–306, 2023.
https://doi.org/10.1007/978-3-031-43612-3_18
 The Snow Grooming Routing Problem 293

rough estimate of 300 million NOK spent every year grooming Norwegian track
networks. These networks are complex, consisting of wide and narrow track
segments for the different styles of cross-country skiing. Popular routes must
be prioritized and groomed more often than peripheral segments solely used by
enthusiasts to maintain a certain standard across the whole network. Grooming
requests from local ski clubs, winter sport events and competitions must be
handled and included in the planning of the operations. These events often set
high standards for the snow grooming, requiring the grooming to be done within
specific time windows. Requests vary from being weekly requests known from
the start of the season to urgent short-term requests. The fact that requested
segments are mandatory complicates the planning.
In 2023, there were more than 700 snowcats grooming track networks and
alpine slopes in Norway [12]. Popular areas have fleets of snowcats grooming
simultaneously to meet demand. Statistics Norway show an increase in cabins
and holiday homes in Norway, where an expected consequence is an increase in
demand for cross-country tracks [14]. Current track networks will therefore prob-
ably meet an increase in skier traffic, requiring additional networks or expansion
of the current ones to be practical. This will further increase the complexity of
the grooming operations. New technology with GPS tracking of snowcats and
websites like skisporet.no has made it easier for skiers to find freshly groomed
tracks, but for the facilitators of the networks, helpful tools for planning groom-
ing operations is hard to find.
The problem of planning the grooming operations, the Snow Grooming Rout-
ing Problem (SGRP) is a rich arc routing problem with profits, see for example
[1,6] and [2] for recent surveys and a book on many aspects of arc routing
problems. The SGRP has multiple depots and a heterogeneous fleet. The track
network consists of wide and narrow segments, which are either requested, pri-
oritized, or regular. A wide segment can be groomed by one large snowcat or two
small snowcats. Requested segments have to be served within a given time win-
dow while the others can be groomed without regard to time. The optional seg-
ments are diversified by priority based on popularity for skiers but are groomed
without precedence to ensure an efficient route through the network. Arc rout-
ing problems for snow plowing and road gritting have similar attributes, see for
example [7–11], and [5] but this constitutes the first known attempt of solving
the Snow Grooming Routing Problem.
This paper aims to give a formal description of the Snow Grooming Routing
Problem (SGRP), formulate the problem as a mathematical model, and use the
model to optimize the snow grooming operations on a set of test instances based
on real-life data. First, we give some background to snow grooming and different
aspects of planning snow grooming operations in Sect. 2. The problem description
follows in Sect. 3 before the mathematical formulation in Sect. 4. Results from
testing the model on five instance are presented in Sect. 5. We conclude the paper
with a short discussion in Sect. 6.
294 H. Andersson et al.

2 Background
Networks of ski tracks vary in size and complexity from place to place. Some areas
have huge and intricate networks, while others have only a couple of closed loops
of tracks. The complexity comes from different track widths, the combination of
prioritized and requested track segments, track crossings, bridges, lakes, swamps
and road crossings. Small networks are usually groomed by a single snowcat,
while large networks have a larger fleet and potentially multiple depots. Norway
has hundreds of track networks, and Fig. 1 shows an overview of these in south-
ern Norway. Figure 2 is a more detailed map of the track network in Bymarka,
Trondheim. This is the main network in the Trondheim area and is the basis
for the problem description and the test instances. The network is operated by
Trondheim Bydrift and consists of more than 100 km of tracks. Three snowcats
in two depots are used to groom the tracks.

Fig. 1. Overview of track networks in Fig. 2. The track network in Bymarka,

southern Norway Trondheim

Track networks with stadiums and tracks meant for competitions are often
very dense, which means that they have many intersections and relatively short
distances between them. Competition tracks are mostly loops stretching 3, 5 or
10 km, and should be groomed in a single route so that skiers can use them
at once. The grooming of dense networks is often more challenging to plan
because of the numerous possibilities, while scattered networks with fewer seg-
ments between intersections and longer track segments pose fewer options.
 The Snow Grooming Routing Problem 295

There are many stakeholders with different agendas regarding the snow
grooming operation, the track network and the surrounding nature. The gen-
eral public is the largest stakeholder, visiting popular cabins serving food as
well as the outskirts of the track network. They are to some extent interested
in quantity over quality when it comes to grooming, and that the track net-
works are large enough. The second largest stakeholder is the ski clubs and
high-schools for athletes. They normally have a set of routes that they regularly
use for organized training activities, and are mostly interested in top quality for
these specific routes. They also hand in grooming requests for special events such
as competitions. The municipalities want to facilitate for an active population,
while at the same time preserve the regional nature and minimize local pollution
in the area.
Decisions about which segments of the network to prioritize early in the
season are made in collaboration with the stakeholders. On a daily basis, it is
up to the snowcat operators to determine which tracks to groom and when to
groom them. Planning longer time-spans is difficult and not robust enough, due
to aspects like weather changes and short-term grooming requests.
During periods of consistent whether a weekly schedule for the network can
be followed, but with an occurrence of precipitation or a change in temperature,
the schedule is of little use. A snowfall resets the problem and all previously
groomed segments will have to be groomed again. Since consistent weather is
the exception for most of the season, the operators experience much overtime.
In some municipalities, the grooming of the tracks can be requested in
advance by stakeholders, either as a repetitive request throughout the season
or as a one-time request on an occasional basis. These requested tracks must be
of a certain standard at a given time of day. To meet the quality requirements,
the requested tracks must be groomed within a given time window. The time
window is set based on weather forecasts and projected traffic of skiers in the
network, a rule of thumb is that the grooming is finished an hour before the start
of the event. The total duration of the time window depends on the length of
the requested route. The possibility of requesting the grooming of certain tracks
results in a further complication of the route planning.
There are some segments that are prioritized, typically popular roundtrips
or routes leading to serviced cabins. Other segments are used less, but still they
have to be groomed regularly because of snowfalls. The goal is, therefore, to
ensure top track quality for the most popular tracks, while at the same time
not neglecting the other segments, substantiating the importance of planning.
Track networks with multiple depots are faced with a more complicated planning
problem. Coordination of routes to minimize overlapping and non-groomed track
segments is obtainable, but hard to plan since the route selection of the operators
is often done in a myopic way with incomplete information of the plans of the
other operators.
With different sizes of snowcats and different widths of tracks, comes the
problem of assigning the right snowcat to the right track segment. Large snowcats
are typically too wide to traverse narrow tracks with only classic skiing, meaning
296 H. Andersson et al.

they can only be used on wide tracks (for both classic and skate). The small
snowcats are used for the narrow tracks, but can also be delegated to grooming
wide tracks. For a small snowcat to groom a wide track, it must traverse it twice.

3 Problem Description
The Snow Grooming Routing Problem (SGRP) is an extension of the arc routing
problem with profits. The problem arises in cross-country skiing facilities around
the world when scheduling the grooming of the track networks. The network is
a set of tracks (edges) and intersections (nodes) between them. Each edge has
a given length and topography and is classified according to width and priority.
The width is either wide or narrow. An edge can be without priority, prioritized
or requested. A requested edge must be groomed and there is a time window
specifying when the edge can be groomed. The characteristics of a track segment
can only change at the intersections. In general, the network is non-directional,
meaning that it does not matter in which direction a snowcat traverses an edge.
A network of ski tracks generally contains many intersections and relatively few
track segments between every pair of intersections.
On narrow tracks, a small snowcat can choose to groom two classic sections
or a combination of classic and skate. This is a dynamic setting that the operator
can set with negligible setup time and cost. For wide track segments, a small
snowcat will have to traverse the segment twice to be able to groom the whole
width. The solution is often to acquire a larger snowcat for networks with many
wide segments, which only have to traverse these segments once. These large
snowcats cannot traverse the narrower segments. Every intersection poses the
possibility of a change in track type, complicating the route planning and choice
of snowcat type for the route. Different snowcats have similar attributes, apart
from width and range. They operate at identical speeds under the same weather
conditions. Within each snowcat size, there are no significant differences. There-
fore, all track segments of the same type can be groomed by the same snowcats.
Figure 3 is a snippet of the larger track network in Bymarka, Trondheim (left),
which is presented in the form of a graph (right). Each segment is represented
by an edge in the graph.
When a track segment is groomed, there is no added benefit by grooming
it over again, and it may even be undesirable for the skiing conditions during
periods of consistent weather conditions. The problem aims to determine how
the track network should be groomed. That is, which snowcats should groom
which track segments, and at what time. The purpose is to ensure the effective
use of the available resources.

4 Mathematical Formulations

Before presenting the notation and the mathematical model, we state some
important assumptions:
 The Snow Grooming Routing Problem 297

Fig. 3. Small part of the track network in Bymarka, Trondheim (left) and the corre-
sponding graph (right). Nodes are intersections and edges are segments. The priorities
are shown: dotted (no priority), dashed (prioritized), and solid (requested). Wide seg-
ments are marked with thick lines.

– Costs of grooming operations are not included in the model. Describing the
problem for daily operations, long-term overhead costs like inventory and
wages are considered sunk.
– The problem considers a single day. The weather conditions are known and
consistent during the day.
– The working hours are fixed, imposed as an upper bound with no possibility
of overtime.
We denote the track segments as edges and the intersections as nodes. The
problem is formulated over a graph G = (N , E) where N is the set of nodes and
E the set of edges. We also introduce A as the set of arcs to track directions,
each edge corresponds to two arcs. A subset E W ⊆ E of the edges are wide
edges, another subset E P ⊆ E of the edges are prioritized edges, and a third
subset E R ⊆ E of the edges are requested edges that must be groomed within a
given time window. We also define K to be the set of snowcats, and denote the
set of small and large snowcats as KS and KL , respectively.
We use (ij) to denote the edge connecting node i and j, with i < j. When
the direction the edge is traversed matter, we use ij to denote the arc starting
in node i and ending in node j. The time used to traverse edge (ij) is T(ij) and
its length is L(ij) . We use [T (ij) , T (ij) ] to denote the time window defined for
the requested edge (ij). The time available for snowcat k is TkK and the node
where it must start and end its route is denote i(k). We introduce a reward
R(ij) for grooming edge (ij) and a cost C per length unit groomed to penalize
unnecessary grooming.
298 H. Andersson et al.

We use a leg-based formulation and introduce the index set L = {0, . . . , L+1}
for the set of legs where L is an upper bound on the number of arcs in the route
of a snowcat. The binary variable xkijl is 1 if snowcat k grooms arc ij on leg
l, and 0 otherwise. The binary variable v(ij) is 1 if edge (ij) is groomed, and 0
otherwise, and wk(ij)l is 1 if edge (ij) ∈ E R is groomed by snowcat k on leg l.
The time variable tkl is the time snowcat k starts leg l and τk(ij)l is the time
when edge (ij) ∈ E R is groomed by snowcat k on leg l.
With this notation, the SGRP can be formulated as follows:
   
max z = R(ij) v(ij) − C L(ij) (xkijl + xkijl ) (1)
(ij)∈E k∈K (ij)∈E l∈L

The objective function (1) maximizes the reward from groomed edges and penal-
izes the traversal of an edge more than once.

xk0i(k)0 = 1 k∈K (2)

 
xkijl = xkji,l+1 k ∈ K, j ∈ N \ {0}, l ∈ L | l ≤ L (3)
ij∈A ji∈A

xki(k)0l = 1 k∈K (4)
l∈L

Constraints (2) and (4) state that snowcat k must start and end its route in
its depot i(k). Node 0 is an artificial starting node with T(0i(k)) = L(0i(k)) = 0.
Constraints (3) ensure the flow of snowcats through the nodes. They state that
if snowcat k enters node j, it must exit the same node on its next leg.
 
v(ij) ≤ (xkijl + xkjil ) (ij) ∈ E \ E W (5)
k∈KS l∈L
  1  
v(ij) ≤ (xkijl + xkjil ) + (xkijl + xkjil ) (ij) ∈ E W (6)
2
k∈KL l∈L S
k∈K l∈L

Constraints (5) and (6) keep track of if an edge (ij) is groomed or not. The right
side of the constraints contains both arc ij and arc ji, so the model becomes non-
directional. Constraints (6) also state that for wide edges, a small snowcat must
groom the edge twice, while large snowcats only need one traversal. Constraints
(5) ensure that only small snowcats can groom narrow edges.

wk(ij)l − xkijl − xkjil ≤ 0 k ∈ K, (ij) ∈ E R , l ∈ L (7)

   
R W
2 wk(ij)l + wk(ij)l = 2 (ij) ∈ E ∩ E (8)
k∈KL l∈L k∈KS l∈L
   
wk(ij)l = 1 (ij) ∈ E R ∩ E \ E W (9)
k∈KS l∈L

Constraints (7) link the x-variable with the w-variable, so that a requested edge
is considered groomed independent of which direction it is traversed. Constraints
 The Snow Grooming Routing Problem 299

(8) state that the prioritized track segments must be groomed, either once by
a large snowcat or twice by a small snowcat. Constraints (9) state that narrow
requested edges must be groomed by a small snowcat.

tk,l+1 ≥ tkl + T(ij) (xkijl + xkijl ) k ∈ K, l ∈ L (10)
ij∈E

Constraints (10) update the starting time for snowcat k on leg l+1 by connecting
it with the starting time on leg l.
τk(ij)l ≥ tkl − M (1 − wk(ij)l ) k ∈ K, (ij) ∈ E R , l ∈ L (11)
R
τk(ij)l ≤ tkl + M (1 − wk(ij)l ) k ∈ K, (ij) ∈ E , l ∈ L (12)
R
T (ij) ≤ τk(ij)l ≤ T (ij) k ∈ K, (ij) ∈ E , l ∈ L (13)
Constraints (11) and (12) link the τ -variable and the t-variable. Constraints (13)
make sure that the requested edges are groomed within their time window.
xkijl ∈ {0, 1} k ∈ K, ij ∈ A, l ∈ L (14)
R
wk(ij)l ∈ {0, 1} k ∈ K, (ij) ∈ E , l ∈ L (15)
v(ij) ∈ {0, 1} (ij) ∈ E (16)
0 ≤ tkl ≤ TkK k ∈ K, l ∈ L (17)
R
τk(ij)l ≥ 0 k ∈ K, (ij) ∈ E , l ∈ L (18)
Constraints (14)–(16) are binary restrictions, while constraints (17) set the
bound on the time used. Constraints (18) ensure non-negativity. Variables xkijl
and wk(ij)l for large snowcats are not generated for narrow edges.

Reducing the Number of Variables Generated

We can reduce the number of variables generated for each snowcat. This strength-
ens the formulation and makes it easier to solve. For each snowcat k, we solve
a shortest path problem from i(k) to all other nodes where the length of each
edge is 1. This gives the least number of edges snowcat k must traverse to reach
each node. Let Dki denote the minimum number of edges snowcat k must tra-
verse to reach node i. This means that no xkjil with l < Dki and no xkijl with
l ≤ L − Dki are generated. If neither xkjil nor xkijl are generated, then wk(ij)l
is not generated either.
Similarly, if we solve the shortest path problem with the traversal time as
the length of each segment, we can calculate the earliest arrival Tki in node i
for snowcat k. If Tki > T (ij) , then neither τk(ijl) nor wk(ij)l are generated. If
Tki + Tij + Tkj > TkK , then xkijl is not generated.

5 Results
We have collected data for Bymarka, Trondheim and created 5 instances. Rel-
evant statistics for each instance are presented in Table 1. The table shows the
300 H. Andersson et al.

name of the instance, the total length of the network in km, the number of
nodes, the number of edges, the number of wide edges, the number of priori-
tized edges, the number of requested edges and the number of small and large
snowcats respectively. The solution time is set to 7200 s for all instances. The
optimization model is implemented in the programming language Mosel [4] and
the models are solved using Xpress-Optimizer Version 9.0. All software is run on
Lenovo ThinkSystem SD530 computers with the following specifications.

CPU: 2x 3.6 GHz Intel Xeon Gold 6244 CPU - 8 core

RAM: 384 Gb
Disk: 250 Gb SATA SSD

Table 1. Statistics over the test instances

Name Length |N | |E| |E W | |E P | |E R | |KS | |KL |

Leinstrand 20.8 20 28 17 5 12 1 0
SkistuaV 15.5 24 32 17 11 0 1 0
SkistuaO 22.1 44 64 47 4 28 1 0
Nilsbyen 22.8 40 55 21 15 19 1 1
N&L 44.1 57 82 37 18 31 1 1

Since each network is connected and there are positive revenues for all edges,
we can assume that each network will be completely groomed given enough time.
Our first test is therefore to run the model on each network without considering
the time windows and the available time for different values on the number of legs
L. A lower bound on the number of legs needed can be calculated as |E| + |E W |
for the instances with one small snowcat. The results are presented in Table 2.
The table shows the name of the instance, the maximum number of legs for each
snowcat, the objective value of the best feasible solution, the dual bound, the
number of edges groomed, and the number of edges traversed.
A first observation when studying the results in Table 2 is that the problem
is very complex and most instances are not solved to proven optimality. Still,
the gaps are low in all cases and we believe that valid concussions can be drawn
based on the results.
If we analyze Leinstrand, the lower bound is 42 and we see that L = 40 is
clearly not enough to groom all edges. Even though L = 50 is higher than the
lower bound, we are still not able to groom all edges. The solutions for L = 60
and L = 70 show that 53 edges (excluding the artificial edges to and from node
0) must be traversed to groom the full network.
The results are the same for the other cases with one small snowcat. The
lower bound on SkistuaV is 45 and 57 edges are needed to completely groom
 The Snow Grooming Routing Problem 301

Table 2. Results for different values on L

 
Name L z z v x
Leinstrand 40 32570 33330 20 40
Leinstrand 50 36160 36644.86 26 49
Leinstrand 60 36410 36550 28 53
Leinstrand 70 36410 36604 28 53
SkistuaV 40 20720 20720 20 40
SkistuaV 50 22850 23120 28 50
SkistuaV 60 23090 23392.9 32 58
SkistuaV 70 23090 23375 32 57
SkistuaO 110 19510 21500 55 102
SkistuaO 120 20680 21540 60 116
SkistuaO 130 20630 21540 60 124
SkistuaO 140 21120 21540 64 127
Nilsbyen 40 58480 58560 47 66
Nilsbyen 50 59820 60132.63 55 75
Nilsbyen 60 59810 60153.57 55 77
Nilsbyen 70 59780 60126.17 55 80
N&L 70 87441.7 90408.84 75 113
N&L 80 87730 90418.64 76 115
N&L 90 88620 90451.2 80 126
N&L 100 88290 90501.09 79 120

the network and the lower bound for SkistuaO is 111 and the number of edges
needed is 127. On the instances with more than one snowcat, the lower bound
on the number of edges is |E|, the small snowcat grooms all narrow edges and
the large snowcat all wide edges, but how to set L based on this is an open
question. The results in Table 2 show that L = 50 is enough for Nilsbyen. For
N&L, L = 90 gives the best results.
Since restricting L causes changes to the solution that do not have to do
with the planning problem in itself, we continue the computational study with
L large enough to not affect the optimal solution of the problem. Still, since the
number of variables and constraints depends on L, keeping it low is preferable.
Next, we analyse the effect of the available time and time windows. We test
this on the Leinstrand instance where the snowcat spends 164.8 min to groom the
complete network. We set the available time to 120 min and test without time
windows. In the solution from Table 2 we note that about half of the requested
edges are groom in the beginning of the route and the rest at the end of the
route. We therefore also test with three different time windows; [0, 60], [30, 90],
and [60, 120]. The results of these tests are presented in Table 3. The table shows
the name of the instance, the maximum number of legs for the snowcat, the
302 H. Andersson et al.

Fig. 4. Two examples of time constrained solutions for the Leinstrand instance. With-
out time windows (left) and with time windows [30, 90] (right). The edges that are not
groomed are shown in orange. The priorities are shown: dotted (no priority), dashed
(prioritized), and solid (requested). Wide segments are marked with thick lines. (Color
figure online)

available time for the snowcat, the time windows [T (ij) , T (ij) ], which are the
same for all requested edges, the objective value of the best feasible solution, the
dual bound, the number of edges groomed, the number of edges traversed, and
the time the snowcat uses to groom.
It is clear that the number of edges that are groomed decreases. In Fig. 4, the
edges that are not groomed are marked in orange. The broad, prioritized edges
are all groomed. The difference between the solutions without time windows and
with early or late time windows is small. In the solution without time windows,
almost all requested arcs are groomed in the first hour, and due to symmetry
it is likely that there is an alternative solution where the remaining edges can
also be groomed within the first hour. Since there are only edges, a route can
be traversed in opposite direction without changing the cost. This means that
a feasible route for the case with time windows [0, 60] can become a feasible
route for the case with time windows [60, 120] by traversing it in the opposite
direction. The solution for the time windows [30, 90] is worse. The snowcat cannot
start grooming the requested edges and must return to do this later. This gives a
situation where the farthest edges cannot be groomed. Not only is the number of
edges lower, but the length of the edges not groomed is also longer. In the time
restricted case without time windows, 8.2 km of the network is not groomed,
while the corresponding number of the case with the time windows [30, 90] is
10.1 km.
In the last analysis, we study the effect of coordination on the two snowcats
instances Nilsbyen and N&L. The variable restrictions (17) set an upper bound
 The Snow Grooming Routing Problem 303

Table 3. Results with restricted time and time windows

 
Name L TkK TW z z v x t1,L+1
Leinstrand 60 ∞ – 36410 36550 28 53 164.8
Leinstrand 60 120 [0,120] 29110 30510 24 44 118.9
Leinstrand 60 120 [0,60] 28080 30400.6 23 48 115.8
Leinstrand 60 120 [30,90] 27240 29327.65 19 44 117.4
Leinstrand 60 120 [60,120] 29100 29990 24 48 119.8

TkK on the available time for snowcat k. This models the case when one driver
per snowcat grooms the track. Another interesting situation is when there is one
driver that operates both snowcats and has an upper bound T K on the total
available time. To model this, we introduce the constraint

tk,L+1 ≤ T K (19)
k∈K

To test coordination together with time windows, we introduce the binary vari-
able o, which is 1 if the small snowcat is operated first and 0 otherwise. The
sequencing of the snowcats is modeled by the following constraints

tkS ,L+1 − tkL ,1 ≤ TkKS (1 − o) (20)

tkL ,L+1 − tkS ,1 ≤ TkKL o (21)
o ∈ {0, 1} (22)

where k S and k L are the indices for the small and large snowcat respectively. If
the fleet of snowcats is larger, a more complex formulation of the sequencing is
needed. Here, we limit the time using the variable bounds

tk,L+1 ≤ T K k∈K (23)

We run Nilsbyen and N&L with less available time than the time used in
the solution from Table 2. We then take the same total available time, but let
the model decide on how to divide the time between the snowcats. We also test
with even less available time to see if coordination can compensate for less time.
The results from these tests are given in Table 4. The table shows the name of
the instance, the maximum number of legs for the snowcat, the available time
for each snowcat when there is no coordination, the total available time when
there is coordination, the time windows [T (ij) , T (ij) ], which are the same for all
requested edges, the objective value of the best feasible solution, the dual bound,
the number of edges groomed, the number of edges traversed, and the time spent
by the small and large snowcat respectively.
304 H. Andersson et al.

Analyzing Nilsbyen, we see that the

reduction to 60 min per snowcat does not
affect the solution much. We also see that
even though the number of groomed edges is
less when the total time can distributed, the
objective function value is higher. Comparing
the 60 + 60 case and the 120 case, i.e. rows 2
and 3 in Table 4, we see that in the 60 + 60
case, 3.5 km of wide edges and 1.4 km of nar-
row edges are not groomed, while the cor-
responding numbers for the 120 case is 2 km
and 2.1 km respectively. In both cases, all pri-
oritized edges are groomed.
When the available time is restricted even
more, we get to a point when it is no longer Fig. 5. The Nilsbyen instance
possible to groom the requested edges and
the problem becomes infeasible without coor-
dination. Table 4 shows that 30 min is not enough for the small snowcat to groom
the requested narrow edges, while when 60 min in total is available a feasible solu-
tion is found. An interesting observation is that the large snowcat is not very
much used when the time is very restricted, even though a wide prioritized edge
can be groomed in one traversal. An explanation for this is the structure of the
network together with the time restriction. The prioritized route shown in Fig. 5
takes too long to traverse and the large snowcat can only traverse a short part of

Table 4. Results with and without coordination

 
Name L TkK T K TW z z v x tkS ,L+1 tkL ,L+1
Nilsbyen 60 1200 – – 59810 60153.57 55 77 80.5 76.7
Nilsbyen 60 60 – – 55840 58524.43 51 68 55.9 53.9
Nilsbyen 60 – 120 56530 58330.12 48 63 48.1 66.7
Nilsbyen 60 – 120 [0,60] 56270 58297.65 48 63 46.1 66.7
Nilsbyen 60 – 120 [60,120] 56560 58526.26 49 64 49.3 69.1
Nilsbyen 60 45 – – 14860 49110.03 35 58 44.8 23
Nilsbyen 60 – 90 – 26950 51431.11 37 56 88.1 0
Nilsbyen 60 30 – – – – – – – –
Nilsbyen 60 – 60 – 11150 26317.42 39 53 51.6 6.2
N&L 90 1200 – – 88620 90451.2 80 126 188.9 147.7
N&L 90 120 – – 76830 89030.64 69 98 84.7 113.6
N&L 90 – 240 – 78050 89444.73 70 100 70.3 147.2
N&L 90 90 – – 43420 81561.08 57 76 58.3 78.3
N&L 90 – 180 – 73690 81552.46 63 84 53.3 110.6
 The Snow Grooming Routing Problem 305

it before it has to return to the depot. When going back and forth on the same
edges, the advantage of the large snowcat is lost.
The results are similar for the N&L instance, the case with a total available
time is better than the case with individual available times. This is especially
true for the 90 + 90 case compared with the 180 case. Figure 6 illustrates this
case. When the total time can be distributed, more time can be given to the
large snowcat so that it can groom all prioritized edges.

Fig. 6. A comparison between solutions without and with coordination for the N&L
instance. In the left figure, each snowcat has 90 min available, in the right figure the
total available time is 180 min. The route of the small snowcat is marked in orange and
the route of the large snowcat is marked in green. The priorities are shown: dotted (no
priority), dashed (prioritized), and solid (requested). Wide segments are marked with
thick lines. (Color figure online)

6 Conclusion
We have introduced a new problem to the arc routing society, the Snow Groom-
ing Routing Problem (SGRP). The SGRP is a rich arc routing problem with
profits including multiple depots, heterogeneous vehicles, time windows, and
edges that can be serviced in different ways. A mixed-integer linear formulation
of the problem is given and tested on five different instances based on real-
life data. The computational study highlights interesting findings regarding the
effect of restricted time and time windows, and coordination. The problem is
complex and hard to solve and developing both efficient heuristics and other
exact solution methods are interesting avenues for future research. The SGRP
studied here is a static version of the problem, extending it to a dynamic prob-
lem where both the initial condition of the tracks and the weather are included,
which is an interesting topic for future research.
306 H. Andersson et al.

References
1. Corberán, A., Eglese, R., Hasle, G., Plana, I., Sanchis, J.M.: Arc routing problems:
a review of the past, present, and future. Networks 77(1), 88–115 (2021)
2. Corberán, A., Laporte, G.: Arc Routing. Society for Industrial and Applied Math-
ematics, Philadelphia (2015)
3. Dagens Næringsliv: Norges beste langrennsløyper (2013). https://www.dn.no/
trening/norges-beste-langrennsomrader/1-1-5010947. Accessed 13 Apr 2023
4. Fair Isaac Corporation: FICOXpress
R Optimization Help (2023). https://www.
fico.com/fico-xpress-optimization/docs/latest/overview.html. Accessed 02 Feb
2023
5. Gundersen, A.H., Johansen, M., Kjær, B.S., Andersson, H., Stålhane, M.: Arc
routing with precedence constraints: an application to snow plowing operations.
In: ICCL 2017. LNCS, vol. 10572, pp. 174–188. Springer, Cham (2017). https://
doi.org/10.1007/978-3-319-68496-3 12
6. Mourão, M.C., Pinto, L.S.: An updated annotated bibliography on arc routing
problems. Networks 70(3), 144–194 (2017)
7. Perrier, N., Langevin, A., Amaya, C.A.: Vehicle routing for urban snow plowing
operations. Transp. Sci. 42(1), 44–56 (2008)
8. Perrier, N., Langevin, A., Campbell, J.F.: A survey of models and algorithms for
winter road maintenance. Part I: system design for spreading and plowing. Comput.
Oper. Res. 33(1), 209–238 (2006)
9. Perrier, N., Langevin, A., Campbell, J.F.: A survey of models and algorithms for
winter road maintenance. Part II: system design for snow disposal. Comput. Oper.
Res. 33(1), 239–262 (2006)
10. Perrier, N., Langevin, A., Campbell, J.F.: A survey of models and algorithms for
winter road maintenance. Part III: vehicle routing and depot location for spreading.
Comput. Oper. Res. 34(1), 211–257 (2007)
11. Perrier, N., Langevin, A., Campbell, J.F.: A survey of models and algorithms for
winter road maintenance. Part IV: vehicle routing and fleet sizing for plowing and
snow disposal. Comput. Oper. Res. 34(1), 258–294 (2007)
12. Skisporet: Skisporet - om oss (2023). https://www.skisporet.no/about. Accessed
13 Apr 2023
13. Statistisk Sentralbyrå: Tabell 12593: Tilrettelegging for friluftsliv (2022). https://
www.ssb.no/statbank/table/12593/. Accessed 13 Apr 2023
14. Statistisk Sentralbyrå: Bygningsmassen (2023). https://www.ssb.no/
bygningsmasse. Accessed 13 Apr 2023
 A Constraint Programming Model
for the Vehicle Routing Problem
with Multiple Time Windows

Florian Linß(B)

Faculty of Business and Economics, TUD Dresden University of Technology,

01069 Dresden, Germany
[email protected]

Abstract. In this paper, we address the vehicle routing problem with

multiple time windows (VRPMTW), which extends the classic vehicle
routing problem to allow customers to be visited only within one or
more time windows. We propose a constraint programming (CP) model
for the VRPMTW and apply model-strengthening techniques such as
fixing variables and adding interval variables to improve the model’s
performance. The model’s effectiveness is demonstrated by comparing its
performance against an existing mixed-integer programming model. The
study also analyzes the differences in solver performance and evaluates
the impact of instance characteristics on the solvability of the problem
instances. Our study demonstrates the benefits of using a CP model
with model-strengthening techniques for solving medium-sized instances
of the VRPMTW.

Keywords: vehicle routing · multiple time windows · constraint

programming

1 Introduction

As e-commerce continues to grow and home parcel delivery volumes increase,

logistics service providers place great importance on flexibility in delivery. One
way to improve flexibility is by offering delivery in multiple time windows. From a
decision support perspective, this can be described as the vehicle routing problem
with multiple time windows (VRPMTW)—an extension of the vehicle routing
problem with time windows (VRPTW). The VRPMTW aims to determine the
minimum cost vehicle routes that serve each customer in one of the specified
time windows while meeting vehicle capacity constraints.
The first mathematical formulation for the VRPMTW was introduced by
[6]. Since then, heuristics have been widely used as solution methods. Among
these heuristics, [2] proposed a variable neighborhood search (VNS) combined
with tabu search, using new benchmark instances derived from the well-known

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 307–321, 2023.
https://doi.org/10.1007/978-3-031-43612-3_19
308 F. Linß

Solomon instances [14]. These benchmark instances have since served as a stan-
dard for evaluating future work for the VRPMTW. Other state-of-the-art heuris-
tics include the adaptive large neighborhood search (ALNS) proposed by [8], the
genetic VNS introduced by [3], the adaptive variable neighborhood search sug-
gested by [5], and the ALNS presented by [13]. More recently, [15] extended
the VRPMTW to the vehicle routing problem (VRP) with availability profiles,
where successful delivery to customers depends on the probability of delivery
within specific time windows. The proposed ALNS heuristic was also evaluated
for the VRPMTW and produced new best-known solutions.
Regarding exact methods, different mixed integer programming (MIP) for-
mulations exist in the related literature. Still, these are mainly problem defini-
tions and have rarely been evaluated as exact approaches for the VRPMTW. [2]
introduced a three-index (vehicle flow) formulation where each vehicle’s flow is
formulated separately. However, since the vehicles are homogeneous, differentia-
tion between vehicles is not necessary. Thus, the formulation can be strengthened
to a two-index formulation introduced by [13], where the flow is formulated aggre-
gated for all vehicles. The authors implemented both MIP formulations using a
standard solver and showed that the latter formulation performed significantly
better. Nevertheless, there were already instances with only 15 customers that
could not be solved optimally. In addition to the comprehensive MIP formula-
tions, [4] presented a column generation approach and a post-optimization VNS
heuristic. For this approach, the pricing problem - which involves the determi-
nation of an elementary shortest path with multiple time windows and capacity
constraints - is solved using a standard solver. The authors were able to optimally
solve instances with up to 17 customers. The rarity of exact approaches for the
VRPMTW contrasts with the adoption of exact approaches for the VRPTW.
As pointed out by [1], many exact methods are available for the latter. In recent
research, [10] suggested that constraint programming (CP) is a viable alterna-
tive to more traditional MIP models and solvers in scheduling problems. This
finding is relevant to the VRP context, as considering time windows and service
times shares similarities to scheduling. [11] previously demonstrated the suit-
ability, who applied CP to solve the traveling salesman problem with multiple
time windows. Additionally, [12] employed a column generation approach for the
VRPTW and integrated CP for solving the pricing problem.
This paper is the first approach that applies a CP formulation in the context
of the VRPMTW. In addition, we show that fixing variables and additional
interval variables improve the performance of the CP model. The CP model
outperforms a comparable MIP formulation regarding the number of instances
optimally solved, the time required to prove optimality, and remaining gap after
a given runtime. We achieve proven optimality for almost all instances with 20
customers and for more than half of the instances with 25 customers. The work is
structured as follows. In Sect. 2, the CP formulation is presented. Computation
tests are discussed in Sect. 3. Finally, a conclusion is presented in Sect. 4.
 A CP Model for VRP with Multiple Time Windows 309

2 Constraint Programming Formulation

for the VRPMTW
2.1 Base Model
The VRPMTW involves the delivery to a set of customers using multiple homo-
geneous vehicles from a single depot. Each customer has a set of multiple time
windows, and the service must begin within one of these time windows. To
ensure that the routes are feasible, the capacity of the vehicles and the maxi-
mum length of the routes must not be exceeded. Similar to [8], we make two
simplifying assumptions in the formulation based on the data provided in the
benchmark instances. Firstly, the depot is assigned only a single time window,
representing the planning horizon. Secondly, the maximum route length is mod-
eled so each vehicle must return to the depot within this time window. The CP
formulation is specifically designed for modeling with the SAT solver provided
by Google OR-tools, but it can be easily adapted to other CP solvers1 .
Table 1 summarizes the necessary parameter and set definitions. We define
G = (V, A) as a digraph with a set of vertices V and a set of arcs A. The
vertex set V consists of a depot vertex 0 and n customer vertices (denoted by
set C). Each vertex i ∈ V has an associated demand μi , a service time σi , and θi
disjoint time windows. It follows that each vertex i ∈ V has a respective set of
time windows Ti = {{αi1 , βi1 }, . . . , {αiθi , βiθi }} in which the service can start. As
described above, the depot has a single time window, i.e., θ0 = 1, and the end
of this time window equals the planning horizon T = β01 .

Table 1. Parameter and set definitions for the CP model

Parameter Description Domain

σi service time at customer i ∈ C integer
μi demand at customer i ∈ C integer
θi number of time windows at vertex i ∈ V integer
αip , βip start and end of time window p at i ∈ V integer
τij , γij time and cost related to the arc (i, j) ∈ A integer
κ vehicle freight capacity integer
γf costs per vehicle used integer
γw costs per unit waiting time integer
Set
V set of vertices {0, 1, . . . , n}
A set of arcs {(i, j)|i, j ∈ V}
C set of customers {1, 2, . . . , n}
K set of vehicles {1, 2, . . . , K}
Ti set of time windows of vertex i ∈ V {{αi1 , βi1 }, . . . , {αiθi , βiθi }}

1
https://github.com/google/or-tools.
310 F. Linß

The arc set is defined as A = {(i, j)|i = j, i, j ∈ V}. Each arc (i, j) ∈ A
is associated with a non-negative travel time τij and travel costs γij . All time
windows ending before the earliest possible start of service and starting after the
latest possible end of service can be removed from Ti of customer i. The earliest
possible start of service is defined as α01 + τ0i and the latest possible end can be
calculated by T − σi − τi0 . We modify the new first and last time windows as
follows: αi1 = max{αi1 , τ0i } and βiθi = min{βiθi , T − σi − τi0 }.
The vehicle set K denotes a homogeneous fleet of K vehicles with a freight
capacity of κ and fixed vehicle costs γ f . Following the definition in [13], the
VRPMTW is considered in two variants, which differ in their objective function.
The first variant aims to minimize vehicle and travel costs. In contrast, the
second variant additionally considers waiting times in case a vehicle arrives at
a customer when there is no time window open. The costs for waiting time
are described by γ w per time unit. To distinguish between both variants, the
parameter B is used to differentiate between them: B = 1 indicates waiting
times are considered, and B = 0 means they are not.

Table 2. Decision variables for the CP model

Variable Description Domain

xij true, if arc (i, j) is used, 0 otherwise boolean
yik true, if customer i ∈ C is served by vehicle k ∈ K boolean
vi vehicle index for customer i ∈ C {1, 2, . . . , K}
r number of vehicles used {1, 2, . . . , K}
si start time of service at vertex i ∈ V Ti
wi waiting time at customer i ∈ C {0, 1, . . . , T }

Table 2 summarizes the definition of the decision variables. The vehicle flow
in the graph is described by the boolean variable xij indicating whether an arc
(i, j) ∈ A is used or not. Additionally, the boolean variable yik defines the assign-
ment of customer i to a vehicle k. For convenience and efficient determination
of the number of vehicles, the variable vi defines the vehicle index for each cus-
tomer i, and the variable r denotes the number of vehicles used. The variable
si describes the start time of a service at a vertex i. The time windows for
each vertex are expressed by specifying multiple disjoint domains in the variable
definition. With this notation, we state a CP model for the VRPMTW:

 
min γ f r + γij xij + B γ w wi (1)
(i,j)∈A j∈C

s.t. MultipleCircuit (A) (2)

ExactlyOne ({yik |k ∈ K}) ∀i ∈ C (3)
max(vi ) = r (4)
i∈C
 A CP Model for VRP with Multiple Time Windows 311

x0j = r (5)
j∈C

μi · yik ≤ κ k∈K (6)
i∈C
yik ⇒ vi = k ∀i ∈ C, k ∈ K (7)
xij ⇒ vi = vj ∀i, j ∈ C : i = j (8)
xij ⇒ sj = si + σi + τij + wj (i, j) ∈ A : i, j = 0 (9)

The objective function (1) minimizes total costs consisting of total fixed vehi-
cle costs, total travel costs, and total waiting costs in case of B = 1. The vehicle
flow is modeled in constraint (2) using the global function MultipleCircuit
implemented in Google OR-tools. The constraint ensures that every customer
vertex has exactly one incoming and one outgoing arc. Additionally, vertex 0
has the same number of incoming and outgoing arcs, and cycles not including
the depot are prohibited. The global functions ExactlyOne in constraints (3)
ensure that exactly one variable of a set of boolean variables must be true. This
relation ensures that only one vehicle is assigned to each customer.
Constraint (4) connects the number of vehicles used with the maximum vehi-
cle index. Constraint (5) guarantees that the number of vehicles used corresponds
to the sum of outgoing arcs from the depot. Constraints (6) ensure that each
vehicle’s capacity is respected. Constraints (7) set vehicle index variables vi
based on binary variables yik . Constraints (8) enforce that the head and tail of a
traveled arc connecting customers have the same vehicle index. The scheduling
of services is respected in constraints (9). If an arc xij is used, the service at its
tail j must start at the arrival time si + σi + τij plus a potential waiting time
wj .

2.2 Model Strengthening

We can strengthen the formulation of the base model by additionally fixing

variables and adding optional interval variables.
Fixing of Variables

x0i ∧ x0j ⇒ vi = vj ∀i, j ∈ C : i = j (10)

¬yik ⇒ vi = k ∀i ∈ C, k ∈ K (11)
v1 = 1 (12)

Constraints (10) defines that if two customers are assigned to two different
routes, i.e., x0i ∧ x0j , their vehicle indices must differ, too. Similarly, we specify
the vehicle index of a customer cannot be k if the customer is not assigned
to vehicle k, see constraints (11). Since all vehicles are homogeneous, the first
customer is assigned to the first vehicle by the symmetry breaking condition (12).
312 F. Linß

Optional Interval Variables. Optional interval variables are increasingly

common in CP-based approaches [7,9]. These variables are introduced to deter-
mine whether a set of customers can be served by one vehicle. In addi-
tion to considering the service time as a fixed component, the travel time
to the next vertex can be added to the interval variables as a variable
component. Thus, the schedule can be modeled more accurately. Since the
length of the interval depends on the next vertex, each interval variable,
denoted as IntervalVar([Start,Length, End)), requires three variables to
model its start, length, and end. Interval variables are defined as always
present. In contrast, the presence of optional interval variables, denoted as
optIntervalVar(Boolean, [Start,Length,End)) is indicated by a boolean
variable.

Table 3. Additional decision variables for defining variable service intervals

Variable Description Domain

s0k start time of vehicle k ∈ K at the depot T0
t0k travel time of vehicle k ∈ K to the first customer {min(τ0j ), max(τ0j )}
j∈C j∈C
a0k arrival time of vehicle k at the first customer T0
ai arrival time at the next vertex after customer i T0
ti travel time from customer i ∈ C to the next vertex {min(τij ), max(τij )}
j∈V j∈V
 
S0k IntervalVar s0k , t0k , a0k for k ∈ K
Sik optIntervalVar (yik , [si , σi + ti , ai )) for k ∈ K, i ∈ C

Table 3 summarizes the definition of the additional variables. Since each vehi-
cle starts at the depot, separate interval variables are defined for each vehi-
cle. More precisely, the integer variable s0k is introduced to define the start
time of a vehicle k at the depot. With integer variables t0k and a0k denoting
the travel time arrival time at the first customer, S0k is defined as
 to and 
IntervalVar s0k , t0k , a0k . Since each customer is visited by one vehicle, exactly
one interval variable is present per customer. Therefore, optional interval vari-
ables are introduced for each vehicle and customer. Their presence depends
on the assignment of customers to vehicles. The variables ti and ai are intro-
duced as travel to and arrival time at the next vertex for each customer i.
Depending on whether a customer is assigned to a vehicle, the interval variable
optIntervalVar (yik , [si , σi + ti , ai )) is only present for vehicle k if yik is true.
In Fig. 1, the definitions above are illustrated by the interval variables of an
example route, see Fig. 1b. The gantt chart in Fig. 1a shows each vertex’s service
time and travel time to the next vertex. Since the start times are only defined
within the time windows, only an overlap of the intervals has to be prevented
for a feasible schedule. This can be enforced efficiently with global constraints
like NoOverlap. The corresponding constraints are listed below.
 A CP Model for VRP with Multiple Time Windows 313

Fig. 1. The schedule of an example route illustrates the definition of interval variables.

xij ⇒ ti = τij (i, j) ∈ A : i = 0 (13)

xij ⇒ ai ≤ sj (i, j) ∈ A : i, j = 0 (14)
x0j ∧ yjk ⇒ t0k = τ0j j ∈ C, k ∈ K (15)
x0j ∧ yjk ⇒ a0k
= sj j ∈ C, k ∈ K (16)
NoOverlap ({Sik |i ∈ V}) k∈K (17)

The length and the end of the optional interval variables of the customers
are defined with the constraints (13) and (14). If j is visited after i, ti is equal
to the travel time τij and the service start sj at j must be after ai . Similarly,
the length and the end of the interval variable at the depot are described by
constraints (15) and (16). If j is the first customer visited by vehicle k, t0k must
be equal to the travel time τ0j . Since each vehicle can start every time within
the planning horizon, the first customer’s service start sj equals a0k , and no
waiting times occur. Finally, overlapping of the intervals can be prevented for
each vehicle with the global constraint NoOverlap, see constraint (17).
314 F. Linß

3 Computational Tests
3.1 Experimental Design

We apply the MIP formulation introduced in [13] to validate and assess the CP
model. We use the CP-SAT solver of Google OR-Tools (9.4) to solve the CP
model and Gurobi (10.0.0) to solve the MIP formulation. Both approaches are
implemented in C++ using MSVC v143. All tests were run with a limit of 8
parallel threads on an AMD EPYC 7513 with 2.6 GHz clock speed and 64 GB
RAM. Each instance is run three times with different seed values to account for
the variability in solver performance. Apart from that, both solvers were run
with their default parameters. The runtime is limited to 3,600 s for each run.
We use the small to medium-sized instances provided by [13] in our experi-
ments. These instances were derived from the instances presented in [2]. Specif-
ically, the first 10, 15, 20, 25, and 50 customers from the instances cm101,
cm105, cm201, cm205, rm101, rm105, rm201, rm205, rcm101, rcm105, rcm201,
and rcm205 were selected to create a total of 60 instances. The instances are
classified according to three metrics for the customers’ location: clustered (cm),
random (rm), and random clustered (rcm). Additionally, the first digit of the
instance name indicates whether the vertices have short (1) or long (2) time
windows. Since the CP model is only defined for integer domains, all parameter
values were scaled with 104 and rounded.

3.2 Results

First, we investigate the impact of the model-strengthening methods and com-

pare the formulations. Next, we examine different lower bounds for the number
of required vehicles, which shed light on the varying performance of the solvers.
After that, we evaluate how solvability is affected by the characteristics of the
instance.

Model Comparison. To evaluate the impact of fixing the variables and adding
interval variables, we solve all instances using the base formulation without
model strengthening (CPBase ) and with model strengthening (CPFull ). In pre-
liminary tests, we observed that CPFull had difficulties finding a feasible solution
for large instances. Therefore, CPBase was solved for maximal 60 s as a starting
solution and passed to CPFull .
The results are shown in Table 4. The average values are grouped by their
instance size, i.e., the number of customers. Apart from the count of optimally
solved runs out of the 36 runs (12 instances × 3 runs per instance), the remaining
gaps and runtime are computed considering the average of all runs. Analyzing the
optimal objective function values shows the validity of the CP model compared
to the MIP model. We were able to prove optimality for the first time in a total
of 18 instances. For this, we refer to Table 6 in the Appendix for more detailed
results for all instances.
 A CP Model for VRP with Multiple Time Windows 315

Table 4. Performance evaluation of MIP, basic CP (CPBase ), and CP with variable

fixing and interval variables (CPFull ); comparison of the number of optimally solved
runs (#O), the average remaining gap (G) in [%] after 3,600 s, and the average runtimes
(T) in [s] per instance size (12 instances × 3 runs). For CPFull , TS indicates the start
solution time in [s].

|C| MIP CPBase CPFull

#O G T #O G T #O G T TS
B=0
10 36 0.0 299.9 36 0.0 0.8 36 0.0 0.2 1.2
15 32 1.7 408.3 36 0.0 50.9 36 0.0 3.0 5.5
20 18 15.7 1,804.0 18 14.0 1,800.5 33 0.6 468.4 30.2
25 17 13.8 1,907.6 18 13.6 1,800.9 24 3.7 1,215.8 30.5
50 6 31.1 3,152.8 6 29.6 3,040.2 6 25.0 3,105.4 59.9
Total 21.8 12.5 1,514.5 22.8 11.4 1,338.7 27.0 5.8 958.6 25.5
B=1
10 36 0.0 202.0 36 0.0 10.7 36 0.0 0.6 8.1
15 33 1.8 339.6 33 2.1 304.6 36 0.0 16.2 11.4
20 18 16.0 1,837.6 18 14.4 1,804.9 33 0.6 607.3 34.6
25 12 14.7 2,484.4 15 14.5 2,138.3 21 7.5 1,654.2 42.3
50 3 31.0 3,344.6 3 30.9 3,357.9 3 25.8 3,425.9 59.9
Total 20.4 12.7 1,641.6 21.0 12.4 1,523.3 25.8 6.8 1,140.9 31.3

Although all models can optimally solve the small instances with 10 cus-
tomers, a comparison between MIP and CPBase reveals that the latter performs
better regarding average gap and runtime over all instances. More considerable
differences can be seen for CPFull . For example, instances with up to 15 cus-
tomers are solved within a few seconds, and, unlike other formulations, all but
one of the 20 customer instances can be solved optimally. However, as the num-
ber of customers exceeds 25, the increasing complexity of the problem becomes
apparent in CPFull . This is reflected by higher runtimes and decreased instances
for which optimality can be proven. Despite these challenges, CPFull is still the
most effective formulation for larger instances.
Comparing the two objective functions shows that including waiting costs
makes the optimality proof more difficult since the runtimes are more extensive
as only minimizing vehicle and travel costs. In summary, the CP model is clearly
better than the MIP formulation due to the use of variable interval variables.
Therefore, we will refer to the CPFull model as the CP model in the following.

Impact of the Lower Bound for the Number of Required Vehicles.

As an example, Fig. 2 shows the detailed solution of instance rm101 with 10
customers. The figure demonstrates that the volume utilization is no more than
316 F. Linß

one–third of the vehicle capacity for both routes in the optimal solution. The
fixed costs in the benchmark instances are very high. Therefore, a solution with
fewer vehicles leads to lower costs [2]. However, the time windows of the depot
and the customers do not permit a feasible sequence with only one vehicle, even
though the volume capacity is sufficient to visit all customers using one vehicle.

Fig. 2. Optimal solution of instance rm101 with 10 customers presented by the sched-
ules and the volume utilization of the routes.

The minimum number of vehicles required can be determined easily using

the volume approximation, see Eq. (18).
 
Vol i∈C μi
LB = (18)
κ
To incorporate the time windows into determining the minimum number of
vehicles required, the CP model was rerun using a modified objective function.
Setting B = γij = 0 and γ f = 1 equals the minimization of vehicles required.
The results of these additional runs with limited runtime of 60 s show that, for 16
out of the 60 instances, the lower bound determined by the CP model (LBCP )
was higher than the lower bound of the volume approximation (LBVol ). This
result was used to divide the instances into two sets with LBVol = LBCP and
LBVol < LBCP . Figure 3 shows these subsets’ average gaps and runtimes for the
MIP and the CP model.
For instances with the same vehicle lower bound, the gap values and runtimes
of the CP model are slightly better than those of the MIP model. However, the
differences are greater when the CP model quickly determines that more than
LBVol vehicles are necessary to serve all customers. The gap values of the CP
model are only a quarter of the values of the MIP model, and the runtimes could
 A CP Model for VRP with Multiple Time Windows 317

Fig. 3. Comparison of MIP and CP solver according to the average gap and runtime.
The instances are divided into a set of 44 instances, where the lower bound for the
number of required vehicles calculated by volume approximation (LBVol ) equals the
lower bound calculated by the CP solver (LBCP ) and a set of 16 instances, where
LBVol < LBCP .

also be halved. It becomes evident that the CP model can determine the LB for
the number of vehicles much better than the MIP. As a result, the CP model
leads to better solutions overall because the number of vehicles greatly influences
the objective function due to high fixed costs.

Impact of Instance Characteristics. We observed varying difficulties in solv-

ing the instances. Therefore, in Table 5, we grouped the instances according to
their metrics and the length of the time windows for both problem variants and
then averaged the gap values and runtimes.

Table 5. Average remaining gap G in [%] after 3.600 s, average runtimes (T) in [s], and
the average number of vehicles used (V) grouped by problem variant and time window
lengths per instance metric (20 instances × 3 runs) are displayed.

Metric B=0 B=1

short TW long TW short TW long TW
G T V G T V G T V G T V
cm 5.9 1,638.1 3.3 7.4 1,457.2 1.7 8.9 1,941.2 3.3 10.5 1,333.0 1.8
rcm 0.0 109.5 3.0 8.6 720.3 1.2 0.2 493.6 3.0 9.0 1,096.0 1.2
rm 6.3 1,086.6 3.5 6.9 720.5 1.2 6.3 1,121.6 3.4 5.8 835.0 1.1
Total 4.1 944.7 3.3 7.6 966.0 1.4 5.1 1,185.5 3.3 8.4 1,088.0 1.4

Interestingly, for both problem variants, the random clustered instance with
short time windows can be solved easier than solving the remaining instances.
318 F. Linß

Due to the spatial structure of several randomly arranged customer clusters,

infeasible customer combinations between these clusters can be excluded more
effectively. This effect is weakened somewhat with long time windows since routes
can also be formed here that visit multiple clusters. The increased complexity
resulting from longer time windows is also reflected in other metrics. Since the
problem is more weakly constrained, more customers can be combined on the
routes. This is apparent from the number of the vehicle used in the solutions.
Specifically, while the instances with short time windows involve an average of
3.3 vehicles, only 1.4 vehicles are used for those with long time windows.

4 Conclusion
This paper studies the vehicle routing problem with multiple time windows
(VRPMTW). The goal is to visit all customers with multiple time windows and
start and end within the depot time window. Two variants of the VRPMTW
are considered with different objective functions: the first variant aims to mini-
mize vehicle and travel costs. In contrast, the second variant considers waiting
times in case a vehicle arrives at a customer when there is no time window
open, in addition to vehicle and travel costs. We developed a constraint pro-
gramming (CP) model with additional model strengthening by fixing variables
and variable interval variables to address this problem. Experimental tests on
benchmark instances reveal that model strengthening improves the performance
of the CP model. As a result, the CP model performs considerably better than
a comparable mixed–integer programming (MIP) model. In total, we solved 18
instances optimally for the first time. We achieve proven optimality for almost
all instances with 20 customers and more than half with 25 customers. When
reaching the time limit, the gap values are halved compared to the MIP model.
In addition, we performed a more detailed analysis to better explain the differ-
ence in performance between the two models and to account for the influence of
instance characteristics on the CP model’s solution performance. The feasibility
of the routes is defined by two factors: the respect to the vehicle capacity and the
feasible sequence so that each customer is visited one of the time windows. The
CP model is particularly well suited for instances where customer time windows
constrain the number of required routes in addition to the vehicles’ capacity. This
way, solutions with too few vehicles can be excluded more quickly. Nevertheless,
instances with longer time windows and, thus, more possible combinations of
customers are challenging to solve.
Observations reveal that in instances with 50 customers, there are large dif-
ferences in the solutions found by the CP solver between different runs on the
same instance. To improve the upper bound in such instances, embedding other
solution methods, such as heuristics, is a promising approach. Another extension
involves the integration of the CP model into a sophisticated solution approach.
The results regarding the lower bound of the required vehicles motivate using
the CP model to solve the traveling salesman problem with multiple time win-
dows as a feasibility problem on subsets of customers. Efficient feasibility checks
 A CP Model for VRP with Multiple Time Windows 319

through the CP model can determine that these customers must be served by at
least two vehicles if their combination is infeasible. This information can be used
to create cutting planes, e. g. cover inequalities in a branch-and-cut approach.

Appendix

Table 6. Detailed results of the best runs of MIP and CP for both problem variants
according to objective value (λ), remaining gap (G) in [%] after the time limit 3,600 s
(TL), and runtime (T) in [s].

Inst. C B =0 B =1
MIP CP MIP CP
λ G T λ G T λ G T λ G T
cm101 10 485.3 0.0 3174 485.3 0.0 9.9 490.8 0.0 1662 490.8 0.0 62
cm105 10 486.0 0.0 58 486.0 0.0 0.7 488.5 0.0 38 488.5 0.0 4.8
cm201 10 833.1 0.0 0.5 833.1 0.0 0.3 834.9 0.0 0.4 834.9 0.0 0.5
cm205 10 834.9 0.0 0.5 834.9 0.0 0.3 835.3 0.0 0.3 835.3 0.0 0.8
rcm101 10 566.0 0.0 0.4 566.0 0.0 0.3 566.0 0.0 0.2 566.0 0.0 0.4
rcm105 10 566.8 0.0 1.2 566.8 0.0 0.3 569.5 0.0 0.2 569.5 0.0 0.8
rcm201 10 1137.8 0.0 1.0 1137.8 0.0 0.2 1137.8 0.0 0.2 1137.8 0.0 0.3
rcm205 10 1137.8 0.0 0.4 1137.8 0.0 0.2 1137.8 0.0 0.1 1137.8 0.0 0.3
rm101 10 598.2 0.0 54 598.2 0.0 0.5 606.7 0.0 30 606.7 0.0 3.5
rm105 10 595.4 0.0 125 595.4 0.0 1.9 595.4 0.0 40 595.4 0.0 27
rm201 10 1173.0 0.0 0.1 1173.0 0.0 0.2 1189.1 0.0 1.3 1189.1 0.0 0.9
rm205 10 1173.0 0.0 0.1 1173.0 0.0 0.2 1173.0 0.0 0.1 1173.0 0.0 0.3
Mean 10 799.0 0.0 285 799.0 0.0 0.2 802.1 0.0 148 802.1 0.0 0.5
cm101 15 537.8 0.0 9.6 537.8 0.0 3.9 555.3 0.0 263 555.3 0.0 148
cm105 15 538.1 0.0 4.1 538.1 0.0 1.6 538.1 0.0 4.2 538.1 0.0 6.0
cm201 15 870.6 0.0 3.3 870.6 0.0 1.1 873.7 0.0 4.6 873.7 0.0 5.9
cm205 15 883.2 0.0 10 883.2 0.0 4.1 883.2 0.0 4.8 883.2 0.0 7.8
rcm101 15 586.9 0.0 2.3 586.9 0.0 0.7 589.3 0.0 1.0 589.3 0.0 2.0
rcm105 15 588.5 0.0 3.2 588.5 0.0 1.1 594.5 0.0 1.2 594.5 0.0 4.6
rcm201 15 1153.6 0.0 3.9 1153.6 0.0 0.4 1153.6 0.0 0.5 1153.6 0.0 0.8
rcm205 15 1153.6 0.0 28 1153.6 0.0 0.4 1156.2 0.0 0.9 1156.2 0.0 2.4
rm101 15 878.2 19.8 TL 878.2 0.0∗ 86 883.8 21.3 TL 883.8 0.0∗ 86
rm105 15 661.5 0.0 2.9 661.5 0.0 1.3 664.9 0.0 1.1 664.9 0.0 4.2
rm201 15 1229.2 0.0 0.3 1229.2 0.0 0.3 1237.2 0.0 1.4 1237.2 0.0 1.6
rm205 15 1229.2 0.0 0.4 1229.2 0.0 0.6 1229.2 0.0 0.2 1229.2 0.0 0.6
Mean 15 859.2 1.7 306 859.2 0.0 2.9 863.2 1.8 324 863.2 0.0 11
cm101 20 801.7 27.7 TL 801.7 0.0∗ 1236 839.0 29.1 TL 832.0 7.2 TL
cm105 20 801.1 28.0 TL 801.1 0.0∗ 558 802.6 28.5 TL 802.6 0.0∗ 798
cm201 20 1599.0 44.3 TL 1599.0 0.0∗ 65 1599.0 44.3 TL 1599.0 0.0∗ 68
cm205 20 1607.2 44.5 TL 948.7 6.7 TL 1608.7 44.4 TL 948.7 0.0∗ 2059
rcm101 20 883.7 0.0 6.4 883.7 0.0 1.7 887.3 0.0 1.7 887.3 0.0 4.5
rcm105 20 885.2 0.0 4.2 885.2 0.0 2.4 895.8 0.0 6.3 895.8 0.0 9.4
rcm201 20 1219.4 0.0 6.2 1219.4 0.0 0.8 1219.4 0.0 0.9 1219.4 0.0 1.2
(continued)
320 F. Linß

Table 6. (continued)

Inst. C B =0 B =1
MIP CP MIP CP
λ G T λ G T λ G T λ G T
rcm205 20 1219.4 0.0 6.3 1219.4 0.0 0.7 1229.6 0.0 194 1229.6 0.0 111
rm101 20 922.4 21.5 TL 922.4 0.0∗ 76 938.1 22.8 TL 938.1 0.0∗ 103
rm105 20 906.2 21.2 TL 906.2 0.0∗ 76 906.2 21.5 TL 906.2 0.0∗ 74
rm201 20 1262.3 0.0 0.7 1262.3 0.0 0.7 1280.4 0.0 98 1280.4 0.0 17
rm205 20 1262.3 0.0 0.5 1262.3 0.0 1.0 1262.3 0.0 0.7 1262.3 0.0 1.6
Mean 20 1114.2 15.6 1802 1059.3 0.6 443 1122.4 15.9 1825 1066.8 0.6 543
cm101 25 847.2 5.5 TL 825.2 6.3 TL 888.8 8.1 TL 876.3 11.7 TL
cm105 25 823.6 2.0 TL 823.6 5.8 TL 825.2 2.3 TL 825.2 6.0 TL
cm201 25 1629.8 44.6 TL 1626.9 0.0∗ 195 1629.9 44.7 TL 1628.7 0.0∗ 378
cm205 25 1619.7 44.2 TL 1023.3 11.7 TL 1624.8 44.3 TL 1624.8 44.5 TL
rcm101 25 895.0 0.0 5.6 895.0 0.0 3.3 898.6 0.0 4.8 898.6 0.0 5.4
rcm105 25 898.3 0.0 5.5 898.3 0.0 6.3 912.4 0.0 35 912.4 0.0 50
rcm201 25 1226.1 0.0 20 1226.1 0.0 1.2 1235.7 0.0 546 1235.7 0.0 23
rcm205 25 1226.1 0.0 24 1226.1 0.0 1.2 1255.2 1.7 TL 1255.2 2.3 TL
rm101 25 1185.8 35.7 TL 1185.8 19.8 TL 1209.3 36.8 TL 1209.3 21.5 TL
rm105 25 960.1 21.2 TL 960.1 0.0∗ 66 976.2 22.8 TL 976.2 0.0∗ 231
rm201 25 1316.9 0.0 13 1316.9 0.0 2.3 1342.0 0.8 TL 1342.0 0.0∗ 243
rm205 25 1313.3 0.0 1.6 1313.3 0.0 1.6 1315.7 0.0 1.4 1315.7 0.0 5.3
Mean 25 1161.8 12.8 1806 1110.1 3.6 1213 1176.2 13.5 2449 1175.0 7.2 1562
cm101 50 1868.3 27.8 TL 1683.7 21.5 TL 1833.1 25.4 TL 1913.7 31.0 TL
cm105 50 1657.5 18.4 TL 1650.0 19.7 TL 1658.1 17.5 TL 1687.1 21.4 TL
cm201 50 3234.6 46.8 TL 3230.0 23.9 TL 3263.6 46.5 TL 3247.4 24.3 TL
cm205 50 2567.5 32.3 TL 2524.5 30.9 TL 2559.3 31.7 TL 2530.7 31.1 TL
rcm101 50 1525.8 0.0 261 1525.8 0.0 232 1537.3 0.0 440 1537.3 0.0∗ 980
rcm105 50 1548.5 0.0 1363 1548.5 0.0 557 1566.0 1.1 TL 1568.2 2.1 TL
rcm201 50 2398.4 43.3 TL 2398.4 42.9 TL 2448.2 44.0 TL 2423.0 43.5 TL
rcm205 50 2398.4 43.4 TL 2398.4 42.9 TL 2430.5 43.5 TL 2431.1 43.7 TL
rm101 50 1885.5 29.6 TL 2063.6 26.7 TL 2109.7 36.8 TL 1922.5 21.4 TL
rm105 50 1797.0 26.4 TL 1794.2 15.8 TL 1793.7 25.9 TL 1806.2 16.4 TL
rm201 50 2488.6 42.3 TL 2485.7 41.2 TL 2664.3 45.1 TL 2540.5 42.5 TL
rm205 50 2478.1 41.7 TL 1486.7 1.7 TL 2479.7 41.0 TL 1487.0 1.7 TL
Mean 50 2154.0 29.3 3136 2065.8 22.3 3068 2195.3 29.9 3337 2091.2 23.2 3390
Note. Start solution time is excluded for CP. Bold indicates the minimum gap per
instance and variant. If both methods prove optimality, bold indicates minimum
runtime. ∗ denotes instances that the CP model solved optimally for the first time.

References
1. Baldacci, R., Mingozzi, A., Roberti, R.: Recent exact algorithms for solving the
vehicle routing problem under capacity and time window constraints. Eur. J. Oper.
Res. 218(1), 1–6 (2012). https://doi.org/10.1016/j.ejor.2011.07.037
2. Belhaiza, S., Hansen, P., Laporte, G.: A hybrid variable neighborhood tabu search
heuristic for the vehicle routing problem with multiple time windows. Comput.
Oper. Res. 52, 269–281 (2014). https://doi.org/10.1016/j.cor.2013.08.010
3. Belhaiza, S., M’Hallah, R., Brahim, G.B.: A new hybrid genetic variable neigh-
borhood search heuristic for the vehicle routing problem with multiple time win-
dows. In: 2017 IEEE Congress on Evolutionary Computation (CEC). IEEE (2017).
https://doi.org/10.1109/cec.2017.7969457
 A CP Model for VRP with Multiple Time Windows 321

4. Bogue, E.T., Ferreira, H.S., Noronha, T.F., Prins, C.: A column generation and
a post optimization VNS heuristic for the vehicle routing problem with multiple
time windows. Optim. Lett. 16(1), 79–95 (2020). https://doi.org/10.1007/s11590-
019-01530-w
5. Hoogeboom, M., Dullaert, W., Lai, D., Vigo, D.: Efficient neighborhood evaluations
for the vehicle routing problem with multiple time windows. Transp. Sci. 54(2),
400–416 (2020). https://doi.org/10.1287/trsc.2019.0912
6. de Jong, C., Kant, G., Van Vlient, A.: On finding minimal route duration in the
vehicle routing problem with multiple time windows. Manuscript, Department of
Computer Science, Utrecht University, Holland (1996)
7. Laborie, P., Rogerie, J., Shaw, P., Vilı́m, P.: Reasoning with conditional time-
intervals. Part II: an algebraical model for resources. In: FLAIRS Conference, pp.
201–206 (2009)
8. Larsen, R., Pacino, D.: Fast delta evaluation for the vehicle routing problem with
multiple time windows (2019). https://doi.org/10.48550/ARXIV.1905.04114
9. Liu, C., Aleman, D.M., Beck, J.C.: Modelling and solving the senior transportation
problem. In: van Hoeve, W.-J. (ed.) CPAIOR 2018. LNCS, vol. 10848, pp. 412–428.
Springer, Cham (2018). https://doi.org/10.1007/978-3-319-93031-2 30
10. Naderi, B., Ruiz, R., Roshanaei, V.: Mixed-integer programming vs. constraint
programming for shop scheduling problems: new results and outlook. INFORMS
J. Comput. (2023). https://doi.org/10.1287/ijoc.2023.1287
11. Pesant, G., Gendreau, M., Potvin, J.Y., Rousseau, J.M.: On the flexibility of con-
straint programming models: from single to multiple time windows for the traveling
salesman problem. Eur. J. Oper. Res. 117(2), 253–263 (1999). https://doi.org/10.
1016/s0377-2217(98)00248-3
12. Rousseau, L.M., Gendreau, M., Pesant, G., Focacci, F.: Solving VRPTWs with
constraint programming based column generation. Ann. Oper. Res. 130(1–4), 199–
216 (2004). https://doi.org/10.1023/b:anor.0000032576.73681.29
13. Schaap, H., Schiffer, M., Schneider, M., Walther, G.: A large neighborhood search
for the vehicle routing problem with multiple time windows. Transp. Sci. 56(5),
1369–1392 (2022). https://doi.org/10.1287/trsc.2021.1120
14. Solomon, M.M.: Algorithms for the vehicle routing and scheduling problems with
time window constraints. Oper. Res. 35(2), 254–265 (1987). https://doi.org/10.
1287/opre.35.2.254
15. Voigt, S., Frank, M., Fontaine, P., Kuhn, H.: The vehicle routing problem with
availability profiles. Transp. Sci. (2022). https://doi.org/10.1287/trsc.2022.1182
 A Variable Neighborhood Search Algorithm
for the Truck-Drone Routing Problem

Batool Madani(B) and Malick Ndiaye

American University of Sharjah, Sharjah, United Arab Emirates

{g00050500,mndiaye}@aus.edu

Abstract. The rise in e-commerce sales has increased the demand for logistics
companies to handle the problem of last-mile delivery, which is complex and
incurs large operational expenses. The use of drones in the development of modern
distribution networks opens new avenues for improving existing delivery systems.
The integration of drones with traditional delivery trucks that form the truck-drone
systems are being proposed by logistics firms for use in delivery procedures. In
this paper, we consider a single truck and drone delivery system and address the
associated routing problem. As a result, a variable neighborhood search algorithm
is developed and tested on small and medium instances. The algorithm shows
its potential ability to handle instances in a short time. Finally, the introduced
truck-drone delivery system with multiple visits per drone dispatch generates
cost-savings in comparison with the case of a single visit per delivery trip.

Keywords: Routing · truck-drone delivery · variable neighborhood search

1 Introduction
The rapid evolution of e-commerce has intensified the need for logisticians to resolve the
problem of conveying goods from transportation hubs to their final destinations, known as
the Last-Mile Delivery problem (LMD) (Boyer et al. 2009). It is the last leg of a product’s
journey, nevertheless the most challenging, with the highest transportation cost across
all distribution networks (Gevaers et al. 2009). The inclusion of new technologies such
as drones may help in overcoming these challenges.
Drones are now widely used in a variety of applications, including logistics, health-
care, and precision agriculture (Scott and Scott 2017; Mogili and Deepak 2018; Murray
and Chu 2015). In this context, major drone manufacturers such as Zipline, Wing, Prime
Air, Flirtey, and others have expressed interest in incorporating drone technologies into
their operations (Wood 2019). For instance, Prime Air, developed by Amazon, can satisfy
customers’ orders within a 16 km radius of a fulfillment center by navigating using the
onboard global positioning system (Lavars 2013). Similarly, Zipline drones have made
over 13,000 deliveries across Rwanda, demonstrating their humanitarian capabilities by
transporting blood for transfusions (Dragolea 2019).
Although drones have shown potential in improving delivery operations, they are
limited by their technical characteristics such as payload capacity and flying range. As a

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 322–334, 2023.
https://doi.org/10.1007/978-3-031-43612-3_20
 A Variable Neighborhood Search Algorithm 323

result, drones and delivery trucks can work in tandem to provide a more flexible delivery
system, resulting in a truck-drone delivery system.
Truck-drone routing problems are viewed as extended versions of the classical Vehi-
cle Routing Problems (VRP), which calls for the development of efficient solution meth-
ods to address the NP-hard nature of the related routing problem. Therefore, this study
considers a specific setting of the truck-drone delivery system and proposes a solution
method based on the variable neighborhood search (VNS) algorithm to solve the routing
considerations. The main contribution of this work is as follows.
1. Developing an efficient VNS algorithm for a truck-drone delivery system where the
drone is able to perform multiple visits per dispatch.
2. Conducting computational experiments to assess the algorithm’s performance using
small and medium instances.
The remainder of the paper is structured as follows. Section 2 presents a review
of the truck-drone routing problems-related literature. Section 3 provides the proposed
model description while Sect. 4 demonstrates the solution approach. The computational
analysis is given in Sect. 5 and finally, Sect. 6 concludes and provides future research
prospects.

2 Related Literature
The exploration of truck-drone systems is definitely appealing to academics and practi-
tioners. According to UPS, truck-drone systems might save up to $50 million per year
by reducing every driver’s route by one mile every day (UPS 2017). Amazon has also
been granted a patent for a truck-drone system in which the truck functions as a mobile
facility for operating and repairing the drone (Wohlsen 2014).
The truck-drone problems-related literature has gained significant attention due to
the various complexities associated with modeling and solving optimization problems.
Interested readers may refer to (Madani and Ndiaye 2022; Chung et al. 2020; Boysen et al.
2021) reviews, among others, to understand the current state-of-the-art of truck-drone
routing problems.
Researchers’ efforts are devoted to studying the logistical uses of drones in enhanc-
ing the delivery processes and tackling the LMD challenges. These studies take into
consideration the operational limitations of a drone and the routing optimization of a
combined truck-drone system. In this context, Murray and Chu (2015) are among the first
to investigate the flying side-kick traveling salesman problem (FSTSP), which considers
a set of customers to be served once by either a truck or a drone, with the truck only
fulfilling customer deliveries that the drone cannot serve. The FSTSP has been extended
and solved in different ways by (Jeong et al. 2019; Windras Mara et al. 2022; Liu et al.
2022; Boccia et al. 2021; de Freitas and Penna 2018; Dell’Amico et al. 2019; Murray
and Raj 2020), among others. For instance, Dell’Amico et al. (2019) proposed a new for-
mulation for the FSTSP that uses two and three-indexed formulations and strengthened
it by a set of valid inequalities and finally solved using branch-and-cut. This approach
yielded solutions in a relatively shorter time frame. Murray and Raj (2020) considered a
fleet of heterogeneous drones, each of which varies in terms of weight, volume capacity,
and endurance.
324 B. Madani and M. Ndiaye

Murray and Chu (2015) suggested the parallel drone scheduling traveling salesman
problem (PDTSP) as another form of the truck-drone delivery problem. In particular, a
truck and one or more drones perform delivery services independently, departing and
returning to the depot simultaneously. The objective is to reduce the time required to
service all customers and allow all vehicles to return to the depot, which was solved
using exact solution methods and a saving heuristic. Lei and Chen (2022) designed an
improved VNS algorithm that is based on reduced VNS for solving the PDSTP. The
algorithm includes 9 neighborhood structures covering operators such as relocate, 1-1
exchange, 2-2 exchange, and other operators tailored to the specific problem.
In a different approach, Agatz et al. (2018) introduced the traveling salesman prob-
lem with a drone (TSP-D), in which the drone must follow the truck’s road network and
may be launched and returned to the same location, which is not allowed in the FSTSP.
The objective of the TSP-D is to minimize the time required to serve all customers or
minimize the arrival time to the depot. The VRP with a drone (VRP-D) considers the
shared delivery operations between multiple trucks and multiple drones visiting a single
customer location per dispatch. Wang et al. (2017) introduced the VRP-D which deals
with multiple trucks and drones with the goal of minimizing the completion time, in
which the drones can be dispatched from and picked up only at the depot and customers’
nodes. The analysis was conducted on several worst-case scenarios, from which they
propose bounds on the best possible savings in terms of completion time when using
trucks and drones instead of trucks only in delivering parcels to customers. Further devel-
opment of this research was studied by Poikonen et al. (2017), where they extended the
worst-case bounds to more generic distance/cost metrics. Also, they explicitly consider
the limitation of battery life and cost objectives. Sacramento et al. (2019) investigated
the VRP-D with each truck equipped with a single drone and developed an adaptive large
neighborhood search algorithm (ALNS). Kuo et al. (2022) studied a similar problem as
Sacramento et al. (2019) but considered the time-window aspect and developed a VNS
solution.
In the above studies, trucks and drones share the responsibility of fulfilling cus-
tomer deliveries, while permitting the launching/collecting of drones to take place at the
customer nodes visited by the truck.
Another truck-drone configuration is to limit the delivery role to the drone while the
truck acts as a moving warehouse for launching, collecting, and refilling with the drone.
In such cases, the drone’s launching and collection operations occur at non-customer
nodes such as drone stations/hubs, nodes along the truck route, and others. This system
configuration can maximize the potential use of drones as they are not restricted to
being launched from customer nodes which may not have all the proper facilities for
launching/collecting the drone. It is also important to stress that few studies limited the
launching/collection to non-customer nodes (Karak and Abdelghany 2019; Poikonen
and Golden 2020a, b; Marinelli et al. 2018; Carlsson and Song 2018; Wang and Sheu
2019). In addition, the majority of the literature considers a single drone capacity while
it would be interesting to investigate the case of allowing the drone to perform multiple
visits per dispatch.
 A Variable Neighborhood Search Algorithm 325

Therefore, this study considers the second configuration of the truck-drone deliv-
ery system while considering drone characteristics such as payload capacity and flying
endurance.

3 Model Description
Given a single truck equipped with a single drone, the task is to deliver packages to
a given set of customers. Each customer must be visited exactly once by the drone
operating in conjunction with the truck while the truck acts as a platform for refilling,
launching, and collecting the drone. The truck with its drone must depart and return to a
single depot. The launching and collecting operations can take place at a set of potential
locations. The objective is to determine the sequencing of the deliveries by drone while
minimizing the total traveling cost of the delivery system. Several assumptions are taken
into account and are described below.
• Drones can handle multiple deliveries per dispatch, according to their maximum
payload capacity Q and flying endurance E.
• Customer demands are based on the weight of the packages.
• Trucks are assumed to have a large enough capacity to transport both parcels and
drones during the whole operation.
• The drone may be launched/collected from the depot or at any of the launch-
ing/collecting potential locations.
• The drone can be collected at the same location from where it was launched or at any
other location.
• The truck must precede the drone at the collection location for safety purposes. In
other words, the arrival time of the truck at the collection location must be before that
of the drone.
Figure 1 demonstrates the delivery operation performed by the truck and the drone.

4 Proposed Variable Neighborhood Search Algorithm

Truck-drone routing problems are regarded as extended variants of traditional VRPs.

When considering a single truck stop and a single drone with a single delivery path, the
truck-drone routing problem reduces to a traveling salesman problem, which is known to
be NP-hard (Madani and Ndiaye 2019). As a result, the truck-drone problem is NP-hard,
and the computing time necessary to reach optimality using exact solutions may increase
exponentially as the problem size increases (Madani and Ndiaye 2019). This highlights
the importance of developing efficient metaheuristic algorithms capable of providing
acceptable solutions for large-scale instances (Salhi 2017). Therefore, an approximate
solution method such as the VNS is developed.
The VNS, developed by Mladenović and Hansen (1997), works on a systematic
neighborhood transformation in which a local search is included in the search. Add to
that, it has the benefit of being simple and straightforward to apply, as it only requires
the definition of neighborhoods (Salhi 2017).
326 B. Madani and M. Ndiaye

Fig. 1. Truck-drone delivery model configuration

The first step in developing the VNS algorithm is to examine the VNS literature in
the context of truck-drone routing models. Table 1 summarizes the literature on VNS for
solving the routing considerations of truck-drone systems. It can be clearly observed that
only a few research have adopted the VNS algorithm. This necessitates the use of VNS
in such problems to investigate its capabilities in the context of truck drone problems. In
a study conducted by Kuo et al. (2022), the VNS proves its ability in providing quality
solutions for problems related to truck-drone routing compared to other metaheuristics
such as the ALNS. Therefore, it is worth exploring the adoption of VNS in this study.

Table 1. VNS-related literature in the context of truck-drone routing problems

Reference VNS variant No. neighborhoods Size of instance

Schermer et al. (2019) Basic VNS 3 Up to 50 nodes
de Freitas and Penna (2018) RVND 5 Up to 100 nodes
de Freitas and Penna (2020) General VNS 7 Up to 250 nodes
Salama and Srinivas (2022) Basic VNS 6 Up to 50 nodes
Kuo et al. (2022) Basic VNS 8 Up to 50 nodes
Gu et al. (2022) VND 6 Up to 200 nodes
Luo et al. (2022) VND 6 Up to 80 nodes
Lei and Chen (2022) Reduced VNS 9 Up to 100 nodes

The structure of the VNS algorithm is divided into three elements, namely, the
initial solution, neighborhood structures, and finally local search strategy. Next, each of
the three elements is presented in detail.
 A Variable Neighborhood Search Algorithm 327

4.1 Initial Solution

The initial solution procedure is based on selecting the launching and collecting nodes
of each drone trip according to the drone’s characteristics (payload capacity and flying
endurance). The approach of the initial solution chooses the depot as the launching node
of the first drone’s delivery trip and assigns a customer node. If no customer node exists
within the flight endurance of the drone from the depot, the next potential location is
chosen, and so on. Following the selection of the launching point, and assuming that
the payload limit has not been exceeded, a customer node can be added to the drone’s
delivery trip, or the current drone trip is terminated. After that, the collecting location
of the drone is selected according to the drone’s flying endurance limit. A new drone
delivery trip can be initiated, and the same steps are repeated until all customers are
served by the drone.

4.2 Neighborhood Structures

In this algorithm, eight neighborhood structures (Kmax = 8), which are a combination
of traditional neighborhoods and problem-specific neighborhoods, are used. It is worth
noting that some of these neighborhoods correspond to the well-known VRP local search
operators. The list of neighborhoods is presented below.
• N 1 : Select a random customer node from two drone delivery trips and swap them.
This neighborhood is similar to 1-1 exchange operators where the positions of a pair
of nodes are swapped.
• N 2 : Select a random drone trip and split it into two delivery trips. In particular, a
random trip is chosen, and a new stopping location is added at random between the
initial launching and collecting nodes, splitting the initial trip.
• N 3 : Select two random drone delivery trips and combine them into a single delivery
trip.
• N 4 : Remove a random customer node from a drone delivery trip and insert it in
another drone trip. This move is related to relocate operator in that a node is injected
from its original position and relocated to another.
• N 5 : Select two random drone delivery trips and swap their order in the list of drone
trips.
• N 6 : Select a random drone trip and swap the launching and collecting locations. If
they are similar, this neighborhood is neglected.
• N 7 : Select a random drone trip and swap two customer nodes. This is comparable to
the 1-1 exchange operator, except it occurs within the same route (exchange intra-
route).
• N 8 : Select a random drone trip where the launching and collecting nodes are different
and make the launching location similar to the collecting one.

4.3 Local Search

The local search, which is based on the exploration of a neighborhood structure, is an

improvement procedure used within the VNS algorithm.
328 B. Madani and M. Ndiaye

In the proposed algorithm, the solution generated in the shaking procedure Nk (X ) of

neighborhood Nk is exhaustively examined. In specific, all the possible swaps between
two customers within a drone delivery trip are performed and the solution with the best
improvement is selected.
As there is a restriction on the drone’s payload capacity, flying endurance, and arrival
time at the collection site, the feasibility testing of the solution is obtained during the
construction of the initial solution and shaking procedures. In other words, the solution
is ignored if a movement results in violating the aforementioned constraints.
The structure of the VNS algorithm is summarized in Algorithm 1. The stopping
criterion is based on the maximum number of iterations.

5 Computational Experiments
In this section, the performance of the proposed VNS algorithm is assessed using
instances derived from Sacramento et al. (2019). We first present the description of
instances and the inputs used in the algorithm followed by the performance evaluation
and additional analysis.
The algorithm is coded using the C++ programming language; all the runs and
experiments were performed on a 64-bit operating system with an x64-based processor
that is Intel(R) Core™ i7-8565U CPU @1.80 GHz and 8 GB memory.

5.1 Parameters and Datasets

The parameters used in the algorithm solution are as follows. Drone traveling cost is
assumed to be 30% of the truck traveling cost (Campbell et al. 2017). The traveling
distance of both vehicles is assumed to be Euclidean. Also, the drone is assumed to have
 A Variable Neighborhood Search Algorithm 329

a maximum payload capacity Q of 5 kg (∼10 lbs) and a flying range E of 30 min, which
are the typical characteristics of a drone. Finally, the drone’s speed is set to 50 mph
(Sacramento et al. 2019).
The instances used in the VNS performance evaluation are based on Sacramento
et al. (2019) instances, which are formed on a grid size of 2d × 2d where d is 5, 10, 20,
30, or 40 and follow a uniform distribution. Finally, customer packages are set to weigh
less than 5 kg and follow a uniform distribution.
Given that Sacramento et al. (2019) presented the VRP-D where the launch-
ing/collecting nodes are customer nodes, we added additional nodes representing the
potential launching/collecting location to each instance. We followed the p-median app-
roach to generate the potential locations used as possible launching/collecting nodes.
The number of potential locations is selected based on the number of customer nodes in
each instance size. In this analysis, we chose 30 instances of size 6, 10, 12, 20, and 50
customer nodes to apply the modification on. The modified instances are named S[no.
potential locations (including depot)].[no. customer nodes].[grid dimension].[instance
number].

5.2 Performance Evaluation

To assess the performance of the VNS algorithm, we use modified instances and obtain
the solution for each. Each instance is computed 10 times (10 runs) and the average C Avg
and best C best solutions are recorded. In addition, the average computational time is docu-
mented t Avg in seconds. Finally, the maximum number of iterations (maximumiteration)
is fixed to 100 for all instances. Table 2 summarizes the results of the 30 modified
instances.
The results show that the VNS generates solutions for instances with up to 50 cus-
tomer nodes in less than 1 s. Add to that, the algorithm provides relatively consistent
results given the difference between C best and C Avg in each instance. In terms of compu-
tational time, the suggested approach outperforms the previous VNS algorithms found
in truck-drone literature (Lei and Chen 2022; de Freitas and Penna 2018 2020; Schermer
et al. 2019; Kuo et al. 2022). Nevertheless, comparing to other heuristics based on com-
putational time is insufficient and requires evaluation based on solution quality. Indeed,
the suggested algorithm must be benchmarked against other algorithms or exact solution
methods to assess its solution quality, which will be part of our future work.

5.3 Drone Capacity Impact: Single vs. Multiple Visits per Drone Dispatch

The ability of the drone to perform multiple visits may have a significant impact on the
delivery cost.
According to Meng et al. (2023), allowing the drone to visit several customers in
one route can bring cost savings.
In this part, we analyze the impact of the drone’s payload capacity by preventing
the drone from visiting more than a single customer per dispatch. The analysis is per-
formed on S(3.6.5.1), S(5.10.5.1), S(6.12.5.1), S(8.20.5.1), and S(10.50.10.3) instances.
Table 3 compares the objective function value for both cases using the above-mentioned
330 B. Madani and M. Ndiaye

Table 2. Performance evaluation of the VNS algorithm

Instance C best C Avg t Avg

S(3.6.5.1) 4.23 4.40 0.0165
S(3.6.5.2) 4.12 4.32 0.0171
S(3.6.10.1) 8.66 9.07 0.0156
S(3.6.10.2) 9.07 9.19 0.0158
S(3.6.20.1) 14.80 15.16 0.0154
S(3.6.20.2) 14.65 16.45 0.0151
S(5.10.5.1) 5.20 6.07 0.1248
S(5.10.5.2) 5.24 6.32 0.1282
S(5.10.10.1) 13.49 15.64 0.1069
S(5.10.10.2) 10.50 11.42 0.1217
S(5.10.20.1) 20.54 23.92 0.1064
S(5.10.20.2) 24.80 33.04 0.1055
S(6.12.5.1) 6.40 7.78 0.2385
S(6.12.5.2) 6.04 8.05 0.2693
S(6.12.10.1) 12.66 14.38 0.3100
S(6.12.10.2) 12.65 15.40 0.3579
S(6.12.20.1) 30.78 36.85 0.3083
S(6.12.20.2) 33.15 43.42 0.2872
S(8.20.5.1) 12.03 12.88 0.3701
S(8.20.5.2) 12.51 13.94 0.3218
S(8.20.10.1) 20.76 23.46 0.3254
S(8.20.10.2) 29.23 31.61 0.3204
S(8.20.20.1) 51.57 61.80 0.3200
S(8.20.20.2) 53.91 59.59 0.3900
S(10.50.10.3) 65.69 73.40 0.6400
S(10.50.10.4) 65.61 68.05 0.6582
S(10.50.20.3) 140.48 153.71 0.6281
S(10.50.20.4) 134.76 147.08 0.6295
S(10.50.30.3) 148.23 171.89 0.6208
S(10.50.30.4) 180.38 198.22 0.6212

instances. The findings indicate that enabling the drone to service multiple customers
can better leverage the drone’s flying endurance, resulting in a shorter distance traveled
by the truck to retrieve the drone, and thus generating cost-savings.
 A Variable Neighborhood Search Algorithm 331

Table 3. Performance evaluation of the VNS algorithm based on the drone’s payload capacity

Instance Single visit C Avg Multiple visits C Avg

S(3.6.5.1) 7.50 4.40
S(5.10.5.1) 11.83 6.07
S(6.12.5.1) 16.35 7.78
S(8.20.5.1) 23.97 12.88
S(10.50.10.3) 129.78 73.40

5.4 Drone Flying Range Impact

Using the above instances in Subsect. 5.3, the impact of the drone’s flying range is
evaluated, where E is set to 15, 20, 30, and 40 min. Table 4 summarizes the results
of this experimental analysis based on C Avg . As can be observed, reducing the flying
range increases the traveling cost because the drone must make additional delivery trips
to satisfy the needs of the customer. On the other hand, increasing it will reduce travel
costs to an extent as the drone’s performance is limited by other features such as its
flying capacity.

Table 4. Performance evaluation of the VNS algorithm based on the drone’s flying range

Instance Flying range E

15 min 20 min 30 min 40 min
S(3.6.5.1) 5.18 4.61 4.40 4.23
S(5.10.5.1) 7.79 7.19 6.07 5.56
S(6.12.5.1) 9.78 8.44 7.78 6.43
S(8.20.5.1) 17.90 13.28 12.88 11.47
S(10.50.10.3) 81.31 79.92 73.40 69.35

6 Conclusion and Future Work

This paper proposed a VNS algorithm to solve the routing considerations of a truck-drone
delivery system. In particular, a single truck functioning as a moving hub is in charge
of operating the drone from potential truck locations while the drone is responsible for
serving multiple customers per dispatch. The suggested approach has demonstrated its
ability to solve small to medium instances in a short time.
This research can be further explored in a variety of ways. To begin, the truck-drone
delivery system could look into using multiple trucks in tandem with multiple drones to
service customers. Second, other aspects of the problem, such as time window and energy
332 B. Madani and M. Ndiaye

consumption, may be examined. Third, a mathematical formula for the given problem
description may be presented. Fourth, the suggested algorithm must be compared to
other solutions generated by exact solution methods and/or metaheuristic solutions to
better assess its quality. Finally, extensive computational analysis of a large number of
instances of various sizes may be performed.

References
Agatz, N., Bouman, P., Schmidt, M.: Optimization approaches for the traveling salesman problem
with drone. Transp. Sci. 52, 965–981 (2018). https://doi.org/10.1287/trsc.2017.0791
Boccia, M., Masone, A., Sforza, A., Sterle, C.: A column-and-row generation approach for the
flying sidekick travelling salesman problem. Transp. Res. Part C 124 (2021). https://doi.org/
10.1016/j.trc.2020.102913
Boyer, K.K., Prud’homme, A.M., Chung, W.: The last mile challenge: evaluating the effects of
customer density and delivery window patterns. J. Bus. Logist. 30(1), 185–201 (2009). https://
doi.org/10.1002/j.2158-1592.2009.tb00104.x
Boysen, N., Fedtke, S., Schwerdfeger, S.: Last-mile delivery concepts: a survey from an operational
research perspective. OR Spectr. 43(1), 1–58 (2021). https://doi.org/10.1007/s00291-020-006
07-8
Campbell, J.F., Sweeney, D., Zhang, J.: Strategic design for delivery with trucks and drones.
Technical report (2017)
Carlsson, J.G., Song, S.: Coordinated logistics with a truck and a drone. Manag. Sci. 64, 4052–4069
(2018). https://doi.org/10.1287/mnsc.2017.2824
Chung, S.H., Sah, B., Lee, J.: Optimization for drone and drone-truck combined operations: a
review of the state of the art and future directions. Comput. Oper. Res. 123 (2020). https://doi.
org/10.1016/j.cor.2020.105004
de Freitas, J.C., Penna, P.H.V.: A randomized variable neighborhood descent heuristic to solve the
flying sidekick traveling salesman problem. Electron. Notes Discret. Math. 66, 95–102 (2018).
https://doi.org/10.1016/j.endm.2018.03.013
de Freitas, J.C., Penna, P.H.V.: A variable neighborhood search for flying sidekick traveling
salesman problem. Int. Trans. Oper. Res. 27(1), 267–290 (2020). https://doi.org/10.1111/itor.
12671
Dell’Amico, M., Montemanni, R., Novellani, S.: Drone-assisted deliveries: new formulations for
the flying sidekick traveling salesman problem. Optim. Lett. (2019). https://doi.org/10.1007/
s11590-019-01492-z
Dragolea, N.: 9 Drones that will revolutionize Healthcare (2019). http://www.doctorpreneurs.com/
9-drones-that-will-revolutionise-healthcare/. Accessed 26 Feb 2021
Gevaers, R., Van de Voorde, E., Vanelslander, T.: Characteristics of innovations in last-mile
logistics-using best practices, case studies and making the link with green and sustainable
logistics. Association for European Transport and Contributors (2009)
Gu, R., Poon, M., Luo, Z., Liu, Y., Liu, Z.: A hierarchical solution evaluation method and a hybrid
algorithm for the vehicle routing problem with drones and multiple visits. Transp. Res. Part C
141 (2022). https://doi.org/10.1016/j.trc.2022.103733
Jeong, H.Y., Song, B.D., Lee, S.: Truck-drone hybrid delivery routing: payload-energy dependency
and No-Fly zones. Int. J. Prod. Econ. 214, 220–233 (2019). https://doi.org/10.1016/j.ijpe.2019.
01.010
Karak, A., Abdelghany, K.: The hybrid vehicle-drone routing problem for pick-up and delivery
services. Transp. Res. Part C 102, 427–449 (2019). https://doi.org/10.1016/j.trc.2019.03.021
 A Variable Neighborhood Search Algorithm 333

Kuo, R.J., Lu, S.-H., Lai, P.-Y., Mara, S.T.W.: Vehicle routing problem with drones considering
time windows. Expert Syst. Appl. 191 (2022). https://doi.org/10.1016/j.eswa.2021.116264
Lavars, N.: Amazon reveals Prime Air drone capable of 30 minute deliveries (2013). https://new
atlas.com/amazon-prime-air-delivery-drone/29982/. Accessed 20 Feb 2021
Liu, Z., Li, X., Khojandi, A.: The flying sidekick traveling salesman problem with stochastic travel
time: a reinforcement learning approach. Transp. Res. Part E: Logist. Transp. Rev. 164 (2022).
https://doi.org/10.1016/j.tre.2022.102816
Lei, D., Chen, X.: An improved variable neighborhood search for parallel drone scheduling trav-
eling salesman problem. Appl. Soft Comput. 127 (2022). https://doi.org/10.1016/j.asoc.2022.
109416
Madani, B., Ndiaye, M.: Autonomous vehicles delivery systems classification: introducing a tsp
with a moving depot. In: 2019 8th International Conference on Modeling Simulation and
Applied Optimization (2019). https://doi.org/10.1109/ICMSAO.2019.8880379
Luo, Q., Wu, G., Mao, Ni., Shi, J.: Hybridization of variable neighborhood search and simulated
annealing for the collaborative truck-drone routing problem. In: Wu, M., Niu, Y., Gu, M.,
Cheng, J. (eds.) ICAUS 2021. LNEE, vol. 861, pp. 1505–1515. Springer, Singapore (2022).
https://doi.org/10.1007/978-981-16-9492-9_149
Madani, B., Ndiaye, M.: Hybrid truck-drone delivery systems: a systematic literature review. IEEE
Access 10, 92854–92878 (2022). https://doi.org/10.1109/ACCESS.2022.3202895
Marinelli, M., Caggiani, L., Ottomanelli, M., Dell’Orco, M.: En route truck-drone parcel delivery
for optimal vehicle routing strategies. IET Intell. Transp. Syst. 12(4), 253–261 (2018). https://
doi.org/10.1049/iet-its.2017.0227
Meng, S., Guo, X., Li, D., Liu, G.: The multi-visit drone routing problem for pickup and delivery
services. Transp. Res. Part E: Logist. Transp. Rev. 169, 102990 (2023). https://doi.org/10.1016/
j.tre.2022.102990
Mladenović, N., Hansen, P.: Variable neighborhood search. Comput. Oper. Res. 24(11), 1097–1100
(1997). https://doi.org/10.1016/S0305-0548(97)00031-2
Mogili, U.M.R., Deepak, B.B.V.L.: Review on application of drone systems in precision agri-
culture. Procedia Comput. Sci. 133, 502–509 (2018). https://doi.org/10.1016/j.procs.2018.
07.063
Murray, C.C., Chu, A.G.: The flying sidekick traveling salesman problem: optimization of drone-
assisted parcel delivery. Transp. Res. Part C 54, 86–109 (2015). https://doi.org/10.1016/j.trc.
2015.03.005
Murray, C.C., Raj, R.: The multiple flying sidekicks traveling salesman problem: parcel delivery
with multiple drones. Transp. Res. Part C 110, 368–398 (2020). https://doi.org/10.1016/j.trc.
2019.11.003
Poikonen, S., Golden, B.: Multi-visit drone routing problem. Comput. Oper. Res. 113 (2020a).
https://doi.org/10.1016/j.cor.2019.104802
Poikonen, S., Golden, B.: The mothership and drone routing problem. INFORMS J. Comput. 32,
249–262 (2020b). https://doi.org/10.1287/ijoc.2018.0879
Poikonen, S., Wang, X., Golden, B.: The vehicle routing problem with drones: Extended models
and connections. Networks 70(1), 34–43 (2017). https://doi.org/10.1002/net.21746
Sacramento, D., Pisinger, D., Ropke, S.: An adaptive large neighborhood search metaheuristic for
the vehicle routing problem with drones. Transp. Res. Part C 102, 289–315 (2019). https://doi.
org/10.1016/j.trc.2019.02.018
Salama, M.R., Srinivas, S.: Collaborative truck multi-drone routing and scheduling problem:
package delivery with flexible launch and recovery sites. Transp. Res. Part E 164 (2022).
https://doi.org/10.1016/j.tre.2022.102788
Salhi, S.: Heuristic Search: The Emerging Science of Problem Solving. Springer, Heidelberg
(2017)
334 B. Madani and M. Ndiaye

Schermer, D., Moeini, M., Wendt, O.: A hybrid VNS/Tabu search algorithm for solving the vehicle
routing problem with drones and en route operations. Comput. Oper. Res. 109 (2019). https://
doi.org/10.1016/j.cor.2019.04.021
Scott, J., Scott, C.: Drone delivery models for healthcare. In: Proceedings of the 50th Hawaii
International Conference on System Sciences (2017)
UPS tests drone delivery system. http://www.businessinsider.com/ups-tests-drone-delivery-sys
tem-2017-2. Accessed 15 Mar 2021
Wang, X., Poikonen, S., Golden, B.: The vehicle routing problem with drones: several worst-case
results. Optim. Lett. 11(4), 679–697 (2017). https://doi.org/10.1007/s11590-016-1035-3
Wang, Z., Sheu, J.-B.: Vehicle routing problem with drones. Transp. Res. Part B 122, 350–364
(2019). https://doi.org/10.1016/j.trb.2019.03.005
Windras Mara, S.T., Rifai, A.P., Sopha, B.M.: An adaptive large neighborhood search heuristic
for the flying sidekick traveling salesman problem with multiple drops. Expert Syst. Appl. 205
(2022). https://doi.org/10.1016/j.eswa.2022.117647
Wohlsen, M.: The next big thing you missed: amazon’s delivery drones could work—they just
need trucks. https://www.wired.com/2014/06/the-next-big-thing-you-missed-delivery-drones-
launched-from-trucks-are-the-future-of-shipping/. Accessed 10 Mar 2021
Wood, L.: $91.5 Billion Autonomous Last Mile Delivery Market, 2030. https://www.prnews
wire.com/news-releases/91-5-billion-autonomous-last-mile-delivery-market-2030--300895
603.html. Accessed 10 Mar 2021
Traffic and Transport
 Prediction and Analysis of Transit Ferries
Travel Time: An Open Data-Based Case
Study

Malek Sarhani1(B) , Mohammed El Amrani2 , and Abdelhak El Achhab1

1
School of Business Administration, Al Akhawayn University in Ifrane,
Ifrane, Morocco
[email protected]
2
Faculty of Sciences, Mohammed V University in Rabat, Rabat, Morocco

Abstract. The prediction of transit travel time using machine learning

is an area of growing interest. Unlike previous research that focused
on buses and rail transit systems, our work is focused on ferry transit.
Specifically, the first contribution of this research is to compare machine
learning techniques for predicting travel time in ferry transit. We then
analyze the impact of various factors using different techniques. Our
results indicate that weather significantly impacts travel time, in addition
to the trip starting time and date.

Keywords: Public transport · Machine learning · Ferry transit ·

Travel time prediction · Open data

1 Introduction
Public transportation plays an increasingly important role in modern transporta-
tion systems and has emerged as one of the most crucial modes of transportation
due to its positive environmental and cost impacts. One of the topics that have
attracted considerable attention in recent years is travel time (TT) prediction,
which refers to the estimation of the time required to traverse a road segment
or a route under current or future traffic conditions. TT prediction is essen-
tial for various applications in intelligent transportation systems, such as route
scheduling [5] and hence can lead to provide a better service to the passengers.
Public transport modes can be divided into buses (including trolleybuses),
rail (including metros and trams), and ferries. While most of the research on
delay and travel time predictions for transit focuses on buses and rail services
(as evidenced by the fact that all papers included in [8] are about these modes), it
is important to note that ferry transit is also a significant mode of transportation.
For instance, the Sydney Ferry network in Australia carries an estimated 1.25
million passengers each month, and the Hamburg ferry system in Germany serves
approximately nine million passengers annually. Therefore, providing accurate
estimates of travel time for ferries should also be given importance, just like
buses and rail transit systems.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 337–349, 2023.
https://doi.org/10.1007/978-3-031-43612-3_21
338 M. Sarhani et al.

Analyzing and predicting the travel time of a ferry is a challenging task due to
several factors that can affect it. That is, ferries are ships, watercraft, or amphibi-
ous vehicles that transport passengers, and as such, their travel time can be influ-
enced by a range of factors. Nevertheless, despite the difficulties in predicting ship
travel times, there have been significant advances in this area in recent years.
Indeed, accurate prediction of ship travel time is crucial for the maritime industry
as it helps in obtaining more precise estimations of the estimated time of arrivals.
The latter has a crucial impact on overall planning, as indicated in [20].
In particular, the use of machine learning (ML) in the transportation sec-
tor and maritime industry has gained considerable momentum due to the large
amount of data generated by this transportation mode, and recent advancements
in ML algorithms have made it possible to effectively analyze and make real-time
predictions based on this data [4]. Indeed, ML algorithms such as support vec-
tor machines (SVM), random forests (RF), gradient boosting (GB), and neural
networks (NN) are examples of ML algorithms that can be used to predict vessel
arrival times and address the challenges in this area. These predictions take into
account factors such as voyage information, vessel performance data, and weather
conditions. This information enables organizations to make more informed deci-
sions and reduces uncertainty about the travel time of ships. Additionally, there
have been recent developments in ML, such as the analysis of feature importance,
which aim to improve model interpretability and understand the impact of differ-
ent features.
In this paper, our goal is to compare different ML algorithms to determine the
most accurate model for estimating and predicting the travel time of ferry transit
using various data sources. We will also evaluate the impact of different features
and highlight the most important ones. These factors can have significant impacts
on transit service quality, reliability, efficiency, and customer satisfaction.
The remainder of this paper is structured as follows: In the next section, we
provide a brief literature review of the topic. Section 3 describes the data used
and the various features employed in addition to the prediction approach. The
experiments are presented in Sect. 4. Finally, we conclude with a discussion of
the results and provide future research directions.

2 Literature Review
At the beginning of this section, we would like to note that we have not found
any previous work that focuses on predicting ferry delays or TT, or analyzing
the impact of different factors on these outcomes. To our knowledge, the main
work that focuses on ferries, proposed by [1], is interested in the observed waiting
time for ferry passengers. More information about the waiting time optimization
problem can be found in [6]. [1]’s results indicate a positive relationship between
waiting time and headway. The related literature review to our paper has primar-
ily focused on ship TT prediction and understanding the factors that impact it.
In this brief literature review, we highlight the different factors affecting transit
or ship TT and delays and discuss some prediction methods.
 Prediction and Analysis of Transit Ferries Travel Time 339

First, as indicated above, predicting TT is a complex and dynamic problem

that involves various factors, such as traffic demand, supply, incidents, weather,
and human behavior. Moreover, reliable and timely data from multiple sources,
such as GPS, sensors, cameras, and social media, are required for effective TT
prediction. Disturbances are one of the main factors affecting transit delays, and
many factors can be included in this category [9]. The COVID-19 pandemic,
which greatly impacted public transport ridership and service provision world-
wide, has been the main disturbance over the past few years [10]. However, there
are other disturbances as indicated below. In our case, the other significant fac-
tor is the dwell time, which is the time spent by a transit vehicle at a stop or
station for boarding and alighting passengers. Dwell time can increase travel
time variability and reduce service regularity [12].
There are several studies that examine the factors affecting ship delays and
transit time (TT). Delays are a major concern for shipping operators and port
authorities, and studying TT is crucial to early detect and mitigate them. Ship
delays can have significant economic and environmental consequences. For an
overview of the importance of detecting delays in real-time and predicting vessel
movement patterns until arrival, we refer to [11]. Therefore, it is essential to
understand the causes of delays and find ways to minimize them.
Marine weather is one of the primary factors that affect ship delays and TT. It
can impact vessel speed and route, as well as the availability and safety of berths
in the port. [7] examines ship traffic and weather data from various sources and
found that ship speed is significantly linked to weather variables such as wind
direction, as well as ship characteristics. He developes a linear regression model
that can estimate ship speed based on these variables, which may be beneficial
for predicting arrival times. However, their model did not consider other factors
such as weather, time and route information, which is why we use more advanced
ML techniques in this paper. In fact, there is a need for more data collection
and analysis to determine the impacts of ship delays on costs, efficiency, the
environment, and customer satisfaction. Hence, further research is necessary to
investigate these concerns and propose solutions to minimize ship delays and
improve shipping reliability. Several ML algorithms, including a NN variant,
RF, SVM, and a decision tree method, were utilized in [20] to establish a vessel
arrival time prediction model. The results indicated that RF performed the best.
In terms of TT prediction, various methods and models have been proposed
to overcome the challenges discussed earlier. These include traditional statistical
and machine learning approaches as well as advanced deep learning and data
fusion techniques. A recent survey by [19] provides an overview of the latest
approaches for predicting ship arrival. From the survey, it is evident that ML
methods hold great promise in addressing these challenges. It is worth noting
that similar or closely related problems, such as ship speed prediction [2] and
trajectory prediction, are also being studied. These problems share the same goal
of real-time ship tracking. For instance, [21] employ a NN method to examine the
trajectories of moving maritime vessels and classified them into three categories
using Automatic Identification System data. In addition, [21] utilized another
340 M. Sarhani et al.

NN variant to detect and track multiple vessels based on radar/laser tracking

data.
To summarize, this literature review has examined several factors that con-
tribute to ship delays, including marine weather and dwell time-related infor-
mation. However, further data analysis is necessary to quantify the factors that
influence TT and delay reasons. ML algorithms show great potential, and their
effectiveness varies depending on the case study. In this study, our objective is
to compare various ML algorithms for ferry prediction and assess the impact of
different factors on TT.
It is important to note that when constructing features for predicting ferries’
TT, we must take into account not only ship-related features such as weather and
dwell time but also features related to transit service such as trips and routes. In
the following section, we will demonstrate how we have integrated these various
features.

3 The Proposed Approach

The purpose of this section is to outline the various data used for predicting
transit delay. Firstly, we will describe how the data was collected, followed by
a description of the different features. Finally, we will illustrate our prediction
approach.

3.1 Data Acquisition

In this paper, we take the public transport ferry network serving the city of Syd-
ney in Australia. Figure 1 shows a graphical overview of the line transit.1 In this
study, different data sources are adopted in order to study the possible delay
patterns of the transit and to extract the different factors affecting its delay.
More specifically, we adopt the NSW Open Data Portal (the data is available
at https://opendata.transport.nsw.gov.au/search/type/dataset) which contain
both GTFS static and GTFS real-time. In particular, for GTFS real-time, we
used the free API provided by the website to extract the real-time information.
For each observation, we provide a matching of the real-time information to the
corresponding static information and the weather information about the corre-
spond day. Weather information can be extracted from Wunderground. Table 1
illustrates the different features used in the study. (We note that the features
used for merging, are available in two categories, e.g, GTFS static and GTFS
real-time, we describe these features in the GTFS real-time category). We note
that the dwell time, mentioned above, is not directly incorporated. However,
related factors such as weekdays are included, as the dwell time is dependent on
the number of passengers, which, in turn, is influenced by weekdays and specific
time periods (in our case, specified by the hour of the day). We note that as no
precipitation was recorded in the studied period, it has not been included in the
study.
1
More information on the network is available in https://transportnsw.info/travel-
info/ways-to-get-around/ferry#/.
 Prediction and Analysis of Transit Ferries Travel Time 341

Fig. 1. Map of the Sydney Ferry System

Table 1. Description of the Features Used in the Study

Feature Description
GTFS static Trip ID A specific trip identification
Route ID A route is a group of trips that are displayed to riders as a single service
Service ID Identifies a set of dates when service is available for one or more routes
Shape ID Identifies a geospatial shape that describes the vehicle travel path for a trip
Trip headsign Text that appears on signage identifying the trip’s destination to riders
Direction ID Indicates the direction of travel for a trip
Route direction Planned arrival time at a specific stop for a specific trip
GTFS real-time Vehicle ID Internal identification for the vehicle
Vehicle Route ID Route of the vehicle (used for merging GTFS real-time with GTFS static)
Vehicle Trip ID Trip of the vehicle (used for merging GTFS real-time with GTFS static)
Vehicle Trip Start Time Start time of the trip of the vehicle (used for merging GTFS real-time with GTFS static)
Vehicle Trip Start Date Start date of the trip of the vehicle (used for merging GTFS real-time with weather and GTFS static)
Weekday Weekday of the trip (extracted from the previous feature)
Latitude Latitude for the vehicle position
Longitude Longitude for the vehicle position
Bearing Bearing for the vehicle position
Weather Temperature Average temperature of the day
Dew Point the temperature the air needs to be cooled to in order to achieve a relative humidity
Humidity Average humidity of the day
Wind Speed Average wind speed of the day
Pressure Average pressure of the day

The data used in this study covered the period from March 5, 2023, to April
7, 2023. Specifically, the data encompasses the time between March 5 and March
7, as well as April 4 to April 7. We made an effort to include all possible data
during these days, considering various weekdays and different weather conditions,
342 M. Sarhani et al.

in order to obtain a more comprehensive assessment of these factors. After pre-

processing, the dataset consists of 8,264 instances.

3.2 The Prediction Process

In this part, we briefly describe the predicted output and the algorithms used
for that. First, the travel time is based on the information sent by every vehicle
every 10 s, which results in a huge stockpile of data. These information include
the Timestamp and the other GTFS real-time are depicted in Table 1. in this
study, we decided to normalize the travel time value as in several similar cases
(e.g., [17] and [16]). So, the normalized travel time is chosen as the target variable
to be predicted. Second, to predict the normalized travel time, we adopt Several
ML methods that can produce satisfactory results for regression (continuous)
problems, including SVM, RF, GB and deep neural networks (DNN). These
approaches, which are commonly used for regression problems, are the focus of
our study in this paper. Below, we provide the setting of the algorithms.

4 Experiment
In this section, we define the experimental setup of our case study, then present
and discuss the results.

4.1 Experimentation Setup and Pre-processing

First, to implement the different algorithms, we adopt the Scikit-learn library
[15]. In particular, for DNN, we use Keras 2.2.5 with TensorFlow as the back-end
for its implementation. The DNN model is trained using the Adam optimizer,
and mean absolute error is used as the loss function. The number of epochs
is fixed at 100. For SVM, the parameter γ (kernel coefficient) is automatically
trained while the values of C (regularization factor) and  are set, respectively,
to 10 and 0.01. We have chosen these values to enable a balance between over-
fitting and underfitting. For the other algorithms, we adopt the default values
provided by the Scikit-learn. That is, for both RF and GB models, the following
hyperparameters were used: Number of trees in the forest is set to 100, subsam-
ple is set to 1, minimum samples required for a split is 2, and minimum samples
required to be at a leaf is 1. Additionally, for the GB model, the learning rate is
set to 0.1.
Second, in order to have a fair assessment of the ML prediction, the data must
be divided into a training set and a test set. In our experiment, 80% of the data is
used for training, and the remaining 20% is used as test data. To ensure a better
assessment, we perform the division five times using 5-fold cross-validation, and
we report the average of the results.
All the experiments are carried out on a computer equipped with an Intel
i7-9750H and 16 GB of RAM. The measures adopted in this paper are: R2
(coefficient of determination), the mean absolute error (MAE), the mean square
 Prediction and Analysis of Transit Ferries Travel Time 343

error (MSE) and the root mean square error (RMSE) and, in addition, the mean
absolute percentage error (MAPE).
We will begin by displaying the correlation plot, as depicted in Fig. 2
Based on Fig. 2, we can observe that most features exhibit a moderate or
low correlation (between -0.8 and 0.8), with the exception of weather-related
features such as dew point and precipitation. We will discuss this later.

Fig. 2. Correlation Plot

4.2 Comparison of ML Predictions for the Transit Travel Time

Prediction

First, for the different algorithms, we display in Table 2 a comparison of their

prediction capabilities in the test set. In other words, we show in Table 2 the R2 ,
MAE, RMSE and MAPE values obtained by comparing the algorithm predic-
tions with the real values in the test set.
344 M. Sarhani et al.

Table 2. Comparison of ML Predictions for the Ferry Travel Time Prediction

GB RF DNN SVM
2
R 0.9334 0.9428 0.6924 0.1664
MAE 0.0061 0.0102 0.092 0.429
RMSE 0.008 0.0132 0.163 0.508
MAPE 0.0989 0.106 0.418 0.760

Moreover, to have a better readability of the results, we depict in Fig. 3 the

same values of MAE, RMSE and MAPE obtained by the different algorithms.

Fig. 3. Results for the Different Approaches

We can draw a preliminary conclusion from Table 2 that the predictive ability
of ML algorithms varies significantly. The best results are obtained using GB,
which is slightly better than RF. Their performance is superior to that of SVM
and DNN. It should be noted that these results are specific to the given case
study and the chosen dataset. However, we can conclude that RF and GB are
effective approaches for this case. These methods have also been adopted to
assess the feature importance, as shown below.
 Prediction and Analysis of Transit Ferries Travel Time 345

4.3 Analysis of the Factors Importance

The aim of this section is to analyze the importance of different features and to
identify the most important ones for the prediction process.
First, we present Table 3, which displays the importance of the features
obtained for different algorithms. Since there are multiple ways to measure fea-
ture importance, and it can be algorithm-specific, we have employed three tech-
niques. The first two correspond to Gradient Boosting (GB) and Random Forest
(RF), in addition to a permutation-based feature importance approach. We have
chosen these two algorithms (GB and RF) as they yielded the best results. Fur-
ther information on these techniques can be found in [3]. We have adopted these
techniques to provide an overview of the impact of the various features.
Specifically, Table 3 shows the corresponding relative value for each feature.
We have separated GTFS and weather features with a line in this table.

Table 3. Feature Importance

Feature RF GB Permutation
Vehicle ID 0.00001 0.00001 0.00061
Vehicle Bearing 0.00003 0.00004 0.000105
Trip ID 0.02048 0.01582 0.025763
Trip Start Time 0.21010 0.21140 0.173025
Trip Start Date 0.20701 0.22431 0.18900
Route ID 0.00563 0.00437 0.069827
service ID 0.00000 0.00000 0.00000
shape ID 0.00000 0.00000 0.00000
Trip_headsign 0.00000 0.00000 0.00000
Direction_id 0.00000 0.00000 0.00000
Route_direction 0.00000 0.00000 0.00000
Weekday 0.03302 0.04368 0.04900
Hour 0.04004 0.04769 0.03529
Latitude 0.11721 0.13140 0.11521
Longitude 0.03560 0.04310 0.00000
Temperature 0.06835 0.03690 0.07
Dew Point 0.05735 0.05067 0.06456359
Humidity 0.0813 0.05098 0.0685
Wind Speed 0.18709 0.16185 0.1734
Pressure 0.06582 0.06377 0.06456359

Furthermore, Fig. 4 displays a bar plot of the Shapley values [14] for the 12
most significant features (which correspond to two of the third used features).
Shapley values have gained popularity for this purpose [18]. They represent the
346 M. Sarhani et al.

average marginal contribution of a feature value across all possible combinations

of features. Figure 4 illustrates the mean absolute value of each feature, indicating
their importance under different feature combinations.
Table 3 and Fig. 4 indicate that most of the GTFS static features do not have
a significant impact. Specifically, features such as shape ID, route direction, and
trip headsign Type have little to no impact. This also holds true for certain
GTFS real-time information, such as vehicle ID (though this only applies to
the adopted case study). The most important GTFS real-time features are the
trip start time and trip start date. Additionally, some weather features have
demonstrated a significant impact except precipitation. Wind speed is the most
important weather-related feature in this case study. This is expected as it is
known to affect the travel time of ships and vessels. Other important weather
features include temperature, dew point, humidity, and pressure. We should note
that as shown in Fig. 2, dew point feature is highly correlated to other weather
features. These are the primary findings, and further research should focus on
utilizing this information to gain a better understanding of the factors that
influence travel time.

Fig. 4. Bar plot of the Shapley values

 Prediction and Analysis of Transit Ferries Travel Time 347

5 Conclusion
Travel time prediction for transit has recently become a recurring topic in the
intelligent transportation literature. Transit travel time is influenced by several
factors. This study proposes a systematic approach to predicting ferry transit
travel time using publicly available timetables, real-time transit service feed data
in the GTFS format, and weather data. In this paper, we merged different data
sources to estimate travel time and showed how open data can be used to enhance
the information provided to passengers. We compared various machine learning
techniques for prediction and found that gradient boosting and random forest
are the most suitable methods for the provided case study. Regarding the factors
affecting travel time, the trip time and date were shown to be the most important
features. Additionally, weather features such as wind speed have a significant
impact.
Our proposed approach based on machine learning integrates heterogeneous
data sources and can therefore be useful for a wide range of agencies worldwide.
Thanks to the increasingly standardized data, it can be easily adopted by many
cities.
The findings and accurate travel time predictions can be used for early detec-
tion of delays as in [13]. For this purpose, more information about the planned
travel times is needed. In our data, we relied solely on actual travel times, but
incorporating both actual and planned travel times is necessary for delay analy-
sis. Additionally, travel time prediction can be used to obtain a better estimation
of arrival travel time and improve service, as mentioned earlier [11]. The same
methodology and features can be utilized for predicting both delays and esti-
mated travel times. These topics are gaining increasing interest due to their
importance in improving transit service, and acquiring more data is essential for
addressing them effectively.
Future research should prioritize the acquisition of additional datasets and
the exploration of alternative case studies. In fact, the availability of open data
for ferry services is still limited compared to other transit services. To yield
more comprehensive results, it is advisable to employ longer data periods for a
more thorough evaluation of various factors (e.g., precipitation as no rain was
recorded at that period). Additionally, beyond the findings presented in Fig. 2, it
is essential to gain a deeper understanding of the relationships between different
factors (e.g., whether wind speed has a positive or negative effect on travel time
and to what extent it is linked to the date of the trip). For assessment, other
measures such as adjusted R2 can be used to detect the best combination of
features for the prediction. The goal of this study is to lay the groundwork for
further research in this crucial field.
348 M. Sarhani et al.

References
1. Andersen, S.N., Tørset, T.: Waiting time for ferry services: empirical evidence from
Norway. Case Stud. Transp. Policy 7(3), 667–676 (2019). https://doi.org/10.1016/
j.cstp.2019.04.006
2. Bauer, D., Tulic, M.: Travel time predictions: should one model speeds or travel
times? Eur. Transp. Res. Rev. 10, 1–12 (2018). https://doi.org/10.1186/s12544-
018-0315-7
3. Brownlee, J.: https://machinelearningmastery.com/calculate-feature-importance-
with-python/. Accessed 27 Feb 2022
4. Cammin, P., Sarhani, M., Heilig, L., Voß, S.: Applications of real-time data to
reduce air emissions in maritime ports. In: Marcus, A., Rosenzweig, E. (eds.) HCII
2020. LNCS, vol. 12202, pp. 31–48. Springer, Cham (2020). https://doi.org/10.
1007/978-3-030-49757-6_3
5. Daduna, J.R., Trépanier, M., Voß, S.: From computer-aided transit scheduling to
systems and surveys in public transport. Public Transp. 14(1), 1–3 (2022). https://
doi.org/10.1007/s12469-022-00302-7
6. Daduna, J., Voß, S.: Practical experiences in schedule synchronization. In: Daduna,
J.R., Branco, I., Paixão, J.M.P. (eds.) Computer-Aided Transit Scheduling. LNE,
vol. 430, pp. 39–55. Springer, Heidelberg (1995). https://doi.org/10.1007/978-3-
642-57762-8_4
7. Filtz, E., de la Cerda, E.S., Weber, M., Zirkovits, D.: Factors affecting ocean-going
cargo ship speed and arrival time. In: Persson, A., Stirna, J. (eds.) CAiSE 2015.
LNBIP, vol. 215, pp. 305–316. Springer, Cham (2015). https://doi.org/10.1007/
978-3-319-19243-7_29
8. Ge, L., Sarhani, M., Voß, S., Xie, L.: Review of transit data sources: potentials,
challenges and complementarity. Sustainability 13(20), 11450 (2021). https://doi.
org/10.3390/su132011450
9. Ge, L., Voß, S., Xie, L.: Robustness and disturbances in public transport. Public
Transp. 14(1), 191–261 (2022). https://doi.org/10.1007/s12469-022-00301-8
10. Gkiotsalitis, K., Cats, O.: Public transport planning adaption under the COVID-
19 pandemic crisis: literature review of research needs and directions. Transp. Rev.
41(3), 374–392 (2020). https://doi.org/10.1080/01441647.2020.1857886
11. Heilig, L., Stahlbock, R., Voß, S.: From digitalization to data-driven decision mak-
ing in container terminals. In: Böse, J.W. (ed.) Handbook of Terminal Planning.
ORSIS, pp. 125–154. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-
39990-0_6
12. Hörcher, D., Graham, D.J., Anderson, R.J.: Crowding cost estimation with large
scale smart card and vehicle location data. Transp. Res. Part B: Methodol. 95,
105–125 (2017). https://doi.org/10.1016/j.trb.2016.10.015
13. Kim, S., Kim, H., Park, Y.: Early detection of vessel delays using combined histor-
ical and real-time information. J. Oper. Res. Soc. 68(2), 182–191 (2017). https://
doi.org/10.1057/s41274-016-0104-4
14. Marcílio-Jr, W.E., Eler, D.M.: Explaining dimensionality reduction results using
Shapley values. Expert Syst. Appl. 178, 115020 (2021). https://doi.org/10.1016/
j.eswa.2021.115020
15. Pedregosa, F., et al.: Scikit-learn: machine learning in Python. J. Mach. Learn.
Res. 12, 2825–2830 (2011)
16. Petersen, N.C., Rodrigues, F., Pereira, F.C.: Multi-output bus travel time predic-
tion with convolutional LSTM neural network. Expert Syst. Appl. 120, 426–435
(2019). https://doi.org/10.1016/j.eswa.2018.11.028
 Prediction and Analysis of Transit Ferries Travel Time 349

17. Taghipour, H., Parsa, A.B., Mohammadian, A.K.: A dynamic approach to pre-
dict travel time in real time using data driven techniques and comprehensive
data sources. Transp. Eng. 2, 100025 (2020). https://doi.org/10.1016/j.treng.2020.
100025
18. Wagner, F., et al.: Using explainable machine learning to understand how urban
form shapes sustainable mobility. Transp. Res. Part D: Transp. Environ. 111,
103442 (2022). https://doi.org/10.1016/j.trd.2022.103442
19. Yan, R., Wang, S., Zhen, L., Laporte, G.: Emerging approaches applied to mar-
itime transport research: past and future. Commun. Transp. Res. 1, 100011 (2021).
https://doi.org/10.1016/j.commtr.2021.100011
20. Yu, J., Voß, S.: Towards just-in-time arrival for container ships by the integration
of prediction models. In: 56th Hawaii International Conference on System Sciences
(2023). https://hdl.handle.net/10125/102778
21. Zissis, D., Xidias, E.K., Lekkas, D.: Real-time vessel behavior prediction. Evolving
Syst. 7, 29–40 (2016). https://doi.org/10.1007/s12530-015-9133-5
 A Bi-Objective Column Generation
Approach for Real-World Rolling Stock
Circulation Planning Problems

Paul Päprer(B) , Janis S. Neufeld, and Udo Buscher

Faculty of Business and Economics, TUD Dresden University of Technology,

01069 Dresden, Germany
{paul.paeprer,janis_sebastian.neufeld,udo.buscher}@tu-dresden.de
https://tu-dresden.de/bu/wirtschaft/bwl/lim

Abstract. To make the planning in rail transport more efficient, this

work deals with a real-world Rolling Stock Circulation Problem. In this
study, sequences of trips and empty runs are formed for each traction
unit to cover all scheduled trips. Practical restrictions for station-wise
balanced planning are integrated into a Set Covering Problem formula-
tion which is solved with a column generation approach. The decision
makers give two objectives, the number of traction units and empty run
kilometers. They have different priorities at different planning stages,
and it is hard for the decision makers to quantify the cost of one trac-
tion unit or one empty run kilometer. A bi-objective column generation
approach is built by integrating the epsilon constraint method, which is
recognized as a classical method to handle multi-objective optimization
problems. To evaluate the relation between both criteria, the algorithm
is tested using real-world use cases. The generated circulation plans are
presented in solution fronts, illustrating the targets’ mutual influence.
The identified trade-off between fewer traction units or fewer empty run
kilometers can serve as decision support for planners in railway systems.

Keywords: Railways · Column Generation · Bi-objective

1 Introduction

In railway systems, the Rolling Stock Circulation Problem (RSCP) is a cru-

cial step of the sequential planning cascade. It takes place after line planning,
train timetabling and train platforming and is followed by train unit shunting,
crew scheduling and crew rostering. Consequently, it not only allocates the most
expensive resources of railway companies but also has a significant influence on
further planning stages [9]. In the RSCP, the type and number of railway vehicles
for every scheduled trip are defined. Therefore, the outcome of this task is given
by circulations, i.e., sequences of trips and empty runs, for all railway vehicles
used [3].

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 350–364, 2023.
https://doi.org/10.1007/978-3-031-43612-3_22
 Column Generation Approach for Circulation Planning 351

In real-world situations, decision makers often aim to minimize both the use
of traction units and empty run kilometers, as they are major cost drivers. How-
ever, it is generally difficult to assign specific cost values to both aspects, which
are necessary when integrating them into one objective function. This is due
to the fact that they have different priorities at different planning stages. The
contribution of this work is the development of a bi-objective column generation
approach for the RSCP, whose solution front allows an evaluation of the rela-
tion between the two objectives and enables the decision makers to choose the
currently most suitable circulation plan.
The remainder of this paper is organized as follows. Section 2 provides a
brief literature review focusing on different models and approaches for RSCPs.
Section 3 describes the real-world problem and its particularities verbally and
mathematically. The bi-objective column generation approach is introduced in
Sect. 4, and the results for real-world use cases are presented in Sect. 5. Finally,
Sect. 6 concludes this work and gives a summary and possibilities for future
research.

2 Literature Review

Many authors presented publications on the topic of RSCPs. First, the articles
can be distinguished by the transport mode in rail passenger transport, e.g., [2],
and rail freight transport, e.g., [13]. The next possible differentiation is based
on which kind of railway vehicle the particular authors study. [13] deal with
the assignment of locomotives to scheduled trips, [2] consider railcars, and [7]
examine the routing of carriages.
An RSCP is usually formulated as a mixed integer problem. Most publica-
tions provide an arc-flow formulation in a time-space graph, e.g., [2]. Another
option for modeling an RSCP is a Set Covering Problem (SCP) formulation as
in [10]. The third formulation can be referred to as hyperarc-flow formulation in
a hypergraph [11]. The difference from regular graphs is that in hypergraphs, an
edge can join any number of nodes, not only two.
In order to generate optimal circulations for railway vehicles, appropriate
evaluation criteria for the solutions must be established in advance. Most objec-
tive functions target the efficiency of the formed plans as in [2]. For example,
the total costs incurred from allocating the vehicles, the driven kilometers, or
the number of necessary railway vehicles are minimized. Other frequently used
objectives can be combined into the quality of service, e.g., [7]. An example of
this is the minimization of kilometers with seat shortages. The third group of
criteria is intended to ensure highly robust solutions. For example, [12] inter-
actively use optimization and simulation to focus on robustness. The criteria
considered in this paper are the number of traction units and the number of
empty run kilometers. Both are evaluated in [6] with the difference that costs
are given to the criteria to combine them in one objective function. However, in
practice this is often impossible since the decision makers can hardly quantify
those values and thus want different solutions with various ratios between both
352 P. Päprer et al.

objectives. This relationship has not been extensively studied in the literature
so far.
Because there is an enormous diversity in the practical requirements of the
particular use case, the applied methods to obtain reasonable solutions differ
strongly. As a result, they can be distinguished into integer programming meth-
ods [5], heuristics [2], and column generation [1].

3 Problem Description
3.1 Real-World Rolling Stock Circulation Planning

The studied problem is based on a practical RSCP from the Austrian Federal
Railways (ÖBB). The first sub-task of rolling stock circulation planning, i.e., the
rolling stock assignment, is already given. This means that the traction unit type
for every trip is known. Only the second sub-task, the train routing, needs to be
solved here. Consequently, circulations must be created for a homogeneous set
of vehicles to cover all scheduled trips requiring this type. The focus lies hereby
on traction units, so carriages are not considered.
The railway network and a train schedule for a specific planning horizon are
given and form the basis for the modeling. Every trip from the schedule has an
origin station, a destination station, a start time, and an end time. The traction
units are allowed to perform empty runs. These are unproductive trips that do
not appear in the timetable and only serve to transfer a traction unit between
two locations. To calculate the duration and distance for possible empty runs,
the duration and distance between specific network points are used. Based on
these fundamentals, a verbal formulation of the real-world circulation problem
is given below, showing the objective function and relevant constraints.
There are two objectives when performing the circulation planning for the
ÖBB. Besides the number of used traction units, minimizing empty run kilo-
meters is important. Since the traction units are the most expensive resources
of a railway company, purchasing new vehicles represents considerable acquisi-
tion costs. Therefore, circulation plans should strive for maximum efficiency in
using the existing traction units, with the aim of minimizing the total number of
units required. As the second part of the objective function, a small number of
empty run kilometers is desired. Because traction unit drivers are an expensive
resource, too, the ÖBB aims their drivers to perform as many productive, i.e.,
planned, trips as possible. Empty runs only cost money but do not generate any
profit, so they should be avoided whenever possible.
There are several practical requirements in the real-world circulation prob-
lem. First, all trips from the previously planned timetable must be covered by
at least one traction unit. Second, the planners aim at a round or cyclic plan-
ning solution. This means, in every station, the number of traction units at the
beginning must equal the number of traction units at the end. Such a station-
wise balanced circulation plan automatically puts the whole system back in its
initial state. It is possible to perform the same circulation plan again in the next
 Column Generation Approach for Circulation Planning 353

period. A third restriction limits the available traction units. Finally, all circula-
tions must be spatially and temporally feasible. This means that after covering
one trip, the next trip has to start later than the arrival of the previous one.
Moreover, if the next trip has a different origin station, the time between both
trips must be greater than the empty run needed to come to this station.

3.2 Mathematical Formulation

As shown in Sect. 2, an RSCP can be formulated as an SCP. The model con-
structed here is developed following existing approaches like in [10] and adapted
to the specific real-world problem. The notation is given in Table 1.

Table 1. Notation for mathematical formulation

Type Symbol Description

Sets P set of possible circulations p
T set of scheduled trips t
S set of stations s
Parameters cp cost of circulation p
ap,t 1 if circulation p covers trip t; 0 otherwise
bp,s 1 if station s is used by circulation p as origin station;
−1 if used as destination; 0 if used as both or otherwise
n number of available traction units
Variables xp binary variable; 1 if circulation p is used; 0 otherwise

The basic idea of the depicted SCP formulation is to have a set T of the
trips to be covered and another set P of feasible circulations. The aim is now
to identify the smallest or cost-minimal sub-collection of the circulation set to
cover all elements of the trip set.


min cp · xp (1)
p∈P

subject to ap,t · xp ≥ 1 ∀t ∈ T (2)
p∈P

bp,s · xp = 0 ∀s ∈ S (3)
p∈P

xp ≤ n (4)
p∈P

xp ∈ {0, 1} ∀p ∈ P (5)
Equation (1) is the objective function of the model formulation and minimizes
the total cost. This happens by summing up the cost of all used circulations.
354 P. Päprer et al.

Note that the term cost can represent both objectives and is defined explicitly
in Sect. 4. Constraints (2) use the parameter ap,t to guarantee that every trip is
assigned to at least one circulation. At this point, the character of the problem
in the sense of an SCP formulation becomes clear because there is the possibility
of over-covering. The constraints (3) ensure a station-wise balanced rolling stock
circulation planning. These restrictions count the circulations which start (+1)
and/or end (−1) at the specific station and provide a balanced overall plan
(balance = 0). The methodology employed in this paper for modeling station-
wise balanced plans draws from [1]. Constraint (4) ensures that the number of
used traction units does not exceed the number of available ones. Finally, the
decision variables have a binary character, as modeled in Eqs. (5). With this 0–1
integer linear programming model, the problem can be solved optimally when
the set of all feasible circulations is built before. Since this is inefficient for large
instances, a column generation procedure is applied.

4 Solution Approach
4.1 General Procedure
To solve the RSCP, column generation is used. The basic procedure of this
solution method is described in [4]. It consists of a constructive heuristic, a
Master Problem (MP), a Pricing Problem (PP), and an integrality step. The
general process of the particular algorithm is depicted in Fig. 1. The picture
shows the interaction of the four critical components in the algorithm.

Fig. 1. Procedure of the applied column generation

The MP selects the most efficient combinations of columns, i.e., circulations,

from a given circulation pool. For this purpose, it is necessary to have an initial
pool of circulations that can lead to a feasible solution. Therefore, a constructive
heuristic is presented, which builds an initial solution. After choosing the best
 Column Generation Approach for Circulation Planning 355

circulations from the current circulation pool in the MP, the PP tries to construct
new improving columns, i.e., circulations.
The applied column generation terminates if one of the following conditions
is met: the best-found circulation in the PP has non-negative reduced cost, or
the total number of 200 iterations is reached, or the objective value of the MP is
improved less than 0.1 % in 20 consecutive iterations. Since the MP works with
continuous variables, it is necessary to convert the found solution into an integer
one after that.

4.2 Constructive Heuristic

The first step of the process of generating a starting solution is that all trips
are sorted by their start times in ascending order, resulting in a list T ordered . In
addition, for each station s a sorted list Tsordered is created in the same way, only
considering the trips departing at s. Afterward, the construction of circulations
can begin. For this purpose, the first trip t of T ordered is placed at the beginning of
a new circulation p and removed from all trip lists. Now, the algorithm attempts
to fill the current circulation with as many trips as possible. To search for suitable
trips, the algorithm looks at the list of trips starting from the destination station
of the current trip t. The first found trip u here, whose start time is greater than
or equal to the current trip t’s end time, is added to p. This procedure ensures
that the resulting circulations are feasible in time and space. As explained above,
every newly added trip is removed from the overall trip list and the ordered
trip list of its origin station. If no additional trip can be added to the current
circulation p, it is closed and added to the circulation plan P . After that, the
algorithm starts with a new iteration, putting the first trip from the remaining
ones in T ordered at the beginning of a new circulation. If T ordered is empty, the
algorithm terminates, and the solution is given by the overall circulation plan
P , consisting of all built circulations.

4.3 Master Problem

The mathematical formulation for the MP is an adjusted version of the SCP for-
mulation discussed Sect. 3.2 and is depicted below. First, the equality constraint
is split into two inequality constraints, (8) and (9), being more efficient and
suited for column generation. Second, the solution of the constructive heuristic
will always cover all trips but may not balance all stations or keep the available
number of traction units. To have a feasible solution, balancing and capacity
constraints (8)-(10) get auxiliary variables us , uv , and w, converting them from
hard to soft constraints. Exceeding these constraints is tolerated but punished in
the objective function (6) by multiplying the auxiliary variables with a penalty
cost parameter cpunish . A comparatively high-cost parameter will make the col-
umn generation seek true feasibility before optimizing the solution. In addition
to these structural changes, two variable adoptions must be done. First, the
Restricted Master Problem (RMP) is defined with the set of columns being a
subset P  of all possible columns, i.e., circulations. The second step is to build
356 P. Päprer et al.

the linear relaxation. Therefore, the integrality constraints for xp , us , vs , and w

are removed. They can be positive real numbers, yielding the following relaxed
Restricted Master Problem (rRMP). To obtain the necessary integer solution in
the end, the variables can be set to integer (us , vs , w) and binary (xp ) before
solving the RMP model.

 
 
min cp · xp + (us + vs ) + w · cpunish (6)
p∈P  s∈S

subject to ap,t · xp ≥ 1 ∀t ∈ T (7)
p∈P 

bp,s · xp ≤ 0 + us ∀s ∈ S (8)
p∈P 

bp,s · xp ≥ 0 − vs ∀s ∈ S (9)
p∈P 

xp ≤ n + w (10)
p∈P 

xp ∈ R+
0 ∀p ∈ P  (11)
us , vs ∈ R+
0 ∀s ∈ S (12)
w∈ R+
0 (13)

4.4 Pricing Problem

After choosing the best circulations from the current circulation pool in the MP,
the PP tries to construct new improving columns, i.e., circulations. The reduced
cost of new variables evaluates their potential to improve the objective value.
Therefore, dual values are required, which arise during the solution of the rRMP.
Table 2 presents the dual values and to which restriction they belong.

Table 2. Dual values and their associated restrictions

Dual values Associated constraints Equation

πt covering constraints (7)
σsl less-than-or-equal station balancing constraints (8)
σsg greater-than-or-equal station balancing constraints (9)
λ capacity of traction units (10)

Given this information, the equation for the reduced cost of a variable can be
derived. The formula always starts with the cost of a new variable. Besides this
term, every restriction of the rRMP appears in the equation by its calculated
dual value and the decision variable coefficient. Note that parameters ap,t and
 Column Generation Approach for Circulation Planning 357

bp,s from the MP are now variables. Now, the PP aims to find the variable with
the lowest reduced cost. Consequently, the equation for the reduced cost has to
be formulated as follows.

  
cp = cp − πt · ap,t − σsl · bp,s − σsg · bp,s − λ (14)
t∈T s∈S s∈S

For constructing new circulations, we aim to build an exact subproblem.

Thus, a new program with new constraints must be set up. The basis for this
exact PP model formulation is a directed graph, i.e., digraph G = (V, E). The
set of nodes V is given by all scheduled trips T merged with an artificial start (i)
and end (j) node. When building the set of edges E for the digraph, two trips
are connected if one of the following two conditions is fulfilled:
1. Two trips can be connected without an empty run if the first’s destination
station is equal to the second’s origin station and the second’s start time is
greater or equal to the first trip’s end time.
2. Two trips with unequal destination station of the first and origin station of the
second can be connected by an empty run. Therefore, the difference between
the first trip’s end time and the second trip’s start time has to be greater
than or equal to the duration of an empty run between these stations.
This setting allows empty runs between two trips. However, to meet the
balancing constraints, it can be necessary to insert empty runs at the beginning
or end of the circulation, too. For this purpose, two empty run nodes are created
for all possible connections between two stations. One can be used right after
the source node and one right before the sink node. Consequently, a set of empty
run nodes R is added to the set of nodes V .

Fig. 2. Example of a representation of trips in a network

The connection between an empty run start node and a trip node is only
possible when the empty run’s destination station is the same as the trip’s origin
358 P. Päprer et al.

station. Analogously, a trip can only be followed by an empty run end node when
the trip’s destination station equals the empty run’s origin station. Furthermore,
the trip’s start and end times must be considered when connecting empty run
nodes and trips. Sometimes a trip starts right after the beginning of the planning
period. In this case, inserting an empty run before the trip is impossible. The
network representation resulting from this process is depicted in Fig. 2.
Now, the PP can be formulated as a shortest path problem [10], where the
goal is to find a path from source i to sink j that generates the lowest reduced
cost. For this purpose, a variable xt,t indicates whether node t is followed by node
t . This variable is created for every directed edge of the network. Additionally,
some other sets of nodes are needed. V s is the subset of the node’s successors
in the network, and V p is the subset of the node’s predecessors. W combines
the scheduled trips T and the empty runs R. The following constraints of the
respective shortest path problem can find the cost-minimal circulation, i.e., the
circulation with the lowest reduced cost.

min cp (15)
⎧
⎪
⎨ 1 if t = i
 
subject to xt,t − xt ,t = 0 if t ∈ W ∀t ∈ V (16)
⎪
⎩
t ∈V s t ∈V p −1 if t = j

xt,t = ap,t ∀t ∈ T (17)
t ∈V s
 
xi,t − xt,j = bp,s ∀s ∈ S (18)
t∈Wsori t∈Wsdes

xt,t ∈ {0, 1} ∀(t, t ) ∈ E (19)

ap,t ∈ {0, 1} ∀t ∈ T (20)
bp,s ∈ {−1, 0, 1} ∀s ∈ S (21)

The first group of constraints (16) are the typical flow conservation con-
straints of the shortest path problem. Only one outgoing arc is allowed for the
source, and there will be exactly one ingoing arc for the sink. All other nodes
must have an equal number of ingoing and outgoing arcs. The next two con-
straints (17) and (18) are the linking constraints between the variables xt,t ,
ap,t , and bp,s . If an arc is used, the respective trip is part of the circulation, and
hence, ap,t needs to be 1 (17). For defining the bp,s variable, it is crucial to deter-
mine whether a station is used as an origin and/or destination station. For this
purpose, a set of all trips or empty runs starting at the particular station (Wsori )
and ending there (Wsdes ) is formed. With the help of these two sets, the balance
for every station can be calculated, which is then equal to the bp,s variable of the
station s (18). The corresponding circulation is added to the circulation pool if
the minimum reduced cost is negative.
 Column Generation Approach for Circulation Planning 359

4.5 Bi-Objective Column Generation

Because two objective criteria are considered for the circulation plans, two col-
umn generation variants are created, each focusing on a different objective.

– Column generation CG-TU: minimize traction units

– Column generation CG-ER: minimize empty run kilometers

They differ in formulating the objective function of the MP and the formu-
lation of the reduced cost for the PP. We apply the ε-constraint method to the
column generation approach to provide a front of different solutions. The general
procedure of that method can be found in [8]. The developed method is depicted

Fig. 3. Bi-objective column generation

360 P. Päprer et al.

in Fig. 3. It starts with two independent runs of both column generation variants.
CG-TU will minimize the number of traction units T Umin leading to an upper
bound of empty run kilometers ERmax because of neglecting this target. In con-
trast, CG-ER minimizes the number of empty run kilometers ERmin , providing
an upper bound for the used traction units T Umax . Between those two extreme
points, other solutions are calculated. Given two objective functions, one has
to be chosen as the single objective making the other a constraint. Solving the
column generation variant CG-TU is much more time-consuming because many
solutions have the same objective value, i.e., number of traction units. For this
reason, the second, more runtime-efficient column generation variant, CG-ER, is
taken. The other objective of minimizing the number of traction units becomes
a constraint, which is already part of the model as constraint (10). Starting from
the CG-TU solution, CG-ER is run. The right-hand side of the capacity con-
straint starts with the minimal number of traction units T Umin and is increased
step by step until the highest number of traction units, i.e., the one of the inde-
pendent CG-ER run T Umax , is reached. This means, the algorithm can use more
traction units, gradually reducing the number of empty run kilometers.

5 Computational Results

To test the column generation, it is applied to two real-world use cases given by
the ÖBB (see Table 3). Some descriptive parameters are observed there to get an
overview of the use cases. Furthermore, the graph in the bottom row visualizes
the use case, where every node belongs to one station. The edges represent the
scheduled trips in the style of an undirected graph. The thicker the line, the
more often a trip occurs over the planning horizon. For this purpose, the trips
are sorted into five classes whose borders are set by quantiles of the absolute
frequencies.
The linear and integer programming formulations of the RSCP are imple-
mented with Gurobi 10.0.1 using its C# API. C# is also used for the implemen-
tation of the proposed constructive heuristic and the column generation frame-
work. All tests are run on an AMD EPYCTM 7513 32-Core Processor using up
to 24 physical cores and 128 GB RAM. The relevant parameters for the column
generation algorithm are set to the following values: cpunish is 100,000, and n is
500.
To evaluate the proposed column generation approach, a distinction is made
between the results of the constructive heuristic, CG-TU, and CG-ER. For all
three solution types as well as both use cases, Table 4 shows the number of
traction units, the empty run kilometers, the best rRMP value, and the run-
time in seconds. The number of infeasibilities for the constructive heuristic only
refers to the station balancing constraints and the capacity of traction units. An
overused traction unit is one infeasibility. In addition, whenever there is a differ-
ence between departing and ending traction units for a station, this difference
is added to the number of infeasibilities. The best rRMP value represents the
best real solution before setting all variables to binary or integer. Consequently,
 Column Generation Approach for Circulation Planning 361

Table 3. Structure of the given use cases

Use case A Use case B

freight transport passenger transport
200 trips 1318 trips
38 stations 61 stations

it refers to the traction unit number in CG-TU and empty run kilometers in
CG-ER.

Table 4. Results of column generation for both use cases

Use case A Use case B

Constructive heuristic Traction units 89 72
Empty run kilometers 0 0
Infeasibilities 156 54
Runtime [s] 0.01 0.01
CG-TU Traction units 25 76
Empty run kilometers 77,306 91,018
rRMP objective value 15.06 53.16
Runtime [s] 307.23 6,952.91
CG-ER Traction units 52 84
Empty run kilometers 8,372 1,096
rRMP objective value 6,678.04 661.87
Runtime [s] 130.12 3,375.97

As explained in Subsect. 4.5, the ε-constraint method is adjusted and applied

to column generation. This well-known multi-objective optimization procedure
will create solutions that are efficient with respect to both objectives of the
362 P. Päprer et al.

real-world problem, the number of traction units and the number of empty run
kilometers. This means no weights are given to the objectives to obtain one single
best solution but a set of non-dominated solutions are determined. Thereby, a
solution S1 dominates another solution S2 if it is not worse than S2 in both
objectives and better in at least one objective [14]. Figure 4 visualizes the ten
non-dominated solutions obtained by our method for use case A. It can be seen
that a hyperbolic solution front is created. In the top left corner is a solution with
a small number of traction units but many empty run kilometers. The opposite
can be found in the bottom right corner, a solution with very low empty run
kilometers and many traction units.

Fig. 4. Solution front of use case A

The assumption that the two objective criteria influence each other can be
strengthened. The improvement of one criterion leads, ceteris paribus, to a dete-
rioration of the other criterion. If up to 44 % more traction units are gradually
deployed, the number of empty kilometers can be progressively reduced to 29 %.
Now, it is up to the decision maker to decide which solution he/she chooses.
The same procedure of multi-objective column generation is applied to the
large use case B. The resulting front of the solutions is shown in Fig. 5. Interest-
ingly, allowing more traction units substantially decreases empty run kilometers
between 73 and 79 traction units. The decision maker should consider using
more traction units here to take the sizeable empty run advantages. Using more
than 79 traction units also improves the empty run kilometers. However, the
improvements are significantly smaller compared to the initial savings.
The figures illustrate the benefits of the approach presented. It helps to cre-
ate solution fronts and allows an analysis of the relationship between the two
objectives. This is very important for decision makers: they can focus on one
direction or the other at different planning stages.
 Column Generation Approach for Circulation Planning 363

Fig. 5. Solution front of use case B

6 Conclusion and Future Research

This work presents a bi-objective column generation method to solve a real-world
Rolling Stock Circulation Problem. The formulation is obtained by integrating
a station balancing constraint in the classical Set Covering formulation for this
problem. The ε-constraint method is applied to column generation to provide
solutions with different quality regarding both objective criteria. The analysis
reinforces the assumption that the two targets compete with each other. Hyper-
bolic solution fronts support this relationship.
Possible developments would be in the field of accelerating the solution pro-
cess. It is necessary to find ways to solve significantly larger instances in a rea-
sonable time, e.g., by solving the Pricing Problem heuristically or applying a
fixing heuristic to reach the integer solution. In addition, the bi-objective col-
umn generation approach could also be an interesting approach to identify and
examine other conflicting goals in rail transport or completely different fields.

Acknowledgements. This work was funded in the course of the project VIPES (FFG
project number 893963) by the Federal Ministry for Climate Protection, Environment,
Energy, Mobility, Innovation and Technology (BMK) as part of the call for proposals
Mobilität der Zukunft. FFG is the central national funding organization and strength-
ens Austria’s innovative power. We also want to thank Matthias Wastian, Senior Data
Scientist at dwh GmbH, for providing all the necessary information to formulate the
model and use cases to run the algorithm.

References
1. Bach, L., Gendreau, M., Wøhlk, S.: Freight railway operator timetabling and engine
scheduling. Eur. J. Oper. Res. 241(2), 309–319 (2015). https://doi.org/10.1016/j.
ejor.2014.08.036
364 P. Päprer et al.

2. Cacchiani, V., Caprara, A., Toth, P.: An effective peak period heuristic for rail-
way rolling stock planning. Transp. Sci. 53(3), 746–762 (2019). https://doi.org/10.
1287/trsc.2018.0858
3. Caprara, A., Kroon, L., Monaci, M., Peeters, M., Toth, P.: Passenger railway opti-
mization. In: Handbooks in Operations Research and Management Science. vol.
14, pp. 129–187 (2007). https://doi.org/10.1016/S0927-0507(06)14003-7
4. Desrosiers, J., Lübbecke, M.E.: A primer in column generation. In: Desaulniers,
G., Desrosiers, J., Solomon, M.M. (eds.) Column Generation, pp. 1–32. Springer,
Boston (2005). https://doi.org/10.1007/0-387-25486-2_1
5. Fioole, P.J., Kroon, L., Maróti, G., Schrijver, A.: A rolling stock circulation model
for combining and splitting of passenger trains. Eur. J. Oper. Res. 174(2), 1281–
1297 (2006). https://doi.org/10.1016/j.ejor.2005.03.032
6. Frisch, S., Hungerländer, P., Jellen, A., Primas, B., Steininger, S., Weinberger,
D.: Solving a real-world locomotive scheduling problem with maintenance con-
straints. Transp. Res. Part B: Methodological 150, 386–409 (2021). https://doi.
org/10.1016/j.trb.2021.06.017
7. Lingaya, N., Cordeau, J.F., Desaulniers, G., Desrosiers, J., Soumis, F.: Operational
car assignment at via rail Canada. Transp. Res. Part B: Methodological 36(9),
755–778 (2002). https://doi.org/10.1016/S0191-2615(01)00027-3
8. Mavrotas, G.: Effective implementation of the ε-constraint method in multi-
objective mathematical programming problems. Appl. Math. Comput. 213(2),
455–465 (2009). https://doi.org/10.1016/j.amc.2009.03.037
9. Neufeld, J.S., Scheffler, M., Tamke, F., Hoffmann, K., Buscher, U.: An efficient
column generation approach for practical railway crew scheduling with attendance
rates. Eur. J. Oper. Res. 293(3), 1113–1130 (2021). https://doi.org/10.1016/j.ejor.
2020.12.058
10. Nishi, T., Ohno, A., Inuiguchi, M., Takahashi, S., Ueda, K.: A combined column
generation and heuristics for railway short-term rolling stock planning with regular
inspection constraints. Comput. Oper. Res. 81, 14–25 (2017). https://doi.org/10.
1016/j.cor.2016.11.025
11. Reuther, M., Schlechte, T.: Optimization of rolling stock rotations. In: Borndörfer,
R., Klug, T., Lamorgese, L., Mannino, C., Reuther, M., Schlechte, T. (eds.) Hand-
book of Optimization in the Railway Industry. ISORMS, vol. 268, pp. 213–241.
Springer, Cham (2018). https://doi.org/10.1007/978-3-319-72153-8_10
12. Rößler, M., et al.: Simulation and optimization of traction unit circulations. In:
2020 Winter Simulation Conference (WSC), pp. 90–101. IEEE (2020). https://doi.
org/10.1109/WSC48552.2020.9383926
13. Scheffler, M., Neufeld, J.S., Hölscher, M.: An MIP-based heuristic solution app-
roach for the locomotive assignment problem focussing on (dis-) connecting pro-
cesses. Transp. Res. Part B: Methodological 139, 64–80 (2020). https://doi.org/
10.1016/j.trb.2020.05.020
14. Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C.M., Da Fonseca, V.G.: Per-
formance assessment of multiobjective optimizers: an analysis and review. IEEE
Trans. Evol. Comput. 7(2), 117–132 (2003). https://doi.org/10.1109/TEVC.2003.
810758
 An Effective Matheuristic Approach
for Robust Bus Driver Rostering
with Uncertain Daily Working Hours

Abtin Nourmohammadzadeh(B) and Stefan Voß

Institute of Information Systems, University of Hamburg, Hamburg, Germany

{abtin.nourmohammadzadeh,stefan.voss}@uni-hamburg.de

Abstract. The bus driver rostering problem (BDRP) is a significant

problem in public transportation that aims at assigning drivers to bus
routes over a given time period while reducing system costs as well as
inconveniences. In this work, a mixed-integer linear mathematical model
is presented for the problem with the objective of minimizing the total
assignment and driver dissatisfaction costs. Uncertainty is considered for
the maximum daily working hours of drivers, which appears on the right-
hand side of a block of constraints. A set of scenarios regarding the uncer-
tain parameters and robust optimization are considered. A matheuris-
tic approach is proposed consisting of a hybridization of particle swarm
optimization (PSO), simulated annealing (SA) and mathematical pro-
gramming of sub-problems based on the concept of partial optimization
metaheuristic under special intensification conditions (POPMUSIC). It
is shown that the proposed matheuristic is an effective approach for
solving the BDRP by examining its performance on a set of benchmark
instances.

Keywords: Bus driver rostering · Matheuristics · Partial optimization

metaheuristic under special intensification conditions (POPMUSIC) ·
Particle swarm optimization (PSO)

1 Introduction
The driver rostering problem (DRP) is a critical problem in the transporta-
tion industry, particularly for bus companies that must efficiently create sched-
ules for their drivers. The problem involves assigning drivers to shifts or duties
over a specific period, such as a week or a month, while satisfying various con-
straints, including the availability of drivers, work regulations, and customer ser-
vice requirements. The DRP is a complex and challenging optimization problem
that has received significant attention from researchers in the fields of operations
research, computer science, and transportation.
In recent years, researchers have developed various approaches to solve the
DRP, including exact methods, heuristics, and metaheuristics. Exact methods
are designed to solve the DRP to optimality but are computationally expensive
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 365–380, 2023.
https://doi.org/10.1007/978-3-031-43612-3_23
366 A. Nourmohammadzadeh and S. Voß

and may not be practical for large-scale problems. Heuristics are fast and effi-
cient but do not guarantee optimal solutions. Metaheuristics are a compromise
between exact methods and heuristics, offering the ability to find near-optimal
solutions in a reasonable amount of time. Metaheuristics are particularly useful
for solving large-scale problems with complex constraints. Besides these methods,
combining elements of mathematical programming techniques with heuristics or
metaheuristics has shown promise under the acronym matheuristics [15,16].
Matheuristics are able to find high-quality solutions in a relatively short
amount of time, making them a popular choice for real-world applications where
time and computational resources are limited. They are particularly effective
when traditional optimization algorithms fail to find a solution or when a solution
is required within a tight time frame. These methods are also highly adaptable
and can be tailored to specific problem structures, making them a flexible tool
for a wide range of optimization problems.
There are several related works which investigate the bus driver rostering
problem (BDRP) with different assumptions and objectives. [21] presents a case
study on the driver rostering problem for a Finnish bus transportation com-
pany. The authors use a mathematical programming model to create an initial
solution, and then applied a local search heuristic to improve the solution. The
results show that the proposed approach significantly improves the quality of
the rosters compared to the manual process used by the company. [20] proposes
a bi-objective evolutionary heuristic to the problem. The proposed method con-
siders two objectives: minimizing the number of drivers needed and minimizing
the number of working hours. The authors tested their approach on real-world
data from a Portuguese bus company and demonstrated that their approach
could provide high-quality solutions in reasonable computation times. [14] used
a genetic algorithm to solve the balanced bus crew rostering problem, which is a
variant of the BDRP where each bus trip requires two drivers to work together.
The authors showed that their approach could generate high-quality solutions
and outperformed other approaches from the literature.
[3] proposes a column generation metaheuristic for the DRP. The authors
use a decomposition approach that allows for the creation of a large number of
rosters in a short time while still maintaining quality. [9] propose an integrated
and sequential solution method for the cyclic BDRP. The authors use a combi-
nation of column generation and branch-and-price. Their test instances are from
a Canadian bus company.
[17] describes a heuristic method for solving the problem, which involves
assigning drivers to shifts while taking into account their preferred days off. The
proposed method uses multi-commodity flow models to decompose the problem
into smaller subproblems, and then, uses a fixing process to optimize the solu-
tions. [24] presents a mathematical model for solving the integrated driver roster-
ing problem in public bus transport. The paper proposes a duty-block network
approach that divides the scheduling process into two stages: duty assignment
and block construction. The proposed model is compared with a traditional
linear programming model, and the results show that the duty-block network
 Bus Driver Rostering with Uncertain Daily Working Hours 367

approach is more efficient and has better performance in terms of computational

time and quality of the solution.
In [25], the authors propose an integrated approach that considers both the
scheduling of drivers and the assignment of vehicles to routes, taking into account
factors such as driver preferences, vehicle types, and passenger demand. The
paper presents a mathematical model for the problem and describes a heuristic
algorithm for solving it. The approach is tested on real-world data from a bus
company in Germany, and the results show that the integrated approach can
lead to significant improvements in efficiency and cost-effectiveness compared to
traditional methods of driver scheduling.
There are two types of crew rostering, namely cyclic and non-cyclic. Cyclic
rostering involves repeating patterns of shifts within a defined cycle, providing
predictability and equal distribution of shifts. Non-cyclic rostering, on the other
hand, offers stability and consistency with fixed shifts over a longer duration.
While both methods have their advantages, our focus in this work is on non-
cyclic rostering to provide employees with a stable and predictable schedule that
promotes work-life balance and consistency.
[27] addresses the cyclic and non-cyclic crew rostering problems in public bus
transit. The authors propose a mathematical model to describe these problems
and present an exact algorithm to solve them. The paper also includes a case
study based on real-world data from a German public transportation company
to indicate the effectiveness of the proposed algorithm.
[26] presents a metaheuristic approach to solve a personalised crew rostering
problem in public bus transit. The authors propose a hybrid metaheuristic that
uses ant colony optimization, simulated annealing, and tabu search to gener-
ate high-quality solutions to this problem. The results show that the proposed
algorithm outperforms the other methods in terms of both solution quality and
computational efficiency.
In our current BDRP, we present a model for the problem which seeks to
minimize the total assignment costs and the total number of undesired duties
assigned to drivers. The aim of the latter is to increase the satisfaction of drivers.
Since the maximum working hours of each driver on each day are usually not
deterministic in the real applications and drivers can be unavailable during some
periods because of health problems or personal engagements, we consider this
parameter to be uncertain. Therefore, drivers cannot be assigned to duties which
must be done during their unavailability periods. On the other hand, maybe
the drivers are able to compensate these missing hours on some other days.
This forms a model where the uncertainty emerges at the right-hand side of a
constraint. To the best of our knowledge, this special characteristic has not yet
been sufficiently investigated.
Short-term staff shortages, particularly among train and bus drivers, have
become a common issue in the transportation sector. Several sources, such as
[1] and [2], highlight the prevalence of this problem. The International Road
Transport Union (IRU) conducted a survey revealing that in 2021, 7% of bus
and coach driver positions in Europe were unfilled, with an expected increase to
368 A. Nourmohammadzadeh and S. Voß

8% in 2022. Furthermore, the survey noted a lack of young drivers, with only
3% of bus and coach drivers under the age of 25, contributing to a demographic
crisis in the industry. Additionally, gender diversity in the profession is limited,
as less than 3% of truck drivers and only 12% of bus and coach drivers in Europe
are women. The primary cause of driver shortages, as reported by road transport
operators, is a scarcity of skilled drivers in various regions.
Some previous attempts in right-hand side uncertainties are for example: [22],
which discusses the bi-level knapsack problem with stochastic right-hand sides,
exploring the optimal solution to this problem using mathematical modeling
techniques, and [13], which addresses the maximization problem of monotone
submodular functions under an uncertain knapsack capacity. [18] explores the
complexity of 2-stage robust linear programming with right-hand side uncer-
tainty and its applications. The paper provides useful insights for researchers
and practitioners dealing with uncertainty in linear programming problems. The
same author researches into the complexity of two-stage robust linear program-
ming problems with right-hand side uncertainty in [19]. It is proven that the
problem becomes NP-hard when the uncertainty set is defined as an ellipsoid,
which is a more general set than the polyhedral set considered in the previous
paper. The paper provides a detailed proof of the complexity result and dis-
cusses its implications for practical applications. Closely related to our paper
are also issues to solve integrated vehicle and crew scheduling problems with
simple uncertainty measures in [10,11] regarding the number of vehicles.
A robust optimization approach is chosen to deal with right-hand side uncer-
tainties in our work. The concept of robustness is essential to address uncer-
tainties inherent in the system. The motivation behind employing robust opti-
mization techniques is to ensure that the proposed solutions remain of high
quality and effective even when faced with various input variations or scenar-
ios. By considering the uncertainties in the inputs, the objective is to develop
strategies or models that can withstand these uncertainties and still perform
optimally. Robustness measurement is a way to quantify the ability of a solution
or approach to maintain its effectiveness and quality under different conditions
or uncertainties. It allows for a more comprehensive evaluation and comparison
of different strategies, taking into account their performance across a range of
possible scenarios.
Due to the complexity of this problem, a matheuristic approach comprising
the particle swarm optimization (PSO) [12] and partial optimization metaheuris-
tic under special intensification conditions (POPMUSIC) [23] using consecutive
mathematical optimization of sub-models of the problem, is devised. A set of
benchmark instances and a peer algorithm from the literature are chosen to test
the ability of our solution methodology.
The remainder of this paper is organized as follows. In Sect. 2, the problem
is described and its mathematical model is presented. Section 3 explains our
solution methodologies. The computational experiments are covered in Sect. 4.
Finally, Sect. 5 is dedicated to the conclusions and directions to extend the work.
 Bus Driver Rostering with Uncertain Daily Working Hours 369

2 Problem Description
In the BDRP, there is a set of duties which have to be performed with a set of
drivers within a planning horizon. Each duty has a duration and each driver has
a set of desired duties. For the assignment of each duty to each driver, a specific
cost incurs. There is also a cost corresponding to the assignment of a duty which
is not included in the set of desired duties of the driver. The objective of this
version of the BDRP is to minimize a weighted sum of the assignment costs
and drivers’ dissatisfaction. A big challenge is that the maximum time of the
availability of each driver is not certain during each day. In other words, drivers
can be unavailable in some periods but may make up the missing hours on other
days. Therefore, the daily working time of drivers may become smaller or larger
than the considered standard 8 h. Drivers do not usually work for more than a
maximum number of consecutive days without taking any day off in between.
This is regarded as another major restriction in rostering.
In the context of the BDRP, an additional restriction has been introduced to
ensure efficient scheduling of duties while considering driver preferences. This
constraint, referred to as the “no early duty after late duty” constraint, is
designed to avoid assigning an early duty to a driver who has done a late duty on
the previous day. The early duties are those which are done before 11:00 while the
late duties start after 18:00. By incorporating the “no early duty after late duty”
constraint into the BDRP, the scheduling algorithm can optimize the assignment
of duties while respecting the drivers’ preferences and promoting their well-being
by acknowledging the need for appropriate rest and recovery periods. This con-
straint enhances the fairness and efficiency of the rostering process, contributing
to improved driver satisfaction and overall operational performance.
The rostering constraints are respecting the maximum daily working hours
of drivers as well as the maximum number of consecutive working days, which
is here assumed to be 5. The notations used in the model are given in Table 1.
The mathematical formulation of our BDRP is as follows:
  
M in w1 cd,u + w2 xd,u (1)
d u∈U d u∈P
/ d

xd,u = 1 u∈U (2)
d

xd,u Hu ≤ M DW Hd,t d ∈ D, t ∈ T (3)
u∈Ut
 
xd,u ≤ 1 − xd,u d ∈ D, t ∈ T (4)
u∈Ul ∩Ut u∈Ue ∩Ut+1

ad,t |T | ≥ xd,u d ∈ D, t ∈ T (5)
u∈Ut


t+5
ad,t ≤ 5 d ∈ D, 1 ≤ t ≤ |T | − 5 (6)
t
xd,u , ad,t ∈ {0, 1} d ∈ D, u ∈ Ut , t ∈ T (7)
370 A. Nourmohammadzadeh and S. Voß

Table 1. Notations used in the mathematical model

Parameters Explanation
T The set of days of the planning horizon; t ∈ T
D The set of drivers; d ∈ D
U The set of all duties; u ∈ U
Ut The set of duties to be assigned to the drivers on day t; u ∈ Ut
Ue The set of early duties; u ∈ Ue
Ul The set of late duties; u ∈ Ul
Hu The duration of duty u in hours
M DW Hd,t The maximum daily working hours of driver d on day t (an uncertain
parameter)
Pd The duties which driver d desires to do
w1 The weight of the total assignment cost for the bus company
w2 The weight related to the importance of driver satisfaction for the
company
cd,u The assignment cost of duty u to driver d
Variables Explanation
xd,u Binary, =1, if driver d is assigned to duty u
ad,t Binary, =1, if driver d is active on day t

In the objective function (1), we calculate the weighted sum of the assignment
costs and the number of undesired duties assigned to the drivers. Constraint (2)
enforces each duty to be assigned to only one driver. Respecting the maximum
working days of each driver is implied by Constraint (3). Constraint 4 prevents
the assignment of an early duty to a driver if they were assigned a late duty on the
previous day. This is an important constraint in our model because M DW Hd,t
are uncertain parameters and appear on the right-hand side. Constraint (5)
activates variable ad,t if driver d works on day t. Constraint (6) avoids exceeding
the maximum consecutive working days of each driver. The variable domains are
expressed by (7).

3 Solution Methodology

Our solution methodology consists of a particle swarm optimization (PSO)

framework hybridized with simulated annealing (SA). A random population of
solutions are generated and regarded as particles. These particles are interac-
tively moved according to the classical PSO rules as follows

vi (t + 1) = wvi (t) + c1 r1 (pbesti − xi ) + c2 r2 (gbest − xi ) (8)

where vi (t) is the current velocity of particle i, pbesti is the personal best position
of the particle, and gbest is the best position of all particles in the swarm. The
 Bus Driver Rostering with Uncertain Daily Working Hours 371

values r1 and r2 are random numbers between 0 and 1, and w is the inertia
weight that controls the balance between exploration and exploitation.
Then, the positions are updated by:

poi (t + 1) = poi (t) + vi (t + 1) (9)

where poi is the current position of particle i.

A proportion (RSA) of the particles are chosen based on their fitness by
the roulette wheel method (see [6]) to be improved by SA. In SA, a number of
neighboring solutions (N BS) of the chosen solution is evaluated in each iteration
and we move to a neighboring solution if it has a better objective value; otherwise
this movement happens with the probability:
−(f (s)−f (s ))
PSA = e T (10)

wheref (s) and f (s ) are the objective values of the current and the neighboring
solution, respectively. T is the current temperature. We begin from an initial
temperature T0 and reduce it after each iteration by dividing it by CT . When
we move to a new solution, searching the remaining neighboring solutions of the
iteration is continued around it. The termination condition of the SA is passing
a number (M ISA) of consecutive iterations without observing any improvement
in the objective value.
Following the algorithms’ procedure, a proportion of the best particles
(P M B) and a proportion of other particles (P M O) are also improved according
to the POPMUSIC framework. The latter group is selected again by the roulette
wheel method. In the POPMUSIC, solutions are divided into K parts; then, each
time, one of these parts (called sseed ) is selected at random and an area includ-
ing it and the D adjacent parts to it called altogether A is processed by solving
the mathematical model of the problem. So, we solve each time a sub-model of
the problem by fixing all variables except for those existing in A at their values
in the incumbent solution. The Gurobi solver is used to solve the sub-models.
The POPMUSIC process on a solution stops when no more improvement can be
done by choosing any part as sseed .
The whole matheuristic terminates if the number of consecutive stagnant
PSO iterations, which are iterations without any improvement in the objective,
exceeds a limit (M CU I). This algorithm has several hyper-parameters, which
have to be tuned properly in the beginning with an appropriate method. The
pseudocode of this matheuristic algorithm is presented in Algorithm 1.
The solutions are encoded as strings. Each cell of the string belongs to a duty
and is filled with the label of one driver. However, as PSO works usually with
continuous domains, the cell contents are real numbers in the interval [0,|D|].
To convert the real numbers to the driver labels, assuming that cc is the cell
content, we calculate cc + 1. Figure 1 shows the encoding of a simple BDRP
with 5 drivers and 10 duties.
372 A. Nourmohammadzadeh and S. Voß

Algorithm 1: The proposed matheuristic

Data: Problem inputs
Result: A feasible solution expected to be of good quality
1 Set the parameters of the algorithm
2 Generate nP op random candidate solutions as the initial particles.
3 Evaluate the population.
4 It = 0.
5 Initialize a variable called BO, which keeps the best objective value found so far.
6 while It ≤ M CU I do
7 Move the particles based on their new speed.
8 Evaluate the particles.
9 Choose P SA × nP op candidates from the population by the roulette
wheel selection.
10 Improve the selected candidates by neighborhood searches based on SA.
11 Evaluate the SA results, and then, merge them with the previous
population.
12 Sort the total population based on the objective values.
13 Choose P M B × nP op from the best solutions and P M O × nP op based
on the roulette wheel selection from other solutions and input them in a set
called P SET .
14 for Each solution spm ∈ P SET do
15 Decompose spm in K parts, i.e., the set of parts is
H = {part1 , ..., partK }
16 Set P = ∅
17 while P = {part1 , ..., partK } do
18 Select a seed part sseed ∈ H and sseed = P at random
19 Build a sub-problem R composed of sseed and its nearest D parts
(corresponding to area A)
20 Optimize R through solving the mathematical formulation by the
Gurobi solver
21 if R has been improved then
22 Update solution S
23 P ←∅
24 end
25 else
26 Include sseed in P
27 end
28 end
29 end
30 Merge the POPMUSIC results with the population.
31 Sort the population based on the objective value and choose the first nP op
best of them as the new population.
32 Set BO = Objective value of the first solution of the population.
33 if BO is improved in comparison to its previous value then
34 It = 0
35 end
36 else
37 It = It + 1
38 end
39 end
40 Report BO.
 Bus Driver Rostering with Uncertain Daily Working Hours 373

Fig. 1. Solution encoding

In the implementation of PSO, whenever the content of any cells violates

the allowable domain, it is reflected back into the feasible area. In the SA, a
neighboring solution is defined as one which differs from the current solution
in the content of one cell. Therefore, to build a neighboring solution, a cell is
chosen randomly and its content is changed to another random value.
For the implementation of the POPMUSIC, each cell is considered as a part.
So, each time a cell is selected as sseed , and area A is build by it and its nearest D
cells located on both of its sides, i.e., one cell is selected from the right, and then,
one from the left and this continues until D parts are completed. Whenever the
beginning or end of the string is reached, the remaining parts are picked from
the opposite side. The variables corresponding to these cells constitute the sub-
model to be solved by mathematical programming. A simple example for the
application of POPMUSIC on an encoded BDRP solution is illustrated in Fig.
2. sseed is shown in dark grey, D is equal to 5, so the last part (cell) is picked
from the right. The D-nearest genes to sseed are shown in grey. It means that the
variables related to the assignment of the third to sixth duties are reset and the
drivers of these duties are optimally determined by solving the corresponding
mathematical sub-model.

Fig. 2. Applying POPMUSIC to a chromosome

The constraints of the model are handled by generating the variables’ values
in their feasible interval. For this sake, the variables are filled with a feasible
value one by one in the ascending order of the duties. When a variable is filled,
its value is entered in the corresponding constraints and determines the feasible
intervals for the next variables.
As the right-hand side of Constraint 3 is uncertain, the problem is solved once
based on each combination of M DW Hd,t values. Hence, we have one solution
and objective value for each scenario. The solution corresponding to the worst
objective value is considered as the robust solution because no scenario can lead
to any objective value worse than that. This is in accordance with the definition
of robustness presented in [4].
374 A. Nourmohammadzadeh and S. Voß

4 Computational Experiments
In this section, firstly, the used test instances, and then, the computational
settings are explained. The important step of parameterization of the algorithm
is declared next. Finally, the obtained results are presented and analyzed.

4.1 Test Instances

To test our proposed solution methodology, we use the 12 real-word benchmark

instances given in [27] and also tackled in [26]. The number of drivers is consid-
ered to be  |U |
50 . The default maximum daily working hours of each driver is 8 h.
The uncertainties are generated by choosing between 20% to 30% of the drivers
and changing their working hours on 10% to 20% of the days by reducing them
by 1 to 4 h, and subsequently, adding the same amount of time to the maximum
working hours of another day chosen randomly. For each chosen driver, 5 differ-
ent scenarios are build. For each unavailability scenario, the affected parameters
and variables are set in the model. Regarding the assignment variables, some
of them must be fixed at zero before starting the solution process because the
driver is not available at the time of the corresponding duty. In the objective
function (1), the cost of each undesired assignment is assumed to be five units,
so we have w1 = 1 and w2 = 5.
The characteristics of these instances are shown in Table 2.

Table 2. Characteristics of instances

Instance Nb. of duties Nb. of drivers

48-75-6 1313 27
52-73-6 1288 26
52-75-6 1321 27
393-45-37 5420 109
392-45-37 5815 117
397-40-37 4917 99
96-70-8 3200 64
87-70-8 3273 66
89-70-8 3260 66
221-45-30 4214 85
211-45-34 4003 81
214-45-34 3966 80
 Bus Driver Rostering with Uncertain Daily Working Hours 375

4.2 Computational Settings

The mathematical model of the BDRP is programmed in Python and Gurobi is
called. The considered time limit for the solver is 12 h. To examine the capabili-
ties of our algorithm in comparison to other alternative methods, the pure math-
ematical programming (MP), the pure hybrid metaheuristic of PSO-SA (MH),
i.e. our algorithm without the POPMUSIC, a counterpart algorithm from [26]
called MMASPara (S), which is an ant-colony-based method with paralleliza-
tion, and finally, our matheuristic, are applied to each instance. Since S is used
for a different version of the BDRP, we re-implemented it for our case. Except
MP, other methods are run five times for each instance and the average objective
value and execution time are referred due to the possibility of a different result
after each run. These algorithms are also coded in Python. All the computa-
tional experiments of this paper are executed in a Computer with an Intel Core
i7-4790 3.60 GHz processor and 8 GB of RAM.

4.3 Parameterization of the Algorithm

Parameter setting is an important phase in the implementation of (meta/mat)-
heuristics algorithms because their abilities are extremely affected by the values
of their hyper-parameters. In this work, we use the response surface method
(RSM) [7], which works based of design of experiments (DOE). In the RSM,
an initial interval is considered for each parameter in which we are likely to
find the best value for that parameter. Then, experiments based on different
combinations of parameters are done and according to their results, the RSM
chooses a value for each parameter. We implement the RSM in R language. The
empirically considered initial values for the parameters of our algorithm are as
follows. In PSO, w : [0, 1], nP op : [100, 500], c1 : [0, 1], c2 : [0, 1], in SA, RSA :
[0.1, 0.5] T0 : [1000, 100000], CT : [1.5, 10], N BS : [50, 200], M ISA = [10, 50],
in POPMUSIC: D = [10, 30], P M B = [0.1, 0.5], P M O = [0.1, 0.5], and for the
whole algorithm M CU I = [10, 50]. The parameter tuning phase is executed for
each instance separately and the required time is included in the total execution
times of the algorithms for each instance, which are presented later. However,
on average about 75% of the execution times belongs to the parameterization
phase. The values of the algorithm’s parameters set for each instance are shown
in Table 3.
The RSM is similarly used for the parameter setting of any other algorithm
used in this work.

4.4 Results and Comparisons

The results of the tested algorithms are summarized in Table 4. For each method,
we have the (average) objective value and elapsed time to reach it. The objec-
tive value is shown as the (average) objective value found by the corresponding
method divided by the best objective value found among all runs of all methods.
These best objective values are shown as BVF in the last column.
376 A. Nourmohammadzadeh and S. Voß

Table 3. Parameter tuning of our algorithm

Instance w nP op c1 c2 RSA T0 CT N BS M ISA D P M B P M O M CU I

48-75-6 0.47 173 0.58 0.40 0.24 65547 5.02 128 22 16 0.31 0.20 29
52-73-6 0.52 119 0.55 0.57 0.26 77873 4.02 98 22 16 0.23 0.21 26
52-75-6 0.56 116 0.56 0.51 0.26 38120 5.41 103 26 15 0.23 0.30 20
393-45-37 0.47 262 0.58 0.45 0.26 84825 3.14 121 21 24 0.27 0.28 22
392-45-37 0.40 304 0.44 0.62 0.28 21263 3.11 105 28 26 0.33 0.35 31
397-40-37 0.58 341 0.66 0.60 0.32 31823 3.72 184 37 22 0.34 0.32 36
96-70-8 0.51 331 0.51 0.52 0.31 21264 2.58 150 42 22 0.28 0.30 43
87-70-8 0.47 340 0.59 0.47 0.32 23183 3.57 86 42 21 0.31 0.34 41
89-70-8 0.48 464 0.62 0.68 0.33 47419 2.51 124 45 17 0.38 0.36 40
221-45-30 0.53 411 0.54 0.64 0.32 86611 4.20 153 49 23 0.36 0.35 33
211-45-34 0.45 409 0.42 0.63 0.31 95091 2.53 188 40 25 0.34 0.35 32
214-45-34 0.57 201 0.42 0.65 0.29 94764 4.86 154 47 18 0.23 0.30 31

Table 4. Results of the examined methods (MP: mathematical programming, MH:

hybrid metaheuristic of PSO-SA, S: state-of-the-art method, MMASPara of [26]),
OA: our proposed matheuristic algorithm, BVF: best objective value found, Obj:
(average) objective value
Best objective value f ound (BV F )
, Time: (average) execution time in seconds

Instance MP MH S OA BVF
Obj Time Obj Time Obj Time Obj Time
48-75-6 1.00 12154.21 1.32 42.10 1.17 45.20 1.04 62.19 8.86 × 108
52-73-6 1.00 10534.53 1.26 37.15 1.06 48.62 1.02 66.73 8.92 × 108
52-75-6 1.00 11543.14 1.28 40.21 1.11 46.71 1.03 65.21 6.32 × 108
393-45-37 1.00 11397.19 1.31 41.64 1.33 182.27 1.07 63.31 4.19 × 108
392-45-37 1.00 13249.95 1.33 44.11 1.13 178.42 1.03 71.09 0.63 × 109
397-40-37 1.00 11864.44 1.25 39.45 1.33 202.82 1.08 68.91 4.07 × 108
96-70-8 1.00 10764.16 1.23 35.20 1.29 188.90 1.07 64.15 2.33 × 108
87-70-8 1.00 14177.99 1.29 41.72 4.10 219.62 1.18 72.63 2.08 × 107
89-70-8 2.65 43200 1.20 34.39 1.41 295.83 1.09 59.87 8.16 × 108
221-45-30 3.33 43200 1.23 36.58 1.23 478.62 1.05 62.42 1.42 × 1011
211-45-34 1.00 15692.70 1.35 45.82 2.16 296.08 1.21 70.10 6.84 × 1010
214-45-34 1.00 16480.40 1.24 43.17 2.25 352.52 1.19 67.06 6.21 × 1010

It is observed that mathematical programming can find the optimal values

in long computational times for the first 8 instances. However, the solver cannot
find the optimal solutions within the considered time limit for instances 89-
70-80 and 221-45-30. So, the time limit (43200 s or 12 h) and the best (last)
proportional objective value of the solver are tabulated. It can be deduced that
the gaps from the best known results are extremely large in terms of these two
instances. The two final instances can be solved to optimality in longer execution
times compared to the first 8 ones.
 Bus Driver Rostering with Uncertain Daily Working Hours 377

The PSO-SA (MH) and MMASPara (S) can find solutions slightly distant
from the bests but in much shorter execution times compared to MP. On the
other hand, our matheuristic spends a little more time than MH but achieves
solutions of considerably better quality. It is also perceived that the POPMUSIC
plays an important role in the success of our matheuristic since the solution
quality significantly drops when we exclude it and use only PSO-SA (MH).
The proportional objective values of all runs are illustrated for the four
examined algorithms as well as the model without uncertainties solved by our
matheuristic (non-robust: shown as NR) as box plots in Fig. 3. It is evident that
the four quartiles related to our approach are lower or better than the three first
and this indicates its satisfactory performance and efficiency. In addition, we see
that the costs of solving the robust model are not high because the box plot of
NR does not considerably differ from OA.

Fig. 3. Box plots of proportional objective values based on all runs of all instances

In the next step, we aim to conduct a statistical test to determine if the

difference of the average objective values in any pair of the used algorithms
are statistically significant. For this sake, a non-parametric test called Friedman
with Bergman-Hommel post-hoc procedure ([5]) is used. [8] shows that this test
is very appropriate to compare more than two methods in pairs. The obtained
p-values are given in Table 5.

Table 5. P-values of the pairwise comparisons of the methods

MP MH S OA
MP – 1.58 × 10−4 2.61 × 10−4 2.71 × 10−4
MH 1.58 × 10−4 – 3.73 × 10−5 3.12 × 10−5
−4 −5
S 2.61 × 10 3.73 × 10 – 4.90 × 10−3
OA 2.71 × 10−4 3.12 × 10−5 4.90 × 10−3 –
378 A. Nourmohammadzadeh and S. Voß

According to the very low p-values, we can reject the null hypothesis claiming
the equality of averages for all pairs in the significance level of 0.01.
Generally, it can be deduced that this robust optimization with regard to the
existence of uncertainties on the right-hand side of the model can be properly
handled by the proposed matheuristic. So, we are able to provide driver rosters
that remain effective even when the daily working hours of drivers change, which
is very likely to happen over the course of the planning horizon.

5 Conclusions
In this work, we addressed the important problem of driver rostering of bus
companies with the consideration of uncertain daily working hours of drivers.
Concerning the fact that it is an uncertainty in right-hand side parameters of the
corresponding mathematical model, a robust optimization problem which is less
focused in the literature was surveyed. A matheuristic algorithm is proposed con-
sisting of proper exploration and exploitation with PSO, SA and POPMUSIC.
The results indicate that the combination of metaheuristics and mathematical
programming techniques can provide a strong optimization power in a short
time, which cannot be found in pure approaches of these types.
For extending this research, we can recommend applying the proposed
method to other similar assignment problems. Another direction is to enhance
the performance of the method by adding extra effective operators to it or to
try other concepts of embedding mathematical optimization in metaheuristics.
Investigation into fast and efficient parameter tuning can also be of great value.

Acknowledgements. This research is funded by dtec.bw - Digitalization and Tech-

nology Research Center of the Bundeswehr. dtec.bw is funded by the European Union
- NextGenerationEU.
The helpful comments of two reviewers are gratefully acknowledged.

References
1. Current traffic situation. https://www.der-metronom.de/fahrplan/aktuelle-
verkehrslage/. Accessed 20 June 2023
2. New survey reveals shocking shortage of bus drivers. https://www.unitetheunion.
org/news-events/news/2021/november/new-survey-reveals-shocking-shortage-of-
bus-drivers/. Accessed 20 June 2023
3. Barbosa, V., Alvelos, F., Respı́cio, A.: Bus driver rostering by column genera-
tion metaheuristics. In: Fonseca, R.J., Weber, G.-W., Telhada, J. (eds.) Compu-
tational Management Science. LNEMS, vol. 682, pp. 225–231. Springer, Cham
(2016). https://doi.org/10.1007/978-3-319-20430-7 29
4. Ben-Tal, A., Nemirovski, A.: Robust solutions of uncertain linear programs. Oper.
Res. Lett. 25(1), 1–13 (1999). https://doi.org/10.1016/s0167-6377(99)00016-4
5. Bergmann, B., Hommel, G.: Improvements of general multiple test procedures for
redundant systems of hypotheses. In: Bauer, P., Hommel, G., Sonnemann, E. (eds.)
Multiple Hypothesenprüfung / Multiple Hypotheses Testing, vol. 70, pp. 100–115.
Springer, Heidelberg (1988). https://doi.org/10.1007/978-3-642-52307-6 8
 Bus Driver Rostering with Uncertain Daily Working Hours 379

6. Blickle, T., Thiele, L.: A comparison of selection schemes used in evolutionary algo-
rithms. Evol. Comput. 4(4), 361–394 (1996). https://doi.org/10.1162/evco.1996.4.
4.361
7. Box, G.E.P., Draper, N.R.: Response Surfaces, Mixtures, and Ridge Analyses, 2nd
edn. Wiley, Hoboken (2007)
8. Derrac, J., Garcı́a, S., Molina, D., Herrera, F.: A practical tutorial on the use of
nonparametric statistical tests as a methodology for comparing evolutionary and
swarm intelligence algorithms. Swarm Evol. Comput. 1(1), 3–18 (2011). https://
doi.org/10.1016/j.swevo.2011.02.002
9. Er-Rbib, S., Desaulniers, G., Elhallaoui, I., Munroe, P.: Preference-based and cyclic
bus driver rostering problem with fixed days off. Public Transp. 13(2), 251–286
(2021). https://doi.org/10.1007/s12469-021-00268-y
10. Ge, L., Kliewer, N., Nourmohammadzadeh, A., Voß, S., Xie, L.: Revisiting the
richness of integrated vehicle and crew scheduling. Public Transp. (2022). https://
doi.org/10.1007/s12469-022-00292-6
11. Ge, L., Nourmohammadzadeh, A., Voß, S., Xie, L.: Robust optimization for inte-
grated vehicle and crew scheduling based on uncertainty in the main inputs. In:
The Fifth Data Science Meets Optimisation Workshop at IJCAI-22. Vienna (2022).
https://sites.google.com/view/ijcai2022dso/. Accessed 30 Apr 2023
12. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of
ICNN’95 - International Conference on Neural Networks, vol. 4, pp. 1942–1948
(1995). https://doi.org/10.1109/ICNN.1995.488968
13. Kawase, Y., Sumita, H., Fukunaga, T.: Submodular maximization with uncertain
knapsack capacity. arxiv (2018). https://arxiv.org/pdf/1803.02565
14. Ma, J.H., Liu, T., Guan, W.: Solving balanced bus crew rostering problem with
genetic algorithm. Appl. Mech. Mater. 361–363, 2070–2074 (2013). https://doi.
org/10.4028/www.scientific.net/AMM.361-363.2070
15. Maniezzo, V., Boschetti, M.A., Stützle, T.: Automatic design for matheuristics.
In: Matheuristics. EURO Advanced Tutorials on Operational Research, pp. 35–57.
Springer, Cham (2021). https://doi.org/10.1007/978-3-030-70277-9 2
16. Maniezzo, V., Stützle, T., Voß, S. (eds.): Matheuristics: Hybridizing Metaheuristics
and Mathematical Programming. Springer, New York (2010). https://doi.org/10.
1007/978-1-4419-1306-7
17. Mesquita, M., Moz, M., Paias, A., Pato, M.: A decompose-and-fix heuristic based
on multi-commodity flow models for driver rostering with days-off pattern. Eur. J.
Oper. Res. 245(2), 423–437 (2015). https://doi.org/10.1016/j.ejor.2015.03.030
18. Minoux, M.: On 2-stage robust LP with RHS uncertainty: complexity results and
applications. J. Global Optim. 49(3), 521–537 (2011). https://doi.org/10.1007/
s10898-010-9645-2
19. Minoux, M.: Two-stage robust LP with ellipsoidal right-hand side uncertainty is
NP-hard. Optim. Lett. 6(7), 1463–1475 (2012). https://doi.org/10.1007/s11590-
011-0341-z
20. Moz, M., Respı́cio, A., Pato, M.V.: Bi-objective evolutionary heuristics for bus
driver rostering. Public Transp. 1(3), 189–210 (2009). https://doi.org/10.1007/
s12469-009-0013-x
21. Nurmi, K., Kyngäs, J., Post, G.: Driver rostering for a Finnish bus transportation
company. In: IAENG Transactions on Engineering Technologies, vol. 7, pp. 15–35.
World Scientific, Singapore and Hackensack NJ (2012). https://doi.org/10.1142/
9789814390019 0002
380 A. Nourmohammadzadeh and S. Voß

22. Özaltın, O.Y., Prokopyev, O.A., Schaefer, A.J.: The bilevel knapsack problem with
stochastic right-hand sides. Oper. Res. Lett. 38(4), 328–333 (2010). https://doi.
org/10.1016/j.orl.2010.04.005
23. Taillard, É.D., Voß, S.: POPMUSIC – partial optimization metaheuristic under
special intensification conditions. In: Ribeiro, C., Hansen, P. (eds.) Essays and
Surveys in Metaheuristics, vol. 15, pp. 613–629. Springer, Boston (2002). https://
doi.org/10.1007/978-1-4615-1507-4 27
24. Xie, L., Suhl, L.: A duty-block network approach for an integrated driver rostering
problem in public bus transport. In: Longhurst, J., Brebbia, C.A. (eds.) Urban
Transport XVIII, pp. 295–306. WIT Transactions on The Built Environment, WIT
Press Southampton, UK (2012). https://doi.org/10.2495/UT120261
25. Xie, L., Kliewer, N., Suhl, L.: Integrated driver rostering problem in public bus
transit. Procedia. Soc. Behav. Sci. 54, 656–665 (2012)
26. Xie, L., Merschformann, M., Kliewer, N., Suhl, L.: Metaheuristics approach for
solving personalized crew rostering problem in public bus transit. J. Heuristics
23(5), 321–347 (2017). https://doi.org/10.1007/s10732-017-9348-7
27. Xie, L., Suhl, L.: Cyclic and non-cyclic crew rostering problems in public bus
transit. OR Spectrum 37(1), 99–136 (2015). https://doi.org/10.1007/s00291-014-
0364-9
 Beyond Cargo Hitching: Combined People
and Freight Transport Using Dynamically
Configurable Autonomous Vehicles

Joris J. A. Kortekaas2 , Breno A. Beirigo1 , and Frederik Schulte2(B)

1
Department of High-tech Business and Entrepreneurship, University of Twente,
Hallenweg 17, 7522 NH Enschede, The Netherlands
[email protected]
2
Department of Maritime and Transport Technology, Delft University of
Technology, Mekelweg 2, 2628 CD Delft, The Netherlands
[email protected], [email protected]

Abstract. A Dynamically Configurable Autonomous Vehicle (DCAV)

is a new class of autonomous vehicle concept using a separable design
of lower and upper parts—carriers and modules—to allow more flex-
ible operation. A fleet of DCAVs consists of a set of carriers and a
set of compatible modules. Different, possibly crowd-sourced, modules
can increase the number of use-cases for DCAVs, possibly leading to
disruptive changes in the transport sector. This study investigates the
use of DCAV system operating on an Autonomous Mobility-on-Demand
(AMoD) scenario, combining passenger and freight transport flows. The
novel problem is denoted as the Dynamically Configurable Autonomous
Vehicle Pickup and Delivery Problem (DCAVPDP). We propose a mixed-
integer linear programming (MILP) model aiming to minimize DCAV-
fleet size and distance traveled. We compare the performance of a DCAV
fleet to the performance of a typical single-purpose fleet (consisting of
dedicated passenger and freight vehicles). The numerical study, with 360
instances for each fleet type, considering four people-and-freight demand
distribution scenarios, the inclusion of ridesharing, module-and-carrier
(de)coupling locations, and different simulation horizon lengths, shows
that the proposed modular DCAV system can fulfill a mixed people-and-
freight demand using, on average, 18.77% fewer carriers than a regular
AMoD system comprised of single-purpose vehicles while increasing on-
duty fleet utilization by 4.82%.

Keywords: Dynamically Configurable Autonomous Vehicles · Shared

Autonomous Vehicles · Cargo Hitching · Combined Passenger and
Freight Transport · Ridesharing

1 Introduction
Dynamically Configurable Autonomous Vehicles (DCAVs) are a new class of
autonomous vehicles using a separable design concept consisting of drive and
transport units. The drive unit (also called the carrier, driveboard, or chassis)
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 381–395, 2023.
https://doi.org/10.1007/978-3-031-43612-3_24
382 J. J. A. Kortekaas et al.

is a standardized platform housing self-driving equipment (e.g., sensors, motors,

battery). The transport unit (also referred to as module, capsule, pod, or body)
is a switchable add-on part that can be dynamically and autonomously out-
fitted onto drive units to accommodate any commodity type (see Fig. 1). Such
a modular nature of DCAVs increases operational flexibility, enabling various
use-cases: transport providers can assign diverse modules to carriers to fulfill
different demands.

Fig. 1. Example of a Dynamically Configurable Autonomous Vehicle (DCAV) system

[15]. The drive unit (carrier) can autonomously outfit diverse transport units (modules)
on the fly to service heterogeneous demands such as freight and mobility.

By separating drive and transport units, providers can increase the utilization
of self-driving carriers (presumably, the most expensive assets) while increasing
profits, adapting to changing demand profiles such as people and freight on the
operational level. Typically, vehicle acquisition is a long-term investment, which
makes deciding on adequate fleet size and mix a complex strategical decision
[2,7]. Therefore, relying on a DCAV fleet may also mitigate losses due to dras-
tic demand shifts. For example, the COVID outbreak led to a surge in delivery
services while social distancing decreased mobility, exposing the fragility of trans-
portation systems based on single-purpose vehicles [12]. Ultimately, investing in
additional freight modules to accommodate a higher freight demand is expected
to be considerably cheaper than investing in entirely new freight vehicles.
Combined people and freight transportation systems have been gaining atten-
tion in the literature [3,9]. For example, solution approaches have been designed
for settings where people and parcel share rides on taxis (see, e.g., [8]) and
compartmentalized autonomous vehicles (see e.g., [1,13]). However, there is no
study on optimizing the operations of a DCAV-based transportation system
that dynamically switches transport units to fulfill heterogeneous demand. Only
concept designs have been evaluated (see, e.g., [10,11,15]) and presented by com-
panies such as the Continental BEE [4], the Mercedes Benz Vision Urbanetic [5]
and the Toyota e-palette [14].
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 383

This study presents a mixed-integer linear programming (MILP) formula-

tion for the Dynamically Configurable Autonomous Vehicle Pickup and Delivery
Problem (DCAVPDP) arising on a transport system that services both passen-
ger and freight demands. Our formulation extends the formulation by [6] for the
vehicle routing problem with trailers and transshipments (VRPTT) by adding
service level constraints as formulated in dial-a-ride problems (DARPs). Addi-
tionally, we consider the integration of passenger and freight transport, the use of
parking locations, and allowing multiple trips and ridesharing. The model seeks
to find an optimal routing solution that minimizes vehicle utilization in terms
of linear combination of the distance traveled by carriers and modules.

Fig. 2. Example of a DCAV fleet operation throughout a typical day.

Figure 2 shows an operational example of a DCAV fleet using different mod-

ules throughout the day to fulfill passenger and freight requests. The carrier
starts at 7:15 at the depot and couples a passenger module to transport pas-
sengers during the morning rush hour. As long as the vehicle capacity is not
exceeded, the vehicle is allowed to make multiple pickups and have passengers
share the ride. When the latest passenger of the first trip is dropped off at 10:15
the vehicle gets called back to the depot. Next, the carrier switches its passen-
ger module for a freight module and starts servicing freight requests as demand
for passenger transport is low at the time. After picking up several packages
the vehicle returns to the depot once more at 14:30 to switch over to a pas-
senger module to transport passengers for the remainder of the afternoon. To
meet demand the carrier has made three separate trips to and from the depot,
384 J. J. A. Kortekaas et al.

changing modules for each trip. Even though demand fluctuates throughout the
day, the carrier is being used to serve requests for either passengers or parcels,
resulting in high carrier utilization.
We benchmark the performance of the DCAV fleet (consisting of carriers and
passenger and freight modules) with a typical single-purpose fleet (consisting of
dedicated passenger and freight vehicles). We consider carriers and dedicated
vehicles of modular and single-purpose fleets have to start and end their route
at a central depot. Vehicles are not allowed to arrive early and wait at a pickup
location since requests are made on-demand, and there is little room for parking
in city centers. However, parking locations are available for vehicles to wait until
a pickup is requested; these locations can also be used for modular vehicles to
switch modules. In contrast, modules do not have to start or end their route
at the depot. They can also start and end their route from a parking location
and do not have to return to the same location as where they originated from.
We assume there is no limit to the number of times a module can be switched,
except occupied modules, which are not allowed to be switched.

2 Problem Description

This study considers a Dynamically Configurable Autonomous Vehicle Pickup

and Delivery Problem (DCAVPDP), which focuses on the daily operations of
an on-demand transportation provider that uses a DCAV fleet to pick up and
deliver requests of both passengers and parcels. We assume carriers are required
to start and end their route at a central depot and are not allowed to arrive early
and wait at a request location—there is little room for parking in an urban city
environment. However, parking locations are available for vehicles to wait until
a pickup is requested. These locations can also be used for modular vehicles to
switch modules. Modules do not have to start or end their route at the depot
and can be repeatedly switched when empty. Ridesharing is allowed between
requests of the same commodity, with detours limited by maximum extra ride
times.
Figure 3 highlights the difference between a modular and a single-purpose
fleet when servicing one passenger and one freight request. The single-purpose
fleet uses two dedicated vehicles traveling direct routes from the depot to the
pickup and delivery. In contrast, the DCAV fleet uses available parking locations
to gather modules closer to request locations and uses a single carrier to service
both requests. The carrier leaves the depot and travels to a parking location
where it couples a freight module to service the freight request. After reaching
the destination of the freight request, it drives to another parking location to
decouple the freight module and travel to the final available parking location to
couple a passenger module. The passenger module is used to service the passenger
request and returns together with the carrier to the depot.
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 385

Fig. 3. Difference between single-purpose and modular DCAV fleets when servicing
the same passenger and freight request set.

3 Problem Formulation
The DCAVPDP is defined on a directed graph G = (V, A), consisting of vertices
V and arcs A, and requires synchronisation in time and space of the movement
of carriers and modules. We consider a grid service area where a total number
of nr requests is generated. The set of all vehicles is indicated by F , which is a
collection of all vehicles that are available to the model. The set of all vehicles is
further separated in the sets Fa , Fq , and Fm . All motorized vehicles in the model
are indicated by Fa , which is a combination of all passenger and freight vehicles
and the carriers. Fq is the set of all vehicles that have load-carrying capability,
which consists of the passenger and freight vehicles and the passenger and freight
modules. The final set regarding vehicles is Fm , which is a set of all modules in
the model. Where V k are all reachable vertices for vehicle k, VPk is a set of all
reachable pickup vertices for vehicle k ∈ F .
The set of traversable arcs Ak is further specified in the following sets: Am C,
APD , and AkW . The set Am
k
C describes all traversable arcs module m ∈ Fm can
travel along when being coupled to carrier c ∈ C, since modules are not capa-
ble of traveling between real-world locations without the assistance of a carrier.
However, modules are allowed to travel to couple vertices on their own to allow
for the distribution of modules across all parking locations. The number of park-
ing locations in the model is specified by np and f and 0 correspond to the
start and final vertex respectively. This ensures a module can be coupled at all
available parking locations on the grid after being decoupled from a carrier. AkPD
386 J. J. A. Kortekaas et al.

Table 1. Sets, parameters, and variables of the DCAVPDP and dedicated people and
freight integrated problem.

Sets
C, Mp , Mf Carriers and passenger and freight modules (modular fleet)
Kp , K f Passenger and freight vehicles (dedicated fleet)
F = Kp ∪ Kf ∪ C ∪ Mp ∪ Mf . Set of all vehicles (fleet)
Fa = Kp ∪ Kf ∪ C. Set of powered vehicles (autonomous)
Fq = Kp ∪ Kf ∪ Mp ∪ Mf . Set of vehicles with load carrying capability
Fm = Mp ∪ Mf . Set of all modules
Vk Set of all reachable vertices for vehicle k ∈ F
VPk Set of all reachable pickup vertices for vehicle k ∈ F

VP = k∈F Vpk
k
Vpark Set of all vertices where vehicle k is allowed to wait
Ak Set of all traversable arcs for vehicle k ∈ F
AmC Set of arcs module m ∈ Fm can only traverse while spatially synchronised with a
compatible carrier c ∈ C
AkPD Set of traversable arcs between request pairs for vehicle k ∈ F
AkW Set of arcs for vehicle k ∈ F where waiting is not allowed at the arrival vertex
Parameters
Vehicles
nk Number of available vehicles with load carrying capability
nc Number of available carriers
Qk Capacity of vehicle k ∈ F
vk Speed of vehicle k ∈ F

Requests
nr Number of requests
qi Demand of request at vertex i ∈ VP
ei Earliest pick up time of request at vertex i ∈ VP
δi Maximum waiting time of request at vertex i ∈ VP
Δi Maximum extra ride time of request at vertex i ∈ VP

Model
0 Start vertex for all vehicles
f Final vertex for all vehicles
np Number of parking locations
si Service time at vertex i ∈ V
di,j Travel distance from vertex i to vertex j
ti,j Travel time from vertex i to vertex j
wm Weight of module travel distances
wrejection Requests rejection penalty
S Simulation start time
H Simulation horizon

Variables
xki,j (Binary) 1 if vehicle k traverses arc (i, j), 0 otherwise
τik Arrival time of vehicle k at vertex i
ωik Load of vehicle k after vertex i
rik Ride time of request i on vehicle k
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 387

is a set of traversable arcs for vehicle k ∈ F that are composed of a pickup and
delivery pair. Finally, AkW is a set of traversable arcs for vehicle k ∈ F on which
arriving early is not allowed since vehicles are not allowed to wait at request
locations.
The total number of requests in the model is specified by nr . Each request
has a pickup vertex i ∈ VP , which has a demand size of qi . The corresponding
delivery vertex has a demand size of −qi , effectively emptying the vehicle again.
The pickup time window of request i is determined by the earliest pickup time
ei of the request and its maximum waiting time δi . The earliest pickup times of
requests are being generated during simulation horizon H, starting from simula-
tion start time S. Requests do not have a set delivery time window, instead, each
request has a maximum extra ride time Δi . The destination of a request has to
be reached within the sum of the travel time and the maximum extra ride time.
Additionally each vertex i ∈ V has a service duration of si seconds. nk specifies
the number of available vehicles in the model that have a load-carrying capa-
bility (passenger and freight vehicles as well as passenger and freight modules),
while nc specifies the number of carriers available to the model. All vehicles have
a set capacity of Qk and an average speed of v k . Table 1 shows an overview of
all sets and parameters used in the mixed-integer linear programming model.
The formulation of the MILP is as follows:
Minimize:
   
xki,j di,j + wm xki,j di,j
k∈Fa (i,j)∈Ak k∈Fm (i,j)∈Ak
   
− xk0,f + (1 − xki,j ) ∗ wrejection (1)
k∈Fa i∈VP k∈Fq (i,j)∈Ak

Subject to:
 
xki,j ≤ 1 ∀i ∈ VP (2)
k∈Fq (i,j)∈Ak
 
xki,h − xkh,j = 0 ∀(i, j) ∈ AkPD , ∀k ∈ Fq (3)
(i,h)∈Ak (h,j)∈Ak

xk0,j = 1 ∀k ∈ F (4)
0,j∈Ak

xki,f = 1 ∀k ∈ F (5)
i,f∈Ak
 
xkh,i − xki,j = 0 ∀i ∈ V k \{0, f}, ∀k ∈ F (6)
(h,i)∈Ak (i,j)∈Ak

xm
i,j − xki,j = 0 ∀(i, j) ∈ Am
C , ∀m ∈ Fm (7)
k∈C
388 J. J. A. Kortekaas et al.

xki,j = 1 ⇒ τjm ≤ τjk ∀(i, j) ∈ Am

C , ∀k ∈ C, ∀m ∈ Fm (8)
xm k m
i,j = 1 ⇒ τj ≤ τj ∀(i, j) ∈ Am
C , ∀k ∈ C, ∀m ∈ Fm (9)
τjk ≥ (τik + si + ti,j )xki,j ∀(i, j) ∈ Ak , ∀k ∈ F (10)
ei ≤ τik ≤ ei + δi ∀i ∈ VPk , ∀k ∈ F (11)
τjk ≥ τik ∀(i, j) ∈ AkPD , ∀k ∈ Fq (12)
xki,j = 1 ⇒ τjk − τik ≤ ti,j + si ∀(i, j) ∈ AkW , ∀k ∈ F (13)
rik ≥ τjk − τik ∀(i, j) ∈ AkPD , ∀k ∈ Fq (14)
ti,j ≤ rik ≤ ti,j + Δi ∀(i, j) ∈ AkPD , ∀k ∈ Fq (15)
ωjk ≥ ωik + qi xki,j ∀(i, j) ∈ Ak , ∀k ∈ Fq (16)
max(0, qi ) ≤ ωik ≤ Qk k
∀i ∈ V , ∀k ∈ Fq (17)
ωik ≤ 0 k
∀i ∈ Vpark , k∈F (18)
xki,j ∈ {0, 1} ∀i, j ∈ Ak , ∀k ∈ F (19)
τik ∈N ∀i ∈ V k
(20)
ωik ∈ N ∀i ∈ V k (21)
rik ∈ N ∀i ∈ VP (22)

The objective function (1) aims to minimize the travel distance of all motor-
ized vehicles, as well as the total number of motorized vehicles used to find an
optimal solution. Secondly, the travel distance of modules is also minimized,
however, it is deemed less important and is offset by a factor of wm . To allow
the model to reject requests a penalty of wrejection is added if a pickup vertex
is not visited. Constraints (2) to (6) are the general routing constraints, and
(7) to (9) are spatial and temporal synchronisation constraints to couple the
movement of carriers and modules. The time window constraints are described
by constraints (10) to (13) and subsequently constraints (14) and (15) are the
ride time constraints. Finally, constraints (16) to (18) are the vehicle load con-
straints. Equations (19) to (22) describe the decision variables in the model.
Constraint (2) ensures that every request’s pickup vertex is visited by at most
one vehicle, however, every vertex does not need to be visited. A vehicle that
has visited the pickup vertex of a request must also travel to the request’s des-
tination vertex, which is ensured by (3). Constraints (4) and (5) guarantee that
every vehicle starts it’s route at the start vertex and ends at the final vertex
respectively. To preserve flow conservation, constraint (6) enforces every vehicle
to leave the same vertex it has entered. Constraint (7) is a spatial synchronisation
constraint to ensure modules are only able to traverse arcs between two different
real-world locations (Am C ) while being coupled with a carrier. Constraints (8)
and (9) are temporal synchronisation constraints that impose identical arrival
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 389

times of coupled carriers and modules. The arrival time of vehicles is defined by
constraint (10). Constraint (11) ensures that a request is picked up after their
earliest pick-up time and within their maximum waiting time. A vehicle should
arrive at the destination vertex of a request pair after the pickup vertex has been
visited, which is guaranteed by (12). Vehicles are not allowed to arrive early and
wait at request pickup locations, constraint (13) ensures that for the set of arcs
on which waiting is not allowed (AkW ) vehicles won’t arrive before the earliest
pickup time. Constraint (14) defines the ride time of a request, and (15) makes
sure the maximum extra ride time of a request is not exceeded. Constraint (16)
defines the vehicle load and (17) enforces the load of a vehicle not to exceed the
maximum capacity of the vehicle. Constraint (18) guarantees that vehicles are
empty when arriving at parking or decoupling vertices, to ensure no occupied
vehicles or modules are parked or decoupled. Finally, constraints (19) to (22)
define the model’s decision variables, which are the traversed arcs, arrival time,
vehicle load and ride time variables respectively.

4 Numerical Study
In this section, the configuration of the MILP for the different instances will be
discussed, as well as the generation of the passenger and freight request data.
Next, the general parameters of the model will be given, followed by a discussion
of key performance indicators of the transport systems.

4.1 Instance Design

We test the model on two fleet configurations (viz., single-purpose and DCAV
fleets) using the same request data. We consider four different demand scenarios
throughout the simulation horizon H:

– Balanced scenario (BLS): Both request types are uniformly distributed

over the horizon.
– Passenger peak scenario (PPS): Passenger requests are generated during
the first half of the horizon, while freight requests are uniformly distributed
throughout the whole horizon.
– Freight peak scenario (FPS): Freight requests appear only during the
second half of the simulation horizon, while passenger requests are uniformly
distributed.
– Complementary scenario (COS): Passenger requests appear during the
first half of the horizon and freight requests during the second half.

Besides, we test the effect of allowing ridesharing (same commodity requests

can share a ride) and vary the number of parking locations np (0, 1, or
3), where carriers can park and switch modules during their route. We also
determine 15 value combinations for simulation horizon H ∈ {1, 8}, the total
fleet size nk ∈ {2, 4, 6}, and the number of requests nr ∈ {4, 8, 12, 16, 24}.
390 J. J. A. Kortekaas et al.

Values nk and nr are divided equally between passenger and freight vehicles
and requests, respectively. Additionally, for each module in the model, there is a
carrier available to be used, thus, the number of carriers nc is equal to the number
of vehicles nk . In the worst-case scenario, the DCAV fleet could behave similarly
to the single-purpose fleet by assigning a module to each carrier and having the
same routes as the single-purpose vehicles. The parameter groups (H, nk , nn )
are as follows: (1, 2, 4), (1, 2, 8), (1, 2, 12), (1, 4, 4), (1, 4, 8), (1, 4, 12), (1, 6, 4),
(1, 6, 8), (1, 6, 12), (8, 2, 12), (8, 2, 16), (8, 2, 24), (8, 4, 12), (8, 4, 16), (8, 4, 24).
An overview of the 720 instances generated can be found in Fig. 4.

Fig. 4. Instances for the DCAVPDP and the benchmark single-purpose problem (ded-
icated vehicles for each commodity).

4.2 Model Parameters

The values of general parameters are shared by all instances. The grid on which
the requests are being generated is 2, 000 by 2, 000 meters, with a distance
between nodes of 100 m. The depot is located at (600 m, 600 m), and the three
available parking locations are located at (600 m, 1400 m), (1400 m, 600 m), and
(1400 m, 1400 m). When only one parking location is accessible for the fleet, it
is located at (1400 m, 1400 m).
The locations of all request pickup and delivery locations are uniformly dis-
tributed over the rest of the nodes of the grid. Each request demands the trans-
portation of one passenger or parcel (i.e., request size qi = 1). Passenger and
freight requests have different values for maximum waiting time δi , extra ride
time Δi , and service duration si . All distances are calculated using the Man-
hattan distance formula and travel times are calculated using an average vehicle
speed v k = 5 m/s. To ensure vehicles will be able to reach the first request on
time, we set the simulation start time S = 900 s since the longest possible travel
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 391

time between opposing corners of the grid is 800 s. Passenger and freight vehicles
and modules have a capacity Qk = 2 for both the single-purpose and the DCAV
fleet, respectively. Typically, passengers are more sensitive to waiting times and
extra ride times, this is reflected in the model by giving more strict values to
passenger requests. The maximum waiting time δi for passenger requests is 300
s and the maximum extra ride time Δi is 900 s. For freight requests, δi = 900
s and Δi = 1800 s. Regarding the service duration si , we assume a vehicle has
to wait 120 s when servicing passengers and 300 s when picking up a parcel.
Besides, we consider the modular DCAV fleet has an additional service duration
si = 120 s to switch (i.e., couple or decouple) modules. Finally, rejecting requests
incurs a penalty wpenalty = 105 and the traveling of modules is weighted down
using wm = 10−3 .

4.3 Performance Metrics

Our numerical study aims to assess whether a fleet of DCAVs improves vehicle
utilization and occupancy while maintaining an equivalent service quality com-
pared to a traditional single-purpose fleet. In order to highlight the differences
between these two fleets, we consider the following fleet management perfor-
mance metrics:

– Utilization rate (on-duty): The time a vehicle is being actively used to

service requests (i.e., traveling to pick up and deliver, service times at request
locations), divided by the total simulation run time.
– Utilization rate (total): Total time vehicles spent traveling divided by the
total simulation run time. For the DCAV fleet, this time includes trips to
parking locations to (de)couple modules and module replacement times.
– Occupancy rate: The fraction of the utilization time that vehicles are trav-
eling loaded, taking into account their occupation ratio.
– Service quality: The service rate, the average pickup time of requests, and
the average ride time of requests.
– Fleet size: Total number of autonomous vehicles (dedicated AVs or carriers).
– Fleet capacity: Total number of people-and-freight transportation upper
parts. In the dedicated fleet, fleet capacity equals fleet size.
– Distance traveled: Total distance traveled by the whole fleet.

4.4 Computational Settings

The numerical study has been performed on an AMD Ryzen 5 3600 4Ghz CPU
with access to 32GB of RAM. Python version 3.8.8 and commercial solver Gurobi
Optimizer 9.1.2 were used to implement the DCAVPDP. All instances were run
to optimality.
392 J. J. A. Kortekaas et al.

5 Results and Discussion

Table 2 shows the relative performance of the DCAV fleet compared to the single-
purpose fleet. The first column features key fleet management performance indi-
cators and the subsequent column “Avg. Diff.” shows the overall average per-
formance difference for all instances for both fleets. The columns “BLS”, “PPS”,
“FPS”, and “COS” show the performance difference for the corresponding scenar-
ios. The last column shows the number of instances in which better results were
achieved using the DCAV fleet. The on-duty utilization rate shows an increase
of 4.82% for all scenarios with 241 instances resulting in better values for the
DCAV fleet. Most of the performance gain is made in the complementary scenario
(COS), where carriers can service both demand types consecutively. The occu-
pancy rate of vehicles is slightly lower for all scenarios, which can be attributed
to the increased use of modules and the difference in extra ride times and travel
distance when compared to the single-purpose fleet. In 329 instances the per-
centage of requests served is similar or better than the single-purpose fleet. This
difference might be caused by the coupling and decoupling delays of the DCAV
fleet, costing additional time and resulting in time windows of requests not being
met. Serving fewer requests can also contribute to the −1.30% average decrease
in vehicle occupancy. The DCAV fleet show a decrease in the average waiting
times of requests, except for the complimentary scenario. The latter can again
be attributed to the couple and decouple times of the DCAV fleet, having to
switch modules when the generated requests change from passenger to freight
requests. Lower waiting times for the other scenarios can be explained by the
7.58% increased use of load-carrying vehicles by the DCAV fleet. When time win-
dows allow, the DCAV fleet uses more modules to be able to handle requests of a
similar type more quicker. The complementary scenario (COS) therefore shows
a much smaller increase in the number of load-carrying vehicles. The DCAV
average fleet size is 18.77% lower, with the complementary scenario (COS) fea-
turing the largest drop—42.60% DCAV are needed. This lower total fleet size is
reflected in the total vehicle utilization: the active carriers are used 23.75% more
than the active single-purpose AVs.
The average fleet size and total travel distance across scenarios can be seen
in Fig. 5. Comparing the number of vehicles shown in the first graph we see that
the modular fleet uses more modules of each type while using fewer carriers on
average. Due to the coupling and decoupling time, the number of modules might
be higher than the equivalent single-purpose vehicles, however, the difference is
mostly caused by an increase in the use of passenger modules. Another explana-
tion could be that the modular fleet enables the use of more passenger modules
when carriers are available to more quickly serve passenger requests since the
model is not minimizing the number of modules used. The difference in total
modules used and the number of carriers used does show that the modular sys-
tem makes use of its flexibility. The lower average travel distance of the modules
can be explained by the use of parking locations since modules are being coupled
closer to request locations.
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 393

Fig. 5. Average fleet size and distance traveled for all fleet types across the four demand
distribution scenarios.
Table 2. Performance difference of the modular DCAV fleet compared to the single-
purpose fleet. The final column shows the number of ran instances that show better
performance for the DCAV fleet.

Performance Indicator Avg. Diff BLS PPS FPS COS #DCAV best
Utilization rate (on-duty) 4.82% −4.80% 3.82% −1.14% 24.36% 241/360
Utilization rate (total) 23.75% 2.90% 16.82% 7.78% 74.71% 326/360
Occupancy rate −1.30% −1.49% −0.47% −2.35% −0.97% 240/360
Service rate −0.08% −0.15% 0.00% −0.15% 0.00% 329/360
Avg. pickup time −9.70% −22.23% −10.27% −8.17% 3.28% 219/360
Avg. ride times −4.57% −9.73% −1.72% −8.15% −1.51% 330/360
Fleet capacity 7.58% 13.6% 5.75% 9.74% 1.81% 266/360
Fleet size −18.77% −4.80% −17.62% −8.24% −42.60% 346/360
Distance traveled −1.52% −1.27% −1.32% −1.21% −2.38% 218/360

6 Conclusion
This study has presented a MILP formulation for the Dynamically Configurable
Autonomous Vehicle Pickup and Delivery Problem (DCAVPDP). The model pro-
vides the first quantitative method investigating the benefits of a DCAV-based
transport system, which leverages automation and a novel automotive modu-
lar design concept to service multiple demand types by dynamically configuring
vehicle upper compartments. We compare the performance of a single-purpose
fleet consisting of dedicated passenger and freight vehicles and a DCAV fleet con-
394 J. J. A. Kortekaas et al.

sisting of motorized autonomous carriers and switchable passenger- and freight-

tailored compartments to service a mixed pickup and delivery heterogeneous
demand comprised of passenger and freight requests considering various scenar-
ios. Since autonomous carriers are expected to be the most expensive component
of a DCAVPDP system, we have aimed at achieving high utilization rates for
these assets. In total, 360 different instances were run for each fleet configura-
tion. The average performance of all tested mixed-demand scenarios shows an
improvement in average modular fleet utilization (carriers coupled to modules)
in relation to average single-purpose mixed fleet utilization up to 24.36% when
there is a distinct separation between demand types throughout the operation
horizon (complementary scenario). The proposed model can be adapted to sup-
port more compartment types with varying capacities and features, giving rise
to a highly customizable transportation service. Future research will focus on
developing a dynamic problem formulation, the adoption of crowd-sourced het-
erogeneous lower and upper parts, and the repositioning of carriers and modules
throughout parking locations anticipating future demand.

References
1. Beirigo, B.A., Schulte, F., Negenborn, R.R.: Integrating people and freight
transportation using shared autonomous vehicles with compartments. IFAC-
PapersOnLine 51, 392–397 (2018). https://doi.org/10.1016/j.ifacol.2018.07.064
2. Beirigo, B.A., Schulte, F., Negenborn, R.R.: A learning-based optimization app-
roach for autonomous ridesharing platforms with service-level contracts and on-
demand hiring of idle vehicles. Transp. Sci. 56(3), 677–703 (2022). https://doi.
org/10.1287/trsc.2021.1069
3. Cheng, R., Jiang, Y., Nielsen, O.A.: Integrated people-and-goods transportation
systems: from a literature review to a general framework for future research, pp.
1–24 (2023). https://doi.org/10.1080/01441647.2023.2189322
4. Continental: Continental launches bee in the urban world of the future (2017).
https://www.continental.com/en/press/press-releases/bee-96328. Accessed 28
June 2023
5. Daimler: Vision urbanetic (2019). https://media.mercedes-benz.com/article/
c518445b-9050-42ec-9be1-3446e1bf3973. Accessed 28 June 2023
6. Drexl, M.: On some generalized routing problems. Ph.D. thesis, Faculty of Busi-
ness and Economics, RWTH Aachen (2007). http://publications.rwth-aachen.de/
record/62536
7. Hörl, S., Ruch, C., Becker, F., Frazzoli, E., Axhausen, K.: Fleet operational policies
for automated mobility: a simulation assessment for Zurich. Transp. Res. Part C:
Emerg. Technol. 102, 20–31 (2019). https://doi.org/10.1016/j.trc.2019.02.020
8. Li, B., Krushinsky, D., Reijers, H.A., Van Woensel, T.: The share-a-ride problem:
people and parcels sharing taxis. Eur. J. Oper. Res. 238(1), 31–40 (2014). https://
doi.org/10.1016/j.ejor.2014.03.003
9. Mourad, A., Puchinger, J., Chu, C.: A survey of models and algorithms for optimiz-
ing shared mobility. Transp. Res. Part B: Methodol. 123, 323–346 (2019). https://
doi.org/10.1016/j.trb.2019.02.003
10. Münster, M., et al.: U-shift vehicle concept: modular on the road. In: Bargende,
M., Reuss, H.-C., Wagner, A. (eds.) 21. Internationales Stuttgarter Symposium.
 Combined Transport Using Dynamically Configurable Autonomous Vehicles 395

P, pp. 333–346. Springer, Wiesbaden (2021). https://doi.org/10.1007/978-3-658-

33466-6_23
11. Şahin, T., Raulf, C., Kızgın, V., Huth, T., Vietor, T.: A cross-domain system
architecture model of dynamically configurable autonomous vehicles. In: Bargende,
M., Reuss, H.-C., Wagner, A. (eds.) 21. Internationales Stuttgarter Symposium,
pp. 13–31. Springer, Wiesbaden (2021). https://doi.org/10.1007/978-3-658-33521-
2_2
12. Sun, S., Wong, Y.D., Liu, X., Rau, A.: Exploration of an integrated automated
public transportation system. Transp. Res. Interdisc. Perspect. 8, 100275 (2020).
https://doi.org/10.1016/j.trip.2020.100275
13. van der Tholen, M., Beirigo, B.A., Jovanova, J., Schulte, F.: The share-a-ride prob-
lem with integrated routing and design decisions: the case of mixed-purpose shared
autonomous vehicles. In: Mes, M., Lalla-Ruiz, E., Voß, S. (eds.) ICCL 2021. LNCS,
vol. 13004, pp. 347–361. Springer, Cham (2021). https://doi.org/10.1007/978-3-
030-87672-2_23
14. Toyota: E-Pallete (2023). https://www.toyota-europe.com/startyourimpossible/e-
palette. Accessed 28 June 2023
15. Ulrich, C., Friedrich, H.E., Weimer, J., Schmid, S.A.: New operating strategies
for an on-the-road modular, electric and autonomous vehicle concept in urban
transportation. World Electr. Veh. J. 10(4), 91 (2019). https://doi.org/10.3390/
wevj10040091
 Impact of Public Transport Development
on Health Care Services in Rural Areas

Joachim R. Daduna and Daniel Philipp(B)

Berlin School of Economics and Law, Badensche Straße 52, 10825 Berlin, Germany
{daduna,Daniel.Philipp}@hwr-berlin.de

Abstract. In the context of cohesion policy, one of the main points is the divergent
development between urban and rural areas. One critical problem is the undersup-
ply of rural areas due to low demand potential. In addition, there are inadequate
public transport services for many residents who, for various reasons, do not
have access to private motorized transport. With technical advances in the field of
road transport vehicles and comprehensive digitization, autonomous driving has
opened up opportunities in recent years for development of new (flexible) forms of
service based on quasi-individual mobility. The technical foundations required in
this regard as well as the organizational and legal framework conditions are high-
lighted and critically analyzed. Using the example of health care in rural areas, a
mobility concept is outlined that offers a suitable solution in conjunction with a
hierarchically centralized location structure of medical care facilities. This app-
roach is also suitable for reducing further structural deficits and supporting the
creation of equal living conditions on a broader level, also with a view to broad
participation in social and community life in rural areas.

Keywords: Autonomous Driving · Health Care · Rural Areas

1 Introduction
With the Second Industrial Revolution [17, 96], profound changes began (especially in
the central European regions and North America) in the settlement and economic struc-
tures that had until then been largely dominated by agriculture and handicraft businesses
The starting point for industrial developments was, in addition to new technologies, espe-
cially innovations in the transport sector. Thus, with the emergence of rail transport ([99]
74–86; [103] 565–573; [54] 234–244) for the first time an efficient transport mode was
available for freight and passenger transport, the network design and operation of which
were largely independent of geographical, topographical and weather-related influences.
Very quickly, the (technical) performance capabilities as well as cost reductions of about
90% became apparent in comparison to the previously existing modes of land and water
transport ([102] 95–96).
This was associated with the emergence of industrial areas and later also of service
centers with a constantly increasing demand for labor, which had to be met mainly from
rural residents or by immigration. This led to an extreme population growth in urban

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 396–415, 2023.
https://doi.org/10.1007/978-3-031-43612-3_25
 Impact of Public Transport Development on Health Care Services 397

regions ([54] 135–136), which inevitably brought about changes in social, economic
and work structures. An essential prerequisite, however, was a fundamental change in
agriculture ([46] 635–654). Significant increases in production with a lower demand for
labor due to efficiency improvements on the agricultural estates and the transition from
subsistence farms, which were widespread at the time, to commercial farming ([54]
141–157) or, through corresponding imports, to ensure the food basis in urban regions.
These developments have led to a divergence of living conditions between urban and
rural areas over many years, which has resulted in considerable inequalities in social par-
ticipation [68] and economic structures. This situation has been and can still be observed
worldwide, with varying degrees of intensity. However, it is politically undesired, espe-
cially because of the inherent potential for social conflict, so that suitable concepts
and approaches have been discussed for some time with the objective of eliminating
structural disproportions and enabling equal living conditions. Thus, one of the prior-
ity objectives in the European Union (EU) is the implementation of a cohesion policy
[27], increasingly also in the context of national measures of a regional policy [19]. For
Germany, this has been written down in a legally binding way in Sect. 1 of the Regional
Planning Act (ROG) for several years. One of the measures in this regard is to ensure
that mobility for residents living in rural areas meets their needs, which is currently
inadequate, ([6] 29–32). The emerging transition to autonomous driving in road traffic
[79] and the possibility of a quasi-individualization of mobility in public transport (see
Sect. 3) is a development step to enable profound changes.
In this paper, we show the extent to which new forms of mobility can be used in rural
areas, to sustainably improve existing structural problems in health care, since these are
of major importance for the population there. First, some aspects of spatial planning, in
rural areas are considered. Subsequently, the operational and technical design options
for improving mobility through the expansion of local and regional public transport
will be addressed, with a focus on autonomous driving vehicles. In the following, it
will be analyzed to what extent these measures can guarantee demand-oriented access
to health care services for all population groups in rural areas. A key point here is the
extent to which decentralized structures are desirable and feasible on the one hand, and
centralization is necessary on the other, also from an economic and technological point
of view. Finally, expectations and further development are discussed.

2 Spatial Planning and the Theory of Central Places

Due to political and also economic considerations, at the end of the 1960s [30] the devel-
opment objectives for rural areas were already aligned with the hierarchically oriented
centrality principle as described by Christaller ([16], 63–85) (see Fig. 1). According to
§ 2 para. 2, no. 2 ROG, this principle has also been a legally mandatory regulation in
spatial planning in Germany since 2008. The underlying idea is the centralization of
supply options and services ([39, 70], 223–241), taking into account existing settlement
and demand structures and the associated holding costs, also with a view to the provision
of public services, as well as the realization of a politically desired cohesion [23]. The
provision of services to the inhabitants in this concept refers not only to the central places
themselves, but also to their surrounding area, which is assigned to the respective hier-
archical level. It is not unproblematic here that access to services at higher levels can be
398 J. R. Daduna and D. Philipp

associated with increasing distances for parts of the inhabitants, depending on the loca-
tion of the respective higher-level locations. With regard to accessibility, the availability
of sufficient mobility, not only in the local area, is therefore absolutely mandatory.

Fig. 1. Basic structure of the Christaller model

The problem of service provision in rural areas is clearly shown by the structural
allocation of areas at the European level, starting from the NUTS classification (Level
3: Small regions for specific diagnoses) ([87] 13). This shows that densely populated
as well as rural regions with aggregation approaches clearly dominate the European
settlement structures. However, there is a certain problem with this (relatively rough)
classification, which does not reflect the significant regional specific differentiation. In
addition, there are also a number of different classification concepts, also for Germany.
Moreover, it must be taken into account that the framework conditions in the respec-
tive counties differ greatly in some cases, depending on demographic developments as
well as on local labor market situation and economic framework conditions ([6] 30–33).
In addition, there are regional location factors, especially with regard to dominant forms
of economic use. An additional influencing factor is the spatial closeness of a county to
(industrial) agglomerations or densification spaces, which can have inducing centrality
effects. In order to ensure an adequate service structure for the population in rural areas,
the availability of sufficient mobility is absolutely necessary. However, there are con-
siderable deficits in this regard. One major reason for this is the technical developments
in the field of individual mobility since the mid-1950s ([59, 97] 30–32). The steadily
growing availability of private cars has led to reductions in public transport, as demand
for it has become increasingly low. In 2017, the modal split in rural areas with small town
and village structures was only 5%, compared to 70% for private motorized transport
([75], 40). Also, more than one third of all families here own at least two cars (30% have
two cars, 6% three and more) ([75], 35).
This shows that a not inconsiderable share of the population in rural areas has no
or only limited access to mobility [58]. Essentially, these are people without a driver’s
license (especially young people), people who cannot drive a vehicle due to their health
situation or age, and people who do not own a car for financial reasons [68, 90] An
 Impact of Public Transport Development on Health Care Services 399

expansion of the actually only rudimentary public transport service is not economically
feasible due to the subsidies required for this [35]. This means that ultimately there is
no sufficient mobility available and thus the framework conditions are not at all in place
to be able to guarantee even approximately equivalent living conditions.
For some years now, developments toward autonomous driving have been achieved
in the motor vehicle sector, also in conjunction with significant advances in digitization.
Based on the current status, this technology will be fully ready for the market before the
end of this decade, as shown by the ongoing research work worldwide as well as many
test projects and initial implementations [67]. On this basis, there is the possibility of
significantly and sustainably improving public transport in rural areas with new services
and thus coming more closely to the goal of equal living conditions.

3 Impact of Autonomous Driving on Public Transport in Rural

Areas

The known public transport services with minibuses and standard buses cannot be an
adequate solution for operating in rural areas, both in terms of suitable structures and, in
particular, for financial reasons. In their current structures, they cannot ensure sufficient
quality of service ([6, 78] 29–32). For this reason, there have already been attempts
several decades ago to develop better alternatives based on demand-oriented services
for rural areas [9, 21, 66], but also in the suburban transition zones [86, 100]. A number
of concepts have been developed, not only in Germany, and have also been tested in
some cases, but often not implemented in regular operations.
The main reasons for this were the complexity of operational monitoring and control
as well as the inadequate communication capabilities between provider and customer.
In addition, there were labor law restrictions regarding to the driving personnel, which
led to limitations in the operational planning of vehicle and duty schedules and also to
considerable costs [66]. The necessary political support from local authorities was not
always available due to the introduction of competing services for the already econom-
ically weakened public transport companies owned by the municipalities. In addition,
there was resistance from the trade unions due to the flexible working time models
associated with the modified forms of service [66].
For a number of years, therefore, considerable effort has been made to develop
on-demand services with autonomous driving vehicles [67]. This also applies to broad
research projects in the field of general public road traffic, such as the Niedersachsen
Test Field project [3]. Standard legal framework conditions as a basis for the certification
and operation of such vehicles have now been laid down in the Road Traffic Act and in
the Autonomous Vehicles Approval and Operation Regulation of June 24, 2022.
Significant advantages for a service design based on such innovative operating con-
cepts in public transport are the elimination of costs for the driving personnel as well as the
partly (time-related) very restrictive labor law limitations. Added to this are far-reaching
developments in vehicle technology, widespread digitization in operational monitoring
and control as well as regarding information management. On this basis, more and more
projects have been initiated worldwide [1, 42], where autonomous driving plays a sig-
nificant role. One project in Germany is the public transport service with minibuses in
400 J. R. Daduna and D. Philipp

the area of Bad Birnbach [7]. In the following, the development of autonomous driving
is outlined first and then the design options are discussed.

3.1 State of Development of Autonomous Driving

The most important basis for the definition of autonomous driving is a classification
based on the J 3016 standard of SAE International [88]. This includes six levels from
SAE Level 0 to SAE Level 5, which define the different functions between the driver and
the automated system (0: No form of automation, 1: Use of automated driver assistance
systems (ADAS), 2: Partial automation, 3: Conditional automation, 4: High automation,
5: Full automation (autonomous driving). The final objective is to bring vehicles with
level 5 into service [52] that means, all vehicle movements are controlled and monitored
by automats. The essential aspect, which is to be seen as the basis of a quasi-individual
mobility, is that point-to-point connections are possible, in which a vehicle can drive
directly to a customer depending on demand.
The main foundations have been laid in recent years as part of the development
of ADAS components [1, 4], that are already being used in practice. These include
supporting functions such as the automatic speed adaptation, lane departure warning,
adaptive cruise control, blind spot information system, collision avoidance, moving
object detection, intelligent speed adoption and traffic sign recognition. In addition,
there are ad hoc communication networks such as vehicle-to-vehicle (V2V) or vehicle-
to-infrastructure (V2I) networks for a more efficient and environmentally oriented traffic
flow management and to increase traffic safety [8, 107].
Automated detection systems installed at specific positions on the vehicle are used
to continuously record the surrounding data. These include optical systems (cameras),
radar-based sensors, and lightwave-based 3D techniques (e.g., 3D laser systems and
LiDAR systems) [1, 47, 108]. The destination positioning as well as the routing are based
on data from (digital) Geographic Information Systems (GIS) [48, 50] in connection with
Global Navigation Satellite Systems (GNSS), such as the US American Differential
Global Positioning System (DGPS) or the system Galileo developed in the EU [45, 48,
50]. Applications of autonomous driving in open area are also already possible at a
contemporary time [80]. Based on the current state of development and the first steps
toward realization, it can be assumed that the essential prerequisites for autonomous
driving are in place in view of the technical developments. Deficits exist primarily in
some necessary adjustments to the legal framework and also in the not always existing
social acceptance, in spite of all the significant advantages.

3.2 Design Options in Public Transport

On the basis of autonomously driving vehicles, completely new design options arise
for public transport, especially for services in rural areas [15, 71, 94]. In addition to
the available technical bases, it is the replacement of the driver by intelligent automats
that enables the independence from, in particular, labor law restrictions and drastically
reduces the previously incurred personnel costs. The essential step here is a demand-
oriented 24/7 service, through which quasi-individual mobility can be provided in rural
 Impact of Public Transport Development on Health Care Services 401

areas. However, some legal issues need to be clarified regarding the integration of these
services into the legal framework of the Passenger Transportation Act (PBefG) [83].
In the foreground are four basic types of services that can be applied here, point-to-
point connections, on-demand based line and line band services as well as area services.
For point-to-point connections, services can also be provided by robotaxis [43, 55] and in
the framework of free-floating carsharing services [49] while in the other cases the focus
will be on the use of minibuses and midibuses [7, 74]. In addition, ridesharing concepts
[33, 64] can also be integrated, especially with regard to local feeder transport. However,
the introduction of these services cannot imply that the existing ones are replaced, as it
has been formulated, for example, by [31] with the convergence theory. In fact, they are
a necessary supplement for a service in the area, but not a substitution of the existing
regional bus services and of services in the regional rail passenger transport, which will
continue to be absolutely necessary within the framework of commuter services due to
the demand structures. However, exceptions exist [91] if demand on certain lines or line
sections does not allow for economic operation and cannot be expected to do so in the
future.
This means that in order to strengthen public transport in rural areas, it is absolutely
necessary to combine the existing scheduled with possible on-demand services in order
to be able to guarantee an adequately efficient public transport there. The main objective
must be an extension of the service, especially in regions where it has not been feasible
so far due to the existing framework conditions. With the new forms of service, suitably
adapted organizational and operational structures are also required. The essential aspects
in this context are outlined in a rough framework in the following section.

3.3 Necessary Framework Conditions for an Implementation of Changed

Structures in Public Passenger Transport
In order to successfully implement the presented concept for improving public transport
in rural areas, the necessary framework conditions must be provided. This concerns
the legal framework [1, 56] as a mandatory prerequisite for the start of operations in
the changed structures, insurance and liability aspects [63, 93, 95], cybersecurity, and
also the legal question regarding the use of the recorded data [2, 38, 53]. However,
the sometimes highly specified technical content, which requires continual adaptation
due to further developments, is problematic. In addition, the necessary organizational
structures and a suitable legal basis for cross-modal cooperation must be developed [13,
29, 104]. Linked to this are also the questions of financing and tariff structure design
[35, 57], especially with regard to the economic conditions. Finally, the operational
structures as well as the control and monitoring of the operational processes [20] must
be designed and a case-specific (internal and external) information management [20]
must be developed.
Only if appropriate framework conditions can be developed and established as well
as stable structures can also be ensured in the long term it will be possible to implement
attractive public transport services in rural areas. The resulting advantages are multilay-
ered because in addition to the broad services of a quasi-individual mobility for residents
without own access to private car use, a change is offered for many previous non-users
of public transport. This leads to a reduction in car ownership and mileage, as well
402 J. R. Daduna and D. Philipp

as a reduction in environmental impact when using vehicles with e-traction. Positive

financial effects also result for households in rural areas, where the previously necessary
multiple ownership of private cars (see the data in [75], 35) can be reduced.

3.4 Acceptance of Autonomous Driving and Ethical Discussions

In addition, the question of acceptance by potential users also plays a decisive role with
regard to successful market penetration in the future. A certain skepticism is evident in
(mostly hypothetical) surveys [5, 11, 60, 77]. Focus here is on the one hand on potential
benefits, but on the other hand risks are also stated, although it must be considered
that the discussions are being held in a far-reaching changing in social, technical and
economic environment. Furthermore, from a psychological point of view, the importance
of private vehicle ownership as a decades-old symbol of individual freedom must also be
seen, which will be lost in a sharing economy. In the context of surveys accompanying
projects [10], however the result of the surveys looks somewhat different, possibly due
to direct contact with the new form of (quasi-individual) mobility. In addition, it must
also be taken into account whether there is a sufficient supply of public transport or
whether, as is the case for a not negligible proportion of the population in rural areas,
use is largely restricted and there is a significant improvement in the situation from the
point of view of those affected.
Recurring claims regarding an increased safety risk due to autonomous driving also
do not meet reality. Looking at the current situation, the introduction of autonomous
driving is more likely to have the opposite effect. Evaluations of accidents show, among
other things also worldwide, that with more than 90%, the overwhelming share of acci-
dent causes lies in human error [72]. For Germany, in 2021, for police recorded accidents
involving personal injury, the share was 90.7%, with 88.0% caused by vehicle drivers
and 2.7% caused by pedestrians ([26] 49).
The probability of such accidents occurring in autonomous driving vehicles is
marginal, since they are based on clearly defined (AI-based) reaction patterns that are
designed on the basis of risk-averse control processes. This is based on high-performance
technical components of ADAS [1, 4], which are clearly over human information taking
and processing in critical situations. As a result, there is the possibility of significant
damage limitation with considerable reductions in accident-related costs as well as costs
to society as a whole in the health and social sectors, which amount to billions of euros.
However, complete accident prevention will not be possible since human error of traffic
participants can in principle not be completely avoided.
However, the situation is problematic in the case of automated decision-making in
conflict situations [62, 85], which are repeatedly mentioned from an ethical perspective
as an argument against the introduction of autonomous driving [37, 44, 92]. However,
this overlooks the fact that dilemmatic ad hoc decisions are not a problem inherent of
autonomous driving [61]. They occur in manual driving too, although it can be assumed
that the responsiveness of automats is significantly better, that means, the extent of dam-
age will be much smaller. In principle, no criminal law relevance can be assumed in
connection with the programming of automats, as the remarks in [18] show. With regard
to autonomous driving, however, there is also a mandatory need for clear political deci-
sions and the resulting legal regulations and measures. Objections from an ethical (and
 Impact of Public Transport Development on Health Care Services 403

also ideological) point of view may not play a role in this context, because philosophical
brain games cannot ultimately replace decisions and their implementation, nor can they
cause damage.

4 Opportunities for Structural Changes in the Design of Health

Care in Rural Areas
The general practitioner as a country doctor continues to play a central role in rural
health care, although it has not yet been possible to achieve the desired level of care
despite funding programs. In Mecklenburg-Vorpommern (MV), a proportion of medical
university seats is allocated as an advance quota to applicants who undertake to work
for at least ten years in general practitioner care in under-served or potentially future
underserved regions, as well as regions with a special local demand for care, after
completing their studies and subsequent further training as a specialist.
In addition, the increasing ageing of the population has to be seen, which poses further
challenges to the comprehensive health care system. Technically possible innovations
in telemedicine, such as the use of virtual consultation hours, can be seen as support
for general health care, but do not replace personal doctor contact ([32], 141–142). It
should be noted, however, that the acceptance of this form of medical care has increased
among both the population and physicians, essentially due to the contact restrictions
during the COVID 19 pandemic. In 2017 less than 2% of doctors in MV offered video
consultations, whereas in May 2020 this proportion had risen to around 50% ([32], 103).
Whether the use of virtual consultation hours will continue to the same extent, however,
remains to be seen. An essential point here is to what extent and when the mobility
offers outlined above will be available to visit a physician’s office. In order to achieve a
continuation and possibly also an expansion of telemedical services after the pandemic,
a comprehensive digital infrastructure is mandatory and the existing technical deficits
have to be solved. In addition, there is the problem of insufficient knowledge in han-
dling of technical devices, especially of older people. To further improve acceptance and
feasibility, it must be ensured that recorded data and images can be transmitted in suffi-
cient quality. Nevertheless, examinations and treatments which, due to their specificity,
require comprehensive specialist experience and possibly also technical equipment that
is associated with higher costs, are in need of other structures. Besides the already more
centralized structures in the area of inpatient care, additional outpatient forms must be
developed and established.
The outlined problem will be explained in the following using three examples, the
organization of dialysis, the use of imaging techniques such as Magnetic Resonance
Imaging (MRI) and Computer Tomography (CT) and AI-based skin cancer screening.
With regard to the care structures in rural areas, the focus is on the conflict between a
high level of service expected by the population, sufficient services as well as economic
efficiency of this provision.

4.1 Cost-Intensive Treatments and the Tendency Towards Centralization

In order to implement the postulate of equal living conditions, the necessary infrastruc-
ture facilities (hospitals and other forms of medical services) must also be accessible to
404 J. R. Daduna and D. Philipp

patients in rural areas at a reasonable cost. In order to ensure nationwide coverage in

Germany, planners have long sought to decentralize hospital operations and distribute
smaller care units over large areas [73]. This led to the German hospital landscape being
characterized by a high number of small and less specialized hospitals by international
standards, many beds (approx. 50% more than the EU average [36], many inpatient
cases (also approx. 50% more than the EU average [13] and correspondingly high staff
requirements. However, such a structure can no longer be effective in view of the grow-
ing disparity in the population structure be-tween rural and urban regions and increasing
costs due to improved treatment methods. This has been recognized in recent years, and
more recent reform efforts are again aiming at merging and centralizing essential care
services [89]. The population development influences not merely the supply in health
care system, since sufficient availability of specialized personnel within decentralized
structures is not given, but also the demand, since expected medical expenditure increases
quantitatively and qualitatively with patient age [34].
Organization and Implementation of Dialysis Measures. Chronic kidney disease
is an individual risk factor for cardiovascular diseases, which can lead to physical and
cognitive impairment and also significantly reduce life expectancy. Despite general prac-
titioner care, people with impaired kidney function are at risk of developing end-stage
renal disease or specific complications. For these patients, kidney transplantation or
lifelong renal replacement (dialysis) is often the only treatment option. Both treatments
cause very high costs, with terminal renal failure being considered one of the most
expensive chronic diseases. Projections show that in Germany, about 100,200 patients
were permanently in need of dialysis in 2017 [41], a rate of 1 in 800 citizens. The average
annual costs are between 43,000 e and 54,000 e per patient [41]. Of this, an average
of between 6300 e and 9800 e is accounted for by the necessary travel costs. The total
costs that burden just the statutory health insurance amount to about 4.73 billion e annu-
ally. Forecasts assume an increase in the number of patients in dialysis by 20 to 23% by
2040 [41]. This implies an increase in the affected patient group to 120,000 to 123,000.
The most common dialysis methods are peritoneal and haemodialysis. While peritoneal
dialysis can also be performed with mobile devices, the latter one usually has to be per-
formed as a stationary procedure. Most machines for this are large and of-ten designed
for use in a clinic [105]. For this reason, haemodialysis in Germany usually takes place
as some kind of center dialysis [40, 41], yet it is the dominant form in Germany. The
group of over 80-year-olds shows a particularly high prevalence; the needs-based care
of this group must also be viewed critically from the aspect that they often have severely
limited mobility [82].
However, according to analyses, the total annual costs, as far as the technical service
part of the practice is concerned, are on average about 990,000 e per service contract.
This does not include the costs for private patients. The Standardized Assessment Scale
(EBM) is only based on costs of about 740,000 e [28]. This casts doubt on the economic
efficiency of provision by low-frequency outpatient practices. In addition to the provision
of outpatient care in dialysis centers, a large proportion of patients already receive
their bloodwashing in hospitals. For MV, the Klinikliste 2023 of the internet platform
Klinikradar refers to twenty-one clinics for dialysis in twenty cities. Therefor Klinikradar
uses the quality reports of the hospitals of the Joint Federal Committee (G-BA). For
 Impact of Public Transport Development on Health Care Services 405

the year 2020, 922 cases are indicated in the overview for Rostock, whereas the least
frequented clinics each have less than 10 treatments to report.
In addition to the burden of high costs, there is an increasing shortage of specialists
and expertise to secure adequate care for patients. A look at the development of the
overall situation of nursing staff in the nephrology sector in relation to the increasing
numbers of elderly people shows a clear discrepancy. At the same time, there are fewer
and fewer classical nephrology nursing staff in relation to the patients to be cared for,
also because of the now increasing retirement of veteran nursing staff [84]. As there will
tend to be more and more patients requiring renal replacement therapy such as dialysis
due to demographic change, nationwide care is a problem.
Radiological Methods of Treatment. According to the German Radiology Network,
about 6,800 of the licensed physicians in Germany have completed training as special-
ists in radiology/diagnostic radiology and also work in this field. With about 2% of all
licensed physicians, they form one of the smaller groups of specialists. Approximately
3,100 radiologists are active in the established sector, with further dovetailing between
the outpatient and inpatient sectors. Looking at the different practice sizes, it becomes
clear that the number of radiology individual practices is declining. Of the approximately
1,100 radiology centers in Germany, fewer than 300 are still run as individual practices.
This is a significant decrease compared to the approximately 400 individual practices in
2016 [76], thus a trend toward the necessary centralization is evident here. The remain-
der are already group practices. Around 600 radiologists work in group practices with
two partners and just under 900 radiologists are organized in group practices with 3–5
partners. With almost 350 radiologists, a not insignificant proportion is accounted for
large group practices with six or more partners. Such large practices can be compared to
medium-sized companies. They often have up to 50 employees and operate equipment
worth well over 10 million euros in some cases. Private practice radiology is currently in a
state of flux, and consolidation is taking place with the formation of large, supra-regional
practice conglomerates [76].
Nationwide radiological care without teleradiology is already hard to implement
today [69]. One of the main reasons for this success lies in the working methodology
in the field of radiology. For example, an essential part of the work of radiologists is
predominantly the interpretation of medical images, which are produced by medical-
technical assistants and transmitted digitally by the equipment, be it conventional X-ray
machines, CT or MRI devices, to be viewed and evaluated on a suitable monitor. Here it
becomes clear that radiological care, in comparison to other medical disciplines, depends
less on the universal availability of specialists, but rather on the availability of sufficiently
trained (nursing) staff and the necessary technical equipment. Few radiological proce-
dures, such as interventional radiology, conventional X-ray fluoroscopy or sonography,
absolutely require the on-site presence of radiologists [69]. Since the number of radio-
logical examinations in Germany is growing by about 10% annually, but only 2–3% of
newly trained radiologists are coming on board [81], we already see modern approaches
to solutions here. However, the basis for this is the provision of technical equipment
both in the field of data collection as also in data processing and transmission. Another
essential point is to ensure data security both for the transmission and the archiving of
406 J. R. Daduna and D. Philipp

data. Combined with the growing need for medical-technical assistants [25], this form
of universal care is facing problems.
For example, if one considers the imaging diagnostic procedures performed in MV
alone, the extent of the demand becomes apparent. The Federal Health Report con-
tains approx. 6,794,000 CT examinations and approx. 1,919,000 MRI examinations of
full inpatients in hospitals for 2021. Compared to 2020, this is an increase of about
321,000 CT examinations and 50,000 MRI examinations in one year. This might hint
that the importance of imaging diagnostics, especially CT examinations, has increased
in Germany in the recent years. The path already taken towards partial centralization
of radiological treatments in centers or larger practices offers the possibility of exploit-
ing the synergies associated with the increasing size of the practices in terms of staff
deployment, equipment procurement, service contracts, etc. [76].
Skin Cancer Detection. Malignant neoplasms of the skin (black skin cancer) are among
the most common types of cancer worldwide. The frequency is increasing rapidly, but
it is easily treatable if detected early [24]. Routine diagnostics for melanoma detec-
tion include full-body visual inspection, often supplemented by dermoscopy, which can
significantly increase the diagnostic accuracy of experienced dermatologists [24]. A pro-
cedure that is additionally offered in some practices and clinics is Total Body Digital
Photography (TBDP) in combination with digital dermoscopy. TBDP is also used in
smartphone apps that serve the self-assessment of the skin by the user. The problem here
is the lack of consistency in the image and light quality of the images used [65].
In recent decades, several non-invasive methods have been developed for the ex-
amination of suspicious pigmented lesions that have the potential to allow improved
and, in some cases, automated assessment of these lesions. Currently, the most promising
approach may be the analysis of routine macroscopic and dermoscopic images by AI [51].
In the classification of pigmented skin lesions based on macroscopic and dermoscopic
images, AI, especially in the form of neural networks, has achieved diagnostic accuracy
comparable to dermatologists in numerous experimental studies. Prerequisites for the
transfer of such diagnostic systems into dermatological routine are the comprehensibility
of the system decisions by the user as well as a uniformly high performance of the
algorithms on image data from other clinics and practices.
As seen in recent studies, computer-aided diagnosis systems unfold the greatest
benefit primarily as assistance systems [12, 101]. The physician/machine combination
achieves the best results and AI-based diagnostic systems help to record morphological
features quickly, quantitatively, objectively and reproducibly and could thus, in addition
to medical experience, provide a more objective basis for analysis. Looking at the num-
bers delivered by the National Association of Statutory Health Insurance Physicians
(KBV) from 2020 illustrates the problem. If one compares, for example, the density
of dermatologists in the spatial planning regions of Mecklenburgische-Seenplatte (MS),
Mittleres Mecklenburg/Rostock (MM), Vorpommern (VP), Westmecklenburg (WM) and
Berlin, at first glance the picture is homogeneous; MS has 6.1 dermatologists per 100,000
inhabitants, MM 7.6, VP 7.4, WM 5.2 and Berlin 7.6. However, if one converts the physi-
cian density to the area, different ratios result. Berlin has 0.3125 dermatologists per km2 ,
while the region of MS has 0.0029 dermatologists per km2 (MM 0.009; VP 0.0048; WM
0.0034). This means that, calculated on the area, dermatologist care in Berlin is 109 times
 Impact of Public Transport Development on Health Care Services 407

higher than in the area of MS. This implies significantly longer journeys for the patients
concerned, while mobility deficits are also evident here.
Ongoing Hospital Reform. In 2018, the G-BA defined minimum structural require-
ments for emergency care in three levels for the first time. This was an important first
step towards a nationwide uniform definition of care structures and the creation of trans-
parency. In December 2022 the Government Commission for Modern and Needs-Based
Hospital Care has presented a reform proposal for hospital remuneration and a fixed
definition of structural prerequisites in the sense of a minimum required hospital-wide
structural quality [14]. The reform proposals of the government commission refer to
inpatient and emergency care. Nevertheless, it also responds to problems, experiences
and developments that can also be found in the outpatient sector. The politically forced
restructuring of the hospital landscape, insofar as it is implemented, leads to a further
centralization of hospital locations. The three examples described above are primar-
ily long-term plannable and recurring treatments. Nevertheless, the coming structural
change is an opportunity to further dovetail outpatient and inpatient care.

4.2 Concept of Area-Wide Hierarchically Structured Service

The implementation of a network with the objective of a comprehensive (hierarchically
oriented) location structure for ensuring medical care for the population in rural areas is
an essential and also necessary step within the framework of the realization of a targeted
cohesion policy. The initial point is to consider central-location structures with a suitable
differentiated scope of care [16, 23]. The essential condition in such location planning
is the specification of the observance with a maximum (L 2 norm based) distance r (=
radius), which is specified for the inhabitants to reach (potential) health care facilities.
These are set covering problems [22, 106], where catchment areas are defined, i.e., from
any location within a given area, the respective facilities must be accessible according
to the specifications. Insofar as information is available regarding the road distances or
travel times, the assignment to the locations can also be calculated on spatially aggre-
gated settlement clusters [98, 106]. Outgoing from this an online-based management of
medical care capacities can be established in a later phase, with respective planning of
the necessary transport processes for the patients affected.
In the problem presented in this context, two specific framework conditions must
be taken into account. For one thing, the location planning for the individual levels is
carried out top down, whereby the locations of level i are set for the planning of level
i-1, that means, the functions of a lower level are in case included in those above it ([16]
63–85). In addition, it must be taken into account that the area under consideration,
the federal state of MV, is not spatially isolated, but is bordered by Schleswig-Holstein
and Niedersachsen to the west and Brandenburg to the south. As a result, it may make
sense that, to a certain extent, health care in the outlying areas can be provided from
appropriate locations outside the borders of the region, also from the point of view of
cost and distance. A possible three-tier structure with r = 75 km, r = 40 km and r = 30
is shown in Figs. 2, 3 and 4.
This approach shows that (largely) covering the area of the federal state can be
ensured with the given radii on the three levels at r = 75 km with two sites, at r =
408 J. R. Daduna and D. Philipp

Fig. 2. Included locations at r = 75

Fig. 3. Included locations at r = 40

Fig. 4. Included locations at r = 30

40 km it is additionally six and at r = 30 km again nine more (see Table 1). Only in the
western and southern peripheral areas are there smaller gaps, but these can be covered
in cooperation with the neighboring federal states of Schleswig-Holstein (Lübeck) and
Niedersachsen (Lüneburg) in the west and Brandenburg (Prenzlau) in the south. In the
area of the Baltic Sea coast, on the other hand, complete coverage is required, due among
other things to the geographic conditions, so that the Fischland-Darß-Zingst peninsula
and the island of Rügen can be included in the coverage.
This solution does not include the existing settlement structures in detail, the distance-
based accessibility of the locations within the existing transport networks and also
 Impact of Public Transport Development on Health Care Services 409

Table 1. Locations included at the respective central levels

Radius r Included locations

75 km Schwerin, Greifswald
40 km Stralsund, Rostock, Torgelow, Neubrandenburg, Waren (Müritz), Ludwigslust
30 km Wolgast, Bergen, Ribnitz-Damgarten, Wismar, Hagenow, Güstrow, Parchim,
Neustrelitz, Demmin

(regional political) influences. In this respect, the outlined structure must be understood
as an initial basis for discussion, which must be specified more precisely at the political
level and the responsible technical levels.

4.3 Necessary Framework Conditions for the Implementation

of Demand-Oriented Structures in Health Care in Rural Areas
Fundamental changes in health care as proposed in the outlined concept require (long-
term) political decisions. The results must be implemented administratively and, what
is often forgotten, also enforced (politically). The central basis is the necessary legal
framework with clear regulation of responsibilities at the different administrative levels,
as well as at the local authority level.
In this context, however, additional measures are required to create the necessary
framework conditions. This may involve structural changes to the healthcare system
and its long-term financing, the establishment and expansion of a suitable digital infras-
tructure for data management, and also communication between (potential) patients and
healthcare institutions. It is also necessary to regulate the control and monitoring of
transport processes in public transport in order to ensure the necessary (and also barrier-
free) access to healthcare for all potential patients. It must also be taken into account
that the structures to be developed will be influenced by ongoing adjustments due to
changing framework conditions. In the foreground here are developments in population
and settlement structures as well as in the demand structures (for example due to demo-
graphic factors). In addition, there are often (technical) developments in diagnostics and
therapeutic methods that can have an impact on the care structures, that means, no static
structure should be developed; instead, permanent further development must be ensured
in the light of changing situations.

5 Expectations and Further Development

The main objective in connection with the introduction of autonomous driving as part
of a redesign of mobility in rural areas is to improve living conditions of the local
population and to increase the attractiveness and economic performance of these regions.
A key point here is the health care service considered in this article, which, also in the
view of demographic trends that are apparent, must be ensured for the coming years
to the necessary extent and with the needed efficiency. In view of the EU’s plans to
regularly review the driving licenses of older citizens, the need for the availability of
quasi-individual mobility as an alternative to private car use will increase significantly.
410 J. R. Daduna and D. Philipp

In addition, the above-mentioned changes in public transport must also be seen in

terms of meeting the needs of rural areas in a wide variety of sectors. This concerns the
accessibility of retail facilities and (commercial) services as well as public administra-
tion. In the field of education and access to cultural and leisure facilities, too, completely
new opportunities arise for the inhabitants of rural areas. Substitution by online-based
offerings, as has been propagated many times and is still being pushed today, cannot
provide a sustainable solution here, since personal contact is of decisive importance in
many situations. However, it can be achieved through the mobility services outlined
above, resulting in substantial positive effects in terms of participation in social life in
rural areas. In this way, an important step can also be taken toward ensuring equal living
conditions, which has been called for over and over again.

References
1. Ainsalu, J., et al.: State of the art of automated buses. Sustainability 10(9), 3118 (2018)
2. Arzt, C., Kleemann, S., Plappert, C., Rieke, R., Zelle, D.: Datenverarbeitung und Cyber-
sicherheit in der Fahrzeugautomatisierung - Rechtliche und technische Anforderungen im
Verbund. Zeitschrift für IT-Recht und Recht der Digitalisierung 25(7), 593–614 (2022)
3. Asbach, L., Ortgiese, M.: Automatisierte Fahrzeuge in Europa sicher betreiben. Interna-
tionales Verkehrswesen 75(1), 22–26 (2023)
4. Ayyasamy, S.: A comprehensive review on advanced driver assistance system. J. Soft
Comput. Paradigm 4(2), 69–81 (2022)
5. Avermann, N., Schlüter, J.: Determinants of customer satisfaction with a true door-to-door
DRT service in rural Germany. Res. Transp. Bus. Manag. 32, 100420 (2019)
6. Baba, L., Wilbert, K.: Soziale Stadt-und Ortsentwicklung in ländlichen Räumen. Bundesin-
stitut für Bau-, Stadt-und Raumforschung (BBSR) (2022)
7. Barillère-Scholz, M., Büttner, C., Becker, A.: Mobilität 4.0: Deutschlands erste autonome
Buslinie in Bad Birnbach als Pionierleistung für neue Verkehrskonzepte. In: Riener, A.,
Appel, A., Dorner, W., Huber, T., Kolb, J.C., Wagner, H. (Hrsg.) Autonome Shuttlebusse
im ÖPNV, pp. 15–22. Springer, Heidelberg (2020). https://doi.org/10.1007/978-3-662-594
06-3_2
8. Bento, L.C., Parafita, R., Rakha, H.A., Nunes, U.J.: A study of the environmental impacts
of intelligent automated vehicle control at intersections via V2V and V2I communications.
J. Intell. Transp. Syst. 23(1), 41–59 (2019)
9. Berg, J., Ihlström, J.: The importance of public transport for mobility and everyday activities
among rural residents. Soc. Sci. 8, 58 (2019)
10. Bernhard, C., Oberfeld, D., Hoffmann, C., Weismüller, D., Hecht, H.: User acceptance of
automated public transport - valence of an autonomous minibus experience. Transp. Res. F
70, 109–123 (2020)
11. Bissell, D., Birtchnell, T., Elliott, A., Hsu, E.L.: Autonomous automobilities - the social
impacts of driverless vehicles. Curr. Sociol. 68(1), 116–134 (2020)
12. Brinker, T.J., Schlager, G., French, L.E., et al.: Computerassistierte Hautkrebsdiagnose.
Hautarzt 71, 669–676 (2020)
13. Busse, R., Berger, E.: Vom planerischen Bestandsschutz zum bedarfsorientierten Kranken-
hausangebot. In: Klauber, J., Geraedts, M., Friedrich, J., Wasem, J. (Hrsg.) Krankenhaus-
Report 2018 - Schwerpunkt Bedarf und Bedarfsgerechtigkeit, pp. 149–170. Schattauer,
Stuttgart (2018)
 Impact of Public Transport Development on Health Care Services 411

14. Busse, R., Karagiannidis, C., Augurzky, B., Schmitt, J., Bschor, T.: Der Vorschlag der
Regierungskommission für eine grundlegende Reform der Krankenhausvergütung. In:
Klauber, J., Geraedts, M., Friedrich, J., Wasem, J. (Hrsg.) Krankenhaus-Report 2023 - Schw-
erpunkt - Personal, pp. 267–280. Springer, Heidelberg (2023). https://doi.org/10.1007/978-
3-662-66881-8_17
15. Chen, X., Ouyang, Y., Arian, E., Yang, H., Ba, X.: Modeling and testing autonomous and
shared multimodal mobility services for low-density rural areas. Civil Engineering Studies,
Illinois Center for Transportation, Series No. 22-013 (2022)
16. Christaller, W.: Die zentralen Orte in Süddeutschland. Gustav Fischer, Jena (Neudruck 1968:
Wissenschaftliche Buchgesellschaft, Darmstadt) (1933)
17. Coluccia, D.: The second industrial revolution (late 1800s and early 1900s). In: Zanda,
G. (ed.) Corporate Management in a Knowledge-Based Economy, pp. 52–64. Palgrave
Macmillan, London (2012)
18. Cramer, T.: Strafrechtsdogmatische Auflösung dilemmatischer Situationen beim autonomen
Fahren - Entwicklung innovativer Leitlinien für die Programmierung von Notstandsalgorith-
men in Lebensnotstandssituationen. Nomos, Baden-Baden (2022)
19. Crescenzi, R., Fratesi, U., Monastiriotis, V.: Back to the member states? Cohesion Policy
and the national challenges to the European Union. Reg. Stud. 54(1), 5–9 (2020)
20. Daduna, J.R.: Impacts of deregulation on planning processes and information management
design in public transit. In: Voß, S., Daduna, J.R. (eds.) Computer-aided scheduling of public
transport, pp. 429–441. Springer, Berlin (2001). https://doi.org/10.1007/978-3-642-56423-
9_24
21. Daduna, J.R.: Evolution of public transport in rural areas - new technologies and digitization.
In: Marcus, A., Rosenzweig, E. (eds.) HCII 2020. LNCS, vol. 12202, pp. 82–99. Springer,
Cham (2020). https://doi.org/10.1007/978-3-030-49757-6_6
22. Daskin, M.S.: A maximum expected covering location model - formulation, properties and
heuristic solution. Transp. Sci. 17(1), 48–70 (1983)
23. Danielzyk, R., Priebs, A.: Die Sicherung der Daseinsvorsorge durch Klein-und Mittelstädte
als Beitrag des zentralörtlichen Systems zu gleichwertigen Lebensverhältnissen. Europa
Reg. 26(3), 7–21 (2020)
24. de Buhr, Y., Bunde, H., Großmann, E., et al.: Prävention von Hautkrebs. Onkologe 28, 49–54
(2022)
25. Deutscher Bundestag: Reformbedarf bei Medizinisch-Technischen Assistenten in Kranken-
häusern und Arztpraxen. Drucksache 19/24629 (2020)
26. destatis - Statistisches Bundesamt: Verkehr - Verkehrsunfälle - Fachserie 8 Reihe 7 (2022)
27. Di Caro, P., Fratesi, U.: One policy, different effects - estimating the region-specific impacts
of EU cohesion policy. J. Reg. Sci. 62(1), 307–330 (2022)
28. Drabinski, T.: Nephrologische Versorgung in Deutschland - Bestandsaufnahme, Vergütung
und Datengrundlage. Schmidt & Klaunig, Kiel (2014)
29. Eckhardt, J., Lauhkonen, A., Aapaoja, A.: Impact assessment of rural PPP MaaS pilots. Eur.
Transp. Res. Rev. 12(1), 1–14 (2020)
30. Einig, K., Zaspel-Heisters, B.: Das system Zentraler Orte in Deutschland. In: Greiving,
S., Flex, F. (Hrsg.) Neuaufstellung des Zentrale-Orte-Konzepts in Nordrhein-Westfalen,
pp. 3–19. Akademie für Raumforschung und Landesplanung (ARL) Hannover (2016)
31. Enoch, M.P.: How a rapid modal convergence into a universal automated taxi service could
be the future for local passenger transport. Technol. Anal. Strateg. Manag. 27(8), 910–924
(2015)
32. Enquete-Kommission Zukunft der medizinischen Versorgung in Mecklenburg-
Vorpommern: Abschlussbericht. Stellungnahmen 7. Sitzung. Kommissionsdrucksache
7/46 (2021)
412 J. R. Daduna and D. Philipp

33. Farhan, J., Chen, T.D.: Impact of ridesharing on operational efficiency of shared autonomous
electric vehicle fleet. Transp. Res. C 93, 310–321 (2018)
34. Frank, J., Ryll, A.: Demographischer Wandel, Morbidität und Inanspruchnahme - Zur
Projektion des ambulanten Leistungsvolumens. Gesundheitswesen 77(12), 932–938 (2015)
35. Gail, M.M., Götz, G., Herold, D., et al.: Staatliche Eingriffe zur Erfüllung der Daseinsvor-
sorge - Der öffentliche Personennahverkehr. In: Gail, M.M., Götz, G., Klotz, P.A., et al.
(Hrsg.) Staatliche Eingriffe in die Preisbildung - Darstellung, Erklärung und Wirkungs-
analyse, pp. 49–66. Springer, Wiesbaden (2023). https://doi.org/10.1007/978-3-658-406
74-5_5
36. Geissler, A., Wörz, M., Busse, R.: Deutsche Krankenhauskapazitäten im internationalen
Vergleich. In: Klauber, J., Geraedts, M., Friedrich, J. (Hrsg.) Krankenhaus-Report 2010,
pp. 25–40. Schattauer, Stuttgart (2010)
37. Gill, T.: Ethical dilemmas are really important to potential adopters of autonomous vehicles.
Ethics Inf. Technol. 23(4), 657–673 (2021). https://doi.org/10.1007/s10676-021-09605-y
38. Girdhar, M., Hong, J., Moore, J.: Cybersecurity of autonomous vehicles - a systematic
literature review of adversarial attacks and defense models. IEEE Open J. Veh. Technol.
(2023, in press)
39. Greiving, S., Flex, F., Terfrüchte, T.: Vergleichende Untersuchung der Zentrale-Orte-
Konzepte in den Ländern und Empfehlungen zu ihrer Weiterentwicklung. Raumforsch.
Raumordn. 73(4), 285–297 (2015)
40. Haag-Weber, M.: Peritonealdialyse. Dialyse aktuell 21(06), 290–295 (2017)
41. Häckl, D., Kossack, N., Schoenfelder, T.: Prävalenz, Kosten der Versorgung und Formen des
dialysepflichtigen chronischen Nierenversagens in Deutschland - Vergleich der Dialysever-
sorgung innerhalb und außerhalb stationärer Pflegeeinrichtungen. Das Gesundheitswesen
83(10), 818–828 (2021)
42. Hänsch, R., Hoelzmann, J., Mielke, S.: Flächendeckend Rufbusse? Eine Untersuchung zur
Verbreitung alternativer Mobilitätskonzepte. Der Nahverkehr 37(1+2), 39–43 (2019)
43. Hallewell, M., et al.: Deriving user experience design dimensions for future autonomous
taxi interface design. J. Usability Stud. 17(4), 140–163 (2022)
44. Hansson, S.O., Belin, M.Å., Lundgren, B.: Self-driving vehicles - an ethical overview. Philos.
Technol. 34, 1383–1408 (2021)
45. Hein, G.W.: Status, perspectives and trends of satellite navigation. Satell. Navig. 1(1), 1–12
(2020). https://doi.org/10.1186/s43020-020-00023-x
46. Henning, F.W.: Handbuch der Wirtschafts-und Sozialgeschichte Deutschlands, Band 2,
Deutsche Wirtschafts- und Sozialgeschichte im 19. Jahrhundert. Schöningh, Paderborn
(1996)
47. Holzhüter, H., Bödewadt, J., Bayesteh, S., Aschinger, A., Blume, H.: Technical concepts of
automotive LiDAR sensors - a review. Opt. Eng. 62(3), 031213 (2023)
48. Hou, Z., Zhou, F.: Assessing the performance of precise point positioning (PPP) with the
fully serviceable multi-GNSS constellations. Remote Sens. 15, 807 (2023)
49. Illgen, S., Höck, M.: Establishing car sharing services in rural areas - a simulation-based
fleet operations analysis. Transportation 47(2), 811–826 (2020)
50. Jing, H., Gao, Y., Shahbeigi, S., Dianati, M.: Integrity monitoring of GNSS/INS based
positioning systems for autonomous vehicles - state-of-the-art and open challenges. IEEE
Trans. Intell. Transp. Syst. 23(9), 14166–14187 (2022)
51. Jutzi, T., Krieghoff-Henning, E.I., Brinker, T.J.: Künstliche Intelligenz auf dem Vormarsch
- Hohe Vorhersage-genauigkeit bei der Früherkennung pigmentierter Melanome. Laryngo-
Rhino-Otologie 84–91 (2022)
52. Khan, M.A., Sayed, H.E., Malik, S., et al.: Level-5 autonomous driving - Are we there yet? A
review of research literature. ACM Comput. Surv. (CSUR) 55(2), 1–38 (2022)
 Impact of Public Transport Development on Health Care Services 413

53. Khan, S.K., Shiwakoti, N., Stasinopoulos, P., Warren, M.: Modelling cybersecurity regula-
tions for automated vehicles. Accid. Anal. Prev. 186, 107054 (2023)
54. Kiesewetter, H.: Industrielle Revolution in Deutschland - Regionen als Wachstumsmotoren.
Franz Steiner, Stuttgart (2004)
55. Kim, S., Chang, J.J.E., Park, H.H., et al.: Autonomous taxi service design and user
experience. Int. J. Hum.-Comput. Interact. 36(5), 429–448 (2020)
56. Kleemann, S., Arzt, C.: Das Gesetz zum autonomen Fahren in Deutschland - Automatisiertes
Fahren findet seinen Weg im StVG. RAW Recht-Automobil-Wirtschaft 99–195 (2021)
57. Knauff, M.: Verfassungs-, europa-und personenbeförderungsrechtliche Aspekte des Vor-
rangs der Eigenwirtschaftlichkeit. In: Knauff, M. (Hrsg.) Vorrang der Eigenwirtschaft-
lichkeit im ÖPNV, pp. 11–28. Nomos, Baden-Baden (2017)
58. Knierim, L., Schlüter, J.C.: The attitude of potentially less mobile people towards demand
responsive transport in a rural area in central Germany. J. Transp. Geogr. 96, 103202 (2021)
59. Kopper, C.: Der Durchbruch des PKW zum Massenkonsumgut 1950–1964. Jahrbuch für
Wirtschaftsgeschichte 51(1), 19–36 (2010)
60. López-Lambas, M.E., Alonso, A.: The driverless bus - an analysis of public perceptions and
acceptability. Sustainability 11(18), 4986 (2019)
61. Lucifora, C., Grasso, G.M., Perconti, P., Plebe, A.: Moral dilemmas in self-driving cars.
Rivista Internazionale di Filosofia e Psicologia 11(2), 238–250 (2020)
62. Lütge, C., Kriebitz, A., Max, R.: Ethische und rechtliche Herausforderungen des autonomen
Fahrens. In: Mainzer, K. (Hrsg.) Philosophisches Handbuch Künstliche Intelligenz, pp. 1–18.
Springer, Wiesbaden (2019). https://doi.org/10.1007/978-3-658-23715-8_30-1
63. Lutz, L.S.: Haftung am Beispiel automatisierter Fahrzeuge. In: Hartmann, M. (Hrsg.) KI &
Recht kompakt, pp. 117–142. Springer, Berlin (2020). https://doi.org/10.1007/978-3-662-
61700-7_3
64. Manik, D., Molkenthin, N.: Topology dependence of on-demand ride-sharing. Appl. Netw.
Sci. 5, 1–16 (2020)
65. Marek, A.J., Chu, E.Y., Ming, M., Khan, Z., Kovarik, C.L.: Impact of a smartphone appli-
cation on skin self-examination rates in patients who are new to total body photography: a
randomized controlled trial. J. Am. Acad. Dermatol. 79(3), 564–567 (2018)
66. Meyer, M.: Bedarfsorientierte ÖPNV-Bedienung in der Fläche - Erfahrungen mit neuar-
tigen Betriebsformen in der Bundesrepublik Deutschland und in den Niederlanden.
Verkehr+Technik 41(7), 280–289 (1988)
67. Michalik, D., Kohl, P., Krause, M., Kummert, A.: Virtueller Demonstrator zur Unterstützung
der Einführung des autonomen Fahrens am Beispiel automatisierter Shuttlebusse. In: Proff,
H. (Hrsg.) Transforming mobility - What next? - Technische und betriebswirtschaftliche
Aspekte, pp. 399–418. Springer, Wiesbaden (2022). https://doi.org/10.1007/978-3-658-
36430-4_24
68. Milakis, D., van Wee, B.: Implications of vehicle automation for accessibility and social
inclusion of people on low income, people with physical and sensory disabilities, and
older people. In: Antoniou, C., Efthymiou, D., Chaniotakis, E. (eds.) Demand for Emerging
Transportation Systems, pp. 61–73. Elsevier, Amsterdam (2020)
69. Möller, T.B.: Teleradiologie in Deutschland. In: Fischer, F., Krämer, A. (Hrsg.) eHealth in
Deutschland, pp. 295–305. Springer, Heidelberg (2016). https://doi.org/10.1007/978-3-662-
49504-9_15
70. Nakoinz, O.: Zentralität - Theorie, Methoden und Fallbeispiele zur Analyse zentraler Orte.
Edition Topoi, Berlin (2019)
71. Nanchen, B., Ramseyer, R., Grèzes, S., et al.: Perceptions of people with special needs
regarding autonomous vehicles and implication on the design of Mobility as a Service to
foster social inclusion. Front. Hum. Dyn. 3, 751258 (2022)
414 J. R. Daduna and D. Philipp

72. Naveen, M.K., Lokesh, M., Varma, K.N.J.L., Teja, P.S.V.: A review on autonomous vehicles
and its components. J. Pharm. Negative Results 13(7), 6916–6922 (2022)
73. Neubauer, G., Beivers, A.: Zur Situation der stationären Versorgung: Optimierung unter
schwierigen Rahmenbedingungen. In: Klauber, J., Geraedts, M., Friedrich, J. (Hrsg.)
Krankenhaus-Report 2010: Schwerpunkt Krankenhausversorgung in der Krise?, pp. 3–11.
Schattauer, Stuttgart (2010)
74. Ng, M.T., Mahmassani, H.S.: Autonomous minibus service with semi-on-demand routes in
Grid Networks. Transp. Res. Rec. 2677(1), 178–200 (2023)
75. Nobis, C., Kuhnimhof, T.: Mobilität in Deutschland - MiD Ergebnisbericht. (Bundesminister
für Verkehr und digitale Infrastruktur) (FE-Nr. 70904/15) Bonn, Berlin (2018)
76. Notogamiprodjo, M., Remplik, P., Schricke, U.: Radiologische Praxis - Terra incognita.
Radiologe 61, 671–675 (2021)
77. Othman, K.: Public acceptance and perception of autonomous vehicles - a comprehensive
review. AI Ethics 1(3), 355–387 (2021)
78. Pangbourne, K.: Challenge, coordination, and collaboration for effective rural mobility solu-
tions. In: Amaral, A.M., Barreto, L., Baltazar, S., Silva, J.P., Gonçalves, L. (eds.) Implications
of Mobility as a Service (MaaS) in Urban and Rural Environments - Emerging Research and
Opportunities, pp. 83–108. IGI Global, Hershey (2020)
79. Parekh, D., et al.: A review on autonomous vehicles - progress, methods and challenges.
Electronics 11(14), 2162 (2022)
80. Patoliya, J., Mewada, H., Hassaballah, M., Khan, M.A., Kadry, S.: A robust autonomous
navigation and mapping system based on GPS and LiDAR data for unconstraint environment.
Earth Sci. Inf. 15, 2703–2715 (2022)
81. Quinsten, A.S., Bontke, B., Saur, A.: Radiologie im Wandel - Remote Scanning als Lösung
gegen den Fachkräftemangel. Radiopraxis 15(02), 62–71 (2022)
82. Rechel, B., Grundy, E., Robine, J.M., et al.: Ageing in the European Union. Lancet
381(9874), 1312–1322 (2013)
83. Saxinger, A.: Genehmigungsvoraussetzungen und -wirkungen von Linienbedarfsverkehren.
In: Knauff, M. (Hrsg.) Neuerungen im Rechtsrahmen für den ÖPNV, pp. 75–93. Nomos,
Baden-Baden (2022)
84. Schäfer, C.: Werden alle Möglichkeiten genutzt? Dialyse aktuell 26(9), 385 (2022)
85. Schäffner, V.: Wenn Ethik zum Programm wird - Eine risikoethische Analyse moralischer
Dilemmata des autonomen Fahrens. Zeitschrift für Ethik und Moralphilosophie 3(1), 27–49
(2020)
86. Schlüter, J., Bossert, A., Rössy, P., Kersting, M.: Impact assessment of autonomous demand
responsive transport as a link between urban and rural areas. Res. Transp. Bus. Manag. 39,
100613 (2021)
87. Schmidt-Seiwert, V., Binot, R., Kiel, L., et al.: Die Unterschiede bestimmen die Vielfalt in
Europa - Ein Atlas ausgewählter Aspekte der räumlichen Strukturen und Entwicklungen.
Bundesinstitut für Bau-, Stadt- und Raumforschung (BBSR), Bonn (2019)
88. Schreurs, M.A., Steuwer, S.D.: Autonomous driving - political, legal, social, and sustain-
ability dimensions. In: Maurer, M., Gerdes, J.C., Lenz, B., Winner, H. (eds.) Autonomous
Driving - Technical, Legal and Social Aspects, pp. 149–171. Springer, Heidelberg (2016).
https://doi.org/10.1007/978-3-662-45854-9_8
89. Schreyögg, J.: Verbesserung der Personalallokation durch Strukturwandel. In: Klauber, J.,
Wasem, J., Beivers, A., Mostert, C. (Hrsg.) Krankenhaus-Report 2023 - Schwerpunkt Per-
sonal, pp. 197–211. Springer, Heidelberg (2023). https://doi.org/10.1007/978-3-662-66881-
8_13
90. Shrestha, B.P., Millonig, A., Hounsell, N., McDonald, M.: Review of public transport needs
of older people in European context. J. Popul. Ageing 10, 343–361 (2017)
 Impact of Public Transport Development on Health Care Services 415

91. Sieber, L., Ruch, C., Hörl, S., Axhausen, K.W., Frazzoli, E.: Improved public transportation
in rural areas with self-driving cars - a study on the operation of Swiss train lines. Transp.
Res. A 134, 35–51 (2020)
92. Siegel, J., Pappas, G.: Morals, ethics, and the technology capabilities and limitations of
automated and self-driving vehicles. AI Soc. 1–14 (2021)
93. Simmler, M.: Maschinenethik und strafrechtliche Verantwortlichkeit. In: Bendel, O. (Hrsg.)
Handbuch Maschinenethik, pp. 453–469. Springer, Wiesbaden (2019). https://doi.org/10.
1007/978-3-658-17483-5_30
94. Sörensen, L., Bossert, A., Jokinen, J.P., Schlüter, J.: How much flexibility does rural public
transport need? - Implications from a fully flexible DRT system. Transp. Policy 100, 5–20
(2021)
95. Spindler, G.: Haftung für autonome Systeme. In: Arroyo Vendrell, T., Gsell, B., Kindl, J.
(Hrsg.) Die Schadensersatzhaftung, pp. 139–184. Nomos, Baden-Baden (2022)
96. Stearns, P.N.: The Industrial Revolution in World History, 4th edn. Westview Press, Boulder,
Co. (2020)
97. Südbeck, T.: Motorisierung, Verkehrsentwicklung und Verkehrspolitik in der Bundesrepublik
Deutschland der 1950er Jahre. (Steiner) Stuttgart (1994)
98. Tansel, B.C., Francis, R.L., Lowe, T.J.: State of the art—location on networks: a survey. Part
I: The p-center and p-median problems. Manag. Sci. 29(4), 482–497 (1983). https://doi.org/
10.1287/mnsc.29.4.482
99. Taylor, G.R.: The Transportation Revolution 1815–1860. Rinehart, New York (1951)
100. Thao, V.T., Imhof, S., von Arx, W.: Demand responsive transport - new insights from peri-
urban experiences. Travel Behav. Soc. 31, 141–150 (2023)
101. Tschandl, P., Codella, N., Akay, B.N., et al.: Comparison of the accuracy of human readers
versus machine-learning algorithms for pigmented skin lesion classification - an open, web-
based, international, diagnostic study. Lancet Oncol. 20(7), 938–947 (2019)
102. van Suntum, U.: Verkehrspolitik. (Vahlen) München (1986)
103. Voigt, F.: Verkehr - Die Entwicklung des Verkehrssystems (Band 2, Teil 1). Duncker &
Humblot, Berlin (1965)
104. Wachinger, L.: Mittel in Kooperationen und Verbünden. In: Knauff, M. (Hrsg.) Rechtsan-
wendung und Finanzierung im ÖPNV, pp. 193–204. Nomos, Baden-Baden (2020)
105. Wieringa, F.: Die Zukunft der Nierenersatztherapie. Nephrologie aktuell 27(1), 26–30 (2023)
106. Ye, H., Kim, H.: Locating healthcare facilities using a network-based covering location
problem. GeoJournal 81, 875–890 (2016)
107. Ye, L., Yamamoto, T.: Evaluating the impact of connected and autonomous vehicles on
traffic safety. Physica A 526, 121009 (2019)
108. Zhou, J.: A review of LiDAR sensor technologies for perception in automated driving. Acad.
J. Sci. Technol. 3(3), 255–261 (2022)
 Ridesharing in Rural Areas
with Autonomous Electric Vehicles
and Interrelated Trips

Marvin Soth, Lennart C. Johnsen, Sebastian Scholz(B) , and Frank Meisel

School of Economics and Business Management, Supply Chain Management,

Kiel University, Olshausenstr. 40, 24098 Kiel, Germany
[email protected]

Abstract. Passenger cars are responsible for 44% of the greenhouse gas
emissions within Europe’s transportation sector. Especially rural areas
are still fundamentally dependent on private motor vehicles, which is why
in particular the younger and older population in such regions is limited
in their mobility. The purpose of this paper is to determine the poten-
tial of shared, electric, autonomous vehicles to close the aforementioned
mobility gap in an eco-friendly manner. It provides recommendations for
the implementation of the associated transport services in terms of vehi-
cle battery size and charging station infrastructure. For this purpose,
we present a model for the electric autonomous dial-a-ride problem with
interrelated trips (e-ADARP-IT), which is solved by a commercial solver
and through a Variable Neighborhood Search heuristic. Computational
experiments reveal the superior performance of the heuristic. It is shown
that large shares of requested trips can be served efficiently at low opera-
tional cost. Results further indicate that expanding vehicle battery capac-
ity seems more relevant than expanding the number of charging stations,
even though both developments clearly have to go hand in hand.

Keywords: Dial-a-ride problem · electric autonomous vehicles · rural

area · interrelated trips · variable neighborhood search

1 Introduction
The transport sector is responsible for more than a quarter of the EU’s green-
house gas emissions. 44% of its share is generated by passenger vehicles and,
consequently, has a crucial impact on achieving the 1.5◦ C target of the Paris
Climate Agreement [13]. The combination of the three innovations of electrifi-
cation, autonomous driving, and shared mobility has the disruptive potential to
significantly change the personal mobility sector [44]. Together they form the
potential for shared autonomous electric vehicles (SAEVs). With the help of
SAEVs, greenhouse gas emissions per kilometer can be reduced by over 90% and
mobility cost reach 0.16–0.26 e/km, where both these performance measures lie
significantly below those of private vehicles with combustion engines [4].
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 416–434, 2023.
https://doi.org/10.1007/978-3-031-43612-3_26
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 417

The route planning problem associated with SAEV services is the dial-a-ride
problem (DARP), a combinatorial optimization problem. The DARP consists of
customer requests for transportation from given origin locations to given desti-
nation locations, which have to be scheduled on a finite number of vehicles. The
objective could be to maximize a service quality measure or to minimize trans-
portation costs [5]. SAEV services as considered in this paper combine multiple
DARP-variants within one problem: The dial-a-ride problem with interrelated
trips (DARP-IT), which synchronizes customer requests that are related to each
other by means of a common arrival time of multiple customers or a round trip
with a desired stay time at a location [27], and the electric autonomous dial-a-ride
problem (e-ADARP) that deals with electric and autonomous vehicles including
battery management and recharging [5]. This article combines these two vari-
ants, resulting in the electric autonomous dial-a-ride problem with interrelated
trips (e-ADARP-IT).
Most studies on DARPs for autonomous and electric vehicles focused on
urban areas [4,5,30]. In contrast to this, [27] considered the DARP-IT for rural
areas, where interrelated trips are of particular importance as customers also
demand guaranteed return trips, synchronized arrivals for joint meetings, or
other complex service patterns. The paper at hand also focuses on rural areas
and interrelated trips. Mobility in such areas is characterized by relatively long
transport distances and a very limited availability of public transport options.
This effects that the population so far relies on privately owned vehicles, which
heavily restricts those people that do not own or cannot operate a car on their
own, such as young or elderly people, see [27]. We, therefore, explore the potential
of SAEV in rural areas by combining the approaches of [27] for the ADARP-
IT and of [5] for the e-ADARP within a holistic model and solution approach.
The advances over [27] lie in the integration of battery management into the
dial-a-ride problem with interrelated trips (ADARP-IT) and thus bridges the
gap between these two research directions in the domain of dial-a-ride prob-
lems. A heuristic solution is adopted from [27] and compared to the solutions
derived from solving the optimization model directly. In our experiments, we
then focus on a rural area in the federal state of Schleswig-Holstein, Germany.
The associated test instances vary in terms of the number of customer requests,
the vehicles being available and their battery capacity, as well as the number
of charging stations. The influence of the number of charging stations and the
size of the battery capacity is then analyzed w.r.t. the service quality for the
interrelated trips in this rural area setting.
Therefore, the implementation of an autonomous ridesharing system in rural
areas can provide several advantages, for instance by offering mobility services
to people that cannot operate a vehicle themselves, by avoiding operational
limitations such as personnel costs and service time constraints of conventional
mobility services, and by allowing for more efficient vehicle sharing as well as
a provision of services during off-peak hours. As rural areas generally lack an
offer of public transport options, such a system can close the gap between the
demand for mobility and the supply of mobility services.
418 M. Soth et al.

The remainder of this paper is organized as follows. Section 2 gives an

overview of relevant SAEV literature. Section 3 formalizes the e-ADARP-IT
planning problem and presents the corresponding optimization model. Com-
putational experiments in Sect. 4 contrast the optimization model-based results
with a heuristic solution approach, focusing on the effects of battery capacity
and the number of available charging stations. Section 5 concludes the paper.

2 Literature Review
SAEV services combine features of electrification, autonomous driving, and
shared mobility, where autonomous vehicles offer economic and service time
advantages due to not having a driver [20]. This typically allows for uninter-
rupted service provision and greater flexibility in the deployment [5]. Thereby
’shared mobility’ refers to the on-demand use of vehicles by multiple customers
in contrast to privately owned vehicles that are exclusively used by their own-
ers. Shared mobility can take various forms such as ridesharing, carsharing, or
bike-sharing, and allows for cost savings, reduced greenhouse gas emissions, and
less reliance on vehicle ownership [43].
The route planning problem of shared autonomous vehicles (SAVs) is treated
within the DARP. The combinatorial optimization problem minimizes the trans-
portation cost and/or maximizes the service quality [5].
[36] classify the literature on SAV research based on booking type and sharing
system. The booking type divides into reservation-based and on-demand-based
services. In the latter, the customer can book his/her trip in real-time, and those
requests have to be integrated into the ongoing vehicle service process, see e.g.
[25]. The use of priority rules has been established for this purpose by [15].
In contrast, in reservation-based services, customer requests are known before
the vehicle routes are to be determined. This means that the DARP is solved
using the data is available up to that time, which typically leads to a static
perspective on this optimization problem, see [24] and [5]. Due to the longer
distances and longer relocation times of vehicles in rural areas, reservation-based
services appear suitable for mobility systems in such regions [27].
The mobility systems for SAVs can be divided into ridesharing, carshar-
ing, and mixed systems based on the way vehicles are shared in the operations
[36]. In carsharing, customers typically use a car one after the other, such that
requests are serviced in a sequential manner. In ridesharing, several customers
that demand transportation from/to similar locations might jointly use the same
vehicle. In this context, [9] investigate how to identify suitable meeting points
to which customers can walk over short distance from their actual origin loca-
tion or desired destination location. They incorporate this feature into a DARP
in order to reduce detouring of vehicles while offering attractive service routes
to customers. In [29], the authors focus on the dial-a-ride problem in an urban
road network with ride-sharing by autonomous taxis under consideration of traf-
fic congestion. They propose a non-linear integer programming model embedded
within a rolling horizon framework to optimize the routing of the taxis while
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 419

maximizing the profit of the system. A mixed system is characterized by cus-

tomers deciding whether to share the vehicle with other customers or use it
on their own. The subdivision of SAVs described above also applies to SAEVs.
Electric vehicles are more energy efficient and reduce CO2 emissions compared
to vehicles with internal combustion engines [14]. In the context of integrating
electric vehicles into public transport, finding routes of minimal energy consump-
tion, scheduling vehicle charging processes, and optimizing battery life have been
identified as novel challenges [3]. Related issues for SAEVs are battery manage-
ment and the provision of charging station infrastructure [30]. Aspects of electric
vehicle battery management have been investigated in many studies on different
vehicle routing problems. Examples include the electric and green vehicle routing
problem, e.g. [11,19,41], the hybrid electric traveling salesman problem [1,12],
and the vehicle routing problem with time windows and charging stations [2,22].
Also charging policies and charging functions have been investigated in detail
for various routing problems. The charging policy determines the amount of bat-
tery capacity that can or must be restored at a charging station [35]. Established
policies for the charging process include full charging [41], battery swapping [33],
or partial charging [11,16]. Next to these charging policies, the assumption can
be made that a charging process can be terminated prematurely when a new cus-
tomer request arrives [4,30]. To deal with the large number of factors affecting
the charging process, linear approximations of the charging function are often
adopted in research [16,28,35]. Furthermore, the majority of approaches assume
that the discharge of the battery occurs constantly over the travel time and that
the corresponding discharge rate can be derived from an energy consumption
model [17,19,39]. This assumption is typically valid if the vehicle load is neg-
ligible in comparison to the weight of the empty vehicle and if vehicles go at
a quite constant speed [5,11,40]. Regarding the deployment of charging station
infrastructure, different strategies exist, too. For example, [26] place charging
stations at cab stands or other points of interest whereas [4] use an elimination
strategy: charging stations are initially placed at all permissible locations and,
then, those stations with the least impact on the system are eliminated in an
iterative process. [31] proposes a two-stage solution approach, addressing the
dynamic vehicle charging scheduling problem with the objective of minimizing
the costs associated with daily charging operations for the fleet. The approach
consists of two components: the scheduling of daily vehicle charging and the
online assignment of vehicles to chargers. The above-mentioned elements of bat-
tery management are focused in [5] among others. The authors investigate the
use of SAEVs in urban environments by modelling an e-ADARP and solving it
through a branch-and-cut algorithm.
In our paper, we combine existing approaches from the literature. We inves-
tigate the e-ADARP-IT from the perspective of a mobility service provider, who
uses autonomous e-vehicles for offering reservation-based ridesharing services in
rural areas. We take up the modeling of battery management decisions of [5]
and include it into the ADARP-IT of [27]. The ADARP-IT has been established
to investigate SAV-services with interrelated trips in rural areas. Through the
420 M. Soth et al.

inclusion of aspects of electric vehicles management, we derive the e-ADARP-IT

that closes the gap between these two directions of DARP-research.
In addition, the research conducted opens a potential link to existing publica-
tions in the area of demand-responsive transportation. While [10] investigate the
benefits of meeting point usage for customer pick-up and drop-off, [23] evaluate
the performance of a dial-a-ride service using simulation. With this, operators of
transport systems can be provided with guidelines for designing their dial-a-ride
services by identifying the parameters that significantly influence the evaluation
criteria.

3 Modelling the E-ADARP-IT

In this section, we present the mathematical formulation for the e-ADARP-IT.
Subsection 3.1 introduces the different types of customer requests that are to
be considered. This is followed by introducing the used notation in Subsect.
3.2 and the presentation of the mixed-integer linear optimization model for the
e-ADARP-IT in Subsect. 3.3.

3.1 Classification of Customer Requests

In DARPs, customers place requests for being transported from a specified start-
ing point (pick-up location) to a specified destination point (drop-off location)
[8]. In classical variants of the DARP, each customer is assumed to place one
such request, without and interrelations among these requests existing, see [7].
We call such requests, conventional requests.
In our work, we also consider more complex trip request types, where two
or more trips are interrelated with each other. Such complex customer requests
were introduced by [27] and are distinguished into requests with serial relations
and requests with parallel relations. Serial relations are characterized by an
outbound and a return trip, combined with a stay time at the destination of the
outbound trip. The drop-off location of the first trip and the pick-up location of
the second trip are, thus, identical. Such trips play a role if a customer demands
a guaranteed return trip, which is particularly relevant in urban area settings.
The challenge in route planning is to schedule the drop-off of the first trip and
the pick-up of the second trip such that the desired stay time is guaranteed.
The second type of complex requests are those with parallel relations, which can
be further subdivided into parallel drop relations and parallel pick relations. In
this paper, only parallel drop relations are considered for reasons of brevity. A
characteristic of such a parallel request is that two or more trips share the same
drop-off location, for example as the involved customers want to meet there. The
complexity of parallel drop requests results from the need to have customers
arrive at the common drop-off location at approximately or exactly the same
time. As is described by [27], these mentioned customer request types can be
flexibly combined in a modular way, resulting in even more complex mobility
request types.
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 421

Table 1. Notation

Sets
K = {1, ..., m} Set of vehicles
O = {o1 , ..., om } Set of origin locations of vehicles
R = {1, ..., n} Set of trips
P = {p1 , ..., pn } Set of pick-up locations
D = {d1 , ..., dn } Set of drop-off locations
CS = {cs1 , ..., csl } Set of charging station locations
V = O∪P ∪D∪CS Set of all locations
Si ⊂ R Set of trips that serially follow trip i
Gi ⊂ R Set of trips with parallel drop-off relation with trip i
Parameters
P Ck Passenger capacity of vehicle k ∈ K
bk Energy storage capacity of vehicle k s battery
ok Origin location of vehicle k
pi Pick-up location of trip i ∈ R
di Drop-off location of trip i
qi Number of passengers boarding or disembarking at location i ∈ V
si Service time for passenger boarding and disembarking at location i
ei Earliest starting time for the service at location i
li Latest starting time for the service at location i
ūi Maximum travel time for trip i
δimin , δimax Minimum / maximum stay time at drop-off location of serial related trip i
ωimax Maximum waiting time for parallel related services of trip i
ti,j Vehicle travel time from location i to location j
ci,j Vehicle travel cost from location i to location j
βi,j Vehicle battery consumption for traveling from location i to j
αcs Charging rate at charging station cs ∈ CS (in kWh per minute)
γ Desired battery charging level
r Objective weight for maximizing the number of fulfilled trips
Decision variables
yi 1 if trip i ∈ R is served, 0 otherwise
xki,j 1 if vehicle k drives directly from i ∈ V to j ∈ V , 0 otherwise
zik 1 if vehicle k ends its tour at location i, 0 otherwise
Tik Arrival time of vehicle k at location i
Lki Number of passengers on vehicle k after leaving location i
Bik Battery charge level of vehicle k when arriving at location i
k
Ccs Charging time of vehicle k at charging station cs

3.2 Notation
We use the following notation for modelling the e-ADARP-IT (see also summary
in Table 1): The mobility service provider has a fleet of vehicles K = {1, ..., m},
where vehicle k ∈ K has a passenger capacity P C k , a battery capacity bk ,
and starts its operations from the origin location ok . These locations form set
422 M. Soth et al.

O = {o1 , . . . om }. Following the concept of a reservation-based service system, the

relevant information about customer requests is known to the service provider at
the time of the planning. More precisely, we denote by R = {1, ..., n} a set of trip
requests with corresponding pick-up location pi and drop-off location di of trip
i ∈ R. The pick-up and drop-off locations of all trips form sets P = {p1 , ..., pn }
and D = {d1 , ..., dn }, respectively. The set of available charging stations for
the electric vehicles is denoted CS = {cs1 , ..., csl }. The set of all locations in
a problem is denoted V = O ∪ P ∪ D ∪ CS. Due to the ridesharing concept, a
vehicle can serve different requests from multiple customers simultaneously. Each
location i ∈ V is thus characterized by a change in the number of customers qi
that occurs at this location. For pick-up locations pi , qpi takes a positive value
because customers are boarding. At drop-off locations di , customers leave the
car which is reflected by a corresponding negative value qdi = −qpi . For all
locations i ∈ O ∪ CS, we set qi = 0. In addition, service times si are incurred for
customers boarding at pick-up location pi and being dropped off at destination
location di . Furthermore, we denote by ei and si the earliest and latest time for
starting service at a location i. Also, the overall travel time for a trip i ∈ R is
bounded to ūi time units in order to restrict the prolonged travel time that the
customer might experience due to the detouring in the ridesharing system. For
a serial-related outbound trip i, we state by Si ⊂ R the set of those trip(s) that
have to follow on trip i subsequently. Furthermore, we denote by δimin and δimax
the least and the maximum stay time of the customer at the drop-off location
of trip i, respectively, before the subsequent trip(s) can be started. Finally, we
denote by Gi ⊂ R the set of those trips that have a parallel drop-off relation
with trip i, i.e., that should arrive at the destination at similar time. For this,
we denote by ωimax the maximum waiting time for the arrival of the customer(s)
belonging to trip i and accordingly for the trips in Gi .
Each charging station cs ∈ CS has a charging rate αcs . This rate defines the
electricity amount (in kWh per minute) that is taken up by a vehicle charging
k
at cs. The charging time Ccs of vehicle k at station cs is a decision variable
and determines the charged electricity quantity as αcs · Ccs k
. Furthermore, we
prescribe a target charging level γ that is not to be exceeded when charging
a vehicle to proactively restrict long charging times if desired. Finally, ci,j , ti,j ,
and βi,j represent the travel cost, travel time, and battery consumption of a
vehicle that goes directly from location i to location j. We assume that battery
consumption βi,j depends linearly on the travel time ti,j and is not influenced
by any further factors.
The decisions to be made are modeled through the following decision vari-
ables. The binary variable yi equals 1 if request i ∈ R is serviced, 0 otherwise.
Variable xkij equals 1 if vehicle k ∈ K travels directly from location i ∈ V to
location j ∈ V , 0 otherwise. Variable zik equals 1 if vehicle k ends its trip at
location i, 0 otherwise. The continuous variable Tik expresses the arrival time
of vehicle k at location i. The number of customers in vehicle k after leaving
location i is expressed by variable Lki . While continuous variable Bik denotes
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 423

k
the charge level of vehicle k’s battery when arriving at location i, variable Ccs
denotes the charging time of vehicle k at charging station cs ∈ CS.

3.3 Optimization Model for the E-ADARP-IT

The e-ADARP-IT is modeled through formulas (1) to (33). The objective func-
tion (1) primarily seeks to maximize the number of serviced requests, followed by
the secondary objective of minimizing the operational service cost. Coefficient r
serves as a weight in this combined objective function and ensures a hierarchy
of the two goals if set to a sufficiently large value [27].
  
max Z = r yi − xki,j · ci,j (1)
i∈R k∈K i∈V j∈V


xkok ,j = 1 k ∈ K (2)
j∈V

xuok ,j = 0 k ∈ K (3)
j∈V u∈K
u=k
 
xkj,i − xki,j = 0 k ∈ K, i ∈ P (4)
j∈V j∈V
j=i j=i
 
xkj,i − xki,j − zik = 0 k ∈ K, i ∈ D ∪ CS (5)
j∈V j∈V
j=i j=i

zik = 1 k ∈ K (6)
i∈D∪{ok }
 
xkpi ,j − xkj,di = 0 k ∈ K, i ∈ R (7)
j∈V j∈V
j=pi j=di

xkpi ,j ≥ yi i∈R (8)
j∈V k∈K

xki,i = 0 k ∈ K, i ∈ P ∪D∪CS (9)

Tjk ≥ Tik + si + ti,j − Mi,j (1 − xki,j ) k ∈ K, i, j ∈ V, i = j (10)
Tdki ≥ Tpki + spi + tpi ,di k ∈ K, i ∈ R (11)
ei ≤ Tik ≤ li k ∈ K, i ∈ V (12)
ui ≥ Tdki − Tpki − spi k ∈ K, i ∈ R (13)
Lkj ≥ Lki + qj − P C k (1 − xki,j ) k ∈ K, i, j ∈ V, i = j (14)
Lkj ≤ Lki + qj + P C k (1 − xki,j ) k ∈ K, i, j ∈ V, i = j (15)
0 ≤ Lki ≤ P C k k ∈ K, i ∈ P ∪ D (16)
Lki = 0 k ∈ K, i ∈ O ∪ CS (17)
y i = yj i ∈ R, j ∈ Si ∪ Gi (18)
424 M. Soth et al.
 
Tpuj ≥ Tdki + δimin −Md i,p (2 − xkl,di − xupj ,l ) k, u ∈ K, i ∈ R, j ∈ Si (19)
j
l∈V l∈V
 
Tpuj ≤ Tdki + δimax −Md i ,pj (2 − xkl,di − xupj ,l ) k, u ∈ K, i ∈ R, j ∈ Si (20)
l∈V l∈V
 
Tduj ≤ Tdki + ωimax −Md i ,dj (2 − xkl,di − xul,dj ) k, u ∈ K, i ∈ R, j ∈ Gi (21)
l∈V l∈V

0 ≤ Bik ≤ bk k ∈ K, i ∈ V (22)
Bokk = bk k∈K (23)
k ∈ K, i ∈ V \CS,
Bjk ≤ Bik − βi,j + bk (1 − xki,j ) (24)
j ∈ V \{ok }, i = j
k ∈ K, i ∈ V \CS,
Bjk ≥ Bik − βi,j − bk (1 − xki,j ) (25)
j ∈ V \{ok }, i = j
k ∈ K, j ∈ V,
Bjk ≤ Bcs
k
+ αcs · Ccs
k
− βcs,j + bk (1 − xkcs,j ) (26)
cs ∈ CS, j = cs
k ∈ K, j ∈ V,
Bjk ≥ Bcs
k
+ αcs · Ccs
k
− βcs,j − bk (1 − xkcs,j ) (27)
cs ∈ CS, j = cs
k
Ccs ≥ 0 k ∈ K, cs ∈ CS (28)
bk · γ ≥ k
Bcs + αcs · k
Ccs k ∈ K, cs ∈ CS (29)
k ∈ K, cs ∈ CS,
Tik ≤ Tcs
k
+ tcs,i + Ccs
k
+ Mcs,i
k
(1 − xkcs,i ) i ∈ D∪P ∪CS ∪{ok }, (30)
i = cs
k ∈ K, cs ∈ CS,
Tik ≥ Tcs
k
+ tcs,i + Ccs
k
− Mcs,i
k
(1 − xkcs,i ) i ∈ D∪P ∪CS ∪{ok }, (31)
i = cs
xki,j , zik ∈ {0, 1} k ∈ K, i, j ∈ V (32)
yi ∈ {0, 1} i∈R (33)

According to Constraints (2) and (3), each vehicle starts its tour at its own
origin location ok . Constraints (4) ensure the vehicle flow at pick-up locations.
Constraints (5) define this for drop-off locations and charging stations. This
constraint provides flexibility for the continuation of a route, as a vehicle k
may continue its tour to some other location j (xkij = 1) or end its tour at the
current location i (zik = 1). Constraints (6) then state that each vehicle route
has a defined end location. This location can also be the origin location ok if
the vehicle did not move at all (zokk = 1). Constraints (7) ensure that both the
pickup and the drop-off of a request are performed by one and the same vehicle.
Constraints (8) ensure that a request is considered served, only if the pickup
location has actually been visited by a vehicle. Constraints (9) avoid that this
requirement is fulfilled via trivial sub-cycles.
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 425

Constraints (10) to (13) address the time components of the solution. Con-
straints (10) propagate the vehicle arrival time from one visited location to the
next, where the value Mi,j is set to max(0, li + si + tij − ei ) as in [5]. Con-
straints (11) denote that a drop-off location di must be visited after the asso-
ciated pick-up location pi . Constraints (12) ensure the preset time windows for
visiting the locations. With i = ok and a corresponding value for eok , the time
at which vehicle k becomes available for the service system can be considered
in this constraint, too. The maximum travel duration of a trip is bounded by
Constraints (13).
Capacity restrictions (14) and (15) derive the number of customers Lkj
onboard vehicle k after leaving a visited location j. Constraints (16) respect
the vehicle capacity and Constraints (17) ensure that all vehicles are empty at
their origin locations and while charging.
From Constraints (18), complex requests are always served completely or not
at all, but never partially. More precisely, all interrelated trips j ∈ Si ∪ Gi are
served if trip i is served. Constraints (19) and (20) focus on serially related trips.
They ensure the minimum and maximum customer stay times between drop-off
and subsequent pick-up, where Md i ,pj = max(ldi +δimin −epj , lpj −edi −δimax , 0),
see [27]. Note that the trips in a serial request can be served by different vehicles
k and u (k = u). For example, vehicle k can drop off the customer at drop-
off location di and another vehicle u can later collect the customer at pick-up
location pj for the subsequent trip. Constraints (21) focus on the second type
of complex requests, the parallel related trips. The restriction synchronizes the
arrival times of parallel related trips i and j at the drop-off locations, where
Md i ,dj = max(ldj − edi − wimax , 0). For this, trip j ∈ Gi at the drop-off location
dj = di must not occur later than wimax time units after trip i. By setting wimax
and wjmax , flexibility or strictness can be reflected for these trips. For example, if
wimax = wjmax = 0, the customers must arrive at the drop-off location at exactly
the same time.
Constraints (22) to (31) are for battery management. According to Con-
straints (22), the battery level of a vehicle k must not be negative nor greater
than the battery capacity bk at any location i. Constraints (23) define that the
battery of vehicle k is fully charged at the origin location ok . The depletion of the
battery level when going from some location i (which is not a charging station)
to a subsequent location j is computed by Constraints (24) and (25). Constraints
(26) and (27) calculate the battery status of a vehicle that just visited a charging
station cs before going to a subsequent location j. These constraints take care of
the charged quantity αcs · Ccs k
. Charging times Ccs k
are non-negative according to
Constraints (28). Additionally, through Constraint (29), charging of a vehicle can
be restricted to not exceed a target level γ. This helps, for example, to proactively
avoid increasing charging times when approaching the maximum battery level.
Furthermore, the charging times of a vehicle at a station cs are included in the
propagation of the arrival time at the subsequently visited location i through
426 M. Soth et al.

Constraints (30) and (31), where Mi,s k

= max(0, li + si + ti,s − es ). Eventu-
ally, Constraints (32) and (33) assure the binary character of variables xki,j , zik ,
and yi .
The e-ADARP-IT model may be solved by any MIP-solver, provided that
the instances are not too large. As an alternative solution approach, we have
adopted the ADARP-IT VNS-heuristic of [27]. In general, VNS is a metaheuristic
originally proposed by [21] that applies a systematic change of neighborhood
structures within a local search scheme. In this paper, we take up the VNS of
[27] and apply it to the e-ADARP-IT, by integrating battery management with
the concept of a battery matrix derived from the matrix scheme proposed by [32].
This monitors chronologically the battery status of the vehicles and allows to
include visits at charging stations suitably into the vehicle routes. The algorithm
terminates after a predefined number of iterations without improvement of the
current solution or after a prescribed number of fixed runs, whichever occurs
first.

4 Computational Experiments
Section 4.1 describes the used test instances. Section 4.2 emphasizes a comparison
of the solution approaches being available as well as the effects of battery capacity
and charging station infrastructure on the solution quality.

4.1 Description of Test Instances

We use test instances from [27] for our computational experiments. These
instances have been generated for the ADARP-IT and are supplemented here
by data for an electric mobility system to use them for the e-ADARP-IT. The
instances cover a rural area in the federal state of Schleswig-Holstein, Germany.
In more detail, this rural area is located near the cities of Husum, Rendsburg,
and Heide with the characteristic river Eider. This area is of a size of 45 × 25 km
(approximately 1,125 km2 ) and has an average population of 85 inhabitants per
km2 [37]. The map material of the test instances has been created with the help
of OpenStreetMap [18]. Figure 1 presents a map of the studied region, where the
areas in light gray indicate the low population density and its spread over the
whole region.
The pick-up and drop-off locations of the test instances are derived from the
distribution of households. Customer requests are, thus, based on population
density. Travel times ti,j are calculated by the Open Source Routing Machine [38]
for the fastest route between the involved locations. The test instances contain
conventional and complex customer requests. For each conventional customer
request, a conventional trip with a randomly selected drop-off time within the
planning period is created. The associated time window [edi , ldi ] is 15 min. For
trips with parallel relations, it is assumed that each customer request consists
of two trips from two customers. The customers will be picked up at their indi-
vidual residences and dropped off at the same drop-off location. The customers’
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 427

Fig. 1. Area under study.

maximum waiting time wimax for each other is randomly selected from the time
interval U [1, 10] minutes. For trips with serial relations, it is assumed that each
customer request consists of two trips, one outbound and one return trip, where
the drop-off location of the outbound trip is identical to the pick-up location of
the return trip. The trips are associated with a stay time at the destination. The
least stay time δimin is drawn from the distribution U [20, 60]. The maximum
stay time δimax is by 10 min greater than δimin . The maximum travel time ui
for each trip i is 150% of the direct travel time. The examined planning period
of each test instance comprises 12 h. It is assumed that customer requests are
evenly distributed over the planning horizon.
We assume a homogeneous vehicle fleet. Each vehicle has a capacity of PCk =
6 passengers, which is in line with ridesharing provider MOIA [42]. The vehicles’
origin locations are randomly selected from the set of relevant locations within
the considered area. The number of customers boarding or alighting at a location
is qi = 1/-1 for all requests. Customers can be picked up and dropped off within
negligible time (si = 0). Each vehicle has a battery capacity of bk = 40 kWh and
is fully charged at the beginning of the planning horizon. The vehicles constantly
consume 0.206 kWh per kilometer driven. All vehicles generate a travel cost of
0.1 monetary units per kilometer. The objective function primarily maximizes
the number of served trips and secondarily minimizes the total cost. To ensure
this, we set an objective weight of r = 50. Note that complex requests are only
accepted if all involved trips can be served.
All charging stations support fast charging, meaning that a vehicle can be
charged up to 80% within 40 min. This percentage value is also used for the target
charge level γ in Constraint (29). During the charging process, the vehicles are
supplied with a charging quantity of αcs = 0.8 kWh per minute at all stations.
The locations of the charging stations are determined using a grid heuristic
[34]. Figure 2 illustrates the distribution of charging stations in the studied area
for a total of 8 stations.
Using the above configurations, we employ eight sets of test instances for our
experiment, each involving four individual instances. Table 2 provides an overview
of these instance sets, illustrating that the number of trips per instance ranges
428 M. Soth et al.

Fig. 2. Distribution of charging stations in the studied area.

from 10 to 50, with 2 to 5 vehicles, and 4 to 16 charging stations. The requests are
a mix of conventional and complex requests. The abbreviations ’CLE’ and ’MOI’
that are added to the name of the last four instance sets refer to two mobility
service providers operating in Germany, from which particular battery capacity
values of vehicles have been taken as is described in the next subsection.
The VNS heuristic has been implemented in Python 3.7 and all computational
experiments were performed on a computer with a 3.2 GHz CPU and 32 GB mem-
ory. For the exact solution of the optimization model, the MIP solver CPLEX 12.9
was applied with a maximum runtime per test instance of ten hours (36,000 s).

4.2 Influence of Battery Capacity and Charging Infrastructure

on Solution Quality
The experiments first contrast the performance of VNS and CPLEX and after-
wards address the impact of the number of charging stations as well as the

Table 2. Composition of the test instances.

Instance set Trips Vehicles Charging stations

MIX10V2CS4 10 2 4
MIX25V5CS8 25 5 8
MIX40V5CS12 40 5 12
MIX50V5CS12 50 5 12
MIX50V5CS12CLE 50 5 12
MIX50V5CS16CLE 50 5 16
MIX50V5CS4MOI 50 5 4
MIX50V5CS8MOI 50 5 8
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 429

capacity of vehicle batteries. We first evaluate the performance of VNS and

CPLEX for the first four instance sets with 10 to 50 trips, see Table 3. The
results reported in a row are the averages over the four instances belonging to
each set. The first column of the table identifies the instance set. Columns 2 to
6 report the results of the CPLEX solver and columns 7 to 11 the results of the
VNS heuristic. Here, the abbreviation Acc denotes the percentage of accepted
trips, i.e., those trips that are actually served in a solution, which refers to the
primary goal of the objective function. The secondary objective, total transporta-
tion cost, is shown in column Co. U B represents the theoretic upper bound of the
overall objective function value as determined by CPLEX. Performance measure
GAP = (obj − U B)/obj is the relative deviation of the actual objective function
value obj of a solution to the upper bound. Value obj is not shown explicitly
in the table but can be computed from the number of accepted requests and
the objective weight r = 50 minus the operational cost. The calculation time
is shown in column CP U in units of thousand seconds (k sec.). Columns 7, 8,
and 11 (Acc, Co, CP U ) of the VNS are analogous in meaning to the CPLEX
columns. Furthermore, ΔAcc maps the difference in accepted trips Acc between
the VNS solution and the CPLEX solution. In order to take into account the
different acceptance rates ΔCo represents the relative change in cost per served
trip between CPLEX and VNS. Note that the CP U time of the VNS heuristic
is composed of the total time of three runs with different initial solutions, which
is why it may exceed the computation time limit set for CPLEX in rare cases.

Table 3. Results of the computational experiments for varied instance size.

Instance set Solver VNS

Acc Co UB GAP CP U Acc Co ΔAcc ΔCo CP U
[%] [%] [k sec.] [%] [%] [%] [k sec.]
MIX10V2CS4 90.0 29.0 461 11.4 18 90.0 29.3 0.0 +1.7 0.14
MIX25V5CS8 96.0 71.8 1,221 8.4 36 99.0 70.7 +3.0 –4.8 1.1
MIX40V5CS12 53.0 88.6 1,963 105.8 36 72.5 89.5 +19.5 –22.2 6.5
MIX50V5CS12 37.0 80.8 2,462 203.3 36 84.0 132.0 +47.0 –27.3 22.5

The GAP s in Table 3 show that the CPLEX solver is hardly able to find opti-
mal solutions. Actually, it only solves two out of the instances with 10 requests
and one with 25 requests to optimality. Furthermore, even for the integer feasible
solutions obtained for the other instances, the acceptance rate drops significantly
the larger these instances are. The VNS algorithm produces comparable results
for the very small instances but within negligible computation time. For the
instances with 25 and more requests, it clearly outperforms CPLEX with much
larger acceptance rates that increase by up to 47% points. At the same time, the
operational costs per served trip are much lower (see column ΔCo), indicating
that VNS not just manages to serve many more requests than CPLEX but also
in a way more efficient fashion and within lower CP U times. Based on the VNS
430 M. Soth et al.

results, SAEVs can serve a large portion of the diverse mobility demands in rural
areas, with an average acceptance rate of 86% over all considered test instances.
We next address the question of whether SAEV providers should prioritize
a high-density charging infrastructure or vehicles with large battery capacities.
Two strategies are developed for investigating this. The first strategy uses vehi-
cles with a battery capacity of 40 kWh and a larger number of charging stations
(12 and 16). The battery capacity is derived from the vehicles used by the ride-
pooling provider CleverShuttle, which is why the test instances are marked with
the abbreviation ‘CLE’ (MIX50V5CS12CLE and MIX50V5CS16CLE) [6]. For
the setting with 12 charging stations, these instances are identical to the ones
reported in the last row of Table 3. In contrast, the second strategy uses vehi-
cles with a larger battery capacity of 86 kWh and fewer charging stations (4
and 8). The battery capacity is similar to those of vehicles used by ridesharing
provider MOIA [42]. The corresponding test instances are, thus, identified by
abbreviation ‘MOI’ (MIX50V5CS4MOI and MIX50V5CS8MOI).
Table 4 reveals the effects of battery capacity and charging infrastructure on
solution quality. The CPLEX solver cannot reach optimality within the given
runtime for any solution of these test instances. Its results all have low to mod-
erate acceptance rates and very high GAP s. Best results with an acceptance
rate of 59.0% and the only GAP below 100% are obtained for test instances
MIX50V5CS4MOI, clearly indicating that the larger the number of charging sta-
tions gets, the more CPLEX struggles. If we consider the test instances in ascend-
ing order of the number of charging stations, the negative impact of this number
on the performance of CPLEX becomes very clear. The average acceptance rate
of 59.0% for MIX50V5CS4MOI strictly decreases to 47.5% (MIX50V5CS8MOI)
to 37.0% (MIX50V5CS12CLE) and to 20.5% (MIX50V5CS16CLE). At the same
time, the GAP s increase strictly from 85.9% up to 604.5%.
In contrast, the acceptance rates of the VNS solutions are above 80% for
all test instances, even reaching 98% for instance set MIX50V5CS4MOI. At the
same time, the cost per served trip is up to 42.8% below the cost of the CPLEX
solutions, confirming the superior performance of VNS under all experimental
settings investigated here. However, VNS also exhibits high computation times,
especially for the CLE-instances with many charging stations. Based on the
results of the VNS, it can be seen that the test instances of the vehicles with a

Table 4. Results of the computational experiments for varied charging infrastructure

and battery capacities.

Instance set Solver VNS

Acc Co UB GAP CP U Acc Co ΔAcc ΔCo CP U
[%] [%] [k sec.] [%] [%] [%] [k sec.]
MIX50V5CS12CLE 37.0 80.8 2,462 203.3 36 84.0 132.0 +47.0 –27.3 22.5
MIX50V5CS16CLE 20.5 49.9 2,466 604.5 36 83.0 118.7 +62.5 –42.8 60.2
MIX50V5CS4MOI 59.0 110.4 2,461 85.9 36 98.0 155.3 +39.0 –14.9 4.4
MIX50V5CS8MOI 47.5 83.8 2,461 126.5 36 97.5 145.9 +50.0 –11.9 3.3
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 431

higher battery capacity (MOI) provide results with better solution quality com-
pared to the CLE-instances. The average acceptance rate of the MOI-instances
is about 15% higher than for the CLE-instances. For the CLE-instances with the
lower battery capacities, more charging processes (about 40) take place in the
solutions (not shown in the table). For the MOI-instances MIX50V5CS4MOI
and MIX50V5CS8MOI, just eight and two charging processes are required,
respectively. Furthermore, for both strategies the number of charging processes
decreases as the number of charging stations increases. From this, the average
cost per trip served decreases for both strategies as the number of charging
stations increases, which reveals the relevance of dense charging infrastructure.
In summary, it can be stated that the number of charging stations has no
substantial impact on the acceptance rate but on the cost per served trip. The
battery capacity, on the other hand, has a clear positive influence on the accep-
tance rate and, thus, a significant impact on the quality of the solutions. This
finding is consistent with [30], who identified battery range as a critical compo-
nent of SAEV service systems. Hence, based on the computational experiments,
a strategy focusing on large battery capacity seems more advantageous with
regards to the service rate of mobility requests, whereas additional charging
stations can help reducing the operational cost of the system.

5 Conclusions

This paper has presented the electric autonomous dial-a-ride problem with inter-
related trips (e-ADARP-IT). The associated mixed integer linear optimization
model includes, in addition to the properties of the classical dial-a-ride prob-
lem, synchronization constraints on the interrelated trips and constraints on the
battery charging management of electric, autonomous vehicles. A variable neigh-
borhood search (VNS) algorithm has been applied for the efficient solution of
larger instances of the e-ADARP-IT. For the computational experiments, test
data from a rural area in the German federal state of Schleswig-Holstein was
used. The results show that small test instances with ten requests can to some
extent be solved optimally using CPLEX. With increasing problem size, the
solution quality obtained by CPLEX decreases drastically, such that the share
of served trips drops to 37% for instances with 50 requests. The limitations of
the performance of CPLEX increase with a larger number of charging stations.
In contrast, the VNS heuristic performs consistently good over all instance sets.
It finds solutions that efficiently serve up to 99% of the trip requests at low
operational cost. The experiments furthermore show that the battery capacity
of vehicles is a crucial issue in SAEV service systems. Larger batteries allow to
serve a much higher share of trip requests. Clearly, sufficient charging infrastruc-
ture needs to be available, too. Nevertheless, increasing the number of charging
stations beyond that point merely reduced the operational service cost in our
experiments but did not result in larger acceptance rates for the trip requests.
Altogether, this paper contributes to identifying the potential of electric
autonomous vehicles for closing mobility gaps in rural areas. It also opens up
432 M. Soth et al.

opportunities for further research. The VNS has proven to be suitable for solving
the e-ADARP-IT but it might be too slow for usage in large service systems.
There are also further aspects at the operational level that require more research.
As an example, if charging stations are occupied by other vehicles, it needs to
be decided whether the SAEVs accept a waiting time or are rerouted to another
charging station. A further area of future research could be to combine the ride
sharing approach with public transport services. Here, an integration of school
transport would be possible as well as an offer as feeder for rail commuter lines.

Acknowledgements. The authors gratefully acknowledge the financial support of

this project through the German Research Foundation (DFG) under the reference
number 418360126.

References
1. Arslan, O., Yildiz, B., Karaşan, O.E.: Minimum cost path problem for plug-in
hybrid electric vehicles. Transp. Res. Part E Logist. Transp. Rev. 80(1), 123–141
(2015)
2. Baldacci, R., Battarra, M., Vigo, D.: Valid inequalities for the fleet size and mix
vehicle routing problem with fixed costs. Netw. Int. J. 54(4), 178–189 (2009)
3. Barco, J., Guerra, A., Muñoz, L., Quijano, N.: Optimal routing and scheduling of
charge for electric vehicles: a case study. Math. Probl. Eng. 2017, 1–6 (2017)
4. Bauer, G.S., Greenblatt, J.B., Gerke, B.F.: Cost, energy, and environmental impact
of automated electric taxi fleets in Manhattan. Environ. Sci. Technol. 52(8), 4920–
4928 (2018)
5. Bongiovanni, C., Kaspi, M., Geroliminis, N.: The electric autonomous dial-a-ride
problem. Transp. Res. Part B Methodol. 122, 436–456 (2019)
6. Clevershuttle: CleverShuttle fährt in Dresden On-Demand mit MOBIshuttle.
https://www.clevershuttle.de/presse/clevershuttle-dresden-mobishuttle. Accessed
04 Apr 2023
7. Cordeau, J.F.: A branch-and-cut algorithm for the dial-a-ride problem. Oper. Res.
54(3), 573–586 (2006)
8. Cordeau, J.F., Laporte, G.: The dial-a-ride problem (DARP): variants, modeling
issues and algorithms. 4OR - Q. J. Belg. Fr. Ital. Oper. Res. Soc. 1(2), 89–101
(2003)
9. Cortenbach, E.: The dial-a-ride problem with meeting points: a problem formula-
tion for shared demand responsive transit. Master’s thesis, University of Twente
(2023). http://essay.utwente.nl/94545/. Accessed 25 June 2023
10. Czioska, P., Kutadinata, R., Trifunović, A., Winter, S., Sester, M., Friedrich, B.:
Real-world meeting points for shared demand-responsive transportation systems.
Public Transp. 11, 341–377 (2019)
11. Desaulniers, G., Errico, F., Irnich, S., Schneider, M.: Exact algorithms for electric
vehicle-routing problems with time windows. Oper. Res. 64(6), 1388–1405 (2016)
12. Doppstadt, C., Koberstein, A., Vigo, D.: The hybrid electric vehicle - traveling
salesman problem. Eur. J. Oper. Res. 253(3), 825–842 (2016)
13. Earl, T.: Roadmap to decarbonising european cars (2018). https://www.
transportenvironment.org/sites/te/files/publications/2050_strategy_cars_
FINAL.pdf. Accessed 27 Mar 2023
 Ridesharing with Autonomous Electric Vehicles and Interrelated Trips 433

14. Emadi, A.: Transportation 2.0. IEEE Power Energy Mag. 9(4), 18–29 (2011)
15. Fagnant, D.J., Kockelman, K.M.: Dynamic ride-sharing and fleet sizing for a system
of shared autonomous vehicles in Austin, Texas. Transportation 45(1), 143–158
(2018)
16. Ángel Felipe, Ortuño, M.T., Righini, G., Tirado, G.: A heuristic approach for the
green vehicle routing problem with multiple technologies and partial recharges.
Transp. Res. Part E Logist. Transp. Rev. 71, 111–128 (2014)
17. Genikomsakis, K., Mitrentsis, G.: A computationally efficient simulation model for
estimating energy consumption of electric vehicles in the context of route planning
applications. Transp. Res. Part D Transp. Environ. 50, 98–118 (2017)
18. Geofabrik GmbH: Openstreetmap data extracts. https://download.geofabrik.de/.
Accessed 04 Apr 2023
19. Goeke, D., Schneider, M.: Routing a mixed fleet of electric and conventional vehi-
cles. Eur. J. Oper. Res. 245(1), 81–99 (2015)
20. Greenblatt, J.B., Shaleen, S.: Automated vehicles, on-demand mobility, and envi-
ronmental impacts. Current Sustain. Renew. Energy Rep. 2(3), 74–81 (2015)
21. Hansen, P., Mladenović, N.: Variable neighborhood search. In: Glover, F., Kochen-
berger, G.A. (eds.) Handbook of Metaheuristics, pp. 145–184. Springer, US,
Boston, MA (2003). https://doi.org/10.1007/0-306-48056-5_6
22. Hiermann, G., Puchinger, J., Ropke, S., Hartl, R.F.: The electric fleet size and mix
vehicle routing problem with time windows and recharging stations. Eur. J. Oper.
Res. 252(3), 995–1018 (2016)
23. Häll, C.H., Lundgren, J.T., Voß, S.: Evaluating the performance of a dial-a-ride
service using simulation. Public Transp. 7, 139–157 (2015)
24. Hyland, M.F., Mahmassani, H.S.: Taxonomy of shared autonomous vehicle fleet
management problems to inform future transportation mobility. Transp. Res. Rec.
J. Transp. Res. Board 2653(1), 26–34 (2017)
25. Hyland, M.F., Mahmassani, H.S.: Dynamic autonomous vehicle fleet operations:
optimization-based strategies to assign AVs to immediate traveler demand requests.
Transp. Res. Part C Emerg. Technol. 92, 278–297 (2018)
26. Jäger, B., Agua, F.M.M., Lienkamp, M.: Agent-based simulation of a shared,
autonomous and electric on-demand mobility solution. In: 2017 IEEE 20th Inter-
national Conference on Intelligent Transportation Systems (ITSC), pp. 250–255
(2017)
27. Johnsen, L.C., Meisel, F.: Interrelated trips in the rural dial-a-ride problem with
autonomous vehicles. Eur. J. Oper. Res. 303(1), 201–219 (2022)
28. Keskin, M., Çatay, B.: Partial recharge strategies for the electric vehicle routing
problem with time windows. Transp. Res. Part C Emerg. Technol. 65, 111–127
(2016)
29. Liang, X., de Almeida Correia, G.H., An, K., van Arem, B.: Automated taxis’
dial-a-ride problem with ride-sharing considering congestion-based dynamic travel
times. Transp. Res. Part C Emerg. Technol. 112, 260–281 (2020)
30. Loeb, B., Kockelman, K.M., Liu, L.: Shared autonomous electric vehicle (SAEV)
operations across the Austin, Texas network with charging infrastructure decisions.
Transp. Res. Part C Emerg. Technol. 89(4), 222–233 (2018)
31. Ma, T.Y.: Two-stage battery recharge scheduling and vehicle-charger assignment
policy for dynamic electric dial-a-ride services. PLoS ONE 16(5), 1–27 (2021)
32. Mankowska, D.S., Meisel, F., Bierwirth, C.: The home health care routing and
scheduling problem with interdependent services. Health Care Manag. Sci. 17(1),
15–30 (2014)
434 M. Soth et al.

33. Masmoudi, M.A., Hosny, M., Demir, E., Genikomsakis, K.N., Cheikhrouhou, N.:
The dial-a-ride problem with electric vehicles and battery swapping stations.
Transp. Res. Part E Logist. Transp. Rev. 118, 392–420 (2018)
34. Mattfeld, D.C., Vahrenkamp, R.: Logistiknetzwerke: Modelle für Standortwahl und
Tourenplanung, 2nd edn. Springer Gabler, Wiesbaden (2014). https://doi.org/10.
1007/978-3-8349-6912-5
35. Montoya, A., Guéret, C., Mendoza, J.E., Villegas, J.G.: The electric vehicle routing
problem with nonlinear charging functions. Transp. Res. Part B Methodol. 103,
87–110 (2017)
36. Narayanan, S., Chaniotakis, E., Antoniou, C.: Shared autonomous vehicle services:
a comprehensive review. Transp. Res. Part C Emerg. Technol. 111, 255–293 (2020)
37. Nord, S.: Kreise und Städte in Schleswig-Holstein im Vergleich: Bevölkerung 2018
(2019). https://www.statistik-nord.de/fileadmin/Dokumente/NORD.regional/
Schleswig-Holstein.regional/Band_1_-_Bevoelkerung/SH_regional_Band_1_
2018.pdf. Accessed 04 Apr 2023
38. Open Source Routing Machine: OSRM (2020). https://project-osrm.org/.
Accessed 04 Apr 2023
39. Pelletier, S., Jabali, O., Laporte, G.: Charge sheduling for electric freight vehicles.
Transp. Res. Part B Methodol. 115, 246–269 (2018)
40. Pelletier, S., Jabali, O., Laporte, G., Veneroni, M.: Battery degradation and
behavour for electric vehicles: review and numerical analyses of several models.
Transp. Res. Part B Methodol. 103, 158–187 (2017)
41. Schneider, M., Stenger, A., Goeke, D.: The electric vehicle-routing problem with
time windows and recharging stations. Transp. Sci. 48(4), 500–520 (2014)
42. Schwarzer, C.M.: Volkswagen-Ridesharing: Moin, MOIA! MOIA startet im
April den größten elektrischen Ridesharing-Dienst Europas (2019). https://www.
electrive.net/2019/03/14/volkswagen-ridesharing-moin-moia/. Accessed 04 Apr
2023
43. Shaheen, S., Chan, N.: Mobility and the sharing economy: potential to facilitate
the first- and last-mile public transit connections. Built Environ. 42(4), 573–588
(2016)
44. Sprei, F.: Disrupting mobility. Energy Res. Soc. Sci. 37, 238–242 (2018)
Combinatorial Optimization
 Operational Integration of Supply Chain
Activities with Earliness and Tardiness
Considerations

Ece Yağmur(B) and Saadettin Erhan Kesen

Konya Technical University, Konya, Turkey

{ecyagmur,sekesen}@ktun.edu.tr

Abstract. The operational integration of the production and distribution activities

in a supply chain requires concurrent consideration of two major problems known
as production scheduling and vehicle routing. While the production environment
of the integrated problem under study consists of identical parallel machines,
the distribution environment has a limited number of homogeneous vehicles that
deliver completed orders to customers. The objective function is to minimize the
sum of total earliness and tardiness according to the time windows specified by
each customer. In this paper, we first describe the integrated problem as a Mixed
Integer Programming (MIP) model and subsequently present an Iterated Local
Search (ILS) algorithm to find optimal or near-optimal solutions in a reasonable
time. The performance of the ILS is evaluated against the MIP-based solutions
obtained by applying a standard solver like CPLEX. According to the computa-
tional findings, the suggested ILS can find optimal or near optimal solutions for
randomly generated test instances in just a few seconds.

Keywords: parallel machine scheduling · vehicle routing · iterated local search ·

earliness · tardiness

1 Introduction
Today, the management of supply chain operations is a trending topic for many businesses
where competition is of utmost importance. Thus, it is inevitable to provide effective
coordination between the production and distribution phases of the supply chain. Fulfill-
ing customer demands with the required quantity on time becomes an extremely difficult
task, especially for companies which apply the zero-inventory model and make-to-order
businesses. The integrated approach is also beneficial for time management of time-
sensitive or perishable products. Considering all these aspects, operational integration
for the two main stages is carried out by making decisions simultaneously for production
scheduling and vehicle routing. According to the integrated approach, as the output of
either does not affect the other, the decisions on production and distribution stages will
not limit each other.
Recently, there has been a remarkable growth in the online shopping sector [2]. Due
to the abundance of competitors in the market and the ease of communication between

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 437–449, 2023.
https://doi.org/10.1007/978-3-031-43612-3_27
438 E. Yağmur and S. E. Kesen

customers and companies, firms are under increasing pressure to ensure customer sat-
isfaction. For this reason, they have turned to customer-oriented performance measures
rather than classical measures such as cost and profit. According to the current appli-
cations, customers know when they can receive the orders and even specify their own
preferences regarding delivery time, which is often handled by time windows in the
vehicle routing literature.
A current survey, which covers the enlarged literature on integrated production
scheduling and outbound vehicle routing problems, can be found in [4]. In the survey,
the literature is classified in terms of the number of operations, the number of vehicles,
and the number of trips. According to this classification, the problem under study can
be represented in a section of parallel machines, several vehicles, and single trips, as in
the studies of [3, 5, 6, 11, 12, 18].
Although there are several criteria (i.e., objectives) identified by researchers, in gen-
eral, objectives can be divided into two main groups: cost and service. In addition,
sometimes multiple criteria can be handled simultaneously as a summation of objectives
in a single form [5, 9, 21, 22] or multi-objective structure [7, 14, 23]. While many criteria
that are related to the problem under study have been defined in the literature [4], the
objective of this study can be expressed as minimizing the sum of total earliness and
tardiness. Ullrich [18] considers a variant of the integrated problem with the objective of
minimizing the total tardiness and proposes a genetic algorithm to produce optimal or
near optimal solutions in a reasonable amount of time. Wu et al. [20] formulate a math-
ematical model for a production scheduling-based routing problem with time windows
and setup times for minimizing total tardiness. Hou et al. [10] address an integrated
problem where the production stage has multiple factories, and each factory has multi-
ple machines based on the flow shop environment. For the objective of minimizing total
weighted earliness and tardiness, they propose a swarm intelligence approach which is
called brainstorm optimization algorithm.
In the integrated problem under investigation, the distribution phase has a homoge-
neous fleet that can only be utilized once, while the production environment consists
of identical parallel machines and each job has a single operation that needs to be per-
formed on any machines. It is very important for customer satisfaction to deliver orders
according to the predefined time windows. In addition, vehicles that arrive earlier must
wait idle until the lower time bound of any customer. So, the objective is to minimize
the sum of total earliness and tardiness. After formulating the problem, we propose a
metaheuristic approach, Iterated Local Search (ILS), to obtain good quality solutions in
a reasonable time.
The contribution of the paper can be summarized as follows: (i) The sum of the
total earliness and tardiness is studied simultaneously for the first time for the integrated
problem at hand. (ii) A new mixed integer programming formulation is developed. (iii)
A metaheuristic algorithm, Iterated Local Search (ILS), is presented, and a perturbation
mechanism is proposed for the only production environment adapted for the integrated
problem.
The remainder of the study is organized as follows: In Sect. 2, we define the MIP
formulation for the integrated problem with an illustrative example. While in Sect. 3,
the proposed ILS structure is given, in Sect. 4, computational experiments are discussed
 Operational Integration of Supply Chain 439

based on the different problem parameters. Finally, in Sect. 5, conclusions and future
directions are discussed.

2 Problem Definition

This section is dedicated to two subsections including the MIP model and an illustrative
example for visual presentation.

2.1 MIP Model

The assumptions of the problem can be summarized as follows: There is a single pro-
duction facility in which a set of identical parallel machines is used to produce customer
orders. After the production of a particular batch is completed, a limited number of
vehicles is delivering the consolidated orders to associated customers by considering the
time window of each customer. Vehicles can be used only once (i.e., multiple use is not
allowed). Vehicle capacities are homogeneous, and the capacity of a vehicle is enough to
serve all customers. As a result, there is no requirement to use all vehicles in the system.
The sequence of production and distribution does not have to be identical. The objective
is to minimize the sum of total earliness and tardiness penalties. The notations used in
the MIP formulation are shown in Table 1.
Based on the formal definition and notations, we formulate an MIP model for the
problem as follows:
 
minimize Ej + Tj (1)
j∈NC

Subject to;

Wjm ≤ 1 ∀m ∈ M (2)
j∈NC

Zji + Lj = 1 ∀j ∈ NC (3)
i∈NC
 
Wjm + Zij = 1 ∀j ∈ NC (4)
m∈M i∈NC
 
Cj ≥ pj − H 1 − Wjm ∀j ∈ NC ; m ∈ M (5)

 
Cj ≤ pj + H 1 − Wjm ∀j ∈ NC ; m ∈ M (6)

 
Ci − Cj + HZij + H − pi − pj Zji ≤ H − pj ∀i, j ∈ NC ; i = j (7)

X0jk ≤ 1 ∀k ∈ K (8)
j∈NC
440 E. Yağmur and S. E. Kesen

Table 1. Notations used in the MIP model.


Xijk = 1 ∀j ∈ NC (9)
k∈K i∈N
 Operational Integration of Supply Chain 441

 
Xujk = Xjik ∀j ∈ N ; ∀k ∈ K (10)
u∈N i∈N
 
Fuj − Fji = dj ∀j ∈ NC (11)
u∈N i∈N

dj Xijk ≤ Fij ∀i, j ∈ N ; i = j (12)
k∈K

Q − di + Xijk ≥ Fij ∀i, j ∈ N ; i = j (13)
k∈K
⎛ ⎞

Rjk = ⎝ Xjik ⎠ + Xj0k ∀j ∈ NC ; ∀k ∈ K (14)
i∈NC
 
Ci − Cj ≤ H 1 − Gij ∀i, j ∈ NC ; i = j (15)

Cj − Ci ≤ HGij ∀i, j ∈ NC ; i = j (16)

 
Vk ≥ Ci − H 1 − Xijk ∀i, j ∈ N ; i = j; ∀k ∈ K (17)

Vk ≤ Cj + H (1 − (Rjk − AX 3jik ∀j ∈ NC ; ∀k ∈ K
 
i∈N C
(18)
+ Gij + Gji − 1 ))
i∈NC
 
Vk − Yj + s0 + t0j ≤ H 1 − X0jk ∀j ∈ NC ; ∀k ∈ K (19)

 
Yj − Vk − s0 − t0j ≤ H 1 − X0jk ∀j ∈ NC ; ∀k ∈ K (20)


Yi∗ − Yj + si + tij ≤ H 1 − Xijk ∀i, j ∈ NC ; i = j (21)
k∈K


Yj − Yi∗ − si − tij ≤ H 1 − Xijk ∀i, j ∈ NC ; i = j (22)
k∈K

Tj ≥ Yj∗ − bj ∀j ∈ NC (23)

Ej ≥ aj − Yj ∀j ∈ NC (24)

 
Yj − aj ≤ H 1 − Oj ∀j ∈ NC (25)

aj − Yj ≤ HOj ∀j ∈ NC (26)
442 E. Yağmur and S. E. Kesen

Yj∗ = Yj − AX 4j + aj Oj ∀j ∈ NC (27)

The objective function minimizing the sum of total earliness and tardiness is given in
(1). It is determined whether a job is the successor of any job on the same machine or the
first/last job to be processed on any machine by Cons. (2)–(4). While Cons. (5) and (6)
are used to determine the production completion time of the first jobs on each machine,
Cons. (7) defines the production completion time for predecessor and successor jobs
on each machine. Cons. (8) identifies the first customer on any tour. Cons. (9) and (10)
guarantee that each customer must be visited exactly once, and the number of entering
and leaving arcs must be the same for any node, respectively. Cons. (11)–(13) are based on
a single-commodity flow formulation described by [8]. This set of constraints guarantee
a connected tour. So, we define Fij as a continuous variable indicating the amount of load
flowing between nodes i and j. In order to assure subtour elimination in the formulation,
we impose tight bounds on the flow variables Fij in addition to a set of flow conservation
constraints. Cons. (11) and (12) indicate that the total amount of load on any vehicle
decreases as it delivers the order to the related customer. According to Cons (13), the total
amount of load on any vehicle should not exceed its capacity. While Cons. (14) defines
the variable of Rjk , , Cons. (15) and (16) define the variable of Gij . . According to Cons.
(17) and (18), each vehicle must wait in production facility until the production process
of the batch assigned completes. Cons. (19) and (20) specify the arrival time to the first
customer on any tour. Cons. (21) and (22) state the arrival time to the two consecutive
customers on any tour. Cons. (23) and (24) calculate tardiness and earliness, respectively,
if either of them occurs. Cons. (25) and (26) define the variable of Oj . . According to
Cons. (27), the service starting time of any customer is equal to the maximum value of
the lower bound of the time windows and arrival time for that customer.
As can be seen, the developed model is not linear because it consists of the product
of two variables. So, while we use the Cons. (28)–(30) for AX 1ijk , AX 2ij and AX 3ijk
where both x and y are binary variables, Cons. (31)–(33) are used for the linearization
of AX 4j where x is a binary variable and y is a non-negative continuous variable, and M
is a sufficiently large number.

z≤x (28)

z≤y (29)

z ≥x+y−1 (30)

z ≤ xM (31)

z≤y (32)

z ≥ y − (1 − x)M (33)
 Operational Integration of Supply Chain 443

2.2 Illustrative Example

Based on the formal definition given in the previous section, we present an illustrative
problem for visualizing the integrated approach. In the example, whose parameters are
listed in Table 2, there are nine customers, four identical parallel machines, and two
homogenous vehicles with enough capacity to serve all customers. Table 1 contains the
descriptions of parameters listed in the first row of Table 2.

Table 2. Parameters of an illustrative example

j pj sj aj bj tij 0 1 2 3 4 5 6 7 8 9
0 – 19 – – 0 0 3 13 5 18 10 12 22 18 10
1 39 1 69 75 1 3 0 13 7 19 12 10 23 16 9
2 20 9 55 61 2 13 13 0 14 30 19 22 11 19 5
3 91 5 128 138 3 5 7 14 0 18 7 15 21 23 12
4 52 2 199 201 4 18 19 30 18 0 14 13 38 31 27
5 85 14 114 135 5 10 12 19 7 14 0 16 26 27 18
6 13 9 108 126 6 12 10 22 15 13 16 0 33 18 18
7 18 9 61 106 7 22 23 11 21 38 26 33 0 29 16
8 2 15 54 81 8 18 16 19 23 31 27 18 29 0 15
9 71 6 132 139 9 10 9 5 12 27 18 18 16 15 0

A feasible solution for the given example is illustrated in Fig. 1. According to Fig. 1,
jobs 7 and 5 are processed on the first machine. Jobs assigned to other machines are
depicted similarly. While customers 2, 7, 8, and 6 are served by the first vehicle, the
second vehicle delivers orders 1, 9, 3, 5, and 4, respectively. Shipping boxes show the
service time spent on each customer. Vehicle one (v1) departs from the depot when the
production process of the first batch completes at time 22, then goes directly from the
depot to customer 2 at time 54. Based on the time window of customer 2, the vehicle must
wait 1 min until the lower bound of the time window (i.e., 55). This process proceeds
in the same way for other customers. Based on the following feasible solution, earliness
occurs for customers 2 and 4, while tardiness occurs for customers 1, 3, 5, 6, and 8, as
seen from Fig. 1. So, the objective value of the problem is equal to the 156 which is a
summation of 6 earliness and 150 tardiness.
444 E. Yağmur and S. E. Kesen

m1 7 5

m2 1 4

m3 8 2 9

m4 6 3

v1 0 2 7 8 6

v2 0 1 9 3 5 4

0 20 40 60 80 100 120 140 160 180 200 220

Fig. 1. A Gantt chart of a feasible solution (also optimal solution) for the illustrated example.

3 Iterated Local Search

The idea of Iterated Local Search (ILS) as an extension of the local search procedure
was first introduced in [13]. Basically, ILS is formed by incorporating a perturbation
mechanism into the local search process to avoid getting stuck in local optima. ILS has
been used for many combinatorial optimization problems in the literature [15––17, 19].
The pseudo code of the proposed ILS is given in Fig. 2.
In the perturbation process, we use the regeneration operator proposed by [1]. In
the selection process, parents are selected as the current local optimal solution and the
neighbor solution with the best objective. The encoding scheme developed by Afralizad
and Rezaeian [1] is also adapted to the integrated problem. According to the adapted
chromosome, while the first and third rows represent the permutation of the jobs and
customers, respectively, the second and fourth rows demonstrate the assignments of
machines and vehicles for each job and customer in the corresponding position. The
encoding scheme for the illustrated problem is depicted in Fig. 3. In the local search
process, we use swap, insert, and 2-opt operators for both production and distribution
phases so as to consider inter and intra machines and vehicles.
 Operational Integration of Supply Chain 445

Fig. 2. The pseudocode of the proposed ILS.

Job order: 7 5 1 4 8 2 9 6 3
Machine assignment: 1 1 2 2 3 3 3 4 4
Customer order: 2 7 8 6 1 9 3 5 4
Vehicle assignment: 1 1 1 1 2 2 2 2 2

Fig. 3. The encoding scheme

4 Computational Results

For computational experiments, we generate a total of 486 instances randomly based

on the number of customers, machines, and vehicles. While the number of customers
includes six levels from 5 to 10, machines and vehicles are set to three levels covering
2, 3, and 4. In addition, for each factor level of the parameters, nine different random
instances are generated.
The CPLEX solver embedded in GAMS is used as the MIP-solver. The maximum
computational time for the CPLEX solver is set to 10800 s, and the stopping criterion
for ILS is run 1000 × N iterations, where N represents the number of customers.
Table 3 reports the comparative results of CPLEX and ILS based on the problem
parameters. The column Parameters describes the level of problem parameters based on
the number of customers, machines, and vehicles, respectively. In the CPLEX column,
while Obj. Indicates the average objective value (i.e., upper bound), LB represents the
lower bound found by CPLEX in a given time limit of 3 h. The number of optimal solu-
tions found by CPLEX and ILS are given in column #Opt. Finally, the solution times in
446 E. Yağmur and S. E. Kesen

seconds for CPLEX and ILS are indicated by the CPU column. The deviation of objec-
−ObjILS
tive values from each other (DEV%) is calculated by ObjCPLEX
ObjCPLEX ×100 if the solution
−ObjCPLEX
found by CPLEX is worse than ILS; otherwise, it is calculated by ObjILSObj ILS
× 100.
We summarize the following remarks from Table 3: CPLEX can find all optimal
solutions for all instances of six and seven customers. As the number of customers
increases, the ability of CPLEX to find optimal solutions declines dramatically. For ten
customers, CPLEX can only find seven optimal solutions out of 81 instances. Like the
first remark, the CPU time also increases drastically when the number of customers
increases from 5 to 10 and reaches to about 10,000 s for the level of 10 customers.
In addition, when the effect of the number of machines and vehicles are examined, we
reach the following results: As the number of machines increases, CPLEX’s ability to find
optimal solutions increases. This is a result of the reduction of the tardiness component in
the objective function due to the simultaneous use of resources in the production phase.
Similarly, as the number of vehicles increases, the average objective value decreases.
However, the number of vehicles has not as much an effect as the number of machines
in terms of the number of optimal solutions found and CPU times. As can be seen from
Table 3, CPLEX is able to find the optimal solution in 331 instances out of 486 instances
within 3 h. When the same instances are solved with ILS, it is seen to provide optimal
solutions for 329 instances out of 331 instances whose optimality are proven by CPLEX.
Especially in instances where the number of customers is set to 9 and 10, the solutions
obtained by ILS are of higher quality, and the percent deviation of CPLEX from ILS
(DEV%) reaches to 77%. As a result, ILS has shown superior performance in terms of
both solution quality and solution time.

Table 3. The comparison of CPLEX and ILS based on the number of customers, machines, and
vehicles.
 Operational Integration of Supply Chain 447

5 Final Remarks and Future Directions

In this study, we develop a new MIP model for a variant of the integrated production
and distribution problem that consists of parallel machine scheduling and homogeneous
vehicle routing problems with time windows. The objective of the model is formed by
minimizing the sum of total earliness and tardiness. The performance of the exact method
is found to be poor, even for only seven customers. We also examine the performance
of CPLEX based on the number of customers, machines, and vehicles. According to
the results, the increase in the number of customers reduces the performance of CPLEX
dramatically. Especially for nine and ten customers, the exact method can no longer
be considered a reasonable solution method for an operational-level problem. On the
contrary, the increase in the number of machines positively affects the performance of
CPLEX. The reason for this is that the departure times of the vehicles are moved earlier
due to the simultaneous use of identical parallel machines in the production environment.
Thus, the total tardiness in the objective function tends to decrease. When the number
of vehicles increases, we see that the objective function value decreases, but there is
no significant difference when CPU time and the number of optimal solutions found is
examined. For a customer number of 10, CPLEX can only find optimality in 7 out of 81
instances in a given time limit of 10800 s, and the deviation between the upper and lower
bounds found by CPLEX is calculated at 96.7%. Then, we propose the ILS algorithm
for solving the integrated problem efficiently. According to the comparative results, the
proposed ILS algorithm performs quite well in terms of both solution quality and time.
The CPU time of ILS for a level of ten customers is about 8 s, even for 10000 iterations.
In addition, ILS is found to produce 77% better solutions in terms of objective function
value when compared to CPLEX for ten customers.
Future work may focus on larger instances, and the performance of ILS can be
compared to other metaheuristics such as genetic algorithms, particle swarm optimiza-
tion, etc. By considering specific situations of problem assumptions, such as hetero-
geneous vehicles, unrelated parallel machines, and sequence dependent set-up times, a
new variant of an integrated problem can be handled. In addition, sustainable objectives
such as energy-efficient strategies at both the production and distribution levels can be
considered.

References
1. Afzalirad, M., Rezaeian, J.: Resource-constrained unrelated parallel machine scheduling
problem with sequence dependent setup times, precedence constraints and machine eligi-
bility restrictions. Comput. Ind. Eng. 98, 40–52 (2016). https://doi.org/10.1016/j.cie.2016.
05.020
2. Alzoubi, H., Alshurideh, M., Kurdi, B., Alhyasat, K., Ghazal, T.: The effect of e-payment
and online shopping on sales growth: evidence from banking industry. Int. J. Data Netw. Sci.
6(4), 1369–1380 (2022). https://doi.org/10.5267/j.ijdns.2022.5.014
3. Belo-Filho, M., Amorim, P., Almada-Lobo, B.: An adaptive large neighbourhood search for
the operational integrated production and distribution problem of perishable products. Int. J.
Prod. Res. 53(20), 6040–6058 (2015). https://doi.org/10.1080/00207543.2015.1010744
4. Berghman, L., Kergosien, Y., Billaut, J.-C.: A review on integrated scheduling and outbound
vehicle routing problems. Eur. J. Oper. Res. (2023). https://doi.org/10.1016/j.ejor.2022.12.036
448 E. Yağmur and S. E. Kesen

5. Chang, Y.-C., Li, V.C., Chiang, C.-J.: An ant colony optimization heuristic for an integrated
production and distribution scheduling problem. Eng. Optim. 46(4), 503–520 (2014). https://
doi.org/10.1080/0305215X.2013.786062
6. Farahani, P., Grunow, M., Günther, H.-O.: Integrated production and distribution planning for
perishable food products. Flex. Serv. Manuf. J. 24, 28–51 (2012). https://doi.org/10.1007/s10
696-011-9125-0
7. Ganji, M., Kazemipoor, H., Molana, S.M.H., Sajadi, S.M.: A green multi-objective integrated
scheduling of production and distribution with heterogeneous fleet vehicle routing and time
windows. J. Clean. Prod. 259, 120824 (2020). https://doi.org/10.1016/j.jclepro.2020.120824
8. Gavish, B., Graves, S.C.: The travelling salesman problem and related problems (1978). http://
hdl.handle.net/1721.1/5363
9. He, P., Li, K., Kumar, P.R.: An enhanced branch-and-price algorithm for the integrated pro-
duction and transportation scheduling problem. Int. J. Prod. Res. 60(6), 1874–1889 (2022).
https://doi.org/10.1080/00207543.2021.1876941
10. Hou, Y., Fu, Y., Gao, K., Zhang, H., Sadollah, A.: Modelling and optimization of integrated
distributed flow shop scheduling and distribution problems with time windows. Expert Syst.
Appl. 187, 115827 (2022). https://doi.org/10.1016/j.eswa.2021.115827
11. Kesen, S.E., Bektaş, T.: Integrated production scheduling and distribution planning with
time windows. Lean Green Supply Chain Manage: Optim. Models Algorithms, 231–252
(2019).https://doi.org/10.1007/978-3-319-97511-5_8
12. Liu, H., Guo, Z., Zhang, Z.: A hybrid multi-level optimisation framework for integrated
production scheduling and vehicle routing with flexible departure time. Int. J. Prod. Res.
59(21), 6615–6632 (2021). https://doi.org/10.1080/00207543.2020.1821927
13. Lourenço, H.R., Martin, O.C., Stützle, T., Iterated local search. In: Handbook of Metaheuris-
tics, pp. 320–353. Springer, Cham (2003). https://doi.org/10.1007/0-306-48056-5_11
14. Mohammadi, S., Al-e-Hashem, S.M., Rekik, Y.: An integrated production scheduling and
delivery route planning with multi-purpose machines: a case study from a furniture manufac-
turing company. Int. J. Prod. Econ. 219, 347–359 (2020). https://doi.org/10.1016/j.ijpe.2019.
05.017
15. Penna, P.H.V., Subramanian, A., Ochi, L.S.: An iterated local search heuristic for the hetero-
geneous fleet vehicle routing problem. J. Heuristics 19(2), 201–232 (2013). https://doi.org/
10.1007/s10732-011-9186-y
16. Stützle, T.: Applying iterated local search to the permutation flow shop problem. Technical
Report AIDA-98–04, FG Intellektik, TU Darmstadt (1998)
17. Stützle, T.: Iterated local search for the quadratic assignment problem. Eur. J. Oper. Res.
174(3), 1519–1539 (2006). https://doi.org/10.1016/j.ejor.2005.01.066
18. Ullrich, C.A.: Integrated machine scheduling and vehicle routing with time windows. Eur. J.
Oper. Res. 227(1), 152–165 (2013). https://doi.org/10.1016/j.ejor.2012.11.049
19. Vansteenwegen, P., Souffriau, W., Berghe, G.V., Van Oudheusden, D.: Iterated local search for
the team orienteering problem with time windows. Comput. Oper. Res. 36(12), 3281–3290
(2009). https://doi.org/10.1016/j.cor.2009.03.008
20. Wu, G.-H., Cheng, C.-Y., Pourhejazy, P., Fang, B.-L.: Variable neighborhood-based cuckoo
search for production routing with time window and setup times. Appl. Soft. Comput. 125,
109191 (2022). https://doi.org/10.1016/j.asoc.2022.109191
21. Yağmur, E., Kesen, S.E.: A memetic algorithm for joint production and distribution scheduling
with due dates. Comput. Ind. Eng. 142, 106342 (2020). https://doi.org/10.1016/j.cie.2020.
106342
 Operational Integration of Supply Chain 449

22. Yağmur, E., Kesen, S.E.: Multi-trip heterogeneous vehicle routing problem coordinated with
production scheduling: memetic algorithm and simulated annealing approaches. Comput. Ind.
Eng. 161, 107649 (2021). https://doi.org/10.1016/j.cie.2021.107649
23. Yağmur, E., Kesen, S.E.: Bi-objective coordinated production and transportation scheduling
problem with sustainability: formulation and solution approaches. Int. J. Prod. Res. 61(3),
774–795 (2023). https://doi.org/10.1080/00207543.2021.2017054
 Constrained Multi-agent Path Planning
Problem

Ali Maktabifard1(B) , Dávid Földes1 , and Bendegúz Dezső Bak2

1
Faculty of Transportation Engineering and Vehicle Engineering,
Department of Transport Technology and Economics,
Budapest University of Technology and Economics, Budapest, Hungary
[email protected], [email protected]
2
Faculty of Mechanical Engineering, Department of Fluid Mechanics,
Budapest University of Technology and Economics, Budapest, Hungary
[email protected]

Abstract. Planning the most efficient routes in a cooperative manner

is a challenge for many mobility and logistics service providers. In this
paper, a new methodological approach is presented based on the Multi-
Agent Path Planning (MAPP) problem which is a variant of the classi-
cal Multiple Traveling Salesmen Problem (MTSP). Given a team of m
agents that must visit n targets, the optimal paths plan (a set of m paths)
should be determined such that each target is visited only once. Mini-
mizing the time of this cooperative operation is the optimization goal
in this study. Thus, the plan is optimal, if the longest path in the plan
is the shortest possible (Min-Max problem). In order to deal with more
practical situations, two additional constraints are applied: the maxi-
mum number of targets each agent is allowed to visit, and the maximum
range for each agent. Accordingly, a so-called Constrained Multi-Agent
Path Planning (CMAPP) problem is elaborated in this paper. An easy
to apply Genetic Algorithm (GA) is presented, which improves the paths
using genetic-like operators and a heuristic method. The applicability of
the approximate solution was tested in four random scenarios where it
showed a decent performance. The developed method can be used for
route planning of mobility and logistics services in which several desti-
nations must be reached by a fleet of vehicles (e.g., group ride-sharing,
last-mile delivery).

Keywords: Multi-Agent Planning Problem · Genetic Algorithm ·

Routing Problem · Optimization · Metaheuristic · MAPF · MTSP

1 Introduction

The importance of cooperative path planning for vehicle fleets has been demon-
strated by various applications such as logistics and delivery services [19,26],
ride-sharing services [13,25], and mission planning for autonomous/unmanned
vehicles [4,28]. In general, a multi-agent planner is required for cooperative path
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 450–466, 2023.
https://doi.org/10.1007/978-3-031-43612-3_28
 Constrained Multi-agent Path Planning Problem 451

planning of vehicle fleets. This planning problem can be defined as finding a set
of paths that allows a fleet of agents to reach a specified number of targets in the
minimum amount of time. This problem is similar to the classical Multiple Trav-
eling Salesmen Problem (MTSP) [1], and its variants such as the Multi-Agent
Path Planning (MAPP) problem [16]. The classical MTSP is generally defined
as follows. Given n cities (targets) and m salesmen (agents), the aim is to find
m tours (closed paths) starting and ending at a depot (initial position of the
agents) such that each target is visited only once and the total cost of visiting
all targets is minimized. The cost metric can be expressed in terms of distance,
time, etc., [1]. The MAPP problem is similar to the classical MTSP with two
differences:

– Subtours (open paths) are considered in the MAPP problem, so each agent
starts and ends its path at two distinct points.
– Agents can have different initial positions.

More practical situations can be modeled by applying additional restric-

tions to these combinatorial optimization problems. These restrictions affect
the agents mostly by restricting their path length, workload, or operational
range [13,19,25,26]. As a novelty, in this study the cooperative path planning
for vehicle fleets is modeled as a so-called Constrained Multi-Agent Path Plan-
ning (CMAPP) problem. This problem is formed by applying two additional
constraints to the MAPP problem: the maximum number of targets each agent
is allowed to visit, and the maximum range for each agent. This paper presents
a two-step solution method composed of an initial solution and a complex solu-
tion for this problem. The initial solution benefits from a new targets assignment
algorithm. The complex solution uses a Genetic Algorithm (GA) to improve the
result of the initial solution. The additional constraints are applied in both ini-
tial and complex solutions. The aim of this study is to show the applicability of
the developed method.
The remainder of the paper is organized as follows. Section 2 reviews solution
methods for similar problems. The CMAPP problem is described in detail in
Sect. 3. The methodology of the proposed solution is described in Sect. 4; then
results are presented and discussed in Sect. 5. Finally, conclusions are drawn in
Sect. 6.

2 Review of Solution Methods

Solving the MTSP and the MAPP problem is difficult because of their complex
combinatorial character (NP-hardness). In general, two types of approaches are
used to tackle the MTSP [12]: exact and heuristic-based approaches.
The exact approaches are based on either the transformation of the MTSP to
an equivalent Traveling Salesman Problem (TSP), or relaxing some constraints
of the problem [1,12]. The problem is solved by applying exact methods such
as Branch-and-Bound [6], Cutting Planes [20], and Integer Linear Programming
452 A. Maktabifard et al.

Table 1. Literature instances on the MTSP having similarities to the CMAPP problem
considered in our study

Literature Solution method Multi-depot Min-Max Min. and Max.

MTSP MTSP No. of targets
for each agent
[17] Integer linear programming × ×
formulations + new bounding and
Subtour Elimination Constraints
(SECs)
[18] A heuristic approach based on an × ×
Evolution Strategy (ES)
[8] An ACO algorithm × ×
[30] Two variants of Parthenogenetic × Only Min.
Algorithm (PGA)
[15] An Ant Colony-Parthenogenetic × ×
Algorithm (AC-PGA)
[14] A heuristic approach based on a ×
graph simplification method and
the 2-OPT algorithm

Formulations [17]. The exact approaches are restricted to the MTSPs with rea-
sonable sizes (Euclidean and Non-Euclidean problems up to 100 and 500 cities
(targets), respectively [6]) because their performance is highly dependent on the
size of the problem. Accordingly, the solution runtime encounters an exponen-
tial rise with increasing the problem size [1,12]. Additionally, transforming the
MTSP to an equivalent TSP might result in an even more difficult problem to
solve, especially using the exact approaches [6,20].
The heuristic-based approaches solve the problem by applying approximate
heuristic methods such as Ant Colony Optimization (ACO) [8,12,15], Genetic
Algorithm (GA) [12,15,29,30], Simulated Annealing [24], Neural Networks [23],
and Tabu Search [22]. The heuristic-based approaches can achieve near-optimal
solutions in a reasonable amount of time even for larger problems [12].
The problems tackled in several previous papers on the MTSP have some
similarities to the CMAPP problem considered in our study (see Table 1). These
similarities include considering multiple depots for the MTSP, minimizing the
length (cost) of the longest tour (Min-Max MTSP), and applying additional
constraints.
There are only a few studies on the variants of the MAPP problem. In [9], a
GA was introduced to solve a so-called Subtour problem which is similar to the
MAPP problem with one difference: there is only one agent. This method was
developed for solving the MAPP problem in [10], and a similar method was pro-
posed in [16] to address the MAPP problem. In [27], a heuristic approach based
on a graph simplification method was presented to solve a multi-depot open
 Constrained Multi-agent Path Planning Problem 453

tours MTSP which is basically the MAPP problem with Min-Sum optimization
objective.
Concluding the literature review, the MTSP is a well-studied problem, while
there are only a few studies on the MAPP problem. Furthermore, previous stud-
ies have not solved the MAPP problem with our additional constraints.

3 Constrained Multi-agent Path Planning Problem:

General Terms and Notation

In this section, the required foundation from Graph theory, and the notations
of the classical MTSP and the MAPP problem are presented. In addition, our
additional constraints for the MAPP problem are defined.
According to Graph theory, a graph is an ordered pair G = (V, E) consisting
of: V = {v1 , ..., vm }, a set of m vertices (nodes), and E = {(vi , vj ) | vi , vj ∈
V, i = j}, a set of edges (links) connecting vertices vi and vj . This type of
graph may be precisely referred to as simple graph which means multiple edges
connecting the same two vertices are not allowed. In the context of the classical
MTSP and the MAPP problem, the targets and the points where the agents
begin their journey, are considered as vertices. In this study, only undirected
graphs are taken into consideration. In an undirected graph, edges are comprised
of unordered pairs of vertices where (vi , vj ) = (vj , vi ). Moreover, if in a graph
all vertices of V are connected to each other, the graph is a complete graph and
it is designated by Km (V ) where m is the number of vertices constituting the
vertex set V .
A path/cycle is a sequence of edges connecting a sequence of vertices. If the
sequence of vertices is composed of distinct vertices, a simple path/cycle is given
with the exception that for simple cycle the starting and ending vertices are
repeated. Here we only consider simple paths/cycles, so henceforth paths/cycles
simply refer to simple paths/cycles. P = (V1 , E1 ) is a path in G = (V, E) if:
V1 = {v1 , ..., vk } ⊂ V , and E1 = {(v1 , v2 ), (v2 , v3 ), ..., (vk−1 , vk )} ⊂ E. This
means a path consisting of k vertices is a sequence of k − 1 edges connecting
these vertices in which each two consecutive edges share a vertex in common.
Likewise, C = (V2 , E2 ) is a cycle in G = (V, E) if: V2 = {v1 , ..., vk } ⊂ V , and
E2 = {(v1 , v2 ), ..., (vk−1 , vk ), (vk , v1 )} ⊂ E. In other words, a cycle comprising
k vertices is a sequence of k edges connecting these vertices in which each two
consecutive edges as well as the first and the last edge share a vertex in common.
Hence, a cycle starts and ends at the same vertex (tour), while a path starts and
ends at different vertices (subtour).
The number of edges in a path or cycle is called the length of that path or
cycle. For G = (V, E), the set of all paths and cycles with length k are designated
by Pk (G) and Ck (G), respectively. A weight (cost) w(vi , vj ) can be assigned to
an edge. If every edge of a graph has a weight, the graph is called a weighted
graph. In a weighted graph if w(vi , vj ) = w(vj , vi ) is valid for every edge, the
weighted graph is called symmetric. In this study, for simplicity, the Euclidean
distance between two vertices of each edge is assigned as the weight of that edge
454 A. Maktabifard et al.

w(vi , vj ) = w(vj , vi ) = |r(vi ) − r(vj )|; note that the problems discussed here
are not restricted to Euclidean distance, so any desired measure (e.g., rectilinear
distance, actual driving distance) can be assigned as the weight of edges. For
a path P ∈ Pk (G), the sum of its edges’ weights is called the total cost of the
path:
 k
c(P ) = w(vi , vi+1 ) (1)
i=1

Similarly, the total cost of a cycle C ∈ Ck (G) is:

k−1

c(C) = w(vi , vi+1 ) + w(vk , v1 ) (2)
i=1

If there is no associated weight to each edge, the total cost of a path or cycle is
simply equal to the length of the path or cycle.
The classical MTSP and the MAPP problem can be formulated on the basis
of the foundation from Graph theory described above. Consider A = {a1 , ..., am }
and T = {t1 , ..., tn } as the set of m agents and n targets, respectively. The agents
and targets are located in Euclidean space. The location of the ith agent and the
j th target is therefore determined by r(ai ) and r(tj ), respectively. The classical
MTSP is formulated as follows. Consider a as the depot for all agents, so ∀ ai = a
and ∀ r(ai ) = r(a). The configuration space of the problem is the complete graph
Kn+1 (V ) with the vertex set V = T ∪ a. Consider Ci as a cycle with length ki
that starts and ends at vertex a (the depot). Let C = {C1 , ..., Cm } be the set
of m cycles Ci with length ki . The aim is to determine C such that each target
is visited only once (visitation only by one agent), and the total cost of C (Eq.
(3)) is minimized (Min-Sum problem).
m

c(C) = c(Ci ) (3)
i=1

The total cost of C is the sum of the lengths (costs) of all m cycles Ci constituting
C (see Eq. (3)).
Likewise, the MAPP problem is formulated as follows. For each agent, the
configuration space of the problem is the complete graph Kn+1 (Vi ) with the
vertex set Vi = T ∪ ai . The Euclidean distance between two vertices of each edge
(e.g., vx and vy ) is assigned as the weight of that edge; w(vx , vy ) = w(vy , vx ) =
|r(vx ) − r(vy )| where vx , vy ∈ Vi . As a result, Kn+1 (Vi ) is a symmetric weighted
graph. Consider Pi as a path with length ki starting at vertex ai . Let P =
{P1 , ..., Pm } be the set of m paths Pi with length ki that each pair of them
do not have any vertex in common (except possible same position for starting
points). The aim is to determine P such that each target is visited only once
(visitation only by one agent), and the cost of the path with the largest cost in
P (Eq. (4)) is minimized.
m
cm (P) = max c(Pi ) (4)
i=1
 Constrained Multi-agent Path Planning Problem 455

Therefore, the MAPP problem is a Min-Max problem in which a team of m

agents must visit n targets in the minimum amount of time. The whole team
of agents should be used, so each agent must visit at least one target (ki ≥ 1).
However, the number of targets visited by each agent can be different.
In this study, our additional constraints for the MAPP problem are applied
independently of each other. These constraints are defined as follows. In the
MAPP problem, if the maximum number of targets that can be assigned to
each agent is limited, the solution would be affected since the maximum length
of each agent’s path is restricted. Accordingly, we introduced the maximum
number of targets each agent is allowed to visit as the first additional constraint.
Let Q = {q1 , ..., qm } be the set of the maximum number of targets for each
agent, so the maximum number of targets that can be assigned to the ith agent
is qi . As a result, the length of each agent’s path is limited to: 1 ≤ ki ≤ qi .
In the case of applying this additional constraint, if the sum of all elements of
m
Q is less than the number of targets ( i=1 qi < n), then the fleet of m agents
cannot visit all n targets. Hereafter, this situation is referred to as the lack-of-
capacity issue. Additionally, if the maximum number of targets for each agent
is the same for all agents (∀ qi = q), the fleet of agents is called a homogeneous
capacity fleet. In this study, for solving the problem with the first additional
constraint, a homogeneous capacity fleet was assumed.
Furthermore, we defined the maximum range of the ith agent as the maximum
Euclidean distance from the position of ai that can be accessed by the ith agent.
This range is unlimited without applying any constraint on it, so each agent
can access all targets regardless of their Euclidean distance from its position.
However, if this range is limited, the only accessible targets for each agent are
located within its maximum range. This not only impacts the assignment of
targets to agents, but also can influence the total cost of each agent’s path.
Accordingly, we introduced the maximum range for each agent as the second
additional constraint. Let R = {r1 , ..., rm } be the set of the maximum range
for each agent, so the maximum range of the ith agent is ri . Consequently, the
j th target can be accessible for the ith agent only if: |r(ai ) − r(tj )| ≤ ri . In the
case of applying this additional constraint, if there would be at least one target
that is not in the maximum range of any agent (|∀ r(ai ) − r(tj )| > ri ), then the
fleet of m agents cannot visit all n targets. Henceforth, this situation is referred
to as the out-of-range issue. Moreover, if the maximum range for each agent is
the same for all agents (∀ ri = r), the fleet of agents is called a homogeneous
range fleet. In this study, for solving the problem with the second additional
constraint, a homogeneous range fleet was assumed.

4 Methodology of the Approximate Solution

In this study, a two-step solution based on the method developed in [16] is pre-
sented for the CMAPP problem. The solution method is composed of an initial
solution and a complex solution. The complex solution improves the result of
the initial solution using a GA that applies genetic-like operators and a heuristic
456 A. Maktabifard et al.

method. A solution is a set of paths that is called Plan (P) in this context. The
Plan is optimal if the longest path in the Plan is the shortest possible (Min-Max
problem). Thus, the solution algorithm is as follows:

– Step 1: a Plan P is generated by the initial solution.

– Step 2: the complex solution is applied to the Plan P to minimize the cost
cm (P) (see Eq. (4)) and eventually obtain the near-optimal Plan.

4.1 Initial Solution

The initial solution generates the Initial Plan (a starting set of paths). Consider
A = {a1 , ..., am } and T = {t1 , ..., tn } as the set of m agents and n targets,
respectively. A Plan is viable if each pair of paths do not have any vertex in
common (except possible same starting points). For the order of planning, we
assume that the targets are assigned to the agents in the following order: a1 →
a2 → · · · → am . The targets assignment algorithm is as follows:

– Step 1: a target ti1 ∈ T1 = T is selected and assigned to the first agent.

– Step 2: a target ti2 ∈ T2 = T1 − ti1 is selected and assigned to the second
agent.
..
.
– Step m: a target tim ∈ Tm = Tm−1 − tim−1 is selected and assigned to the
mth agent.

A viable Plan P = {P1 , ..., Pm } is generated by iterating this algorithm until all
targets would be assigned to the agents.
However, a target selection approach is also required for selecting a target in
every step of the targets assignment algorithm. Therefore, different initialization
methods can be employed by adopting various target selection approaches. In
this study, two initialization methods were employed:

1. Greedy initialization method:

It selects targets on the basis of the nearest neighbor heuristic and assigns
them to the agents. For each agent, the nearest unassigned target to the
previously assigned target to that agent is selected in every iteration of the
targets assignment algorithm. In order to select the first target for each agent,
the nearest unassigned target to the position of that agent (ai ) is selected.
2. Random initialization method:
It selects targets randomly and assigns them to the agents. For each agent, a
random unassigned target is selected in every iteration of the targets assign-
ment algorithm.

The solution for the non-constrained MAPP problem was also considered in
order to enable us to evaluate the effect of applying our two additional con-
straints that were applied independently of each other. Hence, six different solu-
tion modes were analyzed in this study (see Table 2).
 Constrained Multi-agent Path Planning Problem 457

Table 2. Analyzed solution modes

Solution Initialization method Additional constraint

mode Greedy Random Max. number of Max. range for
targets for each agent each agent
1 ×
2 × ×
3 × ×
4 ×
5 × ×
6 × ×

4.2 Complex Solution

In order to evolve a Plan P towards a near-optimal Multi-Agent Plan, the com-

plex solution is applied to the paths Pi ∈ P. The complex solution uses a GA
applying three genetic-like operators and another operator employing a heuristic
method to minimize the cost cm (P) (see Eq. (4)). This reduces the length of the
longest path in the Plan, and eventually the near-optimal Plan is obtained. The
algorithm of the complex solution is the same as a classical GA [11] with one
exception: there is only one Plan P generated by the initial solution to be evolved
(the population size is 1).
In every evolutionary iteration, the developed operators are applied sequen-

tially to a Plan P and generate a new Plan P , then:
 
– If cm (P ) < cm (P), Plan P is kept for the next evolutionary iteration and
Plan P is discarded.

– If cm (P ) > cm (P), Plan P is kept for the next evolutionary iteration and Plan

P is discarded.

In every evolutionary iteration, the following operators are applied in the

same order:

1. Crossover operator:
It is stochastically applied with the probability of Pcrossover . It selects two
paths stochastically either by the best-worst selection with the probability of
Pbest−worst or by random selection with the probability of 1 − Pbest−worst ,
then each of them is randomly cleaved (cleaving position can be different for
each of them) into two parts and the parts are swapped.
2. Mutation operator:
It is stochastically applied with the probability of Pmutation . It randomly
selects two paths and swaps two randomly selected targets between them.
3. Migration operator:
It is stochastically applied with the probability of Pmigration . It randomly
selects two paths Pi and Pj , then it removes a randomly selected target from
458 A. Maktabifard et al.

Pi and adds it to Pj . Obviously, in this procedure the length of Pi reduces,

while the length of Pj enlarges.
4. Boosting operator:
It is stochastically applied with the probability of Pboost . It enhances the
quality of paths by applying the 2-OPT algorithm [2]. This algorithm exam-
ines whether the inequality (5) between each four targets vi , vi+1 , vj , vj+1
of a path is valid or not. If it is valid, edges (vi , vi+1 ) and (vj , vj+1 ) are
substituted with edges (vi , vj ) and (vi+1 , vj+1 ), respectively. Applying this
algorithm generates a shorter path.

c(vi , vi+1 ) + c(vj , vj+1 ) > c(vi , vj ) + c(vi+1 , vj+1 ) (5)

5 Results and Discussion

In order to demonstrate the applicability of the developed solution, four random
scenarios with 100 test cases for each one were considered (see Table 3). In all
test cases, the agents and targets were spread out randomly in a dimensionless
domain (domain area = 1 × 1). The cost reduction Δcm was defined as the
cost reduction of the Final Plan compared to the Initial Plan (see Eq. (6)). The
following evaluation parameters were considered: Δcm is the average Δcm for
100 test cases; cm (Pf inal ) is the average final cost for 100 test cases; qi is the
maximum number of targets for each agent; ri is the average maximum range
for each agent for 100 test cases; and T100 is the runtime for 100 test cases.
cm (PInitial ) − cm (PF inal )
Δcm = × 100% (6)
cm (PInitial )

Table 3. Test scenarios

Scenario Number of agents (m) Number of targets (n)

1 10 100
2 9 100
3 20 200
4 10 200

In terms of the first additional constraint, in each test case

 the maximum
n
number of targets for each agent was determined by qi = m to ensure that
the lack-of-capacity issue would not occur. Regarding the second additional con-
straint, in each test case the developed solution numerically calculates the min-
imum value for the maximum range for each agent such that the out-of-range
issue would not occur. As a result, here the maximum range for each agent is a
case-sensitive variable.
The operators of the complex solution were applied with the probabilities
of: Pcrossover = 0.7, Pbest−worst = 0.5, Pmutation = 0.4, Pmigration = 0.6, and
 Constrained Multi-agent Path Planning Problem 459

Pboost = 0.3. Simulations were run for 150000 evolutionary iterations for each
test case by means of MATLAB R2021b. The hardware configuration of the PC
used was as follows. CPU: AMD RyzenTM 5-5500U, RAM: 8 GB.
The computational results are shown in Table 4. Accordingly, if the maximum
number of targets for each agent was restricted:

– in the case of employing Greedy initialization, the average final cost (except
in scenario 1) and the runtime decreased, while the average cost reduction
(except in scenario 1) increased compared to the non-constrained solution
with Greedy initialization.
– in the case of employing Random initialization, the runtime (except in sce-
nario 1) and the average cost reduction decreased, whereas the average final
cost increased compared to the non-constrained solution with Random ini-
tialization.

Moreover, if the maximum range for each agent was limited:

– in the case of employing Greedy initialization, the average cost reduction and
the average final cost decreased compared to the non-constrained solution
with Greedy initialization. A general trend was not observed regarding the
runtime.
– in the case of employing Random initialization, the average final cost in the
scenarios with more targets (200) and the average cost reduction decreased,
while the average final cost in the scenarios with fewer targets (100) increased
compared to the non-constrained solution with Random initialization. A gen-
eral trend was not observed regarding the runtime.

Additionally, in the absence of any additional constraints, Random initial-

ization led to a smaller average final cost compared to Greedy initialization in
scenarios with fewer targets (100). However, in the presence of additional con-
straints, Random initialization always led to a larger average final cost compared
to Greedy initialization. An interesting
√ result was that ri was always less than
34% of the domain diameter ( 1 + 1). Furthermore, if the number of agents and
n
targets increased while m remained constant, the average final cost decreased in
the case of applying the additional constraints along with Greedy initialization.
For 10 agents, when the number of targets increased by 100%, the average final
cost experienced an increase averaging 47.4% and 59.6% across solution modes 1
to 3, and solution modes 4 to 6, respectively. For 100 targets, when the number
of agents decreased by 10%, the average final cost underwent an increase averag-
ing 5.4% across all solution modes. For 200 targets, when the number of agents
decreased by 50%, the average final cost experienced an increase averaging 61%
and 45.2% across solution modes 1 to 3, and solution modes 4 to 6, respectively.
In order to visualize some instances of the evolution accomplished by the
approximate solution, Fig. 1 and Fig. 2 illustrate the Initial Plans and the Final
Plans generated by three solution modes for two different test cases. In these
figures, the yellow star is the initial position of the agent with the longest path;
the green squares are the initial positions of other agents; the paths are shown
460 A. Maktabifard et al.

Table 4. Computational results

Scenario Solution Δcm [%] cm (Pf inal ) qi ri T100 [s]

mode
1 1 42.2161 1.1562 4275.8578
2 39.5385 1.2115 10 3994.7323
3 32.4922 1.1229 0.4466 4401.3488
4 82.4312 1.1477 3702.8841
5 80.3621 1.2843 10 4007.1264
6 74.9173 1.1801 0.4466 4243.5765
2 1 41.7696 1.2622 4547.9586
2 43.6597 1.2183 12 4446.8247
3 32.6158 1.2054 0.4748 4146.6348
4 83.0654 1.2370 4387.3790
5 82.8877 1.2500 12 4157.2272
6 76.1382 1.3021 0.4748 4302.6578
3 1 36.8756 1.1599 6314.7105
2 38.3951 1.1295 10 6094.6654
3 31.0760 0.9202 0.3416 6569.2436
4 79.5629 1.3924 6533.8671
5 77.7243 1.5373 10 6235.4490
6 73.1099 1.0898 0.3416 6850.0604
4 1 31.1765 1.8080 7881.1427
2 32.6695 1.7667 20 7332.4831
3 28.0342 1.5705 0.4537 7552.6412
4 85.0111 1.8575 7835.9091
5 82.7509 2.1332 20 7334.5086
6 80.2113 1.7799 0.4537 7824.6231

by solid lines and the longest path is highlighted in all illustrated Plans; and
dashed-line circles depict the maximum range of the agents. Figure 1 shows the
Plans obtained for a test case with 9 agents and 100 targets. In this test case the
solution achieved the largest Δcm (see Eq. (6)) among 100 test cases of scenario
2 in the case of applying Greedy initialization along with the constraint on the
maximum number of targets for each agent. In the absence of any additional
constraints, the cost cm (P) (see Eq. (4)) decreased from 3.16 in the Initial Plan
(Fig. 1(a)) to 1.24 in the Final Plan (Fig. 1(b)). If the maximum number of
targets for each agent was restricted to qi = 12, the cost cm (P) reduced from 3.16
in the Initial Plan (Fig. 1(c)) to 1.17 in the Final Plan (Fig. 1(d)). Lastly, if the
maximum range for each agent was limited to ri = 0.38, the cost cm (P) dropped
from 1.59 in the Initial Plan (Fig. 1(e)) to 1.17 in the Final Plan (Fig. 1(f)).
Furthermore, the Plans generated for a test case with 10 agents and 200 targets
 Constrained Multi-agent Path Planning Problem 461

Fig. 1. Initial Plans (left) vs. Final Plans (right) generated by solution modes 1, 2 and
3; scenario 2; the test case in which the solution achieved the largest Δcm in the case
of applying Greedy initialization along with the constraint on the maximum number
of targets for each agent. (Color figure online)
462 A. Maktabifard et al.

Fig. 2. Initial Plans (left) vs. Final Plans (right) generated by solution modes 1, 2 and
3; scenario 4; the test case in which the solution achieved the largest Δcm in the case
of applying Greedy initialization along with the constraint on the maximum range for
each agent. (Color figure online)
 Constrained Multi-agent Path Planning Problem 463

are illustrated in Fig. 2. In this test case the solution attained the highest Δcm
among 100 test cases of scenario 4 if Greedy initialization was applied together
with the constraint on the maximum range for each agent. In the absence of
any additional constraints, the cost cm (P) reduced from 2.86 in the Initial Plan
(Fig. 2(a)) to 1.74 in the Final Plan (Fig. 2(b)). If the maximum number of
targets for each agent was limited to qi = 20, the cost cm (P) dropped from 2.86
in the Initial Plan (Fig. 2(c)) to 1.95 in the Final Plan (Fig. 2(d)). Finally, if
the maximum range for each agent was restricted to ri = 0.42, the cost cm (P)
decreased from 2.60 in the Initial Plan (Fig. 2(e)) to 1.52 in the Final Plan
(Fig. 2(f)).
Comparing our developed method with previously proposed methods using
a similar GA for the MAPP problem, our method is novel because despite the
similarity in utilizing the mutation operator and the 2-OPT algorithm in [10],
their initialization phase (initial solution) and crossover operator function dif-
ferently. Moreover, our additional constraints were not applied in [16] and their
initialization phase uses a different targets assignment algorithm. The presented
method can be extended by applying other practical constraints realizing time
windows and priorities [5,13,19,25], conflict-free paths [3,7], and multi-modal
itinerary planning [21].

6 Conclusions

This paper presents a cooperative path planner based on the Multi-Agent Path
Planning (MAPP) problem with additional constraints to model more practi-
cal situations. The proposed solution uses a Genetic Algorithm (GA) applying
genetic-like operators and a heuristic method. It evolves an Initial Plan towards
a near-optimal Multi-Agent Plan by minimizing the length of the longest path.
The applicability of the approximate solution was tested in four random
scenarios. The computational results show that restricting the maximum number
of targets for each agent can mostly improve the performance of the solution if
the Initial Plan is generated by Greedy initialization. Furthermore, on average
the cost of the Final Plan can be decreased by limiting the maximum range
for each agent if the Initial Plan is generated by Greedy initialization. This
trend is observed in the scenarios with more targets (200 in this study) even if
the Initial Plan is generated by Random initialization. These results show that
using a simple yet efficient initial solution, the developed method can potentially
enhance the efficiency of route planning for a fleet of vehicles when the aim is
to optimize open paths (subtours) in a cooperative manner, and the vehicles of
the fleet have a limited operational range or a balance between their workload
is desired. This situation can be observed in a variety of mobility and logistics
services such as crowd-sourced delivery (e.g., TOURMIX), group ride-sharing,
general home delivery, and airport shuttles.
464 A. Maktabifard et al.

The ultimate goal of this work is to simulate the challenges observed in

cooperative path planning for vehicle fleets. Thus, it can be extended by applying
other practical constraints like time windows and priorities. Our future work will
focus on developing a Multi-Agent route planner for shared mobility services
using the method presented here but by calculating the length of paths based
on map data (i.e., actual driving distance).

References
1. Bektas, T.: The multiple traveling salesman problem: an overview of formulations
and solution procedures. Omega 34(3), 209–219 (2006). https://doi.org/10.1016/
j.omega.2004.10.004
2. Bentley, J.L.: Experiments on traveling salesman heuristics. In: Proceedings of the
First Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 1990), USA,
pp. 91–99. SIAM (1990)
3. Bose, J., Reiners, T., Steenken, D., Voss, S.: Vehicle dispatching at seaport con-
tainer terminals using evolutionary algorithms. In: Proceedings of the 33rd Annual
Hawaii International Conference on System Sciences (HICSS), USA, vol. 2, pp. 1–
10. IEEE (2000). https://doi.org/10.1109/HICSS.2000.926669
4. Chen, J., Ling, F., Zhang, Y., You, T., Liu, Y., Du, X.: Coverage path planning of
heterogeneous unmanned aerial vehicles based on ant colony system. Swarm Evol.
Comput. 69, 101005, 1–11 (2022). https://doi.org/10.1016/j.swevo.2021.101005
5. Fournier, S.M.R., Hülse, E.O., Pinheiro, É.V.: A*-guided heuristic for a multi-
objective bus passenger Trip Planning Problem. Public Transp. 13, 557–578 (2021).
https://doi.org/10.1007/s12469-019-00204-1
6. Gavish, B., Srikanth, K.: An optimal solution method for large-scale multiple trav-
eling salesmen problems. Oper. Res. 34(5), 698–717 (1986). https://doi.org/10.
1287/opre.34.5.698
7. Gawrilow, E., Köhler, E., Möhring, R.H., Stenzel, B.: Dynamic routing of auto-
mated guided vehicles in real-time. In: Krebs, H.-J., Jäger, W. (eds.) Mathematics -
Key Technology for the Future, pp. 165–177. Springer, Heidelberg (2008). https://
doi.org/10.1007/978-3-540-77203-3 12
8. Ghafurian, S., Javadian, N.: An ant colony algorithm for solving fixed destina-
tion multi-depot multiple traveling salesmen problems. Appl. Soft Comput. 11(1),
1256–1262 (2011). https://doi.org/10.1016/j.asoc.2010.03.002
9. Giardini, G., Kalmár-Nagy, T.: Genetic algorithm for combinatorial path planning:
the subtour problem. Math. Probl. Eng. 2011, 483643, 1–31 (2011). https://doi.
org/10.1155/2011/483643
10. Giardini, G., Kalmar-Nagy, T.: Genetic algorithm for multi-agent space explo-
ration. In: Proceedings of AIAA Infotech@Aerospace 2007 Conference and Exhibit
2824, USA, pp. 1–15. AIAA (2007). https://doi.org/10.2514/6.2007-2824
11. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learn-
ing. Addison-Wesley Publishing Company Inc., Boston (1989)
12. Harrath, Y., Salman, A.F., Alqaddoumi, A., Hasan, H., Radhi, A.: A novel hybrid
approach for solving the multiple traveling salesmen problem. Arab J. Basic Appl.
Sci. 26(1), 103–112 (2019). https://doi.org/10.1080/25765299.2019.1565193
13. Herbawi, W., Weber, M.: The ridematching problem with time windows in dynamic
ridesharing: a model and a genetic algorithm. In: Proceedings of 2012 IEEE
Congress on Evolutionary Computation, USA, pp. 1–8. IEEE (2012). https://doi.
org/10.1109/CEC.2012.6253001
 Constrained Multi-agent Path Planning Problem 465

14. Hou, M., Liu, D.: A novel method for solving the multiple traveling salesmen
problem with multiple depots. Chin. Sci. Bull. 57(15), 1886–1892 (2012). https://
doi.org/10.1007/s11434-012-5162-7
15. Jiang, C., Wan, Z., Peng, Z.: A new efficient hybrid algorithm for large scale
multiple traveling salesman problems. Expert Syst. Appl. 139, 112867, 1–11 (2020).
https://doi.org/10.1016/j.eswa.2019.112867
16. Kalmár-Nagy, T., Giardini, G., Bak, B.D.: The multiagent planning problem. Com-
plexity 2017, 3813912, 1–12 (2017). https://doi.org/10.1155/2017/3813912
17. Kara, I., Bektas, T.: Integer linear programming formulations of multiple salesman
problems and its variations. Eur. J. Oper. Res. 174(3), 1449–1458 (2006). https://
doi.org/10.1016/j.ejor.2005.03.008
18. Karabulut, K., Öztop, H., Kandiller, L., Tasgetiren, M.F.: Modeling and opti-
mization of multiple traveling salesmen problems: an evolution strategy approach.
Comput. Oper. Res. 129, 105192, 1–19 (2021). https://doi.org/10.1016/j.cor.2020.
105192
19. Király, A., Abonyi, J.: Redesign of the supply of mobile mechanics based on a novel
genetic optimization algorithm using Google Maps API. Eng. Appl. Artif. Intell.
38, 122–130 (2015). https://doi.org/10.1016/j.engappai.2014.10.015
20. Laporte, G., Nobert, Y.: A cutting planes algorithm for the m-salesmen problem.
J. Oper. Res. Soc. 31(11), 1017–1023 (1980). https://doi.org/10.1057/jors.1980.
188
21. Redmond, M., Campbell, A.M., Ehmke, J.F.: Data-driven planning of reliable
itineraries in multi-modal transit networks. Public Transp. 12, 171–205 (2020).
https://doi.org/10.1007/s12469-019-00221-0
22. Ryan, J.L., Bailey, T.G., Moore, J.T., Carlton, W.B.: Reactive Tabu Search in
unmanned aerial reconnaissance simulations. In: Proceedings of 1998 Winter Sim-
ulation Conference (Cat. No. 98CH36274), USA, vol. 1, pp. 873–879. IEEE (1998).
https://doi.org/10.1109/WSC.1998.745084
23. Somhom, S., Modares, A., Enkawa, T.: Competition-based neural network for the
multiple travelling salesmen problem with minmax objective. Comput. Oper. Res.
26(4), 395–407 (1999). https://doi.org/10.1016/S0305-0548(98)00069-0
24. Song, C.-H., Lee, K., Lee, W.D.: Extended simulated annealing for augmented TSP
and multi-salesmen TSP. In: Proceedings of the International Joint Conference on
Neural Networks (IJCNN), USA, vol. 3, pp. 2340–2343. IEEE (2003). https://doi.
org/10.1109/IJCNN.2003.1223777
25. Tong, Y., Zeng, Y., Zhou, Z., Chen, L., Ye, J., Xu, K.: A unified approach to
route planning for shared mobility. Proc. VLDB Endow. 11(11), 1633–1646 (2018).
https://doi.org/10.14778/3236187.3236211
26. Torabbeigi, M., Lim, G.J., Kim, S.J.: Drone delivery scheduling optimization con-
sidering payload-induced battery consumption rates. J. Intell. Robot. Syst. 97(3),
471–487 (2020). https://doi.org/10.1007/s10846-019-01034-w
27. Wang, X., Liu, D., Hou, M.: A novel method for multiple depot and open paths,
multiple traveling salesmen problem. In: Proceedings of 2013 IEEE 11th Interna-
tional Symposium on Applied Machine Intelligence and Informatics (SAMI), USA,
pp. 187–192. IEEE (2013). https://doi.org/10.1109/SAMI.2013.6480972
28. Yu, H., Meier, K., Argyle, M., Beard, R.W.: Cooperative path planning for tar-
get tracking in urban environments using unmanned air and ground vehicles.
IEEE/ASME Trans. Mechatron. 20(2), 541–552 (2015). https://doi.org/10.1109/
TMECH.2014.2301459
466 A. Maktabifard et al.

29. Yuan, S., Skinner, B., Huang, S., Liu, D.: A new crossover approach for solving
the multiple travelling salesmen problem using genetic algorithms. Eur. J. Oper.
Res. 228(1), 72–82 (2013). https://doi.org/10.1016/j.ejor.2013.01.043
30. Zhou, H., Song, M., Pedrycz, W.: A comparative study of improved GA and PSO
in solving multiple traveling salesmen problem. Appl. Soft Comput. 64, 564–580
(2018). https://doi.org/10.1016/j.asoc.2017.12.031
 UAV Path Planning for Area Coverage
and Energy Consumption in Oil and Gas
Exploration Environment

Salim Sulaiman Maaji(B) and Dario Landa-Silva

School of Computer Science, University of Nottingham, Nottingham NG8 1BB, UK

{salim.maaji2,dario.landasilva}@nottingham.ac.uk

Abstract. This paper proposes a model for unmanned aerial vehicles (UAV) grid-
based coverage path planning, considering coverage completeness and energy con-
sumption in complex environments with multiple obstacles. The work is inspired
by the need for more efficient approaches to oil and gas exploration, but other
application areas where UAVs can be used to explore unknown environments can
also benefit from this work. An energy consumption model is proposed that con-
siders acceleration, deceleration, and turning manoeuvres, as well as the distance
to obstacles, to more accurately simulate the UAV’s movement in different envi-
ronments. Three different environments are modelled: desert, forest, and jungle.
The energy-aware coverage path planning algorithm implemented seeks to reduce
the energy consumption of a single drone while increasing coverage completeness.
The model implementation and experiments were performed in the ROS/Gazebo
simulation software. Obtained results show that the algorithm performs very well,
with the drone able to manoeuvre itself in a combination of hills, valleys, rugged
terrain, and steep topography while balancing coverage and energy consumption.

Keywords: UAV Path Planning · Drone Energy Optimisation · Coverage Path

Planning · Multiple Obstacle Avoidance · Exploration in Unknown Environment

1 Introduction

Environmental concerns have brought challenges to the oil and gas industry, including
pressure to reduce and stabilise costs for competitive advantage, improve its environ-
mental blueprint, and optimise its performance. In recent years, the production of most
onshore oil fields has declined [12] in part because, after a certain level of recovery, pro-
duction costs do not justify further investment. Oil and gas exploration often takes place
in difficult-to-reach environments where the use of robotic systems can be useful [37].
Petroleum deposits are generated by a natural process that commonly occurs at great
depth and is often poorly understood and predicted by earth scientists. Hence large areas
with significant oil and gas potential remain unexplored [25]. In 2015, Shell Oil decided
to abandon efforts to find and develop hydrocarbon resources in the Chukchi Sea, despite
having spent billions of dollars in exploration, which in the end turned out to be a dry

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 467–481, 2023.
https://doi.org/10.1007/978-3-031-43612-3_29
468 S. S. Maaji and D. Landa-Silva

hole [38]. This type of risk cannot be eliminated, but it can be reduced through tech-
nological innovations. An example of such technology is magnetic exploration which
measures variations in the earth’s magnetic field to identify rocks containing hydrocar-
bon [1]. Aeromagnetic surveys are used to measure magnetic anomalies using low-flying
aircraft carrying a high-precision magnetometer, a sensor for magnetic anomaly detec-
tion [2]. However, aeromagnetic surveying with low-flying aircraft brings challenges in
terms of safety, terrain complexity, quality of data, and costs. An attractive alternative
is to conduct magnetic exploration with UAVs [38]. UAVs can fly close to the surface
at optimal speed with efficient coverage for gathering good-quality magnetic data [42].
UAVs are fast becoming an attractive solution for many scenarios like search and rescue
[19], precision agriculture [31], and delivery systems [26].
Magnetic exploration with UAVs requires effective path planning in unknown envi-
ronments and with possibly multiple obstacles (e.g., trees, large rocks, etc.). The robotic
path planning problem can be divided into two main categories: motion path planning
and coverage path planning. In motion path planning, there is a clear start point and an
end point; the goal is to optimally cover the distance between the start to the endpoint
at minimum cost while avoiding obstacles [35]. Coverage path planning (CPP) finds an
optimal collision-free path that a robot must take to pass over each point in an area of
interest in the given environment [10]. CPP is related to the covering salesman problem
where an agent must travel a minimum-length tour covering the subsets of given cities or
customers [46, 24]. CPP can be offline or online depending on the availability of a priori
information about the area of interest [18], and several algorithms exist to tackle this
problem [40]. In CPP, the area of interest can be decomposed into smaller regions or not
at all. Simple and regular-shaped environments require no decomposition, and simple
geometric patterns such as back-and-forth (BF), zigzag movement, or spiral patterns
can be used to solve the problem [34]. When the exploration area is irregular-shaped
and complex, decomposition can be done in different ways, like exact cellular decom-
position [28], approximate cellular decomposition [8, 21], or grid-based decomposition
[18]. For the scenarios investigated in this paper, grid-based decomposition was used to
ensure that every cell within the area of interest is visited once. This method requires
computational power to represent the cells at higher resolution grids [24]. After the grid
is produced, a traveling salesman algorithm is applied to generate a sequence of nodes
or subregions to visit [4]. The coverage path is generated by connecting these nodes in
sequence using back-and-forth motion perpendicular to the sweep direction from the
start to the goal region [9, 28].
Magnetic surveying with UAVs for oil/gas exploration requires complete coverage of
the area of interest. Performing CPP with UAVs in an environment with multiple obsta-
cles is energy-demanding due to the need for obstacle avoidance, and limited energy
capacity is a feature of UAVs [20]. The approach in this paper minimised energy con-
sumption by limiting accelerations, decelerations, and turning manoeuvres by the drone
while at the same time maximising coverage. Aided by simultaneous localisation and
mapping (SLAM), a model for coverage path-planning in a regular and irregular-shaped
environment is developed in ROS/Gazebo platform. The regular-shaped environment
has a perfect square-shaped setting with no mountains, hills, or valleys. In contrast,
 UAV Path Planning for Area Coverage and Energy Consumption 469

the irregular-shaped environment has a complex terrain with no specific shape. Exper-
iments were conducted to study the tradeoff between area coverage and energy con-
sumption. Section 2 outlines related research on coverage path planning. Section 3
describes the methodology and cost function. Section 4 presents the experimental result
and discussions, and concluding remarks are presented in Sect. 5.

2 Related Research

Coverage path planning (CPP) has been extensively studied in the literature but is still
an open problem in robotics [40]. Different approaches have been adopted to classify
the problems into (i) increasing the coverage completeness, (ii) reducing the path over-
lapping, (ii) reducing the energy consumption, (iv) optimising the number of turns, and
(v) reducing the time to completion. A randomised approach does not require sensors
and algorithms for localisation, but such an approach is inefficient when dealing with
large coverage areas requiring more energy and time to complete the coverage [14].
A distributed strategy and model for UAVs oil spill mapping used randomness and
probabilistic guessing to avoid visiting all the cells and hence reduce the total distance
travelled, but coverage completeness decreased [32]. Different geometric patterns and
shapes for coverage patterns have been studied, including back-and-forth (BF), Hilbert
curves, LMAT (chain of equilateral triangles), S curves, and spiral patterns [5]. An opti-
mal line-sweep decomposition path planner to minimise the time required for covering
the area with obstacles in an unknown environment was presented in [4]. The authors
claimed that changing the sweep direction, as shown in Fig. 1, helps in minimising the
number of turns and hence reducing the time to completion. Similarly, reducing the
number of turns using optimal sweep direction and coverage pattern was considered in
[3, 24] in a quest to reduce energy consumption. Energy-aware CPP is an active area of
research aiming at minimum energy consumption and maximum area coverage. Some
studies considered energy optimisation based on the UAV structure, aerodynamic prop-
erties, rotor efficiency, and energy consumed in onboard data processing and controls
[17–44].

Fig. 1. Coverage path-planning with sweep direction changed to reduced turning angles.

Energy-aware coverage path planning algorithms with a high emphasis on trajecto-

ries that reduce power consumption during operations have been proposed [15, 36]. It
has been found that the length of the trajectory, the UAV speed, and the way to make turns
along the trajectory are the main energy sinks during the operation. In the literature, the
performance index for energy consumption was to reduce the path length traveled during
the coverage to minimise the energy consumption [45, 47], improve time-to-completion
470 S. S. Maaji and D. Landa-Silva

for both single and multi-robot coverage [13, 20, 30], and reduce the number of turns [11,
41]. An algorithm for single-robot coverage path planning under constrained energy was
presented in [39]. The robot was aware of its energy limitation, and hence the total dis-
tance traveled to refueling was minimised. Boustrophedon cellular decomposition using
back-and-forth or “ox-plow” motions and Dijkstra’s algorithm were used. Although
reducing the total distance to refueling is important, most of the energy is often con-
sumed during the coverage and turning manoeuvres. Grid-based coverage path planning
exhibits excellent performance in several robotic applications in irregular shapes and no
fly-zones environments [5, 7]. In the work proposed by [43], the irregular-shaped area of
interest was decomposed into regular cells of equal sizes, using an approximate cellular
decomposition technique. The authors introduced a cost function to minimise the num-
ber of turning manoeuvres to save energy but did not take into account important factors
like optimal energy and energy consumption due to accelerations and decelerations. The
energy-aware algorithm proposed by [6] drastically reduced the energy consumption
for the entire coverage. The model was based on optimising the turning manoeuvres,
and avoiding sharp angles while reducing speed, acceleration, and deceleration at the
turning angles. An improved cost function aimed at minimising energy consumption in
an irregular shape was proposed in [7]. The model was good in energy saving, and it may
scale well in a simplistic environment, especially those applications that do not require
a high degree of coverage completeness. Only a single occlusion point was considered
in their experiment, but multiple obstacles can have a significant effect on the coverage
completeness.
Then, optimising the coverage completeness, energy consumption, and distance to
obstacles in complex oil and gas environments with multiple occlusions is an interesting
problem that has not been investigated to the best of our knowledge. The problem
is unique because magnetic survey for oil/gas exploration requires near optimum line
spacing and flight height, and the coverage completeness can be easily affected when
trying to avoid obstacles. The environment is often mixed with regular and irregular
settings of different terrains. In addressing this problem, the following question arises:
Can an intelligent coverage path-planning algorithm that reduces the number of sharp
turns for obstacle avoidance in a highly irregular-shaped environment reduce the amount
of energy consumed by the UAV without reducing the coverage completeness?

3 Proposed Methodology

The proposed model was implemented in the ROS/Gazebo simulation software using
a drone package hector quadrotor noetic [33]. As a proof of concept for the magne-
tometer, the drone is integrated with a magnetic sensor to read the magnetic field of
the Gazebo environment. The modelled exploration environment is rugged terrain with
steep topography, variations in the surface elevations, and thick vegetation cover, as
illustrated in Fig. 2. The environment E is then partitioned into a grid with multiple
cells of equal size Mx × My as illustrated in Fig. 3. Each cell (x, y) has an associated
value u(x, y, t) ∈ [0, 1] that represents the UAV’s uncertainty about the target distri-
bution in the cell at time interval t, for occupancy grid mapping, implemented using the
slam_gmapping package provided in ROS [22]. The distance between each grid cell for
 UAV Path Planning for Area Coverage and Energy Consumption 471

cell-to-cell movement can be calculated using Eq. (1) where xi , yi are the target goal
point [27].
m
d (x, y) = |xi − yi | (1)
i=1

Fig. 2. Example of forest environment modeled in the Gazebo simulation software for oil and
gas exploration.

3.1 Energy Model

To derive the energy model, we first analysed energy consumption data obtained from
real flights of a HEIFU drone [29] with a 22,000 mAh LiPo battery, and the nominal
voltage and current were at 24 V and 60 A, respectively. The drone consumes around 23.5
V for hovering, 20.3 V for liner movement at a constant speed, 21.7 V when turning at a
soft angle, and 22.6 V at a sharp bend. The nominal voltage drops drastically at around
35 min of flight from 25.2 V to 18.6 V; hence the drone has to come back for refueling.
The drone’s weight was 7.5 kg with maximum payload of 6 kg for a maximum take-off
weight of 13.5 kg. To develop the energy cost function, we split the UAV path into a
set of segments as in [7, 16]. The first segment consists of the path for constant speed
(straight line movement), which does not consume much energy, compared to the second
segment, the variable speed of the UVA, which includes hovering and angular movements
for maneuvering and obstacle avoidance. The variable speed consumes energy due to
the acceleration, deceleration, and point of discontinuity along the path. It is important
to note that the energy model proposed by [43] and extended by [7] does not consider
the effect of minimising the energy on coverage completeness, especially in multiple
obstacle environments. Following the same idea, the cost function has been extended to
consider coverage completeness and energy minimisation as two conflicting objective
functions.
Figure 3 illustrates the concept of obstacle avoidance in coverage path planning. It
shows a UAV flying at a fixed height h from the ground acquiring magnetic data from
a cell in the grid. The field of view is the magnetometer’s projected area of size 1 × 1
472 S. S. Maaji and D. Landa-Silva

Fig. 3. Drone coverage path-planning with grid decomposition in an occluded environment.

square meter. An obstacle in the form of a tree is shown at the center of the exploration
environment. Three possible waypoints are shown, marked as a, b, and c. Waypoint
(a) maximises coverage in the occluded region but consumes more energy due to the
sharp angles at the path. Waypoint (b) minimises energy consumption due to the reduced
distance by the resultant vector but still consumes energy due to turning manoeuvres.
Waypoint (c) consumes less energy because of the smoothed angle on the waypoint but
with reduced coverage. The problem lies in optimising Delta , the distance between
the drone and the obstacle during the coverage as a constraint to be imposed for obstacle
avoidance by the UAV. Minimising the delta increases the coverage, and maximising it
reduces energy consumption.
Given an initial position (xo , yo ) and target goal points (xi , yi ), a collision-free path
ensures that at no time should the UAV enter into the no-fly-zone region. The equation
for collision avoidance is presented as follows:
∀p ∈ [1 . . . N ], ∀i ∈ [0 . . . T]
 
xip − xk  ≥  (2)
 
yip − yk  ≥  (3)
In the Eqs. (2), (3) above, xip , yip are the drone position in the x and y axis at the T
time instance, and xk , yk are the obstacle position at the x-axis and y-axis, respectively.
Since most obstacles in the forest can be approximated to the shape of a sphere or circle
in 2D, Eqs. (2) and (3) can be written as Eq. (4).
   
xip − xk  sin θ + yip − yk  cos θ ≥  (4)
The instantaneous power for the flight scenarios is calculated with Eq. (5), where V
is an instantaneous voltage, and I is the instantaneous current consumed by the drone.
P = VI (5)
 UAV Path Planning for Area Coverage and Energy Consumption 473

The total energy consumption can be calculated by integrating the power over the
period of the mission, as shown in the energy Eq. (6), where E is the energy consumed
in joules, P is the instantaneous power consumed in watts, and t is the time interval for
the flight mission.
 t  t
E= Pdt = VIdt (6)
0 0
The energy consumed when moving at a straight line to cover a distance d at constant
speed v can be calculated with Eq. (7).
d /v d
Ec = ∫ Pc dt = Pc (7)
0 v
The energy consumed for variable speed due to acceleration and deceleration can be
calculated with Eq. (8).
di /vi do /vo
Evar = ∫ Pacc dt = Pv (t2 − t1 ) + ∫ Pdecc dt = Pv (t2 − t1 ) (8)
do /vo di /vi

The energy consumed for hovering is calculated using Eq. (9).

 h/vclimb
Eh = Ph dt = Ph (t2 − t1 ) (9)
0
In this work, the drone is restricted not to flying above trees for quality data collection
as well as for manoeuvability. Hence the maximum height is set to 1 m as a constraint.
Finally, the energy consumed during the rotation, maneuvering, and turning at a certain
angle θ, with angular speed ω is calculated with Eq. (10).
 θ/ωi
θ
Eturn = Pturn dt = Pturn (10)
θo/ωo ω
The cost function can be written by summing the total energy consumed in straight
line movement, which also has sub-components of the constant speed and variable speed
due to acceleration and deceleration, and the energy consumed in the rotational move-
ment for turning and maneuvering in obstacle avoidance. The cost function is given by
Eq. (11).
  0  d /v  θ/ωi 
m vi
Ec = Pacc dt + Pdec dt + Pc dt + Pturn dt (11)
i=1 0 vi 0 θo/ωo

While the total energy needed for the coverage can be compud with Eq. (12).
  0  d /v
m vi
Etotal = Pacc dt + Pdec dt + Pc dt
i=1 0 vi 0
  
h/vclimb θ/ωi
+ Ph dt + Pturn dt (12)
0 θo/ωo
474 S. S. Maaji and D. Landa-Silva

3.2 Coverage-Path Pseudocodes

Given the above energy model for the UAV, the coverage path planning shown below was
implemented. The algorithm can be summarised in three sequential main steps: decom-
position, planning, and execution. First, a cellular decomposition technique is applied
over the irregular-shaped area to discretise the map to regular equal size cells. Second,
a coverage path planning algorithm finds the nearest cell in back-and-forth movement
based on the coverage goal received. The solution to the movement is according to the
predefined cost function and obstacle avoidance constraint in the model. Finally, the
resulting path is executed, and the coverage mission is completed.

The UAV energy model, environment model, and coverage path planning approach
were all implemented in the ROS/Gazebo simulation software. Several experiments
were conducted for various types of exploration scenarios. The scenarios consist of a
perfect square regular-shaped environment and an irregular-shaped environment char-
acterised by arbitrarily shaped obstacles. We investigated the tradeoff between coverage
completeness and energy consumption in these environments.
 UAV Path Planning for Area Coverage and Energy Consumption 475

4 Experimental Result
Three different environments usually seen in oil/gas exploration were modelled: a desert
with low obstacles, including hills and valleys; a forest with a combination of hills, val-
leys, and moderate vegetation; and a jungle depicted rugged terrain and steep topography.
A survey conducted by [48] shows that flying drones at varying altitudes of 0.5 m, 1.3
m, and 2.2 m above the ground with line spacing of 1 m provides a good result. The
flight altitude was set to 1 m, and the speed was set to v = 2 m/s for the entire experi-
ment. The experiment in a desert environment is considered a benchmark to evaluate the
coverage quality, assuming the drone could successfully cover the cell with minimum
energy consumption without obstacles.

4.1 Exploration Scenarios

First, an experiment was conducted in a rectangular-shaped environment, and three
flights were performed to compare the algorithm’s effectiveness in deserts, forests, and
jungles. The environment is first decomposed into 20 cells of equal size (5 × 4) and delta
set to 0.55 m. The obtained result shows that exploration in the desert consumes less
energy and provides full coverage. The energy consumption doubled in the forest, and
the coverage was 19 cells out of 20 target cells. The worst performance was recorded in
the jungle exploration, in which the mean coverage was 11 cells out of 20 target cells,
and the energy consumption was still high. A cross-section of the simulation result from
the ROS-rviz software is shown in Fig. 4.

Fig. 4. The red dots represent the area in which the drone performs coverage path planning. The
gray area represents a region in which the UAV is certain there is no obstacle, i.e., path-planning
can be performed on it. The dark edges in the gray regions are obstacle or no-fly zones identified
based on the principle of occupancy grid mapping, and the remaining part of the graph in light
green represents the unexplored area.

To further assess the performance of the algorithm, another experiment was con-
ducted using the same parameters but in a much bigger and irregular map illustrated in
Fig. 5. The area was decomposed into 110 cells of equal size, and the algorithm scales
well in the desert, and also performs well in the forest with few obstacles (Fig. 5b),
as well as in the jungle with more obstacles (Fig. 5c). In the jungle exploration, the
number of uncovered cells was 31 compared to 36 in the forest. In the second sce-
nario, the delta was s et to 0.75 m, and an experiment was also conducted again in a
476 S. S. Maaji and D. Landa-Silva

regular-shaped environment, decomposed into 20 cells of equal size (5 × 4). It was also
run in an irregular-shaped environment decomposed into 110 cells of equal size. Like
the first scenario, three flights were performed to compare the algorithm’s effectiveness
in deserts, forests, and jungles. In the third scenario, the delta was set to 1.0 m, and
the experiment was conducted similarly to scenarios one and two. The regular-shaped
environment was decomposed into 20 cells of equal size (5 × 4). The irregular-shaped
environment was decomposed into 110 cells of equal size. Three flights were performed
to compare the algorithm’s effectiveness in deserts, forests, and jungles. Setting the delta
to 1.0 m makes the coverage very difficult in an obstacle environment. In most cases,
the drone has to abandon the mission because it was trapped for over 8 min facing an
obstacle, i.e., it cannot pass through a narrow path because the delta is too big. Hence,
the energy consumed in the jungle was less than what was consumed in the forest. The
trees in the forest along the way point were dense, and the drone was trapped hence
abandoning the mission before completion.

Fig. 5. Coverage path planning in desert, forest, and jungle for more complex environments.

In some cases, the drone was only able to cover 7 out of the 20 cells in a regular-shaped
forest environment and 29 out of 110 cells in an irregular-shaped forest environment.
Tables 1 and 2 show the average value of the energy consumption and coverage for the
three flights performed in deserts, forests, and jungles.

Table 1. Mean energy (J) vs. Coverage (sq m) regular-shaped environment.

Environment Delta = 0.55 m Delta = 0.75 m Delta = 1.0 m

Mean Mean Mean Mean Mean Mean
Energy Coverage Energy Coverage Energy Coverage
Desert 411.5 20 417.0 20 402.1 20
Forest 1120.9 19 393.5 19 724.7 7
Jungle 1198.2 11 523.6 11 259.2 4
 UAV Path Planning for Area Coverage and Energy Consumption 477

Table 2. Mean energy (J) vs. Coverage (sq m) irregular-shaped environment.

Environment Delta = 0.55 m Delta = 0.75 m Delta = 1.0 m

Mean Mean Mean Mean Mean Mean
Energy Coverage Energy Coverage Energy Coverage
Desert 2451.6 110 2388.8 110 2434.3 110
Forest 3462.7 73 942.7 26 2070.6 29
Jungle 2639.8 78 1380.5 28 1256.9 24

4.2 Evaluation of Experimental Results

The algorithm’s performance was evaluated in terms of increased coverage and reduced
energy consumption. The coverage results for regular-shaped and irregular-shaped envi-
ronments are compared in Fig. 1 and Fig. 2, where the vertical axis is the number of
cells covered, and the horizontal axis shows the various delta values for experiments in
the desert, forest, and jungle. As can be seen, the algorithm achieved better coverage
results in a regular-shaped environment when the delta was set to 0.55 m and 0.75 m, but
the result for forest and jungle got worst when the delta was set to 1.0 m. The findings
from these studies suggest that delta plays an important role and it has significance on
energy consumption and coverage completeness. It allows the drone to maneuver freely
out of the obstacle. That is, the higher the delta, the less energy consumption; the lower
the delta, the more coverage completeness. As expected, there is a tradeoff between the
number of cells covered and energy consumption in the occluded environment. Figure 6
and 7 compare energy consumption and coverage completeness for both regular and
irregular-shaped environments.

120
100
Cells(sq meter)

80
60
40
20
0
Desert Forest Jungle Desert Forest Jungle

Regular-shaped envir. Irregular-shaped envir.

Delta =0.55m Coverage Delta =0.75m Coverage Delta =1.0m Coverage

Fig. 6. Coverage completeness for regular-shaped and irregular-shaped environments.

478 S. S. Maaji and D. Landa-Silva

4000
Energy (Joules) 3500
3000
2500
2000
1500
1000
500
0
Desert Forest Jungle Desert Forest Jungle

Regular-shaped envir. Irregular-shaped envir.

Delta =0.55m Energy Delta =0.75m Energy Delta =1.0m Energy

Fig. 7. Mean energy (J) in regular and irregular-shaped environments.

5 Conclusion

This paper described a model for UAV Path Planning considering the tradeoff between
area coverage and energy consumption. The work is inspired by magnetic data sur-
veying for oil/gas exploration, where environments are usually unknown and present
multiple obstacles. Simulations were implemented in ROS/Gazebo for three environ-
ments: desert, forest, and jungle. A parameter delta is used to set the constraint for
obstacle avoidance, and experiments with different values (0.5 m, 0.75 m, and 1.0 m)
were conducted. Experimental results show that the drone effectively maneuvered itself
in different environmental settings, achieving high coverage and moderate energy con-
sumption when the delta was set to 0.5 m in a desert environment. As the complexity
of the environment increases, energy consumption increases due to obstacle avoidance
while the coverage optimality reduces. The overall best results were obtained when the
delta was set to 0.5 m, skipping a few cells while performing coverage in forest and
jungle but with added energy consumption compared to exploration in the desert. The
worst coverage result was obtained when the delta was set to 1.0 m in the forest and
the jungle. These results show that the delta value influences energy consumption and
coverage completeness. The larger the delta value, the less energy consumption due to
the smoothed angle and less acceleration and deceleration along the obstacle avoidance
path. The lower the delta value, the more coverage completeness. There was a tradeoff
between the number of cells covered and energy consumption in the occluded environ-
ment, and in some cases, the drone was trapped by the obstacles. Further work will look
at more sophisticated path-planning algorithms and multiple drones.
 UAV Path Planning for Area Coverage and Energy Consumption 479

References
1. Abubakar, R., et al.: Formation of magnetic minerals at hydrocarbon-generation conditions.
Mar. Pet. Geol. 68, 509–519 (2015). https://doi.org/10.1016/j.marpetgeo.2015.10.003
2. Aminzadeh, F., Dasgupta, S.: Geophysics for petroleum engineers: chapter 3. Fundamentals
of Petroleum Geophysics, 1st edn, vol. 60, pp. 1–282. Elsevier Inc., UK (2013). ISBN: 978-
0-444-50662-7
3. Araujo, J., Sujit, P., Sousa, J.: Multiple UAV area decomposition and coverage. In: 2013 IEEE
Symposium on Computational Intelligence for Security and Defense Applications (CISDA),
pp. 30–37 (2013). https://doi.org/10.1109/CISDA.2013.6595424
4. Bähnemann, R., Lawrance, N., Chung, J., Pantic, M., Siegwart, R., Nieto, J.: Revisiting bous-
trophedon coverage path planning as a generalized traveling salesman problem. In: Ishigami,
G., Yoshida, K. (eds.) Field and Service Robotics. Springer Proceedings in Advanced
Robotics, vol. 16, pp. 277–290. Springer, Singapore (2021). ISBN: 978-981-15-9459-5.
https://doi.org/10.1007/978-981-15-9460-1_20
5. Cabreira, T.M., Brisolara, L.B., Paulo R, F.J.: Survey on coverage path planning with
unmanned aerial vehicles. Drones 3(1) (2019). https://doi.org/10.3390/drones3010004
6. Cabreira, T., Franco, C., Ferreira, P., Buttazzo, G.: Energy-aware spiral coverage path planning
for UAV photogrammetric applications. IEEE Robot. Autom. Lett. 3(4), 3662–3668 (2018).
https://doi.org/10.1109/LRA.2018.2854967
7. Cabreira, T., Ferreira, P., Di Franco, G. Buttazzo, C.: Grid-based coverage path planning with
minimum energy over irregular-shaped areas with UAVs. In: 2019 International Conference on
Unmanned Aircraft Systems (ICUAS), pp. 758–767, (2019). https://doi.org/10.1109/ICUAS.
2019.8797937
8. Cai, C., Ferrari, S.: Information-driven sensor path planning by approximate cell decomposi-
tion. IEEE Trans. Syst. Man Cybern. Part B 39(3), 672–689 (2009). https://doi.org/10.1109/
TSMCB.2008.2008561
9. Choset, H., Acar, E., Rizzi, A., Luntz, J.: Exact cellular decompositions in terms of critical
points of Morse functions. In: Proceedings ICRA. Millennium Conference. IEEE International
Conference on Robotics and Automation. Symposia Proceedings, vol. 3, pp. 2270–2277, USA
(2000). https://doi.org/10.1109/ROBOT.2000.846365
10. Choset, H., Pignon, P.: Coverage path planning: the boustrophedon cellular decomposition.
In: Field and Service Robotics, pp. 203–209 (1998). ISBN: 978-1-4471-1275-4
11. Choton, C., Prabhakar, P.: Optimal multi-robot coverage path planning for agricultural fields
using motion dynamics. In: 2023 IEEE International Conference on Robotics and Automa-
tion(ICRA), pp. 11817–11823. IEEE (2023). https://doi.org/10.1109/ICRA48891.2023.101
60265
12. Dou, L., et al.: Analysis of the world oil and gas exploration situation in 2021. Pet. Explor.
Dev. 49(5), 1195–1209 (2022). https://doi.org/10.1016/S1876-3804(22)60343-4
13. Duan, L., Wang, J. and Sun, Y.: Multi-robot online complete coverage based on collaboration.
In: Yang, S., Islam, S. (eds.) Web and Big Data. APWeb-WAIM 2022 International Workshops.
APWeb-WAIM 2022. CCIS, vol. 1784, pp. 219–231. Springer, Singapore (2023). https://doi.
org/10.1007/978-981-99-1354-1_19
14. Engelsons, D., Tiger, M., Heintz, F.: Coverage path planning in large-scale multi-floor urban
environments with applications to autonomous road sweeping, In: 2022 International Confer-
ence on Robotics and Automation (ICRA), pp. 3328–3334 (2022). “https://doi.org/10.1109/
ICRA46639.2022.9811941”
15. Franco, C., Buttazzo, G.: Energy-aware coverage path planning of UAVs. In: 2015 IEEE
International Conference on Autonomous Robot Systems and Competitions, pp. 111–117
(2015). https://doi.org/10.1109/ICARSC.2015.17
480 S. S. Maaji and D. Landa-Silva

16. Franco, C., Buttazzo, G.: Coverage path planning for UAVs photogrammetry with energy and
resolution constraints. J. Intell. Robot. Syst. 83(3), 445–462 (2016). https://doi.org/10.1007/
s10846-016-0348-x
17. Frew, E., Argrow, B., Houston, A., Weiss, C., Elston, J.: An energy-aware airborne dynamic
data-driven application system for persistent sampling and surveillance. Procedia Comput.
Sci. 18, 2008–2017 (2013). https://doi.org/10.1016/j.procs.2013.05.370
18. Galceran, E., Carreras, M.: A survey on coverage path planning for robotics. Robot. Auton.
Syst. 61(12), 1258–1276 (2013). https://doi.org/10.1016/j.robot.2013.09.004
19. Gallego, A., Pertusa, A., Gil, P., Fisher, R.: Detection of bodies in maritime rescue operations
using unmanned aerial vehicles with multispectral cameras. J. Field Robot. 36(4), 782–796
(2019). https://doi.org/10.1002/rob.21849
20. Ghaddar, A., Merei, A., Natalizio, E.: PPS: energy-aware grid-based coverage path planning
for UAVs using area partitioning in the presence of NFZs. Sensors 20(13) (2020). https://doi.
org/10.3390/s20133742
21. Guruprasad, K., Ranjitha, T.: CPC algorithm: exact area coverage by a mobile robot using
approximate cellular decomposition. Robotica 39(7), 1141–1162 (2021). https://doi.org/10.
1017/S026357472000096X
22. GvdHoorn, slam_gmapping, gmapping, 2019. http://wiki.ros.org/gmapping. Accessed 07
Mar 2023
23. Jensen-Nau, K., Hermans, R, Leang, T., Leang, K.: Near-optimal area-coverage path planning
of energy-constrained aerial robots with application in autonomous environmental monitor-
ing. IEEE Trans. Autom. Sci. Eng. 18(3), 1453–1468. IEEE (2020). https://doi.org/10.1109/
TASE.2020.301627
24. John, R., Current, Schilling, D.: The covering salesman problem. Transp. Sci. 23(3), 208–213
(1989). https://doi.org/10.1287/trsc.23.3.208
25. Jones, C.M.: The oil and gas industry must break the paradigm of the current exploration
model. J. Pet. Explor. Prod. Technol. 8(1), 131–142 (2017). https://doi.org/10.1007/s13202-
017-0395-2
26. Kuru, K., Ansell, D., Khan, W., Yetgin, H.: Analysis and optimization of unmanned aerial
vehicle swarms in logistics: an intelligent delivery platform. IEEE Access 7, 15804–15831
(2019). https://doi.org/10.1109/ACCESS.2019.2892716
27. Kyaw, P., Paing, A., Thu, T., Mohan, R., Le, A., x Veerajagadheswar, A.: Coverage path plan-
ning for decomposition reconfigurable grid-maps using deep reinforcement learning based
travelling salesman problem. IEEE Access 8, 225945–225956 (2020). https://doi.org/10.1109/
ACCESS.2020.3045027
28. Li, Y., Chen, H., Er, M., Wang, X.: Coverage path planning for UAVs based on enhanced
exact cellular decomposition method. Mechatronics 21(5), 876–885 (2011). https://doi.org/
10.1016/j.mechatronics.2010.10.009
29. Merz, M., Pedro, D., Skliros. V., Bergenhem, C. Himanka, M.: Autonomous UAS-based
agriculture applications: general overview and relevant European case studies, Drones. MDPI
6(5), 128 (2022). https://doi.org/10.3390/drones6050128
30. Mukhamediev, R., et al.: Coverage path planning optimization of heterogeneous UAVs group
for precision agriculture. IEEE Access 11, 5789–5803 (2023). https://doi.org/10.1109/ACC
ESS.2023.3235207
31. Murugan, D., Garg, A., Singh, D.: Development of an adaptive approach for precision agri-
culture monitoring with drone and satellite data. IEEE J. Sel. Top. Appl. Earth Obs. Remote
Sens. 10(12), 5322–5328 (2017). https://doi.org/10.1109/JSTARS.2017.2746185
32. Odonkor, P., Ball, Z., Chowdhury, S.: Distributed operation of collaborating unmanned aerial
vehicles for time-sensitive oil spill mapping, Swarm Evol. Comput. 46, 52–68, (2019). https://
doi.org/10.1016/j.swevo.2019.01.005
 UAV Path Planning for Area Coverage and Energy Consumption 481

33. Rafalamao, hector-quadrotor-noetic. https://github.com/RAFALAMAO/hector-quadrotor-

noetic. Accessed 7 Mar 2023
34. Rezazadeh, J., Moradi, M., Ismail, A., Dutkiewicz, E.: Impact of static trajectories on local-
ization in wireless sensor networks. Wirel. Netw. 21, 809–827. Springer (2015). https://doi.
org/10.1007/s11276-014-0821-z
35. Sarvesh, A., Carroll, A., Gopalswamy, S.: Reshaping local path planner. IEEE Robot. Autom.
Lett. 7(3), 6534–6541 (2022). https://doi.org/10.1109/LRA.2022.3174366
36. Shivgan, R., Dong, Z.: Energy-efficient drone coverage path planning using genetic algorithm.
In: 2020 IEEE 21st International Conference on High Performance Switching and Routing
(HPSR), pp. 1–6 (2020). https://doi.org/10.1109/HPSR48589.2020.9098989”
37. Shukla, A., Karki, H.: Application of robotics in onshore oil and gas industry—a review Part
I. Rob. Auton. Syst. 75, 490–507 (2016). https://doi.org/10.1016/j.robot.2015.09.012
38. Shukla, A., Karki, H.: New Frontiers in Oil and Gas Exploration, 1st edn. Springer, Cham
(2016). ISBN: 978-3-319-40122-5. https://doi.org/10.1007/978-3-319-40124-9
39. Strimel, G., Veloso, M., Coverage planning with finite resources. In: 2014 IEEE/RSJ Inter-
national Conference on Intelligent Robots and Systems, pp. 2950–2956 (2014). https://doi.
org/10.1109/IROS.2014.6942969
40. Tan, C., Mohd-Mokhtar, R., Arshad, M.: A comprehensive review of coverage path planning
in robotics using classical and heuristic algorithms, IEEE Access (2021). https://doi.org/10.
1109/ACCESS.2021.3108177
41. Torres, M., Pelta, D., Verdegay, J., x Torres, J.: Coverage path planning with unmanned aerial
vehicles for 3D terrain reconstruction, Expert Syst. Appl. 55, 441–451 (2016). https://doi.org/
10.1016/j.eswa.2016.02.007
42. Vadim, T., Alexander, P., Vasily, A., Dmitry, K.: Unmanned airborne magnetic survey tech-
nologies: present and future. In: Nurgaliev, D., Shcherbakov, V., Kosterov, A., Spassov, S.
(eds.) Recent Advances in Rock Magnetism, Environmental Magnetism and Paleomagnetism,
pp. 523–534. Springer Geophysics. Springer, Cham (2019). ISBN:978-3-319-90436-8.
https://doi.org/10.1007/978-3-319-90437-5_36
43. Valente, J., Sanz, D., Del Cerro, J., Barrientos, A., de Frutos, M.: Near-optimal coverage
trajectories for image mosaicing using a mini quadrotor over irregular-shaped fields. Precis.
Agric. 14, 115–132 (2013). https://doi.org/10.1007/s11119-012-9287-0
44. Wenceslao, E., Shaw-Cortez, Frew, E.: Efficient trajectory development for small unmanned
aircraft dynamic soaring applications. J. Guid. Control Dyn. 38(3), 519–523 (2015). https://
doi.org/10.2514/1.G000543s
45. Wu, C., Dai, C., Gong, X., Liu, Y., Wang, J., Gu, X., Wang, C.: Energy-efficient coverage
path planning for general terrain surfaces. IEEE Robot. Autom. Lett. 4(3), 2584–2591 (2019).
https://doi.org/10.1109/LRA.2019.2899920
46. Xie, J., Carrillo, L., Jin, L.: An integrated traveling salesman and coverage path planning
problem for unmanned aircraft systems. IEEE Control Syst. Lett. 3(1), 67–72 (2018). https://
doi.org/10.1109/LCSYS.2018.2851661
47. Xu, A., Viriyasuthee, C., Rekleitis, I.: Optimal complete terrain coverage using an unmanned
aerial vehicle. In: 2011 IEEE International Conference on Robotics and Automation,
pp. 2513–2519 (2011). https://doi.org/10.1109/ICRA.2011.5979707
48. Zheng, Y., Li, S., Xing, K., Zhang, X.: Unmanned aerial vehicles for magnetic surveys: a
review on platform selection and interference suppression. Drones 5(3), 93 (2021). https://
doi.org/10.3390/drones5030093
 Minimizing Peak Energy Demand
in Flexible Job Shops

Michael Eley(B)

Faculty of Engineering, TH Aschaffenburg, Aschaffenburg, Germany

[email protected]

Abstract. In the flexible job shop scheduling problem under consider-

ation a set of jobs consisting of different operations has to be scheduled.
The operations can be processed on different machines that have different
energy demands. The problem is formulated as a bi-objective optimiza-
tion problem. A solution approach based on variable neighborhood search
that minimizes peak energy demand for a given, minimal makespan is
developed. Test results are presented based on modified benchmark prob-
lems from the literature. The results are compared with an alternative
solution approach, where the problem is split up into the two subprob-
lems, i.e. makespan minimization and peak energy demand minimization,
that are solved sequentially to optimality. The results demonstrate that
the proposed algorithm outperforms the benchmark approach.

Keywords: scheduling · flexible job shop · peak energy demand ·

variable neighborhood search

1 Introduction

In the flexible job shop problem (FJSP) a given set of jobs has to be scheduled.
Each job consists of an ordered set of operations and for each operation a set
of different machines is available to carry out the required processing. In most
papers that address this classical version of the flexible job shop problem some
given economic or production criteria like cost, total (weighted) tardiness or
makespan have to be optimized while meeting some production constraints like
adhering to the sequence of operations or preventing pre-emption.
In green scheduling this problem formulation is extended by considering addi-
tional aspects like energy cost, energy demand or carbon footprint. These addi-
tional aspects can be integrated by additional constraints or by adopting the
objective function, for example in a multi-objective optimization approach. In this
paper a hierarchical two-objective approach is pursued for a flexible job shop prob-
lem. This approach was motivated by a case study carried out at a mechanical
engineering company. Besides the primary objective function that minimizes the
makespan an additional secondary objective function that minimizes peak energy
demand (peak load) is introduced. We assume that energy demand is determined

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023

J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 482–499, 2023.
https://doi.org/10.1007/978-3-031-43612-3_30
 Minimizing Peak Energy Demand in Flexible Job Shops 483

solely by electricity demand. Other forms of energy such as thermal energy or com-
pressed air are not considered. As for the classical FJSP, the operations can be car-
ried out on different machines, which, however, have different energy demands.
Obviously, the two objective functions are in conflict with one another. For
example, in order to minimize peak energy demand, it could be advantageous not
to use machines with high energy demand at the same time. However, this can
lead to an increased makespan if machines with high energy demand are not used
in parallel. In order to ensure high capacity utilization and adherence of delivery
dates, makespan minimization is often the most important objective of a company.
The motivation for considering peak load minimization as the secondary
objective is that in the German electricity market bulk consumers for energy
like industrial companies are mainly charged according to their peak energy
demand, as power providers have to guarantee that peak energy demand can
be satisfied. Thus, minimizing peak demand results in lower energy costs for a
company. Merely minimizing the peak energy demand makes no sense from an
economic point of view since the proportion of energy costs in relation to total
costs is low and, as described above, other criteria must also be considered in
production planning.
The remainder of the paper is organized as follows. The relevant literature
to the problem under consideration is reviewed in the next section. In the third
section the problem is formulated as a non-linear integer programming problem.
This problem formulation will later on be used to evaluate the performance
of a solution approach that is presented in section four. The main focus of the
proposed approach is to minimize the peak energy demand for a given makespan.
The problem of minimizing the makespan in a flexible job shop is not addressed
in this paper. We refer to the large number of articles in the literature addressing
that problem (e.g. [1,2]). Results on comprehensive tests on extended benchmark
problems taken from the literature are discussed in section five. The last section
summarizes the main results and puts them into a perspective.

2 Literature Review
Below we review the literature on the flexible job shop problem with respect to
papers dealing with energy aspects. In addition, we will also discuss the literature
on minimizing peak energy demand. As the number of papers addressing this
objective for the flexible job shop problem is rather limited, we will discuss peak
load minimization for different variants of scheduling problems. A comprehensive
survey on the literature on different types of green scheduling problems like single
or parallel machine problems, job shop problems as well as flow shop problems
can be found in the paper of [16].

2.1 Green Scheduling Approaches for Flexible Job Shops

The flexible job shop problem in which operations can be carried out on differ-
ent machines, that also have different energy demands seems somehow predes-
tined for the consideration of energy aspects. Thus, several authors have recently
484 M. Eley

addressed green scheduling for different variants of the flexible job shop problem.
Surveys on flexible job shop approaches can be found in the papers of [8] or of
[23]. The latter one also includes green scheduling approaches. Besides makespan
optimization or minimization of total tardiness in most approaches minimizing
total energy demand (or total energy cost) is used as the secondary objective in
a multi-objective optimization approach.
Many papers on green scheduling assume that energy demand at a given
time depends on the state of the machine during that time. Therefore, states
like idle, setup, loading, warm-up or processing are considered. For most of these
states, the energy demand is known in advance and constant over time for each
operation. Only total energy demand during idle times is not known in advance as
the length of idle times depends on the result of the scheduling. Thus, in many
approaches the objective of minimizing total energy demand can be reduced
to minimizing idle energy demand or even idle times (cf. the approaches of
[10,15,18,19,21,24,34,39,42–44]).
In some approaches the production speed for the operations is an additional
decision variable, for example in the articles of [42] or [15]. With a higher produc-
tion speed processing times for operations can be reduced, resulting for example
in a lower makespan, but energy demand is increased.

2.2 Approaches Minimizing Peak Energy Demand

Peak energy demand aspects for different types of scheduling problems have
recently been addressed by several authors. In these approaches, restricting peak
energy demand can be imposed by a constraint that limits a given maximum
value (threshold) for each single period or by the objective function that mini-
mizes peak energy demand.
[4,27] as well as [36] addressed problems from the process industry. Single
and parallel machine scheduling problems were discussed in the articles of [37]
and [3] as well as [40], respectively. [14] as well as [25] addressed the general
flow shop problem, whereas [6] and [12,13] focused on flexible flow shops. Peak
energy demand minimization for job shop problems were discussed by several
authors, e.g. [32,41,45] (reactive scheduling), [22] and [26].
As mentioned in Sect. 1, when peak energy demand aspects are considered,
the energy demand of a machine must be modeled more precisely. The same
applies if energy prices vary over time, e.g. for time-of-use tariffs. Therefore, [38]
proposed using energy load profiles instead of average energy demand values
particularly when considering peak energy demand. Furthermore, [35] recom-
mended a discretization of the energy demand as it avoids the modelling of
complex functions and thus reduces the problem’s complexity.
Thus, it is interesting to see that in most papers energy demand solely
depends on the factors like the operation, the machine, the machine load, the
state of the machine or the chosen speed of the machine, but it is kept con-
stant over the processing time (cf. [3,4,6,12–14,25,26,29,30,37,40]). Only a few
authors consider fluctuating energy demand over time.
 Minimizing Peak Energy Demand in Flexible Job Shops 485

[27] were the first authors who examined variable energy demand (steam
energy) in their simulation model for the process industry. [36] consider varying
energy demand during processing a job for a batch production problem in the
process industry. [45] and [22] divided each operation of their job shop schedul-
ing problem into two parts. Energy demand is constant for both parts, but the
energy demand for the first part is higher than for the second part. [32] model
energy demand of machines by an integral of the energy during processing times.
The authors address a reactive job shop scheduling problem for a flexible man-
ufacturing system. [41] approximate the energy profile of a machine for a job
shop problem by a mathematical power series.
In summary, it can be stated that the flexible job shop problem with different
energy-oriented objectives and restrictions has been dealt with in the literature
in recent years. The main focus, however, was on minimizing total energy con-
sumption. There is also extensive literature on the topic of minimizing peak
energy load, with various problems being discussed. In contrast, the intersection
of the two subject areas green scheduling for the flexible job shop problem and
minimization of peak energy load is hardly considered. Furthermore, if any, other
optimization models than the problem presented in Sect. 1 are discussed. This
is surprising insofar as it is precisely in this case that peak load minimization
is promising if the machines that can carry out the processing of the operations
have different energy demands, and if the energy demand also changes over the
course of the processing.
Below we will introduce a new solution approach to the flexible job shop prob-
lem where we model the operation-dependent energy demand of the machines
by step functions. In the objective functions we will consider minimization of the
makespan and peak energy demand. We will implement a hierarchical approach
minimizing in the first stage the time-based objective function and in the second
stage peak energy demand. As far as the author is aware, this question has so
far not been investigated for the flexible job shop problem. Similar optimization
approaches have only been suggested for different scheduling problems. [6] pro-
posed this hierarchical approach for the flow shop problem; [3] addressed it to
the parallel machine problem.

3 Problem Formulation
The problem under consideration consists of n jobs that have to be scheduled on
a set of m machines. We will pursue a hierarchical two-objective optimization
approach. The primary objective function minimizes the makespan. Minimiza-
tion of peak energy demand serves as a secondary objective. Each job j consists
of hj operations. For each operation Oj,h a sub-set of the machines is permitted
to carry out the operations. We assume that each operation can be processed
by only one machine at a time and that each machine can only carry out one
operation at a time. Pre-emption is not allowed.
Before stating the problem formally, we introduce some notation in Table
1. In the problem formulation used here, the use of time-indexed variables is
486 M. Eley

dispensed with. The variables used instead lead to a non-linear constraint on the
one hand. On the other hand, they increase the comprehensibility and make it
clear how the sub-problems are separated from one another within the framework
of the hierarchical solution approach. It should also be noted that the use of
time-indexed variables does not improve the solvability of the problem since the
number of variables increases with the number of periods. For example, in the
test cases used in Sect. 5, this leads to a few million variables.
Table 1. Parameters and decision variables used

N number of jobs
M number of machines
hj number of operations for job j
l maximal number of operations assigned to a machine
T length of planning horizon
L a large
 number
1 if operation Oj,h can be performed on machine i,
ai,j,h =
0 otherwise
pi,j,h processing time of operation Oj,h if performed on machine i
Ei,j,h (d) energy demand (wattage) for processing operation Oj,h on
machine i after elapsing of d = 1, ..., pi,j,h periods
Γ makespan
Π peak energy demand
tj,h start time of the processing of operation Oj,h
T Mi,k start
 working time of the k-th operation of machine i
1 if machine i is selected for carrying out operation Oj,h ,
yi,j,h =
0 otherwise

1 if operation Oj,h is the k-th operation on machine i,
xi,j,h,k =
0 otherwise

1 if r = T Mi,k ,
wi,k,r =
0 otherwise

min Γ (1)

min Π (2)
The two objective functions (1) and (2) minimize the makespan and the peak
energy demand, respectively, subject to the following constraints.

M
yi,j,h = 1 j = 1, ..., N ; h = 1, ..., hj (3)
i=1

yi,j,h ≤ ai,j,h i = 1, ..., M ; j = 1, ..., N ; h = 1, ..., hj (4)

 Minimizing Peak Energy Demand in Flexible Job Shops 487


N 
hj
xi,j,h,k ≤ 1 i = 1, ..., M ; k = 1, ..., l (5)
j=1 h=1


N 
hj

N 
hj
xi,j,h,k ≥ xi,j,h,k+1 i = 1, ..., M ; k = 1, ..., l − 1 (6)
j=1 h=1 j=1 h=1


l
xi,j,h,k = yi,j,h i = 1, ..., M ; j = 1, ..., N ; h = 1, ..., hj (7)
k=1

Constraint (3) guarantees that each operation Oj,h is assigned to exactly one
machine. This assignment can only take place if the corresponding operation can
be performed on the considered machine i. This is imposed by constraint (4).
The following constraint (5) limits the number of operations that can be assigned
on the k-th position of a machine to at most one. By constraint (6) the positions
of a machine are assigned sequentially. Note, that this constraint is not required
as one can always find an optimal solution that satisfies this constraint. Never-
theless, this constraint limits the number of possible assignments of operations
to machine positions. Finally, each operation assigned to one machine must also
be assigned to one position in the sequence of operations on that machine; cf.
constraint (7).


M
tj,h + yi,j,h · pi,j,h ≤ tj,h+1 j = 1, ..., N ; h = 1, ..., hj − 1 (8)
i=1

T Mi,k ≤ tj,h + (1 − xi,j,h,k ) · L

(9)
i = 1, ..., M ; j = 1, ..., N ; h = 1, ..., hj ; k = 1, ..., l
tj,h ≤ T Mi,k + (1 − xi,j,h,k ) · L
(10)
i = 1, ..., M ; j = 1, ..., N ; h = 1, ..., hj ; k = 1, ..., l


N 
hj
T Mi,k + pi,j,h · xi,j,h,k ≤ T Mi,k+1
j=1 h=1
(11)
i = 1, ..., M ; k = 1, ..., l − 1
Constraint (8) ensures that the available start time of operation Oj,h+1 is
greater than or equal to the completion time of the previous operation of that
job j. The following two constraints (9) and (10) guarantee the consistency of
the time variables of operation Oj,h , i.e. tj,h , with the time variables of machine
i, i.e. T Mi,k . Both constraints are only restrictive in the event that the operation
is carried out on the considered machine as the k-th operation, i.e. if xi,j,h,k = 1.
Finally, constraint (11) guarantees that the time span between the starting times
of two consecutive operations on a machine are large enough to carry out the
operation. Again, this constraint is only restrictive if operation Oj,h is the k-th
operation on machine i.
488 M. Eley


M 
N 
hj

l p
i,j,h

xi,j,h,k · wi,k,t−d+1 · Ei,j,h (d) ≤ Π

i=1 j=1 h=1 k=1 d=1
(12)
t = 1, ..., T


M
tj,hj + yi,j,hj · pi,j,hj ≤ Γ j = 1, ..., N (13)
i=1

The non-linear constraint (12) imposes an upper bound on the energy

demand for every single period of the planning horizon. If operation Oj,h is
assigned to the k-th position in the sequence of machine i, i.e. xi,j,h,k = 1, and
the starting time of that operation was in period t − d + 1, i.e. wj,k,t−d+1 = 1,
then an amount of Ei,j,h (d) must be added to the total energy demand in period
t. In addition, the makespan is restricted by the delivery dates of the jobs; c.f.
constraint (13).

tj,h ≥ 0 j = 1, ..., N ; h = 1, ..., hj (14)

T Mi,k ≥ 0 i = 1, ..., M ; k = 1, ..., l (15)

yi,j,h ∈ {0, 1} i = 1, ..., M ; j = 1, ..., N ; h = 1, ..., hj (16)

xi,j,h,k ∈ {0, 1}
(17)
i = 1, ..., M ; j = 1, ..., N ; h = 1, ..., hj ; k = 1, ..., l

wi,k,r ∈ {0, 1} i = 1, ..., M ; k = 1, ..., l; r = 0, ..., T − 1 (18)

The remaining technical constraints (14) to (18) restrict the values of the
decision variables.
Hereinafter we will focus on optimizing the problem with the secondary objec-
tive function, i.e. problem (2) to (18). Minimizing the first objective function
corresponds to the classical FJSP, which is already an NP-hard combinatorial
optimization problem. Several approaches have been proposed in the literature
to approximately solve this problem. Instead of proposing a new solution algo-
rithm for makespan minimization for the FJSP we refer to the comprehensive
literature (cf. [8,23]).
The focus of our investigation is to examine the potential of peak energy
demand minimization under a strict makespan restriction. This corresponds to
the hierarchical optimization approach, where minimizing the makespan is the
main objective.
 Minimizing Peak Energy Demand in Flexible Job Shops 489

4 Solution Approach
The solution approach pursued in this paper is based on the variable neigh-
borhood search (VNS). VNS, originally proposed by [17,28], is a meta-
heuristic method for solving combinatorial optimizations problems. A pseudo-
code description of the solution approach is shown in Algorithm 1.
In a first step an initial solution is generated by solving a classical FJSP with
respect to makespan minimization. For test runs carried out to assess the quality
of the proposed approach we used known optimal solutions for test cases from
the literature. It goes without saying that the proposed approach also works
when the makespan is not set to its minimal value.
This starting solution is set as the incumbent solution s. Within the proposed
approach a solution s is represented by a string that is a permutation of all
operations supplemented by a feasible machine μ and a starting time tj,h for
N
each operation Oj,h , i.e. si (j, h, μ, tj,h ) for i = 1... j=1 hj . One should keep in
mind that in order to reduce peak energy demand it is not sufficient to consider
only active schedules where all operations are scheduled as early as possible.
Therefore, the starting time of each operation must also be specified in the
solution string. The transformation of a solution string into a solution is carried
out by the local search procedures (cf. Sect. 4.2).
It should be noted that only permutations in which the specified sequence of
operations is adhered to for all jobs represent feasible solutions.
In the VNS approach a predetermined number of iterations imax is carried
out, each of them consisting of a shaking procedure and a local search step. In
every iteration, a random solution s is generated in the current neighborhood
N Hg (s) of s (shaking). Therefore, a total of gmax neighborhoods are available.
It starts in the first iteration with the first neighborhood, i.e. g := 1. Whenever
a solution s is replaced by a new solution the value of g is reset to 1; otherwise g
is increased by 1 in the subsequent iteration, i.e. g := g + 1. If g reaches a value
bigger than a predetermined value gmax then g is set to 1.
A subsequent local search procedure is applied to s transforming this solution
to s . If s is better than s with respect to a specific objective function f (·) or
s meets another acceptance criterion then s replaces s. At the end of the VNS
approach, the best solution sbest found during all iterations is returned as the
final solution.
The implemented neighborhoods, the local search procedures as well as the
acceptance criteria are presented in detail below.

4.1 Implemented Neighborhoods

Neighborhoods are used to generate new, potentially better solutions. As pro-

posed by [33] two different types of neighborhoods are used within the approach.
For both types the parameter n represents the neighborhood width.
The first type of neighborhood (NH1) changes the assignment of operations
to machines. Each time this neighborhood is invoked, n operations are randomly
490 M. Eley

Algorithm 1. variable neighborhood search

1: generate initial solution s;
2: sbest ⇐ s, g ⇐ 1 and iteration ⇐ 1;
3: while iteration ≤ imax do
4: shaking, i.e. rand. s ⇒ s , s ∈ N Hg (s);
5: apply local search to s ⇒ s ;
6: if f (s ) < f (s) or other criteria then
7: s ⇐ s and g ⇐ 0;
8: end if ;
9: if s is feasible and f (s ) < f (sbest ) then
10: sbest ⇐ s ;
11: end if ;
12: g ⇐ (g modulo gmax ) + 1 and iteration ⇐ iteration + 1;
13: end while;
14: return sbest ;

selected from the set of operations that can be processed on at least two different
machines and assigned to new machines. Thereby, all potential operations and all
selectable machines for an operation have the same probability of being selected.
After changing the machine assignment of n operations in the solution string, the
starting times of all operations must be updated. This is done by the procedure
presented in the next section.
The second type of neighborhood (NH2) changes the sequence of operations
in the solution string. Whenever this neighborhood is invoked, up to n oper-
ations are shifted one after another. An operation Oj,h is selected by roulette
wheel selection, where probabilities for selecting an operation are calculated
using values π(j, h). When shifting operations to earlier or later periods in a
classical FJSP while minimizing makespan there is always a direct relationship
to the objective function value as the makespan is also measured in time peri-
ods. Unfortunately, there is no direct relationship between time and objective
function value when minimizing peak energy demand is pursued. For this reason,
we have to employ other mechanisms that detect which operations are difficult
to plan and which are likely to cause high peak energy demands. This selection
mechanism based on the π-values has been shown in test runs to be superior to
neighborhoods in which the operations were selected according to their maximum
energy demand. Variants in which operations were preferably selected that took
place at peak times were also significantly worse. The π(j, h) values are updated
by the local search procedures (cf. explanation below). The new position of the
operation to be shifted is selected randomly following a uniform distribution for
all positions that do not violate the processing sequence of the corresponding
job. After shifting one operation the local search procedure is started in order
to update the starting time of all operations in the solution string.
Because of the high computing effort for the generation of solutions, we have
restricted ourselves to inserting individual operations and have neglected more
complex operators in which several operations swap positions. In addition, we
 Minimizing Peak Energy Demand in Flexible Job Shops 491

had to restrict the neighborhood width n to at most 3. It turned out that larger
values for n did not improve the solutions any more while just increasing required
CPU times. Thus, the value of gmax was set to 6. The neighborhoods are applied
in the following order: NH2/1, NH1/1, NH2/2, NH1/2, NH2/3, NH1/3, where
the number after the slash represents the value of the neighborhood width n.

4.2 Local Search Procedures

A solution string can be transformed into a solution (Gantt chart) by scheduling
the operations one by one starting with the first operation in the string. The
optimal solution for minimizing the makespan always includes an active schedule
in which all operations are scheduled as early as possible. To generate solutions
with low makespan we limit ourselves to generating only active schedules when
decoding a solution string into a solution. Therefore, each operation is scheduled
as early as possible on the designated machine without any delays. It should
be noted that an operation of a job can only be scheduled after the preceding
operation of the job is completed. Additionally, scheduling an operation on a
machine must not delay the beginning of another operation already scheduled
on that machine.
In order to minimize the secondary objective function (peak energy demand)
the following three different local search procedures have been implemented.

– Smoothing: Delaying the start of operations

– Shaving: Reducing energy demand in peak periods
– Updating machines: Changing the assignment of operations to machines

They are invoked in the displayed order. It should be noted that the first two
procedures do not change the sequence of operations on the machines. With all
three local search procedures shifts are carried out in such a way that makespan
of a solution s generated by shaking does not increase. As a result, optimization
is no longer limited to only active schedules.

Smoothing. Starting from a given solution string where the sequence of oper-
ations is already given for every machine, this procedure tries to reduce peak
energy demand by delaying the beginning of operations. Remember that every
operation has been scheduled as early as possible when decoding a solution
string. If, for example, a machine is not used after the completion of an opera-
tion, the start of processing that operation may be delayed in order to reduce
peak energy demand. Figure 1 shows an illustrative example of this situation
with three jobs. Each job consists of up to two operations that are denoted by
Oj,h . The Gantt chart on the left shows the initial situation. Energy demand
for the three jobs and total energy demand for all periods is shown in the table
above the Gantt chart. The Gantt chart on the right shows the situation when
the beginning of the first operation of job 1 on machine 2, i.e. operation O1,1 , is
delayed by three periods. This shift results in a reduction of peak energy demand
from 14 to 10.
492 M. Eley

Fig. 1. Reducing peak energy demand from 14 (left) to 10 (right) by postponing the
start of the first operation of job 1, i.e. operation O1,1 , on machine 2 by three periods

The procedure tries to shift operations one by one beginning with the opera-
tion that determines the makespan, i.e. the operation with the latest completion
time. Thereafter, in every iteration of the smoothing procedure the operation
with the latest completion time that has not yet been considered is identified and
the beginning of that operation is delayed as long as makespan is not increased
or the next operation scheduled on that machine had to be delayed.

Shaving. This local search procedure considers the whole planning horizon.
Starting from the incumbent solution s the procedure tries to shift opera-
tions that are processed during the peak period in order to reduce peak energy
demand. At the same time, however, makespan must not be increased.

Let tpeak denote the peak period of a solution s . Energy demand for every
period t of the planning horizon and therefore the peak period can be calculated
by the left hand side of Eq. (12) as for a given solution s the values of decision
variables xi,j,h,k and wj,k,t−d+1 are known.
Within the procedure the set P contains all operations that are carried out
in solution s during the peak period. In addition, let us define by M (j, h) the
machine an operation Oj,h is assigned to. The operations from set P are exam-
ined one by one and the beginning of processing that operation is gradually
increased by one period. Note that when shifting an operation, it might also be
necessary to shift the succeeding operations belonging to the same job as well as
operations of other jobs that are assigned in a later period to the same machine.
These resulting shifts are also considered when examining the effects on peak
energy demand. Whenever shifting an operation causes the completion of an
operation (either that one or a different one) to be postponed beyond the end
of the specified makespan, the examination of further shifting this operation is
stopped and the starting period of that operation is set to its original period.
A second stopping criterion is a reduction in peak energy demand. In this case
the incumbent solution is replaced by the newly generated solution, the value of
the peak period as well as the set P are updated and the procedure starts again
 Minimizing Peak Energy Demand in Flexible Job Shops 493

in trying to postpone the beginning of operations that are processed during the
newly generated peak period. Eventually, the local search procedure terminates
when all operations from set P have been checked without finding any further
improvement.
During the execution of the shaving procedure values π(i, j) for an operation
Oj,h that are used in the second type of neighborhood are updated. At the begin-
ning of the VNS approach all values of π(i, j) are initialized with 1. Whenever
an increase of the peak energy demand is observed during a shift in the shav-
ings procedure then we identify the operations that are processed in the peak
period. For each of these operations it is compared how high the energy demand
was at the peak period before and after the shift. In this way, we can identify
the operations that contribute to the peak energy demand. For all contributing
operations we increase the value π(i, j) of that operation by 1. Thus, in the long
run we can identify operations that are somehow critical and therefore should
be considered during the shaking phase.

Updating Machines. A third local search procedure tries to reduce machine

load by changing the assignment of operations to machines. Therefore, all oper-
ations are examined with respect to the feasible machines for each operation.
Additionally, the smoothing and the shaving local search procedures are invoked
for the newly generated assignment. Whenever an improvement in the objective
function is found, the new assignment is realized and the search starts again
from the beginning. The procedure terminates when all operations have been
checked and no further improvement could be found.

4.3 Acceptance Criteria

To decide whether a newly generated solution s should be accepted, an objec-

tive function f () that consists of two summands is employed. The first summand
represents the peak energy demand of the solution. The second summand penal-
izes makespan violations of operations. Therefore, whenever an operation is not
completed within the given makespan a penalty is incurred that is calculated
by multiplying the length of the makespan violation with a penalty factor. This
factor is initialized by 1 at the beginning. Whenever a new solution is accepted
that has makespan violations then the penalty factor is increased by multiply-
ing it with a factor randomly chosen from the interval [1.05, 1.1]. In contrast,
when the newly accepted solution has no makespan violations then the penalty
factor is decreased by dividing it again by a factor randomly chosen from the
same interval. This type of acceptance criteria was first proposed by [9] for the
dial-a-ride problem and has proven to be very effective in finding new feasible
solutions for highly restricted optimization problems.
The newly generated solution s replaces the incumbent solution s if f (s )
has a smaller value than f (s). Additionally, also deteriorating solutions s may
become incumbent solutions with probability exp(−[f (s ) − f (sbest )]/τ ). The
initial temperature τ is then set such that if f (s ) is 0.5% larger than f (sbest ),
494 M. Eley

s is accepted with a probability of 0.2. Thereafter, τ is linearly decreased to

0 until the maximum number of iterations has been reached. Carrying out test
runs it could be observed that after a strong reduction in the objective function
value the temperature seemed to be too high, resulting in the acceptance of
too many deteriorating solutions. Therefore, the temperature τ is additionally
recalculated whenever a new best solution has been found based on this new
solution. The temperature is again linearly decreased to 0 over the remaining
number of periods.

5 Computational Results
The use of the proposed approach as part of the above-mentioned case study at
the mechanical engineering company led to a reduction in peak energy demand
of 23.6%. However, this result says little about the performance of the proposed
approach in general. In order to enable a broader evaluation of the proposed app-
roach, modified test cases from the literature were also used (cf. [31]). Therefore,
the test cases were extended by adding randomly generated energy demands for
all operations on their eligible machines. The profiles were generated in such a
way that the difference between the machine with the lowest and the highest
total energy demand for each operation is at most 20%.

5.1 Benchmark Approach

In order to assess the performance of the proposed approach from Sect. 4, we
implemented the following benchmark approach. Therefore, the problem (2) to
(18) is decomposed into two subproblems that are solved sequentially.
In the first subproblem the makespan is minimized subject to the classical
FJSP constraints without considering energy demand aspects, i.e. this problem
consists of the objective function (1) and constraints (3) to (11) and (13) to (18).
We used the optimal solution as the starting point for minimizing the second
subproblem that minimizes the objective function (2) subject to the constraints
(3) to (18). With the solution of the first subproblem, the remaining problem
can be simplified considerably.
This second subproblem could be solved for the test cases under considera-
tion using the commercial solver CPLEX 12. In the following, this approach in
which both subproblems are solved sequentially is referred to as the benchmark
approach.
This benchmark approach corresponds to the procedure in many indus-
trial companies where optimization is initially carried out subject to minimiz-
ing makespan. Only in the second step is an attempt made to minimize peak
energy demand. Compared to the benchmark approach, the approach proposed
in Sect. 4 possesses additional flexibility as operations might be rescheduled,
processed in different orders on the machines or reassigned to other machines as
long as makespan is not increased and the sequence of operations for each job is
respected. The test results are intended to show whether or not this additional
flexibility can be used to reduce peak energy demand.
 Minimizing Peak Energy Demand in Flexible Job Shops 495

Table 2. Aggregated results for the test cases of [5, 7, 11, 20]

test case class # problems # VNS better # ties Min. Avg. Max.
[7] 21 18 0 0.719 0.814 0.902
[11] 7 4 3 0.801 0.841 0.892
[5] 9 8 1 0.719 0.814 0.902
edata set of [20] 51 39 1 0.744 0.863 0.972
rdata set of [20] 45 45 0 0.353 0.398 0.475
vdata set of [20] 40 40 0 0.214 0.242 0.289

5.2 Test Results

The proposed approach of Sect. 4 was coded in Delphi. Each test case was solved
20 times. The number of iterations was set to 20,000 in order to limit the com-
puting effort.
Table 2 summarizes the aggregated results for the different classes of test
cases. The values in columns two through four indicate the number of considered
test cases. In addition, it is indicated how often the proposed approach achieved
a better solution than the benchmark approach and how often both approaches
led to the same results, respectively. In the remaining three columns the best,
average and worst results are displayed that were obtained by applying the
proposed solution approach. The given values represent the share in the objective
function value achieved. For example, if the benchmark approach achieves an
average objective function value of 1000 for all test problems of this class and
the value in the column min is 0.95 then the average of the best objective function
values obtained by the proposed approach was 950. Thus, values in columns 3
to 5 that are below 1 represent improvements obtained by the proposed solution
approach over the benchmark.
The results can be summarized as follows. The proposed algorithm clearly
outperformed the benchmark approach for most of the 173 test cases. With five
test cases, namely three test cases of Dauzère-Pérès and Paulli and one test cases
of Brandimarte and Hurink et al. respectively, both approaches achieved the
same solution. For 14 test cases a better solution than the benchmark approach
could not be found in any of the twenty test runs - these were three test cases of
Chambers and Barnes and eleven test cases from the edata set of Hurink et al.
In the remaining 159 test cases, however, almost consistently better results were
achieved compared to the comparison method, in the majority of cases with a
very clear difference based on the best or worst individual values as well as on
the average values for all 20 test runs.
Of the test cases of [20] the edata test cases were the most difficult ones to
solve. For the rdata and even more for the vdata test cases more machines are
available to be chosen for the operations. Thus, this kind of flexibility can help
to reduce peak energy demand considerably.
496 M. Eley

Finally, further experiments with the proposed approach were carried out to
investigate whether and if so to what extent a relaxation of the strict limitation
of the makespan has an impact on peak energy demand. Therefore, the optimal
value of makespan, i.e. the value of Γ in Eq. (13), was increased in 1% steps
from 1% to 15%. Test runs were carried out for 30 selected test cases from the
edata set of Hurink et al., i.e. abz5 and abz6, la01 to la20, mt06, mt10 and
mt20 and orb1 to orb10. Each test case was solved ten times for each step.
With a larger makespan peak energy demand can be reduced considerably. The
maximum decrease in peak energy demand is almost 20% if at the same time
makespan is increased by 10%. Further reductions in peak energy demand could
not be realized even with higher increases in makespan.

6 Conclusion

In this paper an approach to minimize peak energy demand in a flexible job shop
considering narrow specifications for makespan was presented. In particular, the
last results from Sect. 5.2 show that peak energy demand can be reduced if
the makespan is increased. Obviously, there is a tradeoff between these two
objectives. But, as the results of the test cases also show, there is even a strong
potential for minimizing peak energy demand when the makespan is kept at
its lowest value. The greater the number of alternative machines (with different
energy requirements) available for the processing of individual operations, the
greater this potential is.
Another finding from the computational tests is that a greater reduction in
peak energy demand is possible if both objectives are considered at the same
time. This becomes clear by comparing the results of the proposed algorithm
with the benchmark approach. In the latter approach, the optimal solution
for the objective of minimizing the makespan was first determined and then,
based on this solution, peak energy demand was minimized in the second step.
Although optimal solutions were generated for both subproblems, the approach
presented in Sect. 4 outperformed the benchmark considerably. This is because
the VNS approach can change the assignment of operations to machines, as well
as the order of operations on the machines, provided that the makespan does not
increase. And this additional flexibility is sufficient to reduce the peak energy
demand significantly.
Apart from very energy-intensive industries, with today’s energy prices, many
companies do not view minimizing energy costs as a top priority. The proposed
optimization approach could be of interest to these companies, especially if they
want to cut down peak energy for environmental reasons or if energy prices rise.
Another application would be in the context of a demand side management in
which the local search procedures used in the article can also be used to smooth
the demand.
Improvements would also be possible on the algorithmic side. Due to the very
high computing times that are due to the optimization of the energy requirement,
an obvious possibility would be to consider the parallelization of the calculation.
 Minimizing Peak Energy Demand in Flexible Job Shops 497

Acknowledgement. The research was funded by the Bavarian State Ministry for
Economic Affairs, Regional Development and Energy as part of the R&D program
“Information and communication technology” of the Free State of Bavaria, project
title: KAnIS: Cooperative Autonomous Intralogistics Systems.

References
1. Amirghasemi, M.: An effective parallel evolutionary metaheuristic with its appli-
cation to three optimization problems. Appl. Intell. 53, 5887–5909 (2023). https://
doi.org/10.1007/s10489-022-03599-w
2. Amirghasemi, M., Zamani, R.: An effective asexual genetic algorithm for solving
the job shop scheduling problem. Comput. Ind. Eng. 83, 123–138 (2015). https://
doi.org/10.1016/j.cie.2015.02.011, https://www.sciencedirect.com/science/article/
pii/S0360835215000686
3. Artigues, C., Lopez, P., Haı̈t, A.: The energy scheduling problem: industrial case-
study and constraint propagation techniques. Int. J. Prod. Econ. 143(1), 13–23
(2013). https://doi.org/10.1016/j.ijpe.2010.09.030
4. Babu, C.A., Ashok, S.: Peak load management in electrolytic process industries.
IEEE Trans. Power Syst. 23, 399–405 (2008). https://doi.org/10.1109/TPWRS.
2008.920732
5. Brandimarte, P.: Routing and scheduling in a flexible job shop by tabu search.
Ann. Oper. Res. 41(3), 157–183 (1993). https://doi.org/10.1007/BF02023073
6. Bruzzone, A., Anghinolfi, D., Paolucci, M., Tonelli, F.: Energy-aware scheduling for
improving manufacturing process sustainability: a mathematical model for flexible
flow shops. CIRP Ann. 61(1), 459–462 (2012). https://doi.org/10.1016/j.cirp.2012.
03.084
7. Chambers, J., Barnes, J.: Flexible job shop scheduling by tabu search. Technical
report, The University of Texas, Austin, TX, Technical Report Series ORP96-09,
Graduate Program in Operations Research and Industrial Engineering (1996)
8. Chaudhry, I., Khan, A.: A research survey: review of flexible job shop scheduling
techniques. Int. Trans. Oper. Res. 23, 551–591 (2016). https://doi.org/10.1111/
itor.12199
9. Cordeau, J.F., Laporte, G.: A tabu search heuristic for the static multi-vehicle
dial-a-ride problem. Transp. Res. Part B: Methodol. 37, 579–594 (2003). https://
doi.org/10.1016/S0191-2615(02)00045-0
10. Dai, M., Tang, D., Giret, A., Salido, M.: Multi-objective optimization for energy-
efficient flexible job shop scheduling problem with transportation constraints.
Robot. Comput.-Integr. Manuf. 59, 143–157 (2019). https://doi.org/10.1016/j.
rcim.2019.04.006
11. Dauzère-Pérès, S., Paulli, J.: An integrated approach for modeling and solving the
general multiprocessor job-shop scheduling problem using tabu search. Ann. Oper.
Res. 70, 281–306 (1997). https://doi.org/10.1023/A:1018930406487
12. Fang, K., Uhan, N., Zhao, F., Sutherland, J.: A new approach to scheduling in
manufacturing for power consumption and carbon footprint reduction. J. Manuf.
Syst. 30(4), 234–240 (2011). https://doi.org/10.1016/j.jmsy.2011.08.004
13. Fang, K., Uhan, N., Zhao, F., Sutherland, J.: A new shop scheduling approach
in support of sustainable manufacturing. In: Hesselbach, J., Herrmann, C. (eds.)
Glocalized Solutions for Sustainability in Manufacturing, pp. 305–310. Springer,
Berlin (2011). https://doi.org/10.1007/978-3-642-19692-8 53
498 M. Eley

14. Fang, K., Uhan, N., Zhao, F., Sutherland, J.: Flow shop scheduling with peak
power consumption constraints. Ann. Oper. Res. 206(1), 115–145 (2013). https://
doi.org/10.1007/s10479-012-1294-z
15. Gong, G., Deng, Q., Gong, X., Liu, W., Ren, Q.: A new double flexible job-shop
scheduling problem integrating processing time, green production, and human fac-
tor indicators. J. Clean. Prod. 174, 560–576 (2018). https://doi.org/10.1016/j.
jclepro.2017.10.188
16. Grahm, C., Denz, F., Dirr, M., Tuma, A.: Energy efficient scheduling in manu-
facturing companies: a review and research framework. Eur. J. Oper. Res. 248,
744–757 (2016). https://doi.org/10.1016/j.ejor.2015.07.017
17. Hansen, P., Mladenovic, N., Moreno-Perez, J.: Variable neighbourhood search:
methods and applications. Ann. Oper. Res. 175(1), 367–407 (2010). https://doi.
org/10.1007/s10288-008-0089-1
18. He, Y., Li, Y., Wu, T., Sutherland, J.: An energy-responsive optimization method
for machine tool selection and operation sequence in flexible machining job shops.
J. Clean. Prod. 87, 245–254 (2015). https://doi.org/10.1016/j.jclepro.2014.10.006
19. He, Y., Liu, B., Zhang, X., Gao, H., Liu, X.: A modeling method of task-oriented
energy consumption for machining manufacturing system. J. Clean. Prod. 23, 167–
174 (2012). https://doi.org/10.1016/j.jclepro.2011.10.033
20. Hurink, J., Jurisch, B., Thole, M.: Tabu search for the job-shop scheduling problem
with multi-purpose machines. OR Spektrum 15(4), 205–215 (1994). https://doi.
org/10.1007/BF01719451
21. Jin, L., Zhang, C., Fei, X.: Realizing energy savings in integrated process planning
and scheduling. Processes 7, 120 (2019). https://doi.org/10.3390/pr7030120
22. Kemmoe, S., Lamy, D., Tchernev, N.: Job-shop like manufacturing system with
variable power threshold and operations with power requirements. Int. J. Prod.
Res. 55, 1–22 (2017). https://doi.org/10.1080/00207543.2017.1321801
23. Li, X., Xie, J., Peng, K., Li, H., Gao, L.: A review on flexible job shop scheduling.
IET Collaborative Intell. Manuf. 1(3), 63–77 (2019). https://doi.org/10.1049/iet-
cim.2018.0009
24. Liu, Q., Chekem, F., Zhan, M., Shao, X., Ying, B., Sutherland, J.: A hybrid fruit fly
algorithm for solving flexible job-shop scheduling to reduce manufacturing carbon
footprint. J. Clean. Prod. 168, 668–678 (2017). https://doi.org/10.1016/j.jclepro.
2017.09.037
25. Luo, H., Du, B., Huang, G., Chen, H., Li, X.: Hybrid flow shop scheduling con-
sidering machine electricity consumption cost. Int. J. Prod. Econ. 146, 423–439
(2013). https://doi.org/10.1016/j.ijpe.2013.01.028
26. Masmoudi, O., Delorme, X., Gianessi, P.: Job-shop scheduling problem with energy
consideration. Int. J. Prod. Econ. 216, 12–22 (2019). https://doi.org/10.1016/j.
ijpe.2019.03.021
27. Mignon, D., Hermia, J.: Peak utility load reduction in batch processes operated
periodically and under uncertainty. Comput. Chem. Eng. 20(3), 249–263 (1996).
https://doi.org/10.1016/0098-1354(95)00041-0
28. Mladenovic, N., Hansen, P.: A parallel iterated tabu search heuristic for vehicle
routing problems. Comput. Oper. Res. 24(11), 1097–1100 (1997). https://doi.org/
10.1016/S0305-0548(97)00031-2
29. Mokhtari, H., Hasani, A.: An energy-efficient multi-objective optimization for flex-
ible job-shop scheduling problem. Comput. Chem. Eng. 104, 339–352 (2017).
https://doi.org/10.1016/j.compchemeng.2017.05.004
 Minimizing Peak Energy Demand in Flexible Job Shops 499

30. Moon, J.Y., Park, J.: Smart production scheduling with time-dependent and
machine-dependent electricity cost by considering distributed energy resources and
energy storage. Int. J. Prod. Res. 52, 3922–3939 (2014). https://doi.org/10.1080/
00207543.2013.860251
31. OR library (2020). http://people.brunel.ac.uk/mastjjb/jeb/info.html
32. Pach, C., Berger, T., Sallez, Y., Bonte, T., Adam, E., Trentesaux, D.: Reactive and
energy-aware scheduling of flexible manufacturing systems using potential fields.
Comput. Ind. 65, 434–448 (2014). https://doi.org/10.1016/j.compind.2013.11.008
33. Pezzella, F., Morganti, G., Ciaschetti, G.: A genetic algorithm for the flex-
ible job-shop scheduling problem. Comput. Oper. Res. 35(10), 3202–3212
(2008). https://doi.org/10.1016/j.cor.2007.02.014, https://www.sciencedirect.
com/science/article/pii/S0305054807000524
34. Piroozfard, H., Wong, K.Y., Wong, W.P.: Minimizing total carbon footprint and
total late work criterion in flexible job shop scheduling by using an improved multi-
objective genetic algorithm. Res. Conserv. Recycl. 128, 267–283 (2018). https://
doi.org/10.1016/j.resconrec.2016.12.001
35. Rager, M., Gahm, C., Denz, F.: Energy-oriented scheduling based on evolutionary
algorithms. Comput. Oper. Res. 54, 218–231 (2015). https://doi.org/10.1016/j.
cor.2014.05.002
36. Seid, E., Majozi, T.: Heat integration in multipurpose batch plants using a
robust scheduling framework. Energy 71, 302–320 (2014). https://doi.org/10.1016/
j.energy.2014.04.058
37. Shrouf, F., Ordieres-Meré, J., Garcı́a-Sánchez, A., Ortega-Mier, M.: Optimizing the
production scheduling of a single machine to minimize total energy consumption
costs. J. Clean. Prod. 67, 197–207 (2014). https://doi.org/10.1016/j.jclepro.2013.
12.024
38. Suwa, H., Samukawa, T.: A new framework of energy-efficient manufacturing sys-
tems based on energy load profiles. Procedia CIRP 41, 313–317 (2016). https://
doi.org/10.1016/j.procir.2015.10.011
39. Wang, H., Jiang, Z., Wang, Y., Zhang, H., Wang, Y.: A two-stage optimization
method for energy-saving flexible job-shop scheduling based on energy dynamic
characterization. J. Clean. Prod. 188, 575–588 (2018). https://doi.org/10.1016/j.
jclepro.2018.03.254
40. Wang, Y.C., Wang, M.J., Lin, S.C.: Selection of cutting conditions for power con-
strained parallel machine scheduling. Robot. Comput.-Integr. Manuf. 43, 105–110
(2017). https://doi.org/10.1016/j.rcim.2015.10.010
41. Weinert, N., Chiotellis, S., Seliger, G.: Methodology for planning and operating
energy-efficient production systems. CIRP Ann. 60(1), 41–44 (2011). https://doi.
org/10.1016/j.cirp.2011.03.015
42. Wu, X., Sun, Y.: A green scheduling algorithm for flexible job shop with energy-
saving measures. J. Clean. Prod. 172, 3249–3264 (2018). https://doi.org/10.1016/
j.jclepro.2017.10.342
43. Zhang, L., Tang, Q., Wu, Z., Wang, F.: Mathematical modeling and evolutionary
generation of rule sets for energy-efficient flexible job shops. Energy 138, 210–227
(2017). https://doi.org/10.1016/j.energy.2017.07.005
44. Zhang, Z., Wu, L., Peng, T., Jia, S.: An improved scheduling approach for mini-
mizing total energy consumption and makespan in a flexible job shop environment.
Sustainability 11(1) (2019). https://doi.org/10.3390/su11010179
45. Zheng, H.Y., Wang, L.: Reduction of carbon emissions and project makespan by a
pareto-based estimation of distribution algorithm. Int. J. Prod. Econ. 164, 421–432
(2014). https://doi.org/10.1016/j.ijpe.2014.12.010
 Carbon-Aware Mine Planning
with a Novel Multi-objective Framework

Nurul Asyikeen Binte Azhar1,2(B) , Aldy Gunawan1 , Shih-Fen Cheng1 ,

and Erwin Leonardi2
1
School of Computing and Information Systems, Singapore Management University,
80 Stamford Road, 178902 Singapore, Singapore
{nurula.a.2020,aldygunawan,sfcheng}@smu.edu.sg
2
Rio Tinto Ltd., Tower 3, 12 Marina Boulevard, Marina Bay Financial Centre,
018982 Singapore, Singapore
{asyikeen.azhar,erwin.leonardi}@riointo.com

Abstract. The logistical complication of long-term mine planning

involves deciding the sequential extraction of materials from the mine pit
and their subsequent processing steps based on geological, geometrical,
and resource constraints. The net present value (NPV) of profit over the
mine’s lifespan usually forms the sole objective for this problem, which is
considered as the NP-hard precedence-constrained production scheduling
problem (PCPSP) as well. However, increased pressure for more sustain-
able and carbon-aware industries also calls for environmental indicators
to be considered. In this paper, we enhance the generic PCPSP formula-
tion into a multi-objective optimization (MOO) problem whereby carbon
cost forms an additional objective. We apply the Non-Dominated Sort-
ing Genetic Algorithm II (NSGA-II) to this formulation and experiment
with variants to the solution generation. Our tailored application of the
NSGA-II using a set of real-world inspired datasets can form an approx-
imated Pareto front for planners to observe stipulated annual carbon
emission targets. It also displays that tailored variants of the NSGA-II
can produce diverse solutions that are close to the true Pareto front.

Keywords: Precedence-constraint production scheduling · Resource

capacity optimization · Multi-objective evolutionary algorithm ·
Sustainability

1 Introduction
The five phases of mining present a number of logistical challenges (Fig. 1).
During the planning, implementation, and production phases, logistics manage-
ment must forecast, plan, and schedule tasks from strategic, tactical, and opera-
tional angles. During the planning phase, environmental concerns are addressed
through the environmental impact assessment (EIA) and reclamation plan.
Supported by Enterprise Singapore.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 500–516, 2023.
https://doi.org/10.1007/978-3-031-43612-3_31
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 501

These are submitted with the mining plan to the respective government author-
ities before the implementation phase [7]. As the mining phases mature, the
mining plans, EIA, and reclamation plans are updated and can become progres-
sively detailed. The reclamation plan is updated and reviewed periodically in
certain jurisdictions such as Western Australia. However, regulatory rigor and
intra-governmental coordination vary across jurisdictions and may be improved,
such as integrating the EIA and reclamation plan processes [12]. These can be
classified as the strategic perspective for mine planning.

Fig. 1. Summary of activities for each mine phase and their influences on costs.

Similarly, the tactical and operational perspectives of mine planning are

updated periodically and have the potential to incorporate environmental con-
cerns. Planning from the tactical perspective [13] over the whole mine lifespan
(decades) has been tackled as the precedence-constrained production scheduling
problem (PCPSP) and acknowledged as NP-hard. The logistical challenge in
the PCPSP involves the movement of extracted raw materials from the mine
pit through the broad sequence of processing and refinement steps to obtain
the desired products. To support these activities, there are various operational
facilities for processing (e.g., crushing and grinding), refining (e.g., hydrometal-
lurgy), storage (i.e., stockpile), and waste (e.g., tailings pond and dump). These
facilities have their own set of resources and corresponding capacity constraints.
When modeling mines, planners discretize the buried materials into three-
dimensional blocks to decide the sequence of extraction. Each block would have
its own set of preceding blocks that require prior extraction due to the geology
and geometry unique to each mine site. This is known as precedence constraints
and is illustrated in Fig. 2. There are then diverse resource constraints for each
block at the downstream processing and refining facilities. For the efficient and
effective movement of extracted materials, these sequence and processing deci-
sions are considered holistically with the approximate net present value (NPV)
of profit throughout decades of the mine’s lifespan.
Barring profit, environmental considerations and the push for net zero car-
bon emissions are reverberating within the mining industry due to its exten-
sive processing and operations scale. Hence, the monitoring of carbon emissions
502 N. A. B. Azhar et al.

Fig. 2. Block extraction sequence example from year to year following precedence and
resource constraints for MineLib’s Newman1 instance [3]

throughout the value chain of the PCPSP is increasingly necessary and ties in
with schemes such as the emission trading systems (ETS) and emission taxes. In
this paper, we boost previous research on the PCPSP by enhancing the generic
PCPSP formulation [6] with carbon cost. This allows planners to consider the
NPV of carbon cost alongside profit.
Instead of solely maximizing the NPV of profit in the generic PCPSP formu-
lation, we convert it into a multi-objective optimization (MOO) problem with
another objective that minimizes the NPV of carbon cost. Next, we apply the
NSGA-II [5] whereby components of the algorithm - initial solution generation,
crossover and mutation - are tailored to the PCPSP. As far as we know, earlier
works of evolutionary algorithms (EA) to the PCPSP did not propose similar
frameworks. In fact, details on the initial solution generation, crossover, and
mutation are quite lacking. Furthermore, earlier works for the PCPSP employed
EA such as genetic algorithm [1] and differential evolution [8] for a single objec-
tive formulation problem whereas we use NSGA-II for an MOO problem. We
then test variants of our approach with a real-world mine dataset from an oper-
ating mine and benchmark instances from Minelib [6]. It turns out that variants
to the solution generation can improve results. Furthermore, we demonstrate how
our approach can produce approximated Pareto fronts for planners to observe
carbon emissions targets when deciding on mining plans.

2 Related Work
The angles of strategic, tactical, and operational issues are commonly used for
research on the logistics of mining. Lately, qualitative and quantitative research
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 503

on mine planning that incorporates sustainability elements has increased. How-

ever, quantitative research based on operations research, artificial intelligence,
and machine learning trails behind qualitative research [21]. Furthermore,
research that incorporates carbon emissions from a tactical angle remains sparse.
From the strategic angle, one of the research covers carbon emissions indi-
rectly by increasing efficiency and thus minimizing haulage costs. The main focus
of Rimélé, Dimitrakopoulos, and Gamache [14] is to minimize land disturbance
by optimizing the order of ore extraction. To achieve this, they propose an in-pit
waste disposal approach, which involves depositing unprofitable extracted ores
in available areas within the pit instead of moving them to temporary dumps.
Meanwhile, Xu et al. [20] model carbon emission cost directly as one of the
undesirable environmental outputs. In their multi-objective approach, they aim
to minimize undesirable environmental outputs while simultaneously maximiz-
ing both the NPV of profit and social benefits. It focuses solely on the pit limit,
which determines whether or not to extract the ores while ignoring the sequenc-
ing and processing decisions thereafter.
From the operational angle, the objective is to explicitly decrease expenses
associated with carbon emissions in operational facilities and transportation net-
works. Valderrama et al. [18] utilize a mixed integer programming (MIP) model
to analyze carbon emissions from inter-facility transportation and operating
facilities. Similarly, Attari and Torkayesh [2] employ a multi-objective MIP to
examine carbon emissions in transportation between facilities and customers. In
comparison, Canales-Bustos, Santibañez-González, and Candia-Véjar [4] develop
a multi-objective hybrid particle swarm optimization algorithm to minimize
investment costs, transportation costs, deviations between product quality and
goals, and carbon emissions from facilities and vehicles. These operational angles
ignore the tactical decisions of extraction and sequencing of blocks. Lastly, Wang
et al. [19] indirectly model carbon emissions by comparing resource efficiency and
the NPV of profit with an NSGA-II. Their formulation also pre-defines extrac-
tion decisions and focuses only on the processing of extracted ores.
Research from the tactical angle directly models carbon emission cost when
examining trade-offs between profit and sustainability. Azhar et al. [3] consider
extraction, sequencing, and processing decisions of ores as per the PCPSP. They
enhance the generic PCPSP formulation with an additional constraint of carbon
emission cost [20] to produce an approximated Pareto front.
Our research is based on the tactical angle of the PCPSP, focusing on deci-
sions of block extraction, extraction sequence, and its processing steps [3]. We
also directly model the trade-off between the NPV of profit and the carbon
emission cost by adopting the carbon costing framework by Xu et al. [20]. How-
ever, we differ from Azhar et al. [3] by including the carbon emission cost as an
additional objective function instead of a constraint.

3 The PCPSP Definition and Formulation

The PCPSP determines the entire mining process, including extraction and
processing decisions for valuable mineral ores, and provides investors with an
504 N. A. B. Azhar et al.

estimate of a mine’s value [7]. The mine consists of several components, such
as the pit, dump, stockpiles, processing plants, and heavy machinery. In the
pit, mineral ore deposits are divided into blocks of the same size for modeling
purposes. Each block has a unique value and a set of precedence constraints
based on geology, which affects the overall extraction sequence over time and
how the ore is processed, as illustrated in Fig. 2. After extraction, the block is
transported to processing facilities where it undergoes various treatments, such
as crushing, grinding, and screening, to reduce it according to the requirements.
Then, the material is refined to improve its quality and obtain various desired
end products. Unfortunately, this value chain consumes significant raw materials
(e.g., energy, water, gases) and generates harmful by-products (e.g., water and
air pollution, chemical waste).

3.1 Carbon Costing Formulation

Our main aim is to augment the PCPSP formulation with carbon costing. To
measure carbon emissions, we use the metric by Xu et al. [20], which defines
the cost of carbon emissions Ci,e from energy consumption. This cost reflects the
amount of carbon dioxide that is absorbed during ore excavation, processing, and
refining. The formula includes two key quantities: the amount of ore extracted
from the pit and sent for further processing Qi,o , and the amount of waste
material extracted and treated Qi,w . These quantities are multiplied by the
energy consumed using coal to extract one unit tonne of material (either ore
or waste) from the pit em , and the energy consumed to process one unit tonne
of ore ep . These values are then multiplied by the carbon factor of coal fc , the
conversion coefficient of carbon dioxide from carbon fa , and the absorption cost
of carbon dioxide Cc .

(Qi,o + Qi,w )em + Qi,o ep

C= fc fa Cc (1)
1000

3.2 Enhanced Multi-objective PCPSP Formulation

The PCPSP is a problem in scheduling mining activities that aims to maxi-
mize the NPV of profit while satisfying several requirements related to mineral
ore grade, equipment availability, and processing plant capacity [6]. This requires
expertise in multiple domains such as geology, chemistry, engineering, economics,
and customer relations. Firstly, geologists provide information on the ore com-
ponents and grades based on multiple drill samples and the structure of the
surrounding materials. Secondly, mining engineers use this information to assess
the structure, methods, and equipment needed to access the ore. Next, geologists
and chemists determine the type of processing and refining required for differ-
ent ores, which can result in varying products. Then, economists estimate the
economic value of each block based on demand and supply worldwide. Lastly,
customer relations evaluate the demand of current and potential customers. All
these considerations make mine scheduling a complex task.
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 505

Mine scheduling research usually relies on real-world case studies because

mining operations are unique and shaped by geo-metallurgical factors [13]. How-
ever, this approach leads to solution techniques that cannot be directly compared
with others. The MineLib library [6] provides a set of generalized mathematical
formulations and instances for three problem variants, including the PCPSP,
which is the most complex problem. By adopting the PCPSP formulation, our
work enables other researchers to build upon it.
The generic formulation for the PCPSP [6] defines B as the set of blocks, Bb as
the subset of predecessors for block b ∈ B, and D as the set of destinations. The
profit p̃bdt is obtained by extracting a block b and processing it at a destination d
during a specific period, using qbdr units of operational resource r ∈ R. A binary
decision variable xbt is used to indicate whether block b is extracted during period
t. A continuous decision variable ybdt represents portions of block b delegated to
destination d during period t. The augmented multi-objective PCPSP has two
objective functions. The first objective maximizes the NPV of profit for periods
T , while the second objective minimizes the carbon emission cost pegged to the
extraction and processing of mineral ores.
The first objective function derives the profit p̃bdt for a given period t ∈ T
pbd
with (1+α) t , where α denotes the discount factor. The estimated value of a block

is determined by geologists using its ore composition and grade. The summation
of the NPV of profit throughout the mine’s lifespan emphasizes the significance
of extracting blocks with higher value earlier.

(Objective function 1) Z1 = max p̃bdt ybdt (2)
b∈B d∈D t∈T

The second objective function calculates the discounted carbon cost c̃bdrt
from using operational resource r ∈ R to extract or process block b at desti-
nation d ∈ D during period t ∈ T . It is derived using (1+α)
cbdr
t . Practically, this

function can be crucial within the framework of carbon credit trading, a market-
based instrument aimed at reducing carbon dioxide emissions. Under this system,
economies that exceed their allocated emissions can purchase credits from those
that have reduced their emissions below their carbon emission permits.

 
(Objective function 2) Z2 = min c̃bdrt ybdt (3)
b∈B d∈D r∈R t∈T

Constraint (4) sets conditions for the order in which blocks can be extracted,
and it applies to all blocks and periods. It states that block b must be extracted
in the same or an earlier period than block b, as b is a predecessor of b. This
constraint is decided by mining engineers based on the materials surrounding
the ore, including sand, silt, and clay, as well as the ore’s type and composition.
The latter is determined by geologists.
 
xbτ ≤ xb τ ∀ b ∈ B, b ∈ Bb , t ∈ T (4)
τ ≤t τ ≤t
506 N. A. B. Azhar et al.

Constraint (5) specifies that a block should be completely dispatched to one

or multiple destinations if it is mined. Otherwise, it should not be sent to any
destination. The selection of destination is influenced by the ore’s grade and
composition, as well as the demands of customers.

xbt = ybdt ∀ b ∈ B, t ∈ T (5)
d∈D

Constraint (6) limits block extraction to only once during the mine’s lifespan.

xbt ≤ 1 ∀ b ∈ B (6)
t∈T

Constraint (7) guarantees that for each time period t, the use of every opera-
tional resource r is within the limits of minimum Rrt and maximum R̄rt . These
resources are managed by mining engineers and technicians, and they include
diggers, haulage trucks, grinders, and processing plants.

Rrt ≤ qbdr ybdt ≤ R̄rt ∀ r ∈ R, t ∈ T (7)
b∈B d∈D

Constraint (8) represents side constraints with lower bound a and upper
bound ā. It can represent various mining situations, including grade constraints.

a ≤ Ay ≤ ā (8)
Finally, constraints (9) and (10) reflect the range of values.

xbt ∈ {0, 1} ∀b ∈ B, t ∈ T , (9)

ybdt ∈ [0, 1] ∀b ∈ B, d ∈ D, t ∈ T . (10)

4 NSGA-II for the Augmented Multi-objective PCPSP

Diverse techniques are available for MOO problems with their own advantages
and disadvantages [11]. Due to the NP-hard nature of the PCPSP, exact methods
such as the MIP and constraint programming [15], are found to be efficient for
smaller instances but become intractable as the instance size increases. Hence,
alternatives have to be explored. One of them is the extension of the genetic
algorithm framework. Genetic algorithms use a population of randomly gener-
ated solutions that are evaluated for improvement at each iteration, making it
possible to converge on the entire Pareto set in one run. Multi-objective evolu-
tionary algorithms (MOEAs) belong to the class of genetic algorithms used for
MOO problems. They utilize additional advanced methods to maintain a varied
population of Pareto optimal solutions throughout the iterations. The MOEAs
are differentiated by their fitness assignment, diversity mechanism, elitism, and
the use of external population [9].
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 507

We utilize NSGA-II [5] as it remains a popular choice due to its effective

mechanisms of non-domination sorting, crowding distance sorting, and elitism,
which contribute to a diverse Pareto optimal set and accelerated convergence.
Its wide application across other industries, including medicine [10] and man-
ufacturing [17], further demonstrates its strength. Our implementation of the
NSGA-II for the augmented multi-objective PCPSP is summarized in Fig. 3.

Fig. 3. Overview of NSGA-II implementation.

4.1 Initial Population of Solutions with Cone Sets

The repair of precedence constraint violation is computationally intensive and

hence, it should be prevented as much as possible. To do so, we compute the
complete set of preceding blocks that need to be extracted before the current
block of interest with a directed acyclic graph (DAG). The shape of this set,
together with the block of interest, can be described as a cone set, illustrated in
Fig. 4. For each cone set, we compute the resources required, the profit, and the
earliest period the cone set can be extracted based on the cumulative resource
capacity for each period t. The apportionment of blocks to destinations at each
time period, ybdt is determined by the profit associated with each destination;
we utilize argmax and softmax (details in Sect. 5). These components form the
heuristic for the initial solution, shown in Algorithm 1.
508 N. A. B. Azhar et al.

Fig. 4. Cross-sectional view of mine pit when excavating different target blocks.

Algorithm 1 produces a population of solutions Ωj where each solution j is

the set of blocks extracted at each time period xbt . Each solution is processed
progressively through each time period t ∈ T (line 6). At each time period, blocks
are pre-selected based on the earliest period of extraction (line 9) with allowance
for resource violation (line 15) using an upper bound resource multiplier ρ. The
resource multiplier is randomly generated where ρ ∈ [0, 1] (line 5). This allows
variants amongst solutions in the population set Ωj . Then, the resources required
for the solution are checked against the upper bound of resources for the time
period t. If there are no violations, a local improvement heuristic is run (line
20). Otherwise, a repair operator is invoked (line 22).
Both the improvement heuristic and the repair operator rely on finding fringe
blocks. These are blocks situated at the periphery of the solution that may
prevent major disruptions to the current solution when added or removed. If
there is no resource violation, the priority is to add more profitable blocks.
Otherwise, blocks that consume the least resources are added. Conversely when
removing blocks, blocks that are least profitable are prioritized. Finally, the
solutions are ranked based on the two objective functions with non-domination
sorting and crowding distance [5].

4.2 Reproduction

The reproduction of offspring from the initial solution or parent solution consists
of the binary tournament operator, mutation, and crossover. Once offspring solu-
tions are produced, they are again ranked based on the two objective functions
with non-domination sorting, and crowding distance. This reproduction step is
run till the maximum number of generations (i.e., the termination threshold).
The binary tournament operator randomly selects two individuals (or solu-
tions) from a population and chooses the best (based on the rank function of
NSGA-II) for the next generation. This procedure is repeated until the desired
number of individuals for the next generation is obtained. The set of parent
solutions from this step then leads to the crossover operator.
Crossover facilitates the creation of new offspring solutions by combining
genetic material from two parent solutions. The process involves selecting a
crossover point in the parents’ genetic code and exchanging genetic informa-
tion beyond that point to generate two new offspring solutions. The crossover
rate parameter η determines whether this step occurs for the parent pair. A high
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 509

Algorithm 1. Generation of the population of initial solutions.

Input: Block model B, destinations D, predecessor edges E, time periods T , resource types R,
resource capacity required per block at each destination and resource type qbdr , resource bounds for
each resource type R̄rt , profit of block when sent to destination p̃bd , apportionment of the block to
each destination ybdt upper bound resource multiplier ρ, population size Ω
Output: Population of solutions Ωj where each solution is block extracted at each time period
xbt
1: G ← ConstructDAG(B, E)
2: θ ← ConeSetComputaions(G, D, T , R, qbdr , R̄rt , p̃bd
3: Ωj ← ∅
4: for j in Ω do
5: ρ ← RAN DOM (0, 1)
6: for t in T do
7: B̂ ← ∅
8: for b in θ do
9: if b earliest extraction period = t then
10: B̂ ← B̂ ∪ b
11: end if
12: end for
13: for b in B̂ do
14: Î ←∅ 
15: if b∈B̂ d∈D qbdr ybdt ≤ R̄rt ∗ ρ then
16: Î ← Î ∪ b
17: end if
18: endfor 
19: if b∈Î d∈D qbdr ybdt < R̄rt then
20: xbtj  ← ImproveSolution(G,
 qbdr , p̃bd , Î)
21: else if b∈Î d∈D qbdr ybdt > R̄rt then
22: xbtj ← RepairResourceV iolation(G, qbdr , p̃bd , Î)
23: end if
24: end for
25: Ωj ← Ωj ∪ xbtω
26: end for
27: return Ωj

crossover rate reflects a higher chance that crossover occurs. Due to the multi-
period structure of the PCPSP, we design an interdependent-period single-point
crossover that accounts for the entire mine lifespan. It combines two parent solu-
tions by exchanging portions of their periodic components. We use a crossover
 cut
ζ ∈ [0, 1], drawn randomly, to determine the crossover index using ζ ∗ t∈T xbt .
This index can fall in any time period. Figure 5 illustrates a crossover between
two feasible solutions. The last step in this illustration is an immediate fix to
prevent block extraction across multiple periods.
Subsequently, mutation introduces a small random alteration to the genetic
code of an individual. In the mutation step, we choose blocks for extraction
(i.e., xb t = 1) with no precedence blocks. The current time period t is mutated
randomly to another time period t ∈ {T \ t }. The mutation step occurs for all
parent pairs, but we used the mutation rate γ to select the percentage of blocks
with no precedence |{b̂}| to mutate. The subset of blocks with no precedence is
selected uniformly using SAM P LE(RAN DOM (0, γ) ∗ |{b̂}|, {b̂}).
Solutions may become infeasible after crossover and mutation. The crossover
step may result in precedence violation whereas the mutation step may result
in resource violation. Hence, the repair operator is invoked. It uses the pre-
computed cone sets that are akin to the initial solution generation.
510 N. A. B. Azhar et al.

Fig. 5. Crossover between two feasible solutions.

5 Experiments
We utilize a small-sized real-world inspired dataset from an operating copper
and gold mine, Wilma1. It was adapted for the generic formulation and mocked
up. We complemented it with benchmark instances from MineLib - Newman1
and Kd. Newman1 is small-sized whilst Kd is medium-sized (Kd is a copper
mine in North America). Table 1 informs the number of blocks |B|, precedence
|Bb |, time periods |T |, destinations |D| and operational resource constraints |R|.

Table 1. Key characteristics of the dataset.

Name Block Precedence Periods Destinations Resources

Newman1 1,060 3,922 6 2 2
Wilma1 1,960 7,112 4 3 3
Kd 14,153 219,778 12 2 2

5.1 Experimental Setup

There are two NSGA-II variants - based on the argmax and softmax functions –
for the portions of block b delegated to destination d during period t, ybdt ∈ [0, 1].
This directly affects Objective Function 1. The first variant, using the argmax
function, apportions a block fully to the destination with the most profit. Mean-
while, the softmax function variant, apportions a block across all destinations
with a multiplier λ on the profit of that block when sent to each destination,
pbd :

eλpbd
softmax = D (11)
λpbd
d=1 e
We run the two variants three times, each with 100 solutions in a population
and generations ranging from eight to ten. From the six populations of solution
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 511

sets, we derive the true Pareto front of non-dominated solutions. Then, the solu-
tion sets from each population are compared using the ratio of non-dominated
individuals (RNI) [16], distance, and diversity metrics [5]. Firstly, the RNI met-
ric yields the proportion of best-known solutions φ (i.e., solutions that form part
of the Pareto front) that exist in a population Ωj with N solutions:

|φj |
Ratio of non-dominated individuals (RNI) = (12)
N
Secondly, the distance metric assesses how closely a solution j in population
Ωj converges to the true Pareto front. Initially, the metric calculates the mini-
mum Euclidean distance between a solution from population Ωj and all solutions
k from the true front Ωk . This is then averaged across all N solutions. A value
close to zero is desired.
N
j=1 min djk
Distance metric = (13)
N
Finally, the diversity metric evaluates the even spread of solutions over the
true Pareto front using the Euclidean distance. It considers the distance between
successive solutions di , the average distance d¯ of all di , distances between extreme
solutions of the true Pareto-optimal front dk , and distances between extreme
solutions in a population set dj . A value close to one indicates better diversity.
N −1 ¯
dk + dj + i=1 |di − d|
Diversity metric = (14)
dk + dj + (N − 1)d¯
The NSGA-II model for the augmented multi-objective PCPSP was built
using Python. The algorithm was executed on a Linux operating system with
3.5 GHz 3rd generation Intel Xeon Scalable processor, 128 vCPUs and 128 Gb
memory. The initial solution was produced with the upper bound resource mul-
tiplier ρ of 1.2. In the reproduction, the mutation γ and crossover η rates were
set to 0.2 and 0.6, respectively. Lastly, the softmax multiplier λ was 4.

5.2 Results

The performance metrics of RNI, distance, and diversity for the argmax and
softmax variants are summarized in Table 2. Overall for Newman1 and Wilma1
instances, the argmax variants, compared to the softmax variants, generate sets
of solutions that are closer to the true Pareto front (distance metric), but with
less assortment (diversity metric) and lesser individual solutions that form part
of the true Pareto front (RNI). For Newman1, the argmax and softmax average
distance metrics are 0.290 and 0.304, respectively, while for Wilma1, their values
are 0.056 (argmax) and 0.074 (softmax). Next, the diversity metric averages
0.753 (argmax) and 0.830 (softmax) for Newman1, and 0.863 (argmax) and
1.017 (softmax) for Wilma1. Finally, the RNI averages 0.003 (argmax) and 0.007
(softmax) for Newman1, and 0.007 (argmax) and 0.037 (softmax) for Wilma1.
512 N. A. B. Azhar et al.

Table 2. Performance metrics of experiment variants across datasets.

Experiment Softmax Argmax

Generation RNI Distance Diversity RNI Distance Diversity
Newman1
8 0 0.264 0.862 0 0.254 0.780
9 0 0.332 0.820 0 0.295 0.767
10 0.02 0.317 0.807 0.01 0.321 0.713
Average 0.007 0.304 0.830 0.003 0.290 0.753
Wilma1
8 0.09 0.082 1.037 0.01 0.039 0.803
9 0.01 0.078 0.977 0 0.070 0.875
10 0.01 0.063 1.038 0.01 0.058 0.911
Average 0.037 0.074 1.017 0.007 0.056 0.863
Kd
8 0 0.332 0.749 0.03 0.322 0.946
9 0 0.251 0.684 0 0.324 0.905
10 0 0.236 0.685 0.01 0.333 0.925
Average 0 0.273 0.706 0.01 0.326 0.925

Fig. 6. Newman1 Pareto front across generations for two NSGA-II variants.

Meanwhile for the medium-sized instance of Kd, these observations between

the argmax and softmax variants are reversed. The softmax variants, compared
to the argmax variants, generate sets of solutions that are closer to the true
Pareto front (distance metric), but with less assortment (diversity metric) and
lesser individual solutions that form part of the true Pareto front (RNI). The
average values are found in Table 2.
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 513

Fig. 7. Wilma1 Pareto front across generations for two NSGA-II variants.

Fig. 8. Kd Pareto front across generations for two NSGA-II variants.

A Pareto optimal front is also produced for each population variant. This
visual aid allows mine planners to appreciate the trade-off between the NPV of
profit and carbon cost for decision making. Figure 6, Fig. 7 and Fig. 8 display
the Pareto front of the population variants across generations for Newman1,
Wilma1, and Kd respectively. The computation time for each population variant
was less than an hour for Newman1. However, when the model is applied to larger
datasets, the computation time increases. Hence, the traversal of the search space
may be improved for this framework to be applied to much larger instances.
514 N. A. B. Azhar et al.

6 Conclusion
In this paper, we implement the NSGA-II model to balance the NPV of profit
with carbon emissions when managing the movement of ores from the mine
pit to the production facilities. Our framework for the open pit mine addresses
sustainability concerns by augmenting the generic PCPSP into a multi-objective
problem that can be catered to diverse environmental concerns. It is applied to
real-world instances of an operating mine and MineLib that have been extended
for carbon emission considerations. Our approach demonstrates the effective
formation of Pareto fronts with profit and carbon cost axes so that mine planners
can consider both aspects concurrently.
Previous work has used the generic PCPSP for carbon emissions trade-off
but through an additional constraint. To the best of our knowledge, this is the
first reformulation into a MOO problem. Future work can focus on improving the
computation time to extend this proof of concept to larger datasets. Additionally,
other MOEAs can be evaluated against the NSGA-II model and the objective
functions can be further expanded for other environmental concerns.

Acknowledgements. This work is supported by Enterprise Singapore under the

grant 20-IPPII-T-001-B-1 and Rio Tinto Ltd.

References
1. Alipour, A., khodaiari, A.A., Jafari, A., Tavakkoli-Moghaddam, R.: A genetic algo-
rithm approach for open-pit mine production scheduling. Int. J. Mining Geo Eng.
51(1), 47–52 (2017). https://doi.org/10.22059/ijmge.2017.62152
2. Attari, M.Y.N., Torkayesh, A.E.: Developing benders decomposition algorithm for
a green supply chain network of mine industry: case of Iranian mine industry. Oper.
Res. Perspect. 5, 371–382 (2018). https://doi.org/10.1016/j.orp.2018.11.002
3. Azhar, N.A.B., Gunawan, A., Cheng, S.F., Leonardi, E.: A carbon-aware planning
framework for production scheduling in mining. In: Computational Logistics: 13th
International Conference, ICCL 2022, Barcelona, Spain, 21–23 September 2022,
Proceedings, pp. 441–456. Springer, Heidelberg (2022). https://doi.org/10.1007/
978-3-031-16579-5 30
4. Canales-Bustos, L., Santibañez-González, E., Candia-Véjar, A.: A multi-objective
optimization model for the design of an effective decarbonized supply chain in
mining. Int. J. Prod. Econ. 193, 449–464 (2017). https://doi.org/10.1016/j.ijpe.
2017.08.012
5. Deb, K., Agrawal, S., Pratap, A., Meyarivan, T.: A fast elitist non-dominated sort-
ing genetic algorithm for multi-objective optimization: NSGA-II. In: Schoenauer,
M., et al. (eds.) PPSN 2000. LNCS, vol. 1917, pp. 849–858. Springer, Heidelberg
(2000). https://doi.org/10.1007/3-540-45356-3 83
6. Espinoza, D., Goycoolea, M., Moreno, E., Newman, A.: MineLib: a library of open
pit mining problems. Ann. Oper. Res. 206(1), 93–114 (2012). https://doi.org/10.
1007/s10479-012-1258-3
7. Hustrulid, W.A., Kuchta, M., Martin, R.K.: Open Pit Mine Planning and Design,
Two Volume Set & CD-ROM Pack. CRC Press, Boca Raton (2013)
 Carbon-Aware Mine Planning with a Novel Multi-objective Framework 515

8. Khan, A., Niemann-Delius, C.: A differential evolution based approach for the
production scheduling of open pit mines with or without the condition of grade
uncertainty. Appl. Soft Comput. 66, 428–437 (2018). https://doi.org/10.1016/j.
asoc.2018.02.010
9. Konak, A., Coit, D.W., Smith, A.E.: Multi-objective optimization using genetic
algorithms: a tutorial. Reliabil. Eng. Syst. Saf. 91(9), 992–1007 (2006). https://
doi.org/10.1016/j.ress.2005.11.018
10. Luong, N.H., Alderliesten, T., Bel, A., Niatsetski, Y., Bosman, P.A.: Application
and benchmarking of multi-objective evolutionary algorithms on high-dose-rate
brachytherapy planning for prostate cancer treatment. Swarm Evol. Comput. 40,
37–52 (2018). https://doi.org/10.1016/j.swevo.2017.12.003
11. Marler, R.T., Arora, J.S.: Survey of multi-objective optimization methods for engi-
neering. Struct. Multidisc. Optim. 26(6), 369–395 (2004). https://doi.org/10.1007/
s00158-003-0368-6
12. Morrison-Saunders, A., McHenry, M.P., Sequeira, A.R., Gorey, P., Mtegha, H.,
Doepel, D.: Integrating mine closure planning with environmental impact assess-
ment: challenges and opportunities drawn from african and australian practice.
Impact Assess, Project Apprais. 34(2), 117–128 (2016). https://doi.org/10.1080/
14615517.2016.1176407
13. Newman, A.M., Rubio, E., Caro, R., Weintraub, A., Eurek, K.: A review of opera-
tions research in mine planning. Interfaces 40(3), 222–245 (2010). https://doi.org/
10.1287/inte.1090.0492
14. Rimélé, M.A., Dimitrakopoulos, R., Gamache, M.: A stochastic optimization
method with in-pit waste and tailings disposal for open pit life-of-mine produc-
tion planning. Res. Policy 57, 112–121 (2018). https://doi.org/10.1016/j.resourpol.
2018.02.006
15. Soto, R., Crawford, B., Almonacid, B., Johnson, F., Olguı́n, E.: Solving open-pit
long-term production planning problems with constraint programming a perfor-
mance evaluation. In: 2014 9th International Conference on Software Engineering
and Applications (ICSOFT-EA), pp. 70–77. IEEE (2014). https://doi.org/10.5220/
0005093900700077
16. Tan, K.C., Lee, T.H., Khor, E.F.: Evolutionary algorithms for multi-objective opti-
mization: performance assessments and comparisons. Artif. Intell. Rev. 17(4), 251–
290 (2002). https://doi.org/10.1023/a:1015516501242
17. Touzout, F.A., Benyoucef, L.: Multi-objective multi-unit process plan generation in
a reconfigurable manufacturing environment: a comparative study of three hybrid
metaheuristics. Int. J. Prod. Res. 57(24), 7520–7535 (2019). https://doi.org/10.
1080/00207543.2019.1635277
18. Valderrama, C.V., Santibanez-González, E., Pimentel, B., Candia-Vejar, A.,
Canales-Bustos, L.: Designing an environmental supply chain network in the min-
ing industry to reduce carbon emissions. J. Clean. Prod. 254, 119688 (2020).
https://doi.org/10.1016/j.jclepro.2019.119688
19. Wang, X., Gu, X., Liu, Z., Wang, Q., Xu, X., Zheng, M.: Production process
optimization of metal mines considering economic benefit and resource efficiency
using an NSGA-II model. Processes 6(11), 228 (2018). https://doi.org/10.3390/
pr6110228
516 N. A. B. Azhar et al.

20. Xu, X.C., Gu, X.W., Qing, W., Zhao, Y.Q., Wang, Z.K.: Open pit limit opti-
mization considering economic profit, ecological costs and social benefits. Trans,
Nonferrous Metals Soc. China 31(12), 3847–3861 (2021). https://doi.org/10.1016/
s1003-6326(21)65769-2
21. Zeng, L., Liu, S.Q., Kozan, E., Corry, P., Masoud, M.: A comprehensive interdisci-
plinary review of mine supply chain management. Res. Policy 74, 102274 (2021).
https://doi.org/10.1016/j.resourpol.2021.102274
 Multi-product Lot-Sizing Problem
with Remanufacturing, Lost Sales
and Sequence-Dependent Changeover
Cost

Lucas Gana1 , Sebastián Dávila-Gálvez1,2 , and Franco Quezada1,2(B)

1
Faculty of Engineering, Industrial Engineering Department, University of Santiago
of Chile (USACH), Santiago, Chile
{lucas.gana,sebastian.davila,franco.quezada}@usach.cl
2
Faculty of Engineering, Program for the Development of Sustainable Production
Systems (PDSPS), University of Santiago of Chile (USACH), Santiago, Chile

Abstract. This work studies a lot-sizing problem motivated by a tex-

tile remanufacturing company in Chile. In particular, we investigate a
multi-product lot-sizing with remanufacturing, lost sales, and sequence-
dependent changeover costs. The problem is first formulated as a mixed-
integer linear program. Then, we adapted a known family of valid
inequalities and proposed a new exponential family of valid inequalities
taking advantage of the problem structure. We use them in a branch-and-
cut algorithm to solve the problem. The preliminary numerical results
show the proposed inequalities’ usefulness in strengthening the proposed
formulation’s linear relaxation and show that the method outperforms
the generic branch-and-cut algorithm embedded in a stand-alone math-
ematical solver.

Keywords: Lot-sizing and Scheduling · Remanufacturing ·

Changeover cost · Valid inequalities · Branch-and-Cut algorithm

1 Introduction
The alarming rise of environmental issues and climate change has shifted how
things are done. Industrial companies, which share partial responsibility for the
current environmental crisis, face pressure from governments and consumers
alike to adopt more eco-friendly practices. In Chile, a public law known as The
Extended Liability of the Producer (ERP) has been promoting waste manage-
ment since 2016 to create a circular production model that retains the util-
ity and worth of products and their components. One way to achieve this is
through remanufacturing, which involves transforming waste materials into like-
new products, materials, or substances for the original or other purposes [15].
This research was partially supported by Dicyt projects 062217QV and 062217DG,
Vicerrectorı́a de Investigación, Desarrollo e Innovación, Universidad de Santiago de
Chile.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 517–532, 2023.
https://doi.org/10.1007/978-3-031-43612-3_32
518 L. Gana et al.

Remanufacturing can effectively reduce pollution emissions and natural resource

consumption by reusing embedded materials and components in used products,
making production processes more environmentally friendly.
This work focuses on a remanufacturing system in which end-of-life prod-
ucts are collected, sorted by material, and transformed into raw materials to
create new products. We aim to optimize production planning for this system
over a multi-period horizon while meeting customer demand for remanufactured
products most cost-effectively. This problem is motivated by a Chilean textile
remanufacturing company called ECOCITEX, which specializes in producing
high-quality recycled cotton yarn. The company’s main objective is to con-
tribute to developing a circular economy by reusing textile waste that would
otherwise end up in landfills. ECOCITEX’s remanufacturing process involves
collecting textile waste from various sources, including used clothing, industrial
scraps, and post-consumer waste. The collected waste is then sorted, cleaned,
and processed into high-quality recycled cotton yarn that can be used for vari-
ous applications, including textiles, clothing, and other products. One of the key
benefits of ECOCITEX’s remanufacturing process is that it significantly reduces
the environmental impact of textile production by minimizing the use of water,
energy, and other natural resources. The company’s products are also free from
harmful chemicals and toxins, making them a safe and sustainable choice for
consumers.
We are studying a scenario where each finished product is made from a unique
bill-of-material of end-of-life products, and the production process takes place on
a single production line of machinery. As a result, whenever the production line
needs to switch from producing one finished product to another, there are setup
costs for cleaning the machinery during product changeovers. These operations
consume both time and financial resources, impacting the profitability and pro-
duction capacity of the system. Naively, reducing setup costs is best achieved by
running large lot sizes. However, this approach leads to desynchronized patterns
between customer demand and production plans, resulting in costly high inven-
tory levels. Lot-sizing models thus aim at reaching the best possible trade-off
between minimizing the setup costs and the inventory holding costs, taking into
account both the customers’ demand satisfaction and the practical limitations
of the system. Finally, as the system’s production is implicitly limited by the
quantity of returned end-of-life products by customers, we consider an additional
lost sales cost to be paid when the customers’ demand is not satisfied on time.
We thus investigate a multi-product lot-sizing problem with remanufacturing,
lost sales, and sequence-dependent changeover costs.
Only a few works have addressed multi-product lot-sizing production sys-
tems with remanufacturing and sequence-dependent changeover setups. Heuris-
tic solution approaches lot sizing problem with sequence-dependent setups ,and
setup carry-over in a closed-loop supply chain is investigated in [16]. Simi-
larly, Roshani et al. [14] investigate a MIP heuristic for an integrated lot sizing
and single machine scheduling problem considering sequence-dependent setup
time and costs in a closed-loop supply chain network with an energy-efficient
 Multi-product Lot-Sizing Problem with Remanufacturing 519

objective function. However, these authors investigate a hybrid manufactur-

ing/remanufacturing system, where finished products can be manufactured from
raw materials or remanufactured from end-of-life products, and focus on devel-
oping heuristic solution approaches to solve the problem. In contrast, we investi-
gate exact solution methods for a multi-product lot-sizing within a pure reman-
ufacturing production system with sequence-dependent changeover costs. Exact
solution approaches for a multi-echelon lot-sizing problem with remanufacturing
and lost sales have been addressed, for example, in [12] and [11]. They propose to
solve the problem through a branch-and-cut algorithm based on the new valid
inequalities that explicitly take into account the impact of a limited returns
quantity on the production system. Nonetheless, they consider a single type of
product and do not take into account sequence-dependent or changeover setup
costs. In contrast, we consider a multi-product remanufacturing production sys-
tem, where each finished product is composed of a different bill-of-material of
end-of-life products, and we take into account the changeover setup cost incur-
ring each time switches are made between products in the production line.
To the best of our knowledge, the multi-product lot-sizing for a pure remanu-
facturing system with lost sales and sequence-dependent changeover cost has not
yet been studied in the literature. The present work aims at partially closing this
gap by investigating solution approaches for this problem. Our contributions are
twofold. First, we propose a mixed-integer linear programming formulation for
a practical lot-sizing problem with remanufacturing motivated by a Chilean tex-
tile remanufacturing company. Second, we adopt a well-known family of valid
inequalities to solve this problem. Then, we propose a new family of single-
product valid inequalities that aim to better take into account the fact that
the production quantity that can be processed on a resource is limited by the
availability of its input product. Finally, we develop branch-and-cut algorithms
based on these valid inequalities and assess their computational performance
by comparing them with the one of a stand-alone mathematical programming
solver. The preliminary numerical results show the proposed inequalities’ use-
fulness in strengthening the proposed formulation’s linear relaxation and show
that the method outperforms the generic branch-and-cut algorithm embedded
in a stand-alone mathematical solver.
The remainder of this paper is organized as follows. We first provide a brief
overview of the related literature in Sect. 2. The problem description and its
MILP formulation are provided in Sect. 3. We then present the proposed fam-
ilies of valid inequalities in Sect. 4 and discuss their corresponding separation
algorithms. Computational results are summarized in Sect. 5. Finally, Sect. 6
gives a conclusion and some research perspectives.

2 Related Works
In the past decade, various efforts have been made to enhance the MILP formu-
lation of single-echelon lot-sizing problems that involve remanufacturing. This
has been achieved through the development of extended reformulations and valid
inequalities in the literature.
520 L. Gana et al.

Helmrich et al. [13] discussed several MILP formulations of the uncapac-

itated single-item single-echelon lot-sizing problem with remanufacturing and
introduced new valid inequalities by adapting the previously known (l, S, W W )
proposed by [10] to their problem. Similarly, [5] proposed a multi-commodity
reformulation and a new set of valid inequalities for this problem. In particular,
they further strengthened the (l, S, W W ) inequalities presented in [13] by con-
sidering that the amount of finished products remanufactured in a given period
t is limited by the cumulative quantity of returned products brought back up
to t. Ali et al. [3] enriched the previous works by highlighting a theoretical
property about the equivalence of the shortest path and facility location refor-
mulations. They also carried out a polyhedral analysis of a related sub-problem
based on the single-node fixed-charge network problem, proving the validity of
several flow cover inequalities and their facet-defining conditions. Akartunali
and Arulselvan [2] studied the uncapacitated and capacitated variants of this
single-item single-echelon lot-sizing problem. They showed that the uncapaci-
tated problem could not have a fully polynomial time approximation scheme
(FPTAS) and provided a pseudo-polynomial algorithm to solve the problem.
They also provided valid inequalities based on the flow-cover inequalities for the
problem with a limited production capacity. Authors of [1] studied a capaci-
tated lot sizing problem in a hybrid system. They incorporated three families of
valid inequalities to strengthen the formulation and performed a computational
test to demonstrate that these valid inequalities improved upon the original for-
mulation. In addition, valid inequalities for a multi-echelon lot-sizing problem
with remanufacturing were investigated in [11,12]. Quezada et al. [12] propose to
extend previously known valid inequalities to a stochastic setting by exploiting
the underlying structure of the stochastic process. In contrast, Quezada et al.
[11] introduced new valid inequalities that explicitly consider the impact of a
limited returns quantity on the production system. Other related works are the
ones that deal with defective items that may result from an imperfect production
system and the need for reworking. Recent works on these production systems
can be found, for example, in [7] and [6]. One main distinction between rework
and remanufacturing lies in the nature of the flows involved. Rework focuses on
internal flows within the production process, whereas remanufacturing involves
external flows through the return of products. Finally, we refer the reader to [4]
for a recent survey on single-item lot-sizing problems with remanufacturing.
Most of the works mentioned above focus on single-item, single-echelon
problems in hybrid manufacturing/remanufacturing production systems. They
assume that, due to an uncapacitated manufacturing system, it will always be
possible to satisfy the demand for finished products on time or consider explicit
capacity constraints that can be directly included in formulating valid inequal-
ities. In contrast, our investigation focuses on a pure remanufacturing system.
It considers that the demand will be lost if the quantity and/or quality of the
returned products is insufficient to meet it on time. As a result, most of the
inequalities introduced in the literature need to be more directly valid for our
problem. Finally, to our knowledge, no current work investigates valid inequali-
ties and exact solution approaches for a multi-product lot-sizing problem over a
 Multi-product Lot-Sizing Problem with Remanufacturing 521

remanufacturing production system with sequence-dependent changeover setup

costs and lost sales.

3 Problem Description and Modeling

In this section, we provide a description of the production system and mathe-

matical definitions to formulate a mixed-integer linear program for the problem.

3.1 Production System

We consider a remanufacturing system where a set of end-of-life products are

transformed into a different set of like-new products (or materials). Thus, the
system involves a set J of end-of-life type of products and a set I of finished
products. Each finished product is composed of different types of end-of-life
products. The production system is assumed to be uncapacitated. However, the
system might not be able to satisfy the customer demand on time due to part
shortages if customers do not return enough end-of-life products. In this sit-
uation, the corresponding demand is lost, incurring a penalty cost to account
for the loss of customer goodwill. In addition, we consider the case of the pro-
duction process involving changeover setup cost, i.e., the cost associated with
the time and resources required to switch a production process from producing
one product to another. It includes cleaning the equipment, changing the tools
and machinery, and adjusting the production line for the new product. These
activities are necessary to ensure the quality and consistency of the new product
and represent a cost for the company, as they consume resources and reduce
production capacity.
The time-dependent input parameters of the problem are introduced as fol-
lows. For each period t ∈ T := {1, . . . , T },

– rtj denotes the quantity collected of the end-of-life product j ∈ J .

– dit denotes the customers’ demand for finished product i ∈ I,
– fti denotes the setup cost for processing a product i ∈ I on the machine,
– gti , denotes the unit production cost for the finished product i ∈ I,
– h(s)it denotes the unit inventory cost for finished product i ∈ I,
– h(r)jt denotes the unit inventory cost for end-of-life product j ∈ J ,
– csit denotes the unit lost-sales penalty cost for finished product i ∈ I.

On the other hand, the time-invariant input parameters are defined as follows:

– αij denotes the quantity of the end-of-life product j required for the finished
product i.

3.2 Mixed-Integer Linear Programming Formulation

In order to build a mathematical model for the problem, we introduce the fol-
lowing decision variables for time period t ∈ T :
522 L. Gana et al.

– Xti defines the quantity of finished products i ∈ I,

– Yti ∈ {0, 1} represents the setup variable for finished products i ∈ I,
– I(r)jt denotes the inventory level of the end-of-life products j ∈ J ,
– I(s)it denotes the inventory level of the finished products i ∈ I,

– Wtii ∈ {0, 1} represents the changeover variable from product i to product
i , with i, i ∈ I, i = i ,


– Lit defines the lost sales variables for finished product i ∈ I.

We thus introduce the following MILP formulation for the multi-product lot-
sizing problem with remanufacturing and lost sales to minimize the total cost of
production:
⎛⎛ ⎞ ⎞
  
min ⎝⎝fti Yti + gti Xti + h(s)it I(s)it + csit Lit + cij ij ⎠
t Wt + h(r)jt I(r)jt ⎠
t∈T i∈I j∈I:j=i j∈J
(1a)

s.t. I(r)jt = I(r)jt−1 + rtj − αij Xti ∀t ∈ T ,∀j ∈ J (1b)
i∈I

I(s)it = I(s)it−1 + Xti − dit + Lit ∀t ∈ T ,∀i ∈ I, (1c)

Xti ≤ Mti Yti ∀t ∈ T ,∀i ∈ I, (1d)
Lit ≤ dit ∀t ∈ T ,∀i ∈ I, (1e)
 
Wtii ≥ Yt−1
i
+ Yti − 1 ∀t ∈ T ,∀i, i ∈ I : i = i , (1f)

Yti ≤ 1 ∀t ∈ T , (1g)
i∈I

Yti , Wtii ∈ {0, 1} ∀t ∈ T ,∀i, i ∈ I, (1h)
Xti , Lit , I(s)it ≥0 ∀t ∈ T ,∀i ∈ I, (1i)
I(r)jt ≥ 0 ∀t ∈ T ,∀j ∈ J . (1j)

The objective function (1a) minimizes the total production cost over the
whole planning horizon, i.e., the sum of the setup, production, inventory hold-
ing, lost sales, and changeover costs. Constraints (1b)–(1c) are the inventory
balance constraints. More specifically, constraints (1b) ensure that any remanu-
factured/finished product is either used to satisfy the demand or kept in stock.
if there are not enough remanufactured products to satisfy the demand, the
unsatisfied demand is lost. Constraints (1c) ensure that end-of-life product are
re-manufactured or kept in stock. Constraints (1d) link the production quan-
tity variables to the setup variables. Constraints (1e) restrict the quantity of
lost sales. Constraints (1f) link the setup variable with the changeover variable,
indicating if a transition occurs in production from one product to another.
Constraints (1g) ensure that a maximum of one product per time period is pro-
duced in the production line. Without loss of generality, we assume that the
initial inventory, I(r)j0 and I(s)i0 for each j ∈ J and i ∈ I, is set to 0. Finally,
Constraints (1h)–(1j) provide the domain of the decision variables.
 Multi-product Lot-Sizing Problem with Remanufacturing 523

In order to set the value of each constant Mti using an upper bound on the
quantity that can be processed of each product i at each time period t, this
quantity is limited by two elements: the availability of the end-of-life products
already returned by customers and the future demand for finished products. We
thus define the value of Mti , for each period t and product i, as follows:
⎧ ⎧ ⎫ ⎫
⎨ ⎨  ⎬  ⎬
−1
Mti = min min αij rkj , dik
⎩ j∈J ⎩ ⎭ ⎭
1≤k≤t t≤k≤T

4 Valid Inequalities
We now seek to adapt the single-echelon single-item (k, U ) inequalities intro-
duced in [9] and further investigated in [12] for the case of stochastic demand
and [11] for the case of the remanufacturing system by considering the limited
quantity of returned products available at each time period.
We first recall that a single-echelon single-item (k, U ) is defined as follows:
Proposition 1. Let 1 ≤ k ≤ T be a time period of the planning horizon. Let
U ⊆ {k, ..., T } be a subset of periods and t∗ = max{τ : τ ∈ U } be the last
time period belonging to U . Then, the following inequalities are valid for Prob-
lem (1a)–(1j):
⎛ ⎞
  
I(s)ik−1 + ⎝ dil ⎠ Yti ≥ (dit − Lit ) ∀i ∈ I (2)
k≤t≤t∗ l∈U :t≤l t∈U

Proof. There is a simple proof to show the validity of the inequalities (2) for
Problem (1a)–(1j). Let us start by summing up constraints (1c) from period k
to period T for any product i ∈ I, obtaining the subsequent inequality:
 
I(s)iT = I(s)ik−1 + Xti + (−dit + Lit ) (3)
t≤k≤T t≤k≤T

Then, by the non-negativity of variables I(s)it , we have:

 
I(s)ik−1 + Xti ≥ (dit − Lit ) (4)
t≤k≤T t≤k≤T

Replacing Xti by its upper bound provided by constraints (1d), we have:

⎛ ⎞
   
I(s)ik−1 + ⎝ dil ⎠ Yti ≥ dit − Lit . (5)
t≤k≤T l≤t≤T t≤k≤T
T
 Finally, taking any set U ⊆ {k, . . . , T }, we can replace t=k (dit − Lit ) by
t∈U (dt − Lt ) in the right-hand side of inequality (5) as the former is always
i i

greater of equal to the latter due to the non-negativity of variables Lit . On the
T T T 
other hand, we can replace t=k ( l=t dil )Yti by t=k ( l∈U :l≥t dil )Yti on the
left-hand as we must now satisfy the demand in time periods t ∈ U . 
524 L. Gana et al.

Note that inequalities (2) can be strengthened by taking into account that
the production of each product i is limited by the minimum availability among
all
Tend-of-life products j returned from period one until period t, i.e., we replace
i
d
l=t l by
⎧ ⎧ ⎫ ⎫
⎨ ⎨  ⎬  ⎬
−1
φit = min min j∈J αij rlj , dil . (6)
⎩ ⎩ ⎭ ⎭
1≤l≤t l∈U :t≤l

Proposition 2. Let 1 ≤  < k ≤ T be two periods of the planning horizon. Let

U ⊆ {k, ..., T } be a subset of periods and t∗ = max{τ : τ ∈ U } be the last time
period belonging to U . Then, ∀i ∈ I, ∀j ∈ J the following inequalities are valid
for Problem (1a)–(1j):
⎧ ⎫
 ⎨  ⎬ 
ij −1 −1
α I(r)j−1 + I(s)ik−1 + min αij rlj , dil Yti ≥ (dit − Lit )
⎩ ⎭
k≤t≤t∗ ≤l≤t l∈U :t≤l t∈U

(7)

Note that the proposed inequalities (7) correspond to a simple adaptation of

the single-echelon (, k, U ) inequalities introduced in [11], and its validity can be
proved directly by following the proof in [11].
We now investigate a family of new single-echelon inequalities which seek
to better take into account the fact that the production quantity that can be
processed on a resource is limited by the availability of its input product. More
precisely, the quantity that can be processed in period t is limited by the amount
of used products which has been returned up to period t but not yet processed
in this period. We thus introduce a set of valid inequalities stating that the total
quantity processed over a given time interval [, t] is limited by the sum of the
number of used products already returned and not yet transformed at the end
of the period  − 1 and of the amount of used products returned during interval
[, t].

Proposition 3. Let 1 ≤  ≤ T be a time period of the planning horizon and

S ⊆ {, ..., T } be a subset of periods. The following inequalities are valid for
Problem (1a)–(1j):
⎛ ⎞
  
I(r)j−1 + ⎝ rlj ⎠ Yti ≥ αij Xti ∀i ∈ I, ∀j ∈ J (8)
t∈S ≤l≤t t∈S

Proof. Let (X, Y, I(r), I(s), W, L) be a feasible solution of Problem (1a)–(1j).

We show that this solution complies with inequalities (8) for any period  and
any subset S ⊆ {, ..., T }. This is done by induction on the index of the last
period t that may belong to subset S.
Base step: Set t =  and S ⊆ {}.
We consider two cases:
 Multi-product Lot-Sizing Problem with Remanufacturing 525

 
– S = ∅. In this case, τ ∈S Xτi = 0 whereas I(r)j−1 + τ ∈S φjτ Yτi ≥ I(r)j−1 ≥
0. Inequality (8) is trivially respected. 
– S = {}. If Yj = 0. There is no production in period  so that τ ∈S Xτi =

Xi = 0 whereas I(r)j−1 + τ ∈S φjτ Yτi ≥ I(r)i−1 ≥ 0. Inequality (8) is trivially
respected. If Yi = 1. In this case, the inventory balance constraint (1c) for
period  gives:

I(r)j−1 + rj − Xi ≥ 0 by non-negativity of variables I(r)j .

By noting that, under the current assumptions, φj Yi = rj , we have:

I(r)j−1 + φ Yi ≥ Xi ,

which corresponds to the expression of inequality (8) for S = {}. Therefore,

inequality (8) is thus valid for any S ⊆ {}.

Induction step: Assume that inequality (8) is valid for any S ⊂ {1, ..., t}. We show
that, under this assumption, inequality (8) is valid for any S ⊂ {1, ..., t + 1}. We
again consider two cases:

– t+1∈ / S. Using the induction hypothesis, inequality (8) is valid 

for S.
– t+1 ∈ S. If Yt+1
i
= 0. There is no production in period t+1 so that τ ∈S Xτi =
 i j  j i j 
τ ∈S\{t+1} Xτ whereas I(r)−1 + τ ∈S φτ Yτ = I(r)−1 + φjτ Yτi .
j  τ ∈S\{t+1}j i
Using the induction hypothesis, we have: I(r)−1 + τ ∈S\{t+1} φτ Yτ ≥
 i
τ ∈S\{t+1} Xτ so that inequality (8) holds.

i
If Yt+1 = 1. Summing up the inventory balance constraints (1c) over periods
 to t + 1 and using the non-negativity of variables I(r)j and Xτi gives:


t+1 
t+1
I(r)j−1 + rτj − Xτi ≥ 0
τ = τ =

t+1 
I(r)j−1 + rτj − Xτi ≥ 0.
τ = τ ∈S

t+1 j
By noting that under the current  assumption, φjt+1 Yt+1
i
= τ = rτ and
adding the non-negative terms τ ∈S\{t+1} φjτ Yτi to the left-hand side of the
previous inequality, we obtain:
 
I(r)j−1 + φjτ Yτi − Xτi ≥ 0.
τ ∈S τ ∈S

This shows that if Inequality (8) is valid for any S ⊂ {1, ..., t}, it is also valid
for any S ⊂ {1, ..., t + 1}, which concludes the proof. 
526 L. Gana et al.

4.1 Separation Algorithms

We now briefly discuss the resolution of the separation problem for each proposed
valid inequality. We first recall that this problem consists in finding an inequality
˜ I(s),
violated by a given solution (X̃, Ỹ, I(r), ˜ L̃, W̃) of the linear relaxation of
Problem (1a)–(1j) or prove that no such inequality exists.
In the case of the simple (k, U ) inequalities in Proposition 1, an exact separa-
tion algorithm is known to find the most violated inequality that runs in O(T 2 ).
For each period k = 1, ..., T , the algorithm finds the set U of time periods by
inspection. See [9] for further details on the separation algorithm.
Then, for the case of valid inequalities in Proposition 2, the separation is not
trivial as is mentioned in [11], in particular, because the value of each coefficient
φjt simultaneously depends on U and . We thus consider a heuristic separation
algorithm in our computational experiments as is proposed in [11], which con-
sists of finding the set U as is done for inequalities in Proposition 1 and then
computing the coefficient φjt for each given period . The algorithm has a time
complexity of O(T 2 ).
Finally, we note that inequalities in Proposition 3 form an exponential class
of valid inequalities, i.e., their number grows exponentially fast with the num-
ber of time periods T . However, the corresponding separation problem is poly-
nomially solvable by inspection in O(T 2 ). Namely, to find the most violated
valid inequalities (8) by the current linear relaxation, we need to find, for each
period  ∈ T , the subset U minimizing the difference between the left and the
right-hand
 side of inequality
 (8). This amounts to finding the subset S mini-
mizing t∈S φjt Ỹti − t∈S X̃pt . This can be done by inspection of each period
t ∈ {, . . . , T } and adding the time period t to the set S if φjt Ỹti − X̃ti ≤ 0.
We briefly summarize the proposed valid inequalities and their related fea-
tures in Table 1. Column Class indicates whether the number of valid inequali-
ties in the family grows linearly or exponentially fast concerning the number of
periods in the planning horizon. Column Complexity reports the difficulty of the
corresponding separation problem. Column Sep. alg. then indicates whether
the proposed separation problem is solved by an exact or a heuristic algorithm
in our numerical experiments. Column Time reports the asymptotic time com-
plexity of the implemented separation algorithm, and Column Total # reports
the total possible number of valid inequalities that can be generated. Note that
all these features significantly impact the numerical tractability and efficiency of
a class of valid inequalities.

Table 1. Class and complexity of valid inequalities

Inequalities Class Complexity Sep. alg. Time Total #

Proposition 1 Exponential Polynomial Exact O(|T | ) 2|T | |I|
2

Proposition 2 Exponential unknown Heuristic O(|T |2 ) 2|T | |T ||I||J |

Proposition 3 Exponential Polynomial Exact O(|T |2 ) 2|T | |I||J |
 Multi-product Lot-Sizing Problem with Remanufacturing 527

5 Computational Experiments
In this section, we focus on assessing the performance of the proposed valid
inequalities when used within a branch-and-cut algorithm. We compare the per-
formance of this algorithm with one of the generic branch-and-cut algorithms
embedded in a mathematical programming solver.

5.1 Instance Generation

We considered two sets of instances: Set 1 instances involve |T | = 25 periods

and |I| = |J | = 5 finished/end-of-life products, whereas instances from Set 2
involve |T | = 25 periods and |I| = |J | = 10 finished/end-of-life products. The
instances were randomly generated within each set by adapting the procedure
presented in [8] and extended in [11]. More precisely, we consider one value of
the setup-holding cost ratio f /h ∈ {1000}, two values for the production-holding
cost ratio g/h ∈ {2, 4} and three values of the returns-demand quantity ratio
r/d ∈ {1, 2, 3}. For each set and each possible combination of f /h, g/h, and r/d,
ten random instances were generated, resulting in a total of 120 instances.
For each instance, the value of each problem parameter was set as follows.
Demand dt was uniformly distributed in the interval [0, 100] and the returns
quantity ¯ 1.2(r/d)d],
rt was uniformly distributed in the interval [0.8(r/d)d, ¯ where
 t
d
d¯ = |T | is the average demand per period. The bill-of-materials coefficients αij
were randomly generated following a discrete uniform distribution over interval
[1, 6] for each i ∈ I, j ∈ J . The holding cost h(r)jt for each end-of-life product
j ∈ J was randomly generated following a discrete uniform distribution over
interval [2, 7] and the holding cost h(s)it for each finished product i ∈ I was
randomly generated following a discrete uniform distribution over interval [7, 12].
The production cost gti was 
uniformly

distributed in the interval [0.8(g/h)h̄i ,
h(s)it
1.2(g/h)h̄], where h̄i = t∈T
|T | is the average holding cost for finished
product i ∈ I. The setup i
ft and changeover setup ci,j
t cost were uniformly
distributed in the interval [0.8(f /h)h̄, 1.2(f /h)h̄]. The unit penalty cost for lost
sales, csit , was fixed at 200 per unit. grhttps://github.com/LucasGana/ICCL.

5.2 Results

We carried out extensive numerical experiments in order to assess the compu-

tational performance of the proposed valid inequalities. This was achieved by
solving each considered instance using three alternative branch-and-cut algo-
rithms:

– GRB: the generic branch-and-cut algorithm embedded in GUROBI 9.5.2

using the formulation (1a)–(1j).
– P 1: a branch-and-cut algorithm using the (k, U ) inequalities in Proposition
1 with coefficient φ proposed in (6) to strengthen the formulation (1a)–(1j).
528 L. Gana et al.

Table 2. Performance of GUROBI and branch-and-cut methods over instance in Set 1.

r/d Method R.LPgap R.MIPgap MIPgap C.Time R.Time T.Time # Cuts

1 GRB 2.4 1.2 0.2 0.0 0.2 237.9 –
P1 2.2 0.8 0.2 0.6 0.7 240.4 295
P2 1.8 0.8 0.3 10.0 16.3 243.9 10628
P3 1.8 0.8 0.3 1.3 2.6 240.8 254
P 1+P 3 1.9 0.8 0.2 1.5 2.5 235.3 488
P 2+P 3 1.5 0.0 0.0 10.0 14.3 242.6 9919
2 GRB 3.6 2.1 0.5 0.0 0.0 240.3 –
P1 3.1 1.3 0.5 0.8 0.8 240.4 359
P2 2.5 1.3 0.6 9.2 13.9 242.2 11063
P3 2.8 1.4 0.5 1.3 2.0 240.9 218
P 1+P 3 2.9 1.4 0.5 1.4 2.6 240.8 487
P 2+P 3 2.3 0.0 0.0 10.1 17.1 242.3 11002
3 GRB 4.4 2.6 0.5 0.0 0.0 225.6 –
P1 3.6 1.7 0.5 0.9 1.3 223.2 401
P2 3.0 1.7 0.5 9.2 12.6 220.1 12041
P3 3.6 1.7 0.5 1.1 1.8 205.4 179
P 1+P 3 3.4 1.7 0.5 1.3 2.8 206.0 485
P 2+P 3 2.7 0.0 0.0 11.3 18.9 232.2 11735

– P 2: a branch-and-cut algorithm using the (, k, U ) inequalities in Proposition

2 to strengthen the formulation (1a)–(1j).
– P 3: a branch-and-cut algorithm using the newly introduced (, S) inequalities
in Proposition 3 to strengthen the formulation (1a)–(1j).
– P 1 + P 3: a branch-and-cut algorithm using valid inequalities in Proposition
1 and 3 to strengthen the formulation (1a)–(1j).
– P 2 + P 3: a branch-and-cut algorithm using valid inequalities in Proposition
2 and 3 to strengthen the formulation (1a)–(1j).

Each algorithm is based on the solver GUROBI 9.5.2, and it generates corre-
sponding inequalities through a cutting-plane generation algorithm at the root
node and at intermediate nodes of the branch-and-bound search tree using the
user Callbacks provided by the solver.
All related linear programs and mixed-integer linear programs were solved
using GUROBI 9.5.2 with the solver default settings. The algorithms were imple-
mented in Python 3.9. All tests were run on the computing infrastructure con-
sisting of an Intel(R) Core(TM) i5-10.300H, 2.50 GHz, and 8 GB RAM to solve
each instance. We imposed a time limit of 300 s to find an optimal solution for
each instance.
The corresponding results are displayed in Table 2 for Set 1 instances
and Table 3 for Set 2 instances. Column Method indicates the branch-and-cut
 Multi-product Lot-Sizing Problem with Remanufacturing 529

Table 3. Performance of GUROBI and branch-and-cut methods over instance in Set 2.

r/d Method R.LPgap R.MIPgap MIPgap C.Time R.Time T.Time # Cuts

1 GRB 2.2 1.7 0.9 0.0 0.0 300.1 –
P1 2.1 1.1 0.9 1.1 2.8 300.1 569
P2 1.9 1.2 1.0 37.9 63.8 300.1 37273
P3 1.7 1.1 0.9 6.4 12.0 300.0 691
P 1+P 3 1.8 1.1 0.9 7.3 15.7 300.3 1159
P 2+P 3 1.7 0.0 0.0 53.2 51.4 300.2 36025
2 GRB 2.9 2.5 1.2 0.0 0.0 300.1 –
P1 2.7 1.5 1.2 1.5 4.1 300.0 822
P2 2.4 1.6 1.3 34.5 64.3 300.0 42574
P3 2.4 1.5 1.2 5.2 14.4 300.1 579
P 1+P 3 2.5 1.5 1.2 4.7 15.1 300.3 1221
P 2+P 3 2.2 0.0 0.0 184.4 98.4 300.1 40734
3 GRB 3.6 2.6 1.3 0.0 0.3 300.0 –
P1 3.2 1.9 1.5 1.7 6.2 300.2 952
P2 2.9 2.0 1.6 32.2 97.3 300.1 45887
P3 3.1 1.9 1.5 5.0 11.0 300.0 470
P 1+P 3 3.1 1.8 1.4 4.2 14.4 300.1 1251
P 2+P 3 2.8 0.0 0.0 43.5 187.2 300.0 43018

algorithm used to solve the instances. Column R.LPgap reports the gap between
the value of the linear relaxation strengthened by the corresponding valid
inequalities and the best feasible solution found through the branch-and-bound
search. The GRB method reports the gap between the value of the initial linear
relaxation and the best feasible solution found through the branch-and-bound
search. Column R.MIPgap reports the gap between the lower bound at the root
node (after the generation of GUROBI generic cutting planes) and the best
feasible solution found through the branch-and-bound search. Column MIPgap
reports the gap between the best lower bound and the best feasible solution
found through the branch-and-bound search. The average CPU time in seconds
for the cutting-plane generation of each method is reported in column C.Time,
the CPU time in seconds spent at the root node in column R.Time and the aver-
age total CPU time in seconds in column T.Time. Finally, column # Cuts reports
the number of cuts generated by each method. Note that each line corresponds
to the average value of the corresponding 10 instances.
In general, we observe that the customized branch-and-cut algorithms based
on valid inequalities in Proposition 2 and 3 (P 2 + P 3) outperforms the method
GRB and the other methods, providing solutions of better quality with similar
computation times for all group of tested instances. Specifically, the method
P 2 + P 3 is able to almost close the optimality gap at the root node of the
530 L. Gana et al.

branch-and-bound tree (see column R.MIPgap ) in shorter computation times (see

column R.Time).
Regarding the relative performance of the methods P 1 and P 2, the results
suggest that the branch-and-cut algorithm based on the valid inequalities in
Proposition 2 always provides a better linear relaxation strengthening than the
valid inequalities in Proposition 1 (see column R.LPgap ); however, this improve-
ment is not exhibited by the performance of the branch-and-cut algorithm based
on the same valid inequalities (see column MIPgap and column T.Time). This
might be explained by the large number of valid inequalities generated through
the cutting generation process, which can deteriorate the global performance
of the algorithm. Regarding the method P 3, we observe that the method can
usually provide better linear relaxation strengthening than the methods P 1 and
P 2. However, its performance deteriorates for the instances with the largest
demand-return ratio. This might be explained by the fact that the correspond-
ing instances involve a large amount of returned products as compared to the
demand so that the availability of the returned products does not necessarily
limit the quantity processed on most periods. Thus, method P 2 seems to pro-
vide a better trade-off for this case.

6 Conclusion and Perspectives

We consider a multi-product lot-sizing problem to plan production activities and

determine the product sequence for a remanufacturing system. The problem is
formulated as a mixed-integer linear program ,and we focused on strengthening
this formulation in order to be able to provide optimal or near-optimal solutions
for the problem. Besides the formulation of a new practical lot-sizing and schedul-
ing problem, our main contribution is the adaptation and the development of a
new family of valid inequalities which take into account the limitations of the
production system given by the limited availability of end-of-life products. Com-
putational experiments show that valid inequalities perform well at strengthen-
ing the linear relaxation of the formulation, and the proposed branch-and-cut
algorithm performs well as compared to the generic branch-and-cut algorithm
of the GUROBI 9.5 solver and other studied methods.
The first research direction is to develop a more effective cutting plane gener-
ation algorithm for better selecting valid inequalities to be added in the branch-
and-cut algorithm and acceleration techniques that further improve their per-
formance. Additional computational experiments are also needed to assess the
size of the largest instances that may be solved with the proposed exact solution
approach. On a longer perspective, we could extend the proposed valid inequal-
ities to lot-sizing problems with returns involving complicating features such as
changeover setups, a limited capacity, or backlogging.
In terms of modeling, an interesting research direction will be to extend
the model to a multi-echelon or small-bucket formulation to become closer to
a more realistic and practical problem. In general, assuming that all informa-
tion is certain leads to formulations that might deviate from reality. However,
 Multi-product Lot-Sizing Problem with Remanufacturing 531

these models often serve as a starting point for future research. In our case, the
proposed deterministic approach opens up an interesting discussion about the
returned flows in the textile industry. In a favorable scenario where the returns
are significantly higher than the demand, the deterministic assumption does not
highly impact the production plan since there are enough returned products to
carry out the production plan. Otherwise, as the amount of end-of-life products
available restricts the ability to comply with the planning, it might incur new
costs and lost sales to adjust the production plan. Hence, extending the model
to a stochastic setting is also worth investigating, as demand and end-of-life
product quantities are usually difficult to predict accurately.

References
1. Aishah, S., Ali, S., Ali, F., Talib, M.L., Rozita, W., Din, W.: Valid inequalities for
the capacitated lot-sizing problem in a hybrid manufacturing and remanufacturing
system (2021)
2. Akartunalı, K., Arulselvan, A.: Economic lot-sizing problem with remanufacturing
option: complexity and algorithms. In: Pardalos, P.M., Conca, P., Giuffrida, G.,
Nicosia, G. (eds.) MOD 2016. LNCS, vol. 10122, pp. 132–143. Springer, Cham
(2016). https://doi.org/10.1007/978-3-319-51469-7 11
3. Ali, S.A.S., Doostmohammadi, M., Akartunalı, K., van der Meer, R.: A theoretical
and computational analysis of lot-sizing in remanufacturing with separate setups.
Int. J. Prod. Econ. 203, 276–285 (2018)
4. Brahimi, N., Absi, N., Dauzère-Pérès, S., Nordli, A.: Single-item dynamic lot-sizing
problems: an updated survey. Eur. J. Oper. Res. 263(3), 838–863 (2017)
5. Cunha, J.O., Konstantaras, I., Melo, R.A., Sifaleras, A.: On multi-item economic
lot-sizing with remanufacturing and uncapacitated production. Appl. Math. Model.
50, 772–780 (2017)
6. Goerler, A., Lalla-Ruiz, E., Voß, S.: Late acceptance hill-climbing matheuristic for
the general lot sizing and scheduling problem with rich constraints. Algorithms
13(6), 138 (2020)
7. Goerler, A., Voß, S.: Dynamic lot-sizing with rework of defective items and mini-
mum lot-size constraints. Int. J. Prod. Res. 54(8), 2284–2297 (2016)
8. Guan, Y., Ahmed, S., Nemhauser, G.L.: Cutting planes for multistage stochastic
integer programs. Oper. Res. 57(2), 287–298 (2009)
9. Loparic, M., Pochet, Y., Wolsey, L.A.: The uncapacitated lot-sizing problem with
sales and safety stocks. Math. Program. 89(3), 487–504 (2001)
10. Pochet, Y., Wolsey, L.A.: Polyhedra for lot-sizing with wagner-whitin costs. Math.
Program. 67(1), 297–323 (1994)
11. Quezada, F., Gicquel, C., Kedad-Sidhoum, S.: New valid inequalities for a multi-
echelon multi-item lot-sizing problem with returns and lost sales. In: Mes, M.,
Lalla-Ruiz, E., Voß, S. (eds.) ICCL 2021. LNCS, vol. 13004, pp. 192–207. Springer,
Cham (2021). https://doi.org/10.1007/978-3-030-87672-2 13
12. Quezada, F., Gicquel, C., Kedad-Sidhoum, S., Vu, D.Q.: A multi-stage stochastic
integer programming approach for a multi-echelon lot-sizing problem with returns
and lost sales. Comput. Oper. Res. 116, 104865 (2020)
13. Retel Helmrich, M.J., Jans, R., van den Heuvel, W., Wagelmans, A.P.: Economic
lot-sizing with remanufacturing: complexity and efficient formulations. IIE Trans.
46(1), 67–86 (2014)
532 L. Gana et al.

14. Roshani, A., Paolucci, M., Giglio, D., Demartini, M., Tonelli, F., Dulebenets, M.A.:
The capacitated lot-sizing and energy efficient single machine scheduling problem
with sequence dependent setup times and costs in a closed-loop supply chain net-
work. Ann. Oper. Res. 321(1–2), 469–505 (2023)
15. Suzanne, E., Absi, N., Borodin, V.: Towards circular economy in production plan-
ning: challenges and opportunities. Eur. J. Oper. Res. 287(1), 168–190 (2020)
16. Torkaman, S., Ghomi, S.F., Karimi, B.: Multi-stage multi-product multi-period
production planning with sequence-dependent setups in closed-loop supply chain.
Comput. Ind. Eng. 113, 602–613 (2017)
 A Radius-Based Approach
for the Bi-Objective p-Center
and p-Dispersion Problem

Niels De Walsche, Carlo S. Sartori(B) , and Hatice Çalık

Department of Computer Science, KU Leuven, Ghent, Belgium

[email protected],
{carlo.sartori,hatice.calik}@kuleuven.be

Abstract. Facility location is a well-known class of problems which

come with a variety of objective functions. In applications such as emer-
gency facility location where human life is at stake, service quality, which
is measured by distance or time to reach facilities, and fairness become
priority over cost. A certain level of service quality is often achieved
through coverage constraints or minimax type of objective functions as
it is the case in the p-center problem. In some other applications such as
location of military bases and franchise stores, in addition to minimizing
maximum service distance it may be desirable also to keep the facilities as
distant as possible for competitiveness or security reasons. A problem of
this sort corresponds to the bi-objective p-center and p-dispersion prob-
lem which requires locating p centers to serve a set of demand points. The
p-center and p-dispersion problems are rather well-studied independently
with efficient approximation algorithms, mathematical models and exact
algorithms. This paper investigates effectiveness of radius-based formu-
lations, approximation algorithms and matrix reduction rules for solving
the complex optimization problem where these two objectives must be
simultaneously optimized. We develop an exact algorithm to produce
a set of efficient solutions for this problem by employing a mathemat-
ical formulation based on state-of-the-art models for both independent
p-center and p-dispersion problems. Our algorithm creates a Pareto set
with multiple solutions presenting different tradeoffs between the two
objectives. A computational study with large-scale instances demon-
strates the improved performance of our method compared to a previous
approach from the literature.

Keywords: Facility location · p-center · p-dispersion · Bi-objective

optimization

1 Introduction
The p-center problem (pC) is an NP-Hard [13] facility location problem (FLP)
that requires location of p centers on a given network such that the maximum
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 533–549, 2023.
https://doi.org/10.1007/978-3-031-43612-3_33
534 N. De Walsche et al.

distance between demand nodes and centers is minimized. The problem arises in
emergency logistics for locating ambulances, fire stations, police patrol units and
shelters; military logistics as well as districting. Similar to pC, the p-dispersion
problem (pD) is also NP-hard [9] and requires locating p facilities on a given
network. Unlike pC, the goal is to maximize the minimum distance between
the selected facilities. The optimal solutions to these two problems could be
significantly different as it can be observed in an OR-Library instance in Fig. 1.
In the pC solution of Fig. 1(a), the facilities are more centralized in their clusters
and in the entire network. The pD facilities in Fig. 1(b), however, are more
dispersed and located at the extremes of the network. It is therefore worthwhile
investigating bi-objective approaches if one wishes to maintain a certain level of
balance between the two objectives.

Fig. 1. Optimal solutions for instance pmed1 from OR-Library (p = 5). Square nodes
indicate facilities and circles indicate demand points clustered with the color of their
closest open facility.

In this paper we consider the Bi-objective p-center and p-dispersion prob-

lem (BpCD) which has been introduced by [20]. The BpCD requires locating p
centers on a given network while simultaneously minimizing the maximum dis-
tance between demand points and centers (pC objective) and maximizing the
minimum distance between centers (pD objective).
A notable application of BpCD is the location of franchises. Kuby [15] notes
that pD is useful to reduce intra-chain competition, which arises when stores
from the same franchise are located too close to each other. When they are close,
their profit margins are reduced. However, Kuby also mentions that pD could
be used in a multi-objective scenario alongside objectives such as the p-center.
In such a case, we would like not only to minimize intra-chain competition, but
also maximize customer reach by means of the p-center objective.
Similarly, one may find applications of BpCD in the positioning of military
bases. On the one hand, we would like to ensure short distance ranges for the
defense of strategic areas (pC). On the other hand, we want to maintain a safe
 A Radius-Based Approach for BpCD 535

distance between bases so as to avoid catastrophic results in case of enemy

attacks (pD). Tutunchi and Fathi [20] also refer to applications in the positioning
of traffic sensors in road networks. The objective is to monitor similar road
segments using a few sensors (pC), while simultaneously diversifying the types
of road segments for the remaining sensors (pD). It is clear therefore that there
are many practical applications for the BpCD which motivated us to study new
solution methods to solve this problem.
The contributions of this paper are twofold. First, we develop a method
to quickly solve BpCD employing state-of-the-art mathematical formulations
for p-center and pD. This includes exploiting the problem structure to develop
a special-purpose algorithm based on decomposition into simpler subproblems
instead of relying solely on general-purpose mathematical programming solvers.
To speed up the methods, we employ effective lower and upper bounds as well
as matrix reduction rules inspired by the well-established p-center literature.
Second, we demonstrate by means of a comprehensive computational study the
improvements obtained with our method, which is capable of solving far larger
problems than previous approaches in the literature. We conclude this paper
with a discussion over our results and directions for future research.

2 Literature Review

The seminal paper by [12] introduces the 1-center problem for the purpose of
locating a police station on a given network so that the maximum distance from
this station is minimum. Both absolute and vertex-restricted 1-center problems
are discussed. An absolute center can be placed anywhere on the network whereas
a vertex center has to be located at a vertex. Minieka [16] extends the problem to
the case where p > 1 and provides an exact algorithm solving set covering prob-
lems for a finite number of radius (distance) values. This method has inspired the
most efficient radius formulations and set covering-based algorithms for solving
various p-center variants which have been reviewed by [2]. As [16] proves that
there exists a finite set of points in the given graph which contains an optimal
set of p centers for the absolute p-center problem and [10] provides a method
for generating this finite set, it has been possible to develop mathematical for-
mulations which can tackle both absolute and vertex p-center variants without
further adaptations ([3,6,8] and [14] are a few examples). Although recently
most p-center studies consider only vertex-restricted problems, absolute variants
remain valid for many real-world applications.
A critical and careful review of the pC literature reveals two main categories
of compact mathematical formulations: (1) Mixed-Integer Linear Programming
(MILP) models based on the classical formulation with two-index allocation
variables first provided by [6]. (2) Radius-based Integer Linear Programming
(ILP) formulations which exploit the finite set of distinct distance values between
the centers and the demand nodes. Radius formulations typically utilize either
a telescopic sum of radius values [8] or enforce selection of exactly one radius
value [3]. When the radius set is large, solving these formulations may require
536 N. De Walsche et al.

a substantial amount of computational resources. One may therefore resort to

rather solving a set covering problem for each radius value, as done by [16], or
within a more efficient binary search framework like [8]. The special structure of
the formulations provided by [3] inspires many efficient search strategies solving
the formulations for a subset of radius values at each iteration. Among them, the
strategies considering two radius values at each iteration (double bound - DB),
enable faster convergence in the presence of high quality upper bounds. The DB
methods can easily be adapted to favor instances with high quality lower bounds.
The computational study by [3] on data sets from the OR-Library [1] indicates
that the formulations introduced by [3] are the most efficient pC formulations.
One model by [3] provides the tightest known lower bound equivalent to that
of [8]. The DB methods successfully solve instances with up to 3038 nodes from
the TSPLIB (with 3 h time limit per model) while this number was restricted to
1817 previously. The efficiency of the DB methods for solving problem instances
at this scale is also confirmed through the recent study by [5] whose algorithm
solves problems with up to a million nodes, with relatively small p values that
are not observed to be very challenging in previous studies.
In addition to the classical pC problems, many variants inspired by real-world
applications have been investigated. We refer to [2] for further reading on several
other pC variants and focus on the pD and BpCD in what follows.
The first discrete FLP with a pD objective can be traced back to [15]. Kuby
[15] introduced MILP formulation for the problem using one-index variables. He
also performed initial computational experiments with small instances and even
considered a limited form of multi-objective optimization with a lexicographic
evaluation where pD was first optimized and then an alternative objective was
considered after fixing pD to its maximum value. Erkut [9] employed an IP
formulation and both branch-and-bound and heuristic methods to solve the pD.
Ravi et al. [17] presented a polynomial time 1/2-approximation algorithm for the
pD, which is based on the same principles as of the 2-approximation algorithm
for the pC [11].
Sayah and Irnich [19] introduced a new formulation for the pD which uses
radius-based variables similarly to the formulation of [8] for the pC. They also
discussed other models to solve the pD, including non-linear formulations. They
performed a computational study using large p-median instances from the OR-
Library and compared the performance of these different models. Results clearly
demonstrated that the new formulation could solve more instances to optimal-
ity compared to other models within a time limit of 1800 s. Nevertheless, the
formulation proposed by Sayah and Irnich [19] was not able to solve all pmed
instances from the OR-Library to optimality. Contardo [4] further expanded
on the research of [19] by employing a decomposition approach for solving the
radius-based formulation instead of directly utilizing a general-purpose solver.
He proposed to solve the problem by iteratively clustering locations, solving
the pD and refining the clusters. This enabled reducing the model size and
solve instances with up to 100,000 nodes under a time limit of 20h. For smaller
instances the speedup compared to [19] was significant, thereby demonstrating
 A Radius-Based Approach for BpCD 537

the importance of not only employing an efficient formulation (the radius-based

one), but also of using an efficient decomposition procedure to solve the problem.
The BpCD is first addressed by [20] who provided IP models for both the
bi- and single-objective problems. As a solution approach, [20] introduced an
incremental algorithm which is similar to the -constraint method. They used the
pC model as base and added a constraint to ensure a minimum dispersion level in
the optimal solution. By doing so, [20] obtained optimal values of pC at different
dispersion levels, thereby creating a set of efficient solutions (Pareto front) for
the bi-objective problem. Tutunchi and Fathi ensure additional computational
efficiency by solving the pC formulation using a binary search approach where
each subproblem is a set cover problem.
Table 1 provides a summary of the relevant literature. Column Prob. specifies
the problem class. Columns Ref., Model and alg. and Notes present information
about each of the main references relevant to our research. Here, we make a
distinction between approaches which use radius-based formulations and those
that use some other form of (M)ILP model, including formulations with one- and
two-index variables. The terms LB and UB denote that the approach makes use
of lower- and upper-bound information during the optimization process. The
term VI stands for Valid Inequalities, while DR stands for Dominance Rules,
which can reduce the size of the formulation. The last two columns, n and p,
indicate respectively the largest number of nodes and the largest number of
facilities to be located in the instances tested by the associated reference.

Table 1. Summary of relevant literature.

Prob Ref Model and alg Notes n p

pC Daskin [6] MILP, bisection search 900 200
Elloumi et al. [8] Radius, binary search LB,UB 1817 150
Çalık and Tansel [3] Radius, DB LB,UB,DR 3038 500
Contardo et al. [5] Row generation alg LB,UB,DR 1,000,000 30
pD Kuby [15] MILP 25 10
Erkut [9] MILP, branch-and-bound LB,UB 40 16
Sayah and Irnich [19] Radius LB,UB,VI 900 200
Contardo [4] Radius, decomposition alg LB,UB 100,000 20
BpCD Tutunchi and Fathi [20] MILP, incremental alg LB,UB,VI 600 30
This work Radius, incremental alg LB,UB,DR 2500 200

It becomes clear from Table 1 that developments for both pC and pD are a lot
more advanced than those for BpCD, which is still a relatively recent problem.
While researchers have been able to solve instances with up to a million nodes
for pC and a hundred thousand nodes for pD, for BpCD only instances with up
to 600 nodes could be solved until now. Table 1 also demonstrates that success-
ful approaches for pC and pD make use of radius-based formulations. Finally,
most methods make use of lower- and upper-bound information to at least some
538 N. De Walsche et al.

extent. However, only a few approaches use valid inequalities or dominance rules,
techniques which can sometimes speed up the solution process.
Based on the information from the relevant literature, it is surprising that
Tutunchi and Fathi [20] did not experiment with radius-based formulations to
solve BpCD. It is our main hypothesis that addressing the BpCD by means of
radius-based formulations allows us to increase the size of the problems that
can be efficiently solved. Indeed, we will demonstrate that employing a suitable
mathematical formulation allows us to solve instances which are 3–4 times larger
than those solved using classic two-index MILP formulation.

3 Mathematical Model
Let G be a given undirected network with I = {1, . . . , n} being the set of demand
nodes and J = {1, . . . , m} being the set of potential centers. The BpCD requires
selecting a subset X ⊂ J such that |X| = p under two objective functions:
(1) minimize the maximum distance between the centers and demand points
and (2) maximize the minimum distance between the centers. Let us define dij
as the shortest path distance between points i and j. Let Dc = [dij ] denote
the n × m distance matrix between demand nodes and potential centers and
Dd = [dhj ] denote the m × m distance matrix between potential centers. The pC
can then be formalized as min max min dij whereas pD can be formalized
X⊂J:|X|=p i∈I j∈X
as max min dhj . It is easy to observe that the optimal values of pC and
X⊂J:|X|=p h,j∈X:h=j
pD will be equal to one of the entries in Dc and Dd , respectively. Consider two
ordered sets R = {r1 , . . . , rK } and Q = {q1 , . . . , qL } of distinct distance values
in Dc and Dd , respectively, and let RI and QI be the corresponding index sets.
Recalling the pC formulations of [3], it is then intuitive to associate a binary
variable for each radius value in each set as follows.
uk = 1 if rk , k ∈ RI is the value of pC objective for BpCD, uk = 0 otherwise.
vl = 1 if ql , l ∈ QI is the value of pD objective for BpCD, vl = 0 otherwise.
We may then formulate the BpCD as follows:

min f1 = rk uk (1)
k∈RI

max f2 = q l vl (2)
l∈QI

s.t. yj = p (3)
j∈J

yj ≥ uk , ∀ i ∈ I, k ∈ RI (4)
j∈J:dij ≤rk

uk = 1 (5)
k∈RI

yh + yj ≤ 1 + vl , ∀ h, j ∈ J (6)
l∈QI :dhj ≥ql
 A Radius-Based Approach for BpCD 539

vl = 1 (7)
l∈QI

uk ∈ {0, 1}, ∀ k ∈ RI (8)

vl ∈ {0, 1}, ∀ l ∈ QI (9)
yj ∈ {0, 1}, ∀j∈J (10)

Functions (1) and (2) determine the objective values for pC and pD selected by
Constraints (5) and (7) from R and Q, respectively. Constraints (4) ensure that
each demand point is covered by a center within the selected pC radius. For any
open center pair, Constraints (6) guarantee that the distance between them is
greater than or equal to the selected pD value. Constraints (3) select exactly p
centers and Constraints (8)-(10) are binary restrictions.
However, this BpCD formulation cannot be directly solved by general-
purpose mathematical programming solvers. The existence of two competing
objective functions requires more sophisticated solution techniques and, more
importantly, it requires that we clarify what a good solution is. In this paper,
we will use a terminology based on the work of Erghot et al. [7]. Let us begin by
defining S as the set of all feasible solutions to BpCD and s ∈ S is a solution (an
assignment of facilities). Let us further denote objective value for pC by f1 (s)
and objective value for pD by f2 (s) for solution s. A solution s ∈ S dominates
another solution s ∈ S if one of the following two conditions is true:

f1 (s) < f1 (s ) and f2 (s) ≥ f2 (s ) (11)

f1 (s) ≤ f1 (s ) and f2 (s) > f2 (s ) (12)

A solution s strictly dominates s if

f1 (s) < f1 (s ) and f2 (s) > f2 (s ) (13)

Furthermore, a solution s ∈ S is called weakly efficient if there is no other

solution s ∈ S which strictly dominates s . Meanwhile, a solution s is called
efficient if there is no other solution s which dominates s . If we project solutions
onto an x-y plane where the x-axis refers to f1 and the y-axis to f2 , the resulting
point of a (weakly) efficient solution is called a (weakly) non-dominated point.
The goal of solving a bi-objective problem is typically that of producing the
set of all (weakly) efficient solutions for which the set of all points in the plane
is called the (weakly) non-dominated frontier. This frontier shows the tradeoffs
between the two objective functions while the resulting set of solutions provides
decision makers a number of options to choose from, rather than a single solution
which is often the case for single-objective problems. The next section describes
an algorithm to produce the set of weakly efficient solutions.

4 A Radius-Based Incremental Algorithm

The proposed Radius-based Incremental Algorithm (RIA) is based on the incre-
mental algorithm from Tutunchi and Fathi [20]. The main difference between
540 N. De Walsche et al.

these two methods lies on how we approached the solution of subproblems. Algo-
rithm 1 provides an overview of RIA.
In lines 1–2, the pC and pD associated with the BpCD instance are solved to
optimality. This is done by employing the algorithms described in Sects. 4.1 and
4.2, respectively. Their results provide 2 sets of bounds for the subsequent incre-
mental algorithm. Bound [cmin , cmax ] on pC objective, and bound [lmin , lmax ] on
pD objective. Note that cmin is the optimal pC objective and lmax the optimal
pD objective. Meanwhile, cmax is the center value obtained from the optimal pD
solution. Similarly, lmin is the dispersion value obtained from the optimal pC
solution. In other words: the optimal solutions of pC and pD form the extreme
points in our (weakly) non-dominated frontier.
After these extreme points are obtained, RIA continues to the main loop
spanning lines 4–7. In each iteration of this loop, another (weakly) efficient solu-
tion is generated in line 5. The MCD-Algorithm is used to solve a pC formulation
with an added constraint on the minimum dispersion required in the optimal
solution. This dispersion is defined by means of parameter l, which changes in
each iteration of the loop. A full description of the MCD-Algorithm can be found
in Sect. 4.3. The algorithm returns both the optimal pC objective given a min-
imum dispersion l, denoted by CenterV al as well as the dispersion DispV al
associated with this solution. Note that DispV al ≥ l. In line 6, the minimum
center bound cmin is updated to speed up subsequent MCD-Algorithm execu-
tions. The dispersion level l is updated in line 7 with the minimum value in Q
that is larger than DispV al. The main loop is executed as long as l ≤ lmax .

Algorithm 1. Radius-based Incremental Algorithm

Require: Network G = (N, E), distance matrix Dc and Dd , and parameter p
1: cmin , lmin ← MC-Algorithm(G, Dc , p)
2: cmax , lmax ← MD-Algorithm(G, Dd , p)
3: l ← lmin
4: while l ≤ lmax do
5: CenterV al, DispV al ← MCD-Algorithm(G, Dc , Dd , p, l, cmin , cmax )
6: cmin ← CenterV al
7: l ← min q ∈ Q : q > DispV al

4.1 MC-Algorithm
In order to solve the pC problem, we use formulation MC which is defined as
min (1)
s.t. (3) − (5), (8), (10)
Optimal solutions for MC are obtained by means of the double bound algorithm
DBR2 [3], which decomposes pC into a series of restrictive feasibility subprob-
lems. It then simply searches for the smallest radius with an associated feasible
solution. To speedup solving pC even further, we utilize the quick approximation
algorithms from [3] to obtain lower- and upper-bounds for pC. This in turn can
potentially reduce the number of radii that must be considered by DBR2.
 A Radius-Based Approach for BpCD 541

4.2 MD-Algorithm
For solving the pD problem, we utilize model MD which is defined as

max (2)
s.t. (3), (6), (7), (9), (10)

We again exploit the structure of the problem to solve pD as a series of restric-

tive feasibility subproblems by means of DBR2. Here, we employ the 1/2-
approximation algorithm introduced by Ravi et al. [17] for pD. This provides
us with lower- and upper-bound for the problem in a similar fashion as for the
pC [3], which speeds up the solving time. Finally, we make the necessary modi-
fications to the DBR2 for solving a maximization problem.

4.3 MCD-Algorithm
The MCD-Algorithm aims solving model MCD which is defined as follows

min (1)
s.t. (3) − (5), (8), (10)
yi + y j ≤ 1 ∀ i, j ∈ J : dij < l, i < j (14)

where l is a parameter representing the minimum level of dispersion required in

the optimal solution of pC. By varying l and solving different MCDs, we can
successfully produce (weakly) efficient solutions for each dispersion level.
Unlike MC- and MD-Algorithm, DBR2 is not used to solve the MCD model.
Instead we devise a similar algorithm which exploits the fact that the pC objec-
tive value will likely be close to the pC value found in a previous execution of
the MCD-Algorithm. The rationale is simple. Suppose that during an iteration
of RIA an objective value of c is found. If the solution is weakly efficient then
the objective value of the next iteration is likely c as well. Since we do not know
whether the solution found is efficient, it makes sense to first consider that the
optimal solution for the model at an increased dispersion level is also c. Once we
start solving model MCD by decomposition to a set cover problem, we first check
whether c leads to a feasible set cover. If so, we can stop the search prematurely
and save significant computational effort.
Based on these observations, we developed an alternative decomposition algo-
rithm which proved to be more efficient for use within the MCD phase in RIA.
This algorithm employs an ordered list of radii T containing all radius values
between cmin and cmax . In this ordering, a radius value rk ∈ T is optimal for
MCD with dispersion level l if rk−1 is an infeasible radius. It is therefore easy to
see how we can quickly obtain an optimal solution by employing binary search,
DBR2 or similar methods. With that in mind, our decomposition algorithm can
be defined with three simple steps.
Step 1 - Find a Feasible Solution. We begin by searching for a radius which
leads to a feasible MCD solution for the given dispersion l. This is achieved by
542 N. De Walsche et al.

starting with an index k = 0 and then repeatedly trying other radius values from
the ordered list T until a feasible one is found. Note that k = 0 is associated
with the optimal radius from the previous RIA iteration (current cmin value).
Whenever the k-th radius fails to provide a feasible solution, we increase k by
k ← k + α. Value α denotes the step size. After each k update, we also update
the step size as α ← 2α (initially we set α = 4). Step 1 terminates as soon as a
feasible radius rk ∈ T is found. We will denote by k  the last k before we found
a feasible solution (if k = 0 then k  = 0).
Step 2 - Reduce Range of Bounds. Once we have determined a feasible
radius rk for the current dispersion level, we try to find a small range of radii
within which the optimal solution exists. This is done by means of a straight-
forward binary search over T . In the binary search, we use two indices: a = k 
and b = k. We then proceed by trying the middle point m = (a + b)/2 and
updating either a or b accordingly. This is done so long as a − b > 4.
Step 3 - Find Optimal Value. Finally, we search for the optimal value, which
we know lies within the (index) range [a, b]. Note that ra ∈ T is infeasible from
Steps 1 and 2. Therefore, if we try ra+1 and it turns out to be feasible, we may
quickly end the search (ra+1 is optimal). Otherwise, we try rb−1 and if it is
infeasible, we can also stop the search (rb is optimal). If neither of the previous
situations applies, we check ra+2 . If feasible, then the optimal solution is ra+2 ,
otherwise it must be rb−1 (recall that the range [a, b] has at most five values
because of Step 2).

4.4 Matrix Reduction Rules

In order to speed up RIA, we employ matrix reduction rules which can reduce
the size of the models. This is done whenever we solve a feasibility subproblem
to verify whether a given radius is feasible, for example when employing DBR2,
or our special procedure from Sect. 4.3.
Consider an arbitrary iteration of the MC- or MCD-Algorithm where we are
solving a feasibility problem for radius value rk to verify whether it is a feasible
pC objective. Recall that I denotes the set of demand nodes and J the set of
potential facilities. We construct a matrix A0 as follows.

A0 = [a0ij ] where a0ij = 1 if dij ≤ rk for i ∈ I and j ∈ J (15)

This implies that matrix A0 has a row corresponding to each demand point and
a column corresponding to each potential facility.
Consider an arbitrary iteration of the MD- or MCD-Algorithm, where the
radius value is equal to l for the pD objective. We construct a matrix B 0 as
follows.
B 0 = [b0jm ] where b0jm = 1 if djm ≥ l for j, m ∈ J (16)
This implies that matrix B 0 has a row and column corresponding to each poten-
tial facility. Assuming that Dd is symmetric, B 0 will also be symmetric.
 A Radius-Based Approach for BpCD 543

With the matrices defined, we can employ the following reduction rules
sequentially and repetitively, where relevant and potentially beneficial. One of
the consequences of these rules is removing potential facility locations. Suppose
we know that node m can be excluded, we then update J as J \{m} which reduces
the number of y variables. Where applicable, this change can be reflected in the
matrices as follows. Column m from matrix Ah can be removed to obtain Ah+1
and both row and column m from matrix B h can be removed to obtain B h+1 .
Rule 1 - Center Fixing. Applicable to solve MC and MCD. Suppose for any
row i ∈ I and center matrix Ah the following condition holds:

ahij = 1 (17)
j∈J

This implies that demand node i can only be covered by a single facility j 1
within the given radius rk . Therefore, j 1 should be selected as facility location
in every feasible solution of the problem with radius rk . This can be ensured
by setting yj 1 = 1. In BpCD it also implies that any facility m ∈ J for which
bhj1 m = 0 holds should be excluded. Stated differently, because facility j 1 has to
be selected, any facility m within the dispersion threshold l can not be selected.
Rule 2 - Coverage Feasibility. Applicable to solve MC and MCD. Suppose
for any row i ∈ I and center matrix Ah the following condition holds:

ahij = 0 (18)
j∈J

This implies that demand node i cannot be covered by any facility within a given
radius value rk . Therefore, the problem is not feasible for that radius value (rk )
and will have to be increased. In BpCD facility coverage is not only dependant
on rk but also on the dispersion threshold l, consequently feasibility can be
recovered by increasing rk or lowering l.
Rule 3 - Dispersion Suitability. Applicable to solve MD and MCD. Suppose
for some j ∈ J the following condition holds:

bjm < p − 1 (19)
m∈M

This implies that selecting j as a facility will make the problem infeasible for the
given dispersion value (ql ) since it will not be possible to select p−1 other centers
that are distant enough from j. Therefore, one can safely exclude j from J.
Rule 4 - Dominance. Applicable to solve MC and MCD. Suppose for any
column pair j, m ∈ J the following condition holds:
For each i ∈ I, ahij ≥ ahim (20)
The condition implies that j can serve every demand point that m can and maybe
more. Therefore, one can safely exclude m from the set of potential facilities since
every solution that includes m can instead include j.
544 N. De Walsche et al.

For MD and MCD we can similarly state the following. Suppose for any
column pair j, m ∈ J with bjm = 0 the following conditions hold:

For each i ∈ J \ {j, m}, bij ≥ bim (21)

The condition implies that any potential facility respecting the dispersion thresh-
old when m is selected will remain a potential facility if j is selected instead.
Therefore, one can safely exclude m from J.
While Condition (20) can be applied to MC, and Condition (21) to MD, both
conditions must be simultaneously verified to exclude m when solving MCD.

5 Computational Study
In this section we evaluate the performance of RIA to solve BpCD using instances
from both the OR-Library [1] and the TSPLIB [18]. We implemented RIA in Java
(OpenJDK 19) using CPLEX 22.1 to solve the mathematical programming for-
mulations. All experiments were conducted in a computer with an AMD Ryzen
7 5800X processor at 3.8GHz with 8 cores, 16 GB of RAM and Windows 10
operating system. Furthermore, we have imposed a time limit of 2h per instance
following the settings of Tutunchi and Fathi [20].
Table 2 reports results over the 40 pmed instances from the OR-Library. Col-
umn Instance provides the name of the instance, while columns |N | and p con-
tain the number of nodes in the input graph and the number of facilities that
should be placed, respectively. Columns TC and TD report the execution time,
in seconds, required to solve the first pC and pD, which provide bounds for the
subsequent incremental algorithm. Finally, columns Iter. and TT otal report the
number of iterations of the incremental algorithm and the total execution time in
seconds to solve the BpCD defined by the given instance. Note that the number
of iterations corresponds to the number of weakly efficient solutions. In other
words: it corresponds to the size of the final pareto front.
The first observation we can make from Table 2 is that RIA was able to solve
every OR-Library instance within less than 1600 s seconds of execution time.
Although a precise comparison is not possible due to technological differences,
this brings a great contrast to the results of Tutunchi and Fathi [20] where
instances with |N | = 300 required 1600 s seconds on average to be solved. Fur-
thermore, solving the initial pC and pD problems never required more than 30 s.
Indeed, most of the time is spent in MCD-Algorithm executions where some dis-
persion limits l lead to difficult subproblems. For example, RIA solved the initial
pC and pD for instance pmed40 in 23 s, while the 19 iterations of the incremental
algorithm required, in total, 1500 s seconds.
Let us now turn our attention to larger instances from TSPLIB to verify
the scalability of RIA. Table 3 reports the results for a set of selected TSPLIB
instances which are instances frequently utilized in the pC and pD literature.
The Euclidean distance values are rounded up to the nearest integer. All of
the selected instances contain |N | > 1000 locations. We solved each instance for
p ∈ {5, 10, 20, 40, 100, 200}. The selection of p values was based on the results
 A Radius-Based Approach for BpCD 545

Table 2. Results for pmed instances from OR-Library.

Instance |N | p TC TD Iter. TT otal

pmed1 100 5 0.2 <0.1 22 3.32
pmed2 100 10 0.1 <0.1 31 3.57
pmed3 100 10 <0.1 0.3 27 3.37
pmed4 100 20 <0.1 <0.1 31 1.20
pmed5 100 33 <0.1 <0.1 21 0.49
pmed6 200 5 0.4 <0.1 36 22.28
pmed7 200 10 0.1 0.2 40 24.42
pmed8 200 20 <0.1 0.5 38 19.15
pmed9 200 40 <0.1 0.2 31 3.23
pmed10 200 67 <0.1 <0.1 17 0.63
pmed11 300 5 <0.1 <0.1 33 28.63
pmed12 300 10 0.2 <0.1 43 62.11
pmed13 300 30 0.1 0.3 47 44.68
pmed14 300 60 <0.1 0.3 26 5.79
pmed15 300 100 <0.1 <0.1 20 0.75
pmed16 400 5 0.1 <0.1 39 47.40
pmed17 400 10 0.4 <0.1 47 119.30
pmed18 400 40 0.2 3.7 37 90.33
pmed19 400 80 0.1 0.5 25 7.15
pmed20 400 133 <0.1 <0.1 14 1.23
pmed21 500 5 0.6 <0.1 37 91.71
pmed22 500 10 1.3 0.1 35 144.36
pmed23 500 50 0.3 13.3 28 145.09
pmed24 500 100 0.2 1.1 21 6.91
pmed25 500 167 0.1 <0.1 13 1.62
pmed26 600 5 0.5 <0.1 27 115.97
pmed27 600 10 0.7 <0.1 44 240.65
pmed28 600 60 0.3 6.6 25 105.39
pmed29 600 120 0.2 1.1 21 16.09
pmed30 600 200 0.2 0.1 14 3.05
pmed31 700 5 1.3 <0.1 31 217.85
pmed32 700 10 3.0 <0.1 35 408.70
pmed33 700 70 0.6 11.5 23 316.41
pmed34 700 140 0.4 1.8 15 17.79
pmed35 800 5 0.3 <0.1 28 188.42
pmed36 800 10 2.4 <0.1 37 549.71
pmed37 800 80 0.6 7.4 22 631.15
pmed38 900 5 0.5 <0.1 17 122.80
pmed39 900 10 1.5 <0.1 28 516.80
pmed40 900 90 0.7 22.3 19 1516.84
546 N. De Walsche et al.

of [3] so that we have different degrees of difficulty when solving BpCD. It is

known that extreme values of p (too small or large) have a tendency of being
easy to solve for pC and pD, while values in-between typically make the instance
much harder to solve. Entries with a – denote cases for which RIA reached the
2h time limit and full execution could not be completed. Best lower- and upper-
bounds obtained are reported for those instances under columns CLB , CU B and
DLB , DU B for pC and pD, respectively.

Table 3. Results for selected TSPLIB instances.

Instance |N | p CLB CU B TC DLB DU B TD Iter. TT otal

u1060 1060 5 3456 3456 2.8 6862 6862 1.2 33 1235.69
u1060 1060 10 2273 2273 3.5 4478 4478 1.0 56 763.94
u1060 1060 20 1581 1581 7.0 2865 2865 1.8 56 434.30
u1060 1060 40 1021 1021 2.5 1758 1758 6.9 68 595.76
u1060 1060 100 570 570 1.3 944 944 2.2 59 55.74
u1060 1060 200 360 360 1.2 566 566 1.0 33 7.58
rl1323 1323 5 4543 4543 6.0 9719 9719 2.2 65 4509.53
rl1323 1323 10 3077 3077 8.4 5865 5865 2.4 58 2332.50
rl1323 1323 20 2016 2016 5.8 3612 3612 8.4 70 2159.99
rl1323 1323 40 1352 1352 9.9 2352 2352 34.5 86 3460.95
rl1323 1323 100 787 787 3.8 1270 1270 8.4 115 328.94
rl1323 1323 200 518 518 2.5 783 783 1.1 118 37.58
u1817 1817 5 715 715 7.6 1535 1535 3.4 76 –
u1817 1817 10 458 458 16.0 881 881 5.5 62 –
u1817 1817 20 309 309 97.1 559 559 167.2 40 5319.61
u1817 1817 40 209 209 503.6 364 364 516.8 16 –
u1817 1817 100 127 127 2.1 205 205 16.5 28 3449.97
u1817 1817 200 80 80 3.0 130 130 17.6 16 71.54
pr2392 2392 5 3827 3827 18.7 8086 8086 7.6 32 –
pr2392 2392 10 2581 2581 60.2 4976 4976 11.7 44 –
pr2392 2392 20 1736 1736 1693.1 3150 3150 1176.7 4 –
pr2392 2392 40 1173 1193 – – – – – –
pr2392 2392 100 708 708 4069.6 1204 963 – – –
pr2392 2392 200 481 481 49.7 781 781 2059.7 90 –
d15112-modif-2500 2500 5 5856 5856 27.8 12217 12217 11.2 40 –
d15112-modif-2500 2500 10 3705 3705 27.8 7132 7132 5.0 62 –
d15112-modif-2500 2500 20 2573 2573 237.5 4773 4773 125.4 11 –
d15112-modif-2500 2500 40 1713 1726 – – – – – –
d15112-modif-2500 2500 100 1050 1090 – – – – – –
d15112-modif-2500 2500 200 723 723 2472.8 1146 1119 – – –
 A Radius-Based Approach for BpCD 547

The increased size of the TSPLIB instances poses a challenge to our method.
It is clear from Table 3 that solving the BpCD takes much longer for these
instances, even if the times required to solve the initial pC and pD are still rela-
tively short (most times are under 60 s). Total execution times grow significantly
and in some cases they are close to, or exceed, 1h. These TSPLIB instances also
have larger sets of weakly efficient solutions (Iter.) compared to OR-Library
instances with differences as high as five times. However, this does not seem to
be the cause of increased computation times since larger Iter. values are not
necessarily correlated to longer total execution times.
Another interesting observation from Table 3 is that while solving instances
for p ∈ {5, 10, 20, 40} is difficult, the time to solve the same instance with
p ∈ {100, 200} is often much shorter. This corroborates what Çalık and Tansel [3]
observed for pC, but here we notice the same behavior in the pD and BpCD.
However, while pC and pD can be quickly solved for p ∈ {5, 10}, the same is not
observed when solving BpCD. Indeed, for instance u1817, RIA could not finish
execution for p ∈ {5, 10} while it finished for p = 20. The reason being that some
executions of the MCD-Algorithm took longer for these values of p and certain
dispersion levels, leading RIA to exceed the time limit of 2h.
The positive contribution of the strong lower- and upper-bounds to improve
the efficiency of the radius-based algorithms were confirmed in our computa-
tional study. We also observed a significant speed gain by utilizing the matrix
reduction rules. More specifically, for MCD-Algorithm itself, these rules provided
an average improvement of 22 s and a maximum improvement of 186 s when solv-
ing pmed instances. Employing a list of bitsets for the matrices instead of lists of
Boolean variables played a very important role in this significant improvement
with up to 100 times speed improvement.

6 Conclusion
This study reinforced the efficiency of radius-based formulations and exact
decomposition algorithms based on their restrictive subproblems for solving a
bi-objective location problem where the two objectives have minimax and max-
imin structures (BpCD). The efficient approximation algorithm available for the
p-center (pC) and p-dispersion (pD) problems proved effective in largely reducing
the search spaces for the decomposition algorithms for each individual problem.
Another important improvement component which can be utilized in methods
for solving similar pC and pD variants consists of the matrix reduction rules
which were effectively employed in multiple steps of our radius-based incremen-
tal algorithm (RIA).
Despite our improvements, there are still limitations concerning the size of
BpCD instances which can be effectively solved within a short computation time.
The state-of-the-art methods for both pC and pD can tackle instances with
tens of thousands of nodes. Meanwhile, RIA is the first approach to successfully
breach the one thousand nodes threshold for BpCD, which indicates that there is
a lot of room for additional research in trying to further reduce the gap between
the single- and the bi-objective problems.
548 N. De Walsche et al.

Possible options for future research include exploring a change in the main
formulation within the incremental algorithm. While both our RIA and Tutunchi
and Fathi [20] employed the pC formulation with an additional dispersion con-
straint, how would the performance be impacted if we employed the pD for-
mulation with additional center constraints? Are there instances for which one
formulation outperforms the other? Another open question refers to a capaci-
tated version of BpCD where facilities have an associated maximum capacity.
Although the radius-based formulations are still valid, they require additional
allocation variables which increases the size of the models. Hence, a natural
question appears: can medium- and large-scale instances still be solved exactly
when facilities have capacities?

Acknowledgments. The authors acknowledge the support by Internal Funds KU

Leuven and the Research Foundation Flanders (FWO) through the Strategic Basic
Research project Data-driven logistics (S007318N).

References
1. Beasley, J.E.: OR-Library: distributing test problems by electronic mail. J. Oper.
Res. Soc. 41(11), 1069–1072 (1990). https://doi.org/10.2307/2582903
2. Çalık, H., Labbé, M., Yaman, H.: p-center problems. Location Sci. 51–65 (2019).
https://doi.org/10.1007/978-3-319-13111-5_4
3. Çalık, H., Tansel, B.C.: Double bound method for solving the p-center location
problem. Comput. Oper. Res. 40(12), 2991–2999 (2013). https://doi.org/10.1016/
j.cor.2013.07.011
4. Contardo, C.: Decremental clustering for the solution of p-dispersion problems to
proven optimality. INFORMS J. Optim. 2(2), 134–144 (2020). https://doi.org/10.
1287/ijoo.2019.0027
5. Contardo, C., Iori, M., Kramer, R.: A scalable exact algorithm for the vertex
p-center problem. Comput. Oper. Res. 103, 211–220 (2019). https://doi.org/10.
1016/j.cor.2018.11.006
6. Daskin, M.: Network and Discrete Location: Models, Algorithms and Applications,
2nd edn. Taylor & Francis, Milton Park (2013)
7. Ehrgott, M., Gandibleux, X., Przybylski, A.: Exact methods for multi-objective
combinatorial optimisation. In: Greco, S., Ehrgott, M., Figueira, J.R. (eds.) Mul-
tiple Criteria Decision Analysis. ISORMS, vol. 233, pp. 817–850. Springer, New
York (2016). https://doi.org/10.1007/978-1-4939-3094-4_19
8. Elloumi, S., Labbé, M., Pochet, Y.: A new formulation and resolution method for
the p-center problem. INFORMS J. Comput. 16(1), 84–94 (2004). https://doi.org/
10.1287/ijoc.1030.0028
9. Erkut, E.: The discrete p-dispersion problem. Eur. J. Oper. Res. 46(1), 48–60
(1990). https://doi.org/10.1016/0377-2217(90)90297-O
10. Garfinkel, R.S., Neebe, A.W., Rao, M.R.: The m-center problem: minimax facility
location. Manage. Sci. 23(10), 1133–1142 (1977). https://doi.org/10.1287/mnsc.
23.10.1133
11. Gonzalez, T.F.: Clustering to minimize the maximum intercluster distance. Theor.
Comput. Sci. 38, 293–306 (1985). https://doi.org/10.1016/0304-3975(85)90224-5
 A Radius-Based Approach for BpCD 549

12. Hakimi, S.L.: Optimum locations of switching centers and the absolute centers and
medians of a graph. Oper. Res. 12(3), 450–459 (1964). https://doi.org/10.1287/
opre.12.3.450
13. D Kariv, O., Hakimi, S.L.: An algorithmic approach to network location problems.
Part I: the p-centers. SIAM J. Appl. Math. 37(3), 513–538 (1979). https://doi.
org/10.1137/0137040
14. Kramer, R., Iori, M., Vidal, T.: Mathematical models and search algorithms for
the capacitated p-center problem. INFORMS J. Comput. 32(2), 444–460 (2020).
https://doi.org/10.1287/ijoc.2019.0889
15. Kuby, M.J.: Programming models for facility dispersion: the p-dispersion and max-
isum dispersion problems. Geogr. Anal. 19(4), 315–329 (1987). https://doi.org/10.
1111/j.1538-4632.1987.tb00133.x
16. Minieka, E.: The m-center problem. SIAM Rev. 12(1), 138–139 (1970). https://
doi.org/10.1137/1012016
17. Ravi, S.S., Rosenkrantz, D.J., Tayi, G.K.: Heuristic and special case algorithms for
dispersion problems. Oper. Res. 42(2), 299–310 (1994). https://doi.org/10.1287/
opre.42.2.299
18. Reinelt, G.: TSPLIB-A traveling salesman problem library. ORSA J. Comput.
3(4), 376–384 (1991). https://doi.org/10.1287/ijoc.3.4.376
19. Sayah, D., Irnich, S.: A new compact formulation for the discrete p-dispersion
problem. Eur. J. Oper. Res. 256(1), 62–67 (2017). https://doi.org/10.1016/j.ejor.
2016.06.036
20. Tutunchi, G.K., Fathi, Y.: Effective methods for solving the bi-criteria p-center and
p-dispersion problem. Comput. Oper. Res. 101, 43–54 (2019). https://doi.org/10.
1016/j.cor.2018.08.009
 Author Index

A H
Andersson, Henrik 292 Haug, Helle V. 170
Azhar, Nurul Asyikeen Binte 500 Heilig, Leonard 202
Høyland, Sondre 292
B
Bak, Bendegúz Dezső 450 J
Beirigo, Breno A. 381 Jensen, Rune Møller 105
Buscher, Udo 350 Joe, Waldy 18
Johnsen, Lennart C. 416
Jovanovic, Raka 219
C
Çalık, Hatice 533 K
Cammin, Philip 139 Karami, Farzaneh 36
Cattrysse, Dirk 248 Kascak, Nicholas 56
Cheng, Shih-Fen 500 Kerkenhoff, Tim 36
Kesen, Saadettin Erhan 437
D Kortekaas, Joris J. A. 381
Daduna, Joachim R. 396 Krogstad, Jesper Anker 292
Dávila-Gálvez, Sebastián 517 Kucukoglu, Ilker 248
De Walsche, Niels 533
Dieter, Peter 235 L
Lalla-Ruiz, Eduardo 276
E Landa-Silva, Dario 467
Ehsanfar, Ebrahim 36 Lau, Hoong Chuin 18
El Achhab, Abdelhak 337 Leonardi, Erwin 500
El Amrani, Mohammed 337 Li, Mingyu 170
Eley, Michael 482 Linß, Florian 307
Er, Yuyang 18 Liu, Yijia 85

M
F Maaji, Salim Sulaiman 467
Fagerholt, Kjetil 170, 185 Madani, Batool 322
Flatberg, Truls 292 Main, Alastair Ronald 122
Földes, Dávid 450 Maktabifard, Ali 450
Meisel, Frank 185, 416
G Mes, Martijn 276
Gana, Lucas 517
Grbic, Djordje 105 N
Gullhav, Anders N. 292 Ndiaye, Malick 322
Gunawan, Aldy 260, 500 Nestler, Alexander 56

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2023
J. R. Daduna et al. (Eds.): ICCL 2023, LNCS 14239, pp. 551–552, 2023.
https://doi.org/10.1007/978-3-031-43612-3
552 Author Index

Neufeld, Janis S. 350 Scholz, Sebastian 416

Neugebauer, Julian 202 Schulte, Frederik 154, 219, 381
Nguyen, Dang Viet Anh 260 Shi, Xiaoning 85
Nguyen, Pham Kien Minh 260 Solum, Sigrid H. 170
Norstad, Inge 170 Soth, Marvin 416
Nourmohammadzadeh, Abtin 365
T
O Tang, Xinyu 154
Omholt-Jensen, Simen 185
V
P van Heeswijk, Wouter 276
Pacino, Dario 122 van Steenbergen, Robert M 276
Päprer, Paul 350 van Twiller, Jaike 105
Paternina-Arboleda, Carlos 56 Vansteenwegen, Pieter 248, 260
Philipp, Daniel 396 Voß, Stefan 3, 139, 202, 365
Pour, Morteza Safaei 56
W
Q Warholm, Sanna B. 170
Quezada, Franco 517 Wenzel, Peter 219

R Y
Rodrigues, Filipe 122 Yağmur, Ece 437
Yu, Jingjing 139
S
Sarhani, Malek 337 Z
Sartori, Carlo S. 533 Zhang, Zhiqin 18