0% found this document useful (0 votes)

112 views355 pages

(Book) Pro-Active Dynamic Vehicle Routing

Uploaded by

Aditya Restuaji

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

112 views355 pages

(Book) Pro-Active Dynamic Vehicle Routing

Uploaded by

Aditya Restuaji

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 355

Contributions to Management Science

For further volumes:

www.springer.com/series/1505
Francesco Ferrucci

Pro-active Dynamic
Vehicle Routing

Real-Time Control and

Request-Forecasting Approaches
to Improve Customer Service
Francesco Ferrucci
WINFOR–Business Computing
and Operations Research
University of Wuppertal
Wuppertal, Germany

Additional material to this book can be downloaded from http://extras.springer.com.

ISSN 1431-1941 Contributions to Management Science
ISBN 978-3-642-33471-9 ISBN 978-3-642-33472-6 (eBook)
DOI 10.1007/978-3-642-33472-6
Springer Heidelberg New York Dordrecht London

Library of Congress Control Number: 2013932351

Keywords: Perishable goods, Dynamic vehicle routing, Real-time control, Request forecasting, Stochas-
tic knowledge, Past request information, Tabu Search, Request data quality

© Springer-Verlag Berlin Heidelberg 2013

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. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
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.
While the advice and information in this book are believed to be true and accurate at the date of pub-
lication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.

Printed on acid-free paper

Physica-Verlag is a brand of Springer

Springer is part of Springer Science+Business Media (www.springer.com)
You can see far because you are standing
on the shoulders of giants.

(Inspired by Bernard of Chartres)

I dedicate this book to my mother and
my grandmother
Foreword

Obtaining timely transportation processes is increasingly becoming a crucial fac-

tor in modern supply chains. By implementing nearly bufferless processes, produc-
tion systems strongly depend on a timely delivery of required intermediate prod-
ucts or materials. Since possible delays may result in costly outages, implementing
efficient transportation processes is crucial. Additionally, ensuring the loyalty of
end-consumers requires fulfilling of their rising expectations on a fast order de-
livery. This applies particularly to scenarios where these orders comprise goods
that are perishable and/or are already subsequent deliveries. Besides the existing
urgency, various sources of dynamism, which are present in real-world distribution
processes, make a timely and efficient implementation difficult. Consequently, the
development of suitable adaptive real-time control approaches is a vital and cur-
rent research topic in transportation science. Here, pure deterministic approaches
are distinguished from approaches that additionally integrate stochastic knowledge
about the probability that future requests or events will occur. Since deterministic
approaches only adapt existing tour plans according to already known requests, they
are frequently denoted as reactive real-time concepts. In contrast, the exploitation of
existing stochastic information about the arrival of future requests additionally en-
ables a pro-active guidance of vehicles to request-likely areas. The PhD thesis writ-
ten by Francesco Ferrucci provides significant contributions to this vital research
field. It was mainly motivated by real-world subsequent delivery processes of news-
papers. Due to the high perishability of the transported goods, these distribution
processes have to be executed under considerable urgency. Starting from the appli-
cation of a pure deterministic real-time approach that already enables an efficient
execution of dynamic transportation processes, this research work analyzes whether
and in particular under which circumstances the integration of stochastic knowledge
about expected future request arrivals leads to further improvements. For this pur-
pose, a general cluster scheme for exploiting existing past request information in
an offline step is proposed. Subsequently, the resulting clusters are transformed into
so-called dummy customers with specific time window constraints. These artificial
requests are integrated into the tour plans by applying the deterministic real-time
approach in order to allow for more efficient and robust tour plans. In contrast to

ix
x Foreword

earlier approaches that assume the availability of distributions which can be defined
in a closed form, the utilized stochastic knowledge is systematically attained by the
offline application of a sophisticated cluster analysis that integrates specific quality
restrictions. Therefore, the proposed method can be applied to a larger number of
real-world applications.
In order to analyze the performance of the proposed concept, extensive computa-
tional experiments are conducted. The results of these experiments strongly indicate
that the integration of the generated dummy customers is promising when a mini-
mum temporal as well as spatial diversity of the past request data is given. This
is impressively shown by several specifically designed evaluations. On top of that,
besides the degree of structural distortion, the degree of structural diversity (dosd)
is introduced as a new measure that allows an offline prediction as to whether the
integration of stochastic knowledge is useful for a given past request data set, or not.
Specifically, a strong correlation between the dosd and the improvements that can
be yielded is statistically proven. Furthermore, the PhD thesis provides a detailed
analysis of the impact of allowed vehicle en-route diversions. In contrast to the lit-
erature, where vehicle en-route diversion is either completely allowed or forbidden,
this research work analyzes intermediate scenarios in detail. Here, due to the in-
creased stress level of the drivers that may lead to additional costs, the allowance of
an en-route diversion requires a minimum cost reduction. On the basis of extensive
computational results, this important trade-off is analyzed in detail for the first time.
Having known Francesco for a long time and intensively worked with him on
sophisticated real-time control concepts for vehicle routing, it is not at all surprising
to me that his PhD thesis is one in which I have taken great pleasure and interest
to read. He has always been very ambitious and zealous in trying to understand
research problems thoroughly and has proven to possess the necessary endurance
and scientific understanding to finally obtain the needed insights into the underly-
ing problem structure. Moreover, he reprograms the source codes in order to inte-
grate new ideas immediately. Therefore, after discussing new concepts and possible
workarounds late in the evening, you may receive an email early in the morning in-
forming you about their performance in recent computational tests. All in all, I can
summarize these fruitful years by just saying that it was such a pleasure for me to
be able to supervise him as a member of my team. I am convinced beyond doubt
that we will keep in touch and cooperate in further practical and scientific projects
in future.
Wuppertal, Germany Professor Stefan Bock
July 2012
Preface

In many business relationships, unexpectedly occurring events lead to unfavorable

situations which cause a high inconvenience or high costs for affected customers.
Two examples of such situations are the unsuccessful delivery of newspapers to sub-
scribers and the breakdown of important machines during production processes. In
such situations, affected customers demand urgent assistance from a service com-
pany which resolves the situation by delivering required goods which can be ma-
terial or immaterial, e.g., newspapers or repair services. In order to maintain cus-
tomer loyalty, the minimization of customer inconvenience which increases with
the required time for delivering the demanded goods (denoted as response time)
is increasingly becoming the most crucial objective in such processes. Due to this
urgent demand, these goods are considered as perishable goods in this book. In
real-world applications such as the mentioned subsequent delivery of newspapers
or in repairman companies operating under Service Level Agreements (SLAs), a
quick service of customer requests which dynamically arrive during the execution
of the transportation process is the highest priority. The considered applications are
denoted as the Real-time Distribution Of Perishable Goods (RDOPG).
In this book, three contributions are made to the field of dynamic vehicle rout-
ing. First, based on results that have been achieved by recent marketing research,
an appropriate dynamic routing model for RDOPG applications is formulated. In
this model, customer inconvenience arising along with the response time achieved
for occurring customer requests is considered and its minimization is pursued as
the main objective. Specifically, in this book, we evaluate the impact of a linear
and a quadratic evolution of customer inconvenience depending on response time
combined with a maximum allowed response time on the attained solution quality.
Moreover, a comprehensive introduction to tour planning is given and existing clas-
sifications for routing problems in the literature are presented. Based on these clas-
sifications, a new approach for classifying routing problems according to RDOPG-
specific attributes is proposed. The second contribution represents the main part of
this book. It evaluates the benefits that can be attained in the solution quality through
the integration of stochastic knowledge about expected future customer requests in
RDOPG applications. In order to evaluate these benefits, a new deterministic real-

xi
xii Preface

time control approach which centrally coordinates the transportation process during
its execution is proposed. This real-time control approach considers the concur-
rency of tour plan execution and tour plan adaptation. The transportation process is
adapted during its execution by solving individual problem instances utilizing a pro-
posed stage-based Tabu Search metaheuristic. Moreover, the mathematical model
used for representing the individual problem instances is introduced. The proposed
deterministic real-time control approach is then extended to a pro-active real-time
control approach which additionally makes use of stochastic knowledge about the
spatial and temporal occurrence of expected future customer requests. The utilized
stochastic knowledge is solely generated from past request information by a pro-
posed forecasting approach and no prescribed probability distributions are utilized.
Different quality criteria are considered in the proposed forecasting approach in or-
der to ensure a minimum defined quality of the generated stochastic knowledge. In
order to implement the pro-active real-time control approach, ways of handling and
integrating the stochastic knowledge into the solution process by using dummy cus-
tomers are proposed. According to the considered objective of minimizing customer
inconvenience, the stochastic knowledge is used to guide vehicles to request-likely
areas before real requests arrive there in order to service them quicker and thereby
attain a higher solution quality.
Comprehensive computational experiments are carried out on real-world request
data from a subsequent delivery of newspapers process (S REAL ) as well as on gen-
erated test scenarios (S GEN ). In all experiments, the effects of a variety of parameter
settings are evaluated in order to analyze the effects on the solution quality. In order
to identify scenarios in which the application of the proposed pro-active real-time
control approach achieves superior results compared to the proposed deterministic
real-time control approach, we introduce structural quality for classifying the value
of stochastic knowledge in a given routing problem scenario according to the avail-
able request data. Specifically, we derive the structural quality of a request data set
from two newly proposed criteria denoted as structural distortion and structural di-
versity. In addition to defining both criteria in detail, we also propose methods for
determining the structural distortion and the structural diversity of a given request
data set. Besides evaluating a proposed approach for determining structural distor-
tion on request data sets with different levels of structural distortion, we introduce a
two-dimensional structural diversity classification which is used to generate request
data sets of S GEN with significant differences in structural quality. Computational
results show that in scenarios with high levels of structural diversity, the proposed
pro-active real-time control approach is able to achieve significantly improved re-
sults compared to its proposed deterministic counterpart. Specifically, depending on
the evolution of customer inconvenience, improvements in the solution quality of up
to 26 % (linear2X customer inconvenience) and 47 % (quadratic customer incon-
venience) above the ones attained by the deterministic real-time control approach
can be achieved. Thanks to the proposed classification of structural diversity, dif-
ferences in results obtained among the tested scenarios become explainable. On
the S REAL request data, additional improvements of up to 7 % could be achieved by
using the pro-active real-time control approach. In order to explain the differences
Preface xiii

in the additionally attainable improvements on S GEN and S REAL , we performed a

visual analysis on both types of request data. In this analysis, further characteristics
of scenarios which support an efficient application of the pro-active real-time con-
trol approach could be identified. Moreover, we propose a measure denoted as the
degree of structural diversity (dosd) for determining the structural quality of given,
arbitrary request data sets. We show the effectiveness of the dosd by analyzing cor-
relations existing between the dosd and the improvements attained by the pro-active
real-time control approach on S GEN and S REAL . The results show that there ex-
ists a strong linear correlation between the proposed dosd and the attainable im-
provements. Therefore, the proposed structural diversity and the dosd provide good
starting points for assessing the structural quality of given request data sets which
supports an efficient application of stochastic knowledge in real-time control ap-
proaches under different scenarios. Moreover, the dosd also provides management
with decision support in efficiently customizing the proposed stochastic knowledge
generation approach in given scenarios using results achieved in the computational
experiments.
The third contribution of this book is a method for considering driver inconve-
nience aspects which arise from vehicle en-route diversion activities. Specifically,
a method is proposed which makes it possible to restrict the number of vehicle en-
route diversion activities performed by the real-time control approach by restricting
the allowed vehicle en-route diversion activities to those which improve the solution
quality by at least a defined threshold value. Computational experiments show the
efficiency of the proposed method. Specifically, by utilizing appropriate threshold
parameter values, driver inconvenience can be reduced by significantly lowering ve-
hicle en-route diversion activities and yet still achieving significant improvements
compared to completely forbidding vehicle en-route diversion activities. As also
shown in the second contribution, the pro-active real-time control approach pro-
vides a significantly higher solution quality in request data scenarios comprising
a high structural quality compared to the deterministic real-time control approach.
Even when all vehicle en-route diversion activities are completely forbidden, the
pro-active real-time control approach still achieves a significantly better solution
quality compared to the deterministic real-time control approach with all vehicle
en-route diversion activities allowed.
Wuppertal, Germany Francesco Ferrucci
July 2012
Acknowledgments

During my very exciting time as a research assistant from October 2006 to July 2012
at the Department of Business Computing and Operations Research (WINFOR) at
the University of Wuppertal in Germany, I gathered many terrific experiences in
science, practical research, and life. Motivated by the challenges in a real-world ap-
plication that we got to know from a company, I prepared my PhD thesis originally
entitled “Development and evaluation of pro-active real-time control and forecast-
ing approaches for the distribution of perishable goods utilizing past request infor-
mation to improve quality of customer service and reduce driver inconvenience” on
which this book is based.
My thesis was supervised by Professor Stefan Bock whom I have personally
known for more than 7 years now. While attending Stefan’s lectures during my
study time, I quickly realized that we both eagerly share the same enthusiasm for
developing and evaluating sophisticated algorithms and solution methods for real-
world problems. While writing my diploma thesis under his supervision, Stefan also
supported me with many creative ideas that greatly inspired me. Having known Ste-
fan as someone who is also always keen at discussing ideas and analyzing results in
detail, I decided that it would be a great idea to write my PhD thesis under his su-
pervision. During this time, I had the opportunity of gathering a lot of new expertise
as a teaching assistant, research assistant, project assistant, and project coordinator.
I would like to thank Stefan for insightfully supervising my thesis, always hav-
ing time for exchanging research ideas and also for being both a good lecturer and
friend. Thank you very much, Stefan, for all the possibilities you gave me during
my time writing my thesis. I have enjoyed working with you from the very first day.
One of the highlights during my PhD time was my three months’ visit to Mon-
treal in Canada. In addition to experiencing the beauty of Canada in winter time,
I also had the extraordinary possibility of jointly working on my thesis with Profes-
sor Michel Gendreau, one of the world’s most famous professors in transportation
science at the CIRRELT—the leading research institute for logistics and transporta-
tion. By turning my thesis into an international research project, I had a very en-
riching and fruitful collaboration with Michel. After my research time in Montreal,
I was allowed to give talks about our joint work with Michel at the international

xv
xvi Acknowledgments

conferences of INFORMS in San Diego, INFORMS ALIO in Buenos Aires, EURO

in Lisbon, and OR in Munich. Thanks, Michel, for all of your ideas and fruitful
advice. It has been an honor for me to work with you. Furthermore, I thank Ver-
ena Schmid who I first met in Canada where we had the unforgettable ability to
discover great places in Montreal together. She is always a valuable conversational
partner for scientific and non-scientific discussions.
I would also like to thank all of my colleagues at WINFOR for creating a friendly
as well as very productive work environment. The department has always been a
place with a relaxed atmosphere while working together with highly skilled col-
leagues who are also experts at table top soccer and other great activities. Thank you,
Volker Arendt, Rudolf Bauer, Benjamin Dahmen, Daniel Dittmar, Paul Göpfert, Si-
mon Görtz, Torben Kuschel, Christian Rathjen, and all others for great and helpful
discussions and the wonderful time. I would also like to thank Simon Görtz for his
valuable ideas we discussed in our scientific work group which have been greatly
helpful for my research work and thesis. In the same spirit, my thanks go to my Mas-
ter’s students Britta Kortmann, Sandra Klewer, and Jens Keune for providing high-
quality results which were very useful for my thesis. I also thank Professor Krumke
who provided us with access to his VRP solver. Special thanks also go to DDS Digi-
tal Data Services GmbH in Karlsruhe, Germany for providing me with the real road
network data for my research.
I am greatly indebted to Maria Muwanga for her great ideas and for her dedicated
support in polishing the English of my thesis turning it into a shiny piece of gold.
I would also like to thank my friends who additionally proofread my thesis and also
provided insightful remarks.
Finally, I would like to wholeheartedly thank my family and my friends. All their
love, support, kindness, and encouragement have helped me a lot in preparing my
thesis and mean more to me than words can say.

Wuppertal, Germany Francesco Ferrucci

July 2012
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Categorization of the Considered Problem . . . . . . . . . . . . . 7
1.3 Aims and Contributions of the Book . . . . . . . . . . . . . . . . 9
1.3.1 Central Research Questions to Be Answered in This Book . 9
1.3.2 Contributions of This Book . . . . . . . . . . . . . . . . . 9
1.4 Outline of the Book . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Introduction to Tour Planning: Vehicle Routing and Related Problems 15
2.1 General Task and Definitions . . . . . . . . . . . . . . . . . . . . 16
2.2 Representation of Requests and General Types of Routing Problems 18
2.3 Node-Based Routing Problems . . . . . . . . . . . . . . . . . . . 18
2.3.1 The Traveling Salesman Problem (TSP) . . . . . . . . . . 19
2.3.2 The Vehicle Routing Problem (VRP and CVRP) . . . . . . 20
2.3.3 The General Pickup and Delivery Problem (GPDP)
and Related Variants . . . . . . . . . . . . . . . . . . . . . 22
2.3.4 Complexity of Node-Based Routing Problems . . . . . . . 25
2.3.5 Solution Methods for Node-Based Routing Problems . . . 27
2.4 Extensions to the Vehicle Routing Problem . . . . . . . . . . . . . 32
2.4.1 Time Windows . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2 Backhauls . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.3 Simultaneous Pickup and Delivery . . . . . . . . . . . . . 33
2.4.4 Multi-Depot Problems . . . . . . . . . . . . . . . . . . . . 33
2.4.5 Open Routing Problems . . . . . . . . . . . . . . . . . . . 34
2.4.6 Vehicle Scheduling . . . . . . . . . . . . . . . . . . . . . 34
2.4.7 Multiple Compartments . . . . . . . . . . . . . . . . . . . 34
2.4.8 Multiple Trips . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.9 Time-Dependent Travel Times . . . . . . . . . . . . . . . 35
2.4.10 Legal Driving Time Regulations . . . . . . . . . . . . . . . 35
2.4.11 Heterogeneous Fleet . . . . . . . . . . . . . . . . . . . . . 35
2.4.12 Objective Functions in Vehicle Routing Problems . . . . . 36

xvii
xviii Contents

2.5 Information Revelation in Routing Problems . . . . . . . . . . . . 38

2.5.1 Revelation of Relevant Information in Vehicle Routing
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.2 Approaches in the Literature for Characterizing
Relevant Information and Distinguishing Between Static
and Dynamic Routing Problems . . . . . . . . . . . . . . . 40
2.5.3 A Unified Approach for Classifying Routing Problems
with Regard to Characteristics and Processing of Relevant
Information . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.6 Dynamic Routing Problems Which Utilize a Centralized
Coordination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.6.1 Typical Objectives in Dynamic Routing Problems . . . . . 59
2.6.2 Dynamic Events in Dynamic Routing Problems . . . . . . 59
2.6.3 Technologies Utilized in Real-Time Control Approaches . . 60
2.6.4 The Degree of Dynamism . . . . . . . . . . . . . . . . . . 64
2.6.5 Three-Echelon Classification of Dynamic Routing Problems 68
2.6.6 Evaluating the Performance of Solution Approaches
for Dynamic Routing Problems . . . . . . . . . . . . . . . 73
2.7 A General Classification Scheme for Routing Problems . . . . . . 74
2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3 The Considered RDOPG Applications . . . . . . . . . . . . . . . . . 81
3.1 Problem Description of RDOPG Applications . . . . . . . . . . . 81
3.1.1 Existence of a Real Urban Road Network . . . . . . . . . . 83
3.1.2 Presence of Vehicle On-Board Units and a Central
Dispatching Center . . . . . . . . . . . . . . . . . . . . . 83
3.1.3 Availability and Utilization of Past Request Information . . 83
3.1.4 Allowing Vehicle En-Route Diversion . . . . . . . . . . . 84
3.1.5 Integration of Vehicle Scheduling Decisions . . . . . . . . 84
3.2 Characteristics of RDOPG Applications . . . . . . . . . . . . . . 85
3.2.1 Differences from Other Dynamic Routing Problems . . . . 85
3.2.2 Computational Complexity . . . . . . . . . . . . . . . . . 86
3.2.3 Classifying RDOPG Applications with Regard to
Characteristics and Processing of Relevant Information . . 86
3.2.4 General Classification of RDOPG Applications . . . . . . . 87
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4 Review of the Literature Related to the Considered RDOPG
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.1 First Papers on Dynamic Routing Problems in the Literature . . . . 94
4.2 Selected Reactive Real-Time Control Approaches for Dynamic
Routing Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3 Strategies for Increasing Flexibility in Dynamic Routing Problems 97
4.3.1 Waiting Strategies . . . . . . . . . . . . . . . . . . . . . . 97
4.3.2 Relocation Strategies . . . . . . . . . . . . . . . . . . . . 99
4.3.3 Request Assignment Strategies . . . . . . . . . . . . . . . 100
Contents xix

4.4 Flexibility in Dynamic Routing Without Stochastic Knowledge . . 101

4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge . . . 108
4.5.1 Strategy-Oriented Approaches . . . . . . . . . . . . . . . . 109
4.5.2 Approaches Using Exact Solution Methods . . . . . . . . . 118
4.5.3 Stochastic Modeling Based Approaches . . . . . . . . . . 119
4.5.4 Approaches with Manually Provided Stochastic Knowledge 126
4.5.5 Sampling-Based Approaches . . . . . . . . . . . . . . . . 128
4.6 Related Objective Functions . . . . . . . . . . . . . . . . . . . . . 137
4.7 Other Relevant Factors for RDOPG Applications . . . . . . . . . . 140
4.7.1 En-Route Diversion of Vehicles . . . . . . . . . . . . . . . 140
4.7.2 Road Network . . . . . . . . . . . . . . . . . . . . . . . . 141
4.7.3 Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.7.4 Real-Time Control . . . . . . . . . . . . . . . . . . . . . . 142
4.7.5 City Logistics-Related Decision Support Systems
and Communication Technologies . . . . . . . . . . . . . 144
4.8 Conclusion of the Literature Review . . . . . . . . . . . . . . . . 145
5 A New Deterministic Real-Time Control Approach for RDOPG
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.1 General Categorization of Different Control Concepts in Logistic
and Production Systems . . . . . . . . . . . . . . . . . . . . . . . 149
5.2 The Deterministic Real-Time Control Approach . . . . . . . . . . 151
5.2.1 The Applied Real-Time Control Concept . . . . . . . . . . 151
5.2.2 Update Handling . . . . . . . . . . . . . . . . . . . . . . . 153
5.2.3 Vehicle Scheduling Strategy . . . . . . . . . . . . . . . . . 156
5.2.4 Idle Vehicle Waiting Strategy . . . . . . . . . . . . . . . . 156
5.3 The Mathematical Model of Individual Problem Instances . . . . . 156
5.3.1 Road Network and the Derived Digraph . . . . . . . . . . 156
5.3.2 The Deterministic Dynamic Mathematical Model . . . . . 158
5.4 Characterization of the Deterministic Real-Time Control Approach 163
5.5 Utilization of the Real-Time Control Approach in RDOPG
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6 A New Forecasting Approach for Generating Stochastic Knowledge
from Past Request Information and Utilizing the Stochastic
Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
6.1 Integrating Stochastic Knowledge into the Proposed Real-Time
Control Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 169
6.2 Generation of Dummy Customers Using Segment-Based Clustering 171
6.2.1 Segment-Based Cluster Generation . . . . . . . . . . . . . 172
6.2.2 Cluster Selection and Dummy Customer Generation . . . . 178
6.3 Handling Dummy Customers in the Pro-Active Real-Time Control
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
6.3.1 Dynamic Parameter Updating During the Transportation
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
xx Contents

6.3.2Extended Vehicle Scheduling Strategy with Regard

to Dummy Customers . . . . . . . . . . . . . . . . . . . . 182
6.3.3 Extended Vehicle Waiting Strategy Using Dummy
Customer Information . . . . . . . . . . . . . . . . . . . . 182
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
7 The Proposed Tabu Search Solution Method . . . . . . . . . . . . . . 185
7.1 General Concepts of the Implemented Solution Method . . . . . . 186
7.1.1 The Tabu Search Metaheuristic . . . . . . . . . . . . . . . 186
7.1.2 Description of an Iteration in the Solution Method . . . . . 187
7.1.3 Representation of a Tour Plan in the Solution Method . . . 189
7.1.4 Efficient Calculation of Changes in the Objective Function
Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.2 Neighborhood Operators . . . . . . . . . . . . . . . . . . . . . . . 190
7.2.1 Within Tour Insertion (WTI) . . . . . . . . . . . . . . . . 191
7.2.2 Relocate (REL) . . . . . . . . . . . . . . . . . . . . . . . 192
7.2.3 MultiRelocate (MREL) . . . . . . . . . . . . . . . . . . . 194
7.2.4 Large Neighborhood Search (LNS) . . . . . . . . . . . . . 194
7.2.5 Exchange Between Tours (XBT) . . . . . . . . . . . . . . 195
7.3 The Implemented Tabu List . . . . . . . . . . . . . . . . . . . . . 196
7.4 Construction of the Initial Tour Plan . . . . . . . . . . . . . . . . 198
7.5 Stage-Based Neighborhood Operator Selection Scheme . . . . . . 199
7.6 Integration of the Solution Method into the Real-Time Control
Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
7.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
8 Computational Results . . . . . . . . . . . . . . . . . . . . . . . . . . 203
8.1 The Discrete Event-Based Simulator . . . . . . . . . . . . . . . . 205
8.1.1 Simulation of the Requests . . . . . . . . . . . . . . . . . 207
8.1.2 Simulation of the Vehicles . . . . . . . . . . . . . . . . . . 207
8.2 Test Environment, Parameter Values, and Request Data Classes . . 208
8.2.1 System Environment and Real-Time Control Parameter
Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.2.2 Tabu Search Solution Method Parameter Values . . . . . . 209
8.2.3 Dummy Customer Request Parameter Values . . . . . . . . 210
8.2.4 Request Parameter Values and Request Data Classes . . . . 210
8.2.5 Number of Conducted Computational Experiments . . . . . 216
8.2.6 Performance Evaluation Using the Additionally Attainable
Improvement . . . . . . . . . . . . . . . . . . . . . . . . . 217
8.3 Performance of the Tabu Search Solution Method . . . . . . . . . 218
8.4 Performance Evaluation of Deterministic Real-Time Control
Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
8.5 Performance Evaluation of Allowing Vehicle En-Route Diversion . 224
8.6 Measuring Structural Distortion Using the Goodness-of-Fit Test . . 226
8.7 Results of Generating Stochastic Knowledge for S REAL and S GEN . 231
Contents xxi

8.8 Performance Evaluation of the Pro-Active Real-Time Control

Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
8.8.1 Results on S REAL . . . . . . . . . . . . . . . . . . . . . . 236
8.8.2 Results on S GEN . . . . . . . . . . . . . . . . . . . . . . . 241
8.9 The Degree of Structural Diversity . . . . . . . . . . . . . . . . . 249
8.10 Driver Inconvenience Aspects in the Real-Time Control Approaches 255
8.11 Summary and Answers to the Three Central Research Questions . 262
9 Summary and Outlook on Future Work . . . . . . . . . . . . . . . . 267
9.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
9.2 Outlook on Future Work . . . . . . . . . . . . . . . . . . . . . . . 276
Appendix A Additional Facts Related to Chap. 1 . . . . . . . . . . . . . 281
Appendix B Additional Information on Request Data Sets of Request
Data Class S GEN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
B.1 Parameters for Generating Selected Request Data Sets of S GEN . . 285
B.2 Visualization of All Request Data Sets of S GEN . . . . . . . . . . . 288
Appendix C Request Response Times of the Examined Request Data
Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
C.1 Request Response Times of S REAL . . . . . . . . . . . . . . . . . 289
C.2 Request Response Times of S GEN . . . . . . . . . . . . . . . . . . 291
About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
List of Figures

1.1 Example RDOPG application: The subsequent delivery of newspapers 4

1.2 Layers of the transportation process including the real-time control
layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1 Number of journal publications with regard to VRPs . . . . . . . . . . 16
2.2 Visualization of the tasks in a routing problem . . . . . . . . . . . . . 17
2.3 Infeasible TSP solution due to a subtour in the solution . . . . . . . . 20
2.4 Classification of different solution methods for node-based routing
problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 General process of tour plan generation and adaptation before and
during the execution of the transportation process . . . . . . . . . . . 39
2.6 A new classification of routing problems according to properties and
processing of relevant information . . . . . . . . . . . . . . . . . . . 48
2.7 The transportation process of routing problems with constant relevant
information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.8 The transportation process in Stochastic Vehicle Routing Problems . . 52
2.9 The transportation process in dynamic vehicle routing problems with
a centralized coordination . . . . . . . . . . . . . . . . . . . . . . . . 56
2.10 Example architecture of a real-time control approach . . . . . . . . . 63
2.11 Example architecture for a bidirectional communication system
between the vehicles and the central dispatching center . . . . . . . . 64
2.12 Different arrival times of dynamic requests in two instances of a
dynamic routing problem with the same degree of dynamism . . . . . 65
2.13 The effective degree of dynamism with time windows . . . . . . . . . 67
2.14 Two instances with the same effective degree of dynamism of 0.4 and
significantly different inter-arrival times . . . . . . . . . . . . . . . . 68
2.15 Three-echelon classification of dynamic routing problems . . . . . . . 69
3.1 Illustration of the linear2X and quadratic customer inconvenience . . . 82
3.2 RDOPG applications categorized according to the proposed
classification of routing problems . . . . . . . . . . . . . . . . . . . . 88
4.1 Execution of a transportation process using different waiting strategies 98

xxiii
xxiv List of Figures

4.2 Execution of a transportation process with and without using a

relocation strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.3 Execution of a transportation process using Drive-First and Wait-First
strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.4 Different customer inconvenience and penalty functions . . . . . . . . 138
4.5 Customer inconvenience functions considered by Fagerholt (2000,
2001) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.6 General architecture of real-time control approaches . . . . . . . . . . 144
5.1 The utilized real-time control concept adapting the transportation
process at τ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.2 The update handling of the proposed real-time control approach . . . . 154
5.3 Illustration of the linear2X and quadratic customer inconvenience . . . 162
5.4 Marginal increase in variable customer inconvenience using a
maximum request response time of one hour . . . . . . . . . . . . . . 163
5.5 Integration of the real-time control approach into RDOPG applications 164
6.1 Dividing the service area into subareas and segments . . . . . . . . . 173
6.2 The k-means clustering algorithm and the new proposed cluster
generation approach with different convex bases . . . . . . . . . . . . 174
6.3 Steps of the cluster generation approach generating one valid cluster
using a base consisting of four segments . . . . . . . . . . . . . . . . 174
6.4 Average arrival time of the first request when n request arrivals occur
in cluster ci . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.1 Representation of a tour plan using a doubly linked list . . . . . . . . 189
7.2 Steps for determining the tour of vehicle 1 . . . . . . . . . . . . . . . 189
7.3 Visualization of REL calculation steps . . . . . . . . . . . . . . . . . 193
7.4 Empty tour plan with two real requests and three dummy customer
requests (denoted as r3 , r4 , and r5 ) to be inserted by the construction
heuristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
7.5 Real requests inserted by the construction heuristic. If dummy
customers are used, these are inserted afterwards . . . . . . . . . . . . 198
8.1 Interrelation of event, process, and operation . . . . . . . . . . . . . . 206
8.2 Average request arrival distribution of the request data class S REAL . . 212
8.3 Linear averaging at the example of RegionDiversity (RD) . . . . . . . 213
8.4 Simplified illustration of the road network with request arrival
probabilities p(ai , tj ) for the structural diversity setting
S GEN RD = TD = 1.00 . . . . . . . . . . . . . . . . . . . . . . . . . 215
8.5 Average request arrival distribution of three TD settings of S GEN . . . 216
8.6 Three request data sets of S GEN with different levels of structural
diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
8.7 Adaptation handling of the evaluated deterministic real-time control
approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
8.8 Percentage improvements of TIMELIMIT10 vs. ZEROTIME,
linear2X customer inconvenience on S REAL . . . . . . . . . . . . . . 222
8.9 Percentage improvements of TIMELIMIT10 vs. ZEROTIME,
quadratic customer inconvenience on S REAL . . . . . . . . . . . . . . 222
List of Figures xxv

8.10 Percentage improvements of TIMELIMIT10 vs. ZEROTIME,

linear2X customer inconvenience on S GENall . . . . . . . . . . . . . . 223
8.11 Percentage improvements of TIMELIMIT10 vs. ZEROTIME,
quadratic customer inconvenience on S GENall . . . . . . . . . . . . . . 223
8.12 Impacts of allowing vehicle en-route diversion on request response
times using S REAL and 8 vehicles grouped by intervals of 50 seconds . 225
8.13 Impacts of allowing vehicle en-route diversion on request response
times using S REAL and 10 vehicles grouped by intervals of 50 seconds 225
8.14 Impacts of allowing vehicle en-route diversion on request response
times using S REAL and 12 vehicles grouped by intervals of 50 seconds 225
8.15 Three request data sets with a high structural distortion . . . . . . . . 227
8.16 Average number of generated dummy customers on S REAL and S GEN . 230
8.17 Average Poisson quality on S REAL and S GEN . . . . . . . . . . . . . . 230
8.18 Request response times on S REAL using 8 vehicles and DCminλ = 1.0
grouped by intervals of 50 seconds . . . . . . . . . . . . . . . . . . . 238
8.19 Request response times on S REAL using 8 vehicles and DCminλ = 2.0
grouped by intervals of 50 seconds . . . . . . . . . . . . . . . . . . . 238
8.20 Request response times on S REAL using 10 vehicles and DCminλ = 1.0
grouped by intervals of 50 seconds . . . . . . . . . . . . . . . . . . . 239
8.21 Request response times on S REAL using 10 vehicles and DCminλ = 2.0
grouped by intervals of 50 seconds . . . . . . . . . . . . . . . . . . . 239
8.22 Request response times on S REAL using 12 vehicles and DCminλ = 1.0
grouped by intervals of 50 seconds . . . . . . . . . . . . . . . . . . . 239
8.23 Request response times on S REAL using 12 vehicles and DCminλ = 2.0
grouped by intervals of 50 seconds . . . . . . . . . . . . . . . . . . . 239
8.24 Request arrival rates in the 72 subareas during the considered
planning horizon in S REAL . . . . . . . . . . . . . . . . . . . . . . . 240
8.25 Attainable improvements on S GEN at different levels of structural
diversity using linear2X customer inconvenience . . . . . . . . . . . . 244
8.26 Attainable improvements on S GEN at different levels of structural
diversity using quadratic customer inconvenience . . . . . . . . . . . 244
8.27 Request arrival rates in the 72 subareas during the considered
planning horizon in S GEN RD = TD = 1.00 . . . . . . . . . . . . . . 245
8.28 Request arrival rates in the 72 subareas during the considered
planning horizon in S GEN RD = TD = 0.50 . . . . . . . . . . . . . . 245
8.29 Request response times on S GEN using 8 vehicles and DCmin λ = 1.5
for RD = 0.00 and TD = 1.00 grouped by intervals of 50 seconds . . . 247
8.30 Request response times on S GEN using 10 vehicles and DCmin λ = 1.2
for RD = 0.00 and TD = 1.00 grouped by intervals of 50 seconds . . . 247
8.31 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.5, and 2.0 for RD = 1.00 and TD = 1.00 grouped by intervals of
50 seconds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.32 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.5, and 2.0 for RD = 1.00 and TD = 1.00 grouped by intervals of
50 seconds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
xxvi List of Figures

8.33 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.5, and 2.0 for RD = 1.00 and TD = 1.00 grouped by intervals of
50 seconds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.34 Results of the degree of structural diversity using the linear2X
customer inconvenience function . . . . . . . . . . . . . . . . . . . . 252
8.35 Results of the degree of structural diversity using the quadratic
customer inconvenience function . . . . . . . . . . . . . . . . . . . . 252
8.36 Managerial recommendations based on the dosd and S GEN
demanding a minimum improvement of 10 % . . . . . . . . . . . . . 253
8.37 Managerial recommendations based on the dosd and S GEN
demanding a minimum improvement of 15 % . . . . . . . . . . . . . 253
8.38 Managerial recommendations based on the dosd and S GEN
demanding a minimum improvement of 20 % . . . . . . . . . . . . . 254
8.39 Results of limiting vehicle en-route diversion activities on
S GEN RD = TD = 1.00 with DCminλ = 2.0 using the linear2X
customer inconvenience function and a fleet size of 10 vehicles . . . . 257
8.40 Results of limiting vehicle en-route diversion activities on
S GEN RD = TD = 1.00 with DCminλ = 2.0 using the quadratic
customer inconvenience function and a fleet size of 10 vehicles . . . . 258
8.41 Impacts of assigning penalty costs to vehicle en-route diversion
activities on request response times attained by PROACTIVE10 on
S GEN RD = TD = 1.00 using 10 vehicles and the linear2X customer
inconvenience function grouped by intervals of 50 seconds . . . . . . 260
8.42 Impacts of assigning penalty costs to vehicle en-route diversion
activities on request response times attained by PROACTIVE10 on
S GEN RD = TD = 1.00 using 10 vehicles and the quadratic customer
inconvenience function grouped by intervals of 50 seconds . . . . . . 260
A.1 Market share of domestic and foreign trucks in Germany for domestic
and cross-border transportation from 1991 to 2010 . . . . . . . . . . . 281
B.1 All request data sets of S GEN with different levels of structural diversity 288
C.1 Request response times on S REAL using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience grouped by
intervals of 50 seconds . . . . . . . . . . . . . . . . . . . . . . . . . 289
C.2 Request response times on S REAL using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience grouped by
intervals of 50 seconds . . . . . . . . . . . . . . . . . . . . . . . . . 289
C.3 Request response times on S REAL using 10 vehicles and
DCmin λ = 1.0, 1.2, 1.5, 1.8, and 2.0, linear2X customer
inconvenience grouped by intervals of 50 seconds . . . . . . . . . . . 290
C.4 Request response times on S REAL using 10 vehicles and
DCmin λ = 1.0, 1.2, 1.5, 1.8, and 2.0, quadratic customer
inconvenience grouped by intervals of 50 seconds . . . . . . . . . . . 290
C.5 Request response times on S REAL using 12 vehicles and
DCmin λ = 1.0, 1.2, 1.5, 1.8, and 2.0, linear2X customer
inconvenience grouped by intervals of 50 seconds . . . . . . . . . . . 290
List of Figures xxvii

C.6 Request response times on S REAL using 12 vehicles and

DCmin λ = 1.0, 1.2, 1.5, 1.8, and 2.0, quadratic customer
inconvenience grouped by intervals of 50 seconds . . . . . . . . . . . 290
C.7 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.00
and TD = 0.00 grouped by intervals of 50 seconds . . . . . . . . . . . 291
C.8 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.00
and TD = 0.00 grouped by intervals of 50 seconds . . . . . . . . . . . 291
C.9 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.25
and TD = 0.25 grouped by intervals of 50 seconds . . . . . . . . . . . 291
C.10 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.25
and TD = 0.25 grouped by intervals of 50 seconds . . . . . . . . . . . 292
C.11 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.50
and TD = 0.50 grouped by intervals of 50 seconds . . . . . . . . . . . 292
C.12 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.50
and TD = 0.50 grouped by intervals of 50 seconds . . . . . . . . . . . 292
C.13 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.75
and TD = 0.75 grouped by intervals of 50 seconds . . . . . . . . . . . 293
C.14 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.75
and TD = 0.75 grouped by intervals of 50 seconds . . . . . . . . . . . 293
C.15 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 1.00
and TD = 1.00 grouped by intervals of 50 seconds . . . . . . . . . . . 293
C.16 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 1.00
and TD = 1.00 grouped by intervals of 50 seconds . . . . . . . . . . . 294
C.17 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.00
and TD = 0.00 grouped by intervals of 50 seconds . . . . . . . . . . . 294
C.18 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.00
and TD = 0.00 grouped by intervals of 50 seconds . . . . . . . . . . . 294
C.19 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.25
and TD = 0.25 grouped by intervals of 50 seconds . . . . . . . . . . . 295
C.20 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.25
and TD = 0.25 grouped by intervals of 50 seconds . . . . . . . . . . . 295
xxviii List of Figures

C.21 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.50
and TD = 0.50 grouped by intervals of 50 seconds . . . . . . . . . . . 295
C.22 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.50
and TD = 0.50 grouped by intervals of 50 seconds . . . . . . . . . . . 296
C.23 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.75
and TD = 0.75 grouped by intervals of 50 seconds . . . . . . . . . . . 296
C.24 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.75
and TD = 0.75 grouped by intervals of 50 seconds . . . . . . . . . . . 296
C.25 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 1.00
and TD = 1.00 grouped by intervals of 50 seconds . . . . . . . . . . . 297
C.26 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 1.00
and TD = 1.00 grouped by intervals of 50 seconds . . . . . . . . . . . 297
C.27 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.00
and TD = 0.00 grouped by intervals of 50 seconds . . . . . . . . . . . 297
C.28 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.00
and TD = 0.00 grouped by intervals of 50 seconds . . . . . . . . . . . 298
C.29 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.25
and TD = 0.25 grouped by intervals of 50 seconds . . . . . . . . . . . 298
C.30 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.25
and TD = 0.25 grouped by intervals of 50 seconds . . . . . . . . . . . 298
C.31 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.50
and TD = 0.50 grouped by intervals of 50 seconds . . . . . . . . . . . 299
C.32 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.50
and TD = 0.50 grouped by intervals of 50 seconds . . . . . . . . . . . 299
C.33 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 0.75
and TD = 0.75 grouped by intervals of 50 seconds . . . . . . . . . . . 299
C.34 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 0.75
and TD = 0.75 grouped by intervals of 50 seconds . . . . . . . . . . . 300
C.35 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, linear2X customer inconvenience, RD = 1.00
and TD = 1.00 grouped by intervals of 50 seconds . . . . . . . . . . . 300
List of Figures xxix

C.36 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0,
1.2, 1.5, 1.8, and 2.0, quadratic customer inconvenience, RD = 1.00
and TD = 1.00 grouped by intervals of 50 seconds . . . . . . . . . . . 300
List of Tables

2.1 Comparison of several polynomial and exponential time complexity

functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 General classification scheme for routing problems including
RDOPG-specific attributes (in accordance with, e.g., Bodin and
Golden 1981, Stumpf 1998, Richter 2005, and Eksioglu et al. 2009) . . 75
3.1 The general classification scheme for routing problems applied to
RDOPG applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.1 Stage-based selection scheme of the proposed Tabu Search solution
method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7.2 Applied parameter values for the neighborhood operators MREL and
LNS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
8.1 Parameter values for the time slices utilized in generating S GEN . . . . 216
8.2 Performance of the evaluated deterministic real-time control
approaches on S REAL . . . . . . . . . . . . . . . . . . . . . . . . . . 221
8.3 Performance of different deterministic real-time control approaches
on S GENall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
8.4 Performance of TIMELIMIT10 vs. TIMELIMIT10NODIV on S REAL . 224
8.5 Evaluating the structural distortion on different generated request data
sets using different DCminPQ settings and DCminλ = 1.8 . . . . . . . . 229
8.6 Number of generated dummy customers on S REAL . . . . . . . . . . . 232
8.7 Number of generated dummy customers on S GEN . . . . . . . . . . . 233
8.8 Request realization rates (in percents) in S GEN and S REAL depending
on the parameter DCminλ using DCminPQ = 0.40 . . . . . . . . . . . . 235
8.9 Percentage improvements achieved by PROACTIVE10 vs.
TIMELIMIT10 (i.e., improvements attained by utilizing stochastic
knowledge ) on S REAL . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.10 Percentage improvements achieved by PROACTIVE10 vs.
TIMELIMIT10 (i.e., improvements attained by utilizing stochastic
knowledge) on S GEN . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

xxxi
xxxii List of Tables

8.11 Evaluated response time thresholds and penalty costs included in the
objective function for vehicle en-route diversion activities depending
on the applied customer inconvenience function . . . . . . . . . . . . 256
8.12 Results for TIMELIMIT10 and PROACTIVE10 under different
penalty costs on S GEN RD = TD = 1.00 using 10 vehicles . . . . . . . 259
A.1 Development of transportation performance in billion ton-kilometers
in the EU-27 from 1995 to 2008 . . . . . . . . . . . . . . . . . . . . . 281
A.2 Newspaper advertising expenditures in the United States of America
from 1950 to 2010 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
B.1 Resulting request arrival probabilities of selected request data sets in
S GEN with different levels of structural diversity . . . . . . . . . . . . 286
List of Algorithms

4.1 Basic online optimization . . . . . . . . . . . . . . . . . . . . . . . . 129

7.1 Neighborhood operator Within Tour Insertion (WTI) . . . . . . . . . . 191
7.2 Neighborhood operator Relocate (REL) . . . . . . . . . . . . . . . . . 192
7.3 Neighborhood operator MultiRelocate (MREL) . . . . . . . . . . . . 195
7.4 Neighborhood operator Large Neighborhood Search (LNS) . . . . . . 196
7.5 Neighborhood operator Exchange Between Tours (XBT) . . . . . . . . 196
8.1 Generation of a request data set of S GEN with RD ∈ [0, 1] and
TD ∈ [0, 1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

xxxiii
Abbreviations

Abbreviation Description Defined

on page

ADFSVRPSTW The Asymmetric Dynamic Forecasted Stochastic Vehicle 85

Routing Problem with Soft Time Windows
ADP Approximate Dynamic Programming 123
ADTSPTW The A Priori Dynamic Traveling Salesman Problem with 114
Time Windows
ADW strategy Advanced Dynamic Waiting strategy in Mitrovic-Minic 105
and Laporte (2004)
ASAP The After Salesman Problem 106
B2B Business To Business 1
B2C Business To Customer 1
B&B Branch-and-Bound 28
B&C Branch-and-Cut 28
B&C&P Branch-and-Cut-and-Price 28
B&P Branch-and-Price 28
BRH Branch-and-Regret Heuristic in Hvattum et al. (2007) 134
BUSIEST strategy BUSIEST strategy in Larsen (2000) 115
CCC Capacity-Cut Constraints 21
CCP Chance Constrained Program 53
CRC-32 Cyclic Redundancy Check with a checksum size of 32 bits 197
CRN Common Random Numbers 209
CVRP The Capacitated Vehicle Routing Problem 21
DARP The Dial-A-Ride Problem 24
DDARP The Dynamic Dial-A-Ride Problem 93
DF strategy Drive-First strategy in Mitrovic-Minic and Laporte (2004) 104

A The set of arcs in a directed graph 18

Ai The arrival time of the servicing vehicle at the location of 24
request i in a GPDP
C The loading capacity of the homogeneous vehicles in 21
a CVRP
c1 (P a ) The costs which occur in an SVRP according to the 53
information known before the execution of the
transportation process
c2 (P a ) The costs which are expected to occur in an SVRP during 53
the execution of the transportation process
cij The costs for traveling from location i to location j 18
CRA The competitive ratio of solution approach A 73
Di The departure time of the servicing vehicle at location i in 23
a GPDP
di The demand of request i in a CVRP 21
E The set of edges in an undirected graph 18
ei The start time of the time window of request i in a GPDP 24
f (x) The general objective function of a GPDP 23
G The directed or undirected graph 18
I Instance I of the considered dynamic routing problem in 73
the competitive analysis
K The set of available vehicles in a routing problem 20
K+ The unification of all vehicle start positions in a GPDP 22
K− The unification of all vehicle end positions in a GPDP 22
K The number of available vehicles in a routing problem 20

subregion i in Larsen (2000)
λ Intensity parameter denoting the request arrival rate in 110
Bertsimas and van Ryzin (1991)
M/G/1 Queuing system with a Poisson (Markovian) input, i.e., 111
request arrivals, general service time distribution, and one
server (cf. Kendall 1953 and Tijms 2003, p. 341 et seq.)
PiMAX Penalty costs of request i in the objective functions of 138
Fagerholt (2000) and Fagerholt (2001)
r The chosen request out of the set R to be serviced next 128
in Bent and van Hentenryck (2004c)
S The set of sample scenarios 130
s A sample scenario s ∈ S 130
t The current time during the execution of the transportation 121
process
w The objective function value of the solution of the 129
transportation process in Bent and van Hentenryck
(2004c)
w(r) The gain of request r according to the considered 128
objective function
Chapter 1
Introduction

In many real-world business to customer (B2C) and business to business (B2B) ap-
plications, unexpectedly occurring events may lead to unfavorable situations that
cause inconvenience or high costs for affected customers (or companies; in what
follows, we will use the term customers). For example, considering the case of a
newspaper subscription, a newspaper that has not been delivered in the morning
creates a high amount of inconvenience for the subscriber since reading newspapers
is a common habit in the daily lives of many people. Another example can be found
in manufacturing companies. If an important machine breaks down during the pro-
duction process, high outage costs occur. In order to provide a rough idea about
the dimension of a production outage, consider two of the major automobile brands
BMW (BMW Group) and Mercedes-Benz (Daimler Group) with a total number of
produced cars of 1,236,989 (BMW) and 1,312,456 (Mercedes-Benz) in 2010. One
hour of a production outage in all manufacturing facilities of these brands causes on
average a loss of 140 or 150 produced cars, or, according to their revenue, a financial
loss of over 6 million euros (cf. BMW Group 2011, Daimler Group 2011).
Due to the occurring inconvenience and costs in such situations, affected cus-
tomers want to have these situations quickly resolved. Therefore, customers de-
mand assistance from a specific service company which then resolves the situation
by delivering required goods with a high urgency. These goods may be material or
immaterial, e.g., spare parts, services, or a combination of both. It is assumed that
for the mentioned exceptional situations, a contract with a specific service com-
pany which is responsible for quickly resolving the situation has been negotiated
in advance. The key aspect in these situations is that, due to the high urgency, ser-
vice companies are forced to deliver the required goods to the affected customer
as quickly as possible. Specifically, since the occurrence of the unpleasant situation
has already caused a high amount of inconvenience to the customer, a delayed deliv-
ery of these goods causes additional significant customer inconvenience. Therefore,
this additional arising customer inconvenience decreases the value of the requested
goods with increasing time required for their delivery. Hence, we will refer to these
goods as perishable goods throughout this book. Their flow is always directed from
the service company to the customer so that the triggered service processes can be

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 1

DOI 10.1007/978-3-642-33472-6_1, © Springer-Verlag Berlin Heidelberg 2013
2 1 Introduction

referred to as ad-hoc distribution processes. According to these characteristics, the

real-world processes considered in this book are denoted as Real-Time Distribution
of Perishable Goods (RDOPG) applications.
In the marketing literature, a significant amount of research on customer satis-
faction has been conducted. According to Tax et al. (1998), customer satisfaction
relates to how a company deals with a customer complaint. Moreover, Smith et al.
(1999) state that in case of a failure, a company has the possibility of either restor-
ing customer satisfaction and recovering loyalty or risk customer dissatisfaction.
Specifically, failing to deliver the expected service quality may result in customer
complaints, reduced customer satisfaction, and even in loss of customers. Parasur-
aman (1987) states that in “a market crowded by competitors claiming to be con-
cerned about customer satisfaction, the real winners will be the ones with a truly
customer-oriented corporate culture”. Therefore, an adequate handling of customer
complaints about service quality is crucial for service companies in order to main-
tain customer loyalty (cf. Brady and Cronin Jr. 2001, Pollack 2009). Specifically,
for the service companies considered in RDOPG applications, providing a high ser-
vice quality is considered as the essential factor for successfully maintaining and
expanding their market positions by achieving a high customer loyalty (cf. Parasur-
aman et al. 1985, Zeithaml et al. 1996). Hence, it is the topmost priority of the con-
sidered service companies to fulfill or exceed the service quality expected by their
customers in order to generate high customer satisfaction and to establish long-term
customer connectivity.
The considered RDOPG applications are based on a variant of the Dynamic Ve-
hicle Routing Problem (cf., e.g., Gendreau and Potvin 1998) since new customer
requests or in short requests dynamically arrive at the considered service company
during the planning horizon. A relevant characteristic is that customer inconve-
nience starts at the very moment when the corresponding customer makes a new
request at the service company. This is different from most research on dynamic
routing problems in the literature which focuses on applications where it is suffi-
cient to service requests at any time within assigned time windows. In contrast, in
RDOPG applications, a quick delivery of demanded goods has the highest prior-
ity. Due to the fact that customers become more inconvenienced the longer they
have to wait for the requested goods, the considered objective is to minimize the
total customer inconvenience which evolves with the request response time, i.e., the
time period between the arrival of the customer request and start of service at the
customer request location (cf. Larsen et al. 2007).
In order to attain low customer inconvenience values and thus a high service qual-
ity, the development and evaluation of a new pro-active real-time control approach
for RDOPG applications is the main focus of this book. Specifically, this approach is
used for coordinating the service companies’ daily operative distribution processes
which are also called transportation processes in what follows. The coordination of
these transportation processes consists of efficiently utilizing the available service
workers in real-time. Since it is assumed that each service worker uses a vehicle, in
what follows, the terms service workers and vehicles will be synonymously used.
With regard to the considered objective, an efficient vehicle utilization consists of
1.1 Motivation 3

coordinating the vehicles so that customer inconvenience is minimized. Since cus-

tomer inconvenience depends on the achieved request response time as mentioned
above, using methods that reduce the attained response time of requests is impor-
tant. Therefore, in the proposed pro-active real-time control approach, the available
vehicles are pro-actively coordinated. Specifically, they are guided in advance to ar-
eas in which future requests are expected within a short amount of time so that a
faster service of these newly arriving requests and hence a higher solution quality
can be attained. A challenging aspect here is obtaining such information about ex-
pected request arrivals since this is often not directly available in many real-world
applications. In order to achieve this aim, new methods for automatically generat-
ing information about expected future requests from past request information are
proposed in this book. Since such information contains uncertainty, it is denoted as
stochastic knowledge in what follows. This stochastic knowledge is directly utilized
in the pro-active real-time control approach during the coordination of the trans-
portation process.
In this chapter, a motivation is given in which real-life applications belonging to
the considered RDOPG applications are presented and relevant background infor-
mation is provided. This is followed by a categorization of the considered problem
whose tasks are also classified according to the different layers of transportation
problems. Furthermore, the aims and contributions of the book as well as an outline
of the book are presented.

1.1 Motivation

In B2C and B2B markets, customers expect high service quality from companies,
especially in urgent ad-hoc after-sales services (cf. Parasuraman et al. 1991). The
demand for such services is triggered by a failure in prior service or a failure in a
system for which a service contract has been established. Since customers mainly
rate service quality according to the customer inconvenience that they experience,
which depends on the resulting response times, these increased requirements force
service companies to deliver requested perishable goods under high time pressure.
Specifically, the longer an affected customer waits, the more inconvenience will oc-
cur. On the other hand, quick assistance opens up possibilities for service companies
to improve their image and expand their market position.
In what follows, three real-world examples which motivate the research on
RDOPG applications considered in this book are given. In B2C markets, a typi-
cal application is the subsequent delivery of newspapers. This process occurs when
a subscriber has not received his newspaper in the morning due to a delivery er-
ror (e.g., the delivery person forgot to deliver the newspaper), a delivery failure (e.g.,
drop-out of the corresponding delivery person), or a thievery. In each of these cases,
it is of particular importance to the newspaper publisher to compensate the sub-
scriber. This compensation can either be fulfilled by a voucher or a subsequent
delivery. Since a quick subsequent delivery allows market differentiation for the
4 1 Introduction

Call Request generation Subsequent delivery

Subscriber Dispatch center Vehicles

time
Response time

Fig. 1.1 Example RDOPG application: The subsequent delivery of newspapers

publishing house compared to providing a voucher, the subsequent delivery process

considered in this book is of high practical relevance. Newspapers can be consid-
ered as perishable goods since customers demand a quick subsequent delivery of
their newspaper since they are used to reading it at the same time every day. Specif-
ically, many people are used to and want to read their newspaper during a particular
time period during the day, e.g., while having breakfast or while traveling to work
by public means. In the evening, the news is already outdated and therefore the
newspaper is rather uninteresting at this time of day. Furthermore, emerging low-
cost and convenient alternatives to printed newspapers such as cheaper or even free
newspaper websites of competitors give customers more market options and a better
position to enforce their demands. According to a recent study, these alternatives are
increasingly used by U.S. consumers (cf. The Media Audit 2011). This exerts pres-
sure on publishing houses to satisfy their customers in order to maintain customer
connectivity which is important for two reasons. First, a high number of subscribers
results in high revenues from selling newspapers. Second, a large subscriber base
provides a good negotiating position to the publishing house and enables it to make
good bargains for advertising space. This is especially important due to the emerg-
ing market share of Internet advertisements (see Table A.2 in Appendix A) which
may lead to a decline in revenue of print advertisements as well as revenue from
newspaper inserts (cf. Edmonds 2010).
In Fig. 1.1, the individual activities in the subsequent newspaper delivery process
are illustrated. First, when a newspaper is missing, a subscriber calls the designated
call and dispatch center. Here, a new request is generated which is subsequently ser-
viced by designated vehicles. In order to facilitate this process, Bieding et al. (2009)
propose the application of a system for the electronic communication of required in-
formation as well as utilizing online routing algorithms for coordinating the subse-
quent newspaper delivery process. Moreover, Ferrucci and Bock (2008) have shown
that the utilization of an appropriate solution approach for generating the vehicles’
delivery tours has a high potential of increasing the solution quality in such applica-
tions. A key aspect of this RDOPG application is that, when a vehicle has reached
the customer’s location, the service time required to service the customer, i.e., by
handing over the newspaper, can be assumed to be quite short (about 60 seconds)
as well as homogeneous for all customer requests. This allows a reliable estimation
of the time that a vehicle will require at a customer request. Furthermore, capacity
constraints are also not an issue since customers demand a newspaper to be subse-
quently delivered from a limited set of newspapers. Moreover, since newspapers are
1.1 Motivation 5

light, a sufficient number required for subsequent delivery activities during the day
can be loaded onto the vehicles at the beginning of the day. Therefore, each vehicle
can service any pending customer request. What makes the problem complex is that
newspaper subscribers are usually geographically dispersed so that new requests can
theoretically arise anywhere in the city. Another example of RDOPG applications
in B2C environments is quick road assistance in case of a car failure. This appli-
cation also belongs to the distribution of perishable goods since customers want to
receive quick help in case of problems with their cars. The main difference from the
previously described RDOPG application is that complex technical problems may
occur due to high-tech automobiles so that the service time has a significantly higher
degree of uncertainty.
In B2B markets one application of high economic relevance arises in the area of
repairman companies which provide specialized repair services for severe failures
occurring in production companies (cf. van de Klundert and Wormer 2010). Such
repair services have gained a much higher importance recently since manufacturers
have significantly changed their production habits over the last years in order to
deal with increased complexities in production processes and in order to reduce
costs (cf. Bock 2004, pp. 37–41):
• Decrease of the production depth: Due to technological advances, the com-
plexity of produced goods is steadily increasing. Manufacturers try to cope with
this increased complexity of their production processes by focusing on their core
competencies so that the in-house production depth can be decreased. Another
reason for reducing the production depth arises if required input goods can be
cost-efficiently procured. Possibilities of cheap procurement have arisen in recent
years due to a reduction in transportation costs. Specifically, since the 1990s, the
European transportation market has dramatically changed so that foreign compa-
nies performing truck-based transportation of goods at lower costs are permitted
to enter foreign markets (cf. Stumpf 1998, p. 19 et seq., Bock 2004, p. 30 et seq.).
As shown in Fig. A.1 (see Appendix A), their market share has steadily increased
leading to reductions in transportation costs.
• Just in Time concepts (JiT): In order to reduce storage costs for input goods,
manufacturers instruct component suppliers to deliver input goods just at the time
point when they are required (cf. Faria et al. 2010).
As mentioned above, these developments lead to leaner in-house production pro-
cesses and cost-savings. However, they also make production processes more prone
to errors. Due to the mentioned reductions in the production depth of individual
manufacturers, a significantly increasing demand for transportation services for de-
livering required input goods arises (see Table A.1 in Appendix A). Furthermore, the
mentioned increased complexity of many products due to technological advances
has led to another development:
• Special requirements and globalization: Raw materials and expertise knowl-
edge required for producing specific goods are only available in a limited number
of countries. However, since these goods are required in a variety of production
processes all over the world, strategic supply chain alliances formed by different
6 1 Introduction

companies according to input goods required in production processes are gaining

more importance (cf. Ghiani et al. 2004, pp. 16–17). “As a result of globalization,
more emphasis must be put on the efficient design and management of supply
chains, sometimes at the world level” (Ghiani et al. 2004, p. 17).
If some input goods cannot be delivered as planned, manufacturers may en-
counter problems in keeping their production process running. This situation be-
comes even more critical when using JiT in which such bottlenecks arise much
faster due to reduced storage capacities of required input goods. Furthermore, since
many companies are connected to each other and their production processes have
strong interdependencies, major disturbances throughout the supply chain may oc-
cur in case of failures in a single company in the supply chain (cf. Rice and Caniato
2003).
Since broken-down production lines cause significant financial losses every
minute as mentioned above in the case of automotive companies, companies in sup-
ply chains often make individual contracts which specify expected service levels
and define penalties in case the service level is not met by the supplying com-
pany (cf. Faria et al. 2010). Hence, in case of production failures, quick support
from repairman companies is mandatory. Therefore, repairman companies them-
selves are also often bound to individual contracts with their customers. This type
of contract is referred to as Service Level Agreement (SLA) and defines the type
of problems which are covered as well as terms and conditions, e.g., a maximum
allowed response time to a request of the corresponding customer. If a repairman
company does not service a customer request within this maximum response time,
high penalty costs occur. Moreover, it is advantageous for repairman companies
to service customer requests before the end of this time period for two reasons.
First, customers will experience better service quality so that the repairman com-
pany can attain a better image. Second, since resuming a production process earlier
than stated in the SLA imposes less financial damage to the manufacturer, cost sav-
ings can be shared between both parties. In addition to unknown customer requests,
the time period lengths required for service at different customer request locations
often contain a high degree of uncertainty in repairman applications. Hence, the es-
timation of these time period lengths poses an additional challenge. Furthermore,
for manufacturing companies, the point in time when a broken-down production is
repaired is important. Hence, SLA contracts can also specify a maximum tolerated
downtime (cf. Bertsimas and van Ryzin 1991) consisting of the sum of response
time and required service time. However, since the service time of requests signifi-
cantly depends on the considered RDOPG application and cannot be influenced by
optimizing the transportation process, in this book we focus on customer inconve-
nience and solution quality aspects which depend on the achieved request response
time.

Economic Need for Action As can be derived from the mentioned real-world
applications, the considered RDOPG applications are of high practical relevance.
In these applications, a quick reaction towards dynamically arriving customer de-
mands is the topmost priority so that customer loyalty is maintained and customer
1.2 Categorization of the Considered Problem 7

inconvenience is kept at a low level. Since operational costs incurred by travel activ-
ities of vehicles can be regarded as negligible compared to costs incurred by a late
response to customer requests, it is advantageous for companies to quickly service
pending customer requests even if this requires additional travel activities. Hence, as
mentioned above, it is advantageous to use a pro-active real-time control approach
so that a quicker response to customer requests is attained thereby yielding high
customer satisfaction and quality of service.

1.2 Categorization of the Considered Problem

The considered problem of coordinating the operative transportation processes in
RDOPG applications is part of the service operations management in service com-
panies (Johnston and Clark 2008, p. 3):
Service operations management is concerned with delivering service to the
customers or users of the service. It involves understanding the needs of our
customers, managing the processes that deliver the services, ensuring our ob-
jectives are met, while also paying attention to the continual improvement
of our services. As such, operations management is a central organisational
function and one that is critical to organisational success.
Hence, “Service operations management is the term that is used to cover the
activities, decisions and responsibilities of operations managers in service organisa-
tions. These managers are often called operations managers but many other titles are
used, such as [. . . ], fleet managers in transport companies, [. . . ] and so on” (John-
ston and Clark 2008, p. 4).
Furthermore, the considered RDOPG applications exhibit some characteristics
of short-haul freight transportation processes (cf. Ghiani et al. 2004, p. 247 et seq.).
Specifically, the requested goods are distributed to customers in a relatively small
service area without transshipment. Since the used vehicles can carry more than
one good at the same time, the resulting transportation problem is referred to as a
less-than-truckload (LTL) problem (cf. Ghiani et al. 2004, p. 12). Hence, in such
transportation problems it is likely that several vehicles are close to each other so
that sophisticated methods for assigning and sequencing requests on vehicles can
be advantageously applied. In contrast to classic short-haul freight transportation
processes, there are differences in the applied objective function and the type of
goods to be delivered in the considered RDOPG applications.
In the literature, short-haul transportation processes consist of different activities
divided into three different layers of planning which can also be applied to the con-
sidered RDOPG applications. Besides different activities, each of these layers has a
different time horizon (cf. Ghiani et al. 2004, p. 247 et seq.):
• Strategic layer: In this layer, long-term decisions are made. Generally, in trans-
portation processes, this comprises choosing appropriate depot locations, defin-
ing operation zones as well as choosing transportation services that are offered. In
8 1 Introduction

Strategic layer long-term

Tactical layer medium-term

2 Operative layer short-term

Considered layers in this book

(0
immediate (online
Real-time control layer
during execution)

Fig. 1.2 Layers of the transportation process including the real-time control layer

RDOPG applications, besides establishing appropriate depots, this also includes

choosing the branches of industry for which services are offered, e.g., repair ser-
vices for production systems of automobile manufacturers or subsequent delivery
of newspapers for publishing houses.
• Tactical layer: This layer specifies decisions that are applied for a medium-term
time horizon. In RDOPG applications for example, this involves determining the
available fleet size.
• Operative layer: In this layer, short-term tasks are performed. For example, re-
quests which are already known before the execution of the transportation process
are assigned and sequenced on the available vehicles. In so doing, the tour plan of
the next planning horizon, usually the next working day, is generated (cf. Stumpf
1998, p. 52 et seq.).

As can be seen from these three layers, the described planning activities end
when the vehicles leave the depot in order to service assigned requests. However, in
dynamic routing problems such as the RDOPG applications considered in this book,
this is not sufficient since the majority of requests to be serviced only occur during
the execution of the transportation process.
In order to overcome this limitation, Bock (2004, p. 63) extends these three lay-
ers by introducing a fourth layer denoted as Real-time control layer. This layer
deals with the coordination of the transportation process in real-time parallel to its
execution. Figure 1.2 illustrates the four described layers and their time range. As
can be seen in this figure, the real-time control approaches which are developed in
this book are applied in the operative layer and real-time control layer since an ini-
tial solution is calculated before the execution of the transportation process which
is then adapted during its execution. As mentioned above, since in the considered
RDOPG applications the majority of customer requests only arrive during the exe-
cution of the transportation process, the activities of the real-time control layer are
the main focus of this book.
1.3 Aims and Contributions of the Book 9

1.3 Aims and Contributions of the Book

The described RDOPG applications cover various real-world problems. Motivated
by the challenges faced in the considered real-world problems, a new pro-active
real-time control approach is developed and evaluated in this book. Moreover, since
real-world request data of the subsequent delivery of newspapers process is provided
from a publishing house, this application is used as the representative RDOPG ap-
plication in this book.
In this section, first the aims of this book are presented. Specifically, the cen-
tral research questions to be answered in this book are described. Afterwards, the
contributions of this book are described.

1.3.1 Central Research Questions to Be Answered in This Book

During the evaluation of the proposed pro-active real-time control approach, we

investigate the following three questions:

Question 1: Does the integration of information about expected future request ar-
rivals into a deterministic real-time control approach, thus forming a pro-active real-
time control approach, allow for a significant improvement of the solution quality
by reducing customer inconvenience?

Question 2: If question 1 yields a positive result, under which criteria does this
apply? For example, what impact does the structural quality of existing request data
and the resulting stochastic knowledge which is generated on the basis of this re-
quest data have on the solution quality? Furthermore, which characteristics are suit-
able for measuring the structural quality of available request data?

Question 3: What impact does vehicle en-route diversion have on the solution qual-
ity? Since this increases driver inconvenience, is there a conflict between customer
inconvenience and driver inconvenience and does a useful trade-off between both
objectives exist?

1.3.2 Contributions of This Book

The proposed models, approaches, and methods for pro-actively managing and con-
trolling the operative transportation processes of RDOPG applications can be char-
acterized by the following contributions:1

1 Some of the contributions of this thesis have also been published in concise form in Ferrucci et

al. (2013).
10 1 Introduction

• Maximization of customer loyalty: We integrate results from marketing re-

search dealing with customer loyalty. Specifically, customer inconvenience is
operationalized in the objective function linearly and quadratically. Besides, an
additional penalty occurs for requests that exceed a maximum allowed response
time. This combines the aims of servicing customer requests quickly and main-
taining negotiated conditions, e.g., those defined by SLA contracts.
• Real-time control of time-critical distribution of perishable goods: Since dy-
namically arriving requests have to be immediately serviced due to the consid-
ered objective function, the tour plan execution process needs to be efficiently
controlled in real-time. Therefore, a sophisticated deterministic real-time control
approach is proposed. It is utilized in order to conduct the dynamic transportation
process by solving individual problem instances. Due to the perishability of the
goods to be delivered, this real-time control approach faces a high degree of dy-
namism and urgency (cf. Larsen et al. 2002, Ichoua et al. 2007). Furthermore, the
proposed approach accounts for the concurrency of tour plan execution and tour
plan adaptation during the transportation process. In order to classify the consid-
ered RDOPG applications with regard to other routing problems known in the
literature, a new classification based on suitable classifications in the literature is
proposed.
• Universally applicable forecasting approach for the generation and inte-
gration of stochastic knowledge: In this book, forecasting methods which are
able to automatically extract stochastic knowledge about expected future requests
solely from past request information are developed. Specifically, a customizable
segment-based clustering approach which identifies and selects suitable, well-
defined request-likely areas by analyzing request arrivals of past working days is
proposed. Since underlying request distributions of the available past request in-
formation are unknown, the clustering approach comprises methods which ensure
that only request-likely areas complying with defined quality criteria are selected.
Additionally, methods for integrating the generated stochastic knowledge in form
of dummy customers into the proposed deterministic real-time control approach
are developed. This also includes the definition of handling dummy customers
during the execution of the transportation process. In doing so, the determinis-
tic real-time control approach is extended to the proposed pro-active real-time
control approach.
• Integration of vehicle en-route diversion in a real road network: In order to
increase the practicability of the proposed real-time control approaches, vehicles
move on a fully-detailed urban road network. In order to allow vehicle en-route
diversion in the road network, the implemented discrete event-based simulator
maps vehicle travel activities in detail.
• Tabu Search with a variable neighborhood structure: In order to quickly solve
individual problem instances during the execution of the transportation process,
a Tabu Search metaheuristic with a variable neighborhood structure is proposed.
Depending on solutions explored in previous iterations, intensifying or diversi-
fying neighborhood operators are applied. In order to efficiently and precisely
determine the tabu-active state of solutions, an appropriate tabu-active attribute is
developed.
1.3 Aims and Contributions of the Book 11

• Classification of structural request data quality: In order to support the identi-

fication of scenarios in which the application of the proposed pro-active real-time
control approach is advantageous, we investigate the structural quality of request
data sets. For this purpose, we derive the structural quality of a request data set
from two newly proposed criteria denoted as structural distortion and structural
diversity. In addition to defining both criteria in detail, we also propose methods
for determining the structural distortion and the structural diversity of a given re-
quest data set. By evaluating designed request data sets with different structural
qualities, we show the efficiency of an approach proposed for distinguishing be-
tween request data sets which have a low structural distortion and those that have a
high structural distortion. Note that this is important since a low structural distor-
tion means that request arrivals on the next day follow specific patterns identified
by the forecasting method based on the past request information. This is required
in order to enable reliable predictions of expected future requests which is a pre-
requisite for an efficient utilization of pro-active real-time control approaches.
Furthermore, by examining request data sets with low structural distortion, we
identify structural diversity as a crucial criterion for classifying the value of the
attainable stochastic knowledge thereof. In computational experiments, the im-
pact of different levels of structural diversity on the attainable improvement is
assessed in detail. Specifically, for each tested request data set, the solution qual-
ity improvements attainable by using the pro-active real-time control approach
are evaluated by comparing its results with the results attained by its determin-
istic counterpart. Furthermore, besides a qualitative assessment of the structural
diversity of request data sets, a quantitative approach denoted as the degree of
structural diversity (dosd) is proposed. For request data sets evaluated in the main
computational study, we compare the attained improvements with the values cal-
culated by the proposed dosd. The results show the effectiveness of the dosd in
estimating improvements that can be attained by utilizing the proposed pro-active
real-time control approach compared to its deterministic version. Therefore, the
proposed criteria provide a good starting point for assessing the structural quality
of request data sets in order to efficiently estimate the attainable improvements by
using stochastic knowledge in pro-active real-time control approaches. Further-
more, these criteria allow for an efficient customization of the proposed pro-active
real-time control approach depending on the existing scenario.
• Evaluation of vehicle en-route diversion effects and related driver inconve-
nience: Consistent with results of other research work on dynamic routing prob-
lems in the literature, we show that allowing vehicle en-route diversion has a
significantly positive effect on the attained solution quality. However, allowing
vehicle en-route diversion activities increases the demands on the drivers and
thereby raising driver inconvenience. Since the positive effects on solution qual-
ity make the complete prohibition of vehicle en-route diversion unattractive, an
approach for limiting the number of vehicle en-route diversion activities is pro-
posed and resulting effects on the solution quality are investigated. Reasonable
trade-off parameter settings with regard to the contradicting aims of minimizing
driver inconvenience and customer inconvenience are presented.
12 1 Introduction

In this book, the above-mentioned contributions are achieved by developing,

implementing, and evaluating real-time control and forecasting approaches which
use various methodologies from the area of Operations Research (OR). Hillier and
Lieberman (2008) describe OR as follows:
As its name implies, operations research involves “research on operations.”
Thus, operations research is applied to problems that concern how to con-
duct and coordinate the operations (i.e., the activities) within an organization.
(. . . ) The research part of the name means that operations research uses an ap-
proach that resembles the way research is conducted in established scientific
fields. To a considerable extent, the scientific method is used to investigate the
problem of concern. In fact, the term management science [MS] sometimes is
used as a synonym for operations research. (Hillier and Lieberman 2008, p. 2)
According to the authors, the following steps are required to be undertaken in an
OR project:
In particular, the process begins by carefully observing and formulating the
problem, including gathering all relevant data. The next step is to construct a
scientific (typically mathematical) model that attempts to abstract the essence
of the real problem. It is then hypothesized that this model is a sufficiently
precise representation of the essential features of the situation that the conclu-
sions (solutions) obtained from the model are also valid for the real problem.
Next, suitable experiments are conducted to test this hypothesis, modify it
as needed, and eventually verify some form of the hypothesis. (Hillier and
Lieberman 2008, p. 2)
Another important aspect is that “OR is also concerned with the practical man-
agement of the organization. Therefore, to be successful, OR must also provide pos-
itive, understandable conclusions to the decision maker(s) when they are needed”
(Hillier and Lieberman 2008, p. 3). This is important since executives seem not to
trust systems that make decisions on a black box basis. According to Stenfors et al.
(2007), a survey in Finland’s 500 largest companies which assessed the utilized
tools when making major decisions revealed that only 10 % of the used tools belong
to OR/MS tools. Nevertheless, executives see benefits in using OR tools in order
to make operational processes more efficient. According to Stenfors et al. (2007),
“Tools which support creativity are needed”. In order to attain a higher acceptance of
the proposed methods in practical applications, the process of generating stochastic
knowledge before the execution of the transportation process as well as the vehicle
activities performed during the execution of the transportation process are visualized
in the graphical user interfaces of the developed software tools. However, since the
focus of this book is to evaluate the attainable improvements by utilizing stochas-
tic knowledge in real-time control approaches, the implemented visualizations are
not considered any further in this book. Moreover, skilled users can interactively
influence the stochastic knowledge generation process by defining specific values
for different parameters.
Further information about the field of OR can be found among others in Hillier
and Lieberman (2008) and Domschke and Drexl (2011).
1.4 Outline of the Book 13

1.4 Outline of the Book

This book is organized as follows. After this chapter, a comprehensive introduction

to tour planning approaches is given in Chap. 2. Here, classic static routing problems
as well as its variants known in the literature are described. Furthermore, differences
between static and dynamic routing problems are presented. Moreover, classifica-
tions of routing problems published in the literature are presented and, based on
these classifications, a unified approach for classifying routing problems accord-
ing to RDOPG-specific attributes is proposed. As mentioned above, since RDOPG
applications are based on dynamic routing problems, a major part of this chapter
considers characteristics of dynamic routing problems.
In Chap. 3, a problem description of the considered RDOPG applications is pre-
sented in detail. According to this problem description, in the literature review in
Chap. 4, research on dynamic routing problems which covers aspects relevant for
RDOPG applications is examined. The major focus is on describing approaches
which provide an increased flexibility for integrating newly arriving requests during
the execution of the transportation process.
Chapter 5 introduces the proposed deterministic real-time control approach
which is used to control the transportation process during its execution. Further-
more, the mathematical model utilized for modeling individual problem instances
arising during the execution of the transportation process is described. Afterwards,
the proposed real-time control approach is classified according to a general classifi-
cation scheme for real-time control approaches.
In order to generate stochastic knowledge out of past request information, a new
segment-based clustering approach is proposed in Chap. 6. Specifically, we elabo-
rate on how our forecasting approach generates and subsequently integrates stochas-
tic knowledge into the deterministic real-time control approach thereby creating the
proposed pro-active real-time control approach. Furthermore, the handling of the
generated stochastic knowledge in the real-time control approach during the execu-
tion of the transportation process is described.
In Chap. 7, the developed solution method which is used for solving the individ-
ual problem instances arising in the real-time control approaches is presented. This
is a stage-based Tabu Search metaheuristic which utilizes different neighborhood
operators, depending on solutions visited in previous iterations. The neighborhood
operators as well as the tabu-active attribute utilized in this solution approach are
presented in detail. Furthermore, the stage-based scheme used for selecting a neigh-
borhood operator in each iteration is described.
In Chap. 8, the results of the computational experiments are extensively pre-
sented. Test instances based on real-world and designed request data sets compris-
ing defined structural quality characteristics are presented. Specifically, in order to
evaluate the proposed pro-active real-time control approach under different levels of
structural request data quality, the criteria structural distortion and structural diver-
sity are described in detail. Using these criteria, request data sets with different levels
of structural distortion as well as different levels of structural diversity are designed.
Results from different preliminary computational studies are presented. In the main
14 1 Introduction

computational experiments, the performance of the pro-active real-time control ap-

proach is evaluated by comparing its attained results with the results achieved by
its deterministic counterpart. Moreover, new approaches for assessing the structural
quality of given request data sets with regard to their suitability in our proposed
pro-active real-time control approach are presented. After identifying suitable re-
quest data sets, measures for determining achievable improvements by using the
pro-active real-time control approach are proposed. Furthermore, a new approach
for limiting vehicle en-route diversion activities is introduced. Finally, answers to
the three central research questions are given.
Chapter 9 concludes this book by summarizing the main results and discussing
avenues for further research that are opened by the contributions of this book.
Chapter 2
Introduction to Tour Planning:
Vehicle Routing and Related Problems

In the literature, tour planning problems (or routing problems) are a vital research
area. As illustrated in Fig. 2.1, the number of journal publications on Vehicle Rout-
ing Problems (VRPs, see Sect. 2.3.2) has steadily increased over the years. This in-
creased scientific interest can be explained by several factors. For example, progress
in computational resources has opened new possibilities for modeling more complex
routing problems such as dynamic routing problems (see Sect. 2.5.3.2). Further-
more, new arising real-world applications provide inspiration for developing new
approaches for coordinating complex transportation processes.
In this chapter, a general introduction to tour planning problems as well as as-
pects of specific types of routing problems are given. Note that this chapter mainly
deals with routing problems which comprise characteristics of short-haul routing
problems, i.e., routing problems in which transportation activities are performed in
a bounded service area without transshipment during a planning horizon which usu-
ally has a length of up to one working day (see Sect. 1.2 and cf. Ghiani et al. 2004,
p. 247 et seq.). First, the general task and definitions of tour planning problems are
introduced. Afterwards, the representation of requests and general types of routing
problems, i.e., node-based and edge-based routing problems, are described. Since
the considered RDOPG applications can be modeled as variants of node-based rout-
ing problems, different node-based routing problems known in the literature and
characteristics of these, including solution complexity issues and existing solution
methods, are presented. Furthermore, since the considered RDOPG applications can
be modeled as a variant of the well-known Vehicle Routing Problem, extensions
to the Vehicle Routing Problem which can be found in the literature, e.g., specific
problem extensions or other objective functions, are presented. Afterwards, different
types of information revelation in routing problems are discussed by describing ap-
propriate classifications known in the literature. Moreover, a new classification with
regard to the type of information revelation and coordination of the transportation
process is proposed in order to describe specific attributes which are characteristic
of RDOPG applications.
Since RDOPG applications utilize a centralized coordination of the transporta-
tion process, characteristics of dynamic routing problems which utilize a centralized

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 15

DOI 10.1007/978-3-642-33472-6_2, © Springer-Verlag Berlin Heidelberg 2013
16 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

500
447
450
Number of publications in journals

400

350

300

250
212
200
158
150
101
100
53
50
14 23
2 2 4
0
1954-1959 1960-1964 1965-1969 1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2006
(7 years)
Year

Fig. 2.1 Number of journal publications with regard to VRPs (cf. Eksioglu et al. 2009)

coordination are presented in the following section. In addition to typical objec-

tives considered in dynamic routing problems and a description of different types
of possible dynamic events, a measure known in the literature for determining the
difficulty of solving dynamic routing problems is presented. Furthermore, dynamic
real-world routing applications are described and classified according to the pre-
sented measure. Moreover, two measures for evaluating the performance of solution
approaches for dynamic routing problems are presented. Finally, a general classi-
fication scheme for routing problems known in the literature is described and ex-
tended according to characteristics which are specific for RDOPG applications.

2.1 General Task and Definitions

In general, the tour planning process can be defined as follows: A set of requests
needs to be assigned to a set of resources, for example vehicles that execute the
transportation process. Since resources always represent vehicles in the RDOPG
applications considered in this book, the terms vehicle and resource are used syn-
onymously. For each vehicle, the assigned requests are required to be ordered in a
specific sequence in which the requests will be serviced by the vehicle afterwards.
Performing the assignment and sequencing tasks results in a tour for each vehicle.
Both of these tasks can either be consecutively or simultaneously performed. The
step in a tour which consists of traveling from one request to the next is called a
route. The set of all vehicle tours forms the tour plan which is also denoted as the
solution of the considered routing problem. Note that problem-specific constraints,
e.g., time window restrictions of requests, need to be fulfilled by the tour plan in
order to obtain a feasible solution. Figure 2.2 illustrates the described tour planning
process using five pending requests and two vehicles which start and end their tour
at a depot D.
2.1 General Task and Definitions 17

Assignment of requests to vehicles

r1 r2 r3 r4 r5 Requests

v1 v2 Vehicles

Sequencing of requests

r4 r2 r1 Tour of vehicle v1

r3 r5 Tour of vehicle v2

Resulting tour plan

r4 r2

D r1 Tour of vehicle v1

r3 r5 Tour of vehicle v2

Fig. 2.2 Visualization of the tasks in a routing problem

The generation of a tour plan is connected to a specific aim which is pursued,

e.g., minimizing the total travel distance. Mathematically, this aim is defined by the
objective function which assigns a specific objective function value to each solu-
tion. By using this objective function value, different solutions can be compared
and ordered according to their solution quality. The goal is to find the optimal solu-
tion, i.e., the one with the best objective function value among all feasible solutions.
In what follows, we assume that the aim is always to minimize a given objective
function value, i.e., an optimal solution has the minimum objective function value
among all feasible solutions. Since routing problems usually belong to the class
of strongly non-deterministic polynomial time hard (NP-hard) combinatorial opti-
mization problems (cf. Lenstra and Rinnooy Kan 1981), it is often not possible to
determine an optimal solution in reasonable time due to complexity reasons. Hence,
the aim is to determine a good solution whose objective function value is as close
as possible to the objective function value of an optimal solution. Details on the
18 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

complexity of routing problems are described in Sect. 2.3.4. Furthermore, different

objectives which are pursued in routing problems are presented in Sect. 2.4.12.

2.2 Representation of Requests and General Types of Routing

Problems
In routing problems, requests are represented by relevant locations and are modeled
in a graph. This graph consists of a set of vertices (or nodes) defined by V and
of a set of arcs A or edges E. Specifically, if the graph is directed, it is defined
by G = (V , A). If it is undirected, it is defined by G = (V , E). In the directed case,
direction-dependent travel costs can be assigned so that the costs for traveling from
location i ∈ V to location j ∈ V represented by cij can be different from cj i whereas
in the undirected case, it holds that cij = cj i ∀i, j ∈ V .
Routing problems can be distinguished according to the type of requests that need
to be serviced. Specifically, a distinction is made between arc-based routing prob-
lems and node-based routing problems. In arc-based routing problems, customer re-
quests are represented by arcs or edges, depending on whether the graph is directed
or undirected. A characteristic of these types of problems is that the vehicle is at a
different position before and after servicing a request. In the Chinese Postman Prob-
lem introduced by Kwan (1962), each road is assumed to be undirected so that each
edge of G needs to be visited at least once. Moreover, if G also contains directed
roads, the problem is called the Mixed Chinese Postman Problem. Another variant
is the Rural Postman Problem in which only a given subset of edges E ⊂ E needs
to be serviced. For more information on arc-based routing problems, see Eiselt et al.
(1995) and Golden et al. (2008, p. 37 et seq.). In contrast to arc-based routing prob-
lems, node-based routing problems arise if customers are located at vertices v ∈ V
so that vehicles remain at the same position while servicing a customer request. In
addition to pure node-based and pure arc-based routing problems, mixed routing
problems where requests are located at arcs as well as nodes exist.
Since the RDOPG applications considered in this book belong to the class of
node-based routing problems, we consider node-based routing problems in what
follows.

2.3 Node-Based Routing Problems

In this section, classic node-based routing problems are introduced in their asym-
metric variant so that different costs for both travel directions can be modeled. First,
the Traveling Salesman Problem (TSP) and the Vehicle Routing Problem (VRP and
CVRP) are described. Afterwards, routing problems in which goods or persons have
to be transported between customer locations are presented. These problems consist
of the General Pickup and Delivery Problem (GPDP) and related variants known as
the Pickup and Delivery Problem and the Dial-A-Ride Problem. Furthermore, the
2.3 Node-Based Routing Problems 19

complexity of node-based routing problems as well as solution methods for node-

based routing problems are addressed.

2.3.1 The Traveling Salesman Problem (TSP)

The first routing problem which was considered in the literature is the Traveling
Salesman Problem (TSP, cf. Dantzig et al. 1954). In the TSP, one salesman needs to
service a given set of R customer requests R = {1, . . . , R}, also denoted as cities.
In this routing problem, the set of requests is equal to the set of relevant locations V
in the digraph G = (V , A) which are defined by V = {1, . . . , R}. The salesman
is originally located at one of the customer requests and needs to service the re-
quests by visiting each customer location exactly once and finally returning to its
origin. Traveling from i to j , (i, j ) ∈ A = V × V produces costs denoted by cij . For
example, these costs can represent travel distances (see Sect. 2.4.12). The sought
optimal solution is a cost-minimal tour plan that represents the sequence in which
the requests are to be visited. Specifically, the solution is given by an assignment of
the binary variables X = {xij } (i, j ∈ V , i = j ) which are 1 if the salesman directly
travels from location i to j and 0 otherwise. The TSP is constituted by the following
mathematical model (cf. Laporte 1992b):

min z = cij · xij s.t. (2.1)
i∈V j ∈V

∀i ∈ V : xij = 1 (2.2)
j ∈V

∀i ∈ V : xj i = 1 (2.3)
j ∈V

∀S ⊂ V , 2 ≤ |S| ≤ R − 2: xij ≥ 1 (2.4)
i∈S j ∈{V \S}

The objective function value of a solution is determined by the objective func-

tion (2.1). Constraints (2.2) and (2.3) ensure that each location is visited and left
exactly once. Constraints (2.4) eliminate subtours in the solution (see Fig. 2.3). Note
that this subtour formulation results in 2R − 2R − 2 subtour elimination constraints
which cannot be reasonably handled by using Mixed Integer Programming (MIP)
solvers even for small problem instances (cf. Laporte 1992b). Besides this sub-
tour elimination formulation, other subtour elimination constraint formulations are
possible. For example, Miller et al. (1960) propose a formulation which only con-
tains O(R2 ) restrictions but at the expense of additional variables. Moreover, it may
produce weaker lower bounds than the previous formulation (cf. Desrochers and
Laporte 1991). Other subtour elimination formulations can be found for example in
Laporte (1992b) and in Domschke and Scholl (2010, p. 100).
20 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Fig. 2.3 Infeasible TSP 2

4
solution due to a subtour in
the solution 1 3

5
6
8
7

2.3.2 The Vehicle Routing Problem (VRP and CVRP)

The Vehicle Routing Problem (VRP) was introduced by Dantzig and Ramser (1959).
In its basic version, a set of requests has to be serviced by a fleet of homogeneous
vehicles. Each vehicle begins and ends its tour at the same location called the depot.
The VRP is used to model two different types of real-world business cases:
1. Goods that are loaded onto the vehicles at the depot are delivered to the cus-
tomers.
2. Goods that are located at customer locations are collected and transported to the
depot.
The uncapacitated VRP in which capacity constraints of the vehicles can be ne-
glected is equal to a multiple-salesman variant of the previously introduced TSP (m-
TSP, cf. Laporte and Nobert 1980 and Bektas 2006). In the VRP, R = {1, . . . , R}
defines the set of R customer requests. The set of nodes is defined by V = R ∪ {0},
i.e., customer requests and the depot denoted by node 0. Furthermore, the set K
represents the set of K uncapacitated vehicles K = {1, . . . , K}. As is the case in the
TSP, travel costs on the arcs are defined by cij with i, j ∈ V . Since each request
location is visited exactly once by a vehicle and since the utilized vehicles are ho-
mogeneous, it is not required to specify which tour belongs to which vehicle in the
solution. Hence, it is possible to model the VRP using only a two-index vehicle flow
formulation that uses O(R2 ) binary variables as is the case in the TSP (cf. Toth
and Vigo 2002, p. 30). In this formulation, the solution consists of a tour plan rep-
resented by a set of binary variables X = {xij } (i, j ∈ V , i = j ) equal to 1 if and
only if a vehicle directly travels from location i to location j . The VRP can then
be defined by the following mathematical model (cf. Laporte and Nobert 1980 and
Bektas 2006):

min z = cij · xij s.t. (2.5)
i∈V j ∈V

∀i ∈ R: xij = 1 (2.6)
j ∈{V \i}

∀i ∈ R: xj i = 1 (2.7)
j ∈{V \i}

x0i = K (2.8)
i∈R
2.3 Node-Based Routing Problems 21

xi0 = K (2.9)
i∈R

∀S ⊂ R, 2 ≤ |S| ≤ R − 2: xij ≥ 1 (2.10)
i∈S j ∈{V \S}

The objective function is given by (2.5). Constraints (2.6) and (2.7) ensure that
each customer request is visited and left exactly once. These constraints also de-
termine that exactly one vehicle services each request. Constraints (2.8) and (2.9)
make sure that each vehicle leaves the depot and returns to it exactly once. Finally,
constraints (2.10) eliminate subtours in the vehicle tours.

The Capacitated VRP The Capacitated Vehicle Routing Problem (CVRP) is

closely related to the uncapacitated VRP. In addition, each request i ∈ R has a de-
mand di and all vehicles have the same loading capacity denoted by C. The CVRP
can be modeled using the previously introduced formulation of the VRP and replac-
ing constraints (2.10) by

∀S ⊂ R, 2 ≤ |S| ≤ R − 2: xij ≥ r(S) (2.11)
i∈S j ∈{V \S}

Constraints (2.11) are known as capacity-cut constraints (CCCs) and they ensure
both the elimination of subtours as well as vehicle capacity limits. The value r(S)
defines the minimum number of vehicles required to service all requests of set S.
It can be calculated by solving a bin-packing problem for determining the min-
imum number of required bins, each with capacity C, for loading the items of
set S (cf. Toth and Vigo 2002, p. 7) or by using a trivial lower bound given
i∈S di
by r(S) = C . Note that setting C = ∞ leads to constraints (2.10).
Despite requiring only the same amount of binary variables as in a TSP, the de-
scribed two-indexed formulation has certain drawbacks. Specifically, since no tour-
specific attributes can be integrated, e.g., individual vehicle arrival times at customer
requests, it does not allow to model more complex variants of the VRP. In order to
overcome this limitation, a three-index vehicle flow formulation is utilized (cf. Toth
and Vigo 2002, p. 15). In this formulation, let X = {xij k } be the set of binary vari-
ables that are 1 if and only if vehicle k ∈ K directly travels from location i to loca-
tion j (i, j ∈ V , i = j ). Furthermore, each binary variable of the set Y = {yik } is
1 if and only if vehicle k ∈ K visits location i ∈ V . Using this notation, the CVRP
can be modeled as follows:

min z = cij · xij k s.t. (2.12)
i∈V j ∈V k∈K

∀i ∈ R: yik = 1 (2.13)
k∈K

y0k = K (2.14)
k∈K
22 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

∀k ∈ K, i ∈ V : xij k = xj ik = yik (2.15)
j ∈V j ∈V

∀k ∈ K: di · yik ≤ C (2.16)
i∈V

∀S ⊆ R, h ∈ S, k ∈ K: xij k ≥ yhk (2.17)
i∈S j ∈S
/

Besides the objective function defined in (2.12), constraints (2.13) ensure that each
request is serviced by exactly one vehicle while constraints (2.14) make sure that
the depot belongs to the tour of each vehicle. Constraints (2.15) ensure the connect-
edness of the vehicle tours and by constraints (2.16) vehicle capacity restrictions are
taken into account. Subtours are eliminated by constraints (2.17).

2.3.3 The General Pickup and Delivery Problem (GPDP)

and Related Variants

In this section, we describe the General Pickup and Delivery Problem as well as two
variants denoted as the Pickup and Delivery Problem and the Dial-A-Ride Problem.
In contrast to the previously introduced routing problems, these problems consider
applications in which goods or passengers are picked up at specific locations and
need to be transported to other locations on the same tour. Hence, the vehicles are
empty at the start and at the end of their tour.

2.3.3.1 The General Pickup and Delivery Problem

The General Pickup and Delivery Problem (GPDP) was described by Savelsbergh
and Sol (1995). In the GPDP, each customer request i ∈ R consists of a total cargo
of weight q̄i which needs to be transported from a set of pickup locations Ri+ to a set
of delivery locations Ri− which are individual for each customer request. Each loca-
tion j ∈ {Ri+ ∪ Ri− } has a weight qj assigned. Positive weights assigned to pickup
locations indicate the weight of goods that have to be picked up whereas negative
weights at delivery locations denote
the quantity
of goods that need to be delivered.

Moreover, it holds that q̄i = j ∈R + qj = − j ∈R − qj . The set R + = i∈R Ri+
i i
comprises all pickup locations and set R − = i∈R Ri− is composed of all delivery
locations of all requests. Additionally, set L = R + ∪ R − unifies all request locations.
Requests are serviced by a heterogeneous vehicle fleet K = {1, . . . , K}. Specifi-
cally, each vehicle k ∈ K has an individual capacity Qk , as well as an individual
start location + and end location k − . Furthermore, we define K + = +
k k∈K k ,
− − + −
K = k∈K k , and W = K ∪ K . Hence, in the corresponding digraph G, the
set of nodes V is defined as V = L ∪ W . For all i, j ∈ V , dij denotes the travel
2.3 Node-Based Routing Problems 23

distance, tij the travel time, and cij the costs for traveling from location i to j .
Note that service times required at request locations can be easily integrated into the
travel times, hence they are not considered explicitly.
The model of the GPDP is formulated using four types of variables. Each binary
decision variable zik (i ∈ R, k ∈ K) is equal to 1 if request i is assigned to vehicle k,
0 otherwise. Each binary decision variable xij k ((i, j ) ∈ (L × L) ∪ {(k + , j )|k ∈ K,
j ∈ L} ∪ {(j, k − )|j ∈ L, k ∈ K}, i = j ) is 1 if vehicle k directly travels from lo-
cation i to j , 0 otherwise. Each non-negative variable Di (i ∈ V ) specifies the de-
parture time at location i. Finally, each non-negative variable yi (i ∈ V ) denotes the
current load of the corresponding vehicle when it arrives at location i. The GPDP
can be modeled as follows:

min z = f (x) s.t. (2.18)

∀i ∈ R: zik = 1 (2.19)
k∈K

∀i ∈ R, l ∈ Ri+ ∪ Ri− , k ∈ K: xlj k = xj lk = zik (2.20)
j ∈V j ∈V

∀k ∈ K: xk + j k = 1 (2.21)
j ∈{L∪{k − }}

∀k ∈ K: xik − k = 1 (2.22)
i∈{L∪{k + }}

∀k ∈ K: Dk + = 0 (2.23)
∀i ∈ R, p ∈ Ri+ , q ∈ Ri− : Dp ≤ Dq (2.24)
∀i, j ∈ V , k ∈ K: xij k = 1 ⇒ Di + tij ≤ Dj (2.25)
∀k ∈ K: yk + = 0 (2.26)

∀i ∈ R, l ∈ Ri+ ∪ Ri− : yl ≤ Qk · zik (2.27)
k∈K
∀i, j ∈ V , k ∈ K: xij k = 1 ⇒ yj = yi + qi (2.28)

The objective function is defined by (2.18). The authors state that depending on
the considered application, different functions can be used for f (x), for exam-
ple travel distance, travel time, or customer inconvenience. For details, we refer to
Savelsbergh and Sol (1995) and Sect. 2.4.12. Constraints (2.19) make sure that each
customer request is assigned to exactly one vehicle. Constraints (2.20) ensure that
a vehicle only visits and leaves customer request locations which are belonging to
requests assigned to the vehicle exactly once. Constraints (2.21) and (2.22) define
that each vehicle starts and ends its tour at the specified location. Constraints (2.23)
enforce that all vehicles start their tour at time point 0. Constraints (2.24) form
the request precedence constraints. Specifically, all pickup locations of a request
24 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

have to be serviced before its delivery locations are allowed to be serviced. Con-
straints (2.25) set the correct departure times at visited locations. Note that the less or
equal formulation allows the integration of time windows defined by [ei , li ], i ∈ L
and Di = max(Ai , ei ) with Ai denoting the arrival time of the servicing vehicle at
request location i. Hence, vehicles are allowed to wait at request locations if they
arrive there before ei . Constraints (2.26), (2.27), and (2.28) together form capacity
constraints.
In what follows, two special cases of the GPDP are presented. Note that the
previously described TSP and VRP also represent special cases of the GPDP.

2.3.3.2 The Pickup and Delivery Problem (PDP)

The Pickup and Delivery Problem (PDP) is a special case of the GPDP. Specifically,
in the PDP, each request i ∈ R has exactly one pickup and one delivery location and
only one depot exists at which all vehicles start and end their tour. The PDP can be
formulated by using the model of the GPDP and setting its parameters to |W | = 1
and ∀i ∈ R: |Ri+ | = |Ri− | = 1 so that all vehicles have the same start and end posi-
tion and each request consists of exactly one pickup and one delivery location.

2.3.3.3 The Dial-A-Ride Problem (DARP)

In Dial-A-Ride Problems (DARP, cf., e.g., Cordeau and Laporte 2003, 2007), appli-
cations which deal with the transportation of passengers are considered. The DARP
can be formulated by using the mathematical model of the GPDP and defining its
parameters as |W | = 1, ∀i ∈ R: |Ri+ | = |Ri− | = 1, and ∀i ∈ R: q̄i = 1 since each
request is assumed to consist of exactly one passenger. Common real-world ap-
plications of the DARP are taxi cab related problems or scheduled transportation
of patients to medical services. Since in the DARP people are transported instead
of goods, the objective function often includes specific passenger convenience at-
tributes. For example, the ride time of individual passengers can be restricted or
longer ride times penalized. According to Larsen (2000), the DARP exists in three
different versions:
• Many-to-One (MTO): In this variant of the DARP, only one destination to which
passengers are transported is considered. This case represents for example a real-
life application where different customers from a small residential area are col-
lected and afterwards transported to the same destination.
• Many-to-Few (MTF): In this case, the number of destinations is still relatively
small compared to the number of origins. In real-life applications, this case rep-
resents for example picking up customers from a residential area and transporting
them to different stops at a large shopping mall.
• Many-to-Many (MTM): This variant of the DARP can be considered as a taxi
cab system in which multiple individual passengers with different origins and
destinations have to be serviced.
2.3 Node-Based Routing Problems 25

More information on the DARP and its dynamic variants can be found in
Sects. 2.6.5.1 and 2.6.5.3.

2.3.4 Complexity of Node-Based Routing Problems

All the previously described node-based routing problems represent combinatorial

optimization problems. Specifically, since the corresponding tasks of assigning and
sequencing requests on vehicles can only be performed in a finite number of differ-
ent combinations, the number of existing solutions is also finite. Hence, one method
for determining an optimal solution consists of just evaluating all feasible solutions.
However, as we illustrate in this section, the computational time required to solve
routing problems by using complete enumeration is often too high. Combinatorial
problems can be distinguished by their complexity in order to estimate the required
computational time for solving them. According to Garey and Johnson (1979, p. 6),
a function f (n) is O(g(n)) if and only if there exists a constant c so that it holds
that |f (n)| ≤ c · |g(n)| for all values of n ≥ 0 with n denoting the length of the
input, e.g., the number of customer requests in a routing problem. According to
the authors and this notation, an algorithm is called a polynomial time algorithm if
its computational time complexity function is O(p(n)) for some polynomial func-
tion p. The authors state that all algorithms with no polynomial function bounding
their time complexity are called exponential time algorithms. Hence, the group of
exponential time algorithms covers all algorithms that have at least an exponential
time complexity (cf. Gass 2003, p. 304). Using this classification, it can be decided
whether an algorithm can be used to solve the considered problem with reasonable
computational effort. Specifically, in Table 2.1, the computational time complexity
for algorithms with a polynomial, exponential, and factorial time complexity are
illustrated. In this table, it is assumed that each computational operation consumes
the same amount of time which is defined as one microsecond, i.e., 10−6 seconds.
As can be derived from this table, complete enumeration cannot be used to solve a
given problem which belongs to the exponential classification in reasonable time if
larger problem instances are considered.
In order to link the results illustrated above to the time complexity of algo-
rithms for solving routing problems, we consider the required computational time
for finding an optimal solution of a TSP instance with n customer requests. Since
the number of feasible solutions amounts to (n − 1)!, the number of feasible solu-
tions grows in a factorial manner with n. Therefore, even TSP instances with a small
size cannot be practically solved by enumerating all possible solutions in a reason-
able amount of time (see again Table 2.1). Although more sophisticated solution
methods have been developed for solving TSPs, the TSP still belongs to the class of
strongly NP-hard combinatorial optimization problems (cf., e.g., Garey and Johnson
1979, pp. 211–214, Laporte 1992b, and Korte and Vygen 2012, p. 405). In compu-
tational complexity theory, the problem class P comprises all problems which can
be solved in polynomial time using a deterministic Turing machine while those of
26 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Table 2.1 Comparison of several polynomial and exponential time complexity functions
(cf. Garey and Johnson 1979, p. 7)
Time complexity Problem size n
function 10 20 30 40 50 60

n 0.00001 0.00002 0.00003 0.00004 0.00005 0.00006

seconds seconds seconds seconds seconds seconds
n2 0.0001 0.0004 0.0009 0.0016 0.0025 0.0036
seconds seconds seconds seconds seconds seconds
n3 0.001 0.008 0.027 0.064 0.125 0.216
seconds seconds seconds seconds seconds seconds
n5 0.1 3.2 24.3 1.7 5.2 13.0
seconds seconds seconds minutes minutes minutes
2n 0.001 1.0 17.9 12.7 35.7 366
seconds seconds minutes days years centuries
3n 0.059 58 6.5 3,855 2 · 108 1.3 · 1013
seconds minutes years centuries centuries centuries
n! 3.6288 771 8.4 · 1016 2.6 · 1032 9.6 · 1048 2.6 · 1066
seconds centuries centuries centuries centuries centuries

the problem class NP can only be solved in polynomial time by a non-deterministic

Turing machine (cf., e.g., Garey and Johnson 1978; 1979, pp. 27 and 31). More-
over, a problem is denoted as NP-hard if all problems of the problem class NP are
reducible to the considered problem. According to Lenstra and Rinnooy Kan (1979),
a problem A is defined as reducible to problem B if for each instance of problem A,
an instance of problem B can be generated in polynomial time so that solving the
instance of problem B will also solve the instance of problem A. Hence, the re-
ducibility of problem A to problem B implies that problem A represents a special
case of problem B. Therefore, problem B is at least as hard as problem A. Further-
more, a problem is NP-complete if it is NP-hard and also belongs to problem class
NP. For more information on computational complexity theory, see, e.g., Garey and
Johnson (1978; 1979, pp. 23–33) and Lenstra and Rinnooy Kan (1979).
An NP-hard combinatorial optimization problem is called strongly (or unary)
NP-hard if it remains NP-hard even when all input numbers are encoded in unary,
cf., e.g., Garey and Johnson (1978; 1979, p. 120) and Lenstra and Rinnooy Kan
(1979). Otherwise, NP-hard combinatorial problems are called weakly (or binary)
NP-hard if a pseudo-polynomial time algorithm for solving the problem is known:
“A pseudo-polynomial time algorithm can be useful even when there is no natural
bound on the input numbers we expect. It will display ‘exponential behavior’ only
when confronted with instances containing ‘exponentially large’ numbers, and in-
stances of this sort might be rare for the application we are interested in. If so, this
type of algorithm might serve our purposes almost as well as a polynomial time
algorithm” (Garey and Johnson 1979, p. 91).
2.3 Node-Based Routing Problems 27

Node-based
routing problem
solution methods

Exact methods Heuristic methods

Branch&Bound
Branch&Cut Decomposition Dynamic Construction Improvement
Branch&Price approaches programming heuristics heuristics
Branch&Cut&Price

Column Classic
Lagrangian
generation Simultaneous Successive improvement
relaxation
heuristics

Sequential Route ﬁrst–

Metaheuristics
cluster second

Cluster ﬁrst–
Parallel
route second

Fig. 2.4 Classification of different solution methods for node-based routing problems (in accor-
dance with Richter 2005, p. 42 and Aberle 2009, p. 543)

Since the TSP is a strongly NP-hard combinatorial optimization problem as men-

tioned above, all currently known algorithms for solving TSPs to optimality have at
least an exponential computational time complexity in the worst case. Note that
since all of the previously described routing problems, such as the VRP and the
GPDP, are extensions of the TSP, these problems are also strongly NP-hard com-
binatorial problems. Nevertheless, by utilizing specific characteristics of these rout-
ing problems or by accepting close-to-optimum solutions, various solution tech-
niques which are applicable for solving practical-sized problem instances in reason-
able time have been developed for the considered node-based routing problems. An
overview of various solution techniques is presented in the next section.

2.3.5 Solution Methods for Node-Based Routing Problems

In order to solve node-based routing problems, various exact and heuristic solution
methods have been developed (cf., e.g., Laporte 1992a, Toth and Vigo 2002, Richter
2005, Cordeau et al. 2007b, and El-Sherbeny 2010). Figure 2.4 gives an overview of
several techniques developed for solving node-based routing problems exactly and
heuristically.
28 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

2.3.5.1 Exact Solution Methods

Despite their exponential computational time complexity in the worst case, exact
solution methods determine an optimal solution of a routing problem in finite time.
Using problem-specific characteristics, they can be used to derive an optimal solu-
tion for certain routing problems within reasonable time in the average case. More-
over, they usually provide information about the solution quality of the currently
best known solution during the solution process by calculating the current gap. This
gap denotes the maximum percentage deviation from the optimal objective function
value. Specifically, this gap is determined by calculating the deviation of the cur-
rently best known solution’s objective function value from the value of a calculated
lower bound of the considered problem (cf. Toth and Vigo 2002, p. 159).
One class of exact solution methods are Branch-and-Bound (B& B) and related
approaches which were first introduced by Land and Doig (1960) and extended by
Dakin (1965) for solving MIP problems using tree-search approaches. In general,
first a relaxed version of the MIP problem, e.g., its linear relaxation, is solved in
the root node of the search tree. If all variables in the solution of the relaxed prob-
lem are integer, the procedure stops. Otherwise, an integer variable xk is selected
whose value xk0 is continuous in the current relaxed solution. Using xk , two new
problems are constructed in the branching step by adding the restriction xk ≤ xk0
and xk ≥ xk0 , respectively, to the corresponding relaxed problem, thus creating
two new nodes, i.e., subproblems, in the search tree. By solving these relaxed ver-
sions of the original problem again, lower bounds for the optimal objective function
value of the original problem existing in the corresponding subtree of the current
node can be calculated. In doing so, it is known with certainty that no better solu-
tion can be found in the subtree belonging to this node if the lower bound of the
current node is higher than the objective function value of the currently best known
solution of the original problem. Hence, this subtree can be removed from further
consideration—this activity is called bounding. The aim of B&B approaches is to
find an optimal solution by examining as few nodes as possible in the search tree by
calculating tight lower bounds in each node so that subtrees of the search tree which
do not lead to an optimal solution can be early bounded (cf. Gass 2003, p. 313).
One of the first B&B related approaches for the TSP is the approach of Little et al.
(1963). In Branch-and-Cut (B& C) approaches, additional cutting planes are subse-
quently integrated in order to strengthen the relaxed problem solved in each node
in order to calculate better lower bounds. In Branch-and-Price (B&P, cf. Barnhart
et al. 1998), column generation is applied in each node with the same aim of achiev-
ing better lower bounds. For example, Savelsbergh and Sol (1998) propose a B&P
approach for solving a real-life GPDP. Furthermore, both techniques can be com-
bined to Branch-and-Cut-and-Price (B&C&P) approaches (cf., e.g., Fukasawa et al.
2006).
Lagrangian relaxation and column generation based approaches belong to the
class of decomposition approaches. Lagrangian relaxation (cf. Fisher 2004) is of-
ten used as a sub-procedure in order to determine good lower bounds for B&B and
related approaches. Specifically, by removing certain restrictions from the original
2.3 Node-Based Routing Problems 29

problem, a relaxed problem which can be solved significantly faster than the origi-
nal problem is created. Due to the removed restrictions, an optimal solution of the
relaxed problem is likely to be infeasible for the original problem. Nevertheless, it
represents a lower bound of the optimal solution’s objective function value of the
original problem. In order to reduce such infeasibilities, a violation of the removed
restrictions is penalized in the objective function of the relaxed problem. The chal-
lenge then is to find good Lagrange multipliers by which these penalties are inte-
grated into the objective function of the relaxed problem so that infeasibilities with
regard to the original problem are reduced as much as possible. In so doing, the
calculated lower bounds are tightened, i.e., they are closer to the objective function
value of an optimal solution of the original problem thereby gaining a higher quality.
In order to find such good Lagrange multipliers, subgradient methods are frequently
applied. One approach for computing tight lower bounds for the Symmetric Trav-
eling Salesman Problem (STSP) using Lagrangian relaxation is proposed by Held
and Karp (1970). In this approach, the problem is relaxed to a Minimum Spanning
Tree (MST) problem which can be solved in polynomial time (cf. Sedgewick 1983,
p. 413). Since the solution of an MST often disregards tour connectivity charac-
teristics of a TSP solution, this Lagrangian relaxation-based approach is used as a
sub-procedure of a B&B solution approach proposed in Held and Karp (1971).
Column generation (cf. Desaulniers et al. 2005) is another solution technique for
reducing the average computational effort which is required to compute an optimal
solution of combinatorial optimization problems. Starting with a small subset of
decision variables, duality costs of the generated optimal solution of this reduced
problem are examined in order to add further promising decision variables which
lead to better solutions. Column generation has been successfully applied to VRPs,
e.g., by Krumke et al. (2002) and Westphal and Krumke (2008). In these approaches,
binary decision variables are used to select tours out of a given set of available tours.
Specifically, set partitioning is utilized in order to construct feasible tour plans out
of the currently available tours. New tours, i.e., columns, are subsequently gener-
ated by extending the set of available tours with tours comprising more customer
requests. Since in the practical application considered in both approaches optimal
solutions often consist of tours with a low number of requests, optimal solutions
are often achieved within acceptable time limits. Note that this is one example of
efficiently utilizing the given problem structure in developing practically applicable
solution methods. More information about the mentioned approaches can be found
in Sect. 4.6 of the literature review.
Dynamic programming (cf. Bellman 1954, 2003) can be applied to combinato-
rial problems which can be modeled as multi-stage decision processes. The current
state of such a process is determined by a set of quantities denoted as state vari-
ables (cf. Bellman 1954, p. 503). The current stage is advanced by making deci-
sions which transform the state variables. According to Bellman (1954, p. 503), a
“sequence of decisions [is] called a policy, and a policy which is most advantageous
according to some preassigned criterion [is] called an optimal policy” which repre-
sents an optimal solution of the considered problem. Moreover, the author states the
principle of optimality as follows, “An optimal policy has the property that whatever
30 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

the initial state and initial decisions are, the remaining decisions must constitute an
optimal policy with regard to the state resulting from the first decisions” (Bellman
1954, p. 504). According to Lew and Mauch, this also means that optimal policies
have optimal subpolicies, “That this observation is valid follows from the observa-
tion that, if a policy has a subpolicy that is not optimal, then replacement of the
subpolicy by an optimal subpolicy would improve the original policy” (Lew and
Mauch 2007, p. 5). Hence, the optimal policy is determined by making optimal de-
cisions in the current stage by using the optimal decision of previous stages with
regard to the considered objective function. This procedure is also known in the
literature as the optimal substructure property. In each stage, the size of the consid-
ered subproblems is increased so that an optimal solution of the original problem is
finally calculated. One issue is that in NP-hard routing problems “the cardinality of
the state space is usually exponential in the problem size” (Desrochers et al. 1988).
If problem-specific characteristics can be used to limit the set of feasible subprob-
lems early, i.e., removing states which cannot be extended to a feasible solution of
the original problem, these characteristics can be utilized to reduce the number of
states required to be solved for calculating an optimal solution of the original prob-
lem. For instance, the approach of Dumas et al. (1995) for solving a TSP with hard
time windows uses different feasibility tests in order to detect infeasible states due
to time window violations. At each stage, the set of currently considered requests
is increased until finally an optimal policy for all requests of the original problem
can be derived using the calculated optimal policies for the previously considered
subproblems. If tight time windows are present in a problem instance, the approach
is able to significantly reduce the search space and hence also the computational
time required to solve the problem instance to optimality.

2.3.5.2 Heuristic Solution Methods

In contrast to exact solution approaches, heuristic solution approaches aim at find-

ing good solutions quickly, often without providing information about the gap to
optimality. Heuristic solution methods for routing problems can be divided into con-
struction heuristics and improvement heuristics. Construction heuristics generate a
tour plan from scratch while improvement heuristics improve a given solution.
Construction heuristics can be classified into simultaneous and successive ap-
proaches. In simultaneous approaches, the assignment and sequencing of requests
is performed at the same time. Such approaches can be further distinguished accord-
ing to how the tours are constructed. Specifically, if the tours are constructed one
after another, a construction heuristic is called sequential. If the vehicles’ tours are
constructed at the same time so that a pending request can be inserted in different
tours, the construction heuristic is called parallel. Apart from simultaneous con-
struction heuristics, successive construction heuristics exist in which the assignment
and sequencing tasks are consecutively performed. For example, if a route first–
cluster second heuristic is applied to a VRP, a TSP is solved first resulting in a
so-called giant tour which determines the sequence of all requests. In a second step,
2.3 Node-Based Routing Problems 31

restrictions such as capacity constraints are used to divide the giant tour into indi-
vidual vehicle tours. In cluster first–route second heuristics, first the assignment of
requests to vehicles is performed by considering existing constraints. Afterwards,
a TSP is solved for each vehicle tour. Further information and examples of con-
struction heuristics can be found among others in Bräysy and Gendreau (2005a,
pp. 6–10).
Besides construction heuristics, improvement heuristics are used to modify an
existing solution in order to improve its solution quality. Improvement heuristics
can be categorized into classic improvement heuristics and metaheuristics. What
both types of heuristics have in common is that a given current solution is modified
using one or more neighborhood operators. The set of solutions that can currently
be generated by using a specific neighborhood operator is called the neighborhood
of the current solution. After evaluation of the neighborhood, one solution out of the
neighborhood is selected as the new current solution. Using this solution as the new
current solution, the described neighborhood search process starts all over again.
Classic improvement heuristics (cf. Bräysy and Gendreau 2005a) modify a given
tour plan until a local optimum is reached, i.e., until there exists no solution in the
neighborhood of the current solution with a better objective function value than
the current solution. Besides classic improvement heuristics, various metaheuristic
concepts for routing problems have been developed. The basic idea is to embed one
or more problem-dependent neighborhood operators into a problem-independent
framework which controls the application of these neighborhood operators. The
purpose of these frameworks is to overcome local optima in order to find better
solutions in other areas of the solution space. In Hoos and Stützle (2005), a large
variety of known metaheuristics is described where the most popular are: Vari-
able Neighborhood Descent (VND), Simulated Annealing (SA), Tabu Search (TS),
Greedy Randomized Adaptive Search Procedure (GRASP), Ant Colony Optimiza-
tion (ACO), and Memetic Algorithms (MAs). Furthermore, metaheuristic solu-
tion approaches for routing problems are also based on Genetic Algorithms (GAs,
cf. Mitchell 1998) and Variable Neighborhood Search (VNS, cf. Hansen and Mlade-
novic 2001). Another sophisticated solution technique is Large Neighborhood
Search (LNS) which has for example been applied to routing problems by Bent and
van Hentenryck (2004a). Ropke and Pisinger (2006) utilize LNS in a modified ver-
sion called Adaptive Large Neighborhood Search (ALNS). Further information on
metaheuristics for routing problems can be found in Toth and Vigo (2002), Cordeau
et al. (2005), Bräysy and Gendreau (2005b), Gendreau et al. (2006), and Gendreau
et al. (2008). The Tabu Search metaheuristic developed and utilized in the real-time
control approaches will be presented in Chap. 7 of this book.
The solution of a routing problem can be regarded as a graph in which cus-
tomer requests and the depot represent the nodes of the graph and in which the
arcs represent the routes, i.e., the vehicle travel activities between the relevant lo-
cations. In doing so, neighborhood operators which are applied in improvement
heuristics can be subdivided into arc-exchanging and node-exchanging operators.
In arc-exchanging operators, two or more arcs which, depending on the operator,
belong to the same vehicle tour or to different tours, are exchanged so that one or
32 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

more customer requests are indirectly moved to a different position in the tour plan.
In contrast, node-exchanging operators directly move customer requests to another
position in the same tour or in another tour. The following neighborhood operators
are frequently applied in improvement heuristics for VRPs:
• 2-opt and 3-opt operator: These operators introduced by Lin (1965) exchange
two or three arcs within the same tour.
• Or-opt operator: This operator proposed by Or (1976) exchanges three arcs in
the same tour in order to move a sequence of customer requests to another po-
sition on the same tour while preserving the sequence in which the exchanged
customers are serviced.
• 2-opt* operator: This operator by Potvin and Rousseau (1995) exchanges the
latter part of two different tours while preserving the sequence of the requests in
the exchanged tour parts.
• Relocate, exchange, and cross operator: Savelsbergh (1992) introduces these
three node-exchanging operators. The relocate operator moves one customer to
another tour while the exchange operator swaps two customers in two different
tours preserving their positions in the tours. The cross operator works similar
to 2-opt*.
• CROSS-exchange operator: Badeau et al. (1997) propose the CROSS-exchange
operator. It considers two tours in which a part of each tour is exchanged. Both
exchanged tour parts can have a different number of customers. Since one part
can also consist of zero customers, it is possible to move one or more customers
from one tour to another thereby generalizing other neighborhood operators such
as the previously mentioned relocate, exchange, Or-Opt, and 2-opt* operator.
The neighborhood operators which are utilized in the proposed Tabu Search
metaheuristic will be presented in Sect. 7.2.

2.4 Extensions to the Vehicle Routing Problem

As mentioned above, since RDOPG applications can be modeled as a variant of the

VRP, the remainder of this chapter mainly deals with characteristics and aspects
of VRPs. In this section, several extensions and special cases of the classic VRP
and other objective functions are presented. For further variants and explanations
we refer to Toth and Vigo (2002, p. 177), Baldacci et al. (2010), and El-Sherbeny
(2010).

2.4.1 Time Windows

In VRPs with time windows, a time window [rie , ril ] is assigned to each customer
request i ∈ R. We denote the point in time when the servicing vehicle arrives at
2.4 Extensions to the Vehicle Routing Problem 33

request i by yis . In general, two types of time windows exist. In case of hard time
windows, service of a request is required to start before the end of the time win-
dow, i.e., ∀i ∈ R: yis ≤ ril . Tour plans which do not fulfill these hard time window
restrictions are infeasible. If a vehicle arrives at the location of request i before rie , it
is usually allowed to wait there. This problem variant is called the Vehicle Routing
Problem with Time Windows (VRPTW, cf. Toth and Vigo 2002, p. 8). The sec-
ond variant is the Vehicle Routing Problem with Soft Time Windows (VRPSTW,
cf. Toth and Vigo 2002, p. 179). In the VRPSTW, service of requests is allowed
to be performed after the end of the time window but at the expense of additional
penalty costs. These penalty costs may either be fixed, i.e., occurring for each late
serviced request regardless of the length of the lateness, variable, i.e., depending on
the length of the lateness, or a combination of both. More information on fixed and
variable penalty costs are given in Sect. 2.4.12.
Another variant is the Vehicle Routing Problem with Multiple Time Windows in
which each request i can be serviced during different time windows (cf. Toth and
Vigo 2002, p. 179). A further variant of the VRPTW is the Vehicle Routing Problem
with Time Deadlines (VRPTD, cf. Thangiah et al. 1993), where each request i can
be serviced at any point in time up to a given latest time point denoted by ril . This
variant can be modeled using a VRPTW and setting rie = 0 for all requests i ∈ R.

2.4.2 Backhauls

In the Vehicle Routing Problem with Backhauls (VRPB), both mentioned real-world
business cases of VRPs dealing with delivering and collecting goods at customer
locations are combined. For this purpose, the set of requests is split into the ones
requiring linehauling (delivering goods) and those that require backhauling (col-
lecting goods). In VRPB problems, usually all linehaul customer requests must be
serviced before the backhaul requests (cf. Toth and Vigo 2002, p. 9, Parragh et al.
2008).

2.4.3 Simultaneous Pickup and Delivery

The Vehicle Routing Problem with Simultaneous Pickup and Delivery (VRPSPD)
is an extension of the VRPB. Specifically, in this problem, backhaul requests can
already be collected by a vehicle before it has delivered all linehaul requests (cf. Toth
and Vigo 2002, p. 10).

2.4.4 Multi-Depot Problems

In the Multi-Depot Vehicle Routing Problem (MDVRP, cf. Golden et al. 2008,
p. 164), dispatchable vehicles are located at different depots. There may be special
34 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

restrictions for assigning requests to vehicles. Specifically, certain requests may re-
quire service by a vehicle from a specific depot, for example if goods to be delivered
are only available at certain depots.

2.4.5 Open Routing Problems

In the Open Vehicle Routing Problem (OVRP), vehicles do not need to return to their
depot after completing their assigned tour. This problem is for example considered
in the work of Tarantilis et al. (2004).

2.4.6 Vehicle Scheduling

While in VRPs decisions only comprise the assignment and sequencing of requests
to vehicles, in Vehicle Routing and Scheduling Problems (VRSP) scheduling deci-
sions of vehicles are additionally considered (cf. Bodin and Golden 1981, Solomon
1987, and Mitrovic-Minic et al. 2004). Specifically, vehicles leave the depot and
customer locations only at specified time points. Vehicle scheduling decisions are
useful for different purposes. For example, in order to consider law-regulated driv-
ing time restrictions (cf. Brandão and Mercer 1997 and Sect. 2.4.10), vehicles can
be instructed to wait for a certain amount of time at a specified request location.
Furthermore, this allows the implementation of waiting strategies (see Sect. 4.3).
Note that the integration of scheduling significantly increases the decision space so
that the problem complexity is increased.

2.4.7 Multiple Compartments

In the VRP variant denoted as the Multiple Compartment Vehicle Routing Prob-
lem (MCVRP, cf. Muyldermans and Pang 2010), vehicles have different storage
areas for goods. This is important for transporting goods on the same vehicle which
must not be mixed, e.g., different liquids.

2.4.8 Multiple Trips

In Multiple Trip Vehicle Routing Problems, vehicles can return to the depot during
the planning horizon and set for another trip afterwards. Such intermediate returns
to the depot are useful if distribution problems are considered in which the load
of all requests assigned to a vehicle’s tour exceeds its capacity. Hence, the vehicle
2.4 Extensions to the Vehicle Routing Problem 35

can replenish its capacity and continue to service further requests. Furthermore,
vehicle scheduling decisions may become relevant, e.g., for determining when the
vehicle sets for its next trip. For example, Brandão and Mercer (1997) consider a
Multiple Trip Vehicle Routing and Scheduling Problem (MTVRSP) in which legal
time breaks and maximum legal driving times per day are additionally considered.

2.4.9 Time-Dependent Travel Times

VRPs with time-dependent travel times are denoted as Time-Dependent Vehicle

Routing Problems (TDVRP). A characteristic of these problems is that the same
route requires a different amount of time depending on the current time of the
day (cf. Malandraki and Daskin 1992, Ichoua et al. 2003). Specifically, the travel
time between two relevant locations i, j ∈ V is a function tij (τ ) which depends on
the current time τ at which the travel activity starts at location i. In order to model
time-dependent travel times, the time during the day is often divided into m time in-
tervals T = {T1 , T2 , . . . , Tm }. Within each of these intervals t ∈ T , the travel times
are assumed to be constant. Note that when modeling time-dependent travel times it
is reasonable to consider the first in-first out (FIFO) property, i.e., a later departure
at location i does not lead to an earlier arrival at location j (cf. Ichoua et al. 2003
and Fleischmann et al. 2004a). Depending on the type of problem, time-dependent
travel times can be known in advance with certainty, with some uncertainty, or even
completely unknown (see also Sect. 2.5.2.1). Further approaches which consider
time-dependent travel times can be found in the survey of Flatberg et al. (2005). If
time-dependent travel times are present, scheduling decisions can become important
in order to reduce travel times.

2.4.10 Legal Driving Time Regulations

In order to comply with legal regulations, different VRP approaches consider le-
gal driving time regulations (cf. Brandão and Mercer 1997 and Ferrucci 2006,
pp. 38–42). Depending on the country, these regulations consist of a maximum driv-
ing time per work shift or legal time breaks required after a certain driving time has
been completed.

2.4.11 Heterogeneous Fleet

In the Heterogeneous Vehicle Routing Problem (HVRP), the available vehicles are
divided into different classes. Differences between vehicle classes consist for ex-
ample of different fixed and variable operational costs as well as individual loading
capacities (cf. Gendreau et al. 1999b and Tarantilis and Kiranoudis 2007) and dif-
ferent travel speeds (cf. Ferrucci 2006, pp. 36–37).
36 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

2.4.12 Objective Functions in Vehicle Routing Problems

As mentioned in Sect. 2.3.3, different objective functions are pursued in VRPs ac-
cording to the considered application. In what follows, we present different types of
objective functions. Note that these individual objectives can also be combined by
utilizing a weighted sum of different objective functions.

2.4.12.1 Minimization of Travel-Dependent Parameters

Besides the minimization of operative transportation costs denoted by cij as has

been mentioned for example in the description of the TSP and the VRP, the mini-
mization of travel distances dij or travel times tij is often considered (i, j ∈ V ).

2.4.12.2 Minimization of the Number of Utilized Vehicles

This objective function aims at minimizing the number of vehicles which are uti-
lized to service the requests. This aim is often combined with the minimization of
the total travel distance (cf., e.g., Homberger and Gehring 2005) or the total travel
time (cf., e.g., Hvattum et al. 2006).

2.4.12.3 Minimization of the Sum of Tour Durations

This objective pursues the minimization of the total time utilized by all vehicles in
order to service the requests (cf. Savelsbergh 1992). If time windows are present,
vehicle scheduling aspects become relevant.

2.4.12.4 Minimization of the Completion Time

This objective function aims at minimizing the time point at which the last vehicle
returns to the depot after all requests have been serviced. Note that if in the pre-
viously described objective all vehicles have to depart from the depot directly at
the beginning of the execution of the transportation process, this and the previous
objective become equal (cf. Savelsbergh 1992).

2.4.12.5 Minimization of Lateness Costs

In VRPSTWs, costs occurring for late serviced requests can be classified into fixed
and variable lateness costs. We denote the length of the lateness that has occurred
at request i by rilt = max(0, yis − ril ) where, as mentioned above, yis describes the
2.4 Extensions to the Vehicle Routing Problem 37

start time of service at request i. Furthermore, we define F (x) as a function rep-

resenting the amount of the penalty costs that occur at a lateness of length x. By
using the objective function min z = i∈R F (rilt ), fixed penalty costs of amount P
can be modeled by defining F (x) = P · (x) with (x) = 1 if x > 0 and 0 oth-
erwise. Accordingly, variable penalty costs can be modeled by defining F (x) as a
monotonously increasing function, e.g., a linear or a quadratic function. As men-
tioned above, both types of penalty costs can also be combined.

2.4.12.6 Minimization of the Number of Unserviced Customers

This objective function is related to the previously presented one, but instead of soft
time windows, hard time windows exist. Hence, late service of a customer is not
allowed. Instead, if it holds that yis > ril , request i is removed from the system and
is counted as an unserviced customer. Hence, in order to minimize the number of
unserviced customers, it is the primary objective to service as many customers as
possible within their time windows. In general, this objective can be modeled by
fixed penalty costs which have been introduced in the previous section. However, it
is advisable to provide more information to the solution method about how urgent
a customer is to the solution method, i.e., how far away customer request i is still
from its last allowed time point for start of service denoted by ril . The utilization of
such additional information has also been addressed in van de Klundert and Wormer
(2010). In their work, the authors state that just using fixed penalty costs in order
to fulfill maximum request response times defined by SLAs (which is related to
the considered objective in this book, see also Sect. 1.1) is not advisable. By inte-
grating appropriate additional information into the objective function, the solution
method can precisely discriminate between different solutions in which all customer
requests are feasibly serviced within their time windows. Therefore, the solution
method is able to identify solutions in which requests are currently scheduled for
service with larger buffer times before the end of their time windows. This is also
advantageous in the considered RDOPG applications since it increases the possibil-
ities for integrating future arriving requests into the tour plan within the maximum
allowed response time so that no penalty costs occur (see also Sect. 5.3.2.4).

2.4.12.7 Minimization of Customer Inconvenience and Request Response

Time

As mentioned in Chap. 1, the considered RDOPG applications aim at servicing re-

quests so that customer inconvenience is minimized and a maximum request re-
sponse time whose violation causes penalty costs is maintained if possible. While
a maximum allowed request response time can be represented by using fixed late-
ness costs, customer inconvenience can be modeled by using variable lateness costs
(see Sect. 2.4.12.5) and defining rie = ril = ria with ria denoting the point in time
when customer request i arrives in the system. Depending on the utilized customer
38 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

inconvenience function in the objective function min z = i∈R F (rilt ), different ob-
jectives can be achieved. Specifically, if a linear customer inconvenience function
for F (x) is assumed, the pursued aim is equal to the minimization of the average re-
quest response time. In contrast, by utilizing an over-linear customer inconvenience
function such as a quadratic function for F (x), more fairness with regard to the
request response times of individual requests can be attained since the longer a cus-
tomer waits, the quicker the inconvenience grows. More information on the eval-
uated customer inconvenience functions and the utilized objective function in the
considered RDOPG applications will be presented in Sect. 5.3.2.4. Different effects
on the solution quality resulting from the application of these customer inconve-
nience functions will be analyzed in the computational experiments in Chap. 8.

2.5 Information Revelation in Routing Problems

In all of the previously described routing problems, the solution is obtained by solv-
ing the corresponding model on basis of problem data which is given for the pa-
rameters of the model. In what follows, we define that this data is derived from
the relevant information of the corresponding routing problem. Depending on the
considered routing problem, the problem data is either constant or changes during
the transportation process of the considered routing problem. Moreover, we define
that the problem data changes during the execution of the transportation process if
during its execution different parameter values are present in the model. Depend-
ing on the type of relevant information, such changes in the problem data may be
known in advance with certainty or can contain uncertainty. Moreover, if changes in
the relevant information are not known with certainty, stochastic knowledge about
expected changes may be available. This is explained in detail in Sect. 2.5.3.
Basically, the problem data of a routing problem consists of the following types
of information:
• Customer requests: Information about customer requests comprises the request
location, the demand, start and end times of time windows, and fixed as well as
variable penalty costs.
• Vehicles: In order to efficiently assign the requests to vehicles according to the
given objective, information about the available fleet size and properties of the
vehicles is required.
• Travel times and travel distances: This type of information is provided by the
digraph G for all connections between the considered relevant locations (see
Sect. 2.2).
In this section, we describe differences in the revelation of relevant information
in vehicle routing problems, i.e., how and when changes in the relevant information
occur in the considered routing problem during the execution of its transportation
process. Furthermore, we present approaches in the literature for characterizing rel-
evant information and distinguishing between static and dynamic routing problems.
2.5 Information Revelation in Routing Problems 39

Operative layer Real-time control layer

Initial tour plan Tour plan adaptation during tour plan execution
generation before (depending on the type of routing problem)
tour plan execution
Relevant information Relevant information may change (depending on the problem type)
does not change
time
Beginning of the transportation process

Fig. 2.5 General process of tour plan generation and adaptation before and during the execution
of the transportation process

Based on the results of these approaches, a unified approach for classifying rout-
ing problems with regard to characteristics and processing of relevant information,
which are both important for RDOPG applications, in routing problems is proposed.
Using this approach, different types of VRPs known in the literature are presented
in detail.

2.5.1 Revelation of Relevant Information in Vehicle Routing

Problems

With regard to information revelation in VRPs, a common practice is to distinguish

between the relevant information which is known before the execution of the trans-
portation process and the relevant information which only becomes known during
the execution of the transportation process (see Fig. 2.5).
Before the execution of the transportation process, an initial tour plan is gener-
ated based on the problem data given by the relevant information which is known at
that time point. With regard to the layers of transportation processes (see Sect. 1.2),
this initial tour plan generation corresponds to the operative layer. It is assumed that
the time available for generating the initial tour plan is not restricted and that the rel-
evant information does not change during this generation. After the initial tour plan
generation is complete, the execution of the transportation process starts. Specifi-
cally, the vehicles leave the depot in order to service customer requests according to
the initial tour plan. An important factor is that, depending on the considered rout-
ing problem, the existing relevant information can change during the execution of
the transportation process. If such changes are caused by events whose time of oc-
currence during the execution of the transportation process cannot be precisely pre-
dicted in advance, we denote them as dynamic events—one example of a dynamic
event is the arrival of a new request. Otherwise we state that incomplete relevant in-
formation is present. Specifically, although the relevant information does not change
by itself, some parts of it are unknown during the initial tour plan generation pro-
cess and are only revealed later. For example, the actual customer demand is only
disclosed when the vehicle reaches the corresponding customer location. Routing
problems with changing relevant information and appropriate tour plan adaptation
techniques are described in detail in Sect. 2.5.3.2. In the next section, we deal with
characteristics of relevant information as well as with differences between static and
dynamic routing problems.
40 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

2.5.2 Approaches in the Literature for Characterizing Relevant

Information and Distinguishing Between Static
and Dynamic Routing Problems

In what follows, differences in routing problems according to characteristics of rel-

evant information are presented. Moreover, approaches published in the literature
for distinguishing between static and dynamic routing problems are described.

2.5.2.1 Characteristics of Relevant Information

In Psaraftis (1995), the author introduces four different criteria for classifying rele-
vant information in routing problems:
• Evolution of information: In static routing problems, the relevant information
remains constant while in dynamic routing problems the relevant information
changes during the execution of the transportation process. More information
about differences in static and dynamic routing problems is given in the next
section.
• Quality of information (about future events): Information available about fu-
ture events is of one of four types; future events can either be known with cer-
tainty (deterministic), known with uncertainty (forecasted), follow prescribed
probability distributions (probabilistic information), or unknown (no informa-
tion).
• Availability of information: Changes which occur in the relevant information
during the transportation process can either be locally or globally available. One
example of a local information availability is if a driver is informed of the actual
amount of goods that the current customer requires only when the correspond-
ing customer location is reached and there is no possibility of informing other
vehicles or a central dispatching center about this change. In contrast, in rout-
ing problems where such changes are globally available, they are recorded in a
central system.
• Processing of information: If a centralized processing of relevant information
is applied, all changes in the relevant information are recorded and handled by a
central dispatching center. In a decentralized processing, changes in the relevant
information are autonomously handled by the vehicles.

2.5.2.2 Static and Dynamic Routing Problems

In this section, we describe different approaches proposed in the literature which

classify static and dynamic routing problems. In what follows, we present the twelve
aspects by Psaraftis (1988) in which static and dynamic routing problems can be
distinguished from each other:
2.5 Information Revelation in Routing Problems 41

1. Time dimension is essential: In a static routing problem the current time di-
mension during the execution of the transportation process is not important
since, as stated above, static routing problems only consider the generation of
the initial tour plan before the execution of the transportation process and the
relevant information does not change during its execution. In dynamic rout-
ing problems, time is essential for several reasons. First, since changes in the
relevant information often occur unexpectedly during the execution of the trans-
portation process, depending on the current time, vehicles have a different posi-
tion and status which needs to be considered in order to efficiently adapt the tour
plan according to occurring changes in the relevant information. Furthermore,
during the adaptation of the tour plan according to the new system situation,
time elapses in the transportation process. Hence, further changes in the system
situation that occur during this adaptation need to be carefully considered (see
also Sects. 2.5.3.2 and 4.7.4).
2. Problem may be open-ended: In contrast to static routing problems, in some
dynamic routing problems the length of the planning horizon is not determined
in advance.
3. Future information may be imprecise or unknown: As mentioned above,
in static routing problems all relevant information is known in advance with
certainty. In dynamic routing problems, future events are only known with un-
certainty or are not known at all before the execution of the transportation pro-
cess.
4. Near-term events are more important: In static applications where all rele-
vant information is known with certainty, no drawbacks exist in directly com-
mitting vehicles to decisions which are only taking place in the far future. For
example, a vehicle can be sent to an isolated customer request without negative
consequences, even if it needs to wait there for a long time until this request can
be serviced. In contrast, in dynamic routing problems, postponing far future de-
cisions can be advantageous in order to make better decisions later when more
relevant information is available, for example due to additional requests which
have arrived intermediately.
5. Information update mechanisms are essential: Since in static problems
the relevant information does not change, no tour plan adaptations are re-
quired during the transportation process. In dynamic routing problems, it is
necessary to integrate occurring relevant information updates into the real-
time control approach. Specifically, ongoing processes, e.g., requests currently
being serviced or vehicle movements being performed, have to be consid-
ered in tour plan adaptations. Moreover, in order to inform vehicles about
new decisions made during an adaptation, the existence of a bidirectional
communication between the vehicles and the dispatching center is impor-
tant.
6. Resequencing and reassignment decisions may be warranted: In dynamic
routing problems, changes in the relevant information may render the currently
executed tour plan infeasible so that it needs to be adapted. During this adap-
tation, existing requests may be required to be resequenced or reassigned to
42 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

other vehicles. Moreover, newly arrived requests are required to be integrated

in order to comply with the new system situation.
7. Faster computation times are necessary: As mentioned in Sect. 2.5.1, in
static routing problems, a large amount of computational time is assumed to be
available for generating the initial tour plan with the demanded solution quality.
In contrast, the available time in dynamic problems is limited so that a quick
adaptation of the tour plan according to occurred changes in the relevant infor-
mation is required (see also aspect 1). Hence, adaptations have to be carried out
by sophisticated solution methods which provide a good solution quality in a
short amount of time.
8. Indefinite deferment mechanisms are essential: In dynamic problems, it may
happen that the service of a particular request is theoretically indefinitely post-
poned. Specifically, this happens if a request possesses an unfavorable geo-
graphical location and new requests keep arriving close to the current vehicles’
positions. Hence, mechanisms which prohibit an indefinite deferment of indi-
vidual pending requests are required. For example, time window constraints or a
nonlinear objective function which generates higher penalty costs for excessive
waiting times can be utilized for this purpose.
9. Objective function may be different: In static routing problems, the con-
sidered objective often pursues the minimization of operative transportation
costs, travel distances, or travel times. In addition, different objectives are con-
sidered in specific dynamic routing problems. Specifically, in real-world dy-
namic routing applications it can be observed that the higher the number of
dynamic events, the more important customer-oriented objectives become (see
also Sect. 2.6.5).
10. Time constraints may be different: In dynamic routing problems, the unex-
pected occurrence of new requests may lead to situations in which it becomes
very difficult or impossible to service all pending requests while fulfilling given
hard time window constraints. Hence, in dynamic routing problems, time con-
straints are often soft so that service of requests is also allowed after the end
of the corresponding time windows, but at additional penalty costs. Using hard
time window constraints in dynamic routing problems allows for modeling
business applications in which customers are treated as unserviced if their time
windows are not met (see also Sect. 2.4.12.6).
11. Flexibility to vary vehicle fleet size is lower: In static problems it is often
possible to utilize an arbitrary number of vehicles in order to meet request time
window constraints. In contrast, in dynamic routing problems, a small available
fleet size often results in a lower flexibility for efficiently integrating requests
which dynamically arrive during the execution of the transportation process.
This may lead to the violation of time window constraints and lower quality of
service.
12. Queueing considerations may become important: In dynamic routing prob-
lems with a high number of dynamically arriving requests (also called heavy
traffic scenarios), situations may occur in which the arrival rate of new re-
quests exceeds a certain amount beyond which vehicles cannot service pending
2.5 Information Revelation in Routing Problems 43

requests without significant delays. Psaraftis states that in this case, classic so-
lution approaches for routing problems result in a poor system performance
so that the utilization of methods from the area of queueing theory is recom-
mendable. Bertsimas and van Ryzin evaluate the efficiency of different request
dispatching strategies in light and heavy traffic scenarios by using queueing the-
ory insights in their approaches. Their results will be discussed in the literature
review in Sect. 4.5.1.
Besides Psaraftis’ twelve aspects, Bianchi (2000) proposes another approach to
distinguish between static and dynamic routing problems. In this approach, a routing
problem (or, in general, an optimization problem) is classified as dynamic if it is
uncertain how parts of the relevant information are revealed or updated during the
execution of the transportation process. Hence, impacts of occurring changes in the
problem data are not always known with certainty. As mentioned above, if the tour
plan currently in execution is required to be adapted, this has to be carried out while
time in the system elapses so that the solution process is concurrently executed
with occurring changes in the relevant information. Since it is not sufficient to only
determine an initial solution once, only a strategy or policy, i.e., a solution method,
which is used to solve the problem during the execution of the transportation process
can be specified in advance. If this is not the case, a routing problem is deemed not
dynamic according to this classification. Therefore, routing problems in which one
of the following two characteristics is present are not dynamic according to the
author:
• The problem data changes over time, but these changes are known with certainty
in the relevant information.
• The relevant information is constant, but some parts of it are unknown at the
beginning of the execution of the transportation process and are only revealed
during its execution, for example when a vehicle reaches a customer location.
Moreover, for all of these uncertain parts, prescribed probabilistic information is
given as stochastic knowledge in advance. Hence, an a priori solution which is
only changed in minor ways during the execution of the transportation process
can be determined in advance (see Sect. 2.5.3.2.1).
Another approach for characterizing static and dynamic routing problems is pro-
posed in the work of Ghiani et al. (2003). The authors state that a routing problem is
static if all types of its relevant information are independent of the current time dur-
ing the execution of the transportation process. If this is not the case, then the routing
problem is dynamic. Moreover, if all changes in the relevant information which oc-
cur during the execution of the transportation process are known in advance with
certainty, the routing problem is called deterministic, otherwise stochastic. Conse-
quently, four types of routing problems arise in this characterization approach:
• In static deterministic routing problems, no changes in the relevant information
occur during the execution of the transportation process so that the problem data is
constant. Hence, only the generation of the initial tour plan before the execution
of the transportation process is considered in problems of this type. Problems
44 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

of this type are also denoted as static routing problems which are presented in
Sect. 2.5.3.1.1.
• In static stochastic routing problems, some parts of the relevant information are
only known with uncertainty before the execution of the transportation process.
This uncertain relevant information often belongs to customer attributes, e.g.,
demand of customers or whether customers are required to be serviced at all.
Characteristic for these types of routing problems is that the uncertain relevant
information is given by random variables that follow prescribed probability dis-
tributions which makes the problem stochastic. Moreover, the prescribed prob-
ability distributions are often available as stochastic knowledge before the ex-
ecution of the transportation process. Since the values of these random vari-
ables do not change during the execution of the transportation process, the prob-
lem is characterized as static. The actual values of the uncertain relevant in-
formation are revealed during the execution of the transportation process. For
example, when a vehicle reaches a customer location, its actual demand is re-
vealed. Hence, the problem data changes in this type of problem, but these
changes can be considered in the initial tour plan generation due to the avail-
able stochastic knowledge. Routing problems of this type are known in the liter-
ature as variants of the Stochastic Vehicle Routing Problem which is described in
Sect. 2.5.3.2.1.
• In dynamic deterministic routing problems, the relevant information remains con-
stant, but the problem data changes over time during the execution of the trans-
portation process. Characteristic of such routing problems is that the point in time
when these changes in the problem data happen as well as their effects are known
in advance with certainty. For example, routing problems in which travel times
change during the planning horizon according to a function which is known in
advance belong to this problem type. Routing problems of this type are described
in Sect. 2.5.3.1.2.
• In dynamic stochastic routing problems, the relevant information changes at un-
known time points during the execution of the transportation process when dy-
namic events occur. Moreover, no knowledge about these dynamic events is avail-
able. Such problems are described in Sect. 2.5.3.2.2.
Pillac et al. (2010) propose an approach similar to the previous one. Specifi-
cally, static routing problems are classified as described in Ghiani et al. (2003).
In dynamic routing problems, relevant information changes during the execution
of the transportation process and these changes always inhibit uncertainty. Specif-
ically, dynamic deterministic routing problems are characterized by the fact that
no information about future changes in relevant information is available. Hence,
this group of routing problems is equal to the dynamic stochastic problems in
the aforementioned classification. In contrast, this classification regards dynamic
stochastic routing problems as real-time routing problems in which some stochastic
knowledge about future events is available. Problems of this type are presented in
Sect. 2.5.3.2.2.
2.5 Information Revelation in Routing Problems 45

2.5.2.3 Dynamism of the Problem, Model, and Application in Routing

Problems

In Powell et al. (1995), the authors do not only deal with the type of changes of the
problem data, which they denote as the problem, during the transportation process,
but they also categorize the approach utilized to solve the given problem. Specif-
ically, besides a model which is used to represent the problem in the solution ap-
proach, the application defines the way in which the approach solves the model
during the transportation process. Hence, in a given routing application, the prob-
lem is predetermined while the model and the application are approach-dependent.
In the proposed categorization, the authors distinguish each of these three items
according to the type of dynamism with which each of them exists in a given routing
problem and the applied solution approach:
• Dynamism of the problem: The problem is dynamic if one or more of its param-
eters change over time during the execution of the transportation process accord-
ing to functions of time which may be known or unknown before its execution.
Two types of dynamic problems are distinguished by the authors: Problems with
dynamic data and problems with time-dependent data. In the first type, the rele-
vant information and hence the problem changes unexpectedly, i.e., the mentioned
functions are unknown. For instance, this includes dynamically arriving requests,
unexpected changes in traffic conditions, or unexpected changes of vehicles or
driver statuses. In dynamic problems with time-dependent data, changes in the
problem occur deterministically according to functions which are known in the
relevant information before the execution of the transportation process.
• Dynamism of the model: The model represents the problem in a specific solu-
tion approach. It is dynamic if it explicitly considers changes in the problem dur-
ing the execution of the transportation process whereas a static model does not
consider any occurring changes in the problem. Furthermore, models are subdi-
vided into deterministic and stochastic models so that four types of models exist.
Specifically, a deterministic static model can be used to model static routing prob-
lems in which the relevant information does not change and also the problem is
static. A stochastic static model represents a routing problem in which uncertain
parts in the relevant information are considered by generating a robust tour plan
with regard to expected changes in the problem that occur according to available
probability distributions. By using a deterministic dynamic model, changes which
occur in the problem over time are considered but the model does not consider
the dynamic structure of the problem. Specifically, when solving the model, the
possible occurrence of further changes in the problem is not considered so that
only the currently deterministically known parts of the relevant information are
considered, even in situations where stochastic knowledge is available. In con-
trast, in stochastic dynamic models, decisions are made on the basis of relevant
information known with certainty and do additionally consider available stochas-
tic knowledge about expected future events.
• Dynamism of the application: The application defines how the solution ap-
proach is applied in order to solve the model. The application is dynamic if the
46 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

related model is repeatedly solved thereby considering changes in the problem

during the execution of the transportation process. In contrast, the application of
an approach is static if it does not adapt the solution during the transportation
process even when the problem has changed.
In order to illustrate examples of different solution approaches according to the
utilized model and the application, the authors discuss a routing problem in which
a vehicle has to be guided from an origin to a destination location in a road net-
work. The objective is to find the route with the minimum expected travel time.
It is assumed that the traffic conditions change over time with uncertainty. Hence,
the considered problem is dynamic. Moreover, it is assumed that stochastic knowl-
edge is available in advance by prescribed probability distributions describing the
expected changes in the travel time. The authors present different approaches for
using a specific model and a specific application for this given dynamic problem:
• Deterministic static model and static application: In such an approach, the
route is calculated once by using one particular static representation of the prob-
lem in the model, for example by using average expected travel times for the
roads in the road network. Since the application is static, occurring changes in
the problem during the execution of the transportation process are not considered
so that no adaptation of the solution is performed. In dynamic routing problems,
approaches of this type are likely to result in infeasible tour plans during the
execution of the transportation process (see Sect. 2.5.3.2). Therefore, the utiliza-
tion of this type of approach is only advisable for static routing problems (see
Sect. 2.5.3.1.1).
• Deterministic dynamic model and dynamic application: In approaches of this
type, the model is repeatedly solved during the execution of the transportation
process due to the dynamic application. The model considers changes that have
already occurred in the problem, but the available stochastic knowledge about
future expected changes is not considered. In this book, such approaches are de-
noted as reactive real-time control approaches and deterministic real-time control
approaches (see Sect. 2.5.3.2.2).
• Stochastic dynamic model and static application: These approaches consider
changes that have already occurred in the problem as well as future expected
changes. However, since the application is static, only one initial solution is com-
puted which is not adapted when changes in the problem occur during the execu-
tion of the transportation process. Hence, the aim is to generate an initial tour plan
which is robust, i.e., which is likely to represent a good solution during the exe-
cution of the transportation process even when the problem changes according to
the available prescribed probability distributions. In dynamic routing problems,
such a priori optimization-based approaches are denoted as Stochastic Vehicle
Routing Problems (see Sect. 2.5.3.2.1).
• Stochastic dynamic model and dynamic application: Comparable to the pre-
viously described approach type, the model considers changes in the problem
that have occurred as well as expected future changes. Moreover, the model is
repeatedly solved during the execution of the routing problem so that occurred
2.5 Information Revelation in Routing Problems 47

and expected changes in the traffic conditions are considered. In this book, ap-
proaches of this type are denoted as pro-active real-time control approaches (see
Sect. 2.5.3.2.2).

2.5.3 A Unified Approach for Classifying Routing Problems

with Regard to Characteristics and Processing of Relevant
Information

In the previous section, different approaches in the literature for characterizing rel-
evant information and different types of routing problems were described. Based
on the results of these approaches, in particular those by Powell et al. (1995) and
Psaraftis (1995), we present a unified approach for classifying routing problems
according to RDOPG-specific attributes in this section. This approach gives a struc-
tured overview about the type of relevant information, i.e., constant or changing, and
considers the availability and processing of occurring changes in the relevant infor-
mation in detail. Moreover, in case of changing relevant information, it describes
the type of stochastic knowledge available, if any.
The proposed classification is illustrated in Fig. 2.6. The resulting items of this
classification represent individual classes of routing problems. According to the
main distinction concerning whether the relevant information is constant or com-
prises uncertain changes during the transportation process, the type of problem data
denoted as the problem is identified. If changing relevant information exists, the
availability of resulting changes in the problem is presented. Moreover, the pro-
posed classification considers in which way the problem is processed in order to
coordinate the transportation process. Finally, depending on the quality of informa-
tion, i.e., the type of stochastic knowledge which is available about expected future
events, appropriate variants of VRPs known in the literature are described. In what
follows, characteristics for each of these classes and general solution approaches are
presented.

2.5.3.1 Routing Problems with Constant Relevant Information

In routing problems with constant relevant information, all relevant information

is exactly known before the execution of the transportation process and does not
change during its execution (see Fig. 2.7). In such routing problems, the evolution
of information is static and the constant relevant information is globally available.
Moreover, a centralized processing of this relevant information is used. In doing so,
a centralized coordination of the transportation process is applied. Since only the
initial tour plan generation process is considered in such routing problems, solution
approaches make often use of a static deterministic model and a static application.
Since the relevant information is constant and no future changes occur, the quality
of information about future events is deterministically known.
48 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Relevant Example
information is variant
globally available
and given Deterministic
by constant Centralized knowledge
Constant
relevant
parameters
/ Static problem, processing
/ Centralized available
/ VRP
no changes coordination
information K
KK
G KK M: deterministic static
KK

A: static
KK
K
Relevant information KK
%
is globally available
Dynamic Centralized
Deterministic
knowledge
Relevant information
and given by known
does not change
functions of time problem, processing
/ Centralized available
/ TDVRP
time-dependent coordination
data M: deterministic static
A: static

Prescribed
stochastic
Dynamic Decentralized knowledge
Type of
relevant
problem, local processing
/Decentralized / SVRP
available
knowledge of coordination
information
-- @ F >>
dynamic data >> M: stochastic static
--
E >>
>>
A: static
--

>>
--
>>

Local availability of
-- relevant information
changes, no bidirectional

-- communication available

Decentralized Agent-based
Relevant information -
--

processing and related

changes with approaches

uncertainty over time Decentralized

processing
--

--
Partly local, partly

global availability
of relevant Dynamic Mixed Mixture of
Changing
relevant
information changes
/ problem, processing
/ Mixed / decentralized
mixed coordination and centralized
information knowledge of A approaches
66
66 dynamic data
66
66
66
66 8 DVRP
66 No stochastic
knowledge available
66
66 M: deterministic dynamic
Global availability of 66 A: dynamic
relevant information 66 Mixed processing
changes, bidirectional 66
communication available
66
Prescribed
stochastic
Dynamic Centralized knowledge
problem, processing
/ Centralized available / DSVRP
global coordination
<<
knowledge of << M: stochastic dynamic
dynamic data << A: dynamic
<<
<<
Stochastic knowledge <
<
<<
can be generated
<<
from past request
information <
DFSVRP
N.B.:
M denotes the type of model used in an approach, M: stochastic dynamic
A denotes the type of application used in an approach (see Section 2.5.2.3). A: dynamic

Fig. 2.6 A new classification of routing problems according to properties and processing of rele-
vant information
2.5 Information Revelation in Routing Problems 49

Operative layer Real-time control layer

Initial tour plan No tour plan adaptation during tour plan execution
generation before
tour plan execution
Relevant information Relevant information does not change
does not change
time
Beginning of the transportation process

Fig. 2.7 The transportation process of routing problems with constant relevant information

As illustrated in the classification in Fig. 2.6, two types of routing problems with
constant relevant information exist. In Powell et al. (1995), they are distinguished
according to the dynamism of the problem, i.e., whether the problem is static or
dynamic.

2.5.3.1.1 Static Problem: Static VRPs

In routing problems in which the constant relevant information is given by constant

parameters, the problem data also remains constant during the planning horizon so
that a static problem, in which no changes occur, is present. One example of this
routing problem class is the CVRP which was described in Sect. 2.3.2.

2.5.3.1.2 Dynamic Problem: Time-Dependent VRPs

In routing problems of this class, the existing problem data changes during the exe-
cution of the transportation process. However, since these changes occur according
to known functions of time which are available in the relevant information before
the execution of the transportation process and since the relevant information does
not change, all problem data changes are deterministically known. Therefore, a dy-
namic problem with time-dependent data is present. An example routing problem
is the Vehicle Routing Problem with Time-Dependent Travel Times (TDVRP, see
Sect. 2.4.9).

2.5.3.2 Routing Problems with Changing Relevant Information

In addition to the previously discussed routing problems with constant relevant in-
formation, there are routing problems in which the relevant information changes
during the execution of the transportation process. What is important is that these
changes cannot be predicted with certainty. Hence, a dynamic problem with dy-
namic data in which changes in the problem data cannot be predicted with certainty
is present. According to Psaraftis, the evolution of information is dynamic and the
quality of information about future events is not deterministic and depends on the
type of available stochastic knowledge.
50 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

As mentioned in Sect. 2.5, uncertain changes in the relevant information are

brought about by one of the following two circumstances:
1. Occurrence of dynamic events: Since the time point of occurrence and the im-
pact of dynamic events cannot be predicted with certainty, dynamic events may
lead to an outdated tour plan with regard to the system situation generated by the
dynamic event. For example, a dynamic event by which a new request arrives
in the system requires an integration of this request into the tour plan. More-
over, dynamic events may render the tour plan infeasible, e.g., if a vehicle breaks
down, its assigned requests need to be reassigned to other vehicles. An overview
about different types of dynamic events is presented in Sect. 2.6.2.
2. Presence of incomplete relevant information: In addition to routing problems
in which the relevant information is changed by dynamic events, there are rout-
ing problems in which parts of the relevant information are unknown during the
initial tour plan generation process. In general, these unknown parts are random
variables whose actual values are revealed during the execution of the transporta-
tion process. They follow prescribed probability distributions which are usually
available in the initial tour plan generation process.
What both circumstances have in common is that the problem data changes with
uncertainty during the execution of the transportation process so that specific adap-
tations of the tour plan may become necessary. Therefore, the dimension of the
availability of changes in the relevant information is important since the resulting
availability of changes in the problem mainly determines the types of tour plan
coordinations applicable during the execution of the transportation process. In our
classification, there are three different values for this dimension:
• Local availability of relevant information changes: Local availability of rel-
evant information changes is present if only one particular vehicle is informed
about changes that have occurred and there is no possibility of informing a cen-
tral dispatching center due to the lack of bidirectional communication equip-
ment (see Sect. 2.6.3) or since such a central dispatching center does not exist
in the considered routing problem. Due to the local knowledge of dynamic data,
occurring changes cannot be centrally processed so that a decentralized process-
ing of changes and therefore a decentralized coordination of the tour plan is
inevitable. Hence, the vehicle driver who is informed about the changes is in
charge of processing appropriate adaptations of the vehicle’s tour. In VRPs, these
types of routing problems are represented by variants of the Stochastic Vehicle
Routing Problem presented in Sect. 2.5.3.2.1.
• Global availability of relevant information changes: In such scenarios, a cen-
tralized dispatching center is available and bidirectional communication devices
in the vehicles allow communication between the dispatching center and the ve-
hicles (see Sect. 2.6.3). In this case, a global knowledge of dynamic data exists
in the dispatching center either by centrally collecting changes (e.g., customers
directly call the dispatching center) or by vehicles transmitting locally revealed
changes to the central dispatching center. This allows for different types of co-
2.5 Information Revelation in Routing Problems 51

ordination of the transportation process which, in what follows, are explained by

the example of adapting the tour plan in case of newly arriving requests:
– Decentralized coordination: In this case, some or all of the vehicles are in-
formed about the new requests and autonomously decide which new requests
will be serviced by each vehicle; this results in a decentralized processing of
changes in the relevant information. In our classification, such routing prob-
lems are denoted as agent-based/related approaches. This type of approaches
is not discussed any further in this book. For more information, we refer for
example to the work of Fischer et al. (1996), Kohout and Erol (1999), Weiss
(2000), Perugini et al. (2003), Cubillos et al. (2007), Bloos et al. (2009), Hüls-
mann et al. (2009), and Schönberger (2010).
– Centralized coordination: By making use of a centralized processing of
changes in the relevant information, the dispatching center centrally adapts the
tour plan and transmits resulting changes to the vehicles involved. In the con-
sidered example, the new requests are centrally integrated into the tour plan
and the vehicles to which these new requests are assigned are informed after
the tour plan adaptation is complete. Routing problems which belong to this
class are variants of the Dynamic Vehicle Routing Problem. They are described
in detail in Sect. 2.5.3.2.2.
– Mixed coordination: In systems applying a mixed tour plan coordination,
some types of changes in the relevant information are centrally processed, e.g.,
newly arriving requests, while other changes are locally processed, e.g., re-
routing of a vehicle in case of traffic congestions. Such approaches can be
realized by utilizing a combination of the methodologies for decentralized and
centralized tour plan coordination.
• Partly local, partly global availability of relevant information changes: This
scenario is present if specific types of changes in the relevant information are
globally available and other types are only locally available. In such scenarios,
tour plan adaptations can be realized either by using a decentralized coordina-
tion or a mixed coordination.
In what follows, we present variants of Stochastic Vehicle Routing Problems
which utilize a decentralized coordination and variants of dynamic VRPs using
a centralized coordination of the transportation process. Furthermore, the type of
stochastic knowledge which is available in the individual routing problems is de-
scribed.

2.5.3.2.1 Decentralized Coordination: Stochastic Vehicle Routing Problems

In Stochastic Vehicle Routing Problems (SVRP), the relevant information avail-

able during the initial tour plan generation process is incomplete so that some parts
are unknown (as mentioned in aspect 2 in Sect. 2.5.3.2). In SVRPs, such unknown
relevant information can comprise uncertain servicing of customers, uncertain cus-
tomer demands, stochastic travel times, or stochastic service times (cf. Cordeau et al.
2007a). As mentioned above, an important aspect is that these unknown parts are
52 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Operative layer Real-time control layer

Initial tour plan Adaptation of individual tours during tour plan execution
generation before in case of tour failures by applying an appropriate recourse action
tour plan execution
Relevant information Relevant information does not change, but some parts are only revealed during
does not change the execution of the transportation process
time
Beginning of the transportation process

Fig. 2.8 The transportation process in Stochastic Vehicle Routing Problems

represented by random variables. Their values are unknown in the initial tour plan
generation but do not change during the execution of the transportation process.
They are revealed to individual vehicles during the transportation process which is
discussed in detail below. Moreover, these random variables follow prescribed prob-
ability distributions which are known and can be utilized as stochastic knowledge
during the initial tour plan generation. With regard to the classification of Psaraftis
(1995), the quality of information about future events follows prescribed probability
distributions (probabilistic information). Hence, the unknown parts of the relevant
information are given by distributions known to be correct with certainty (cf. Richter
2005, p. 65).
The transportation process in SVRPs is illustrated in Fig. 2.8. A decentralized
coordination is applied during the execution of the transportation process since no
technical equipment for a bidirectional communication is available. The initial tour
plan is centrally generated irrespective of this lack of technical equipment since all
vehicles are assumed to be located at the depot before the execution of the trans-
portation process. During its execution, the individual vehicles can be required to
adapt their tour when uncertain relevant information is revealed. Note that such an
adaptation is usually only performed if changes in the relevant information make
a tour infeasible; this is denoted as a tour failure. For example, this happens when
the demand of a customer exceeds the currently available capacity of the assigned
vehicle.
In case of a tour failure, adaptations denoted as recourse actions or corrective ac-
tions are required in order to re-establish a feasible tour (cf. Gendreau et al. 1996).
Due to the decentralized coordination, recourse actions are limited to local adap-
tations of the vehicle’s tour. Moreover, since often no computational equipment is
assumed to be available in vehicles, local adaptations are often carried out by uti-
lizing predefined policy-based operations consisting of specific rules which can be
directly applied by the drivers. Due to these restrictions, recourse actions often re-
sult in suboptimal tour plan adaptations which generate high additional costs and
therefore lead to a significantly reduced solution quality. Since the minimization of
operational costs is usually pursued in SVRPs, the reduction of costs generated by
recourse actions during the execution of the transportation process is desired. This
is achieved by including the available stochastic knowledge in the initial tour plan
generation process which is carried out by sophisticated solution methods. Due to
the fact that only policy-based tour plan adaptations are performed during the execu-
tion of the transportation process, SVRPs are also denoted as a priori optimization
problems (cf. Powell et al. 1995).
2.5 Information Revelation in Routing Problems 53

In order to integrate the available stochastic knowledge into the initial tour plan
generation process, SVRPs are often formulated as stochastic programs. According
to Dror (1989), stochastic programming techniques can be distinguished into wait
and see as well as here and now approaches. While in the first type of approaches
a solution is calculated after the disclosure of some uncertain relevant information,
approaches of the second type compute a solution before the uncertain relevant in-
formation is revealed. Since the a priori solution of an SVRP is generated before
unknown information is revealed during the execution of the transportation process,
SVRPs belong to the second type of approaches. For SVRPs, two main solution
techniques exist in the literature (cf. Gendreau et al. 1996):
• Chance Constrained Program (CCP): In this type of solution technique, the
aim is to generate the initial tour plan so that the chance of a tour failure during the
execution of the transportation process is below a defined threshold. A drawback
is that CCPs neglect the costs of a recourse action that are generated when a tour
failure nonetheless occurs during the execution of the transportation process.
• Stochastic Program with Recourse (SPR): This type of solution technique pur-
sues the generation of an a priori tour plan P a with minimum total expected
costs c1 (P a ) + c2 (P a ). In this formulation, c1 (P a ) describes the costs which oc-
cur in the tour plan according to the relevant information known with certainty.
Moreover, utilizing the available stochastic knowledge, c2 (P a ) represents the es-
timated costs of recourse actions which are expected to be required during the
execution of the transportation process. In general, SPRs require more complex
solution methods than CCPs but the considered objective function has a higher
practical relevance since it explicitly considers expected additional costs.
As mentioned above, different recourse actions can be performed in case of a
tour failure. The best choice for a recourse action depends on the way in which the
uncertain relevant information is revealed during the execution of the transportation
process. For example, according to Gendreau et al. (1996), a customer’s demand can
become known when service of the previous customer has been completed or when
the vehicle has arrived at the corresponding customer request location. Moreover,
applicable recourse actions also depend on the type of uncertainty in the relevant
information which is present in the considered SVRP. In the literature, the following
types of uncertainty are conducted in SVRPs (cf. Gendreau et al. 1996, Toth and
Vigo 2002, p. 332, and Cordeau et al. 2007b, p. 45 et seq.):
• Vehicle Routing Problem with Stochastic Demands (VRPSD): In VRPSDs,
the demands of customer requests are unknown. Therefore, infeasible tour plans
can occur due to capacity constraints of vehicles. This SVRP variant was first
studied by Tillman (1969) and Golden and Stewart Jr. (1978) and is by far the
most studied variant of the SVRP. More studies on the VRPSD can be found
among others in the work of Dror (1989), Secomandi (2000), Laporte et al.
(2002), Novoa et al. (2006), Christiansen and Lysgaard (2007), Mendoza et al.
(2009), and Secomandi and Margot (2009).
• Vehicle Routing Problem with Stochastic Customers (VRPSC): Each cus-
tomer request i ∈ R has an occurrence probability pi and therefore requires no
54 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

service with a probability of 1 − pi . If more customers than originally expected

demand service, infeasible tour plans can arise due to capacity constraints of ve-
hicles or time window violations. This variant has been first studied by Jézéquel
(1985) and Jaillet (1987). Note that the set of potential customer requests is
known in advance. This is in contrast to Dynamic Vehicle Routing Problems con-
sidered in Sect. 2.5.3.2.2 in which new customer requests can theoretically occur
anywhere in the considered service region as well as at unknown time points.
• Vehicle Routing Problem with Stochastic Travel Times (VRPSTT): In this
SVRP variant, travel times between the relevant locations are given by random
variables which can lead to unfavorable solutions. For example, in Lambert et al.
(1993) a VRPSTT is considered in which a late service of requests results in
penalty costs.
• Vehicle Routing Problem with Stochastic Service Times (VRPSST): In this
variant of the SVRP, the time required for servicing customer requests is uncer-
tain. If a vehicle is required to spend more time at a customer request location
than planned, postponed service of consecutive requests can occur which can
lead to violations of time window constraints and penalty costs. The VRPSST is
considered among others in Laporte et al. (1992), Hadjiconstantinou and Roberts
(2002), and Verweij et al. (2003).
Moreover, combinations of the described types of uncertainty in the relevant in-
formation are possible. A frequently studied variant is the Vehicle Routing Problem
with Stochastic Customers and Demands (VRPSCD) which is also known as the
Probabilistic Vehicle Routing Problem (PVRP, cf. Larsen 2000, p. 28). It was first
studied by Jézéquel (1985) and Jaillet (1987). Another SVRP variant which con-
siders two types of uncertainty denoted as the Stochastic Vehicle Routing Problem
with Stochastic Travel Times and Stochastic Service Times is analyzed in the work
of Laporte et al. (1992) and Kenyon and Morton (2003).
As mentioned above, the VRPSD is the most studied SVRP variant in the liter-
ature. Hence, various policy-based recourse strategies have been developed for the
VRPSD. In what follows, some of them are described. One of the first recourse
strategies proposed for the VRPSD lets a vehicle return to the depot whenever the
request assigned next in a vehicle’s tour cannot be serviced since the currently re-
maining vehicle capacity is too low. Laporte and Louveaux (1990) and Dror et al.
(1993) describe other recourse strategies in which preventive breaks are performed
at strategic positions in the tour. For example, when a vehicle is near the depot, it
delivers the currently loaded goods there. Another policy-based recourse strategy
is presented by Novoa et al. (2006). In this strategy, more flexibility is integrated
by allowing vehicles to postpone the next assigned request and to service other re-
quests first before returning to the depot if it is likely that those other requests can
be serviced with the currently remaining capacity. Further recourse strategies for
the VRPSD are discussed in Dror (1989), Gendreau et al. (1996), and Cordeau et al.
(2007b).
Besides policy-based recourse actions, more complex recourse strategies are
based on a reoptimization, i.e., adaptation, of the remaining tour of each vehicle
2.5 Information Revelation in Routing Problems 55

when a tour failure occurs. As an advantage, this can improve the system perfor-
mance. However, such recourse strategies complicate the estimation of costs for
expected recourse actions during the initial tour plan generation (cf. Gendreau et al.
1996). In the work of Dror (1989), the author introduces a formulation based on
a Markov Decision Process (MDP, cf. Puterman 2005) for solving a single-vehicle
VRPSD. In this approach, the solution is reoptimized whenever the vehicle reaches
the next request where its demand is revealed. At each customer request, the deci-
sion is made whether to service the request now or to service it later. In the work
of Dror (2005), the author states that reoptimization-based approaches represent
the most promising approaches for solving VRPSDs. According to Secomandi and
Margot (2009), recourse actions should be divided into reactive actions which are
only taken when a tour failure occurs and pro-active actions in which decisions
are taken by anticipating the likelihood of a tour failure. Furthermore, they state
that the opportunities given by technological advances are especially beneficial to
reoptimization-based approaches.

2.5.3.2.2 Centralized Coordination: Dynamic Vehicle Routing Problems

The application of a centralized coordination of the transportation process during

its execution enables the utilization of real-time control approaches. Real-time con-
trol approaches are able to monitor the execution of the transportation process and
perform tour plan adaptations simultaneous to its execution (cf. Bock 2010). Fig-
ure 2.9 illustrates the coordination of the transportation process in dynamic vehicle
routing problems. As can be seen, the tour plan currently in execution is required to
be adapted during its execution. Such adaptation activities belong to the real-time
control layer of the transportation process (see Sect. 1.2). The utilization of real-
time control approaches allows the handling of all types of dynamic events which
are described in Sect. 2.6.2.
An important factor is the decision when an adaptation of the tour plan shall
be performed. Specifically, it is necessary to define how the complete dynamic
problem is divided into subproblems which are solved during the planning hori-
zon (cf. Bianchi 2000). For example, the tour plan can be adapted whenever a
new dynamic event arrives in the system. This is denoted as single event optimiza-
tion (SEO, cf. Bianchi 2000). Another possibility is to adapt the transportation pro-
cess in specific time intervals (cf. Bock 2010). Further characteristics of real-time
control approaches will be discussed in Sects. 4.7.4 and 5.1. More information on
the implemented real-time control approach will be given in Sect. 5.2.
In the literature, different variants of dynamic vehicle routing problems cen-
trally coordinated by real-time control approaches can be found. They can be dis-
tinguished according to the type of available stochastic knowledge, i.e., the quality
of information about future events (see Sect. 2.5.2.1):
1. Stochastic knowledge is unavailable: If no stochastic knowledge is available,
real-time control approaches can only react after new dynamic events have ar-
rived in the system during the execution of the transportation process. In the
56 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Operative layer Real-time control layer

Initial tour plan Tour plan adaptation during tour plan execution according to
generation before occurring dynamic events utilizing a real-time control approach
tour plan execution
Relevant information Relevant information is changed by dynamic events during the execution of
does not change the transportation process
time
Beginning of the transportation process

Fig. 2.9 The transportation process in dynamic vehicle routing problems with a centralized coor-
dination

literature, such routing problems are classified as variants of the Dynamic Vehi-
cle Routing Problem (DVRP, cf. Psaraftis 1988).
2. Stochastic knowledge is probabilistic: In this case, stochastic knowledge is
available and follows prescribed probability distributions which usually describe
all uncertain elements of the relevant information. Since occurring dynamic
events change the relevant information according to the known probability dis-
tributions, the utilization of this type of stochastic knowledge is likely to pro-
vide significant improvements on the solution quality. In the literature, these ap-
proaches are often denoted as variants of the Dynamic Stochastic Vehicle Rout-
ing Problem (DSVRP, cf. Hvattum et al. 2006 and Ichoua et al. 2006).
3. Stochastic knowledge is forecasted: In many practical applications, stochas-
tic knowledge is not directly available. However, if historical data is available,
forecasted stochastic knowledge can be generated by analyzing this data. For
example, by applying sophisticated methods on request arrivals of elapsed days
including arrival time and location, stochastic knowledge can be generated out
of this historical data. Such historical data is denoted as past request informa-
tion in what follows. Since the stochastic knowledge is forecasted, in this book,
we denote such problem scenarios as variants of a Dynamic Forecasted Stochas-
tic Vehicle Routing Problem (DFSVRP) to which also the considered RDOPG
applications belong.
Note that in routing problems which forecast stochastic knowledge it is as-
sumed that common characteristics observed in the past provide good predictions
of future events. Hence, there exist situations in which future dynamic events
may possess significantly different attributes as predicted. Since the stochastic
knowledge is known with uncertainty only, in the worst case, the utilization of
forecasted stochastic knowledge can lead to an inferior system performance com-
pared to applying a deterministic real-time control approach which only consid-
ers information known with certainty. Moreover, it is advantageous if the request
data of the considered routing application possesses specific characteristics. We
denote them as structural quality of a request data set in what follows. Hence, in
practice, it is advantageous to analyze existing real-world request data according
to their suitability with regard to generating stochastic knowledge which can be
advantageously applied. We will discuss this further in the computational exper-
iments in Chap. 8.
2.5 Information Revelation in Routing Problems 57

With regard to Powell et al. (1995), in the first case, a deterministic dynamic
model is utilized. Specifically, it is dynamic because it is updated during the exe-
cution of the transportation process but since it only considers changes which have
already occurred, it is deterministic. In the latter two cases, the utilized model is
a stochastic dynamic model since additional stochastic knowledge about expected
future events is utilized.
As mentioned in Sect. 2.5.3.2.1, tour plan adaptations required by unexpected
changes in the relevant information may lead to significant changes in the currently
executed tour plan. Although the discussed real-time control approaches are able
to perform a centralized adaptation of the tour plan by elaborated solution meth-
ods, dynamic events can nonetheless cause high additional costs. With regard to
RDOPG applications in which new customer requests arrive during the execution of
the transportation process, the arrival of many unexpected customer requests may
require the dispatching of additional vehicles. If no additional vehicles are avail-
able (see aspect 11 of Psaraftis’ approach in Sect. 2.5.2.2), vehicle tours are likely
to require significant adaptations. This negatively influences the execution of the
transportation process but also results in higher travel distances and longer request
response times which often leads to a lower solution quality. Therefore, the aim is to
coordinate the transportation process by real-time control approaches in a way that
provides an increased level of tour plan flexibility. In doing so, tour plans provide
more possibilities for adapting the transportation process according to changes in
the relevant information that have occurred without causing significant distortions.
This increased flexibility can also be regarded as a technique for increasing the ro-
bustness of tour plans towards uncertain dynamic events.
Depending on the availability of stochastic knowledge and the pursued objec-
tive, two types of approaches are useful in providing an increased level of tour plan
flexibility:
• Flexibility-increasing approaches without using stochastic knowledge: Tour
plan flexibility can be increased by least-commitment strategies such as waiting
strategies. By postponing decisions to a later point in time, it is likely that better
decisions with regard to travel distance can be made since more relevant infor-
mation is available then. In what follows, such real-time control approaches are
denoted as deterministic real-time control approaches.
• Flexibility-increasing approaches using stochastic knowledge: Stochastic
knowledge offers additional possibilities for realizing the aforementioned waiting
strategies. Moreover, waiting strategies can be extended to relocation strategies
by actively repositioning idle vehicles to subregions in which future requests are
likely to arrive soon. Furthermore, request assignment strategies can be used for
efficiently coordinating utilized vehicles. Specifically, by estimating future vehi-
cle positions resulting from specific request assignments, expected impacts, e.g.,
request response times, can be calculated for expected future requests. This in-
formation can be directly utilized for evaluating the solution quality of possible
adaptations during the transportation process. Such real-time control approaches
are denoted as pro-active real-time control approaches in this book.
58 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

In theory, stochastic knowledge is available for all types of dynamic events that
have been presented in Sect. 2.6.2. However, in real-world applications, some dy-
namic events can be predicted better than others. For example, since in many
real-world applications recurring arrival patterns of requests can be observed, the
arrival of new requests has been very often considered in approaches in the liter-
ature. Besides, dynamism arising from changes in travel times is nowadays also
well predictable due to floating car data (FCD, cf. Ehmke et al. 2009, 2010). Fur-
thermore, traffic jams are automatically detectable by traffic sensors which are
commonly installed at highways and major roads nowadays (cf. Ghiani et al. 2003
and Attanasio et al. 2007). In contrast to both these sources of dynamism, predict-
ing the location and time of vehicle breakdowns is rather impossible. Hence, in
the literature, most research deals with dynamic events arising from new request
arrivals or changing travel times.

Real-time control approaches which do not provide an increased tour plan flex-
ibility only react towards newly arriving dynamic events. Hence, we denote such
approaches as reactive real-time control approaches. In the literature review, we
will mainly focus on approaches which provide an increased flexibility in the trans-
portation process. Specifically, details on approaches utilizing waiting strategies
will be discussed in Sect. 4.3.1. Moreover, relocation strategies will be presented
in Sect. 4.3.2 and request assignment strategies will be described in Sect. 4.3.3.

2.6 Dynamic Routing Problems Which Utilize a Centralized

Coordination

In the previous section, different variants of routing problems and their character-
istics were presented. In accordance with the considered RDOPG applications, this
section deals with specific dynamic routing problems in detail, in particular those
with a centralized coordination of the transportation process utilizing real-time con-
trol approaches (see Sect. 2.5.3.2.2). First, typical objectives in dynamic routing
problems are described. Afterwards, dynamic events in dynamic routing problems
by which the relevant information is changed are presented. In addition, technolo-
gies utilized in real-time control approaches are introduced. In order to classify
dynamic routing problems according to their difficulty level with regard to finding
good solutions, a measure denoted in the literature as the degree of dynamism is de-
scribed. Based on this measure, practical dynamic routing problems are introduced
using a three-echelon classification of dynamic routing problems. Finally, two mea-
sures for evaluating the performance of solution approaches for dynamic routing
problems are presented.
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 59

2.6.1 Typical Objectives in Dynamic Routing Problems

In general, three different objectives (or combinations of these objectives) are pur-
sued in dynamic routing problems (cf. Larsen et al. 2007):
• Minimization of distribution costs: The minimization of costs which are gen-
erated during the execution of the transportation process (see Sects. 2.4.12.1
and 2.4.12.2) is the most widely used objective in static routing problems. Fur-
thermore, this objective plays an important role in specific dynamic routing prob-
lems, in particular those in which the transportation process is less time-critical
and only a small part of the relevant information changes during the execution of
the transportation process.
• Maximization of service quality: If a high quality of customer service is pur-
sued, dynamically arriving customer requests are required to be quickly serviced.
As described in Chap. 1, high request response times lead to customer incon-
venience and may impose significantly high costs on the service company, for
example due to penalty costs incurred by SLA contracts. This type of objective is
considered in the RDOPG applications which are investigated in this book.
• Maximization of system throughput: This objective is applied in dynamic rout-
ing problems in which the service of as many customers as possible within a given
planning horizon is pursued. In such applications, it may happen that the service
of an isolated customer request is postponed for a long time in favor of servicing a
larger amount of other requests which are closely located to each other. Hence, as
mentioned by Psaraftis in aspect 8 of his approach in Sect. 2.5.2.2, it is reasonable
to integrate methods which ensure that no customer is postponed indefinitely.
Note that the presented objectives cannot be fulfilled to their full extent at the
same time. For example, tour plans which are generated according to the objective
of maximizing quality of customer service usually produce additional travel activ-
ities during the day which makes them suboptimal from a distribution costs per-
spective. This has also been addressed by Richter (2005, p. 200 et seq. and p. 224).
In Sect. 2.6.5, we present various real-world routing problems and their pursued
objective.

2.6.2 Dynamic Events in Dynamic Routing Problems

In dynamic routing problems, dynamic events change the relevant information dur-
ing the execution of the transportation process. In general, the following four types
of dynamic events are considered in the literature:
• New requests: Each newly arriving request has a location in the service area. De-
pending on the considered routing problem, a new request comprises additional
attributes such as demands, time windows, lateness penalties, and service times.
With the arrival of a new request, an adaptation of the tour plan in order to in-
tegrate this request is required. In some applications, it may be allowed to reject
60 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

new requests, often at additional penalty costs. One approach in which this ap-
plies is the approach of Yang et al. (2004). Such a rejection may be beneficial if,
at request arrival, it can already be determined that a request cannot be feasibly
serviced (e.g., due to violations of hard time window constraints) or if the costs of
servicing the request are higher than the costs of its rejection. However, in appli-
cations which aim at maximizing quality of service, for example measured by the
number of serviced customers or the minimization of request response times, all
arriving requests usually need to be serviced. This is considered among others in
Bent and van Hentenryck (2004b) and Kleywegt et al. (2009). According to Pil-
lac et al. (2010), the arrival of new requests is the most studied type of dynamic
events in the literature.
• Changes in request attributes: In some dynamic routing problems, attributes of
known requests change or, similar to SVRPs, are only revealed when the vehicle
reaches the customer location. For example, if the demand of a request turns
out to be different from what was expected, the capacity of the assigned vehicle
may not be sufficient. In this case, the request needs to be reassigned to another
vehicle. This type of dynamism is conducted in Wu et al. (2004) and Hvattum
et al. (2007).
• Traffic congestions and road blockages: This type of dynamic events causes
either increased travel times or a complete blockage of affected roads. The length
of such impacts may be known or unknown. Since the service of requests can be
delayed by such dynamic events, re-routing of vehicles to non-congested roads or
the reassignment of requests to other vehicles may be advantageous. This type of
dynamic event is considered in Fleischmann et al. (2004b), Ferrucci (2006, p. 44),
Attanasio et al. (2007), Barceló et al. (2007), Haghani and Yang (2007), Cortés
et al. (2008), and Bock (2010).
• Vehicle disturbances: This results in a decreased efficiency or a complete failure
of a vehicle due to a breakdown or a driver drop-out. As a consequence, pending
requests may be required to be reassigned to other vehicles. Moreover, if the
affected vehicle is transporting requests when a complete failure occurs, its loaded
requests need to be picked up by other vehicles at the affected vehicle’s current
location. This type of dynamic events is studied in Ferrucci (2006, pp. 44–45), Li
et al. (2009a, 2009b), Bock (2010), and Mu et al. (2010).

2.6.3 Technologies Utilized in Real-Time Control Approaches

Approximately 20 years ago, technical limitations often prohibited the application

of real-time control approaches in transportation processes. In addition to new de-
velopments in computational technology, advances in communication and position-
ing technology which have been achieved in the last two decades also support the
practicability of real-time control approaches.
In this section, we describe technologies which are used in on-board units of
vehicles and those utilized in computer systems in dispatching centers in order to
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 61

enable the coordination of the transportation process in real-time. After describing

these technologies separately, we present an exemplary implementation of a system
establishing a bidirectional communication between the vehicles and the dispatch-
ing center. Moreover, we present an exemplary architecture of a real-time control
approach which makes use of the previously described bidirectional communica-
tion system. In doing so, it is illustrated how the presented technologies can be
jointly used in order to efficiently realize real-time control approaches in real-world
dynamic routing problems.

2.6.3.1 Recent Advances in On-Board Technologies

In order to reasonably adapt the tour plan during its execution, the current vehi-
cle positions are required to be provided to the dispatching center. In particular, if
positions of traveling vehicles are available in real-time at the dispatching center, a
newly arriving request which is close to a vehicle can be directly serviced by this ve-
hicle. In doing so, the required detour and the request response time can be reduced,
compared to returning to this request later. For this purpose, on-board systems are
utilized.
On-board systems allow the transmission of different types of vehicle data, such
as its current position, to the dispatching center in real-time during the execution of
the transportation process. The automatic determination of the current vehicles’ po-
sitions and velocities became possible with the introduction of the Global Position-
ing System (GPS) in 1995. It consists of 24 satellites which are arranged in 6 orbital
planes with 4 satellites each (cf. Kaplan and Hegarty 2006, p. 3). These satellites
broadcast the current time as well as their current position which can be received
and further used by GPS-capable devices. By utilizing the transmitted data of in-
dividual GPS satellites, GPS devices can determine their current position with an
accuracy of a few meters using the GPS operation mode called Standard Position-
ing Service (SPS) which provides sufficient precision for routing applications. SPS
is free of charge for commercial and non-commercial use and is utilized by millions
of GPS devices worldwide (cf. Kaplan and Hegarty 2006, pp. 3–4). For more infor-
mation on GPS and other satellite-assisted positioning systems such as GALILEO,
see Kaplan and Hegarty (2006).
Besides GPS-capable on-board vehicle units, a data transmission technology is
required for establishing an electronic communication between the vehicles and the
dispatching center. In the last ten years, significant technological advances have
emerged in this area. According to Larsen (2000, p. 12 et seq.), in the year 2000,
dedicated radio based communication systems were favorable compared to mo-
bile phone communication systems which had low setup costs but high operational
costs. Over the years, alternative tariff models have made electronic communi-
cation via smartphones much more economical. In the year 2011, mobile phone
contracts providing Internet access at a flat rate are available for less than 30 eu-
ros per month in Germany (cf. Telefónica Germany GmbH & Co. OHG 2011).
Moreover, data transfer rates in mobile phone networks has evolved from GSM
62 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

(Global System for Mobile Communications, before Groupe Spécial Mobile) to

GRPS/EDGE (General Packet Radio Service and Enhanced Data Rates for GSM
Evolution) to UMTS (Universal Mobile Telecommunications System). UMTS pro-
vides data transfer rates with HSDPA (High Speed Downlink Packet Access) and
HSUPA (High Speed Uplink Packet Access). These are comparable to fixed line
Internet connections (cf. Halonen 2003, p. xxvi). The use of mobile Internet-based
data transmission technologies in combination with low priced but powerful GPS-
enabled smartphones and other mobile devices allows for the continuous automated
transmission of current vehicles’ positions to the central dispatching center in real-
time at reasonable costs.
Moreover, vehicle on-board units can be used to provide information about cur-
rent activities to the dispatching center during the execution of the transportation
process. For example, when a service worker is at a customer request, information
about the progress of the service activity can be efficiently transmitted. If unex-
pected difficulties occur during a service activity, the dispatching center can identify
delays in scheduled tour plan activities early and can perform appropriate adapta-
tions in time. This is especially advantageous in applications pursuing customer
service quality aspects in which service times required at customer locations are
uncertain. Furthermore, on-board units can provide a request management system
which displays changes in the vehicle’s tour to the driver and allows the driver to
communicate with the dispatching center, e.g., confirm changes in the tour. More-
over, request management systems can be combined with integrated navigation sys-
tems which automatically update driving route instructions in case of a tour plan
adaptation or when traffic conditions change. Note that on-board units can exist in
different variants, for example installed units, GPS-capable smartphones, or tablet
computers with a mobile Internet connection.

2.6.3.2 Recent Advances in Dispatching Center Technologies

Besides the previously mentioned on-board technologies which allow for the track-
ing and monitoring of vehicle activities in real-time, there has been an advancement
in the technologies of dispatching centers in recent years. Specifically, the coordina-
tion of vehicles in real-time has become possible due to the development of sophis-
ticated real-time control approaches which make use of elaborated solution methods
providing a high solution quality within a short amount of computational time (see
Sect. 4.7.4 and Chap. 5). Hence, tour plan adaptations can be carried out quickly
after dynamic events have occurred in the system. This allows real-time control ap-
proaches to coordinate the transportation process under real-world restrictions.
In order to efficiently handle occurring dynamic events in a real-time control
approach, an electronic transmission of unexpected changes in the relevant infor-
mation (see Sect. 2.5) is advisable. For dynamically arriving customer requests, this
can be realized by electronically recording new customer requests in a call cen-
ter (see also Fig. 1.1). Besides day- and time-dependent travel time profiles, dif-
ferent information sources like traffic flow sensors automatically provide updated
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 63

Real-time control

Adaptation of the executed tour plan according to the current situation in the
transportation process

Adapted tour plan Current situation

in the transportation process

Current
Dynamic events Database data
Vehicle fleet
Dynamic
E.g., incoming requests, events Representation of the Current position and
traffic congestions, Modified
current situation in the tours
current state
road blockages, or
transportation process of each vehicle
vehicle breakdowns

Fig. 2.10 Example architecture of a real-time control approach (cf. Bock 2010)

travel times on roads and highways. In order to quickly calculate travel distances
and travel times between considered relevant locations and vehicles, the utilization
of efficient shortest path approaches in combination with a Geographic Information
System (GIS) which provides road network data is advantageous. Apart from com-
mercially available road networks, OpenStreetMap is a comprehensive community-
based road network data project to which users can make contributions (cf. Haklay
and Weber 2008, OpenStreetMap 2011). For more information about existing ap-
proaches in the literature which make use of GIS systems, see also Sect. 4.7.5 in the
literature review.

2.6.3.3 Interplay of Utilized Technologies and Information Flow in Real-Time

Control Approaches

The joint use of the described technologies allows for the efficient application of
real-time control approaches. In Fig. 2.10, an exemplary architecture for a real-time
control approach is presented (cf. Bock 2010). In this architecture, the current sys-
tem situation is stored in a central database which is supplied with new informa-
tion by two systems. The first system transmits changes in the relevant information
arising from dynamic events (see Sect. 2.6.2) while the second one establishes the
bidirectional communication between the dispatching center and the vehicles. It pro-
vides current information about the vehicles to the database, e.g., their positions and
the current activities of the service workers. Moreover, by using this system, vehi-
cles are informed when changes in their tour have been realized by the real-time
control. The real-time control adapts the current tour plan by transforming the cur-
rent system situation from the database into the considered problem data. After the
tour plan adaptation has been completed, the generated tour plan is stored in the
database and changes in the tours are transmitted to the vehicles.
In order to realize the bidirectional communication system, Larsen (2000, p. 14)
describes a communication architecture based on an Internet data connection. As
illustrated in Fig. 2.11, in this architecture, vehicles are equipped with the previously
64 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

GPS satellites

Position
information

Tour plan updates

Determines position, Vehicle information,

GSM, Central dispatching center
receives tour changes e.g., position and status
UMTS using a real-time control approach
for the centralized coordination
of the transportation process

Fig. 2.11 Example architecture for a bidirectional communication system between the vehicles
and the central dispatching center (cf. Larsen 2000, p. 14)

described GPS-enabled on-board units. Using this data connection, position and
status are sent from the vehicles to the central dispatching center and changes in the
tour plan are transmitted to the vehicles.
To sum up, it can be stated that using a real-time control approach by utilizing the
described systems can significantly support dispatchers by providing decision sup-
port for tour plan adaptations. Specifically, by assessing various tour plan adaptation
possibilities in a short amount of time and by enabling vehicle en-route diversion
activities, the solution quality can be significantly increased (see also Sect. 8.5). De-
tailed information about real-time control approaches will be given in the literature
review in Sect. 4.7.4 and in Chap. 5. Moreover, further research work on the utiliza-
tion of communication technologies in dynamic routing problems will be described
in the literature review in Sect. 4.7.5.

2.6.4 The Degree of Dynamism

Each dynamic routing problem comprises a specific difficulty in coordinating the

corresponding transportation process with regard to the considered objective. We
denote this as the difficulty level in what follows. This difficulty level depends on
many factors. For example, the type of dynamic events considered in a dynamic
routing problem and the considered objective are crucial attributes. Moreover, var-
ious instances of the same dynamic routing problem can also have a significantly
different difficulty level. Since a higher difficulty level makes it more complex to
attain a highly qualitative coordination of the transportation process, this imposes
higher requirements on the solution method utilized for adapting the tour plan. Note
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 65

Fig. 2.12 Different arrival Instance A

times of dynamic requests in
two instances of a dynamic
routing problem with the time
0 r1a r2a r3a r4a r5a r6a T
same degree of dynamism
(cf. Larsen 2000, p. 57) Instance B

time
0 r1a r2a r3a r4a r5a r6a T

that awareness of this difficulty level is beneficial. For example, this information
can be used for assessing the performance of a solution method when applied to
different problem instances comprising individual difficulty levels.
In the literature, an approach for determining such differences in difficulty levels
arising in dynamic routing problems is the degree of dynamism (dod). This measure
considers dynamic routing problems in which dynamism arises by newly arriving
requests which is, as mentioned in Sect. 2.6.2, the most studied type of dynamic
events in the literature. The degree of dynamism exists in three versions. Its basic
version for dynamic routing problems without time windows was introduced by
Lund et al. (1996). For defining the dod, it is assumed that the execution of the
transportation process starts at time point 0 and new requests can arrive up to time
point T . Furthermore, nstat denotes the number of requests which are known during
the initial tour plan generation process while ndyn defines the number of dynamically
arriving requests. With nall = nstat + ndyn , the dod is defined as:
ndyn
dod =
nall
The dod ranges from 0 to 1 where instances with a higher difficulty level have a
higher dod. Specifically, if all requests are known in the initial tour plan generation
process, it holds that dod = 0 indicating a static routing problem. In contrast, if
all requests just arrive during the execution of the transportation process, it holds
that dod = 1 so that a completely dynamic routing problem is present. This points
to the fact that an instance of a dynamic routing problem in which only a small
fraction of all requests dynamically arrives can be considered to be less challenging
and easier to solve than an instance of the same dynamic routing problem in which
all customer requests dynamically arrive. In what follows, we use the term problems
which are easier to solve which refers to problems that make it easier to attain a
solution with a good solution quality in reasonable time. Clearly, the opposite is
true for problems which are more difficult to solve.
Despite its simplicity, other characteristics which also have a significant impact
on the difficulty level are not considered in the dod. For example, it does not con-
sider the time points at which dynamic requests arrive in the system. As illustrated in
Fig. 2.12, two instances with significantly different request arrival properties can in-
hibit the same dod. Specifically, if the minimization of travel distance is the pursued
objective, instance A is easier to solve than instance B since all dynamic requests
66 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

arrive early in instance A (cf. Larsen 2000, p. 57). Hence, after the arrival of the
last request, the remaining routing problem is a static problem. Since at this point in
time it is likely that many other requests are still pending in instance A, a tour plan
with a high solution quality can be generated according to this large amount of still
changeable decisions. In contrast, in instance B, additional changes in the relevant
information are also revealed late. Hence, at this point in time it is likely that the
majority of requests have already been serviced.
In order to consider the mentioned differences in request arrival times, Larsen
(2000, p. 58) extends the degree of dynamism to the effective degree of dynamism
denoted as edod. In the edod, each dynamic request i ∈ {1, . . . , ndyn } arrives at time
point ria in the system. The edod is then defined as follows:
ndyn ria
i=1 ( T )
edod =
nall

The effective degree of dynamism represents the ratio of the arrival times of dynamic
requests to the latest possible time point T at which they could arrive. In accordance
with the previously mentioned objective of minimizing travel distances, if dynamic
requests arrive early, the edod has a low value while later request arrivals result in
higher edod values. Similar to the dod, the edod is a measure between 0 and 1 in
which larger values indicate problem instances which are more difficult to solve.
If time windows are considered in dynamic routing problems, it is reasonable to
consider them in the calculation of the degree of dynamism since tight hard time
windows limit dispatching possibilities making it difficult to attain good solutions.
In case of soft time windows, this leads to less possibilities for servicing requests
without penalty costs compared to problem instances with relaxed request time win-
dow constraints. If the objective of distance minimization is considered, requests can
be fulfilled in more distance-optimized ways. Therefore, such problem instances can
be considered to be easier to solve than those with tight time windows. In order to
consider such time window characteristics of dynamic routing problems, Larsen
(2000, p. 59) proposes a third version of the degree of dynamism denoted as the
effective degree of dynamism with time windows (edodtw ). It is defined as
ndyn
ndyn

1 T − (ril − ria ) 1 r re
edodtw = = 1− i
nall T nall T
i=1 i=1

where rire = ril −ria denotes the available reaction time of request i and represents the
time period from the arrival time of request i to the end of its time window. Hence,
the existence of larger reaction times provides more possibilities for dispatching and
servicing pending requests which leads to a low edodtw and indicates relaxed sys-
tem situations. In Fig. 2.13, the reaction times of two instances of dynamic routing
problems, each comprising two dynamic requests, are illustrated. As can be seen
in this figure, the reaction times are larger in instance A than in instance B making
instance A easier to solve.
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 67

Fig. 2.13 The effective Instance A r2re

degree of dynamism with r1re
time windows (cf. Larsen
2000, p. 60)
time
0 r1a r2a r1e r1l r2e r2l T

Instance B r1re r2re

time
0 r1a r1e r1l r2a r2e r2l T

Note that in dynamic routing problems in which the minimization of customer in-
convenience is pursued, it holds that ril = rie = ria ∀i ∈ {1, . . . , ndyn } since customer
inconvenience represented by penalty costs directly starts upon request arrival in the
system. In such applications, the reaction time rire is 0 for all dynamically arriving
requests. Since almost all customer requests are dynamic in RDOPG applications
considered in this book, these applications have an edodtw close to 1 making them
difficult to solve with a high solution quality.

Further Extensions to the Degree of Dynamism The described dod and its vari-
ants provide efficient measures for determining the difficulty level of routing prob-
lems in which the minimization of travel distance or comparable objective functions
is pursued. With regard to RDOPG applications in which the minimization of cus-
tomer inconvenience is pursued, further extensions are possible out of which two are
described in what follows. For example, considering Fig. 2.12, a different difficulty
level for the two instances illustrated arises. Under this objective, instance A can be
considered to be more difficult to solve than instance B since in instance A, the dis-
patcher has to deal with a large amount of newly arriving requests in a short amount
of time. This can be complex, especially if customer inconvenience depends on re-
quest response times, good and fast tour plan adaptations are required. Hence, with
regard to RDOPG applications, it is advantageous to additionally evaluate other
request arrival characteristics. Specifically, the distribution of request inter-arrival
times is another important aspect. As illustrated in Fig. 2.14, two instances with
the same edod can have a significantly different request arrival distribution. If the
minimization of customer inconvenience is pursued, instance A is more difficult to
solve than instance B since, as mentioned above, the dispatcher is required to make
good decisions for a large amount of requests at the same time. Moreover, in dy-
namic VRPs, the fleet size is often restricted (see aspect 11 of Psaraftis’ approach
presented in Sect. 2.5.2.2). Hence, a large number of requests shortly arriving one
after the other causes longer request response times and hence higher customer in-
convenience values compared to evenly distributed request arrivals.
Another possible extension which is relevant for RDOPG applications is men-
tioned by Pillac et al. (2010). The authors state that the geographical distribution of
newly arriving requests and resulting travel times between requests should be con-
sidered since they significantly determine the complexity of dynamic routing prob-
lems in which the minimization of the request response times is pursued. Related to
68 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Fig. 2.14 Two instances with Instance A

the same effective degree of
dynamism of 0.4 and
significantly different 0 T
time
inter-arrival times 96 97 98 99
1234

Instance B

time
0 10 20 30 40 60 70 80 90 T

this, we will propose a new measure for characterizing differences in dynamic rout-
ing problems with regard to the applicability of stochastic knowledge in pro-active
real-time control approaches in Sect. 8.9.

2.6.5 Three-Echelon Classification of Dynamic Routing Problems

In Larsen (2000, p. 60 et seq.) and Larsen et al. (2007) a two-dimensional classifica-

tion for distinguishing real-world dynamic routing problems is proposed according
to their degree of dynamism and the pursued objective. In this classification, dy-
namic routing problems are, according to the percentage of dynamically arriving re-
quests, subdivided into weakly, moderately, and strongly dynamic systems. Hence,
these echelons represent routing problems with a low, medium, and high degree of
dynamism. Moreover, the pursued objective ranges from the minimization of dis-
tribution costs to the maximization of service quality. As mentioned in Sect. 2.6.1,
these two objectives usually contradict each other so that both objectives cannot
be completely fulfilled at the same time. Nevertheless, there exist dynamic routing
problems in which both objectives are considered with a similar priority.
The proposed classification is illustrated in Fig. 2.15. As can be seen, there is
a strong correlation between the degree of dynamism and the pursued objective in
real-world dynamic routing problems, i.e., the higher the degree of dynamism, the
more important service quality aspects become in the considered routing problem. In
what follows, we describe characteristics of each echelon and present corresponding
real-world applications (cf. Larsen 2000 and Ghiani et al. 2003).

2.6.5.1 Echelon I: Weakly Dynamic Systems

Weakly dynamic systems are characterized by only a small number of customers

dynamically arriving during the execution of the transportation process. According
to the authors, the majority, i.e., at least 80 % of all customer requests are already
known during the initial tour plan generation process (cf. Larsen et al. 2007). Hence,
routing problems in this echelon have a dod < 0.2. In what follows, we present
typical real-world routing problems which belong to this echelon:
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 69

Maximize service quality,

minimize customer inconvenience

Police,
Police, fire
fire fighters,
fighters,
ambulances
ambulances

SLA
SLA repairman
repairman Taxi
Taxi
companies,
companies, cabs
cabs
subsequent
subsequent delivery
delivery
of
of newspapers
newspapers

Dial-a-ride
Dial-a-ride
(Telebus
(Telebus systems)
systems)

Repairman
Repairman
companies
companies

Long-distance
Long-distance
courier
courier services
services

Echelon I: Echelon II: Echelon III:

Weakly Moderately Strongly
dynamic dynamic dynamic
systems systems systems
Dial-a-ride
Dial-a-ride
(Handicapped
(Handicapped
transports)
transports)

Dynamic
Dynamic fleet
fleet
management
management

Vendor-managed
Vendor-managed
distribution
distribution systems
systems

Subscription-based
Subscription-based
distribution
distribution systems
systems

Minimize distribution costs

Fig. 2.15 Three-echelon classification of dynamic routing problems (cf. Larsen 2000, p. 63,
Larsen et al. 2007)

• Subscription-based distribution systems: In general, customers subscribe to

this type of service so that customer requests and their demands are often known
in advance. The authors mention that less than 5 % of all customers request an
additional service or make changes in their pre-ordered demands. In such sys-
tems, the tour plan is usually only adapted in longer time intervals, e.g., on a
monthly basis, in order to save coordination effort and increase efficiency by uti-
lizing driver learning curves. In order to service the few spontaneous customer
requests, some slack time is included in the tour plan. One real-world example
of subscription-based distribution systems is the delivery of groceries to private
households on a regular basis.
• Vendor-managed distribution systems: In such distribution systems, compa-
nies manage the inventory of their customers and automatically replenish their
inventory when necessary. In order to determine which customers are required to
70 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

be serviced during the next planning horizon which is usually the next day, the
distribution company estimates customers’ inventory levels using specific meth-
ods. Hence, the set of pending requests can be well estimated during the initial
tour plan generation process. However, since the actual consumption rates of cus-
tomers may vary from what has been predicted, a small percentage of customers
additionally require service. Moreover, new customers may demand urgent ser-
vice. Besides newly arriving requests, another source of dynamism is given by
the uncertain demand of customers. Hence, capacity limitations of vehicles may
require additional but usually few tour plan adaptations during the day. A real-
world application belonging to such distribution systems is the distribution of
heating oil to private households (cf. Larsen 2000, pp. 16–17). In this application,
the oil consumption is estimated by the degree-days method which estimates con-
sumption rates based on measured outdoor temperatures. The author states that
about 20 % of all customers are dynamic in this real-world problem.
• Dynamic fleet management: In many international freight forwarding compa-
nies, cost aspects often have a higher priority than urgency of requests. Due to
the long-haul character of these transportation processes, customers are usually
required to negotiate transportation requests some time in advance with the trans-
portation company. Nevertheless, tracking vehicles in real-time and allowing the
re-routing of traveling vehicles, which still have available loading capacities, can
help to integrate dynamically arriving requests. In doing so, significant cost sav-
ings can be achieved on long-distance transportation activities.
• Dial-a-ride for handicapped people: Specialized companies offer dial-a-ride
services for transporting handicapped people between their home and their de-
sired destination, usually for medical appointments. The arising transportation
processes can be modeled as dynamic variants of the Dial-A-Ride Problem
(DARP, see Sect. 2.3.3.3). Since customers usually book transportation services
in advance, only few customers request a service on the same day. Hence, it is
possible to generate cost-efficient tour plans which also provide a high customer
service quality by meeting the customers’ desired appointment times and mini-
mizing their ride times. For more information on related routing problems, see,
e.g., Cordeau and Laporte (2003, 2007) and Cordeau et al. (2007a).

2.6.5.2 Echelon II: Moderately Dynamic Systems

This echelon consists of real-world applications in which the percentage of dynami-

cally arriving customer requests during the execution of the transportation process is
noticeable. Nevertheless, there is a considerable proportion of requests still known
before the execution of the transportation process (cf. Larsen et al. 2007). The fol-
lowing real-world applications belong to the class of moderately dynamic systems:
• Long-distance courier services and (express) parcel services: This routing
application occurs in companies which offer services for transporting mail and
parcels. Besides regular customers, there are customers who demand transporta-
tion services on short notice so that the vehicle tours are required to be replanned
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 71

each day. Furthermore, some companies accept late-in requests. In this case, cus-
tomers may demand goods to be picked up at their location after the vehicles
have already started to collect goods so that their tours require real-time adapta-
tions. In order to realize long-distance transportation services, different logistical
structures can be used. For example, shipments are collected at the customer lo-
cations and are transported to a hub. From this hub, the shipments are transported
to individual destination hubs from which they are transported to their final desti-
nations. For more information on such logistical structures, see, e.g., Bock (2004,
p. 48 et seq.) and Crainic and Kim (2007).
Besides standard transportation services, several companies offer next-day ex-
press services which guarantee the delivery of shipments to national and interna-
tional destinations by a specified time point the next day, e.g., 9:30 am, 10:30 am,
or 12:00 am. Moreover, very urgent freights can also be delivered on the same
day (cf. DHL Express Germany GmbH 2010). In order to efficiently coordinate
such time-critical transportation processes, the utilization of sophisticated real-
time control approaches is advantageous.
• Repairman companies with scheduled customers: Specialized companies pro-
vide repair services and maintenance for their customers, e.g., for maintaining
automated teller machines (cf. Larsen 2000, p. 15). A characteristic of this type
of repairman problems is that most requests are known in advance since they arise
from maintenance contracts in which regular service is specified in known time
intervals. In addition to customer requests known in advance, some urgent cus-
tomer requests which require an immediate service can occur during the planning
horizon. Other repairman companies mainly dealing with urgent requests which
are not known in advance are described in the next echelon.

2.6.5.3 Echelon III: Strongly Dynamic Systems

This echelon comprises real-world routing problems in which only few or no re-
quests are known before the execution of the transportation process. In these ap-
plications, the dod amounts to at least 80 % (cf. Larsen et al. 2002). Further-
more, arriving customer requests often require an urgent service. Due to primarily
customer-oriented objectives which often consider customer inconvenience depend-
ing on the request response time, one possibility for improving the solution qual-
ity is the utilization of stochastic knowledge in order to utilize available vehicles
more efficiently. More information about this will be given in the literature review
in Sect. 4.3. The following real-world routing problems belong to the echelon of
strongly dynamic systems:
• Telebus dial-a-ride systems: Larsen (2000, p. 64) presents a variant of a DARP
which can be found in telebus systems. In such systems, buses only operate on
demand, i.e., if customers request for transportation. For example, telebus sys-
tems are used on bus lines, either all day or at times where there is usually a very
low demand. According to Borndörfer et al. (1997), telebuses can also be rented
for handicapped people who cannot use public transportation systems. Similar
72 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

to taxi cab services, most customer requests dynamically arrive during the day.
However, since telebuses are usually required to be ordered some time in advance,
the urgency is lower than in taxi applications described later in this section.
• SLA repairman companies: Besides the previously mentioned repairman com-
panies in which most of the requests are known in advance, there exist repairman
companies which mainly deal with urgent problems arising in case of severe fail-
ures. As described in Chap. 1, example applications comprise professional repair-
man companies which quickly resolve severe problems faced by their customers,
e.g., manufacturing companies which are part of a supply chain. In such appli-
cations, repairman companies often need to comply with the regulations of SLA
contracts negotiated with their customers. Failing to fulfill the SLA results in high
penalty costs so that a quick reaction to dynamically arriving customer requests
is of high importance. Bertsimas and van Ryzin (1991) introduce the Dynamic
Traveling Repairman Problem (DTRP) and evaluate solution strategies which are
described in detail in Sect. 4.5.1.
• Subsequent delivery of newspapers: As introduced in Chap. 1, the subsequent
delivery of newspapers occurs when a subscriber has not received his newspa-
per in the morning as expected. In this case, a quick subsequent delivery of the
missing newspaper is required in order to minimize customer inconvenience and
maintain customer loyalty.
• Taxi cab services: Taxi drivers are often connected to a central call center where
new customer requests are recorded and dispatched. By using appropriate tech-
nical equipment (see Sect. 2.6.3), individual taxi cabs can be located in real-time
and pending customer requests are assigned to them (cf. Horn 2002a, 2002b).
Since people frequently demand instant transportation, quality of service is also
measured in this application by the resulting response time. These processes are
modeled as a Dynamic Dial-A-Ride Problem (cf. Cordeau and Laporte 2007)
since a customer request consists of two locations instead of one. Moreover, the
service time of a request, i.e., the time needed for transporting a customer com-
prises some uncertainty since travel times can only be estimated.
• Emergency services: For emergency services such as police, fire fighters, and
ambulances, typically no requests are known in advance. The solution quality of
an emergency system is often measured by the achieved request response time.
According to Larsen et al. (2007), typical ambulance emergency systems agree
on a certain level of service quality so that, for instance, 90 % of the requests
should be serviced within 5 minutes and the remaining requests within a maxi-
mum of 8 minutes. Due to the high significance of emergency services, there is
a lot of ongoing research on this topic which has also been the case for the last
decades. For example, Larson and Odoni (1981) propose the hypercube queu-
ing model which has been widely used, especially in the coordination of police
patrols. In Brotcorne et al. (2003), the authors study approaches for ambulance
dispatching developed during the past three decades. More information on relo-
cation strategies which have also been applied to this type of applications will be
given in the literature review in Sect. 4.3.2.
2.6 Dynamic Routing Problems Which Utilize a Centralized Coordination 73

As can be derived from this description, RDOPG applications, i.e., SLA repair-
men companies and the subsequent delivery of newspapers (see Sect. 1.1) belong to
the echelon of strongly dynamic systems.

2.6.6 Evaluating the Performance of Solution Approaches

for Dynamic Routing Problems

In what follows, we describe two measures known in the literature for evaluating
the performance of solution approaches for dynamic routing problems.

2.6.6.1 Competitive Analysis

Sleator and Tarjan (1985) introduced a general framework denoted as competitive

analysis. In dynamic routing problems, this framework can be used to define the
competitive ratio of a solution approach A denoted as CRA (cf. Larsen et al. 2007).
Given an instance I of the considered dynamic routing problem, zA (I ) defines the
objective function value of the solution generated by A applied to this instance.
Furthermore, z∗ (I O ) represents the objective function value of the solution gen-
erated by an optimal solution approach when applied to the corresponding offline
instance I O , i.e., the corresponding static instance of I in which the optimal so-
lution approach has access to all relevant information before the execution of the
transportation process. Furthermore, the set M comprises all theoretically possible
instances of the considered dynamic problem. The competitive ratio of solution ap-
proach A is defined as:
zA (I )
CRA = sup
I ∈M z∗ (I O )
By definition, CRA represents the worst-case ratio of solution approach A compared
to an optimal offline solution approach. Therefore, the competitive ratio is a measure
for evaluating the worst-case performance of A on all instances of a given dynamic
routing problem. One issue with applying the competitive ratio is that up to now
many dynamic routing problems can only be empirically evaluated by simulation-
based methods. Hence, for such problems, it is not possible to evaluate all theoret-
ically possible instances I ∈ M. Nevertheless, there exist specific dynamic routing
problems to which the competitive ratio has been successfully applied (cf. Bertsi-
mas and van Ryzin 1991, 1993a, 1993b, de Paepe 2002, and Angelelli et al. 2007).
In these problems, specific assumptions, e.g., about request arrivals and their dis-
tribution, are made. In case of dynamic events, plan adaptations are carried out by
strategy-oriented approaches which allow the competitive ratio to be applied since
the effects of the applied adaptation methods on the solution quality can be analyti-
cally studied. Strategy-oriented approaches will be described in the literature review
in Sect. 4.5.1.
74 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

2.6.6.2 The Value of Information

In order to evaluate the performance of solution approaches on dynamic routing

problems which cannot be evaluated by the competitive ratio previously introduced,
Mitrovic-Minic et al. (2004) propose a measure denoted as the value of information.
It evaluates the worth of having complete knowledge of all relevant information
before the execution of the transportation process. Given a solution approach H
(often a heuristic approach) and an instance I of the considered dynamic routing
problem, zH (I ) represents the objective function value of the solution generated
by H for problem instance I . Furthermore, zH (I O ) represents the objective function
value of the solution which has been generated by H on the corresponding offline
instance of I . The value of information of solution approach H on instance I is
defined as:
zH (I O ) − zH (I )
VH (I ) =
zH (I O )
Following this definition, the value of information represents the maximum possi-
ble gain in solution quality when H is applied on I if all information is known in
advance. Hence, a larger value of VH (I ) indicates instances in which it is especially
beneficial when more relevant information is known before the execution of the
transportation process, for example by utilizing stochastic knowledge about future
expected events. Based on the value of information, we will describe in Sect. 8.2.6
a measure which is used in our computational results for evaluating differences in
the performance of the real-time control approaches developed in this book.
Although the value of information gives an upper bound on improvements in
solution quality which are attainable if more information, e.g., suitable stochastic
knowledge, is known in advance, it does not consider specific characteristics which
are present in the application or in test instances. Moreover, note that the estimation
of this measure is also solution approach-dependent. In this book, we will propose
new measures which take into account specific characteristics present in the consid-
ered routing problem. Hence, these measures allow for estimating the improvements
which are additionally attainable just by analyzing the request data without requir-
ing empirical assessment of the attained solution quality in the online and offline
case and therefore allows for an approach-independent estimation. Moreover, these
measures allow for identifying specific characteristics of request data sets which
contribute to an efficient utilization of stochastic knowledge thereby assessing pos-
sible improvements and supporting estimations more precisely. These measures will
be described in Sects. 8.8.2 and 8.9.

2.7 A General Classification Scheme for Routing Problems

In order to capture all attributes of routing problems, several authors have developed
a general classification scheme (cf., e.g., Bodin and Golden 1981, Stumpf 1998,
Richter 2005, and Eksioglu et al. 2009). In this section, a variant of this scheme
2.7 A General Classification Scheme for Routing Problems 75

Table 2.2 General classification scheme for routing problems including RDOPG-specific at-
tributes (in accordance with, e.g., Bodin and Golden 1981, Stumpf 1998, Richter 2005, and Ek-
sioglu et al. 2009)
Attribute Possible values

Requests
Type Pickup only
Delivery only
Pickup and delivery on the same tour (with backhauls,
see Sect. 2.4.2)
Combined requests (Pickup and delivery)
Location At nodes
At arcs/edges
Mixed
Geographical distribution Distributed with a pattern
Randomly distributed
Degree of dynamism Low
Moderate
High
Percentage of regular customers None
Few
Medium
High
All
Time restrictions No time windows
One-sided time window
Two-sided time window
Fixed time points for service
Hard time windows
Soft time windows
Multiple time windows
Divisibility Non-divisible requests
Arbitrarily divisible requests
Restricted divisible requests (e.g., minimum amount
per part)
Transshipment No transshipment
Transshipment possible
Request size Full-truckload (FT)
Less-than-truckload (LTL, e.g., piece goods)
Types of transportation objects Goods
Services
Persons
76 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Table 2.2 (Continued)

Attribute Possible values

Size of service area Urban

Rural
Regional
National
International
Sequence of requests Arbitrary
Predecessor-/Successor-restrictions
Selection of requests Service of all requests
Possibility to select requests to be serviced, i.e.,
requests can be rejected
Frequency of request occurrence Once
Semi-periodic
Periodic

Fleet
Size One vehicle
Multiple vehicles, restricted fleet size
Multiple vehicles, unlimited fleet size
Consistency Homogeneous (one type of vehicles)
Heterogeneous (different types of vehicles)
Capacity No capacity constraints
One-dimensional capacity constraints
Multi-dimensional capacity constraints
Location One depot
Multiple depots
Freely distributed
Number of tours per vehicle One tour
and planning horizon Multiple tours
Time restrictions No restrictions
Maximum duration of tour
Loading restrictions Arbitrary loading
Restricted set of goods can be loaded
Number of drivers per vehicle (Number)
Driver to vehicle assignment Fixed
Free
Vehicle reliability Reliable
Vehicle slowdowns possible
Vehicle breakdowns possible
2.7 A General Classification Scheme for Routing Problems 77

Table 2.2 (Continued)

Attribute Possible values

Network
Type of network Road network (real distances)
Coordinate system (Euclidean distances)
Data Symmetric (undirected graph)
Asymmetric (directed graph)
Mixed
Travel times Constant
Time-dependent known
Dynamic with uncertainty

Planning horizon
Length One period
Multiple periods
Open-ended
Type Once
Continuous
Repeating

Tour plan
Type Daily changing tour plan
Fixed tour plan
Fixed tour plan with slack time for flexibility
Type of tours Open tours (without return to depot)
Closed tours (with return to depot)
Restrictions Maximum tour duration
Maximum tour length
Worker-dependent restrictions (e.g., maximum
working time, driving time restrictions)

Relevant information
Type of relevant information Constant
Changing
Type of problem Static
Dynamic, changes deterministically known
Dynamic, changes comprise uncertainty
Availability of changes Global
Local
Mixed
Types of considered dynamic events New requests
Changes in request attributes
78 2 Introduction to Tour Planning: Vehicle Routing and Related Problems

Table 2.2 (Continued)

Attribute Possible values

Traffic congestions and road blockages

Vehicle disturbances
Type of data Real-world data
Designed data
Both real-world and designed data

Stochastic knowledge
Type of stochastic knowledge None
Prescribed probability distributions are available
Forecasted by analyzing historical data, e.g., past
request information

Increasing tour plan flexibility

Methods used for increasing None
tour plan flexibility Waiting strategies
Relocation strategies
Request assignment strategies

Objective
Considered objective Minimization of travel length
Minimization of travel time
Minimization of operational costs
Minimization of the number of utilized vehicles
Minimization of the number of unserviced customers
Minimization of response time
Minimization of customer inconvenience

Applied approach
Processing of changes in the relevant Centralized
information Decentralized
Mixed
Type of applied methods Exact
Heuristic
Simulation
Real-time control methods
Type of model Deterministic static
Stochastic static
Deterministic dynamic
Stochastic dynamic
Type of application Static
Dynamic
2.8 Summary 79

extended by attributes which are relevant for the considered RDOPG applications is
presented. The resulting classification scheme is illustrated in Table 2.2.
The RDOPG applications considered in this book will be characterized according
to this general classification scheme in Sect. 3.2.3.

2.8 Summary

In this chapter, we discussed different characteristics and variants of routing prob-

lems. First, an introduction to classic static routing problems was given. Since
RDOPG applications can be modeled as a DVRP variant in which new requests
dynamically arrive, variants of the VRP and the DVRP made up the main part of
this chapter. Specifically, first static VRPs which differ from each other in terms of
problem attributes, such as considered type of requests, number of depots, available
vehicles, and pursued objective function, were described. After this description, dy-
namic routing problems in which the relevant information changes during the exe-
cution of the transportation process were discussed and classifications known in the
literature which deal with various differences between static and dynamic routing
problems were described. With regard to attributes specific for RDOPG applica-
tions, a unified approach for classifying routing problems based on insights from
known classifications in the literature was introduced. In this unified approach, the
type of relevant information and corresponding changes that occur in the problem
data during the execution of the transportation process as well as the availability and
processing of these changes were considered. Moreover, depending on the type of
available stochastic knowledge, different variants of VRPs known in the literature
were described. After this description, in accordance with the considered RDOPG
applications, dynamic routing problems in which a centralized coordination of the
transportation process is performed were considered in detail. This included typi-
cal objectives pursued in dynamic routing problems and a description of dynamic
events considered in dynamic routing problems. Moreover, a measure denoted as
the degree of dynamism which has been proposed by Lund et al. (1996) and Larsen
(2000) was described. This measure is used for determining the difficulty level of
solving individual problem instances of dynamic routing problems. Based on this
measure, a three-echelon classification proposed by Larsen (2000) was introduced
which includes the characterization of various real-world dynamic routing appli-
cations. In order to evaluate the performance of dynamic routing approaches, two
measures known in the literature were presented, in particular the competitive ra-
tio and the value of information. Finally, a general classification scheme for char-
acterizing routing problems was presented and extended with regard to additional
characteristics which are relevant for RDOPG applications.
Chapter 3
The Considered RDOPG Applications

In Chap. 1, different real-world problems which belong to RDOPG applications

considered in this book were introduced. In this chapter, we introduce the problem
description of RDOPG applications in detail. Following this description, character-
istics of RDOPG applications are discussed.

3.1 Problem Description of RDOPG Applications

In the considered RDOPG applications, a daily distribution process has to be con-

trolled in real-time. In this process, a set of R customer requests R = {1, . . . , R}
that occur dynamically during the day needs to be serviced by a set of K homoge-
neous vehicles K = {1, . . . , K} which are originally located at one depot to which
they return after their tour has been completed. During the execution of the trans-
portation process, the relevant information is changed by dynamic events by which
new requests dynamically arrive. As mentioned in Chap. 1, the vehicle tours consist
of delivery activities, i.e., vehicles deliver goods that are loaded onto the vehicles
at the depot before the execution of the transportation process. All of the arriving
requests have to be accepted and serviced by the available vehicles. In other words,
they cannot be rejected or subcontracted. Moreover, vehicle capacity constraints can
be regarded as negligible since a sufficient amount of the goods to be delivered is
loaded onto the vehicles at the depot before the transportation process starts. The
considered planning horizon has a length of up to one working day and the consid-
ered service region consists of a medium to large-sized urban area.
As mentioned in Chap. 1, the goods to be delivered are denoted as perishable
goods because they have to be delivered as quickly as possible once they are re-
quired by customers. Therefore, the considered objective pursues the minimization
of the total customer inconvenience. This total customer inconvenience consists of
the sum of all individual customer request inconvenience values which depend on
the request response times, i.e., the time period between request arrival and start of
service (see also Sect. 2.4.12.7). In the marketing literature, two main approaches to

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 81

DOI 10.1007/978-3-642-33472-6_3, © Springer-Verlag Berlin Heidelberg 2013
82 3 The Considered RDOPG Applications

response time R mrt

Maximum allowed
Customer inconvenience

Lateness penalty R pen

Request arrival

linear2X
quadratic

Response time

Fig. 3.1 Illustration of the linear2X and quadratic customer inconvenience

model customer inconvenience dependent on elapsed response time can be found.

Davis and Maggard (1990) propose that customer inconvenience increases linearly
with resulting response time. In contrast, Kristensen et al. (1992) derive a quadratic
dependency. Since both approaches are reasonable for the considered RDOPG ap-
plications, we evaluate both alternatives in this book. Further information on other
approaches that utilize customer inconvenience aspects in the objective function can
be found in the literature review in Sect. 4.6. Additionally, a maximum request re-
sponse time denoted by R mrt has to be considered for each request, otherwise, high
penalty costs denoted by R pen occur. In this case, the loss of the corresponding
customer is likely, which should be prevented under all circumstances. Moreover,
this integrates an effective method for avoiding an indefinite deferment of individ-
ual customer requests (see aspect 8 of Psaraftis’ approach presented in Sect. 2.5.2.2)
which is desirable if the quality of customer service aspects is to be considered. The
evolution of customer inconvenience used in RDOPG applications is illustrated in
Fig. 3.1.
As mentioned in Sect. 1.3, we investigate the subsequent delivery of newspa-
pers as the exemplary real-world RDOPG application in this book. Nevertheless,
note that other related RDOPG applications can be modeled by setting the problem
parameters accordingly. One important characteristic of the subsequent delivery of
newspapers is that the service time which is required for servicing a request can be
considered to be short and equal for all customer requests. In this book, we assume a
service time of R st = 60 seconds for all requests. Other related real-world problems
for RDOPG applications can be found in Sect. 1.1. In what follows, further elements
which are part of the considered RDOPG problems are introduced.
3.1 Problem Description of RDOPG Applications 83

3.1.1 Existence of a Real Urban Road Network

In order to enable an accurate mapping of vehicle movements on roads and to cal-

culate the resulting travel times between customer requests and current vehicle po-
sitions accordingly, a real-world urban road network is used. In this road network,
different road classes which allow for modeling individual travel speeds are present.
Moreover, one-way roads which restrict possible travel paths exist. Note that using
a road network enables a much higher degree of realism compared to assuming that
vehicle movements are performed on Euclidean distances between relevant loca-
tions. The utilized road network will be introduced in detail in Sect. 5.3.1.

3.1.2 Presence of Vehicle On-Board Units and a Central

Dispatching Center

In order to enable tracking of vehicle positions and to inform vehicles about per-
formed tour adaptations in real-time, on-board communication units are assumed
to be available in the deployed vehicles. This allows a bidirectional communica-
tion between the vehicles and the dispatching center which performs a centralized
coordination of tour plan adaptations during the execution of the transportation pro-
cess (see Sect. 2.6).

3.1.3 Availability and Utilization of Past Request Information

In the considered RDOPG applications, past request information is available. This

past request information consists of arrival time and location of all requests that
have arrived in the system on elapsed working days. Note that such information
is available in many real-world applications. In our proposed approach, this past
request information is used to generate stochastic knowledge in order to anticipate
future requests which are expected to occur with a high probability during the next
planning horizon (see also Sect. 2.5.3.2.2).
With regard to the considered objective of minimizing the total customer incon-
venience, an improved solution quality can be achieved by quickly responding to
dynamically arriving customer requests. This quick response can be attained by
distributing the available vehicles over the considered service region so that they
are efficiently prepared to service expected future request arrivals. Specifically, by
guiding vehicles into request-likely areas in which new requests are likely to appear
and keeping vehicles there for a specified time period, newly arriving requests can
be serviced faster. Note that a request-likely area is represented by a subregion of
the service region during a certain time period only since request arrival rates may
change during the planning horizon. The generated stochastic knowledge represents
such request-likely areas in the proposed pro-active real-time control approach.
84 3 The Considered RDOPG Applications

Note that in our considered problem setting, the stochastic knowledge is gener-
ated by solely analyzing and examining the available past request information. In
particular, information about expected future request arrivals is only derived from
this data. Since no prescribed probability distributions are available for generating
stochastic knowledge, the underlying distributions of the available past request in-
formation are unknown to the approach which forecasts the stochastic knowledge.
This is representative of many practical applications. The proposed forecasting ap-
proaches for generating and integrating stochastic knowledge into the pro-active
real-time control approach will be presented in Chap. 6.

3.1.4 Allowing Vehicle En-Route Diversion

In the considered problem setting, allowing en-route diversion of vehicles enables a

reassignment of the first request in a vehicle’s tour to which the vehicle is currently
traveling. This increases dispatching flexibilities and enables a more direct reaction
to newly arriving requests (cf. Ichoua et al. 2000). Specifically, by allowing vehicle
en-route diversion, a newly arriving request which is close to the current position
of a vehicle can be serviced quickly since the corresponding vehicle only needs
to perform a small detour on its way to its currently assigned first request. Ichoua
et al. (2000) and Richter (2005, p. 219) have shown that this can be advantageously
applied if the considered objective function comprises quality of customer service
aspects, as is the case in the studied RDOPG applications. Note that due to the use
of a real road network, vehicles can only change their direction at the next road
junction. However, since the utilized urban road network is very dense, individual
road parts are quite short so that the next road junction can often be quickly reached
by a vehicle.

3.1.5 Integration of Vehicle Scheduling Decisions

As mentioned in Sect. 3.1.3, by making use of stochastic knowledge, vehicles can

be kept in request-likely areas for specific time periods. Since the length of these
time periods is determined by the proposed pro-active real-time control approach
according to the current request situation, vehicle scheduling decisions become a
part of the considered problem. Therefore, the problem considered in RDOPG ap-
plications belongs to the class of Vehicle Routing and Scheduling Problems (VRSP,
see Sect. 2.4.6) if the available stochastic knowledge, i.e., the pro-active real-time
control approach, is utilized.
3.2 Characteristics of RDOPG Applications 85

3.2 Characteristics of RDOPG Applications

After the problem considered in RDOPG applications has been described, in this
section we examine further characteristics of RDOPG applications. First, differences
from other dynamic routing problems are discussed. After that, the computational
complexity of RDOPG applications is derived. The proposed classification approach
for routing problems described in Sect. 2.5.3 is used for classifying RDOPG appli-
cations with regard to characteristics and processing of relevant information. More-
over, a general classification of RDOPG applications is presented according to the
classification scheme for routing problems which was introduced in Sect. 2.7.

3.2.1 Differences from Other Dynamic Routing Problems

As mentioned above, RDOPG applications consider the distribution of perishable

goods without transshipment to dynamically arriving customer requests using a
known fleet size located at one depot. According to the different variants and charac-
teristics of routing problems which have been presented in Chap. 2, RDOPG appli-
cations can be modeled on the basis of an Asymmetric Dynamic Forecasted Stochas-
tic Vehicle Routing Problem with Soft Time Windows (ADFSVRPSTW) which is
a specific variant of the Dynamic Vehicle Routing Problem with Soft Time Win-
dow constraints (DVRPSTW, cf. Gendreau et al. 1999a). The considered problem is
asymmetric since travel times between relevant locations can depend on the direc-
tion of travel due to one-way streets in the road network. Furthermore, the problem
is dynamic since new requests arrive during the execution of the transportation pro-
cess. In the problem, stochastic knowledge is available which is solely generated
from available past request information resulting in forecasted stochastic knowl-
edge. Moreover, the considered distribution of goods from one depot to customer
locations can be modeled by an uncapacitated VRP (see Sect. 2.3.2). In order to
model the considered objective function, soft time windows with variable and fixed
penalty costs are used. Specifically, variable penalty costs directly starting upon re-
quest arrival represent customer inconvenience while fixed high penalty costs occur
when the maximum response time R mrt is not fulfilled.
Differences from other dynamic routing problems arise from the structure of cus-
tomer requests. An analysis of the real-world request data of the subsequent delivery
of newspapers (see Sect. 8.2.4) has shown that only a negligible part of the cus-
tomers is known before the execution of the transportation process. The largest part
of all customer requests dynamically arrive during the execution of the transporta-
tion process. Specifically, over 90 % of all customer requests are dynamic. Hence,
in comparison to other DVRPs with a low or moderate degree of dynamism, the
considered RDOPG applications are determined by a high degree of dynamism (see
Sect. 2.6.4). Moreover, the utilized objective function distinguishes RDOPG appli-
cations from many other real-time routing problems with soft time windows. As
mentioned in Chap. 1, in many other approaches a significant time period between
86 3 The Considered RDOPG Applications

the request arrival and the start time of penalty costs incurred by violations of soft
time window restrictions exists. In the edodtw presented in Sect. 2.6.4, this time pe-
riod is denoted as reaction time. Since customer inconvenience directly starts with
the arrival of the request in the system, dynamic requests have a reaction time of
0 and therefore no grace period. Owing to the absence of such a grace period,
appropriate computational experiments have revealed a value of information (see
Sect. 2.6.6.2) of about 10 which is high compared to other dynamic routing prob-
lems. For example, Gendreau et al. (1999a) state that a value of information between
1.025 and 1.041 was attained, depending on the utilized request data scenario. Note
that we were able to achieve similar results for the value of information by allowing
a grace period of 30 minutes between request arrival and the start of customer in-
convenience in our problem setting. Due to the high value of information, it can be
concluded that using additional available information about expected future events
is likely to result in a significantly increased solution quality. Hence, in the compu-
tational experiments in Chap. 8, the impact of utilizing stochastic knowledge in the
considered RDOPG applications will be evaluated in detail.

3.2.2 Computational Complexity

In the previous section it was described that RDOPG applications can be repre-
sented as a variant of the DFSVRP. As mentioned in Sect. 2.5.3.2.2, the DFSVRP
is a DVRP which is extended by stochastic knowledge. According to Larsen (2000,
p. 5), “the DVRP is a richer problem compared to the conventional static VRP. If
the problem class of VRP is denoted P(VRP) and the problem class of DVRP is
denoted P(DVRP), then P(VRP) ⊂ P(DVRP)”. Since the VRP belongs to the
class of strongly NP-hard combinatorial problems (see Sect. 2.3.4), the considered
RDOPG applications also belong to the class of strongly NP-hard combinatorial
problems.

3.2.3 Classifying RDOPG Applications with Regard to

Characteristics and Processing of Relevant Information

In this section, we characterize the considered RDOPG applications with re-

gard to our proposed classification for categorizing routing problems presented in
Sect. 2.5.3. According to this classification, RDOPG applications have the following
attributes:
• Type of relevant information and type of problem: During the execution of the
transportation process, dynamic events occur by which new customer requests
arrive in the system. Since these dynamic events and their impacts are not deter-
ministically known in advance, changing relevant information is present. These
3.3 Summary 87

changes in the relevant information lead to modifications of the problem which

cannot be predicted in advance during the execution of the transportation process.
Hence, a dynamic problem is present.
• Availability of changes in the relevant information and type of stochastic
knowledge: As mentioned in Sect. 3.1.2, available on-board units in the vehi-
cles allow for a bidirectional communication with the central dispatching center.
At this dispatching center, new requests are centrally collected so that a global
knowledge of dynamic data is present. Moreover, as mentioned in Sect. 3.1.3,
past request information is available so that stochastic knowledge can be gener-
ated from past request information.
As mentioned above, the RDOPG applications are modeled by a variant of a
DFSVRP. Hence, the applied real-time control approaches which will be presented
in Chaps. 5 and 6 has the following attributes:
• Type of transportation process coordination: Adaptations to the tour plan are
carried out by the dispatching center. Hence, a centralized coordination of the
transportation process is applied.
• Type of model: The occurring changes in the relevant information during the
execution of the transportation process update the model so that a deterministic
dynamic model is present in the real-time control approach presented in Chap. 5.
Moreover, in the pro-active real-time control approach, future expected events
derived from the available stochastic knowledge are also considered in the model.
Hence, a stochastic dynamic model is present.
• Type of application: The problem represented by the stochastic dynamic model
is controlled in real-time during the transportation process. This results in a dy-
namic application.
In Fig. 3.2, the considered RDOPG applications are categorized according to the
classification for routing problems proposed in Sect. 2.5.3.

3.2.4 General Classification of RDOPG Applications

In Sect. 2.7, a general classification scheme for routing problems was introduced.
In Table 3.1, RDOPG applications and the applied pro-active real-time control ap-
proach are classified according to this classification scheme.

3.3 Summary
In this chapter, the problem of coordinating transportation processes of RDOPG
applications in real-time was described in detail. One of the main characteristics
which distinguishes RDOPG applications from several other dynamic routing appli-
cations is the objective function of minimizing customer inconvenience. Customer
88 3 The Considered RDOPG Applications

Relevant Example
information is variant
globally available
and given Deterministic
by constant Centralized knowledge
Constant
relevant
parameters
/ Static problem, processing
/ Centralized available / VRP
no changes coordination
information K
KK
G KK M: deterministic static
KK

A: static
KK
K
Relevant information KK
%
is globally available
Dynamic Centralized
Deterministic
knowledge
Relevant information
and given by known
does not change
functions of time problem, processing
/ Centralized available
/ TDVRP
time-dependent coordination
data M: deterministic static
A: static

Prescribed
stochastic
Dynamic Decentralized knowledge
Type of
relevant
problem, local processing
/Decentralized / SVRP
available
knowledge of coordination
information @ F >>
---- dynamic data >> M: stochastic static
----
E >>
>>
A: static
----

>>
----
>>

Local availability of
---- relevant information
changes, no bidirectional

-- - communication available

---

Decentralized Agent-based
Relevant information ----

processing and related

----

changes with approaches

----

uncertainty over time Decentralized

----

processing

----

--
Partly local, partly

global availability
of relevant Dynamic Mixed Mixture of
Changing information
relevant
changes
/ problem, processing
/ Mixed / decentralized
mixed coordination and centralized
information A approaches
6666 knowledge of
6666 dynamic data
6666
6666
6666
6666 No stochastic 8 DVRP
6666 knowledge available
6666
6666
M: deterministic dynamic
A: dynamic
Global availability of 6666
relevant information
6666 Mixed processing
changes, bidirectional
6666
66
communication available
Prescribed
stochastic
Dynamic Centralized knowledge
problem, processing
+3 Centralized available / DSVRP
global coordination
<<<<
knowledge of <<<< M: stochastic dynamic
dynamic data <<<< A: dynamic
<<<<
<<<<
Stochastic knowledge <
<<<
<<<<
can be generated
<<<<
from past request
information
<<!
DFSVRP
N.B.:
M denotes the type of model used in an approach, M: stochastic dynamic
A denotes the type of application used in an approach (see Section 2.5.2.3). A: dynamic

Fig. 3.2 RDOPG applications categorized according to the proposed classification of routing
problems
3.3 Summary 89

Table 3.1 The general classification scheme for routing problems applied to RDOPG applications
Attribute Possible values

Requests
Type Delivery only
Location At nodes
Geographical distribution Distributed with a pattern
Degree of dynamism High (over 90 %)
Percentage of regular customers None
Time restrictions Soft time windows
Divisibility Non-divisible requests
Transshipment No transshipment
Request size Less-than-truckload (LTL, e.g., piece goods)
Types of transportation objects Goods (material and immaterial, i.e., services)
Size of service area Urban
Sequence of requests Arbitrarily
Selection of requests Service of all requests
Frequency of request occurrence Once

Fleet
Size Multiple vehicles, restricted fleet size
Consistency Homogeneous (one type of vehicles)
Capacity No capacity constraints
Location One depot
Number of tours per vehicle and planning One tour
horizon
Time restrictions No restrictions
Loading restrictions Arbitrary loading
Number of drivers per vehicle 1
Driver to vehicle assignment Fixed
Vehicle reliability Reliable

Network
Type of network Road network (real distances)
Data Asymmetric (directed graph)
Travel times Constant

Planning horizon
Length One period (i.e., one day)
Type Repeating (from day to day)
90 3 The Considered RDOPG Applications

Table 3.1 (Continued)

Attribute Possible values

Tour plan
Type Daily changing tour plan
Type of tours Closed tours (with return to depot)
Restrictions None

Relevant information
Type of relevant information Changing
Type of problem Dynamic, changes comprise uncertainty
Availability of changes Global
Types of considered dynamic events New requests
Type of data Both real-world and designed request data

Stochastic knowledge
Type of stochastic knowledge Forecasted; past request information is
available

Increasing tour plan flexibility

Methods used for increasing tour Waiting strategies
plan flexibility Relocation strategies (pro-active real-time
control approach)
Request assignment strategies (pro-active
real-time control approach)

Objective
Considered objective Minimization of customer inconvenience

Applied approach
Processing of changes in the relevant Centralized
information
Type of applied solution method Exact solution methods used in the proposed
approach for generating stochastic knowledge
out of the past request information
Pro-active real-time control approach utilizing
simulation and heuristic solution methods for
adapting the tour plan during the transportation
process
Type of model Stochastic dynamic, using available stochastic
knowledge
Type of application Dynamic (Real-time control approach utilizing
simulation)
3.3 Summary 91

inconvenience directly starts upon arrival of a customer request in the system and
therefore can be distinguished from penalty costs which only occur if a request is
serviced late. According to results from the marketing literature and in order to eval-
uate different effects, customer inconvenience is modeled as a linear and a quadratic
function. Moreover, a maximum allowed response time is defined in order to avoid
an indefinite deferment of individual customer requests. Another important aspect
is that vehicles move on a real road network and vehicle en-route diversion is al-
lowed during the execution of the transportation process. The availability of past
request information allows for the application of appropriate forecasting methods
for generating stochastic knowledge out of it. This is a prerequisite for applying
the proposed pro-active real-time control approach to RDOPG applications. After
the problem description, various attributes of the considered RDOPG applications,
including their problem complexity, were described. Finally, RDOPG applications
were characterized according to the proposed classification approach for routing
problems (see Sect. 2.5.3) as well as to the general classification scheme which was
presented in Sect. 2.7.
Chapter 4
Review of the Literature Related
to the Considered RDOPG Applications

Routing problems are a vital research area, not least due to their practical relevance.
Recent advances in computational power and available technological equipment
have opened up new areas for research and applications in practice. As illustrated in
Sect. 3.2.1, RDOPG applications can be modeled as a variant of a Dynamic Vehicle
Routing Problem (DVRP). Hence, this literature review focuses on DVRPs that are
controlled by real-time methods in which changes in the relevant information are
centrally available and centrally processed (see Sect. 2.5.3). For static VRPs, dif-
ferent extensive surveys exist in the literature. We refer the interested reader to the
recent and detailed reviews of exact and approximated approaches for static VRPs
by Laporte (1992a), Toth and Vigo (2002), Cordeau et al. (2005), Baldacci et al.
(2007), and El-Sherbeny (2010). Moreover, for a recent and thorough survey of Dy-
namic Pickup and Delivery Problems (DPDPs) as well as of Dynamic Dial-A-Ride
Problems (DDARPs) including recent results from competitive analysis, we refer to
Berbeglia et al. (2010). Further details on static and dynamic DARPs can be found
among others in Cordeau and Laporte (2003, 2007) and Cordeau et al. (2007a). Be-
sides DVRP based approaches, approaches of other types of routing problems will
be presented in this review if they contain specific additional relevant aspects.
In this review, the terms real-time, dynamic, and online will be synonymously
used. The presented literature review is based on the comprehensive surveys by
Psaraftis (1988, 1995), Powell et al. (1995), Séguin et al. (1997), Gendreau and
Potvin (1998), Bianchi (2000), Larsen (2000), Cordeau and Laporte (2003, 2007),
Ghiani et al. (2003), Lackner (2004), Richter (2005), Flatberg et al. (2005), Ichoua
et al. (2007), Cordeau et al. (2007a), Larsen et al. (2007, 2008), Eksioglu et al.
(2009), and Pillac et al. (2010).
As mentioned in Sect. 2.5.3.2, solution approaches for dynamic routing prob-
lems can be divided into reactive, deterministic, and pro-active real-time control
approaches. This literature review primarily deals with deterministic and pro-active
real-time control approaches. These types of real-time control approaches are deter-
mined by providing additional flexibility in the transportation process. With regard
to the considered RDOPG applications, this is very important since the majority of
requests only arrive during the execution of the transportation process so that the

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 93

DOI 10.1007/978-3-642-33472-6_4, © Springer-Verlag Berlin Heidelberg 2013
94 4 Review of the Literature Related to the Considered RDOPG Applications

tour plan frequently needs to be adapted according to the changes in the system sit-
uation. Besides deterministic approaches which do not use stochastic knowledge, in
this literature review, various pro-active approaches are presented according to the
type and the utilization of stochastic knowledge.
This chapter is organized as follows. Some of the first papers on dynamic rout-
ing problems in the literature are introduced. This is followed by a description of
selected reactive real-time control approaches for dynamic routing problems. Fur-
thermore, strategies for increasing flexibility in dynamic routing problems which
can be found in the literature are described. The main part of this literature review
describes solution approaches which provide flexibility in dynamic routing with-
out stochastic knowledge followed by those that allow for providing flexibility in
dynamic routing using stochastic knowledge. Since the objective function in the
considered RDOPG applications differs from many other approaches of dynamic
routing, special attention is also given to approaches in the literature which utilize
related objective functions. After that, approaches in the literature that cover other
relevant factors for RDOPG applications, such as vehicle en-route diversion, real-
time control concepts, or communication technologies, are presented. The chapter
closes with a conclusion of the literature review in which aspects that have not yet
been dealt with in the research on pro-active approaches are also described.

4.1 First Papers on Dynamic Routing Problems in the Literature

The majority of approaches which deal with dynamic routing problems have been
developed during the last two decades while pro-active approaches have only at-
tained increased attention in recent years. Nevertheless, the first solution approaches
were already published more than 30 years ago. One of the first papers that deals
with a dynamic routing problem was presented by Wilson and Colvin (1977). The
authors conduct research on a single-vehicle DARP in which customer requests dy-
namically arrive during the execution of the transportation process. The requests
represent customers that request a trip from a given origin to a given destination.
The authors propose insertion heuristics for solving the considered problem.
Psaraftis (1980) also deals with a DARP utilizing a single capacitated vehicle.
Two cases are considered in this work. First, research on the static case is conducted,
in which all customers are known in advance. The author considers the minimiza-
tion of a weighted combination of the time to service all customers and the total
customer dissatisfaction experienced while waiting for service. The dissatisfaction
of a customer is modeled by a linear function depending on the waiting and ride
times. The second study considers the related dynamic routing problem in which
customers dynamically appear and demand an instant service when they arrive in
the system. In this case, the objective solely consists of minimizing customer dissat-
isfaction. Note that this type of objective function is similar to the one considered
in the RDOPG applications in this book since no grace period between customer
4.2 Selected Reactive Real-Time Control Approaches for Dynamic Routing 95

request arrival and start of customer dissatisfaction exists (see also Sect. 3.1). The
author develops a dynamic programming approach with a special priority rule called
maximum position shift (MPS) which prohibits an indefinite postponement of cus-
tomer requests. Benchmark instances with up to 10 customers are examined and
optimally solved. The author notes that, due to the MPS rule, the size of the search
space is restricted in comparison to the size of the search space that occurs when the
MPS rule is not used. Therefore, the tested benchmark instances could be optimally
solved. However, since the search space exponentially grows, the author encourages
the comparison of the performance of exact approaches to heuristic approaches. Fur-
thermore, the author suggests the use of the presented exact method as a subroutine
in multiple-vehicle variants of the considered problem.
Another early contribution to the dynamic routing literature is the work of Bell
et al. (1983). In the described real-world problem, the delivery of industrial gases
from a depot to a set of customers is considered. The objective function pursues the
maximization of the sum of delivered gas quantities. The tour plan is generated for
the next two to five days but the focus is on the part of the tour plan that is exe-
cuted on the current day. The problem is partly static and partly dynamic since all
requests of the current day are known in advance while additional requests arriving
during the day are only required to be serviced on the days that follow. However,
dynamism also arises on the current day since the quantity of the industrial gas to
be delivered to the individual customers can only be estimated with uncertainty. The
exact amount of gas that a customer needs can only be determined while filling the
tank at the customer’s location so that adaptations of the tour plan may be required
if estimated delivery quantities are not sufficient. The investigated problem utilizes
a real road network for calculating mileage and toll costs involving a heterogeneous
fleet and time windows of customer requests. To solve the problem, the authors pro-
pose a two-step solution method. In the first step, a set of possible one-day tours is
determined. In the second step, tours are selected by decomposing the MIP prob-
lem using a Lagrangian relaxation into a knapsack problem for each vehicle tour
and embedding the relaxed problem into a Branch-and-Bound (B&B) procedure.
According to the authors, the integration of the proposed approach in a real-world
delivery process has achieved savings of 6 % to 10 % in operational costs while
increasing the productivity of the deployed vehicles.

4.2 Selected Reactive Real-Time Control Approaches

for Dynamic Routing Problems

In the last decade, a lot of research on DVRPs and related routing problems has been
published and reactive real-time control approaches have been developed for these
problems. In place of this magnitude of various research work, we present some
research proposing reactive real-time control approaches in this section. Research
work proposing pro-active real-time control approaches will be presented in the
literature review in the sections that follow.
96 4 Review of the Literature Related to the Considered RDOPG Applications

Pankratz (2002) considers a DPDP for the long-distance freight forwarding in

which multiple orders can be simultaneously loaded onto the vehicles. Outsourcing
of requests to subcontracted companies (at additional costs) is allowed and vehicles
can be immediately reused after their tour has ended. Dynamism is considered by
newly arriving requests during the day. In order to solve the routing problem, a
heuristic solution approach based on a Genetic Algorithm (GA) is proposed. For
future research, the author suggests considering additional sources of dynamism
such as vehicle breakdowns and traffic jams.
In Lackner (2004) four metaheuristics are developed and applied to the capaci-
tated Dynamic Vehicle Routing Problem with Time Windows (DVRPTW). These
four metaheuristics consist of two evolutionary strategies, an Ant Colony Opti-
mization based algorithm, and a Simulated Annealing metaheuristic. In the con-
sidered problem, dynamism arises from newly incoming requests. The author an-
alyzes the behavior of the implemented metaheuristics under different degrees of
dynamism (see Sect. 2.6.4) ranging from 10 % to 90 % as well as under different
scenarios that differ in the arrival rate of customer requests, in particular one to nine
new requests per minute. A single event optimization strategy is used, i.e., an adap-
tation of the tour plan by the specified metaheuristic is directly performed whenever
a new customer request arrives. Vehicle en-route diversion is not allowed so that a
vehicle can receive new orders after servicing the currently first assigned request.
Computational results based on the Solomon instances are assessed under the con-
sidered objective function which pursues the minimization of the number of rejected
customers, traveled distance, as well as lateness occurring when servicing customer
requests. The computational results show that the implemented evolutionary strate-
gies perform best among the tested metaheuristics. As an outlook on future research,
the author proposes the combination and parallelization of different metaheuristics.
Richter (2005) examines the adaptation of classic construction heuristics to the
Dynamic Traveling Salesman Problem (DTSP) as well as to the DVRP in their ca-
pacitated versions. The author argues that his main reasons for adapting construc-
tion heuristics to the dynamic case are short required computational times, which
are advantageous in real-time environments, as well as an easy understanding of
how these heuristics operate. In his work, the author also assesses the effects on the
solution quality if vehicle en-route diversion is allowed. In an extensive computa-
tional study, scenarios with 50 to 200 customer requests as well as with a degree
of dynamism between 20 % and 80 % are evaluated using two different objective
functions, specifically the minimization of travel distance and the minimization of
customer waiting times, i.e., response times. Richter concludes that the utilization of
vehicle en-route diversion allows a significant improvement of the solution quality
if the minimization of customer waiting time is pursued.
Bock (2010) considers a multi-modal transportation problem. In this complex
transportation problem, a large variety of dynamic events are considered. These
types of dynamism include the arrival of new requests, breakdowns and slowdowns
of vehicles, and traffic jams. The objective is to minimize the overall operative trans-
portation costs. In order to solve the problem, a real-time control concept which re-
gards the concurrency of tour plan adaptation and tour plan execution is proposed.
4.3 Strategies for Increasing Flexibility in Dynamic Routing Problems 97

In this real-time control approach, a specifically designed improvement heuristic

is applied that makes use of different neighborhood operators which are selected
depending on the quality of solutions found in previous iterations. Computational
results show that the application of the proposed real-time control approach, which
utilizes the complete execution time of the transportation process for adapting the
tour plan, yields the best results among the evaluated real-time control approaches.
After introducing the first papers on dynamic routing problems in the literature as
well as selected research work dealing with reactive real-time control approaches,
the literature review continues with describing basic strategies which are used in
dynamic routing approaches in order to increase the flexibility in the transportation
process. Afterwards, publications in the literature that make use of these strategies
are presented.

4.3 Strategies for Increasing Flexibility in Dynamic Routing

Problems
In this section, we illustrate basic strategies that are used in order to increase the
flexibility which is available during the execution of the transportation process.
Specifically, with regard to the considered RDOPG applications, flexibility accord-
ing to possibilities of integrating dynamically arriving requests is considered. As
mentioned in Sect. 2.5.3.2, an increased flexibility is advantageous when dealing
with dynamic events that occur during the transportation process. These strategies
can be distinguished into waiting strategies, relocation strategies, and request as-
signment strategies. The successive description of approaches that can be found in
the literature provides an explanation about which of these strategies are used in the
individual approaches.

4.3.1 Waiting Strategies

In static VRPs, the common procedure is that a vehicle directly moves to its next
assigned request after it has finished servicing its current request. If time windows
are used in an approach and a vehicle arrives at a request location prior to the start of
the assigned time window, it is often allowed to wait at this request location until the
time window opens. In contrast, in dynamic routing problems in which the objective
of travel distance minimization is pursued, waiting strategies can be advantageously
used to keep vehicles at their last serviced request location for some additional time
after service of the request has been finished. Therefore, waiting strategies can be
especially beneficial in applications in which time windows of requests open in the
far future in relation to the arrival times of requests in the system. In such cases, a
time period denoted as slack time occurs during which a vehicle would need to wait
anyway, for example at the last serviced request location or at the next scheduled
request location.
98 4 Review of the Literature Related to the Considered RDOPG Applications

Waiting at the next request location Waiting at the last serviced request location
A B C D

3 3 3 3

2 2 2 2

4 4 4 4

1 1 1 1

Planned vehicle tour Currently known request

Traveled vehicle tour Future request, currently not known

Fig. 4.1 Execution of a transportation process using different waiting strategies

In Fig. 4.1, the effects of waiting at the last request and at the next request loca-
tion are illustrated. Specifically, when a vehicle directly travels to the next request
location defined by request 2 and waits there (Picture A) and a new request (de-
noted by request 4) arrives, the vehicle has to perform a long additional travel dis-
tance (Picture B). In contrast, by waiting at the last serviced request location defined
by request 1 (Picture C), the decision concerning which request should be serviced
next can be postponed to a later time point. If a new request at a location close
to the vehicle’s last serviced request location arrives as the vehicle is still waiting
there, this new request only requires a small amount of additional travel distance to
be serviced (Picture D). Hence, new requests which arrive near the current vehicle
waiting position can be efficiently serviced with regard to the considered objective
of minimizing the total travel distance. This is especially beneficial in applications
in which no vehicle en-route diversion is allowed since, in this case, no re-routing
of the vehicle is allowed when request 4 arrives even if the vehicle has just left re-
quest 1 in order to travel to request 2. Note that if waiting strategies are used, the
decision space of the routing problem is extended since it has to be decided how
long a vehicle should wait at a specific location. Hence, the VRP is extended to a
Vehicle Routing and Scheduling Problem (VRSP, see also Sect. 2.4.6).
As illustrated above, waiting strategies can be used even when no stochastic
knowledge is available. If stochastic knowledge is available, waiting strategies can
additionally improve the flexibility by deciding where and how long vehicles should
wait according to the stochastic knowledge about expected request arrivals in indi-
vidual subregions in the near future.
Note that although it is likely that applying waiting strategies will have posi-
tive effects on the solution quality in many cases, there are also situations in which
waiting strategies lead to a worse solution quality. For instance, consider a situation
where a large number of new requests with tight hard time windows occurs close to
a customer request which is to be serviced next in a vehicle’s tour. If only a limited
number of vehicles which are already highly utilized is available and vehicles are
ordered to wait at their last request location, this may lead to a larger number of un-
serviced customer requests since they cannot be serviced within their time windows.
Furthermore, waiting strategies can also lead to a worse system performance in spe-
cific applications. For example, in the considered RDOPG applications in which
4.3 Strategies for Increasing Flexibility in Dynamic Routing Problems 99

Without vehicle relocation With vehicle relocation

A B C D

Traveled vehicle tour Currently known request

Relocation position Future request, currently not known

Fig. 4.2 Execution of a transportation process with and without using a relocation strategy

dynamically arriving requests can be directly serviced upon arrival and the objec-
tive is to minimize customer inconvenience, it should be carefully decided whether a
vehicle should wait when it has pending requests. Specifically, this decision should
be taken according to the probability that new requests will arrive soon and in close
vicinity of the considered vehicle. However, if this probability is not taken into ac-
count so that the vehicle waits in any case, the customer inconvenience of pending
requests will further increase. Therefore, approaches which directly service pending
requests may achieve a better solution quality in this case.

4.3.2 Relocation Strategies

In dynamic routing problems in which quality of service aspects are pursued, relo-
cation strategies enable potential solution improvements by actively repositioning
vehicles to strategic locations where future requests are expected to arrive. Reloca-
tion strategies were first assessed by Powell (1988). By using relocation strategies,
new arriving requests in close vicinity of the strategic waiting positions can be ser-
viced faster. Relocation activities can be beneficial in scenarios where a quick re-
sponse to new arriving requests is mandatory and significantly more important than
the total travel distance. For example, this is true for ambulance relocation prob-
lems (cf., e.g., Gendreau et al. 2001, Haghani and Yang 2007, and Schmid 2012)
and, as mentioned above, in dynamic routing problems in which the solution quality
is mainly determined by customer-related quality aspects. Specifically, such aspects
may comprise inconvenience aspects so that a fast reaction to newly arriving re-
quests is pursued (as considered in Larsen 2000 and the RDOPG applications in
this book) or aspects that consider the maximization of the feasibly serviced cus-
tomers within hard time windows (cf., e.g., Bent and van Hentenryck 2007). The
utilized strategic relocation positions may be constant or dynamically determined
during the day and can be the same or individual for the vehicles.
In Fig. 4.2, the difference between letting an idle vehicle wait at its last serviced
request location and relocating it to a strategic location is illustrated in an application
100 4 Review of the Literature Related to the Considered RDOPG Applications

in which the minimization of customer inconvenience which depends on request

response time is pursued. If the vehicle waits at the location of its last serviced
request (Picture A) and a new request arrives, the vehicle may need to travel for a
much longer time before it can reach and start servicing the new customer request if
the new request is far away from the last serviced request location (Picture B). This
may result in a long response time and therefore a high customer inconvenience. In
contrast, if relocation strategies are used, an idle vehicle is relocated to a strategic
waiting position in whose vicinity new requests are expected to arrive in the near
future (Picture C). If a new request arrives close to this waiting position, the vehicle
requires less travel time so that it can service the new request much faster (Picture D)
which results in significantly reduced customer inconvenience. Clearly, the use of
relocation strategies can only be utilized if stochastic knowledge about expected
future request arrivals is available. Approaches that make use of relocation strategies
are presented in Sect. 4.5.
The previously introduced waiting strategies are strictly connected to relocation
strategies. Specifically, when vehicles are relocated to certain positions, it needs
to be decided how long each vehicle should wait at this relocated position under
different circumstances, e.g., while the vehicle is idle or if newly arrived requests
have been assigned to the vehicle. This decision also depends on the urgency of the
new request and the travel time required from the current vehicle’s position to the
location of the new request.

4.3.3 Request Assignment Strategies

The above described waiting and relocation strategies are mainly applied to vehi-
cles which are currently not ordered to directly service any requests, i.e., vehicles
which can be seen as being currently idle. In contrast, request assignment strategies
are utilized to increase flexibility in the transportation process for busy vehicles by
directly influencing the assignment and sequencing of requests on vehicles, i.e., by
actively impacting the tour plan generation and adaptation process.
If no stochastic knowledge is available, request assignment strategies can in-
crease the flexibility in tour plans by restricting the detailed planning process to
short-term decisions since it is assumed that decisions in the long-term will be
re-adapted due to other dynamically arriving requests (cf., e.g., Mitrovic-Minic
et al. 2004). Furthermore, simultaneously storing parts of different valid tour
plans (cf. Gendreau et al. 1999a) or simultaneously storing various complete tour
plans (cf. Bent and van Hentenryck 2004b) makes it possible to make quick decision
about good request assignments. For example, if hard time windows are present, this
enables fast determination about whether or not new pending requests can be feasi-
bly integrated into the currently executed tour plan.
If stochastic knowledge is available, request assignments to vehicles can be per-
formed based on estimated objective function value outcomes. By considering ex-
pected request arrivals in the tour plan generation process, more flexible tour plans
4.4 Flexibility in Dynamic Routing Without Stochastic Knowledge 101

can be created. Specifically, it is likely to be advantageous if the assignment and

sequencing of requests to vehicles is performed under consideration of the resulting
vehicles’ positions and the estimated response times concerning expected future re-
quests. For example, if available stochastic knowledge reveals that request locations
are uniformly distributed, requests can be assigned to vehicles in a way that uni-
formly scatters the vehicles over the considered service area. In this case, it is likely
that expected future requests are serviced within a shorter response time. Moreover,
vehicle tours can be generated in such a way that additional time is reserved in
specific subregions at certain time intervals. This additional time can be used for
integrating expected future requests in the vehicles’ tours without requiring already
planned requests to be rescheduled. Similar to the above strategies, this strategy
is beneficial if the objective pursued considers quality of service aspects such as
in RDOPG applications. As a result, by using stochastic knowledge directly in the
tour plan generation process, future positions of vehicles can be advantageously in-
fluenced. Hence, effects that are similar to the ones achieved if idle vehicles are
relocated can also be achieved for vehicles which are currently busy.
In the following sections, recent publications in the literature which consider
methods for increasing tour plan flexibility in order to allow a better integration of
expected future requests are presented. In some applications, no stochastic knowl-
edge can be utilized either because stochastic knowledge with sufficient forecast-
ing quality is not available or because technical or other restrictions prohibit the
use of available stochastic knowledge. Based on this distinction, research work on
approaches which do not use stochastic knowledge is presented. Afterwards, ap-
proaches which use stochastic knowledge are described.
Unless stated otherwise, all approaches consider applications that service re-
quests by visiting them in a restricted Euclidean plane with Euclidean distances
and travel paths between request locations. Furthermore, the single source of dy-
namism is, by default, represented by newly arriving requests during the execution
of the transportation process. The considered planning horizon usually has a length
of one working day and vehicle en-route diversion is not allowed. Moreover, the
considered problems generally have one depot from which the vehicles start their
tour and to which they return after completion of their assigned tour.

4.4 Flexibility in Dynamic Routing Without Stochastic

Knowledge

In this section, we present approaches in the literature which provide an increased

flexibility in dynamic routing problems without using stochastic knowledge. Ac-
cording to our classification in Sect. 2.5.3.2.2, we denote them as deterministic ap-
proaches.
In the work of Kilby et al. (1998), the authors consider a DVRP where dynamism
arises from newly arriving requests. The considered routing problem is heuristically
solved and a specific waiting strategy is applied, that is, if a vehicle has currently no
102 4 Review of the Literature Related to the Considered RDOPG Applications

further requests assigned, the vehicle waits at its last serviced request location for
further requests which are assigned to its tour. The objective consists of minimizing
the total travel distance regardless of the number of used vehicles. Computational
results are evaluated on modified test instances of Taillard, Christofides, and Fisher.
Besides varying degrees of dynamism, the authors evaluate different lengths of the
commit horizon, i.e., the time period for which unchangeable decisions are made
during the transportation process. The authors test their approaches on the DVRP
case in which goods are picked up at customer locations as well as on the case in
which goods are delivered from the depot to customer locations (see Sect. 2.3.2).
The authors state that in the static case, solving both variants results in the same
solution quality. In contrast, in the considered dynamic problem, the delivery vari-
ant is more costly than the collection variant. This can be explained by the fact that
a request which needs to be collected can be serviced by any vehicle with free ca-
pacity while the delivery of a request can only be performed by the vehicle which
is transporting the specific good. Moreover, this cost gap increases the higher the
degree of dynamism. With regard to the evaluated commit horizon, the authors state
that the use of longer commit horizons results in a worse solution quality. For fu-
ture research, the authors propose, among other research fields, investigating the
integration of past request information into the transportation process.
The work of Gendreau et al. (1999a) considers a Dynamic Vehicle Routing Prob-
lem with Soft Time Windows (DVRPSTW) with a limited fleet size. Requests may
be serviced late, i.e., after the end of the assigned time window, but at additional
penalty costs. The objective aims at minimizing the weighted sum of travel dis-
tance and linear lateness costs incurred by late serviced customers. The dynamic
problem is solved as a series of static problem instances consisting of the currently
pending requests. Each static problem instance is solved by a Parallel Tabu Search
metaheuristic that has been proposed by Badeau et al. (1997). In order to increase
flexibility in the transportation process, the authors utilize the concept of an adaptive
memory in which tours of different tour plans that have been previously found dur-
ing the search process are stored. The advantage of this variety of stored tours is that
when new requests arrive, the variety can be used to achieve a quick construction
of a new tour plan of high quality which integrates the new requests. Specifically, a
new tour plan is constructed by randomly selecting tours from the adaptive memory
while the probability that a tour is selected depends on its solution quality. Since
this typically leads to a tour plan which is constructed out of tours from different
solutions, this approach combines strengths of Tabu Search and GA metaheuristics.
Computational results on modified Solomon instances show that the proposed Paral-
lel Tabu Search metaheuristic produces superior results compared to other evaluated
solution methods. The authors state that it would be interesting to integrate stochas-
tic knowledge about future events such as the time-space distribution of requests into
the solution approach in order to improve decision making during the execution of
the transportation process.
The adaptive memory concept described in the previous approach is general-
ized in the Multiple Plan Approach (MPA) for a DVRPTW proposed by Bent and
van Hentenryck (2004b). Since this problem considers hard time windows, requests
4.4 Flexibility in Dynamic Routing Without Stochastic Knowledge 103

need to be serviced within given time windows. Otherwise, requests cannot be ser-
viced any more and are counted as unserviced customer requests. The objective is to
maximize the number of serviced requests (see also Sect. 2.4.12.6). The MPA gen-
erates and maintains several tour plans which are consistent with the current system
situation while deleting incompatible ones. Individual tour plans are generated by
a solution method for VRPs using Simulated Annealing and Large Neighborhood
Search which has been proposed by Bent and van Hentenryck (2004a). The authors
suggest the continuous application of the solution method in the dynamic problem
setting. Specifically, since decisions are only required at specific events such as at
the completion of the service of a request or at the arrival of a new request, the
intermediate time can be used for improving tour plans stored in the MPA. This is
advantageous since it allows a quick reaction based on high-quality tour plans if
fast decisions are required. Besides the stored tour plans in the MPA, one additional
tour plan denoted as the distinguished plan which is executed in the system is main-
tained and incrementally generated during the transportation process. Specifically,
whenever a vehicle has completed the service of its currently assigned request, the
distinguished plan is extended by selecting the next request to be serviced by the
respective vehicle. This request selection is made according to the stored tour plans
in the MPA by using a dedicated ranking function. In their work, the authors test two
ranking functions. The first ranking function called distance selects the next request
for a vehicle by the least travel distance criteria which is used in the corresponding
approach variant called MPAd . The second ranking function is called consensus and
selects the next request for a vehicle based on similarity aspects. Specifically, for
vehicle v, it is evaluated how often request r is assigned to vehicle v as the next
request in the tour plans stored in the MPA. After evaluating all requests, the re-
quest which is assigned most is selected as the next request for vehicle v. The MPA
approach variant using the consensus ranking function is called MPAc . As can be
seen, both MPA variants iteratively determine the tour for each vehicle. Specifically,
on the basis of the corresponding ranking function the next request to be serviced by
the vehicle is selected whenever a vehicle has completed its current activity. Com-
putational results show that MPAc significantly outperforms MPAd with regard to
the number of requests that can be feasibly serviced within their time windows. Fur-
thermore, the authors extend the MPA to Multiple Scenario Approaches which use
stochastic knowledge. These approaches are described in Sect. 4.5.5.
Fleischmann et al. (2004b) study an urban single-load Dynamic Pickup and De-
livery Problem with Soft Time Windows (DPDPSTW) in which dynamism arises
from newly incoming requests and changes in travel times. Since no requests can
be rejected, all requests have to be serviced. Requests can be serviced also after the
end of their time window but at high penalty costs. The objective pursues primar-
ily the minimization of request delays and, secondarily, operational transportation
costs. Since the transportation process is carried out on a real road network, short-
est path computations are required when travel time information changes during the
execution of the transportation process. Updates in the travel time information are
provided by a traffic management center and shortest paths are calculated using a
modified version of Dijkstra’s algorithm. If appropriate, vehicles are allowed to wait
104 4 Review of the Literature Related to the Considered RDOPG Applications

Drive-First strategy (DF) Wait-First strategy (WF)

A B C D

3 3 3 3

2 2 2 2

4 4 4 4

1 1 1 1

Planned vehicle tour Currently known request

Traveled vehicle tour Future request, currently not known

Fig. 4.3 Execution of a transportation process using Drive-First and Wait-First strategy

at their last serviced request location and service of requests may be postponed.
Hence, a waiting strategy is used in this approach. For assigning the next request
to a vehicle when it has finished its current activity, the authors test three different
strategies: assignment rules, an assignment algorithm, and an insertion heuristic in
various settings. The assignment rules determine the next activity whenever a vehi-
cle has finished the completion of its current activity. Here, only the corresponding
vehicle is considered and the decision with the least cost is chosen. In case of the
assignment algorithm, an assignment problem is solved, i.e., for each vehicle the
next request is determined. Since no vehicle en-route diversion is allowed, this as-
signment determines the next request for each vehicle after the respective vehicle
has completed its current activity. The third tested strategy based on an insertion
heuristic solves the dynamic problem by maintaining a complete tour plan so that
all pending requests are always assigned to vehicles’ tours. The authors state that
test results on instances with 500 requests reveal that the assignment algorithm per-
forms best among the tested strategies. Specifically, higher degrees of dynamism
lead to a bigger improvement compared to the other evaluated strategies. The au-
thors direct future research possibilities to the evaluation of real-time methods on
the multiple-load DPDP.
Mitrovic-Minic and Laporte (2004) evaluate four different waiting strategies for
a DPDP with hard time windows and multiple vehicle starting locations. Moreover,
the available fleet size is not restricted. The pursued objective considers the mini-
mization of total travel distances. The first tested waiting strategy is Drive-First (DF)
which is equal to a policy in which vehicles directly travel to the next request after
service of the current request is completed. Since no vehicle en-route diversion is
allowed, this strategy directly commits vehicles to the next assigned requests. There-
fore, requests which arrive after departure and before service of the next assigned
request begins may lead to an additional detour when further requests arrive (see
Pictures A and B in Fig. 4.3).
The second strategy, Wait-First (WF), forces vehicles to wait at their current loca-
tion up to the point in time at which a departure at the last serviced request location
leads to an arrival at the next scheduled request location directly at the end of its
time window, i.e., at the latest allowed time point. This strategy provides the biggest
4.4 Flexibility in Dynamic Routing Without Stochastic Knowledge 105

flexibility in terms of integrating new arriving requests and leads to the best results
in terms of total travel distance (see Pictures C and D in Fig. 4.3). However, due to
the resulting tight schedule the authors state that it happens often that new arriving
requests require the use of additional vehicles. Because of this, the authors provide
two further strategies called Dynamic Waiting (DW) and Advanced Dynamic Wait-
ing (ADW). For these strategies, the authors introduce a dynamic partitioning of the
pending requests into service zones. Here, each service zone contains a number of
consecutive request locations that are close to each other. What the waiting strate-
gies DW and ADW have in common is that for traveling between requests within
the same service zone, the DF strategy is applied. At the end of a service zone, DW
applies the WF strategy whereas ADW uses only a certain part of the maximum
available waiting time. The authors define this part as the ratio of the service zone
time span to the sum of time spans of all service zones that are assigned to the tour
of a vehicle. The authors test their strategies with instances of between 100 and
1,000 requests. Computational results show the superiority of the waiting strategy
ADW with regard to total distance and number of required vehicles, especially if an
elaborated Tabu Search solution method is applied for adapting the tour plan.
In their subsequent work, Mitrovic-Minic et al. (2004) also consider a DPDP with
hard time windows. An unrestricted fleet size is assumed since, in the considered
real-world problem, a very large fleet size is available and fixed costs for vehicle
deployment are paid by the drivers. The objective consists of minimizing the total
travel distance. The authors propose a concept called double horizon that generalizes
short-term rolling horizon approaches. This generalization is applied in order to
prevent unwanted long-term effects on the tour plan execution process if only good
short-term decisions are made. The double horizon concept is implemented by using
a different objective function for the short-term horizon and the long-term horizon.
Specifically, for the short-term horizon, the objective used is identical with the one
that is pursued in the problem while the objective of the long-term horizon is to
maintain tours in such conditions which facilitate the integration of future arriving
requests. This is achieved by selecting tour plans which exhibit large slack times
in the distant future in their tours. Both horizons are optimized using a simplified
version of the Tabu Search metaheuristic proposed by Gendreau et al. (1998) (later
published in Gendreau et al. 2006). The authors apply some of the waiting strategies
described in Mitrovic-Minic and Laporte (2004). Computational results are carried
out on 100 and 500 customer request instances generated using real-world data from
two courier service companies in Canada. The results show the potential that can be
achieved by applying the double horizon approach compared to classic single rolling
horizon approaches. Furthermore, the authors introduce the definition of the value
of information that assesses the system performance of a given solution approach in
the dynamic case compared to the static case (see also Sect. 2.6.6.2).
Following the concept of waiting strategies, de Oliveira et al. (2008) propose an
approach for a capacitated DVRPTW considering the objective of minimizing travel
distance. The authors evaluate four different metaheuristics: a Simulated Annealing
algorithm, a Greedy Randomized Adaptive Search Procedure heuristic, a Memetic
Algorithm, and a hill climber heuristic. In order to increase flexibility, the authors
106 4 Review of the Literature Related to the Considered RDOPG Applications

integrate a vehicular waiting time heuristic (VWTH). This heuristic is used to deter-
mine appropriate waiting times for vehicles at serviced request locations. As in the
previously described approaches, the purpose of the VWTH is to enable more pos-
sibilities for vehicles to integrate future arriving requests by letting them wait for an
appropriate amount of time. Consistent with the results of Mitrovic-Minic and La-
porte (2004), the computational results show that using the VWTH can significantly
improve the solution quality.
Another concept denoted as request buffering was introduced to dynamic routing
problems by Pureza and Laporte (2008). The authors study an uncapacitated DPDP
with hard time windows. In order to prioritize the quality of service, a hierarchi-
cal objective function is applied. Specifically, the main objective is to minimize the
number of unserviced requests followed by the number of used vehicles and finally
the total travel distance. Dynamism arises from newly arriving requests as well as
from random travel times. Furthermore, time-dependent travel times are present.
Since uncertain travel times may lead to longer travel activities, the authors adapt
the WF strategy proposed by Mitrovic-Minic and Laporte (2004) to a new strategy
called WE. This new strategy schedules vehicle departures in a way that a vehicle
arrives at each customer request location at the start of its time window but not ear-
lier. Since travel activities are carried out on the Euclidean plane, traveling from
location i to location j is always assumed to be performed by using the direct Eu-
clidean connection between both locations defined by the arc (i, j ). Furthermore,
since time-dependent travel times are modeled by modifying the travel time of an
entire arc, the triangle inequality is not valid in this problem. Therefore, the authors
extend the WE strategy by integrating travel time information based on fastest paths
between locations. Specifically, this concept tests whether the destination can be
reached faster by traveling via an intermediate location instead of using the direct
arc between the current location and the destination. In their approach, the authors
construct fastest paths with at most one intermediate location. This time can be used
as additional slack time that increases flexibility for integrating further requests. Ad-
ditionally, another strategy called Request buffering (RB) is proposed. This strategy
postpones the insertion of newly arrived but less urgent requests to a later time.
Since no vehicle en-route diversion is allowed in the considered application, de-
cisions become unchangeable once taken. Hence, this deferment of requests leads
to situations in which a larger number of pending requests, on whose basis better
routing and scheduling decisions can be taken, is present. This is in line with the ob-
servations by Larsen regarding the effective degree of dynamism (see Sect. 2.6.4).
The authors evaluate instances with 100, 200, and 400 requests with varying degrees
of dynamism. Computational results show that using a combination of fastest paths
and the RB strategy yields the best results among the tested strategies.
In the work of van de Klundert and Wormer (2010), the authors consider a real-
world application, which is called the After Salesman Problem (ASAP). This prob-
lem arises from after-sales service demands of customers and has similarities to the
considered RDOPG applications. A real-world case study is conducted at the largest
service organization in the Netherlands in which customer requests are required to
be serviced within a maximum response time which is defined by SLA contracts.
4.4 Flexibility in Dynamic Routing Without Stochastic Knowledge 107

Besides dynamically arriving customer requests, uncertainty arises from uncertain

service time durations which are considered to be given by stochastic variables only.
In this problem, a service area with Euclidean distances is considered. In the eval-
uations, the main pursued objective consists of maximizing responsiveness by min-
imizing the average waiting time of the dynamically arriving customers which is
comparable to applications in which the minimization of a linear customer incon-
venience function is pursued. In order to solve the problem, the authors propose
three solution approaches. The first is based on a periodically performed multiple-
vehicle First-come, first-served (FCFS) strategy. The second approach called match-
ing periodically selects the next request for idle vehicles. It is modeled as a linear
assignment problem which is optimally solved by the Hungarian Method in polyno-
mial time. The third approach called set partitioning periodically assigns all requests
to vehicle tours forming a complete tour plan. In order to solve this set partitioning
problem, an approach similar to the one proposed by Krumke et al. (2002) is applied
(see also Sect. 4.6). In the computational results, the authors show that in instances
with high request arrival rates, the FCFS strategy performs worst since this strategy
does not take advantage of the possibility to service requests which have arrived
later but are close to a vehicle. Out of the tested approaches, the set partitioning
strategy performs best. In another computational study, the authors assess the im-
pact of the interval length after which the reoptimization, i.e., the adaptation of the
tour plan, is performed. According to their results, in light traffic conditions a more
frequent reoptimization leads to small improvements while in heavy traffic condi-
tions, i.e., if high request arrival rates are present, longer intervals result in small
improvements since better decisions can be taken when more pending requests are
available. However, this may trigger additional costs due to a postponement of the
service of waiting customers, thus inconveniencing them more. Furthermore, the
authors consider that solving individual problem instances during the execution of
the transportation process to optimality does not guarantee that the routing problem
considered during the whole planning horizon, denoted as the end-of-day objec-
tive, is also optimally solved. Therefore, the impact of using a different objective
function—for solving the individual problem instances—on the solution quality is
examined. For instance, the authors show that if the aim is to maintain a defined SLA
for the customers by servicing customers within a maximum allowed response time,
it is beneficial to integrate information about the request response times into the ob-
jective function in order to allow the solution method for discriminating between the
quality of different solutions in which all requests fulfill the SLA requirements (see
also Sect. 2.4.12.6). Moreover, they suggest to use a lower time limit for the SLA in
the problem instances than required in the end-of-day objective in order to maintain
some buffer time for exceeding this limit in the dynamic transportation process and
still maintaining the demanded end-of-day SLA. Furthermore, effects of allowing
vehicle en-route diversion are assessed. Computational results show that in heavy
traffic scenarios, allowing vehicle en-route diversion yields an improvement of over
20 % of the solution quality. Moreover, in another study in which the effects of us-
ing additionally available knowledge about the expected request service times are
assessed, results indicate that this knowledge can yield significant improvements
108 4 Review of the Literature Related to the Considered RDOPG Applications

on the solution quality. For future research, the integration of stochastic knowledge
in the considered application is suggested. Note that this research question is con-
ducted in the RDOPG applications considered in this book.

4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge

In this section, we focus on the presentation of approaches that use stochastic knowl-
edge, hence increasing flexibility in the transportation process and providing possi-
bilities to improve the solution quality. Since dynamism in the considered RDOPG
applications arises from newly arriving requests, in what follows, we focus on ap-
proaches which use stochastic knowledge in order to expect future request arrivals.
According to Ichoua et al. (2007), the use of stochastic knowledge in approaches
which consider dynamically arriving requests is motivated by knowledge that a hu-
man dispatcher has. Specifically, having attained valuable knowledge from experi-
ence, dispatchers typically have knowledge about patterns regarding arrival times of
new requests and their location. This allows them to make predictions about when
and where new requests are likely to arrive during the day and allows them to man-
age the vehicles more efficiently in anticipation of these future arriving requests.
Therefore, if stochastic knowledge is available and utilized in real-time control ap-
proaches, additional significant improvements are often attainable when compared
to the previously illustrated methods for increasing flexibility without using stochas-
tic knowledge. According to our classification in Sect. 2.5.3.2.2, we denote such
approaches as pro-active approaches.
In the presented approaches, stochastic knowledge is either manually entered into
the system based on human experience or automatically forecasted by appropriate
approaches that analyze past request information. Bent and van Hentenryck (2004b)
mention several articles in the literature in which authors suggest the utilization of
stochastic knowledge in order to anticipate future requests. This stochastic knowl-
edge can either be given by analyzing past request information using appropriate
methods (as suggested in Kilby et al. 1998 and Gendreau et al. 1999a, 2006) or by
prescribed probabilistic distributions that have been entered into the system in some
other way (cf., e.g., Gendreau et al. 1996 and Yang et al. 2000). Furthermore, in the
previously described thesis of Richter (2005) and in the work of Bock (2010), the
integration of stochastic knowledge is also suggested.
According to Bianchi (2000), approaches for dynamic routing problems can be
differentiated by the way the dynamic problem is divided into individual problem
instances which are solved by a defined solution method during the execution of the
transportation process. Specifically, strategy-oriented approaches evaluate different
ways of dividing the dynamic problem while the solution method utilized for solv-
ing these individual problem instances is often assumed to be available and is not
analyzed in detail. Specifically, it is often assumed that a solution approach which
optimally solves the given individual problem instance is available. This type of
studies often analytically derives lower bounds and other properties of specific so-
lution strategies for dynamic routing problems. However, due to the computational
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 109

complexity, not all types of dynamic routing problems can be analytically evaluated.
In contrast, heuristic-oriented approaches utilize a defined strategy for dividing the
dynamic problem into subproblems and evaluate in detail the system performance
of one or more solution methods applied to solve the individual problem instances.
Studies which develop such approaches often empirically provide computational
results on given problem instances for complex dynamic routing problems.
This section starts by presenting the strategy-oriented approaches. In the fol-
lowing subsections, heuristic-oriented approaches which make use of stochastic
knowledge are presented. We refer to this type of approaches as real-time con-
trol approaches which reoptimize individual problem instances of the transporta-
tion process during its execution. First, approaches using exact solution methods
are illustrated. This is followed by the presentation of stochastic modeling based
approaches (cf. Pillac et al. 2010). These approaches often use multi-stage models
in which the stochastic knowledge is directly incorporated which results in models
which are often very hard to solve to optimality. Due to complexity reasons, a large
part of approaches deals with heuristically inspired solution methods. Accordingly,
heuristic solution approaches with manually provided stochastic knowledge. After
that, sampling-based approaches which reduce complexity issues by generating sce-
narios through sampling probability distributions are described.

4.5.1 Strategy-Oriented Approaches

As mentioned above, strategy-oriented approaches utilize available stochastic

knowledge in one or more strategies in order to assess the impact of each strategy
on the solution quality (cf. Bianchi 2000). Specifically, such approaches perform
tour plan adaptations according to a specific strategy (or rule). Strategy-oriented
evaluations for dynamic routing problems using stochastic knowledge were first as-
sessed by Bertsimas and van Ryzin. Their studies are mainly based on the Dynamic
Traveling Repairman Problem (DTRP) which is introduced in Bertsimas and van
Ryzin (1991). In its simplest form, the DTRP is related to the DTSP. The DTSP was
introduced by Psaraftis (1988) and is defined as follows. In a considered service
area, a set of nodes is given forming a complete graph. New requests are indepen-
dently generated at each node of the graph according to a Poisson process with a
known intensity parameter. The requests have to be serviced by a single salesman
which requires a deterministically known travel time on the arcs between the nodes.
At each node, a known service time is required for servicing the request. The aim
is to minimize a given objective, for example, the sum of customer request waiting
times.
In the DTRP, a single repairman (also denoted as vehicle in what follows) travels
in the Euclidean plane with a constant velocity and needs to service dynamically
arriving customer requests. In contrast to the DTSP, requests in the DTRP can oc-
cur at an arbitrary location within the considered service area. None of the requests
110 4 Review of the Literature Related to the Considered RDOPG Applications

are known in advance so that the problem has the highest possible degree of dy-
namism (see Sect. 2.6.4). As the name of the problem suggests, the DTRP repre-
sents a problem in which time-critical orders arrive dynamically and are required to
be serviced as quickly as possible. According to the authors, this problem belongs
to the set of real-world applications in which the minimization of customer waiting
times is more important than operative travel costs. Therefore, the objective is to
minimize the average expected system time of requests which consists of the ex-
pected waiting time, i.e., the request response time, and the service time required at
the request. Note that due to the considered objective function, the DTSP and the
DTRP are related to the RDOPG applications considered in this book if a linear cus-
tomer inconvenience function, depending on request response times, is utilized (see
also Sect. 3.1).
In the first studied DTRP variant in Bertsimas and van Ryzin (1991), dynamic
requests appear according to a Poisson process with a given intensity parameter λ.
Stochastic knowledge is given since it is known that request arrivals follow the de-
scribed Poisson process and the additionally given information that newly arriving
requests are independently and uniformly distributed in the considered Euclidean
plane. Furthermore, service times of requests are independently and identically
distributed according to a known distribution. Hence, in this approach, stochastic
knowledge is given by prescribed probability distributions. In their study, the authors
derive system characteristics of the described problem which include lower bounds
on the performance by using mathematical methods. Specifically, they present a
model formulation of the problem that allows the combination of approaches from
different areas such as queuing theory, geometrical probability, combinatorial opti-
mization, and simulation. In order to apply results from queuing theory, the authors
model the problem as a stochastic queue median problem (SQMP). According to
Kendall’s notation (cf. Kendall 1953 and Tijms 2003, p. 341 et seq.), this problem
is represented by the notation M/G/1 since the requests that arrive according to a
Poisson process have exponentially distributed inter-arrival times and one vehicle is
available for servicing the pending requests. The aim of the study is to determine
the best strategy for the repairman among a set of tested strategies under light and
heavy traffic conditions, i.e., low and high request arrival rates. The authors test sev-
eral strategies for sequencing the pending requests on the vehicle. In the First-come,
first-served (FCFS) policy, all requests are serviced in their sequence of arrival.
When the vehicle is idle, it is relocated to the median of the service region where
it waits for further requests. Therefore, a relocation and a waiting strategy is used.
The stochastic queue median policy (FCFS-SQM policy) behaves similar to FCFS
but always repositions the vehicle at the median after it has finished service of the
current request before it continues to service other waiting requests. The Partition-
ing policy (PART) partitions the service region into m2 subregions in which arriving
requests are serviced in an FCFS manner. If no requests are pending in the subregion
which is currently serviced by the vehicle, the vehicle continues to service requests
in the next adjacent subregion. In this strategy, the sequence in which subregions
are considered is predetermined. In another strategy denoted as the TSP strategy,
requests are queued and grouped into sets of a predetermined size n. When a new
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 111

set of n requests has been generated, an optimal TSP solution for the requests in
this set is calculated. Afterwards, the requests are serviced according to this optimal
solution. If more than one set is pending at the same time, the sets are considered
in FCFS order. The Space Filling Curve (SFC) policy services requests ordered by
their arrival in repeated clockwise sweeps of a circle that covers the Euclidean ser-
vice region. In the Nearest Neighbor (NN) policy, the request to be serviced next is
determined by choosing the one closest to the vehicle’s current position.
The authors evaluate the performance of the given strategies in light and heavy
traffic scenarios by providing appropriate lower bounds. They show that the FCFS-
SQM policy is optimal in case of light traffic, i.e., in scenarios with low request
arrival rates. In scenarios with increased request arrival rates, the FCFS and FCFS-
SQM policy become unstable. In this case, the authors indicate that the PART policy
performs well. If a heavy traffic condition is present, the authors show that the TSP,
SFC, and NN strategies are the best policies. Specifically, the SFC strategy reduces
the average request time in the system by about 15 % when compared to the TSP
policy. Furthermore, the NN strategy reduces the average request system time by an
additional 10 % compared to the SFC strategy. On the other hand, the authors note
that the TSP and SFC policies tend to be fairer by considering the request arrival se-
quence more than the NN strategy and they additionally offer provable performance
guarantees. Furthermore, the authors identify that the heavy traffic policies achieve
asymptotically the same behavior so that their performance distinguishes from each
other only by a constant factor. For future research, the authors suggest extending
the problem to the multiple-vehicle case as well as considering capacitated problem
variants. Furthermore, other policies could also be investigated.
In Bertsimas and van Ryzin (1993a), the authors extend the previously consid-
ered problem to the multiple-vehicle case in which m vehicles are available for ser-
vicing requests in the Euclidean plane. They analyze two variants of the problem,
in particular the uncapacitated case and the capacitated case in which each vehicle
can only service q requests before it is required to intermediately travel to the de-
pot. In this work, the authors present strategies for the single-depot case but they
mention that the results can be extended to a multi-depot scenario if certain symme-
try conditions exist. When dividing the problem into m subregions, m independent
M/G/1 queueing systems (cf. Kendall 1953 and Tijms 2003, p. 341 et seq.) arise
to each of which the FCFS-SQM policy can be applied. Hence, the authors mod-
ify the FCFS-SQM policy of the single-vehicle case to the FCFS-mSQM policy in
which each of the m vehicles is positioned at one of the m median locations in the
considered service area. Similar to the FCFS-SQM policy, each vehicle returns after
completion of a request to its assigned median before it services the next request in
FCFS arrival order. As also mentioned in Bertsimas and van Ryzin (1991), in light
traffic scenarios, the authors prove that this policy is optimal. For heavy traffic sce-
narios, the authors propose different strategies and lower bounds. The randomized
assignment (RA) policy divides the Poisson request arrival process into m Pois-
son subprocesses by randomly assigning requests that arrive in the original Poisson
process to the m subprocesses. Each subprocess is assigned to one vehicle that ser-
vices delegated requests in the whole service area. The authors show that this strat-
egy provides a constant performance factor guarantee but that this factor increases
112 4 Review of the Literature Related to the Considered RDOPG Applications

with m. Another policy that possesses the same dependency of m is a G/G/m ver-
sion (cf. Kendall 1953 and Tijms 2003, p. 341 et seq.) of the TSP policy which is
an extension of the previously described TSP policy for the single-vehicle case. In
this policy, requests are queued in sets of size n. In an FCFS manner, the next set is
then assigned to the next available vehicle which services the requests of this set in
sequence of an optimal TSP tour. The authors also present two variants of the TSP
policy with a performance that is independent of the number of available vehicles.
In the first policy variant called the modified G/G/m policy, the service area is sub-
divided into k subregions where each subregion forms a wedge of area A/k with A
denoting the whole service area. In each subregion, arriving requests are queued
into sets of size n/k. Whenever a set is complete, the next available vehicle services
the requests assigned to this set, in sequence of an optimal TSP tour while pending
sets are queued in an FCFS manner. The authors state that this strategy can also be
used well in the single-vehicle case since it halves the request system time that the
best strategy evaluated in Bertsimas and van Ryzin (1991) requires. In the second
policy variant, the authors utilize the same subregions as in the previous policy but
arriving requests are queued in each subregion. In specified time intervals, all re-
quests which are queued in the subsequent region are considered in an optimal TSP
tour which is added to the queue of pending tours which are again processed in an
FCFS manner. Since in this policy a new request (consisting of a tour of requests)
arrives in the queue after a predefined amount of time, this forms a D/G/m system
(cf. Kendall 1953) which can be analyzed using the same lower bounds developed
for the previously described G/G/m systems. In another policy called the Pμ pol-
icy, the service area is partitioned into m subregions and one vehicle is assigned
to each of these subregions. Requests which arrive in each subregion are serviced
by the corresponding vehicle according to a chosen single-server policy. Hence, the
authors conclude that for the uncapacitated case, all heavy traffic policies with a
constant factor which were proposed for the single-vehicle case can be extended
to the multiple-vehicle case. Furthermore, for the capacitated case, strategies from
the uncapacitated case are adapted and new strategies are also proposed. Addition-
ally, the authors discuss possibilities to adapt the presented policies to a scenario
where the objective consists not only of minimizing waiting and service times, but
additionally comprises the minimization of travel costs. As an outlook on future re-
search, the authors suggest evaluating dynamic routing problems in road networks
instead of using the Euclidean plane.
In Bertsimas and van Ryzin (1993b), the authors extend their findings to more
generalized situations in which new requests arrive according to a general renewal
process and their locations are determined by an arbitrary distribution. Furthermore,
the authors state that static vehicle routing approaches can yield near-optimal poli-
cies for dynamic routing problems if they are properly adapted. Therefore, the au-
thors state that most of the heuristics and exact methods which have been developed
for static problems are not irrelevant for developing effective policies in dynamic-
stochastic routing problems.
Papastavrou (1996) introduces a new strategy for the single-vehicle case of the
DTRP called the generation policy. This policy positions the vehicle at the median
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 113

of the service region and services requests as soon as they are available. After ser-
vicing, the vehicle returns to the median. If no further requests are pending, the
vehicle waits and the strategy starts all over again. Otherwise, the currently pend-
ing requests are serviced in sequence of the corresponding optimal TSP tour. After
servicing all of the requests of the generated optimal tour, the vehicle returns to the
median and the generation policy is again applied. In contrast to the previously men-
tioned strategies, Papastavrou states that the generation policy performs well in light
as well as in heavy traffic conditions. The author emphasizes that a policy that works
well under all traffic conditions is advantageous for practical use since a change of
the applied policy can be costly and difficult to realize. This is especially important
in scenarios where the request arrival rate often changes. The author mentions that
the generation strategy can be easily adapted to the multiple-vehicle case by posi-
tioning m vehicles at the m medians of the service region if the request locations are
uniformly distributed or at the weighted m medians in routing problems in which a
general request distribution exists.
Swihart and Papastavrou (1999) extend the research on the DTRP to a single-
vehicle DPDP while considering the same objective as in the DTRP. In this research,
two cases are analyzed: The unit-capacity case where the vehicle directly needs to
travel from the pickup to the delivery location of the request before new requests
can be transported and the multiple-capacity case in which the vehicle can trans-
port more than one request at the same time. For the unit-capacity case, the authors
derive new bounds for the system performance in the light and heavy traffic case
and evaluate different policies that are comparable to the ones tested in the DTRP.
Like in the DTRP, in the light traffic case, the authors prove that the SQM policy
is optimal. In the heavy traffic case, different policies are tested. Specifically, in the
sectoring policy, the service area is partitioned into m2 sectors. When the vehicle
has finished servicing the current request, the request that has waited longest in the
current vehicle’s sector is serviced next. If no request is currently pending in this
sector, another sector is randomly chosen. Furthermore, the stacker crane policy is
a variation of the TSP policy in which each node of a pickup and delivery pair is
treated as a single location in an asymmetric TSP. This is possible since once a re-
quest has been picked up, the vehicle is required to directly travel to its delivery
location. Furthermore, a nearest neighbor policy is also evaluated. In order to as-
sess the performance of the strategies, the authors apply mathematical as well as
simulation-based methods since not all strategies are mathematically tractable due
to complexity reasons. Their evaluations show that while the sectoring policy has
a bad performance, the stacker crane policy as well as the nearest neighbor policy
perform better while the nearest neighbor policy performs best out of the tested poli-
cies. In the multiple-capacity case, the light traffic case is not analyzed any further
since it is very unlikely that in this case more than one request is pending at a time
point. For the heavy traffic case, two further strategies are invented, that is the Dual
TSP policy and the multiple-capacity nearest neighbor policy. Computational re-
sults show that the multiple-capacity nearest neighbor strategy performs better than
the Dual TSP policy. According to the authors, two further tested variants of the
Dual TSP policy, i.e., the Partially Sectored Dual TSP policy and the Sectored Dual
114 4 Review of the Literature Related to the Considered RDOPG Applications

TSP policy, yield the best results when only one sector is used so that both variants
become exactly the same as the original Dual TSP policy. For future research, the
authors suggest improving the proposed bounds and considering multiple-vehicle
problem variants.
In his PhD thesis, Larsen (2000) extends the definition of the degree of dy-
namism (see Sect. 2.6.4) which had originally been introduced by Lund et al. (1996).
Furthermore, the author gives detailed information about the generation of customer
requests and request-specific attributes such as appropriate distributions for gener-
ating arrival times, service times, locations, and demands. Moreover, he proposes
heuristic solution methods for three different dynamic routing problems. One of the
evaluations is performed on the DVRPTW under different degrees of dynamism
without using stochastic knowledge while the other two investigated dynamic rout-
ing problems use stochastic knowledge and are described in what follows. In the first
problem, the author extends the single-vehicle case of the DTRP to the The Partially
Dynamic Traveling Repairman Problem (PDTRP). In this problem, a given portion
of all requests to be serviced is known before the execution of the transportation
process. Furthermore, the considered objective function is different from the re-
search on the DTRP described so far since it pursues the minimization of travel
costs. During the execution of the transportation process, the dynamic requests ar-
rive according to a Poisson process whose properties are known and can be used
as stochastic knowledge in order to define appropriate strategies. Service times of
all customers are assumed to be distributed according to a log-normal distribution.
Moreover, vehicle en-route diversion is not allowed. The author evaluates the strate-
gies FCFS, FCFS-SQM, NN, and PART on instances with varying degrees of dy-
namism. The results of the conducted experiments show that the NN policy performs
best among the tested strategies. Utilizing the given objective, the FCFS-SQM pol-
icy performs poorly since it generates additional travel activities to the median of
the service area after each completion of service of a request. Furthermore, using
the NN strategy, the author states that travel costs linearly increase with the degree
of dynamism. The results of this study as well as the extended definition of the
degree of dynamism have also been published in Larsen et al. (2002). For future
work, the author proposes utilizing stochastic knowledge that provides information
about arrival rates of customer requests in specific parts of the service area as well
as information about changes in these rates during the day. For this purpose, the
author outlines an algorithm for integrating dummy customers based on historical
request data. In the second considered application which utilizes stochastic knowl-
edge, Larsen introduces the A Priori Dynamic Traveling Salesman Problem with
Time Windows (ADTSPTW) in which soft time windows are present. Similar to the
previously considered PDTRP, service times of requests are distributed according
to a log-normal distribution and vehicle en-route diversion is prohibited. The ob-
jective function consists of minimizing a weighted sum of travel costs and lateness
costs which occur if a request is serviced after the end of its time window. Lateness
costs depend linearly on the length of the lateness. Dynamically arriving requests
are generated by dividing the considered service area into subregions. In each sub-
region i, requests arrive according to a Poisson process with an individual intensity
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 115

parameter λi . These intensity parameters remain constant over the considered plan-
ning horizon and are assumed to be available as stochastic knowledge. In this study,
the effects of applying different strategies are examined if the available stochastic
knowledge is utilized. Specifically, this knowledge is used to choose an appropriate
waiting position for the vehicle if it is currently idle or if the time window of its
next assigned customer request only starts in the far future. Hence, the proposed
approach makes use of waiting and relocation strategies. Whenever new requests
arrive, a TSP is solved by a local search heuristic in which a given solution is im-
proved by exchanging two customer requests if this leads to an improvement of the
solution. The author mentions that since the applied heuristic utilizes a running time
of less than two seconds, concurrency considerations of adapting and executing the
tour can be neglected due to the shortness of the required time period. In the compu-
tational experiments, different strategies are assessed to select the waiting position
out of a set of predefined waiting locations where one idle point is defined per sub-
region i. Depending on the used strategy, the waiting point for the vehicle is: (a) the
nearest (NEAREST), (b) the one in the subregion i with the highest rate parame-
ter λi (BUSIEST), or (c) the waiting point of the subregion in which it is expected
that the most requests can be serviced within the current idle time of the vehicle,
i.e., the time period which is available before the vehicle departs from the selected
subregion, at the latest, in order to service other currently pending requests with
a time window opening far in the future (HI-REQ). Hence, the HI-REQ strategy
takes into account the distance from the chosen subregion to other customers that
are known but can only be serviced farther in the future. After the potential waiting
location has been selected according to the used strategy, the probability that at least
one more request will occur in the selected subregion within the current idle time
is determined. This probability is used to decide, according to a defined threshold,
whether the vehicle is relocated to this waiting point or not. Clearly, the higher this
threshold, the fewer the relocation activities that are performed. As a reference, the
author defines a strategy denoted as NO REPOS in which the vehicle always waits
at the last serviced customer request location so that no relocations are performed.
In the computational experiments, two variants of the objective function were con-
sidered. In the first variant, the minimization of travel costs is prioritized while in the
second variant, the minimization of lateness costs is the topmost aim. According to
first provided results, small improvements can be attained for reducing the lateness
while a higher threshold reduces the total travel distance due to fewer relocation
activities. The author states that more research on these policies is likely to result
in significant additional solution quality improvements. In Larsen et al. (2004), the
authors continue to study the previously described ADTSPTW. The authors provide
extensive computational results and identify recommendations about which policy
should be used depending on the objective and the degree of dynamism.
Irani et al. (2004) perform research on a variant of the DTRP in which a single
vehicle services dynamically arriving requests which require no service time at their
location. Each arriving request is required to be serviced until a given latest time
point. In their study, the authors use a uniform distribution for modeling the length
of the time period between request arrival and the latest allowed time point. The
116 4 Review of the Literature Related to the Considered RDOPG Applications

considered objective is to maximize the number of feasibly serviced requests up to

the latest allowed time point. The authors develop two strategies called BATCH and
Double-Gain. For both strategies, they present upper bounds of the competitive ratio
as functions depending on the diameter of the considered metric space. Furthermore,
several lower bounds for arbitrary deterministic strategies are presented. For future
work, the authors propose evaluating the Double-Gain strategy in scenarios where
the time period available for feasibly servicing requests is not uniformly distributed.
Branke et al. (2005) consider a DVRP in which exactly one additional customer
request dynamically arrives during the execution of the transportation process. Ve-
hicles have fixed tours for servicing the known requests. Requests have no time
windows nor service times but the vehicles need to arrive at the depot, at the latest,
at a given time point. The objective is to maximize the probability that the additional
occurring request can be feasibly integrated into the fixed tour plan, i.e., so that all
vehicles still feasibly arrive at the depot. For this purpose, the fixed tours comprise
additional slack time which can be used to wait for the additional request as well as
service it. Stochastic knowledge is available in form of prescribed probability dis-
tributions which describe the location and the arrival time of the additional request.
Regarding its location, it is known that the arriving customer request is uniformly
located in the considered service area. For the arrival time, two variants are con-
sidered. In the first variant, it is known that the request only arrives after a specific
time point whereas in the second variant, this time point is equal to the start time
of the transportation process so that the request can arrive at any point in time. In
both cases, the arrival time of the additional request is uniformly distributed after the
specified time point. In order to maximize the probability of feasibly integrating the
additional customer request, the stochastic knowledge is utilized to make vehicles
wait at suitable locations during their tours. Hence, in this problem, the positions of
the vehicles can be influenced by using a waiting strategy. The considered problem
is called the waiting drivers problem (WDP) and the authors show that both vari-
ants of the WDP are NP-hard. The authors evaluate different waiting strategies. In
the single-vehicle case, it is proven that the optimal strategy is not to wait at all. If
two vehicles are available, the authors show that waiting is beneficial and propose
an optimal strategy for this scenario. For the multiple-vehicle case with more than
two vehicles, due to the NP-hardness of the problem, six different heuristic waiting
strategies as well as two variants of an evolutionary algorithm which improves the
solution by examining different waiting time values are proposed. Computational
results for the multiple-vehicle case show that, compared to a strategy in which ve-
hicles do not wait at all, an appropriate waiting strategy can significantly increase
the probability of feasibly servicing the additional customer request as well as re-
duce the additional travel distance required for servicing the dynamically arriving
request. Specifically, one of the tested strategies called “variable” orders vehicles
to service requests without waiting until the point in time at which the remaining
travel time to the depot is equal to the slack time available in the tour. From this
time point, the available waiting time is proportionally distributed to the remaining
customers according to the remaining travel distances. By applying this strategy, the
authors state that the probability of being unable to feasibly service the additional
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 117

request is reduced by approximately 10 % while the average length of the additional

travel distance in case of a successful insertion of the new request can be reduced
by about 35 %. The authors also give insights into variants of the considered prob-
lem, in particular when a non-uniform distribution of customer requests is present
as well as in the case where two or more customer requests dynamically arrive. As
an outlook on future research, the authors suggest considering problem variants in
which a non-uniform distribution of new customers and the arrival of multiple dy-
namic customers is present. Furthermore, they propose evaluating possibilities of
modifying tours during the execution of the transportation process and improving
the proposed evolutionary algorithm.
In Attanasio et al. (2007), a real-time management system for coordinating a fleet
of same-day couriers in London, United Kingdom, is presented. The considered
problem is a DPDP with dynamically arriving requests and uncertain travel times.
The objective is to maximize the quality of customer service as well as the average
number of serviced customers per courier. Customers are divided into three groups
that are distinguished from each other according to the maximum allowed response
time to newly arriving requests. Vehicle en-route diversion is allowed to service dy-
namically arriving requests. Stochastic knowledge about request arrivals is derived
from historical data by using a time series decomposition approach. Specifically, the
service area is divided into zones and for each pair of zones, the expected request
arrival rate is calculated. Travel times are learned from historical data while they
are also updated in real-time using online traffic information during the day. The
deployed couriers can accept or reject requests that are offered to them by the im-
plemented request management system. Requests which are not directly accepted by
couriers are integrated into the tour plan by using a Tabu Search procedure. While
stochastic knowledge about expected travel times directly influences the assignment
of requests to couriers, stochastic knowledge about expected request arrival rates
is used for relocating idle vehicles to zones in which the expected request arrival
rate is high in the near future. While providing no computational results, the authors
mention that the use of the system has increased the quality of service as well as
courier efficiency and that it also provides reduced marginal administration costs
for additional deployed couriers.
In Kleywegt et al. (2009), the authors study a dynamic and stochastic multiple-
period routing problem in which new requests arrive over time but only one dy-
namic request arrives per period. Based on SLAs, the dispatcher can decide whether
a newly arriving request should be serviced in the current period or better still in the
next period. The considered objective is to minimize the average transportation cost
per period over a given planning horizon. It is assumed that probability distributions
of newly arriving requests are available so that stochastic knowledge can directly be
utilized. The authors examine various dispatch strategies. These strategies include
a two-stage myopic policy, a sampling-based policy, a multi-stage sample average
approximation policy, and an optimal policy. Three different types of service areas
are evaluated in the computational experiments, namely an interval, a circle, and a
disk. The relative performance of the strategies is assessed and also compared to
an optimal solution in the corresponding offline case where all information about
118 4 Review of the Literature Related to the Considered RDOPG Applications

future request arrivals is known in advance with certainty. Two different request lo-
cation distributions are tested. In the first setting, request locations are independent
and identically distributed according to the uniform distribution on the interval, and
in the second setting, customer locations are independent and identically distributed
according to a specific discrete distribution. Computational results provided for sev-
eral instances under the different scenarios show that the multi-stage sample average
approximation policy often performs best among the tested approaches. The authors
propose to extend the approaches to the multiple request case per period. First re-
sults for this case can be found in the PhD thesis of Uyar (2008).

4.5.2 Approaches Using Exact Solution Methods

In the work of Yang et al. (2004), the authors consider a multi-vehicle truckload
DPDP in which each vehicle can carry at most one request at the same time. Re-
quests dynamically arrive over time and a small time period for each request is
given during which a newly arriving request may be accepted or rejected. Requests
have soft time windows at pickups and deliveries so that service after the end of a
time window is allowed but at the expense of additional linear lateness costs. Rev-
enue is gained if a request is accepted, otherwise penalty costs have to be paid. The
objective is to maximize the total profit resulting from request profits and penalties
as well as lateness and travel costs. The individual problem instances which arise
during the execution of the transportation process are formulated as a Mixed Inte-
ger Programming (MIP) problem whereas, due to complexity reasons, the number
of considered requests in each MIP problem instance is limited by a pre-selection
routine. The authors evaluate five different policies for solving the dynamic prob-
lem. Among these policies are three local adaptation heuristics and two approaches
which globally reoptimize the tour plan. The latter two approaches solve the MIP
problem with a time limit of 20 seconds. Both approaches are different from each
other in that one reoptimization approach makes myopic decisions based on deter-
ministically available knowledge only while the other approach additionally uses
given stochastic knowledge about expected future request arrivals. This stochastic
knowledge is utilized by integrating expected opportunity costs into the objective
function. These costs are based on the current and future vehicle positions resulting
from assigned requests in relation to expected future request arrivals. Hence, vehi-
cles are more likely to be deployed within the considered service area according to
the expected request arrival rates. Computational results on test instances with 10 ve-
hicles and 1,000 requests with varying degrees of system utilization show that the
two reoptimization approaches perform better than the local adaptation heuristic ap-
proaches. Out of the two evaluated reoptimization-based approaches, the one which
additionally uses stochastic knowledge performs best by yielding about 5 % better
results compared to the myopic reoptimization approach. The authors mention that
further research on methods which utilize stochastic knowledge more efficiently is
likely to have a big potential for additional improvements.
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 119

4.5.3 Stochastic Modeling Based Approaches

In this section, approaches which model the dynamic routing problem by using
specific stochastic formulations are presented. According to Powell et al. (1995),
these models are called multistage stochastic models which directly utilize stochas-
tic knowledge available by prescribed probabilistic distributions. For example, these
models use Markov Decision Processes (MDP) or Approximate Dynamic Program-
ming formulations.
Powell (1988) considers a truckload DPDP in which each vehicle can carry one
request at a time. In this problem, requests have to be picked up at one location and
require transportation to another location. After the request has been delivered, the
next action for the empty vehicle has to be decided. Specifically, possible actions
consist of assigning the vehicle to another pending request, keeping it at its current
location, or relocating it to another location (e.g., a different city) in anticipation
of future requests arriving there. Hence, waiting and relocation strategies are used
in this approach. Stochastic knowledge is given by request arrival rates which vary
in different subregions and may also change during the day. The considered practi-
cal application arises from a business case in which large transportation companies
regularly subcontract requests. According to the author, this problem has a degree
of dynamism of 60 %. Since newly arriving requests are often delegated to compa-
nies that currently have trucks close to the pickup location of a new request, it is
beneficial for carriers to have trucks close to locations where new arriving requests
are expected to occur in the near future. However, each company has to balance
travel costs generated by relocating trucks with the expected profitability of arriv-
ing requests in each region. Moreover, companies may reject requests if they do not
appear profitable. The author names this problem the dynamic vehicle allocation
problem (DVA). In this work, the author gives an overview of different techniques
on how to model the DVA in the single and multiple-vehicle variant. In addition
to deterministic models including static single stage and dynamic time-space mod-
els, stochastic network formulations are presented. Furthermore, the problem is also
modeled using an MDP formulation. The author mentions that the state space grows
very quickly so that the problem becomes intractable for real-life problem instances.
Moreover, a two-stage stochastic programming formulation of the DVA is described
in which the first stage consists of known requests and the second stage represents
the expected value of having additional vehicles in specific regions. Insights into
the multistage deterministic formulation and the two-stage stochastic formulation
are combined to form a hybrid model. Moreover, an overview of alternative models
is given. This work provides one of the first references in the literature for using
stochastic knowledge in order to relocate vehicles and decide about their waiting
times at certain locations where future requests are anticipated.
In a later paper, Powell (1996) continues the research of a long-haul transporta-
tion process in which requests have to be assigned to vehicles in real-time. Besides
reviewing a variety of alternative models for this dynamic assignment problem, the
author develops a new model for approximately solving the stochastic and dynamic
120 4 Review of the Literature Related to the Considered RDOPG Applications

request to vehicle assignment problem. A hybrid model which consists of an as-

signment and a forecasting component that anticipates future requests is presented.
In the proposed model formulation, costs of possibly required recourse actions are
included in the arc costs of the transportation network. The author shows that the
model is computationally fast enough to handle the largest problem instances which
arise in the considered practical application while being flexible enough to integrate
different practical requirements at the same time. Computational results carried out
on experiments with 200 vehicles and 600 requests per week with a degree of dy-
namism of about 60 % demonstrate that the implemented anticipative model outper-
forms standard myopic models.
In the work of Ichoua (2001), a DVRP is considered in which vehicles that have
finished servicing a request are not always directly sent to their next assigned re-
quest. Instead, if stochastic knowledge indicates that further requests are expected
to arrive soon within the vicinity of a vehicle’s last serviced request location, the
vehicle is required to wait at its current position for a given amount of time. The
results of this thesis are integrated into the approach of Ichoua et al. (2006) which
is described below.
The work of Thomas and White III (2004) deals with a routing problem in which
a single uncapacitated vehicle travels on a road network from an origin O to a des-
tination D which needs to be reached by a defined time point at the latest. The arcs
in the road network have different travel costs and dynamism arises from newly ar-
riving customer requests which demand goods to be picked up and transported to
destination D. A special aspect of this approach is that new requests only arise at a
set of known locations, i.e., at a defined set of nodes in the road network. Requests
do not have time windows nor service times. For each potentially arriving request,
the probability of its arrival is given as a function of time. When a pending request
is serviced, defined revenue is granted but requests may also be rejected. The ob-
jective is to determine the optimal path along the road network so that the expected
profit is maximized, i.e., the sum of expected revenues from serviced requests mi-
nus travel costs. Hence, if revenues are small and travel costs are high, the optimal
policy might consist of just traveling on the least cost path from O to D while ne-
glecting all potentially arriving requests. The authors formulate the problem as an
MDP and solve it using stochastic dynamic programming techniques. The solution
is iteratively constructed by making a decision at each node that the vehicle visits
about which node the vehicle should travel to next. The vehicle anticipates future
requests by not only visiting possible request locations but also other nodes in the
road network. Different experiments on a road network with 132 nodes and 202 arcs
are carried out on different probability distribution settings according to the mean
and variance of the applied request arrival probabilities in order to analyze the effect
of the implemented anticipatory routing method. Computational experiments which
are conducted in cases where one customer additionally arrives indicate that the
optimal anticipatory routing policy performs well compared to a deterministic and
purely deterministic version of the solution approach. This is especially true if the
probability that this additional customer appears late has a high mean and a low vari-
ance value. The authors also provide computational results for the multiple request
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 121

case which turns the considered problem into an NP-hard TSP. This fact can also
be observed at the computational times required for calculating an optimal solution.
Specifically, while less than one second is required in the single request case, an
instance with seven potentially arriving customer requests requires over 8,000 sec-
onds of computational time. Hence, the authors state that using the proposed dy-
namic programming approach might encounter problems in solving real-life sized
problem instances so that future research should improve the computational effi-
ciency of the applied solution methods. Furthermore the authors suggest extending
the anticipatory strategy to the multiple-vehicle variant of the considered routing
problem.
In Ichoua et al. (2006), the authors consider an uncapacitated DVRP where a
set of given heterogeneous vehicles needs to service dynamically arriving requests
with soft time windows. If service of a request is performed after the end of its time
window, linear penalty costs occur. All vehicles have to return to the depot by a
given time point at the latest. The objective is to maximize the number of serviced
customers and, as a subordinate goal, minimizing the sum of travel and lateness
costs. Vehicles move in the Euclidean plane and en-route diversion of vehicles is
not allowed. Requests arrive according to time- and space-dependent Poisson pro-
cesses. Specifically, the considered service area is divided into geographic zones and
the considered planning horizon is divided into time periods. Each of these zones
can have a different request arrival rate assigned in each time period. Therefore, re-
quest arrival rates vary in different zones of the considered service area. Stochastic
knowledge is given by having knowledge about the intensity rates of the time-space
Poisson processes in advance. In order to increase the ability of servicing more
requests, vehicles may wait at the location of their last serviced request if the prob-
ability of further arriving requests nearby is high in the near future, even if requests
that are located further away are currently pending. Hence, the VRP is extended to a
VRSP (see Sect. 2.4.6) using a waiting strategy which utilizes the available stochas-
tic knowledge. The authors formulate the problem using a dynamic programming
approach which calculates an optimal solution for assigning requests and scheduling
the vehicles. Since the problem is strongly NP-hard, the authors propose a vehicle-
waiting heuristic to solve the problem. In the proposed heuristic that uses results
from the PhD thesis of Ichoua (2001), for each idle vehicle k at time t it is decided
whether the vehicle should wait for a defined amount of time δtk or directly travel
to the next request. If the probability that new requests will arrive within the neigh-
borhood geographical zone where vehicle k is currently located during the time
interval [t, t + δtk ] is higher than a given threshold and if not too many other idle
vehicles are already waiting in this zone, the vehicle waits for δtk time units. Oth-
erwise, it directly moves to its next assigned request. In order to determine which
request is assigned to a vehicle next, the Parallel Tabu Search metaheuristic pro-
posed in Gendreau et al. (1999a) is applied in a modified version. Specifically, if
a vehicle is ordered to wait, a dummy customer is inserted at the beginning of the
tour of the corresponding vehicle k stating that the vehicle is available for further
activities only at t + δtk . If a new request arrives while a vehicle is currently waiting
in the associated zone, the temporarily inserted dummy customer is replaced by this
122 4 Review of the Literature Related to the Considered RDOPG Applications

new customer request. If no customer request has arrived until time point t + δtk ,
the dummy customer is removed from the vehicle’s tour and the vehicle continues
to travel to its next assigned customer according to the best solution in the adaptive
memory of the Parallel Tabu Search. Computational results on four different sce-
narios resulting from two different request arrival rates (27 and 36 requests per hour
leading to an average of 138.1 and 187 requests per instance, respectively) com-
bined with two different fleet sizes of 4 and 6 vehicles are performed. The proposed
vehicle waiting heuristic-enabled approach is compared to the original Parallel Tabu
Search procedure which does not use any stochastic knowledge. The computational
experiments show that the utilization of the vehicle-waiting heuristic yields signif-
icant improvements, especially when utilized in scenarios with a smaller fleet size
or more pending requests.
Thomas (2007) continues the study on applying waiting strategies in dynamic
routing problems. The author considers a dynamic and stochastic routing problem
with a single uncapacitated vehicle. The objective is to maximize the expected num-
ber of requests which can be serviced. In this problem, stochastic knowledge about
expected future requests is used in order to decide when and where the vehicle
should wait. The author utilizes an MDP formulation and derives analytical charac-
teristics of tested policies. In the scenario in which one customer dynamically ar-
rives, the optimal policy is completely analyzed. These results are used for defining
a real-time heuristic and other heuristics that can be applied in scenarios in which
multiple requests dynamically arrive. Computational results conducted on different
data sets show that using a heuristic strategy which uses stochastic knowledge de-
noted as center-of-gravity longest wait heuristic, performs better when the percent-
age of dynamic customers is 25 % or less. In instances with a higher degree of dy-
namism, the authors state that distributing waiting times across the currently known
requests is an effective heuristic strategy which has also been shown to perform
well in situations where no stochastic knowledge is available (see Mitrovic-Minic
and Laporte 2004 in Sect. 4.4). The authors suggest including the consideration of
the multiple-vehicle case as well as time windows for requests in future research
work.
Secomandi and Margot (2009) consider a single-vehicle capacitated Vehicle
Routing Problem with Stochastic Demands (VRPSD). In this problem, customer
requests which are known with certainty need to be supplied with a specific good.
Uncertainty arises from the fact that the quantity demanded by each request is only
known by prescribed probability distributions. Note that most approaches in the
literature solve this kind of problem by using a priori optimization methods as de-
scribed in Sect. 2.5.3.2.1. However, as mentioned in the same section, some authors
state that the full potential of solving these problems can only by achieved by using
real-time solution methods. Hence, the authors propose a new optimal decision strat-
egy for solving the transportation process during its execution. Having finished the
current request activity, the proposed strategy called partial reoptimization decides
whether the vehicle should move to one of the pending customer requests directly
or first return to the depot and perform a replenishment activity. Furthermore, split
deliveries are possible so that a portion of the customer’s demand can be deliv-
ered before and after the intermediate return to the depot. This decision takes the
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 123

currently remaining vehicle capacity into account. Specifically, the vehicle should
perform the replenishment activity next when its currently remaining load drops
below a threshold. This threshold is based on various factors such as the expected
demands of pending customer requests. The authors propose a new computationally
efficient MDP formulation of the proposed strategy. The formulation is based on the
concept of solving a restricted set of states so that at least one admissible policy,
i.e., a solution in which the demands of all pending requests are entirely serviced, is
determined. For instances with up to 15 customer requests, an optimal solution can
be generated by using the proposed solution approach. In larger instances, memory
limitations occur so that two different methods are introduced in order to select a
subset of states of the MDP. Based on the restricted set of states, the policy is de-
termined by backward dynamic programming. The performance of the approach is
assessed on various instances. The authors show that on instances with up to 90 cus-
tomers several versions of the proposed heuristically operating MDP approaches
outperform existing heuristics in the literature and that the best performing heuristic
has a gap of about 10–13 % compared to a lower bound which has been computed
with perfect knowledge of customer demands given in advance. Furthermore, new
lower bounds for the VRPSD are derived. With the ability of solving instances up
to 15 customers to optimality, the authors show that for these cases, the proposed
lower bounds are quite tight.
As has been mentioned in this section, solving real-life problem size instances of
dynamic routing problems to optimality by using stochastic models is often compu-
tationally intractable. This is due to the generally used recourse function applied in
dynamic programming which determines an optimal solution by first solving a large
number of subproblems and then determining an optimal solution by backward re-
cursion. The issue that frequently occurs is that the state space quickly becomes
too large (cf., e.g., Desrochers et al. 1988 and Powell 2007, p. 47). In order to deal
with this issue, the backward-working dynamic programming has been modified to a
forward-working Approximate Dynamic Programming (ADP). In ADP approaches,
the need to evaluate all possible states is eliminated by approximating the value
function in a forward-manner (cf. Pillac et al. 2010). Further information on ADP
approaches can be found in Powell (2007, 2009). According to Pillac et al. (2010),
ADP has been successfully used in various field of transportation applications. One
of these fields is fleet management. Here, Godfrey and Powell (2002) consider a
stochastic version of a dynamic resource allocation problem in which a set of ho-
mogeneous resources is assigned to randomly arriving new requests which are re-
quired to be picked up at one location and need to be delivered to another location.
Furthermore, some requests are known at the beginning of the planning horizon.
Three characteristics related to routing problems make this problem different from
a usual assignment problem: the integration of costs for preparing resources for ser-
vicing requests, penalty costs for late service of requests, and the occurrence of new
vehicles (besides new requests) over time. It is assumed that stochastic knowledge
which provides arrival rates of expected future events is available. The objective is to
maximize the profit consisting of the expected reward for servicing requests minus
transportation and penalty costs. The problem is solved using an ADP approach with
124 4 Review of the Literature Related to the Considered RDOPG Applications

nonlinear functional approximations which provide information about the expected

value of resources in the future. Computational tests show that the proposed method
significantly outperforms a rolling horizon approach which adapts the transporta-
tion process by using deterministically known information only. The authors state
that by using stochastic knowledge, improvements of 2 % to 16 % can be attained
where higher improvements can be achieved in scenarios with less resources or more
customer requests, i.e., in scenarios with a higher resource utilization. Other appli-
cations to fleet management can be found among others in Powell and Topaloglu
(2005).
Similar to the previously described work of Secomandi and Margot (2009),
Novoa and Storer (2009) introduce a new approach for dealing with a single-vehicle
VRPSD from a real-time control approach perspective. They formulate the problem
as a stochastic shortest path problem that is related to an MDP formulation. Further-
more, the authors propose an ADP approach called the rollout algorithm. Differ-
ent variants of one-step and two-step lookahead stages in the rollout algorithm are
compared to an a priori solution on instances with 60 customers. The authors note
that computing an optimal solution of VRPSDs is currently intractable for instances
with more than 20 customer requests. The experiments show that among the tested
methods, the two-step lookahead rollout algorithm performs best. Specifically, the
routing costs are about 4.8 % less compared to the results achieved by the one-step
rollout algorithm. Moreover, the authors compare the results obtained for the online
instances with results of related offline instances in which perfect information is
given in advance. For one of the two evaluated instance sets, the authors show that
the best tested two-step rollout algorithm is only 3.38 % away from the estimated
perfect information lower bound.
Schmid (2012) considers a practical application dealing with real-time coor-
dination of ambulances in Vienna, Austria using a real road network with time-
dependent travel times. When a new emergency request arrives, an ambulance has
to be chosen among the idle and waiting ambulances and dispatched to serve the
new request under the objective to minimize the response time of this request that
has newly arrived. When an ambulance has finished servicing a request and becomes
idle again, a decision has to be made about where to position this vehicle. For this
purpose, a set of predefined waiting locations in the city are available. The aim is
to relocate the vehicle to a specific waiting location so that the estimated response
time of expected future requests is minimized. Due to legal regulations, the clos-
est vehicle is always sent to a request so that the decision about which ambulance
should be assigned to a request is not considered. In order to be able to quickly
service expected future requests, it can also be advantageous to relocate waiting
vehicles during the day for two reasons. First, the overall and region-dependent re-
quest arrival rates change throughout the day. Second, due to the existence of time-
dependent travel times, the coverage rate defining the area that is reachable from a
given idle position within a maximum amount of time also varies during the day.
However, since another law prohibits the relocation of an ambulance from one wait-
ing location to another while it is idle, as far as the original problem is considered
this decision can only be taken for vehicles which just have finished servicing a
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 125

request. In order to solve the considered problem, the author proposes a new ADP
approach which derives good positioning decisions depending on the current system
situation by analyzing historical real-world request data and given time-dependent
travel time information. Hence, in the described approach, stochastic knowledge is
automatically generated from the available past request and travel time information.
Extensive computational tests on real-world problem instances show that, compared
to the currently applied strategy in which ambulances have a fixed waiting location,
the approach is able to reduce the average response time by 7 % while also reducing
the maximum response time. In a further test study, the regulation which stipulates
that the closest vehicle is always required to be dispatched is relaxed if a new re-
quest is not life-threatening. Using this relaxation, the average response time can
be further reduced by 0.7 %. Future work is directed at allowing the direct en-route
dispatching of idle ambulances which are currently traveling to a waiting location.
Moreover, the evaluation of improvements attainable by relocating idle ambulances
between waiting locations according to changing expected request arrivals during
the day is also suggested.
Another type of state-based solution approaches are Hybrid Adaptive Predictive
Control (HAPC) approaches. In HAPC approaches, individual states involve both
integer and continuous variables. They work in discrete time steps by generating
individual problem instances. A new time step is considered whenever a new rel-
evant event occurs during the execution of the transportation process. Sáez et al.
(2008) develop a HAPC approach for a capacitated multiple-vehicle DPDP that can
be considered as a DDARP in the Euclidean plane. Dynamism arises from newly
arriving customer requests that demand an instant transportation from an origin to a
destination. In the considered problem, up to four customers can be transported in a
vehicle at the same time. As known from DARPs, the objective includes quality of
service factors. Specifically, the objective is to minimize the waiting and travel time
of customers as well as operational costs mapped by idle travel times of vehicles.
The authors analytically develop a problem-specific state space formulation which
uses results published in Cortés et al. (2009). The dynamic transportation problem
is carried out by solving the problem for the complete number of defined antic-
ipation steps and, after the solution process is completed, implementing the next
step of the best found solution. The authors mention the importance of the num-
ber of considered anticipation steps which is heavily problem-dependent. In order
to solve the problem, the authors propose an explicit enumeration scheme as well
as a GA which is evaluated in three variants that can be distinguished from each
other according to the number of considered individuals and generations. Histor-
ical information about requests is available and utilized to automatically generate
stochastic knowledge about expected origin-destination trip patterns using a fuzzy
clustering method (FCM, cf., e.g., Xu 2005). Specifically, the FCM assigns past re-
quests to one or more defined fuzzy zones. In fuzzy zoning, requests can belong to
multiple zones through a so-called membership function which defines to what per-
centage a request belongs to each zone. Stochastic knowledge is then generated by
assigning requests to the fuzzy zone to which this request belongs with a percentage
which is at least a specified threshold. If a request does not occupy a zone with such
126 4 Review of the Literature Related to the Considered RDOPG Applications

a minimum required percentage, it is not considered during the generation of the

stochastic knowledge. In doing so, only stochastic knowledge with a certain quality
is generated. Several test instances are evaluated in which the available historical
data is generated from patterns which have been designed according to observed
request arrivals in a real-world request data set. For small problem instances, the
authors show that applying the explicit enumeration scheme is a possible option for
generating solutions of high quality. For instances of real-world size, the explicit
enumeration scheme produces an impractical runtime so that the use of the pro-
posed GA is advantageous. The GA generates the best results in the tested variant
with the largest population size and the largest number of generations. Compared
to a myopic strategy of the approach which does not use stochastic knowledge, the
application of a two- and three-step ahead version of the selected GA variant results
in significant improvements on tested instances with 120 customers. Specifically,
the three-step ahead version is able to reduce waiting time by more than 22 %. In an
outlook for future research, the authors propose that stochastic knowledge should
be generated as a combination of historical data and updated information which be-
comes available during the execution of the transportation process. Moreover, the
authors suggest extending the objective function by including hard and soft time
windows and a more detailed consideration of operational costs.
In a consecutive paper, Cortés et al. (2008) consider two sources of dynamism,
in particular future request arrivals as well as uncertain travel times in the Euclidean
plane. For both types of dynamism, stochastic knowledge is assumed to be available
in the presented HAPC approach in which it is utilized to generate expected future
costs. Moreover, deterministically known time-dependent travel times are consid-
ered. Computational results show that a two-step ahead approach is able to achieve
significant improvements compared to a myopic one-step approach that does not
make use of the available stochastic knowledge. Furthermore, in Núñez et al. (2008),
the authors extend the objective function by including operational costs. Since the
minimization of total operational costs and the minimization of the previously pro-
posed customer inconvenience aspects are to some extent contradicting aims, the
authors use an evolutionary multi-objective optimization concept in order to cal-
culate the Pareto-optimal front of solutions. Three different criteria for selecting a
solution which is to be implemented in each decision step are proposed. Compu-
tational results show that the proposed approach can be efficiently used to either
prioritize waiting time, i.e., minimization of customer inconvenience, or travel time
and distance, i.e., minimization of operational costs, while also considering the other
aim to some extent.

4.5.4 Approaches with Manually Provided Stochastic Knowledge

In this section, dynamic routing approaches in the literature in which available

stochastic knowledge is manually provided are presented. In van Hemert and
La Poutré (2004), the authors consider a DVRP in which each vehicle has a fixed ca-
pacity and needs to pick up requests which have to be transported to a central depot.
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 127

Pending requests may be rejected and the considered objective is to maximize the
number of serviced requests while considering fixed time windows of the requests
and a latest return time at the depot. The motivation for utilizing stochastic knowl-
edge in the considered approach is that in various real-world applications request
arrivals are often not uniformly distributed but appear in a clustered fashion. In their
approach, the authors denote such clusters as fruitful regions which are character-
ized by a high probability that new requests will arrive there. In the approach, test
instances are generated by first defining a set of fruitful regions, represented by a
set of clusters, around which requests are generated in a subsequent step within a
predefined range. In order to solve the dynamic transportation process, an evolution-
ary algorithm is used. It involves a strategy called anticipated move which relocates
vehicles to fruitful regions in expectation of newly arriving requests. The decision
whether to perform anticipated moves is directly integrated into the objective func-
tion of the evolutionary algorithm. This decision is based on the stochastic knowl-
edge available from information about fruitful regions. Using anticipated moves,
vehicles are ordered to travel to fruitful regions even if no requests have arrived
there yet but are expected to do so with a high probability in the future. Benchmark
results indicate that for certain parameter settings regarding time windows and ve-
hicle capacities, the utilization of the proposed anticipated move strategy is very
beneficial.
Branchini et al. (2009) consider a DVRPSTW in which the fleet consists of ve-
hicles with the same capacity. While some customers are known in advance, dy-
namism arises from newly arriving requests whose location and demand is only
known by probability distributions which are available as stochastic knowledge.
Pending requests have soft time windows so that they may be serviced late at the
expense of linear lateness costs. Furthermore, requests may also be rejected. Each
request has an individual revenue and a variable lateness penalty. Tours of vehicles
are dynamically adapted by a real-time control which pursues the objective of max-
imizing the expected profit regarding request revenues, lateness costs, fixed costs
for each utilized vehicle, and travel costs. A tour plan adaptation by the real-time
control approach takes place whenever a new request arrives and vehicle en-route
diversion is allowed. Tours are constructed using a proposed constructive heuristic
which scatters the utilized vehicles in the service area. Afterwards, the generated
tours are improved by an adaptive granular local search procedure. The distribution
of the vehicles over the service area is achieved by calculating so-called seed cus-
tomers. These seed customers are created by including four request-specific criteria
such as time window incompatibility and vehicle waiting times at requests. Each
seed customer is assigned to a different vehicle tour. Due to the fixed vehicle usage
costs, one aim is to determine the number of required vehicles before the execu-
tion of the transportation process. However, additional vehicles can be deployed at
additional costs during the day if necessary. The real-time control approach makes
use of relocation and waiting strategies. For this purpose, based on historical data,
strategic waiting positions are manually identified and provided to the approach as
stochastic knowledge. The performance of the approach is assessed using designed
test instances and real-world problem instances that are based on transportation pro-
cesses of Brazilian logistic companies. These instances are determined by a degree
128 4 Review of the Literature Related to the Considered RDOPG Applications

of dynamism between 60 % and 80 %. Different settings for the real-time control

approach and the applied flexibility-increasing strategies are evaluated. Computa-
tional results show that the application of all three strategies, i.e., vehicle en-route
diversion, waiting strategy, and relocation strategy in the real-time control approach
which uses the proposed granular local search performs best among the tested solu-
tion approaches.

4.5.5 Sampling-Based Approaches

In comparison to the approaches presented above, sampling-based approaches uti-

lize stochastic knowledge by sampling available probability distributions. This sam-
pling results in one or more scenarios that contain specific realizations of the sam-
pled distributions. When sampling is used to anticipate future request arrivals, this
often generates possible future scenarios that contain different expected requests.
After optimizing each of these instances, analyzing individual solutions generated
for the individual instance can yield insights that lead to robust decisions according
to expected future events.
Bent and Van Hentenryck are among the first to use sampling-based approaches
in dynamic routing applications. The basic problem which is considered by the au-
thors is a DVRPTW in which dynamism occurs by newly arriving requests. In the
past few years, both authors have proposed several sampling-based approaches for
dynamic routing problems which are known as Online Stochastic Combinatorial
Optimization (OSCO, cf. van Hentenryck and Bent 2009 and Bent et al. 2010).
The authors use a general framework for online optimization in all of their pro-
posed strategies. In this framework, two black boxes are used for decision mak-
ing: a problem-specific sampling procedure with a given distribution for generating
stochastic knowledge resulting in sample scenarios and an optimization algorithm
for solving these sampled problem instances (cf. Bent and van Hentenryck 2007).
According to Bent et al. (2010), the idea of using sampled requests comes from the
key observation that in many practical dynamic applications, the arising uncertainty
does not depend on the decisions taken since request arrivals are independent of
generated tour plans and vehicle activities. Hence, the authors state that in these ap-
plications it is not necessary to use more general models that explore sequences of
decisions or scenario trees. Therefore, given distributions can be sampled indepen-
dent of the decisions taken so that expected future scenarios can be provided. As
a consequence, the proposed framework provides significant computational advan-
tages compared to the previously mentioned general models. The proposed frame-
work is depicted in a basic form in Algorithm 4.1.
The proposed framework can be used for different online optimization problems,
but may require problem-specific adaptations. For example, in dynamic routing
problems with hard time windows, it is necessary to check if a new arriving request
can be feasibly integrated into the tour plan. The dynamic problem is considered
as a discrete-time model over a given time horizon H = [H , H ]. In what follows,
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 129

Algorithm 4.1 Basic online optimization (H ) (adapted from Bent and van Henten-
ryck 2004c)
1: R ← ∅
2: w ← 0
3: t ← H
4: while t ≤ H do
5: R ← AVAILABLEREQUESTS(R, t) ∪ NEWREQUESTS(t)
6: r ← CHOOSEREQUEST(R, t)
7: t ← SERVICEREQUEST(r, t)
8: w ← w + w(r)
9: R ← R \ {r}
10: end while

this basic model is considered for a dynamic routing problem in which no vehicle
en-route diversion is allowed. Note that this basic formulation only considers rout-
ing problems in which a single vehicle is present since the described framework can
only make a new decision after the previous one has been completed. The tour plan
is iteratively generated. Specifically, when the vehicle has finished servicing its cur-
rently assigned request at t ∈ H , for each pending request r ∈ R a weight w(r) that
represents the gain according to the given objective function, if request r is chosen
for service at time t, is assigned. Note that this weight may change over time. At
the beginning of the process, t is set to the beginning of the considered time horizon
denoted by H . While the time horizon has not ended, new requests that are avail-
able up to the current time are added to the set of pending requests R. As mentioned
above, one request from this request set is chosen to be serviced next. After this
selection, the request is serviced (line 7) and the time t is advanced to the point in
time when service of request r is completed. Furthermore, the realized gain w(r) is
added to the objective function value of the solution denoted by w and the request is
removed from the set R. In Bent et al. (2010), the authors make several extensions
to this basic framework, among others, for multiple decisions in order to extend the
problem to the multiple-vehicle case.
Regarding the basic strategies in DVRPs for providing increased flexibility as
presented in Sect. 4.3, it can be observed that the described framework considers
choosing the next request as the only type of decision so that the definition of the
function CHOOSEREQUEST(R, t) is the essential task for applying this frame-
work. An implementation of this function may or may not use stochastic knowledge.
The authors mention two implementations of this function that do not use stochas-
tic knowledge. Specifically, CHOOSEREQUEST-G is a greedy heuristic that sim-
ply selects the request that possesses the highest objective function gain out of the
pending requests. In another implementation called CHOOSEREQUEST-LO, it is
assumed that a function called OPTIMALSOLUTION exists which can compute an
optimal solution for the set of currently pending requests. The first request of this
solution is chosen for execution at time t.
In addition to the implementations CHOOSEREQUEST-G and CHOOSERE-
QUEST-LO that do not use stochastic knowledge, the authors propose different im-
130 4 Review of the Literature Related to the Considered RDOPG Applications

plementations of the function CHOOSEREQUEST which make use of stochastic

knowledge. All of these implementations use a concept denoted as Multiple Sce-
nario Approach (MSA) which is based on the MPA (see Sect. 4.4). In MSA-based
approaches, pro-active strategies are implemented by sampling a given probabil-
ity distribution representing expected future request arrivals. By using this given
stochastic knowledge, expected future requests called sampled customers are cre-
ated (cf. Bent and van Hentenryck 2007). By sampling the distribution n times,
n different realizations of an expected future scenario consisting of currently known
and sampled customer requests are generated. These realizations are then used in
the decision process. MSA-based implementations are similar to a sampling-based
approach for online packet scheduling called expectation which was proposed by
Chang et al. (2000). In this approach, for each sample scenario s ∈ S the gain that
is expected if request r ∈ R is serviced next in scenario s is calculated by assum-
ing that r is serviced next and then using the solution method to generate a solu-
tion for requests R\{r}. For each request r, the sum of the expected gains of all
scenarios is calculated and the request with the highest gain is chosen. The advan-
tage is that this is likely to lead to a robust solution since different possible future
outcomes are directly considered in the decision process. A drawback of the expec-
tation approach is the time complexity since the solution method is required to be
performed |S| · |R| times which can raise bottlenecks under time constraints which
exist in real-time control approaches. In order to reduce the required computational
effort, decisions in MSA-based implementations about which request is to be ser-
viced next when a vehicle has finished servicing its current request are performed
similar to the MPA which was described in Sect. 4.4. Specifically, while in the MPA
a solution is constructed out of a set of stored solutions using a specific ranking
function, in the MSA a set of sample scenarios S is created and a solution is gen-
erated for each of them. By utilizing a specific ranking function, the next request
is selected using these |S| generated solutions. The advantage is that only |S| sce-
narios, instead of |S| · |R| scenarios, need to be solved which can be realized more
easily in time-restricted real-time systems. Multiple-vehicle routing problems are
handled in MSAs by storing the decisions about the next request for each vehicle
separately. In the first MSA-based implementations in Bent and van Hentenryck
(2004b), the two ranking functions which have already been shown to be efficient
in the MPA are used for selecting the next request. Specifically, the tested ranking
functions distance and consensus, result in the MSA variants MSAd and by MSAc .
In both MPA variants, the sampled customers are removed after generating the so-
lutions for the sample scenarios providing flexibility for integrating future requests.
Moreover, a real request is always selected by the ranking functions. Computational
results show that each of the two tested ranking functions in the MSA significantly
outperforms the corresponding MPA variant. Furthermore, the authors recommend
the utilization of MSAc instead of MSAd since, especially for high degrees of dy-
namism as well as late arriving dynamic customers, MSAc significantly reduces the
number of unserviced customers compared to MSAd .
In the working paper of Bent and van Hentenryck (2003), the authors evaluate
the previously described MPA and MSA-based approaches on the problem setting
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 131

considered by Larsen et al. (2002) (see Sect. 4.5.1). Computational results show that
in the single-vehicle case, the utilization of the stochastic knowledge in the MSA
can improve the solution quality compared to the original NN heuristic by Larsen
when applied in scenarios with a larger delivery area. In the same paper, the authors
extend the MSAc approach to the MSAc with least commitment (MSA-LCc ) that
makes additional use of a waiting strategy. Specifically, if a sampled customer is
chosen as the next requests in a vehicle’s tour, the corresponding vehicle will wait
at its current position until the sampled customer materializes or is removed from
the first position in the tour, i.e., until a real request is assigned to this vehicle as
its next request. Computational results are carried out in a multiple-vehicle setting
with capacity constraints and hard time windows. The authors show that the MSA-
LCc implementation is able to significantly reduce the total travel distance since it
postpones commitments to requests when future requests are expected to arrive. In
contrast, when the minimization of unserviced customers is pursued, the best results
are achieved by using the NN heuristic proposed by Larsen et al. (2002) as well as an
MSA approach variant in which the NN heuristic is used for solving the individual
sample scenarios.
In the work provided by Bent and van Hentenryck (2004c), the authors continue
to study characteristics of the consensus ranking function. The authors mention that
the consensus ranking function’s main limitation is its elitism, i.e., only the best re-
quest is considered in a sample scenario while other suitable requests which may
also provide robust solutions are neglected. Hence, the authors propose a modified
version called quantitative consensus which calculates the consensus value by the
sum of objective function values of the solved sample scenarios in which this re-
quest is scheduled next. Since this version of the consensus function does not reduce
but rather even increases the observed elitism, the authors present two new ranking
functions. The first one is a combination of the consensus and the expectation rank-
ing function. In this function, first the consensus is applied in order to identify a set
of promising requests. Afterwards, the expectation ranking function is applied to
this set in order to achieve a more precise ranking among them. However, since the
computational effort generated by this ranking function is still considerably high
when compared to the consensus ranking function, the authors propose a second
ranking function which is called regret. By quickly estimating negative impacts on
other requests when a specific request is scheduled as the next request in a vehicle’s
tour, this ranking function provides additional ordering information among requests
without having to solve additional scenarios. Details on this ranking function can
also be found in Bent et al. (2005) and Bent and van Hentenryck (2007). Computa-
tional results show that the regret ranking function performs well compared to other
evaluated ranking functions, especially in scenarios with a high degree of dynamism
and tight time window restrictions.
In Bent and van Hentenryck (2007), the authors propose two extended MSA
variants that make additional use of a waiting and a relocation strategy. Using the
consensus or regret ranking function, the next request of a vehicle is identified while
assigned sampled customers as next requests are regarded as a decision for the wait-
ing action. If the waiting action is chosen by the ranking function, the vehicle waits
132 4 Review of the Literature Related to the Considered RDOPG Applications

at its current position. The relocation strategy is implemented by moving a vehicle

to the next identified request by the ranking function regardless whether this re-
quest is a real or a sampled customer. In doing so, if the next assigned request is a
sampled customer, the vehicle is relocated to sampled customer’s position. Compu-
tational results on the consensus and regret ranking function reveal that the waiting
and relocation strategy can significantly improve the attainable solution quality. Es-
pecially for problem instances that are highly dynamic and have many late customer
requests, the regret ranking function with the relocation strategy achieves the high-
est improvements compared to the earlier MSA variants that do not make use of
the waiting or relocating strategies. The authors conclude that both new presented
strategies achieve the best performance for different objectives. Specifically, while
the waiting strategy is beneficial when the minimization of total travel time is the
main objective, the relocation strategy performs best if the primary aim is the mini-
mization of unserviced customers within given hard time windows.
In all of the approaches described above, the existence and availability of a prob-
ability distribution that represents future request arrivals is assumed. In Bent and
van Hentenryck (2005), the authors propose techniques to learn distributions either
online from observed events during the transportation process or before the trans-
portation process from available historical data. In order to deal with dynamic rout-
ing problems in which request arrival rates change during the planning horizon, the
authors propose historical sampling. Specifically, request arrivals from the historical
data are utilized as sampled customers in MSA-based approaches. First computa-
tional results using the MSA-LC approach are performed on historical sample sizes
of 10 and 1,000 samples utilizing the consensus and regret ranking functions. The
number of unserviced customers is compared to the instances in which the original
distribution is available for sampling. From the obtained results, the authors con-
clude that historical sampling is an efficient approach for providing knowledge from
historical data. For future research, they propose to analyze impacts of the quality
of historical data on the solution quality as well as develop filtering techniques for
identifying relevant parts of the historical data.
In the Master’s thesis of Markus (2009), the author bases his research on the
described findings of Bent and Van Hentenryck. The author proposes discretiza-
tion of the considered service area so that sampled customers in the same subre-
gion are regarded as the same for calculations performed by the ranking functions.
This allows customer requests to be arbitrarily located within the considered service
area. Furthermore, in the regret ranking function the author discusses which requests
should be included in the calculations. For solving the sample scenarios, the author
proposes a solution method based on an attribute-based hill climber metaheuristic.
Computational experiments show that calculating the regret ranking function for
the requests on the same tour only and including sampled requests in these calcula-
tions performs best. Furthermore, the author mentions different ways of performing
the service activity at a sampled customer. For example, a sampled customer can
be treated like a real request by traveling to the sampled customer, waiting for the
beginning of its time window, and also simulating the service activity. Another pos-
sibility is to simply move to the sampled customer and directly drive to the next
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 133

customer afterwards. As an outlook on future research, the author suggests several

improvements for the algorithm and the model, e.g., considering additional sources
of uncertainty.
Besides research work based related to OSCO, there are other approaches that use
sampling-related techniques to utilize stochastic knowledge. Hvattum et al. (2006)
consider a dynamic routing problem that is based on a real-world application in
one of the leading distribution companies in Norway. The problem is modeled as
a single-depot DVRP with Time Windows where a fleet of heterogeneous capaci-
tated vehicles need to pick up goods at requests that arrive dynamically during the
day. Each request has a specific demand and a hard time window. Uncertainty arises
from dynamically arriving requests as well as from time windows and demand not
known with certainty. In the approach, probability distributions for the mentioned
sources of uncertainty that can be used as stochastic knowledge are assumed to be
available and vehicle en-route diversion is not allowed. The considered hierarchical
objective function consists of the topmost aim of minimizing the number of utilized
vehicles and second the minimization of total travel distance. In order to model the
problem, a two-stage stochastic formulation is presented that is similar to SVRP for-
mulations. This two-stage model is then extended to a multistage stochastic model.
The time horizon of the considered transportation process is divided into a predeter-
mined number of intervals and in each interval, a new stage is generated in which
the new requests that have arrived during the previous interval are integrated. Hence,
new arriving requests are queued until the beginning of the next interval. The prob-
lem is solved by a proposed Dynamic Scenario Hedging Heuristic (DSHH) which
gradually constructs the solution during the execution of the transportation process.
Specifically, in each interval, the DSHH first generates a set of sample scenarios us-
ing the existing distributions. Each of these sample scenarios includes the currently
known real requests and a set of expected future requests. In a second step, the sam-
ple scenarios are solved by a least cost insertion heuristic. After the tour plans for
all sample scenarios have been generated, requests that are frequently scheduled in
the generated tour plans within the upcoming time interval are identified and se-
lected. This selection stops when one of the following two conditions is fulfilled:
(a) all requests that have not yet been selected are only infrequently considered in
the next time interval, i.e., below a predefined percentage in the solutions of the
sample scenarios, or (b) a specific number of requests depending on the minimum
number of vehicles used in the sample scenarios has been found. Afterwards, the
sample scenarios that do not consider the selected requests in the next interval are
solved again while enforcing the integration of these requests. At this time, the re-
quests that are considered in the next time interval have been chosen. In the next
step, the tour plan generation which determines the assignment and sequencing of
the chosen requests to the vehicles is performed. For this purpose, a concept that is
related to the previously described consensus ranking function is used. In the gen-
erated solutions of the sample scenarios, it is counted how often each request is
assigned first to a specific vehicle and the most frequently identified request-vehicle
pair is added to the generated tour plan. This selection process is continued until
all of the chosen requests have been scheduled in the generated tour plan. In com-
putational experiments, the performance of the proposed DSHH is compared to a
134 4 Review of the Literature Related to the Considered RDOPG Applications

myopic dynamic heuristic (MDH) that does not use stochastic knowledge and gen-
erates the tour plan using a local search heuristic in each interval. For reference
purposes, an offline version of the problem that has access to all information be-
fore the execution of the transportation process is also considered. In this offline
setting, the solution is solved using the same local search heuristic as in the MDH
approach. In comparison to the MDH, the authors show that the DSHH is able to
reduce travel distances up to 15 % while using a comparable number of vehicles.
If more computational time is available by choosing longer intervals between solu-
tions, the utilization of more samples can yield a better solution quality. The authors
conclude that in a scenario with loose time windows (ten hours) using an interval
length of two hours produces the best results in terms of travel distance. In contrast,
in case of tighter time windows with a length of three to five hours, the DSHH pro-
duces longer travel distances. The authors presume this to happen due to the more
elaborated local search utilized in MDH which can take advantage of finding good
solutions in more constrained problems compared to the insertion heuristic used in
DSHH. For future work, the authors suggest giving a detailed evaluation about why
the DSHH tends to sometimes use an extra vehicle compared to the MDH.
In the following research work by Hvattum et al. (2007), the authors consider the
same routing problem as the one in the work presented before. They address the pre-
viously mentioned issue of additionally used vehicles in the DSHH. They explain
that choosing the request to be serviced next according to the decision in the major-
ity of the sample scenario solutions may require the use of additional vehicles when
a request which was only considered in the minority of the sample scenarios ar-
rives. In order to deal with this issue, they extend the DSHH to a Branch-and-Regret
Heuristic (BRH). Similar to the DSHH, samples in this approach are generated from
given probability distributions. In comparison to the DSHH, a difference is that de-
cisions produced by consulting the sample scenarios in the BRH are only imple-
mented if it is unlikely that they will lead to suboptimal effects. The proposed BRH
works as follows. After producing a solution for each sample scenario, the heuris-
tic iteratively merges the individual sample scenario solutions by evaluating certain
criteria on which to branch, for example if a request should be serviced during the
next interval. This concept is similar to stochastic B&B (cf., e.g., Klein Haneveld
and van der Vlerk 1999) except for the fact that only one path in the branching tree
is considered by the BRH. DSHH and BRH differ mainly in two points. First, for
each request, the DSHH decides whether the request will be serviced in the next
interval based on a percentage threshold. In the BRH this type of decision is made
by solving all sample scenarios twice; once with the request and once without the
request. If the best decision is to service the request in the next interval, it is added
to the set of chosen requests. Hence, in comparison to the DSHH, the threshold pa-
rameter is not required. According to the authors, this is advantageous because the
value of this parameter can be significantly problem-dependent. Second, in order
to choose the next request i for a vehicle k, the assignment proposed in the DSHH
is evaluated in comparison to alternative assignments. The authors evaluate three
different variants: (a) same request i on a different vehicle, (b) same vehicle k with
another request, and (c) the setting of the previous two variants which can be found
4.5 Flexibility in Dynamic Routing Using Stochastic Knowledge 135

4.6 Related Objective Functions

As mentioned in Sect. 3.1, the objective pursued in the considered RDOPG applica-
tions consists of minimizing customer inconvenience in order to maximize quality
of customer service. When dealing with such quality aspects, the objective function
often contains some type of penalty function that makes solutions with poor service
quality less attractive. As mentioned in Sects. 2.4.12.5 and 2.4.12.7, such a penalty
function can contain fixed and variable components. Fixed lateness penalties can
also be used to model DVRPs with hard time window constraints where customer
requests are considered to be lost when they are not serviced within given time
windows (cf. Bent and van Hentenryck 2004b). While fixed lateness penalties are
completely determined by their parameter value, the definition of variable penalty
costs also includes a definition regarding the evolution of the lateness-dependent
penalty. In the marketing literature, two main approaches for modeling customer
inconvenience depending on elapsed request response times can be found. While,
138 4 Review of the Literature Related to the Considered RDOPG Applications

bi ai bi ai bi
Sexton and Bodin (1985) Sexton and Choi (1986) Dumas et al. (1990)

Fig. 4.4 Different inconvenience and penalty functions (cf. Desrosiers et al. 1995)

PiMAX

Ei Ai Bi Li Ei Ai Bi Li Ei Ai Bi Li

(a) (b) (c)

Fig. 4.5 Customer inconvenience functions considered by Fagerholt (2000, 2001)

among others, Davis and Maggard (1990) propose that customer inconvenience in-
creases linearly with resulting response time, other authors such as Kristensen et al.
(1992) derive a quadratic dependency.
In our definition, a customer quality-related objective function is called an incon-
venience function if penalties directly occur with the arrival of a request in the sys-
tem while in a lateness function, penalties only occur when service has not occurred
within specified time windows. Moreover, depending on the application, quality-
related objectives can also be used to penalize the unwanted case in which a request
is serviced too early, i.e., before the start of its time window. In what follows, rout-
ing problem-related approaches known in the literature that consider inconvenience
and lateness functions are presented.
In the routing literature of approaches for static routing problems, different types
of inconvenience and penalty functions can be found. For example in DARPs,
considered service quality consists of time differences experienced by customers
that demand to be picked up at a specified time point (cf. Cordeau et al. 2007a).
Hence, passengers should neither be picked up too early nor too late. Furthermore,
the length of the ride time experienced by passengers is another quality criteria.
Desrosiers et al. (1995) mention three different types of penalty functions (see
Fig. 4.4). The first one described by Sexton and Bodin (1985) denotes a linear de-
creasing inconvenience function with a hard time window at time point bi . The
second function describes a lateness function proposed by Sexton and Choi (1986).
Here, too early or too late service, i.e., service before ai or after bi , is allowed but
linearly penalized. In the third function, an inconvenience function by Dumas et al.
(1990) is presented that quadratically penalizes dispersion from a desired time point
of service while also limiting this dispersion in both temporal directions by hard
time windows.
Besides vehicle-based routing problems, aspects of the service quality are also
considered in other routing applications, for example multi-ship PDPs (cf. Fagerholt
4.6 Related Objective Functions 139

2000, 2001). In this static problem, outer and inner time windows exist. Specifically,
the outer time window of request i defined by [Ei , Li ] can be seen as a hard time
window before or after which no service is possible and, instead, high penalty costs
denoted by PiMAX occur. As illustrated in Fig. 4.5, the inner time window [Ai , Bi ]
represents a soft time window so that service that occurs too early or too late results
in an either (a) linear, (b) fixed, or (c) quadratic penalty.
In dynamic routing problems in which new requests arrive during the execu-
tion of the transportation process, a linear customer inconvenience function is used
among others in the described approaches of Psaraftis (1980), Bertsimas and van
Ryzin (1991, 1993a, 1993b), Papastavrou (1996), in the approaches of Gendreau
et al. (1999a), Swihart and Papastavrou (1999), Richter (2005), Attanasio et al.
(2007), Schmid (2012), and in van de Klundert and Wormer (2010). Moreover, a
linear lateness function is used in the work of Larsen (2000), Yang et al. (2004),
Ichoua et al. (2006), Branchini et al. (2009), and Ghiani et al. (2009). In the work of
Krumke et al., the authors also consider lateness functions in a DTRP encountered
by the German Automobile Club ADAC. For example, in Krumke et al. (2002), a
linear lateness function is considered whereas in other publications, e.g., in West-
phal and Krumke (2008), a quadratic lateness function is used in order to penalize
longer lateness more. The authors formulate the problem as a set partitioning model
that is solved by a column generation approach. Although the considered problem
is strongly NP-hard, the authors show that due to a low request per vehicle ratio in
tested real-world instances, limited tour lengths of vehicles are often present in opti-
mal solutions of the considered instances. Hence, the evaluated real-world problem
instances can be solved exactly or with a very small gap compared to a proposed
lower bound within a couple of minutes. Moreover, the authors report that for all
tested instances the optimal objective function value is at most 0.494 % above the
lower bound so that it can be concluded that the proposed lower bound is quite
tight and hence of high quality. Test instances of 106 requests and 84 vehicles up
to 775 requests and 211 vehicles are tested and the performance of the proposed
exact solution method is evaluated against different variants of a GA provided by
an industrial partner. Since, as mentioned above, the computational time required to
solve the problem instances which occur during the transportation process is very
important in real-time environments, the authors test the achieved solution quality
after computational times of 5, 15, 30, 60, and 120 seconds. The computational re-
sults show that the proposed exact solution method is able to perform significantly
better than the GA in terms of solution quality but also with regard to required
computational times. For all of the tested instances, the exact method provides so-
lutions with an objective function value of at most 1 % above the optimal objective
function value or also often directly an optimal solution within 5 seconds. In the
bigger test instances, the evaluated GAs still have a gap of more than 5 % after
600 seconds computational time. In Westphal and Krumke (2008), the authors im-
prove the approach for the previously presented real-world problem by proposing
a more sophisticated pruning scheme for the search tree used in the column gener-
ation approach. Computational experiments show the effectiveness of the new ap-
plied scheme. Specifically, compared to the old pruning method, the new pruning
140 4 Review of the Literature Related to the Considered RDOPG Applications

approach requires only 38.2 % of the computational time and explores only 24 % of
the nodes in the search tree.
Besides customer quality of service criteria, driver inconvenience aspects are
considered in the objective function in the work of Savelsbergh (1992) and As-
cheuer et al. (2000) by minimizing the completion time of vehicle tours (see also
Sect. 2.4.12.4).

4.7 Other Relevant Factors for RDOPG Applications

In this section, we deal with other relevant factors that are existent in the consid-
ered RDOPG applications in this book. First, approaches that evaluate the impact of
allowing en-route diversion of vehicles are presented. This is followed by a descrip-
tion of approaches which utilize a road network to map vehicle movements more
realistic. In the next section, approaches which describe simulator concepts which
can be used to simulate dynamic transportation processes are illustrated. Further-
more, research work describing real-time control approaches are introduced. Since
the RDOPG applications are carried out in an urban area, city logistics-related de-
cision support systems and communication technologies are also presented.

4.7.1 En-Route Diversion of Vehicles

As mentioned in Sect. 2.6.3, the utilization of technical equipment makes it possi-

ble to locate vehicles and to track their movements in real-time. Hence, when a new
request arrives close to the route that a vehicle is currently traveling to its next as-
signed request, the application of vehicle en-route diversion can redirect the vehicle
to service this new request first. As a result, the overall service quality is increased
if an objective considering the minimization of customer inconvenience is pursued.
Furthermore, allowing vehicle en-route diversion can also be advantageous when
the objective of minimizing total travel distance is considered since it allows for the
reduction of required detours. In Regan et al. (1995, 1998), different vehicle en-route
diversion strategies are investigated and positive effects on system performance in
truckload routing applications are shown. Besides, vehicle en-route diversion is only
considered in a few approaches, e.g., in Ichoua et al. (2000), Richter (2005), Attana-
sio et al. (2007), Branchini et al. (2009), and van de Klundert and Wormer (2010). In
Ichoua et al. (2000), a long-haul courier service routing problem is considered. The
objective consists of minimizing the weighted sum of travel distance and customer
lateness. In order to adapt the tour plans, the Parallel Tabu Search metaheuristic of
Gendreau et al. (1999a) is used and adapted. Compared to the original approach in
which no vehicle en-route diversion is allowed, computational experiments show
that the solution quality can be significantly improved by using vehicle en-route
diversion. Moreover, as mentioned in Sect. 4.2, Richter (2005) concludes that al-
lowing vehicle en-route diversion can significantly reduce customer inconvenience
and therefore increase quality of service.
4.7 Other Relevant Factors for RDOPG Applications 141

4.7.2 Road Network

As can be concluded from the literature review, only a few of the described pro-
active real-time control approaches make use of a road network to map vehicle
movements and to calculate travel distances between relocations, such as Bell et al.
(1983), Fleischmann et al. (2004b), Tarantilis et al. (2004), Thomas and White III
(2004), and Schmid (2012). The utilization of a road network is important for de-
termining accurate travel times from current vehicle positions to request locations
by taking the existing road infrastructure into account. The utilization of a real road
network in practical applications is especially useful when vehicle en-route diver-
sion is allowed. When considering transportation problems in real road networks
that cover a big service area, improvements with regard to the required computa-
tional time compared to the Dijkstra algorithm can be found among others in Bauer
et al. (2008) and Schultes (2008).

4.7.3 Simulator

In order to empirically assess the performance of dynamic real-time control ap-

proaches, mathematical models are used to represent the system, i.e., the considered
transportation process, in terms of logical and quantitative relationships. In general,
“a system is defined to be a collection of entities, e.g., people or machines, that
act together toward the accomplishment of some logical end” (Schmidt and Taylor
1970). Since a system changes over time, its current state is defined to be a “collec-
tion of variables necessary to describe a system at a particular time, relative to the
objectives of a study” (Law and Kelton 2000, p. 3). Simulation-based methods are
often used if mathematical models are too complex to be studied analytically. Using
simulation allows for “numerically exercising the model for the inputs in question
to see how they affect the output measures of performance” (Law and Kelton 2000,
p. 5).
Basically, for modeling the advance of the simulation time in the considered sys-
tem, two types of discrete simulators exist: (a) simulators in which the simulation
time is incremented by a fixed amount of time and (b) simulators in which the simu-
lation time is directly advanced to the time point of the next relevant event (cf. Lar-
son and Odoni 1981, p. 590). In approaches of the latter type, the “[simulation]
clock is advanced only to those instants in time when certain important (for the sim-
ulated system) events take place. Such events might be the arrival or departure of
customers in a queueing system, or the occurrence of a fire alarm in the simulation
of a fire department” (Larson and Odoni 1981, p. 590). According to the authors,
this is also known as advancing the clock simulation. The used term discrete rep-
resents the fact that the state variables of the simulated system only change their
values at specific discrete time points. These time points are determined by events
that do not consume any time (cf. Biethahn et al. 1999, p. 52, Law and Kelton 2000,
142 4 Review of the Literature Related to the Considered RDOPG Applications

p. 6). Hence, by using events, all state changes occurring in the considered dynamic
system are covered.
In dynamic routing problems, applying a discrete event-based simulator is often
appropriate since the relevant system state changes whenever an event occurs. The
advantage of event-based simulators is that no decision has to be taken about time
precision as is the case in simulators which increase the simulation time in fixed time
intervals. Hence, in event-based simulators, the simulation time is always directly
set to the point in time of the next event while time points at which no relevant event
occurs are skipped so that less computational time is required for simulating the
system itself (cf. Larsen 2000, p. 111).
A discrete event-based simulator needs to distinguish between endogenous and
exogenous events during the simulation (cf. Mihram 1972, p. 219 et seq.). The first
type of events are generated by the system, e.g., when a vehicle has finished servic-
ing its current request and requires further orders. These events may trigger other
activities to be performed. For example, in approaches which iteratively generate
the tour plan by calculating the next request for a vehicle when this is demanded, an
activity is required at the occurrence of such an event. The second type of events is
represented by exogenous events which come from outside into the system, such as
the arrival of a new request which are required to be integrated into the transportation
process by an appropriate adaptation activity. Furthermore, when vehicle en-route
diversion is allowed and a real road network is used, the simulator framework is
more complex since vehicle movements need to be modeled explicitly in order to
enable a determination of their current position at any point in time. More infor-
mation about simulation environments can be found among others in Regan et al.
(1998), Larsen (2000), Ichoua et al. (2000), Gendreau et al. (2006), and Barceló
et al. (2007). More information about the discrete event-based simulator which is
developed to simulate the considered RDOPG applications in this book will be pre-
sented in Sect. 8.1.

4.7.4 Real-Time Control

As mentioned in Sect. 2.5, in contrast to static routing problems that only consider
the initial tour plan generation before the execution of the transportation process,
dynamic routing problems are required to be controlled in real-time in order to
monitor the execution of the transportation process and to adapt the tour plan when
necessary. Giaglis et al. (2004) mention that, once a tour plan is generated, classic
planning systems for static VRPs do not take into account the possibility of changes
in the relevant information that may require an adaptation of the constructed tour
plan. Hence, the only option for these classic systems is to discard the current tour
plan in execution and generate a completely new one. This procedure is expected
to raise problems in real-world applications since vehicles may get completely dif-
ferent tours assigned to them than were assigned before. Furthermore, it may not
be possible to consider information about current vehicle positions and statuses as
4.7 Other Relevant Factors for RDOPG Applications 143

well as already loaded goods in such systems. Therefore, the application of real-
time control approaches that are able adapt the current tour plan in order to integrate
changes in the relevant information caused by dynamic events is advantageous.
An important aspect in real-world routing applications is that while tour plans are
adapted according to changes in the relevant information, the transportation process
continues during this adaptation. Since most adaptation methods require an amount
of time that cannot be considered as zero time1 anymore, one important aspect of
real-time control approaches is that they need to take the concurrency of tour plan
execution and tour plan adaptation into account. Hence, since utilized adaptation
methods usually consider individual problem instances that do not change during
the adaptation, it has to be ensured that decisions which are taken during the adapta-
tion can still be made after the adaptation has been completed. This is, for example,
accounted for in Ichoua et al. (2000) which utilizes the Parallel Tabu Search meta-
heuristic by Gendreau et al. (1999a). In the proposed real-time control approach the
current system state at time point t is anticipated to its state in δt time units. In-
compatible tour plans in the adaptive memory are then removed and the remaining
time is used for the improvement method. The authors state that the choice of the
length of δt can be of high importance. Specifically, while a large value of δt may
miss adaptation possibilities, a too small value of δt may prohibit the improvement
method from unfolding its full potential. The authors evaluate different strategies for
setting δt, in particular for setting δt to a fixed value or for dynamically determin-
ing δt according to the system situation at t. Moreover, Hvattum et al. (2006) and
van de Klundert and Wormer (2010) consider the research question about choos-
ing the length for the time intervals in which tour plans should be adapted. As
can be seen from the different results in these research work, this value is highly
application-dependent. The utilized value in the real-time control approaches of this
book will be described in Sect. 8.2.1.
Another real-time control approach that considers the concurrency of tour plan
execution and tour plan adaptation can be found in Bock (2010). The author consid-
ers a real-time control concept that pre-simulates the system situation which exists
at the beginning of each adaptation activity for a specified time period denoted as
the anticipation horizon. After the pre-simulation, the remaining time of the an-
ticipation horizon is used to adapt the tour plan based on the problem data of the
anticipated system state. Since the real-time control approaches developed in this
book are based on this real-time control concept, more information about it will be
described in Chap. 5. Moreover, when uncertain travel times are considered, real-
time control approaches can be extended by methods for estimating vehicle arrival
times proposed by Cathey and Dailey (2003) and Jula et al. (2008).
In Zeimpekis and Giaglis (2005), the authors propose a general architecture of
real-time control approaches applicable for DVRPs (see Fig. 4.6). It can be de-
scribed as follows. In the monitoring state, the current system state which includes

1 Zero time defines a time period during which the system state of the considered dynamic process

can be assumed to remain constant. Note that the length of this time period is highly application-
dependent—in the considered RDOPG applications its length is one second at most.
144 4 Review of the Literature Related to the Considered RDOPG Applications

New dynamic
Current events pending? Adapted
tour plan yes Adaptation of Realization tour plan
Monitoring
the tour plan and execution
no

Feedback (from vehicles)

Fig. 4.6 General architecture of real-time control approaches (cf. Zeimpekis and Giaglis 2005)

the current vehicle positions and their status as well as the arrival of newly arrived
dynamic events. When dynamic events that require an adaptation of the tour plan
occur, the solution method that adapts the tour plan with regard to the occurred
changes in the relevant information is invoked. In this step, also measures for con-
sidering the concurrency of tour plan execution and tour plan adaptation, e.g., the
pre-simulation of the tour plan, are considered. After the execution of the solution
method has been completed, the adapted tour plan is given into execution. Feedback
can be provided from the vehicles after the changes in the tour plan have been com-
municated to them. Besides the acknowledgment of the transmitted changes, such
a feedback can also consist of various other elements. For example, if the approach
allows the rejection of requests, drivers may reject assigned newly requests. Further-
more, it may happen that vehicles cannot execute the adapted tour plan as planned,
for example when vehicles have a different position than has been anticipated during
the adaptation. In such cases, another invocation of the adaptation procedure can be
triggered.

4.7.5 City Logistics-Related Decision Support Systems

and Communication Technologies

Since the considered RDOPG applications can be considered as transportation prob-

lems in urban regions, research on city logistics reflects an important factor for in-
creasing the practicability of the proposed real-time control approaches. Larson and
Odoni (1981) provide detailed information about foundations of appropriate mathe-
matical models as well as suitable scientific methods for handling city logistic prob-
lems. Furthermore, the use of technological advances in routing applications has
been widely researched. This is advantageous in order to realize a real-time control
system at a central dispatching center, the current location of the vehicles needs to
be automatically and accurately updated and transmitted to the dispatching center
in real-time as mentioned in Sect. 2.6.3. The use of a bidirectional communication
system allows the automatic transmission of information between the vehicles and
the dispatching center. Several authors have conducted research on technical sys-
tems that are appropriate for this purpose. Among others, in Giaglis et al. (2004)
and Zeimpekis and Giaglis (2005), several information about available technical in-
frastructure, e.g., positioning systems such as GPS and data transfer systems such as
4.8 Conclusion of the Literature Review 145

GSM and GPRS, are presented. Furthermore, a possible technical implementation of

a real-time system is presented. The application of GPS technology in transportation
problems is also investigated in Mintsis et al. (2004). In addition, the use of infor-
mation about expected changes in travel times is advantageous. In order to obtain
such information, floating car data can be processed (cf. Ehmke et al. 2009, 2010).
Furthermore, traffic sensors which are commonly installed at highways and major
roads allow for an automatic identification of traffic jams (cf. Ghiani et al. 2003 and
Attanasio et al. 2007).
In Tarantilis et al. (2004), a decision support system for an open DVRP is pro-
posed that incorporates a real road network from a geographical information sys-
tem (GIS) and GPS-assisted vehicle location techniques in order to assist the dis-
patcher during the execution of the tour plan process. By using the implemented
decision support system, dynamic events arising from urgent dynamic customer re-
quests, request cancellations, or vehicle damages are automatically handled and the
tour plan is dynamically adapted according to occurred dynamic events. For han-
dling city logistics with vehicle delays and vehicle breakdowns, Zeimpekis et al.
(2005) propose a dynamic fleet management system that monitors the location of
vehicles in real-time and recommends re-routing activities.
In the work of Barceló et al. (2007), a framework for modeling city logistics
applications is developed. This framework includes several data connectors in order
to retrieve necessary data as well as transportation-specific components such as a
simulator and a routing optimization module and a graphical user interface. The
framework is successfully applied in case studies conducted in two Italian cities.

4.8 Conclusion of the Literature Review

According to the presented literature review, it can be observed that there are a large
variety of pro-active approaches in the literature which integrate various strategies
for providing an increased flexibility for the integration of dynamically arriving re-
quests during the execution of the transportation process. As can be seen from the
provided computational results, the presented approaches are able to efficiently uti-
lize available stochastic knowledge so that the solution quality can be significantly
improved compared to using deterministic real-time control approaches. However,
with regard to applying the presented approaches in RDOPG applications, differ-
ent issues arise that need to be additionally considered. First, the considered ob-
jective in the RDOPG applications dealing with the minimization of customer in-
convenience has only been assessed in a few dynamic pro-active real-time routing
approaches (see Sect. 4.6). As mentioned in Sect. 3.1, in the considered RDOPG ap-
plications, both a linear and a quadratic customer inconvenience are evaluated in this
book in order to evaluate the impacts of different customer inconvenience functions
on the attainable solution quality. Furthermore, additional prohibitive penalty costs
occur if a defined maximum response time is not fulfilled which also prevents an
indefinite deferment of customer requests (see aspect 8 of Psaraftis’ approach pre-
sented in Sect. 2.5.2.2). Furthermore, it enables the modeling of routing problems in
146 4 Review of the Literature Related to the Considered RDOPG Applications

which the maximization of the number of dynamically arriving requests within their
hard time windows is pursued. As can be seen in the described literature review, this
is an often considered scenario. Second, since the RDOPG applications represent
inner-city problems, the utilization of a real road network is advisable. Note that the
utilization of a road network is only considered in a few approaches in the literature
while many approaches assume Euclidean connections between the relevant loca-
tions. Moreover, as reported among others by Richter (2005) and van de Klundert
and Wormer (2010), using vehicle en-route diversion is likely to allow a significant
improvement with regard to solution quality when customer quality-related aspects
are considered. Hence, vehicle en-route diversion should be implemented in routing
applications in which a real road network is used and the attainable improvements
should also be assessed. To the best of our knowledge, in the literature, this has only
been performed in Ferrucci (2006) and Attanasio et al. (2007). Furthermore, allow-
ing vehicle en-route diversion may increase driver inconvenience due to the fact that
drivers are obliged to quickly react to occurring changes in their tour while they are
traveling to their next assigned request. We are not aware that this issue has been
dealt with yet in dynamic routing problems. In order to deal with this research ques-
tion, we propose a new approach that additionally considers driver inconvenience
aspects during the execution of the transportation process.
In pro-active approaches existing in the literature, another issue often arises when
utilizing stochastic knowledge. Specifically, while developed pro-active real-time
control approaches offer sophisticated methods for integrating available stochastic
knowledge into the transportation process in order to efficiently adapt the tour plan
according to expected future requests, in these approaches it is often assumed that
available stochastic knowledge is given by prescribed probability distributions. Fur-
thermore, it is often assumed that the dynamic requests exactly arrive according to
these prescribed probability distributions. Note that these are strong assumptions
which are often not present in real-world applications. However, since past request
information is available in many real-world applications, the development of fore-
casting methods which are able to automatically extract stochastic knowledge from
past request information is advisable. In the literature, only few approaches that
present such forecasting methods can be found, e.g., in Bent and van Hentenryck
(2005), Attanasio et al. (2007), and Schmid (2012) but these approaches deal with
applications which have significant differences to the considered RDOPG applica-
tions. Moreover, Bent and van Hentenryck (2005) and Ghiani et al. (2009) suggest
that future research should analyze the impact of the quality of the available past
request information on the performance of pro-active real-time control approaches.
This research question is addressed in this book by determining the structural qual-
ity of a request data set according to its structural distortion and structural diversity
(see also Chap. 8). Specifically, it is advisable to assess the quality of available past
request information with regard to characteristics of this information that are as-
sumed in the forecasting methods and in the pro-active real-time control approach.
For example, if request arrivals based on Poisson processes which are constant from
day to day are assumed to exist, it is advisable to verify whether the available past re-
quest information is likely to fulfill this assumption. The forecasting method for gen-
erating stochastic knowledge from past request information proposed in this book in
4.8 Conclusion of the Literature Review 147

Chap. 6 checks whether the assumed characteristics are likely to exist in the avail-
able past request information. For this purpose, the structural distortion of a request
data set is determined. Computational results on the efficiency of the proposed ap-
proach for identifying structural distortion in given request data sets are provided
in Sect. 8.6. Besides evaluating the structural distortion of a request data set, the
structural diversity of a request data set is defined and its impact on the attainable
solution quality is evaluated. For this purpose, request data sets comprising different
levels of exploitable information for generating stochastic knowledge are designed
(see Sect. 8.2.4) and two approaches for determining the structural diversity of given
request data sets are proposed. The first approach is a qualitative assessment which
makes use of a visual analysis (see Sect. 8.8). The second approach denoted as the
degree of structural diversity (dosd) is a quantitative approach for determining the
structural quality of given request data sets. It is proposed in Sect. 8.9 and its effi-
ciency is shown by using the computational results which have been attained on the
evaluated request data sets.
Two other aspects are connected with the generation of dynamically arriving re-
quests that remain unclear in some pro-active approaches. The first aspect is about
whether dynamically arriving requests appear at arbitrary locations in the considered
service area or if they can only appear at few predefined locations which is similar
to the modeling of uncertain customers in classic SVRPs (see Sect. 2.5.3.2.1). The
second aspect concerns changes in the probability distributions of request arrivals,
in particular changes in request arrival rates during the planing horizon. In order
to classify and compare differences in the structural diversity of individual consid-
ered problem settings, it is advantageous to know whether the utilized distributions
for generating dynamic events such as newly arriving requests remain constant or
whether they change and, if so, in which dimensions, during the planning horizon of
the considered transportation process. Note that this does not only apply to global
changes of request arrival rates which occur during the considered planning horizon.
In particular, it is especially important to take into account changes in the request ar-
rival rates which occur in individual subregions of the considered service area since,
as will be shown, these changes have a significant impact on the structural quality
of request data sets. The proposed dosd presented in Sect. 8.9 explicitly considers
such characteristics of given request data sets.
This book contributes new approaches for dealing with the mentioned aspects.
The proposed approaches and computational studies provide information on defin-
ing the structural quality of arbitrary request data sets and the correlation between
their structural quality and attainable improvements. These contributions provide
promising starting points for new avenues of further research which will be dis-
cussed in Sect. 9.2.
Chapter 5
A New Deterministic Real-Time Control
Approach for RDOPG Applications

In this chapter, the deterministic real-time control approach is introduced. It is used

to coordinate the transportation process of the considered RDOPG applications,
which were presented in Chap. 3, in real-time during its execution. First, a gen-
eral categorization of different control concepts in logistic and production systems is
described followed by the deterministic real-time control approach. It solves the dy-
namic problem by considering consecutive static snapshots, i.e., problem instances
which represent the current system state of the transportation process during its
execution. In order to represent these problem instances in the deterministic real-
time control approach, the mathematical model of individual problem instances is
presented next. Moreover, a characterization of the deterministic real-time control
approach according to the general categorization introduced in Sect. 5.1 is given.
Finally, with regard to the considered layers of transportation processes described
in Sect. 1.2, the utilization of the real-time control approach in RDOPG applications
is presented.

5.1 General Categorization of Different Control Concepts

in Logistic and Production Systems

Control concepts applied to coordinate dynamic processes in logistic and production

systems share common characteristics. According to Bock (2004, p. 67 et seq.), such
control concepts can be classified according to the following five attributes:
1. Adaptation-frequency: This attribute specifies the points in time at which an
adaptation is performed.
• Disturbance-oriented concepts: The adaptation is started after the occurrence
of a disturbance, i.e., a dynamic event.
• Event-oriented concepts: Specific predefined events start the adaptation.
• Periodic concepts: The adaptation is performed in predefined but longer time
intervals.

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 149

DOI 10.1007/978-3-642-33472-6_5, © Springer-Verlag Berlin Heidelberg 2013
150 5 A New Deterministic Real-Time Control Approach

• Continuous concepts: The adaptation is continuously performed taking into

account almost all or, in the ideal case, all decisions that have not yet been
implemented.
2. Adaptation-synchrony: Using this attribute, the continuation of the process dur-
ing its adaptation is defined.
• Planning priority: It is assumed that the considered process is stopped during
the adaptation. By this assumption, it is expected that the system state present
at the beginning of the adaptation still exists after the adaptation has been per-
formed. Since processes usually cannot be stopped at arbitrary points in time
during their execution, this assumption is usually only valid for adaptation
methods which can guarantee adaptation of the process in a time period which
can be still considered as zero time (see also Sect. 4.7.4).
• Process priority: It is accounted for that decisions are executed in the process
during its adaptation. Hence, certain parts of the plan currently in execution
are not able to be changed during the adaptation. This needs to be considered
in the real-time control concept, for example by anticipating the system state
that is assumed to exist after the adaptation stops and by using this anticipated
system state in the adaptation method.
• Simultaneity of process execution and plan adaptation: In this constellation,
plan adaptations are carried out parallel to the process execution. Note that this
is an ideal setting in which all decisions which are currently not in execution
can be adapted.
3. Adaptation-technique: This attribute defines the methods that are used to per-
form the adaptation.
• Priority rules: The plan is adapted by greedy approaches which can usually be
carried out in zero time.
• Improvement procedures: Plan adaptation methods such as local search or
metaheuristics are applied. Note that due to their complexity, the utilization
of such methods usually cannot be performed in zero time.
4. Adaptation-range: This attribute represents the range of decisions that can be
changed in the plan during its adaptation.
• Complete adaptation range: All decisions that have not yet been performed in
the plan at the beginning of the adaptation can be changed.
• Usable range: A reduction of the adaptation range by the time period which is
required to transmit and realize plan adaptations in the process is considered.
Hence, such decisions cannot be changed.
• Time-limit depending range: A reduction of the adaptation range by a specified
time limit is performed. This time limit may consider the time period utilized
for performing the adaptation process and transmitting changes. With this set-
ting, the concurrency between process adaptation and process execution can
be considered.
• Selected range: Predefined subsets of decisions are excluded from adaptation.
For example, all decisions that are made within the next hour are excluded.
5.2 The Deterministic Real-Time Control Approach 151

5. Adaptation-duration: This attribute defines the computational time which is

available for the applied adaptation method.
• Time limit based: The available time is defined by a time limit. Hence, it is
known in advance when the adaptation will be complete.
• Zero time: The available computational time is defined by a time period that
can be interpreted as zero time, i.e., no noticeable changes occur in the con-
sidered process during the adaptation.
• Infinite time: The adaptation method is continuously applied. Note that the
time points in this setting at which performed adaptations are applied in the
considered process are required to be specified.

5.2 The Deterministic Real-Time Control Approach

In this section the deterministic real-time control approach which is used to solve
the problem represented by RDOPG applications is proposed. As mentioned in
Sect. 3.1, in the considered problem, a set of R customer requests R = {1, . . . , R}
which dynamically occur during the day need to be serviced by a set of K homo-
geneous vehicles K = {1, . . . , K} which are originally located at one central de-
pot. During the execution of the transportation process, the current time is denoted
by τ . When the execution of the transportation process starts, it holds that τ = 0.
Let ria denote the arrival time of request i in the system. Note that this arrival time
is not known to the real-time control approach in advance. The set of requests
that are known in the real-time control approach at time point τ is then defined
by Rτ = {i ∈ R|ria ≤ τ }. Since the real-time control approach only utilizes informa-
tion currently available while not making use of information about expected future
requests but provides additional tour plan flexibility by utilizing a waiting strategy,
it is classified as a deterministic real-time control approach (see Sect. 2.5.3.2). In
Chap. 6, it will be extended to a pro-active real-time control approach which uti-
lizes stochastic knowledge.
In what follows, the applied real-time control concept is described first. After-
wards, the update handling procedure consisting of activities which are carried out
in each step of the real-time control approach are presented. Moreover, the vehicle
scheduling strategy as well as the idle vehicle waiting strategy which are utilized in
this deterministic real-time control approach are introduced.

5.2.1 The Applied Real-Time Control Concept

Since the considered RDOPG applications are dynamic, the existing system situa-
tion changes during the day. Hence, the tour plan in execution needs to be adapted
according to incoming dynamic events. As mentioned in Chap. 3, dynamism arises
in the RDOPG from newly arriving requests. In order to integrate these dynamically
152 5 A New Deterministic Real-Time Control Approach

arriving requests, the tour plan currently in execution is coordinated during the exe-
cution of the transportation process by the applied real-time control concept. How-
ever, since the considered problem is NP-hard in the strong sense (see Sect. 3.2.2),
sophisticated tour plan adaptations cannot be carried out by zero time heuristics.
Therefore, appropriate real-time control approaches which consider relevant practi-
cal aspects of dynamic routing problems have to modify the tour plan simultaneous
to its execution, i.e., while the vehicles are continuing to perform assigned activities.
In order to fulfill this requirement, the proposed real-time control approach im-
plements a concept related to the one described by Bock (2010). In this concept,
the real-time control of the considered transportation process during its execution
is conducted by solving a series of individual static problem instances. They arise
from system snapshots of the ongoing transportation process that are generated at
intervals of t a time units each during the transportation process. In what follows,
these intervals are denoted as anticipation horizons. Note that each static problem
instance represents the current system state of the ADFSVRPSTW as described in
Sect. 3.2.1. In order to simplify the following explanations, we designate the tour
plan that is currently in execution as the relevant tour plan. It is denoted as Pτr , where
the index τ denotes the current time at the beginning of the considered anticipation
horizon. Analogously, Pτt defines a theoretical tour plan that is optimized during
this anticipation horizon utilizing a specified solution method. Additionally, Pτbt de-
notes the best theoretical tour plan found during the performed optimization.
At the beginning of each anticipation horizon, the corresponding static problem
instance is generated. In order to consider the concurrency of the tour plan adap-
tation and tour plan execution, the static problem instance at τ is derived from the
future system state which is expected at the end of the current anticipation horizon.
This future system state is attained by pre-simulating Pτr for t a time units. Conse-
quently, after applying the solution method, Pτt and Pτbt are directly implementable
at the end of the respective anticipation horizon. In order to enable a detailed map-
ping of system snapshots, each corresponding static problem instance comprises all
currently known unserviced requests as well as positions and statuses of vehicles.
Since dynamically arriving requests are buffered throughout each anticipation hori-
zon, this problem instance additionally comprises a list of these newly arrived re-
quests. After the pre-simulation that has consumed tτcalc time units is completed, the
remaining time of the current anticipation horizon denoted by t a − tτcalc is available
for optimizing Pτt by the solution method. Clearly, since Pτt contains only deci-
sions that are taken after the elapse of the current anticipation horizon, all decisions
that will be made by Pτr in the current anticipation horizon become unchangeable
at its beginning and thus are executed in the system as planned. Therefore, the as-
signment of a request to a vehicle becomes fixed if service at this request starts
according to Pτr before τ + = τ + t a . Since no uncertainties in travel times or service
times are considered in the RDOPG applications, in what follows, we assume that
the anticipated system state is in agreement with the actual system state at τ + . By
optimizing Pτt throughout the considered anticipation horizon, Pτbt is stored for a
possible subsequent replacement of Pτr . Note that due to the utilization of the pre-
sented pre-simulation concept, all existing tour plans are directly applicable at τ +
and are therefore feasibly interchangeable.
5.2 The Deterministic Real-Time Control Approach 153

Tour plan processing layer simulates the transportation process by executing the current relevant tour plan

Relevant tour plan P τr

Dynamic events Dτ
Queued dynamic events (i.e., requests)
arrived between τ - and τ which are
required to be integrated into the
transportation process at τ

Fixed decisions of P τt Changeable decisions of the theoretical plan P τt

Theoretical plan follows relevant tour Theoretical plan can be changed by the solution method invoked in [τ,τ +)
plan in this anticipation horizon If a new best theoretical plan is found, the relevant tour plan is changed at τ +

Theoretical tour plan P τt

Tour plan adaptation layer improves the theoretical plan by applying the solution method

Current anticipation horizon First subsequent anticipation horizon Second subsequent anticipation horizon time

τ τ + ta τ + 2t a τ + 3t a

Fig. 5.1 The utilized real-time control concept adapting the transportation process at τ

The applied real-time control concept is illustrated in Fig. 5.1. The length of the
anticipation horizon t a can be chosen arbitrarily but is constant during the consid-
ered transportation process. Depending on the considered applications, a trade-off
may arise here: While longer anticipation horizons may allow a more comprehen-
sive adaptation by the applied solution method, shorter anticipation horizons can
more directly react to newly arriving requests. In what follows, the applied update
handling which is performed in each anticipation horizon is presented in detail.

5.2.2 Update Handling

In what follows, we describe in detail how tour plans are adapted and selected in
each anticipation horizon of the previously described real-time control concept. In
each anticipation horizon, the applied update handling comprises the following four
steps (see also Fig. 5.2):
1. As mentioned above, since requests that have arrived in the previous anticipation
horizon have only been buffered, all requests of the set Dτ = {i ∈ Rτ |ria > τ − }
with τ − = τ − t a need to be integrated into the three tour plans Pτr , Pτt , and Pτbt .
Therefore, for all new requests in Dτ , travel times from and to all other active
relevant locations (including current vehicle’s positions) are calculated using Di-
jkstra’s algorithm (cf. Dijkstra 1959). This calculated information is used to up-
date the digraph G (see Sect. 5.3.1) with travel time information of the customer
requests in Dτ . After these calculations, the pending requests are subsequently
inserted in a sequence of non-decreasing arrival times into all three tour plans
using a least cost insertion heuristic. Since this least cost insertion can be per-
154 5 A New Deterministic Real-Time Control Approach

Fig. 5.2 The update handling

of the proposed real-time
control approach Check for pending requests
which have arrived during the
previous anticipation horizon

[New requests pending]

Step 1
Update the digraph G
according to the new requests
and vehicle positions [No new requests]

Least-cost insertion of
pending requests into all
three tour plans

Select the relevant tour plan among

Step 2
the three tour plans according to the best
objective function value of all three tour plans

Pre-simulation of the system state

according to the relevant tour plan
for the length of the anticipation horizon

Step 3
Update vehicle positions and corresponding
travel time information in the digraph G

Remove requests from the theoretical plan

where service has begun at τ +

Step 4
Perform the solution method on the
theoretical tour plan for t a - tτcalc time units

formed in zero time, we obtain the set P̃τ = {P̃τr , P̃τt , P̃τbt } of modified tour plans
that are directly applicable in the transportation process at τ .
2. The tour plan of the set P̃τ with the best objective function value is selected as
tour plan Pτr∗ . This tour plan becomes the new relevant tour plan and there-
fore controls the transportation process until time point τ + . Note that we do not
select P̃τbt directly as Pτr∗ because this would not take into account effects of
inserting the pending requests of set Dτ on the objective function value of all
three tour plans in the set P̃τ . Since this selection can also be performed in zero
time, Pτr∗ remains directly applicable at τ .
3. Based on Pτr∗ , the system state at time point τ + is anticipated. This state is com-
puted by pre-simulating all vehicle activities up to τ + according to Pτr∗ . This
updates the vehicle positions and their availability times, i.e., the time points at
which vehicles can react to new orders (see also Sect. 5.3.2.2). Apart from the
resulting anticipated static problem instance at τ + , the removal of all requests,
where service has begun during the pre-simulation, leads to the relevant tour
5.2 The Deterministic Real-Time Control Approach 155

plan Pτr+ applicable at time point τ + . Clearly, in order to adapt the theoretical
tour plan so that it is also applicable at time point τ + , those requests are also re-
moved from P̃τt . The resulting theoretical tour plan is defined as Pτt∗+ and is used
as the initial solution for the solution method. The solution method is a Tabu
Search metaheuristic which will be described in Chap. 7.
Note that by resuming the search process of the Tabu Search procedure from
the solution examined last in the previous anticipation horizon, a larger diversifi-
cation is attained. This is not the case when the search process is always started
using the solution that is currently in execution. However, in case new requests
have arrived (i.e., Dτ = ∅) or in case requests have been serviced in the previ-
ous anticipation horizon, the Tabu Search procedure starts its search directly on
a solution that is equal to Pτr+ using intensifying neighborhood operators. This
methodology is applied since changes in the system state justify the direct appli-
cation of intensifying neighborhood operators on the tour plan Pτr+ .
4. Based on the tour plan Pτr+ , all travel times from the anticipated vehicle positions
at τ + to all other relevant locations are updated using Dijkstra’s algorithm. Note
that only vehicles need to be updated that have changed the current road in the
road network during this anticipation horizon. Otherwise, only their availability
time (see Sect. 5.3.2.2) at the next street crossing needs to be updated. Simi-
lar to step 1, the digraph G is updated according to this calculated travel time
information that will be used in the solution method after the update handling.

At the end of the four steps, the update handling is completed. Note that one
further important aspect for practical applications should be mentioned. Especially
when a large road network is used, time-consuming calculations, due to Dijkstra’s
algorithm in step 4 of the update handling, may occur. As a consequence, the de-
scribed update handling may consume an amount of time that cannot be considered
as zero time in all situations. Hence, the time used for anticipating the new system
situation at τ is measured by tτcalc . After the update handling has been completed at
time point τ + tτcalc , the solution method is executed on Pτt∗+ . Note that the computa-
tional time that is available for the solution method is not equal to t a since the time
needed for the update handling needs to be considered in order to stay in synchro-
nization with the transportation process that is simultaneously executed. Hence, as
mentioned above, the computational time available to the solution method is given
by t a − tτcalc .
After the solution method has been executed, analogous to the previous antici-
pation horizon, the best solution that has been found during the application of the
solution method is stored as Pτbt+ and the tour plan examined last is stored as Pτt + .
These tour plans are utilized in the next anticipation horizon in order to integrate
new requests and to select the relevant tour plan Pτr∗+ that will be executed during
the next anticipation horizon.
156 5 A New Deterministic Real-Time Control Approach

5.2.3 Vehicle Scheduling Strategy

According to our definition, a vehicle scheduling strategy determines when a vehi-

cle will leave its last service request location in order to travel to its next assigned
request location. The vehicle scheduling strategy that is used in this deterministic
real-time control approach is straightforward. Specifically, since customer incon-
venience values directly start when requests arrive in the system according to the
objective function, vehicles are scheduled to directly drive to their next assigned
request location when they have finished servicing the current request. Clearly, this
only applies when a vehicle currently has further requests assigned to its tour. Oth-
erwise, the idle vehicle waiting strategy is utilized as described in what follows.

5.2.4 Idle Vehicle Waiting Strategy

Since new requests arrive during the day, we utilize the following waiting strategy
for idle vehicles in order to increase flexibility for integrating future requests in
the tour plan. It works as follows. If a vehicle has no further requests assigned to
its tour during the day, it waits at its last serviced request location for new requests.
Preliminary computational experiments show that this strategy is superior compared
to an immediate return of the vehicles to the depot. In other words, in the considered
RDOPG applications, computational results showed that it is more advantageous to
keep the vehicles at positions where past requests have arrived than directly sending
the vehicles back to the depot.

5.3 The Mathematical Model of Individual Problem Instances

In this section, the road network and the derived digraph that models information
regarding the relevant locations of the transportation process are described. After
that, the deterministic dynamic mathematical model that is used to represent each
of the individual problem instances in the proposed real-time control approach is
introduced.

5.3.1 Road Network and the Derived Digraph

As mentioned in Sect. 3.1.1, a real road network is utilized in order to enable the
real-time control approach to consider more realistic vehicle movements in the road
network. The utilized road network R̃ is defined by R̃ = (Ñ , Ã, t˜) with Ñ and Ã
denoting its nodes and arcs, respectively. While the nodes represent crossings and
road endpoints, the arcs represent the roads. The value t˜(ã) of function t˜ : Ã → R
5.3 The Mathematical Model of Individual Problem Instances 157

describes the travel time in the road network required on an arc ã ∈ Ã. In order
to model the current situation in the transportation network at τ , the digraph G that
maps all relevant locations at τ , i.e., request locations, the depot, and current vehicle
positions, is used. It represents an abstraction layer from the road network R̃ and is
used to provide the information that is required during tour plan adaptation to the
solution method. Specifically, since in the solution method only the total travel times
between relevant locations are required in order to calculate arrival times at relevant
locations and thereby resulting objective function values, no information about the
exact routes that vehicles will take is required to be provided to the solution method.
As mentioned above, the set of known requests at time point τ is defined by Rτ
whereas the set of unserviced requests at τ is defined by RτU = {1, . . . , RU τ }, i.e., all
requests i ∈ Rτ where no service has begun up to τ . Furthermore, K = {1, . . . , K}
represents the set of available homogeneous vehicles.
The digraph is given by G = (N, A). It models the relevant locations, i.e., the
depot location, request locations, and vehicle locations, of the system state at τ . The
set N = {0, . . . , RU
τ + K} is the associated set of nodes and A = (N × N ) the corre-
sponding set of arcs. Node 0 represents the depot which is located at node D p ∈ Ñ
and nodes 1, . . . , RU U
τ correspond to the set of unserviced request locations Rτ at τ .
p
Similar to the depot, it holds that each request i ∈ RτU is located at ri ∈ Ñ , i.e.,
at a crossing or at a road endpoint in the road network. The start location of vehi-
cle k ∈ K is represented by node vk+ = (RU τ + k) of the digraph. This node corre-
p
sponds to the location of vehicle k at τ in the road network defined by node vk ∈ Ñ .
p
Note that if vehicle k is not at a node ñ ∈ Ñ at τ but on an arc ã ∈ Ã, vk represents
the node of the road network that vehicle k will reach next. Furthermore, vkδt repre-
sents the availability time of vehicle k when it is available at its start location vk+ .
This availability time is a time duration relative to τ . Hence, this value is larger
than 0 if the vehicle is busy with traveling on a road or with servicing a request at τ
(see also Sect. 5.3.2.2). In the considered RDOPG applications, the end location of
vehicle k denoted as vk− is always the depot that is represented by node 0 of di-

graph G. We define K + = k∈K vk+ and K − = k∈K vk− . A non-negative value tij
is associated with each arc (i, j ) ∈ A. It represents the time needed for a vehicle
to travel from location i to location j in the road network R̃ on the shortest path.1
With regard to the considered RDOPG applications, shortest paths in terms of least
travel times are sought. Specifically, with pij = (pij 1 , pij 2 , . . . , pij n ) ⊂ Ã repre-
senting the sequence of arcs in R̃ that describe the shortest path from node i to
node j , it holds that tij = a∈pij t˜(a). Note that in practical applications, the short-
est path routing information given by pij can be used in order to route a vehicle to
the location of its next assigned customer request. Specifically, further information
regarding traffic jams and current travel times could be integrated into the short-
est path calculations in the dispatching center and then transmitted to the vehicle’s

1 The shortest path is defined as the path in a weighted graph connecting two given nodes i

and j with the property that the sum of the weights of all the edges is minimized over all such
paths (cf. Sedgewick 1983, p. 413).
158 5 A New Deterministic Real-Time Control Approach

on-board unit (see also Sect. 2.6.3). In the conducted empirical studies, the simula-
tion component utilizes this information to simulate vehicle movements in the road
network (see also Sect. 8.1).
In the proposed real-time control approach, two different types of shortest path
calculations are required. Specifically, while highly detailed information about the
shortest paths is required for routing vehicles in the road network, the solution
method only needs information about total travel times between the relevant lo-
cations as mentioned above. In these calculations, the geographical positions of the
p p
relevant locations in the road network are given by D p , ri (i ∈ RτU ), and vk (k ∈ K).
Studies performed in Ferrucci (2006, p. 86 et seq.) show that it is beneficial to
use Single-Point Shortest-Path (SPSP) calculations for the first type of calculations
and Single-Source Shortest-Path (SSSP) as well as Single-Source Shortest-Path Re-
verse (SSSPR ) calculations for the second type of calculations. Specifically, SPSP,
also denoted as one-to-one shortest paths, calculates the shortest path for a given
source-destination pair while SSSP calculates the shortest paths starting from a
given source to a set of destinations. Moreover, since travel time information is re-
quired between the relevant locations in both travel directions in the solution method
and travel time information may be direction-dependent due to one-way streets,
SSSPR is utilized in order to calculate the shortest paths starting from a given set
of sources to one destination. More information on shortest path calculations can be
found among others in Ahuja et al. (1993, p. 93 et seq.) while details on efficient
SPSP and SSSP implementations in road networks can be found in Zhan (1997),
Zhan and Noon (1998), and Zhan and Noon (2000). Further sophisticated shortest
path calculation methods which are suitable for calculating shortest paths in road
networks are proposed, among others, by Bauer et al. (2008) and Schultes (2008).

5.3.2 The Deterministic Dynamic Mathematical Model

In this section, the mathematical model for representing the individual static prob-
lem instances generated from the system snapshot at time τ is introduced. Note that
the problem data of the model is derived from the pre-simulated system state at τ +
in order to consider the concurrency of tour plan execution and tour plan adaptation.
The solution of this model determines the tours of the vehicles in order to service
the currently pending requests so that the value of the considered objective function
is minimized. Since the generated vehicle tours are applied at τ + , they start at the
respective anticipated vehicle positions and take into consideration the anticipated
vehicle status and availability times at this point in time. According to the classi-
fication of Powell (see Sect. 2.5.2.3), the presented model is dynamic since it is
adapted according to the current system state during the execution of the transporta-
tion process. Moreover, the model is deterministic since no stochastic knowledge
about future requests is integrated and the model is only solved on the basis of the
current information that is known with certainty at τ . This model will be extended
to a stochastic dynamic model by integrating stochastic knowledge in Chap. 6. In
5.3 The Mathematical Model of Individual Problem Instances 159

the following, we consider the deterministic dynamic version of the model and its
problem data in more detail.

5.3.2.1 Modeling the Requests

In the RDOPG applications, customer requests have soft time windows as well as
variable and fixed penalty costs. Each unserviced customer request i ∈ RτU+ is unam-
p
biguously defined by the tuple ri = (ria , rie , ril , ris , riw , ri ). In this tuple, ria denotes
e
the arrival time of request i in the system whereas ri and ril define the earliest and
latest valid start of service without fixed penalty costs denoted by R pen occurring.
Hence, the time period between rie and ril represents the maximum allowed response
time as introduced in the problem description in Sect. 3.1. Since in the RDOPG ap-
plications, each customer request is directly waiting for service when it arrives in
the system, its time window simultaneously opens upon its arrival, i.e., rie is equal
to the arrival time ria . Note that both these parameters may have different values
for dummy customer requests which will be introduced in Chap. 6. The customer
inconvenience is modeled by variable penalty costs which start at rie . The time re-
quired to service a request i at its corresponding customer location is defined by
the service time denoted by parameter ris . As mentioned in Sect. 3.1, since the max-
imum allowed response time as well as the service time are assumed to be equal
for all requests in the considered example application given by the subsequent de-
livery of newspapers, we define ∀i ∈ RτU+ : ris = R st , ril = rie + R mrt . Parameter riw
determines the weight factor with which the request is considered in the objective
p
function. Finally, ri determines the aforementioned location of the request in the
road network. Since all requests are located at a node in the road network, it holds
p
that ri ∈ Ñ ∀i ∈ RτU+ .

5.3.2.2 Modeling the Vehicles

Each vehicle k ∈ K is modeled by the tuple vk = (vk+ , vk− , vk , vkδt ). As mentioned

above, parameter vk+ defines the start location of vehicle k in the digraph repre-
sented by node (RU τ+
+ k) whereas vk− defines the end location which is equal to the
p
depot location for all vehicles. Parameter vk ∈ Ñ defines the node in the road net-
work which corresponds to the start location of vehicle k. Moreover, relative to τ + ,
p
parameter vkδt determines the availability time of vehicle k at vk . In what follows,
+
we introduce how three vehicle statuses that are possible at τ are transformed into
p
different values of the parameters vk and vkδt :
1. Vehicle k is busy and drives to its next request: As mentioned in the problem
description in Sect. 3.1.4, the first request may be reassigned due to the allowed
vehicle en-route diversion. Therefore, the current destination of a vehicle may be
changed. However, note that such a change is only possible when the vehicle has
reached a node in the road network, i.e., a road junction or road end point ñ ∈ Ñ .
160 5 A New Deterministic Real-Time Control Approach

p
Thus, vk represents the node in the road network that the vehicle will reach next.
Moreover, τ + + vkδt is defined as the point in time when vehicle k arrives at
p
node vk .
2. Vehicle k is busy and services a request: Assuming that vehicle k is servicing
request i at its location ri ∈ N at time point τ + and let x be the remaining service
p

time of request i at τ + . Then, it holds that vk = ri and vkδt = x since the vehicle
p p

needs to complete servicing request i before it is available to fulfill further orders.

3. Vehicle k is idle and has no assigned requests or it waits unused at the depot:
As mentioned in Sect. 5.2.4, the vehicle waits at the location of its last serviced
request i or at the depot, respectively. Since vehicle k is immediately available
p p
for fulfilling further orders, it holds that vkδt = 0 and vk = ri where i ∈ RτU+
p
represents the request i that vehicle k has serviced last or vk = D p , respectively.

5.3.2.3 Model Formulation

In this section, we describe the model which is utilized to represent the indi-
vidual problem instances which occur during the execution of the transporta-
tion process. In the model, binary decision variables of set X = {xki s } (k ∈ K,
d } (k ∈ K, i ∈ {R U ∪ K + }, j ∈ {R U ∪ K − }, i = j ) define request-
i ∈ RτU+ ) ∪ {xkij τ+ τ+
vehicle assignments as well as vehicle travel activities which form the tours of the
vehicles. Specifically, for each vehicle k ∈ K, xki
s is 1 if and only if vehicle k ser-

vices request i ∈ Rτ + while xkij is 1 if vehicle k directly drives from the relevant
U d

location i to the relevant location j and 0 otherwise. Note that this formulation of
binary decision variables relates to the three-index vehicle flow formulation pre-
sented for the CVRP in Sect. 2.3.2. The variable set Y = {yia } ∪ {yis } determines the
time assignments of the requests. For each request i ∈ RτU+ , yia denotes the arrival
time of the assigned vehicle at request i and yis the start time of service at request i.
Note that the variable set Y depends on set X so that the solution of the model is
completely defined by set X.
Based on these definitions, each static problem instance occurring in the trans-
portation process of the considered RDOPG applications can be modeled as follows:

∀i ∈ RτU+ : s
xki =1 (5.1)
k∈K

∀k ∈ K, i ∈ RτU+ : d
xkj i=
d
xkij = xki
s
(5.2)
j ∈{vk+ ∪R U+ } j ∈{R U+ ∪vk− }
τ τ

∀k ∈ K: d
xkv + =
j
d
xkiv − =1 (5.3)
k k
j ∈{R U+ ∪vk− } i∈{vk+ ∪R U+ }
τ τ

∀i ∈ RτU+ : yis = max rie , yia (5.4)
5.3 The Mathematical Model of Individual Problem Instances 161

∀i, j ∈ RτU+ , k ∈ K: d
xkij = 1 ⇒ yja = yis + ris + tij (5.5)
+
∀j ∈ RτU+ , k ∈ K: d
xkv + = 1 ⇒ yj = τ
j
a
+ vkδt + tv + j (5.6)
k k

In this model, constraints (5.1) ensure an unambiguous assignment of requests to

vehicles. Constraints (5.2) make sure that a vehicle correctly visits all of its assigned
requests. Constraints (5.3) guarantee that a vehicle starts and ends its tour at its
defined start and end location. Constraints (5.4) derive the start time of service at
each request location while restrictions (5.5) define the arrival time of each vehicle
at the request locations of its tour. Constraints (5.6) determine the arrival time of
each vehicle at the first request location assigned to its tour. Besides the required
travel time, this arrival time also depends on the availability time of vehicle k at its
p
start location vk . As mentioned above, depending on the current state of vehicle k
at τ + , this availability time is greater than or equal to τ + .
In what follows, detailed information about modeling the objective functions uti-
lized in the considered RDOPG applications is provided.

5.3.2.4 Objective Function

As introduced in Sect. 3.1, the objective function pursues the minimization of the
total customer inconvenience. This total customer inconvenience consists of vari-
able and fixed inconvenience costs and is operationalized by a function of request
response times (see Sect. 2.4.12.7) as well as fixed lateness costs. As defined above,
the point in time when service begins at request i is defined as yis . We denote the
response time of each request i ∈ RτU+ as tires defined by tires = yis − rie . Since each
request i ∈ RτU+ has to be serviced with certainty, each of these requests obtains
the same weight factor riw by which it is weighted in the objective function. For
these requests that already have arrived in the system, this weight factor is set to 1.
When exceeding the predefined maximum allowed response time R mrt , a prohibitive
increase of the customer inconvenience occurs due to the penalty costs that are de-
fined by R pen . Utilizing the specified response time tires , the objective function that
is pursued in the model of the considered RDOPG applications is defined by

min z = riw · F tires + R pen · tires − R mrt
U

i∈R
τ+ Variable customer Lateness penalty costs
inconvenience (i.e., lateness inconvenience)

with the function (x) = 1 if x > 0 and 0 otherwise. As can be seen, the primary
objective is to avoid a late service of customer requests, i.e., all customers should be
serviced within their maximum allowed response time. Furthermore, the secondary
objective is to minimize the total variable customer inconvenience which depends
on the required request response times. As mentioned in Sect. 3.1, the effects of a
linear and a quadratic customer inconvenience increase are assessed in the RDOPG
applications. Therefore, F is defined either as a linear or as a quadratic customer
inconvenience function.
162 5 A New Deterministic Real-Time Control Approach

response time R mrt

Maximum allowed
Customer inconvenience

101.0
Lateness penalty R pen

1.0
Request arrival

linear2X
quadratic

0.0
Response time

Fig. 5.3 Illustration of the linear2X and quadratic customer inconvenience

As illustrated in Fig. 5.3, both customer inconvenience functions have the same
customer inconvenience contribution of 1.0 at the maximum allowed response time
at R mrt whereas further increase, i.e., a violation of this maximum allowed response
time, are penalized by R pen . The prohibitive effect of this penalization is realized
by the significantly higher value of these lateness penalty costs in comparison to
the customer inconvenience values. The presented objective functions are used in
the real-time control approach for coordinating the transportation process during its
execution. In the computational results presented in Chap. 8, the lateness penalty
costs are separately considered by specifying the number of late customer requests.
Hence, the presented improvements which were attained refer to the variable cus-
tomer inconvenience of the objective function.

5.3.2.4.1 Linear Customer Inconvenience (linear2X)

In the linear case, F is defined by:

1 res mrt res
F tires = · min t i , R + 2 · max 0, t i − R mrt
R mrt
Note that in this function the steepness is doubled after the maximum allowed re-
sponse time R mrt in order to still consider those late requests with high priority.
Hence, in the computational results, this type of evaluated customer inconvenience
is denoted by linear2X.

5.3.2.4.2 Quadratic Customer Inconvenience (quadratic)

t res
In this case, the function F is simply defined as F (tires ) = ( Rimrt )2 . Since the steep-
ness of the quadratic function increases implicitly, no adaptation at R mrt is required.
5.4 Characterization of the Deterministic Real-Time Control Approach 163

0.0007

0.0006
Marginal increase per second

0.0005

0.0004

0.0003

0.0002
linear2X
0.0001
quadratic
0
0 500 1000 1500 2000 2500 3000 3500 4000
Response time [seconds]

Fig. 5.4 Marginal increase in variable customer inconvenience using a maximum request response
time of one hour

In the computational results, this type of customer inconvenience is denoted as

quadratic.
In Fig. 5.4, the marginal increase in variable customer inconvenience, i.e., with-
out the lateness penalty costs R pen , in both evaluated customer inconvenience func-
tions is illustrated. In this figure, the explicitly defined increase of the marginal in-
crease of the linear2X customer inconvenience function for requests serviced after
the maximum allowed request response time as well as the implicit increase of the
quadratic customer inconvenience function can be observed.

5.4 Characterization of the Deterministic Real-Time Control

Approach

According to the general classification of real-time control concepts which has been
introduced in Sect. 5.1, the proposed real-time control concept can be characterized
as follows:
1. The adaptation-frequency is implemented by using a continuous concept. Specif-
ically, in time intervals of equal length which is defined by the anticipation hori-
zon t a , the transportation process is adapted by integrating intermediately ar-
riving dynamic events into consecutive problem instances. Hence, the complete
time which is available during the execution of the transportation process is used
for applying the solution method.
2. The adaptation-synchrony is implemented by a process priority since the trans-
portation process cannot be interrupted while the tour plan adaptations are per-
formed by the Tabu Search metaheuristic. Since this proposed solution method
consumes an amount of computational time which cannot be considered as
zero time, changes in the transportation process which occur during the adap-
164 5 A New Deterministic Real-Time Control Approach

Strategic layer long-term

Tactical layer medium-term

Initial tour plan Operative layer short-term

RDOPG Tour plan
Real-time control adaptations
immediate (online
Real-time control layer
Information about during execution)
newly arriving requests
Bidirectional
and their attributes communication

Dynamic events Vehicle on-board units

(i.e., new requests) (e.g., GPS-smartphones)

Fig. 5.5 Integration of the real-time control approach into RDOPG applications

tation need to be considered. In order to consider such changes, the adaptation

range (see below) is set accordingly.
3. The applied adaptation-technique is an improvement method implemented in the
form of a developed stage-based Tabu Search metaheuristic which will be pre-
sented in Chap. 7.
4. The adaptation-range is implemented as a time-limit dependent range. Specif-
ically, since decisions which are made in the transportation process during the
adaptation of the tour plan cannot be changed after the adaptation, they are ex-
cluded from the available adaptation range. The length of this range is defined
by the length of the anticipation horizon t a .
5. The adaptation-duration is realized by a time-limit based concept. Its length is
individually defined for each problem instance by the remaining length of the an-
ticipation horizon after the pre-simulation has been performed, i.e., by t a − tτcalc .

5.5 Utilization of the Real-Time Control Approach in RDOPG

Applications

In Sect. 1.2, the individual layers of transportation processes were introduced. In

Fig. 5.5, the integration of the proposed real-time control approach into the con-
sidered RDOPG applications is illustrated. As was described before, the proposed
real-time control approach for the RDOPG applications is applied to the operative
and real-time control layer of the transportation process. In accordance with the ex-
ample architecture of a real-time control system which was presented in Fig. 2.10,
the RDOPG real-time control is connected to two systems. The first system pro-
vides information about new dynamic events (new requests) while the second one
establishes the bidirectional communication with the vehicles. With regard to the
5.6 Summary 165

considered layers, the real-time control approach supplies one initial tour plan be-
fore the execution of the transportation process (operative layer) and repeatedly pro-
vides adaptations to the relevant tour plan during the execution of the transportation
process (real-time control layer).

5.6 Summary

In this chapter, the real-time control approach which is utilized for coordinating the
transportation process of RDOPG applications was described. Based on a determin-
istic real-time control approach proposed by Bock (2010), the proposed real-time
control approach solves the transportation process during its execution by dividing
the dynamic problem in individual problem instances. A new problem instance is
generated in time intervals of a specified length defined by the anticipation hori-
zon. New requests which occur during the transportation process are buffered until
the generation of the next problem instance. Hence, each problem instance can be
seen as a static problem instance which is adapted by the developed solution method
which will be presented in Chap. 7. Since the solution method consumes a time pe-
riod which cannot be considered as zero time for adapting the tour plan and since
the transportation process continues its execution during the adaptation, the concur-
rency of transportation process execution and adaptation has to be considered. For
this purpose, before the adaptation, the state of each static problem instance at the
end of the current anticipation horizon is anticipated so that the calculated solution
is directly applicable in the transportation process after the adaptation is complete.
The update handling which is used for integrating newly arrived requests and gener-
ating the individual static problem instances was presented in detail. By making use
of an idle vehicle waiting strategy, idle vehicles wait for further arriving requests
at their last serviced request location instead of directly traveling back to the de-
pot. Moreover, vehicles with pending requests directly drive to their next assigned
request. Furthermore, the mathematical model used to represent the individual prob-
lem instances in the real-time control approach was formulated including constraints
which need to be regarded in the considered RDOPG applications. In this mathe-
matical formulation, both evaluated objective functions comprising a linear and a
quadratic evolution of customer inconvenience were introduced. After the descrip-
tion of the utilized real-time control approach, it was categorized according to the
classification of real-time control approaches proposed in Bock (2004, p. 67 et seq.)
which was introduced at the beginning of the chapter. Since the proposed real-time
control approach does not make use of stochastic knowledge, it is a deterministic
real-time control approach and therefore the utilized mathematical model is a de-
terministic dynamic model. The proposed deterministic real-time control approach
will be extended to a pro-active real-time control approach in the next chapter.
Chapter 6
A New Forecasting Approach for Generating
Stochastic Knowledge from Past Request
Information and Utilizing the Stochastic
Knowledge

The real-time control approach presented in the previous chapter utilizes a determin-
istic dynamic model which has been introduced in Sect. 5.3.2. Apart from the idle
vehicle waiting strategy (see Sect. 5.2.4) no methods for providing an increased flex-
ibility for integrating dynamically arriving requests are utilized in this real-time con-
trol approach. Hence, according to the classification in Sect. 2.5.3.2.2, this real-time
control approach is a deterministic real-time control approach which only adapts
the tour plan according to problem data known with certainty. Specifically, although
past request information is available (see Sect. 3.1) out of which stochastic knowl-
edge can be generated, this deterministic real-time control approach does not make
use of it. Therefore, each time a static problem instance is solved, the performed
adaptations do not take into account the fact that further changes in the relevant
information are likely to occur during the execution of the transportation process.
Consequently, tour plans generated by the deterministic real-time control approach
could in many cases be improved in order to service additionally arriving requests
more efficiently, i.e., with lower response times. This is especially true when newly
arriving requests are located far away from current vehicle positions which happens
when previous requests have only arrived in other regions of the considered ser-
vice area. This may occur if changing request arrival rates in the individual regions
during the planning horizon exist in the considered application.
In order to take advantage of the available past request information, we propose
new forecasting methods for generating and integrating stochastic knowledge into
the deterministic real-time control approach in this chapter. By using these methods,
the stochastic knowledge is generated from the available past request information,
i.e., from request arrivals observed on elapsed working days. The generated stochas-
tic knowledge is integrated into the real-time control approach in order to implement
the three pro-active strategies mentioned in Sect. 4.3:
• Relocation strategy: By using this strategy, vehicles can be relocated to subre-
gions, in which future requests are likely to arrive, at suitable time points and
can be kept there for a specified amount of time. Such promising regions in the
service area are automatically derived from the available request information by

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 167

DOI 10.1007/978-3-642-33472-6_6, © Springer-Verlag Berlin Heidelberg 2013
168 6 A New Forecasting Approach for Generating Stochastic Knowledge

the proposed forecasting methods. Note that keeping vehicles in promising sub-
regions after relocation activities have been performed also involves the use of a
waiting strategy which is described next.
• Waiting strategy: When a vehicle has been relocated to a promising subregion
or is already located in such an area, a decision has to be made about how long
this vehicle will be kept there in order to wait for expected future requests. For
this purpose, a maximum value of this amount of time is defined by the generated
stochastic knowledge. Furthermore, the real-time control approach can dynami-
cally determine to keep vehicles for a shorter amount of time in these subregions
according to the current request situation during the transportation process. Com-
parable to the idle vehicle waiting strategy applied in the deterministic real-time
control approach (see Sect. 5.2.4), idle vehicles with no assigned relocation ac-
tivities wait at their last serviced request location.
• Request assignment strategy: In addition to the previous two strategies which
are primarily applied to improve the utilization of idle vehicles, the available
stochastic knowledge is also directly integrated into the tour plan adaptation pro-
cess. Specifically, since request assignments to vehicles affect future positions of
the vehicles, resulting effects on the response times and customer inconvenience
values of expected future requests are anticipated and directly considered during
tour plan adaptations.
According to our classification in Sect. 2.5.3.2.2, using these pro-active strate-
gies extends the deterministic real-time control approach to a pro-active real-time
control approach. The general aim of the pro-active real-time control approach is to
enable earlier service of expected future requests by efficiently utilizing the available
fleet of vehicles according to the expected requests. This is especially beneficial in
the considered RDOPG applications in which the pursued objective solely consists
of minimizing customer inconvenience. Therefore, an earlier service of pending re-
quests is likely to increase the solution quality significantly.
In order to enable relocation and waiting activities of vehicles according to ex-
pected future requests, their temporal and spatial occurrence has to be forecasted.
For this purpose, we propose a new forecasting method for generating stochas-
tic knowledge from the available past request information. The integration of the
stochastic knowledge into the real-time control approach supports the generation
of tour plans with an increased flexibility for integrating expected future requests
which offers possibilities to increase the solution quality of generated tour plans.
By using the stochastic knowledge during the adaptation of the individual prob-
lem instances, the deterministic dynamic model (see Sect. 5.3.2.3) is extended to a
stochastic dynamic model.
In what follows, the proposed approaches for generating and integrating stochas-
tic knowledge into the real-time control approach are described. We first describe
the general ideas that are used for integrating stochastic knowledge into the pro-
posed real-time control approach. Since the stochastic knowledge is integrated into
the pro-active real-time control approach by utilizing dummy customers, we present
how the generation of dummy customers using segment-based clustering is carried
6.1 Integrating Stochastic Knowledge into the Real-Time Control Approach 169

out by analyzing the available past request information. Finally, methods for han-
dling dummy customers in the pro-active real-time control approach during the ex-
ecution of the transportation process are presented.

6.1 Integrating Stochastic Knowledge into the Proposed

Real-Time Control Approach
In the deterministic dynamic model described in Sect. 5.3.2, the set of unserviced
requests at τ + denoted by RτU+ only comprises requests known with certainty so that
no expected future requests are considered. In order to integrate stochastic knowl-
edge about expected future requests, the model is extended as follows. In addition
to the mentioned requests that are known with certainty, the set RτU+ also contains
dummy customer requests, or in short dummy customers, that additionally have to
be considered during tour plan adaptations and the execution of the transportation
process. As mentioned in Chap. 4, dummy customer-related approaches have been
used among others by Ichoua et al. (2006) and Bent and van Hentenryck (2007). In
contrast to many of these approaches, our forecasting approach generates dummy
customers by the application of a sophisticated offline procedure before the execu-
tion of the transportation process. This procedure provides stochastic knowledge of
customizable quality about expected future requests. Since this stochastic knowl-
edge defines the temporal and spatial position of the generated dummy customers,
only one scenario comprising these high-quality dummy customers has to be solved
during the time-critical adaptations of the transportation process. Note that this is
a significant advantage compared to many other approaches which require solving
numerous sample scenarios simultaneous to the execution of the transportation pro-
cess. Consequently, in order to provide useful decision support, these approaches
have to generate their stochastic knowledge in real-time under high time pressure
which may limit their practical applicability and efficiency. Additionally, our offline
approach allows for extending the proposed forecasting methods to more complex
procedures which may require more computational time that is available before the
execution of the transportation process.
As mentioned above, all dummy customers are generated in an offline step so
that they are known before the execution of the transportation process. Therefore,
all dummy customers are directly considered in the initial tour plan generation to-
gether with the real requests that are already known before the execution of the
transportation process. Dummy customers and real requests are similarly handled
by the real-time control approach except for the following four aspects:
• Arrival time and vehicle scheduling: In contrast to real requests, whose time
window directly begins upon their arrival in the system, time windows of dummy
customers open at a specified time point derived from the past request infor-
mation. Specifically, the start of the time window relates to the expected arrival
time of future requests that are represented by the respective dummy customer.
This also requires an extended vehicle scheduling strategy which is described in
Sect. 6.3.2.
170 6 A New Forecasting Approach for Generating Stochastic Knowledge

• Location and service time: In contrast to real requests that have a specific lo-
cation in the road network, a dummy customer guides a vehicle to the spatial
barycenter of a promising subregion denoted as a request-likely area. As men-
tioned in Sect. 3.1.3, a request-likely area represents a temporally and spatially
defined area in which future requests are expected. Since the locations of requests
that are expected to arrive in a request-likely area are presumably different from
its spatial barycenter, the travel time which the assigned vehicle requires from the
barycenter to the location of the expected future requests needs to be considered.
Hence, dummy customers provide information about the estimated total amount
of time that a vehicle requires within a request-likely area. For each request-likely
area, this amount of time is individually determined by the total sum of travel and
service times required to fulfill all expected future requests represented by the
respective dummy customer of the request-likely area. The service time of the
corresponding dummy customer is set to this calculated time period.
Note that since the service time represents the estimated amount of time that a
vehicle will require for servicing the expected requests in a request-likely area, the
integration of dummy customers into the transportation process makes it possible
to have a more foresighted tour plan generation. Specifically, the service time of
a dummy customer acts as a buffer that increases the robustness of generated tour
plans which may lead to better results. Due to this increased robustness of the
tour plans, substantial tour modifications under high time pressure become less
likely and therefore tour plans with a high solution quality can be more frequently
implemented as planned. Moreover, actively guiding vehicles to request-likely ar-
eas before real requests arrive there enables an earlier service of these requests.
Therefore, the utilization of the generated stochastic knowledge provides the de-
sired increased flexibility for integrating expected future requests.
• Uncertainty of occurrence: As mentioned in Sect. 5.3.2.4, each request i that is
known with certainty is considered in the objective function with a weight factor
of riw = 1. In contrast, a dummy customer i represents the expected arrival of
future requests in the corresponding request-likely area with some uncertainty.
Therefore, it is considered in the objective function with a weight factor riw < 1
which depends on the probability that at least one real request will actually arrive
in the assigned request-likely area. As an advantage, this enables the pro-active
real-time control approach to decide, depending on the current pending request
situation, if and for how long each vehicle should be kept in a request-likely area.
With regard to the considered objective function, this flexible vehicle utilization
is more advantageous compared to defining fixed rules.
• Parameter updating: In contrast to known requests that are removed from the
system after service at their request location is completed, dummy customers
represent place-holders for expected future requests. Since dummy customers are
only assigned to a request-likely area which is temporally restricted, dummy cus-
tomers need to be updated during the transportation process. Specifically, as time
passes by during the execution of the transportation process, the probability that
further requests will arrive in the corresponding request-likely area of a dummy
customer decreases. Hence, the dummy customer parameters representing the ex-
pected service time and the weight factor need to be dynamically updated during
6.2 Generation of Dummy Customers Using Segment-Based Clustering 171

the execution of the transportation process. Furthermore, when the probability

that more requests will occur in the assigned request-likely area is too low, the
dummy customer has to be removed from the system. The parameter updating
and removal of dummy customers is discussed in Sect. 6.3.1.

6.2 Generation of Dummy Customers Using Segment-Based

Clustering

In this section, we propose a new forecasting procedure for generating stochastic

knowledge which is integrated into the pro-active real-time control by using dummy
customers. As mentioned in Sect. 3.1.3, this stochastic knowledge is solely gener-
ated by analyzing past request information that has been observed on a specified
number of previous working days which are assumed to be representative of the
next working day, i.e., the next planning horizon. In what follows, the number of
past working days that is utilized to generate the stochastic knowledge is denoted
by nf .
In the proposed forecasting procedure, no information about underlying request
distributions and arrival rates is explicitly given. It is assumed that request arrival
rates follow certain patterns which may depend on the region as well as on the
time of the day. Specifically, it is assumed that dynamically arriving requests can be
modeled by a set of unknown non-homogeneous time-space Poisson processes. In
order to describe non-homogeneous time-space Poisson processes, we introduce the
following definitions:
• Counting process: A stochastic process {N (t), t ≥ 0} is denoted as a counting
process if N (t) denotes the number of events which have occurred by time point t
(cf. Ross 2010, p. 312 et seq.).
• Poisson process: A counting process is a Poisson process with rate parameter λ
if the following three criteria are fulfilled (cf. Ross 2010, p. 313):
(a) N(0) = 0.
(b) For an arbitrary time interval of length t , the number of events occurring in
this time interval is Poisson distributed with mean λt . Due to the properties of
the Poisson distribution, it holds that the expectancy value is E(N (t )) = λt .
The parameter λ is also denoted as rate parameter since it represents the
number of expected events, i.e., request arrivals, per time unit.
(c) Increments are independent, i.e., the number of events which occur in disjoint
time periods are independent. Hence, in Poisson processes, request arrivals
are independently and uniformly distributed in time (cf. Larson and Odoni
1981, p. 46 et seq. and p. 75).
• Time-space Poisson processes: The previously described Poisson process is also
denoted as a one-dimensional Poisson process in time t. A two-dimensional (spa-
tial) Poisson process is defined by a two-dimensional random variable X ∈ R2
172 6 A New Forecasting Approach for Generating Stochastic Knowledge

on a considered area A with N (B) representing the number of events in a sub-

area B ⊂ A. A three-dimensional Poisson process in which events have an occur-
rence time as well as a two-dimensional point location (i.e., X ∈ R3 ) is denoted
as a time-space Poisson process (cf. Larson and Odoni 1981, p. 196).
• Homogeneous and non-homogeneous Poisson processes: A Poisson process
is homogeneous if its rate parameter stays constant. In contrast, it is called non-
homogeneous if the rate parameter comprises changes in one or more dimensions
of the random variable X (cf. Larson and Odoni 1981, p. 193).
In our application, each of the unknown non-homogeneous time-space Poisson
processes has a temporally and spatially varying rate parameter assigned to them.
This may result in time- and space-varying request arrival rates during the consid-
ered planning horizon. Moreover, the rate parameters of the Poisson processes are
also unknown but they are assumed to remain constant from day to day. Note that
the proposed forecasting approach is also applicable if the specified assumptions are
not given. However in this case, a reduced efficiency of the forecasting approach is
likely. Therefore, we develop methods which check specific quality criteria of the
generated stochastic knowledge and integrate them into our forecasting approach.
In what follows, the proposed forecasting approach for generating stochastic
knowledge is described in detail. First, we first describe the segment-based cluster
generation which is used to generate all valid clusters, i.e., those that fulfill defined
quality criteria. After all valid clusters have been generated, we describe how the
cluster selection and dummy customer generation is performed.

6.2.1 Segment-Based Cluster Generation

In order to derive reliable information about expected future requests, the developed
forecasting methods determine appropriate time-space Poisson processes by ana-
lyzing the available past request information. For this purpose, the temporal and
spatial area of the considered service area during the considered planning hori-
zon is divided into subareas which are further temporally divided into segments.
According to Fig. 6.1, each segment s ∈ S has the same predetermined quadratic
spatial (DCse ∈ R2 ) as well as a temporal (DCte ∈ R) extension. Note that all seg-
ments are disjoint and cover the entire considered service area and planning hori-
zon. For each segment s ∈ S, expected request arrivals in s are modeled by a non-
homogeneous time-space Poisson process ps . As mentioned above, it is assumed
that request arrival rates remain constant from day to day. Therefore, the corre-
sponding rate parameter λ(ps ) is calculated by the average number of past request
arrivals in s over the last nf working days. Clearly, since ps represents the expected
request arrivals in segment s each day, ps is only considered during segment s.
Hence, the time unit of λ(ps ) is given by DCte .
Preliminary computational experiments indicated that defining the spatial and
temporal extension of a segment by DCse × DCte = (2.5 kilometers × 2.5 kilo-
meters) × 1 minute provides a reasonable setting for the considered RDOPG ap-
plication (see also Sect. 8.2.1). The division of the considered service area into
6.2 Generation of Dummy Customers Using Segment-Based Clustering 173

y y

y
Time

x x
Considered service area Dividing the area Dividing each subarea into segments
into subareas segments have a spatial and a temporal dimension

Fig. 6.1 Dividing the service area into subareas and segments

segments with such a highly detailed time resolution allows for a precise tempo-
ral mapping of request arrival rates but usually leads to low individual segment
rate parameter values λ(ps ), s ∈ S. In order to allow more reliable forecasts of ex-
pected future requests, adjacent segments are combined into clusters modeled by
compound Poisson processes. Specifically, n non-homogeneous time-space Pois-
son processes ps , s ∈ S ⊂ S with |S | = n are combined into a new compound

Poisson process pS = s∈S ps with rate parameter λ(pS∗ ) = s∈S λ(ps ). Note
∗

that λ(pS∗ ) represents the total number of expected requests in the segments s ∈ S
on one day. Hence, the time unit of λ(pS∗ ) is defined by T · DCte with T denot-
ing the temporal height of pS∗ , i.e., the number of consecutive temporal levels from
which segments s ∈ S are combined in pS∗ . This combination of adjacent segments
is performed for the purpose of pS∗ reaching a rate parameter which is sufficiently
high, thereby justifying the assumption that real requests will actually arrive within
the generated cluster. Such a cluster represents the aforementioned request-likely
area.
In order to generate clusters with a specified minimum forecasting precision and
a minimum reliability of the occurrence of at least one real request in the cluster
during the execution of the transportation process, adjacent segments are assigned
to a cluster ctemp so that the following three requirements are met:
(i) Appropriate shape and extension: Segments are combined so that the clus-
ter ctemp is a three-dimensional convex area which represents the aforemen-
tioned request-likely area. The assigned segments do not exceed a predefined
maximum spatial and temporal extension DCmse and DCmte , respectively.
(ii) Minimum occurrence probability: The sum of the rate parameter values of
the segments s ∈ Sctemp which are assigned to ctemp reaches a defined minimum
score DCminλ , i.e., λ(ctemp ) = λ(pS∗c ) ≥ DCminλ .
temp
(iii) Suitable request distribution and Poisson quality: Both these quality aspects
are described later on.
Preliminary tests have revealed that clusters of the desired quality defined by the
three requirements above cannot be generated by applying the well-known k-means
algorithm (cf. MacQueen 1967). Specifically, k-means requires the number of clus-
ters that shall be generated. However, for generating stochastic knowledge in our
application, this is a restrictive and inappropriate requirement since the number of
174 6 A New Forecasting Approach for Generating Stochastic Knowledge

k-means clustering The new segment-based clustering approach

2D view
of cluster

3D view
of cluster

base and cluster base has base has base has base has
extension are not 1 segment 2 segments, 2 segments, 4 segments
controllable at all vertical horizontal

Fig. 6.2 The k-means clustering algorithm and the new proposed cluster generation approach with
different convex bases

Time Time Time Time

y y y y
x x x x
Temporal levels = 1 Temporal levels = 2 Temporal levels = 3 Temporal levels = 4

Fig. 6.3 Steps of the cluster generation approach generating one valid cluster using a base con-
sisting of four segments

clusters that can be generated depends on the used clustering method itself and
the structural quality of the available past request information. Moreover, accord-
ing to requirement (i), a generated cluster is only useful if it represents a convex
area and its spatial and temporal extensions do not exceed certain limits. This re-
quirement cannot be controlled in the k-means algorithm. Additionally, clusters
generated by k-means are rarely compact and frequently irregular (see Fig. 6.2).
Therefore, we have developed a new cluster generation method that does not have
the above-mentioned disadvantages.
Our proposed clustering approach works in two phases. It starts by generating
all valid clusters that meet the three requirements mentioned above. In the second
phase, the maximum set of non-overlapping clusters is determined. In order to gen-
erate all valid clusters in the first phase, a new cluster is iteratively generated by
starting from all possible bases. Such a base comprises either 1, 2, or 4 segments
that are adjacently located at the same temporal level (see again Fig. 6.2) and to-
gether form a convex cluster ctemp with a temporal height of one segment. Note that
only bases are considered whose sum of assigned rate parameters is larger than 0.
Note that the definition of these bases provides a compact spatial cluster shape
that guarantees the fulfillment of the convexity and the maximum spatial exten-
sion in requirement (i). In order to simultaneously fulfill requirement (ii), segments
located at the successive temporal levels are iteratively inserted into ctemp (see
Fig. 6.3). This is repeated until the calculated rate parameter λ(ctemp ) is at least
DCminλ or the temporal height of ctemp exceeds DCmte . In the latter case, ctemp
is discarded, since it violates the maximum temporal extension in requirement (i).
Otherwise, if ctemp fulfills both requirements, the additional quality criteria of re-
quirement (iii) are checked by conducting the following two steps:
6.2 Generation of Dummy Customers Using Segment-Based Clustering 175

1. As mentioned above, real requests which arrive in a cluster are likely to have
a location which is different from the corresponding dummy customer location.
Hence, a vehicle currently positioned at the dummy customer location is required
to perform a certain detour which consumes additional time in order to reach
the real request. If the expected detour is too large, positive effects of using a
dummy customer are likely to disappear. Therefore, the quality of clusters with
regard to the expected travel time for such detour activities is assessed. For this
purpose, a suitable location for the dummy customer in ctemp is derived. The spa-
tial barycenter of the past requests that have been observed within ctemp over the
last nf working days is calculated and mapped on the road network by determin-
ing the closest located node nctemp . Since this node may be positioned within an
area consisting of roads with a low speed limit, it is relocated to a node that fulfills
the following two requirements: First, it is located within a maximum predefined
travel time radius DCradiusTT of nctemp . Second, in order to provide a reasonable
road connection, it is located at a road of a minimum desired road class where
lower road classes denote roads with higher speed limits. This node called nctemp
is determined by conducting a Dijkstra search that starts at nctemp and is limited
by the parameter DCradiusTT . Note that such a node may not exist in the consid-
ered radius. In this case, the closest node with minimum road class (i.e., largest
speed limit) is determined. The node nctemp represents the potential location of
the dummy customer assigned to ctemp . After nctemp is calculated, the average
avgTT
travel time ctemp from nctemp to all past requests observed within ctemp over
avgTT
the last nf working days is calculated. If ctemp exceeds a predefined maximum
average travel time DCmaxAvgTT , ctemp is discarded since the estimated detour
in ctemp is too large to provide a reliable forecast for an efficient positioning of
the assigned vehicle.
2. As mentioned above, we assume in our forecasting method that request arrivals
can be modeled by non-homogeneous time-space Poisson processes whose rate
parameters are constant from day to day. In order to satisfy this assumption to a
degree which is practically reasonable,1 the quality of ctemp is analyzed. Specif-
ically, the three requirements of Poisson processes (see Sect. 6.2) are considered
for each cluster ctemp as follows. While requirement (a) is straightforward, re-
quirement (b) is explicitly checked by applying Pearson’s Chi-Square Goodness-
of-Fit test (cf. Kvam and Vidakovic 2007, p. 155 et seq.). Using this test it is
checked whether there exist significant deviations between the number of ac-
tual customer request arrivals in ctemp over the considered nf working days and
those that are expected according to the Poisson distribution with rate parame-
ter λ(ctemp ).
In general, the Chi-Square Goodness-of-Fit test is applied to test the null
hypothesis H0 : FX (x) = F0 (x) where FX (x) is the observed sample X =

1 Asnoted by Larson and Odoni (1981, p. 194), almost any practical system has a non-perfect de-
gree of conformity with the postulates of the Poisson process. Hence, a trade-off between accuracy
and insights possible by utilizing the Poisson model has to be found.
176 6 A New Forecasting Approach for Generating Stochastic Knowledge

{X1 , X2 , . . . , Xn } and F0 (x) represents the distribution to be tested. The domain

of this distribution is partitioned into r non-overlapping intervals I1 , I2 , . . . , Ir .
Moreover, n1 , n2 , . . . , nr represents the number
of samples that have been ob-
served in the corresponding interval with ri=1 ni = n. Furthermore, npi is the
expected value of interval Ii according to the distribution to be tested under H0 .
The Chi-Square test statistic for the sample X is then defined as follows:

r
(ni − npi )2
X =
2
npi
i=1

The X 2 test statistic is asymptotically χ 2 distributed with f = r − k − 1 degrees

of freedom where k represents the number of parameters for the distribution
under H0 that were estimated by using sample X (cf. Kvam and Vidakovic 2007,
p. 156). Let χf2 (x) be the cumulative distribution function of the χ 2 distribution
with f degrees of freedom and value x. After calculating the X 2 test statistic of
the sample X, it is decided whether the null hypothesis is to be rejected or not.
For this purpose, a type I error α which is application-dependent is defined. If it
holds that χr2 (X 2 ) > 1 − α, the null hypothesis is rejected.
In our forecasting approach, the Chi-Square Goodness-of-Fit test is applied in
order to test the null hypothesis H0 : the number of daily request arrivals observed
in ctemp over the last nf working days represented by the sample X are Poisson
distributed with rate parameter λ = λ(ctemp ). Since we estimated the rate param-
eter λ by using the sample X, k is 1 and consequently f = r − 2. We set n = nf
to the number of analyzed working days and accordingly X1 , . . . , Xnf represents
the number of observed request arrivals in ctemp over the past nf working days.
Let PSλ (x) denote the probability mass function of the Poisson distribution with
rate parameter λ and value x. According to the estimated rate parameter λ, we
generate l intervals I1 , I2 , . . . , Il with npi denoting the expected number of days
on which exactly i − 1 requests arrive in ctemp , i.e., npi = n · PSλ(ctemp ) (i − 1).
The interval Il represents the expected number of days on which at least l − 1 re-
quests arrive in ctemp . Consequently, it holds that li=1 pi = 1. According to the
past request information used in the cluster generation process, ni represents the
number of days out of the considered nf working days on which exactly i − 1 re-
quests have been observed in ctemp .
As mentioned above, the X 2 test statistic is asymptotically χ 2 distributed. In
order to use this property also for practical sample sizes, different authors men-
tion that certain data conditions should be met (cf., e.g., Yates et al. 1999, p. 734
and Kvam and Vidakovic 2007, p. 156). Specifically, it is recommended that at
least 80 % of all intervals I1 , . . . , Il have an expected value of npi ≥ 5 while
all of the intervals fulfill npi ≥ 1. In order to fulfill these conditions, we pool
the first and last intervals until both requirements are fulfilled. After this pool-
ing has been performed, we obtain r intervals I1 , . . . , Ir with correspondingly
adapted values of npi and ni . After calculating the X 2 test statistic for ctemp
denoted as Xc2temp , the null hypothesis is rejected and the cluster is discarded
6.2 Generation of Dummy Customers Using Segment-Based Clustering 177

ctemp ) > 1 − α. Clearly, smaller values of χr−2 (Xctemp ) indicate a higher

2 (X 2
if χr−2 2 2

conformity of observed request arrivals in ctemp with those expected by a Poisson

distribution having rate parameter λ(ctemp ). Hence, we define the Poisson qual-
ity of ctemp by PQ(ctemp ) = 1 − χr−2 2 (X 2
ctemp ). Since cluster ctemp is accepted
if PQ(ctemp ) ≥ α we interpret α as the minimum demanded Poisson quality for
valid clusters. For this reason, setting α = 0.05 which is often proposed in the
literature (cf., e.g., Israel 2008, p. xxii, Sapsford and Jupp 2006, p. 217) might
be a value that is too low for our application since our aim is to accept only
clusters whose observed request arrival rates possess a Poisson distribution to a
certain extent. In Sect. 8.6, we will evaluate the impacts of different values for
parameter α on the generated dummy customers.
Finally, requirement (c) of Poisson processes can be analogously checked by
using a Chi-Square Goodness-of-Fit test and testing the request arrival times
which were observed in ctemp over the last nf working days for uniformity. How-
ever, due to the small temporal extensions of generated clusters, we refrain from
performing this test and assume that request arrivals are independent of each
other. This is because the required accuracy in request arrival times is frequently
not fulfilled by practical request data sets and therefore would lead to an unfa-
vorable rejection of clusters.

If both steps have been successfully completed, the cluster ctemp fulfills the de-
manded quality criteria and is therefore added to the set of valid clusters denoted
by C. Furthermore, request parameter values of the corresponding dummy customer
request of ctemp are calculated. The dummy customer request i belonging to ctemp
gets the weight riw defined by:

riw = P (X ≥ 1) = 1 − PSλ(ctemp ) (0)

This value corresponds to the probability that at least one request will arrive in ctemp .
Moreover, the service time ris of dummy customer i is defined by:

avgTT
ris = R st + ctemp · λ(ctemp )

As mentioned above, ris is set to the sum of service and travel times that the assigned
vehicle is estimated to require for servicing all expected requests in ctemp .
Note that for each base (see Fig. 6.2) on which valid clusters can be generated,
our clustering method generates the valid cluster with the smallest temporal height.
Since generating such precise clusters allows for producing two temporally small
clusters instead of a large one, two dummy customers can also be created. This
is advantageous since it increases the flexibility of the pro-active real-time control
approach by allowing to assign the integrated stochastic knowledge to different ve-
hicles.
178 6 A New Forecasting Approach for Generating Stochastic Knowledge

6.2.2 Cluster Selection and Dummy Customer Generation

After generating all valid clusters, the clusters which are integrated as stochastic
knowledge into the transportation process have to be selected. In order to integrate
as much stochastic knowledge as possible, it is pursued to select the maximum num-
ber of clusters that do not overlap from the set C. For this purpose, the following
Mixed Integer Programming (MIP) problem is solved. The set SC ⊆ S comprises
all segments which are assigned to clusters of the set C. The binary parameter cis
is set to 1 if and only if cluster i ∈ C contains segment s ∈ SC . Moreover, parame-
ter cistartTL denotes the temporal level at which cluster ci starts. The binary decision
variable xi is 1 if and only if cluster i ∈ C is selected, 0 otherwise. Furthermore, pa-
rameter M is a big number. Using these definitions, the cluster selection procedure
is formulated as follows:

max z = M · xi − cistartTL · xi s.t. (6.1)
i∈C

∀s ∈ SC : cis · xi ≤ 1 (6.2)
i∈C

xi ≤ C max (6.3)
i∈C

The primary goal of the hierarchical objective function (6.1) is the maximization of
the number of selected clusters. Since this also maximizes the number of dummy
customers that are integrated in the tour plan, it ensures that the maximum stochas-
tic knowledge generatable by our forecasting approach is utilized. The secondary
goal is to start clusters at the earliest point in time in order to support an earlier
service of the requests that are expected in the identified request-likely areas. This
also supports the pursued aim of guiding the vehicles into request-likely areas at
the right time. Constraints (6.2) ensure that all segments s ∈ SC are assigned to, at
most, one cluster i ∈ C. Constraint (6.3) imposes a theoretical upper bound on the
number of clusters that can be selected. Note that this restriction was defined for
technical reasons; since it allows the MIP solver to generate tighter bounds during
the examination process which enables a faster solution process of the considered
cluster selection problem. For the computational experiments, this bound is initially
set to:

s∈SC λ(ps )
C max
=
DCminλ
After solving the cluster selection problem, for each selected cluster a dummy cus-
tomer with the parameter settings defined above is added to the set of initial requests.
These dummy customer requests are integrated into the initial tour plan which is
constructed before the execution of the transportation process. This is described in
Sect. 7.4.
6.3 Handling Dummy Customers in the Pro-Active Real-Time Control Approach 179

6.3 Handling Dummy Customers in the Pro-Active Real-Time

Control Approach

As mentioned in Sect. 6.1, request-likely areas are only present for a specific time
period during the execution of the transportation process. Hence, the proposed uti-
lization of stochastic knowledge in the real-time control approach requires the up-
dating of specific dummy customer parameters and appropriate actions during the
transportation process. In what follows, we describe in detail the performed dy-
namic parameter updating during the transportation process. This is followed by
the presentation of the extended vehicle scheduling strategy with regard to dummy
customers. Specifically, if a dummy customer is the next request in the tour of a
vehicle, it is advantageous to determine its departure according to the start of the
dummy customer’s time window. Finally, an extended vehicle waiting strategy us-
ing dummy customer information is proposed.

6.3.1 Dynamic Parameter Updating During the Transportation

Process

As stated above, a dummy customer i belonging to cluster ci guides a vehicle

to a request-likely area whose spatial and temporal dimensions are equal to the
ones of the corresponding cluster ci at the appropriate time. Analogous to the
definition of cistartTL presented above, we define ciendTL as the temporal level at
which cluster ci ends. Consequently, the start and end time of cluster ci denoted
as cistart and ciend in the transportation process is defined by cistart = cistartTL · DCte
and ciend = (ciendTL + 1) · DCte . From the moment the time point of the anticipated
system situation τ + reaches cistart during the execution of the transportation process,
the values of the parameters of dummy customer i continuously change. This is
because the number of segments, which are assigned to the remaining part of clus-
ter ci that is still in the future at τ + , diminishes so that the rate parameter value of
the remaining assigned compound Poisson process decreases with τ + . Therefore,
whenever a new problem instance which represents the anticipated system situa-
tion at τ + is generated at τ , the remaining rate parameter value of each cluster ci
is updated to λ(ci , τ + ). In particular, this value represents the rate parameter val-
ues of the Poisson processes ps of segments that are assigned to ci and are still in
the future at τ + . Moreover, the remaining rate values of segments, within whose
time dimension the time point τ + is currently situated, are linearly approximated.
Clearly, for each cluster with cistart ≥ τ + , it holds that λ(ci , τ + ) = λ(ci ). Moreover,
for each cluster ci we define the start time of the remaining part of the cluster ci at
τ + as cistart (τ + ) = max(cistart , τ + ). Based on the anticipated system state at τ + and
the resulting rate parameter λ(ci , τ + ), the updating of the three parameters rie , ris ,
and riw of each dummy customer i is conducted as follows:
180 6 A New Forecasting Approach for Generating Stochastic Knowledge

n=4 Average arrival time of the first

request if n requests occur in ci
n=3
Average arrival time of additional
requests if n requests occur in ci
n=2

n=1

time
0 ci start(Ĳ+) ci end

n=4

n=3

n=2

n=1

Fig. 6.4 Average arrival time of the first request when n request arrivals occur in cluster ci

• Beginning of the dummy customer time window (rie ): As mentioned above,

requests represented by dummy customer i are only expected and thus do not
occur with certainty. Hence, the opening of time window rie is set to the expected
arrival time of the first request in the remaining part of cluster ci which is still in
the future at τ + under the assumption that at least one more request will occur in
this remaining part of ci . Clearly, this value is always in the future. The expected
arrival time of the first request is updated as follows:
∞
n=1 PSλ(ci ,τ + ) (n) · n+1
1
+
rie τ = cistart τ + + ciend − cistart τ + · (6.4)
1 − PSλ(ci ,τ + ) (0)

The expected arrival time of the first request in cluster ci is derived by determin-
ing the average arrival time of the first request in all scenarios with 1, 2, . . . , ∞
requests that arrive in the remaining time period of cluster ci . The calculation of
this average arrival time is described in what follows. Due to the assumed proper-
ties of Poisson processes (see Sect. 6.2), request arrivals are on average uniformly
distributed over the time period of cluster ci . Hence, as illustrated in Fig. 6.4,
when n > 0 requests arrive in cluster ci , the remaining time period of cluster ci
is divided into n + 1 intervals of equal length so that the average arrival time of
the first request, relative to the beginning of cluster ci , is n+1 1
· (ciend − cistart (τ + )).
This time point is multiplied by the probability PSλ(ci ,τ + ) (n) which represents the
probability that exactly n requests will arrive in cluster ci . The resulting values
for all n ∈ {1, 2, . . . , ∞} are summed up and divided by the probability that at
least one request will arrive in cluster ci . By adding the current start time point
cistart (τ + ) of cluster ci , the expected arrival time of the first request in cluster ci is
determined (see (6.4)).
6.3 Handling Dummy Customers in the Pro-Active Real-Time Control Approach 181

Since request arrivals are independent of each other in Poisson processes, no

requests which have already arrived in ci need to be considered. Thus, due to
the impact of rie (τ + ) on the objective function value, generated tour plans do not
prefer to visit dummy customer i earlier than the expected occurrence of the next
request within cluster ci if other requests can be visited before.
Note that updating rie in the described way has an additional advantage.
Consider the case in which rie is not updated and remains at a value defined
by rie = cistart . In this case, although a vehicle is already waiting at the dummy cus-
tomer, the corresponding dummy customer nevertheless contributes a customer
inconvenience value to the objective function. This is an unwanted scenario since
it is not taken into consideration whether or not a vehicle has arrived at the dummy
customer location in time. In contrast, by updating rie in the way described above,
it is positively acknowledged when a vehicle arrives at the location of dummy
customer i in time by omitting a contribution of a customer inconvenience value
to the objective function value of the generated tour plan.
• Service time (ris ): As mentioned above, the service time ris of dummy customer i
represents the time period which a vehicle is estimated to require in the corre-
sponding request-likely area. This time period is represented by the sum of av-
erage travel and service time for fulfilling all expected future requests. Since the
number of expected future requests in ci at τ + is given by λ(ci , τ + ), the service
time is updated to:
avgTT
ris τ + = R st + ci · λ ci , τ +

• Weight factor (riw ): For each dummy customer i, the weight factor riw is updated
according to the current remaining probability at τ + that at least one more request
will arrive in cluster ci . Therefore, for each dummy customer i, this probability
can be calculated by:

riw τ + = P (X ≥ 1) = 1 − PSλ(ci ,τ + ) (0)

Furthermore, since dummy customers are placeholders for expected future real re-
quests, dummy customers are not serviced. Hence, the time point when a dummy
customer is removed from the tour plan has to be defined. In our approach, a dummy
customer is removed from the set of unserviced requests RτU+ when it becomes un-
likely that further requests will occur in the considered request-likely area defined
by cluster ci . This time point denoted as tcrem
i
is defined by the time when λ(ci , τ + )
becomes smaller than a predefined minimum threshold DCλrem . Note that since tcrem i
is independent of dynamic events, it can be calculated offline before tour plan exe-
cution. Hence, the point in time when each dummy customer is removed from the
tour plan during the transportation process is exactly defined before the execution
of the transportation process.
182 6 A New Forecasting Approach for Generating Stochastic Knowledge

6.3.2 Extended Vehicle Scheduling Strategy with Regard

to Dummy Customers

During the execution of the transportation process, situations may occur in which a
dummy customer is scheduled as the next request of a vehicle tour but its service is
planned in the far future due to a much later opening of its time window. Using the
vehicle scheduling strategy presented in Sect. 5.2.3, the vehicle would nevertheless
immediately drive to the dummy customer i and wait there for a much longer time
than is intended by the dummy customer. In contrast to a scenario where the current
transportation process execution time τ is after the time point cistart and therefore
placing an idle vehicle in the corresponding request-likely area is reasonable, this
could lead to unwanted scenarios. For instance, vehicles are guided to remotely lo-
cated areas much too early and are therefore not directly available for intermediately
arriving requests in more central regions.
In order to deal with this issue, an extended vehicle scheduling strategy is applied.
Specifically, if a vehicle’s next request is a dummy customer i, the vehicle waits
at its current position as long as the difference between the current transportation
execution time to cistart is larger than the travel time that is needed to reach the
dummy customer i. This is performed because a visit to dummy customer i might
also be useful in the interval [cistart , rie ) since requests are also expected to occur
with a certain probability before the expected arrival time of the next request in the
cluster ci denoted by rie .

6.3.3 Extended Vehicle Waiting Strategy Using Dummy Customer

Information

As mentioned before, dummy customers are used in order to guide vehicles to

request-likely areas and keep them there for a specified amount of time. The length
of this time period is decided by the pro-active real-time control process itself.
Specifically, a vehicle waits at a dummy customer until another request is assigned
to its tour as the first request. This happens when another request is assigned before
the dummy customer to the vehicle’s tour or when the dummy customer is removed
from the tour plan. By allowing the pro-active real-time control approach to dynam-
ically decide about the length of this waiting period, the waiting strategy can be
utilized in a flexible manner according to changing situations during the execution
of the transportation process.
In order to enable a vehicle to wait at the location of a dummy customer for fu-
ture customer requests to arrive, an additional vehicle status has to be defined in
the mathematical model introduced in Sect. 5.3.2. This is due to the fact that while
real requests can always be serviced immediately, the start of the time window of a
dummy customer is always in the future as described in Sect. 6.3.1. In this constel-
lation, the vehicle simply waits at the dummy customer in order to be prepared for
6.4 Summary 183

quickly servicing future arriving requests in the corresponding request-likely area.

The additional vehicle status can be handled analogous to the third status presented
in Sect. 5.3.2.2. Specifically, if, at τ + , vehicle k is currently waiting at a dummy cus-
p
tomer location, the vehicle location vk is set to the location of the dummy customer
δt
and the vehicle availability time vk is set to 0 since vehicle k is directly available
for servicing other requests.

6.4 Summary
In this chapter, a new forecasting approach was presented in order to extend the
deterministic real-time control approach to a pro-active real-time control approach.
The proposed forecasting approach makes it possible to generate stochastic knowl-
edge out of the available past request information. After its generation, the stochas-
tic knowledge is integrated into the real-time control approach by using dummy
customers. In contrast to real requests, dummy customers are used to guide vehi-
cles into request-likely areas in which future requests are expected to occur with a
high probability according to the available stochastic knowledge. The generation of
stochastic knowledge is performed by partitioning the considered service area into
subareas and segments. For each of these segments, request arrivals are modeled by
a time-space Poisson process whose rate parameter is calculated by analyzing the
number of request arrivals in the respective segment on the considered past working
days. Afterwards, dummy customers are created by aggregating segments to clus-
ters that fulfill certain quality criteria, for example that one or more requests actually
arrive in place of a dummy customer with a minimum specified probability. In order
to check the assumption that the number of request arrivals in the segments can be
modeled by Poisson processes, the likelihood that past request arrivals which are
observed in each generated cluster follow a Poisson distribution is evaluated. If this
is too unlikely, the cluster is discarded. In order to generate as many valid dummy
customers as possible, an exact two-step solution method was presented. In the first
step, all valid clusters are generated while in the second step the maximum number
of non-overlapping clusters is selected by making use of an MIP solver. Moreover,
the methods by which the deterministic real-time control approach is extended in
order to integrate and handle dummy customers were described. This includes how
relevant parameters of dummy customers are updated during the execution of the
transportation process and how it is decided when dummy customers are to be re-
moved from further consideration. Furthermore, the extension of the deterministic
real-time control approach included the introduction of an extended vehicle schedul-
ing strategy. This strategy makes a vehicle wait at its current position if its next as-
signed request is a dummy customer and if its departure at the current time in the
transportation process would otherwise make the vehicle arrive earlier than desired
at the dummy customer’s location.
Chapter 7
The Proposed Tabu Search Solution Method

In this chapter, the developed solution method which is utilized to solve individual
problem instances that occur during the execution of the transportation process is
presented. This solution method is a specific stage-based Tabu Search metaheuris-
tic. In order to efficiently integrate intensification and diversification aspects into the
enumeration process, the Tabu Search procedure switches between different stages
that control the currently applied neighborhood operator. In each iteration, depend-
ing on the current stage, a neighborhood operator to be applied is deterministically
selected. If a defined number of iterations has been performed without finding a new
best solution, the Tabu Search algorithm switches into the next higher stage in order
to continue the search with a different neighborhood operator. Neighborhood oper-
ators on lower stages intensify the search in direct vicinity of the current solution,
whereas higher stages apply more diversifying neighborhood operators. If a new
best solution is found, the search returns to the first stage.
As mentioned in Sect. 5.2.1, a static problem instance that is adapted during a run
of the solution method is not modified by dynamic requests that intermediately occur
in the transportation process since these events are buffered until the generation of
the next static problem instance. Therefore, each adaptation of the solution method
is similar to solving a static routing problem. In order to support the approach to
enable a continuous optimization over static problem instances, an appropriate tabu
list structure is developed. In the static problem instance generated at τ , the Tabu
Search solution method is applied for t a − tτcalc time units. After the execution of
the solution method, the tour plans Pτbt+ and Pτt + , i.e., the best and the last tour plan
found during the search process, are generated. Hence, since the computational time
that is available for the solution method is temporally restricted, different concepts
for efficiently computing the objective function are utilized in order to speed up the
solution process.
In this chapter, an introduction to Tabu Search by describing general concepts
of the implemented solution method is given. Specifically, basic concepts of this
metaheuristic are explained as well as the way a tour plan is represented in the de-
veloped solution method. Moreover, it is illustrated how a tour plan is modified by
the application of the neighborhood operators and how the objective function value

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 185

DOI 10.1007/978-3-642-33472-6_7, © Springer-Verlag Berlin Heidelberg 2013
186 7 The Proposed Tabu Search Solution Method

is efficiently updated according to performed modifications. After that, the individ-

ual neighborhood operators that are applied in the proposed Tabu Search solution
method are presented in detail. Furthermore, the implemented tabu list for identify-
ing tabu-active solutions is described. Moreover, the construction of the initial tour
plan is illustrated. In the following section, the stage-based neighborhood operator
selection scheme that controls the application of the individual neighborhood opera-
tors during the execution of the solution method is presented. Finally, the integration
of the solution method into the real-time control approaches is explained.

7.1 General Concepts of the Implemented Solution Method

In this section we introduce the general concepts that are used in the proposed Tabu
Search solution method. First, an introduction to the Tabu Search metaheuristic is
given. Since Tabu Search is an iteration-based solution method, a description of an
iteration in the solution method is given. This is followed by the representation of
a tour plan in the solution method is and, finally, it is explained in which way an
efficient calculation of changes in the objective function value is performed.

7.1.1 The Tabu Search Metaheuristic

In the real-time control approach that is used in this book to coordinate the trans-
portation process of RDOPG applications, a stage-based Tabu Search metaheuris-
tic (cf. Glover 1989 and Glover and Laguna 1998) is applied. The Tabu Search
metaheuristic is an iteration-based heuristic improvement method which performs a
neighborhood based search in the solution space. As mentioned in Sect. 2.3.5, the
neighborhood of a neighborhood operator is defined as the set of solutions which
can be generated from a current solution by applying this operator. In order to over-
come local optima and to enable the metaheuristic to examine a broader part of the
solution space, neighborhood operators are classified into intensifying and diversi-
fying operators (cf. Glover and Laguna 1998, p. 7 et seq.):
• Intensifying neighborhood operators: Such operators often attempt to improve
the current solution by applying small changes to the tour plan, e.g., only rein-
serting one customer to another location within the same tour. It is advisable to
utilize such operators when the current solution is located in a part of the solution
space where better solutions are expected to be attained by making slight changes
in the current solution. Besides this, Glover and Laguna state that these operators
are useful for exploring the neighborhood of elite solutions i.e., new best solu-
tions, found during the search process. This is carried out in order to find local
optima in the solution space that are close to the current solution. Moreover, these
operators can also be used to combine parts of previously found elite solutions in
order to generate a new best solution similar to adaptive memory techniques, e.g.,
7.1 General Concepts of the Implemented Solution Method 187

as applied in Gendreau et al. (1999a). In doing so, previously visited parts of the
solution space can also be re-examined in more detail.
• Diversifying neighborhood operators: Such operators attempt to improve the
current solution by applying significant modifications to it. In so doing, guiding
the search process to unvisited parts of the solution space is the pursued aim so
that these parts are examined for other promising solutions. The utilization of di-
versifying neighborhood operators is advisable when intensifying neighborhood
operators have not found a new best solution during a specific number of past
iterations of the search process.
As a specific criterion of Tabu Search metaheuristics, information about the
progress of the current search process is stored. Specifically, characteristics of so-
lutions that have already been visited are stored in a list denoted as tabu list. Using
this tabu list aims at avoiding the revisiting of solutions within a specified num-
ber of iterations during the solution process. For this purpose, visited solutions are
marked as tabu-active. The utilization of the tabu list supports the search process in
continuing to evaluate other parts of the solution space after a local optimum has
been reached. In order to identify tabu-active solutions, characteristic attributes for
determining the tabu-active state of solutions have to be defined. More information
about the implemented attributes in the proposed Tabu Search metaheuristic is given
in Sect. 7.3.
An iteration of the search process in a Tabu Search metaheuristic is generally
carried out as follows. Its neighborhood is examined by applying a specific neigh-
borhood operator starting from a current solution. After all solutions of this non-
tabu active neighborhood have been evaluated, one neighbor is selected as the new
solution according to specific criteria. The modifications that are applied to the cur-
rent solution to obtain the new solution are called a neighborhood move or in short
move. Note that in situations during the search process in which all neighbored so-
lutions are worse than the current solution, the selected neighborhood move leads
to a solution with an objective function value worse than the current solution. For
example, this scenario happens when a local optimum is reached. In a Tabu Search
metaheuristic, worse solutions are usually accepted in order to overcome these local
optima. Specifically, the application of more diversifying neighborhood operators
is likely to first result in worse solutions since major parts of the current solution
are changed in order to guide the search process into a different part of the solution
space. However, this is accepted since it is expected that this temporal deterioration
leads to better solutions in subsequent iterations.
Since Tabu Search is an improvement metaheuristic, an initial solution has to
be provided to it. In our proposed solution approach, the generation of the initial
solution, i.e., the initial tour plan, is described in Sect. 7.4.

7.1.2 Description of an Iteration in the Solution Method

In each iteration, a neighborhood operator is selected and its neighborhood is evalu-

ated. In the proposed solution method, this selection is deterministically performed
188 7 The Proposed Tabu Search Solution Method

by using a stage-based selection scheme which is described in Sect. 7.5. Once a

neighborhood operator has been selected, it is applied to a given current solution
(i.e., the current theoretical tour plan) by executing the following steps:
1. Selecting the considered request set: Depending on the chosen neighborhood
operator, one or more unserviced requests of the set RτU+ are selected. In what
follows, the set of selected requests is denoted as R ∗ .
2. Removing the considered request set: The requests belonging to the set R ∗
are removed from the current solution.
3. Temporarily reinserting the requests into the theoretical tour plan: Each
request of R ∗ is reinserted into the theoretical tour plan. The position at which
each request is reinserted into the theoretical tour plan is determined by the
neighborhood operator and can be either randomly chosen, exhaustive (i.e., suc-
cessively evaluating all positions), or rule-based (defined by the neighborhood
operator, e.g., at its least cost position). This set of temporarily performed rein-
sertion activities defines the currently evaluated neighborhood move. By this
reinsertion, a tentative new solution is generated.
4. Evaluating the tentative new solution: In order to assess the quality of this
new solution, the change in the objective function value denoted as the move
value is calculated.
5. Storing the best found neighborhood move: If the current neighborhood
move is non-tabu, it is compared to the currently known best neighborhood
move that has been found during the ongoing neighborhood examination in this
iteration. If the current neighborhood move is better than the best found move
so far, it becomes the new best found move. Moreover, the best found move
is stored as escape move which, regardless of whether it is tabu-active or not,
is executed in case all evaluated neighborhood moves are tabu. Note that this
only happens if the set RτU+ comprises few requests so that all neighbored solu-
tions of the current neighborhood operator have already been evaluated in the
previous iterations of the current static problem.
6. Evaluating further neighbors: If the current neighborhood operator evaluates
more than one reinsertion position of the requests of set R ∗ , further reinsertion
positions are evaluated by continuing this procedure at step 3. Otherwise, the
process continues with the next step.
7. Restoring the initial solution: After all reinsertion positions for the requests
of set R ∗ have been evaluated, the considered requests are reinserted at their
original positions so that the theoretical tour plan is now identical with the initial
solution of this iteration.
8. Selecting the next request set to be conducted: If the current neighborhood
operator considers the temporal reinsertion of different request sets R ∗ , the pro-
cess continues with the next set at step 1, otherwise at step 9.
9. Selecting and executing the best neighborhood move: After all possible
neighbored solutions have been evaluated, the best found neighborhood move
(or, in case no non-tabu move has been found, the escape move, see step 5) is
executed. This new generated solution now represents the current solution.
7.1 General Concepts of the Implemented Solution Method 189

D r1 r2 r4 r6 v1 r3 r5 v2 E

Fig. 7.1 Representation of a tour plan using a doubly linked list

_ _ _M

v1 K 7
s _
MD
_ _
B |
7 _ F _ F _ F
x x x
@ABC
GFED / @ABC
GFED
r1 o / GFED
@ABC
r2 o / @ABC
GFED
r4 o / GFED
@ABC
r6 o / v o / @ABC
GFED
r o / GFED
@ABC
r o / v o / GFED
@ABC
D o 1 3 5 2 E

Fig. 7.2 Steps for determining the tour of vehicle 1

10. Marking the current solution as tabu-active: The tabu-active characteristics

of the current solution are stored in the tabu list.
11. Updating the best found solution: If the new current solution possesses a bet-
ter objective function value than the best found solution so far, it is replaced by
the new current solution.
Here, the current iteration of the proposed Tabu Search metaheuristic is com-
pleted and the newly generated current solution is used as starting solution in the
next iteration.

7.1.3 Representation of a Tour Plan in the Solution Method

In this section, an efficient data structure for modeling tour plans in the solution
method is described (cf., e.g., Nanry and Wesley Barnes 2000). In this data structure,
a tour plan is stored as a doubly linked list (see Fig. 7.1). The nodes of this list
represent the relevant locations (i.e., requests and the depot) in the current situation
of the transportation process and the currently utilized vehicles. The first node in this
list is always the depot D and the last node is an artificial end tour plan element E.
Moreover, each request i is modeled by node ri and each vehicle j is modeled by
node vj . The tour of vehicle k is determined by all request nodes that are located
to the left of vehicle k represented by vehicle node vk in the doubly linked list up
to the next vehicle node or the depot node, respectively. Each vehicle starts its tour
at its current position determined by the vehicle node, services the assigned request
nodes and lastly travels back to the depot node.
Note that unused vehicles waiting at the depot are not generally included in the
tour plan and need to be inserted by a neighborhood operator. In the example given
in Fig. 7.1, the tour of vehicle 1 is therefore described as v1 → r1 → r2 → r4 →
r6 → D and vehicle 2 has the tour v2 → r3 → r5 → D. In Fig. 7.2, the tour of
vehicle 1 is illustrated.
190 7 The Proposed Tabu Search Solution Method

7.1.4 Efficient Calculation of Changes in the Objective Function

Value

As stated in step 4 in Sect. 7.1.2, the change in the objective function value, i.e., the
move value, has to be evaluated for each neighbored solution within an iteration.
Clearly, this evaluation can be realized by calculating the objective function value
of the complete tentative tour plan. However, since this consumes computational
time and the solution method is executed under a given time limit, it is advisable
to determine the move value in more computational time-efficient ways. In what
follows, we present two aspects which are utilized in the proposed solution method
for efficiently calculating the move value. The first aspect consists of two important
characteristics of the considered objective function value: (a) the objective function
value of a tour plan is equal to the sum of all tours in the tour plan and (b) the
objective function value of each tour is independent of the other tours. Hence, by
storing the objective function values for each tour instead of storing only one value
of the complete tour plan, the move value can be computed by only calculating dif-
ferences in the objective function of tours which are modified by the current neigh-
borhood move. In the second aspect, systematically examining the neighborhood of
the current neighborhood operator enables the re-use of partial results of computed
objective function values calculated in previously examined neighbored solutions.
This reduces the computational effort of calculating changes in the objective func-
tion value (cf., e.g., Badeau et al. 1997). An example of systematically exploring
the neighbored solutions is shown for the neighborhood operator REL which is de-
scribed in Sect. 7.2.2.

7.2 Neighborhood Operators

In this section, the neighborhood operators that are applied in the proposed Tabu
Search metaheuristic are presented. Note that all utilized neighborhood operators
are node-exchanging operators, i.e., they move one or more requests in the tour plan
to another location within the same vehicle tour or onto another vehicle tour. All
utilized neighborhood operators only consider feasible tour plans which fulfill all
restrictions of the mathematical model defined in Sect. 5.3.2.3.
First, intensifying neighborhood operators are described, i.e., Within Tour In-
sertion (WTI) and Relocate (REL) while neighborhood operators possess more di-
versifying characteristics, i.e., MultiRelocate (MREL), Large Neighborhood Search
(LNS), and Exchange Between Tours (XBT) are presented afterwards. Besides de-
scribing the utilized neighborhood operators in detail, they are also presented in
pseudo-code. In order to comply with the existing time limit, all neighborhood op-
erators frequently check the remaining available computational time during their
execution. Hence, our computational experiments show that the available compu-
tational time limit is always met by the proposed Tabu Search solution method.
In what follows, we use the function U (a, b) which returns a uniformly distributed
continuous number from interval [a, b] as well as the function u(a, b) which returns
a uniformly distributed discrete number from interval [a, b].
7.2 Neighborhood Operators 191

7.2.1 Within Tour Insertion (WTI)

The WTI neighborhood operator removes a single request and reinserts it at its best
position within the same tour. Specifically, for each tour the best performing local
relocation that is not tabu-active is implemented if it leads to a better solution. In
order to provide maximum intensification, this operator only allows request relo-
cations that lead to a better solution. Therefore, within each application, for each
tour of the tour plan, WTI removes a single request and reinserts it at its best po-
sition within the same tour if this leads to a solution improvement. This search is
exhaustive, i.e., an evaluation is made for each request in the tour plan within one
move while storing the best request relocation for each tour. After the evaluation,
the neighborhood move consisting of all stored request relocations is applied to the
tour plan unless this leads to a solution that is marked as tabu. In the latter case,
the best found request relocations are applied successively to the tour plan in order
of non-increasing objective function value improvements. After the application of
each single request relocation, it is checked whether the current tour plan is tabu-
active. If not, the current set of performed request relocations is stored as the current
maximum implementable set WTI max of request relocations that lead to a non-tabu
solution. After evaluating all request relocations, the neighborhood move consisting
of the first WTI max stored tour moves is applied in order of non-increasing objective
function value improvements to the tour plan. The pseudo-code of the WTI neigh-
borhood operator is depicted in Algorithm 7.1.

Algorithm 7.1 Neighborhood operator Within Tour Insertion (WTI)

1: for each request r ∈ RτU+ do
2: Remove r from Pτt∗+
3: Reinsert r at all positions within its original tour
4: Store the best improving non-tabu tour move for each tour
5: end for
6: Sort best tour moves by their objective function value improvement (highest
first)
7: Perform the combined neighborhood move consisting of the consecutive appli-
cation of all stored best tour moves that provide an improvement
8: if combined neighborhood move is tabu-active then
9: Perform first tour move. This move can be executed non-tabu, otherwise, it
would not have been stored
10: Perform improving best tour moves consecutively
11: Store the highest number WTI max of tour moves that lead to a non-tabu solu-
tion
12: Perform the neighborhood move consisting of the WTI max tour moves in con-
secutive order
13: end if
192 7 The Proposed Tabu Search Solution Method

7.2.2 Relocate (REL)

The REL neighborhood operator works identically with the previously described
WTI neighborhood operator except for the fact that REL evaluates all positions
for request relocation that are on tours other than the one the selected request is
currently assigned to. Hence, it also evaluates the benefits of additionally using cur-
rently idle and unused vehicles. Similar to the WTI neighborhood operator, REL
also simultaneously performs multiple request relocations in one iteration, if pos-
sible. Since a neighborhood move of REL modifies two tours for each relocated
request, both modified tours are marked as changed in order to exclude them from
further request relocation operations in the current neighborhood move. Analogous
to the neighborhood operator WTI, all stored request relocations are subsequently
applied in order of non-increasing objective function value improvements. Poten-
tially occurring solutions which are tabu-active are avoided in the same way as de-
scribed in WTI. In order to allow an increased diversification compared to WTI,
in case of non-existing improving request relocations in the current neighborhood,
the best neighborhood move that leads to a non-tabu solution is applied, i.e., the one
that generates the slightest cost increase. The pseudo-code of the REL neighborhood
operator is depicted in Algorithm 7.2.

Algorithm 7.2 Neighborhood operator Relocate (REL)

1: for each request r ∈ RτU+ do
2: Remove r from Pτt∗+
3: Reinsert r at all positions on tours other than the original tour
4: Store the best improving non-tabu move for each tour from which r is re-
moved. Furthermore, remember the second tour in this move on which r is
relocated to
5: end for
6: Sort best tour moves by their objective function value improvement (highest
first)
7: Perform best tour move regardless of improvement value. If further improving
tour moves are available, perform the combined neighborhood consisting of
consecutive application of improving best tour moves that only change tours
which have not already been modified by previously applied tour moves
8: if combined neighborhood move is tabu-active then
9: Perform first tour move. This move can be executed non-tabu, otherwise, it
would not have been stored
10: Perform improving best tour moves consecutively
11: Store the highest number RELmax of tour moves that lead to a non-tabu solu-
tion
12: Perform the neighborhood move consisting of the RELmax tour moves in con-
secutive order
13: end if
7.2 Neighborhood Operators 193

r6 Request r6 to be reinserted into the following tour plan:

D r1 r3 r2 r4 v1 r5 ...

Evaluated insertions of r6 :

D r6 r1 r3 r2 r4 v1 r5 ...

D r1 r6 r3 r2 r4 v1 r5 ...

...

D r1 r3 r2 r4 r6 v1 r5 ...

D r1 r3 r2 r4 v1 r6 r5 ...

D r1 r3 r2 r4 v1 r5 r6 ...

...

Fig. 7.3 The first steps of systematically reinserting request r6 into the tour plan using the neigh-
borhood operator REL

As described in Sect. 7.1.4, the evaluation of the current neighborhood can be

performed in a time-efficient way which is described in what follows at the example
of REL. For this purpose, let c(t) denote the objective function value of tour t of a
tour plan P . The origin tour from which request r will be removed is defined by tro
and trd represents the destination tour to which r will be relocated to. With tro− rep-
resenting tour tro after request r has been removed and trd+ defining trd into which r
is inserted at the next position to be evaluated, according to the used neighbor-
hood operator. Using these definitions, the resulting move value can be calculated
by c(tro− ) + c(trd+ ) − c(tro ) − c(trd ). As can be seen, for a fixed origin and destination
tour, c(trd ), c(tro ), and c(tro− ) remain constant so that only c(trd+ ) changes depending
on the reinsertion position of r. Moreover, by enumerating the insertion positions of
the destination tour in ordered sequence, the computational effort required for cal-
culating c(trd+ ) in each neighbored solution can be reduced. As depicted in Fig. 7.3
at the example of reinserting request r6 , the first part of tour trd+ up to the current
6
insertion position of r6 remains constant and is extended by exactly one request in
each step of the neighborhood evaluation. Hence, the objective function value of
the first part of trd+ up to r6 can be calculated in constant time so that only the sec-
6
ond part needs to be re-evaluated in each step. This efficient type of evaluation can
be found in Badeau et al. (1997) in which also further techniques for approximat-
ing the objective function value of the second part of trd+ in a time-efficient way
6
are described. Since in the considered RDOPG applications tour lengths are quite
194 7 The Proposed Tabu Search Solution Method

short, these further approximation techniques are not implemented in the proposed
approach.

7.2.3 MultiRelocate (MREL)

The neighborhood operator MREL simultaneously removes n requests from the tour
plan where n is randomly drawn from interval [MRELreqMin , MRELreqMax ]. After-
wards, the removed requests are successively reinserted at their least cost positions
in the tour plan. Since a sub-procedure of the REL neighborhood operator is applied
for this reinsertion, MREL also evaluates benefits of using additional idle and un-
used vehicles. Note that the sequence in which the removed requests are iteratively
reinserted may have considerable effect on the resulting solution. Therefore, MREL
considers all possible permutations for reinserting the removed requests into the tour
plan. In order to make a time-efficient use of previously considered permutations,
evaluated permutations are iteratively modified from last positions to first positions.
Clearly, due to the exponential number of possible request removals, an exhaus-
tive examination of all existing subsets comprising n requests in an iteration is not
reasonable. Therefore, the neighborhood operator focuses on requests that are cur-
rently assumed to be inefficiently positioned in the tour plan. These requests are
identified by a high contribution value to the objective function value. This con-
tribution is individually determined for each request r by calculating the objective
function value improvement that occurs when only r is removed from the tour plan.
Subsequently, all requests are ordered in a list denoted as R w by non-increasing
objective function contribution values. Since requests with highest contribution
values are located at the beginning of this list, all requests with a later position
than n · MRELreqSelR are erased from R w . Afterwards, n requests that are consid-
ered by MREL in the current evaluation are randomly chosen from R w forming the
set R ∗ .
Within one move, the described selection, removal, and reinsertion of requests
is repeated c times, where c is randomly chosen from interval [MRELexecMin ,
MRELexecMax ]. The best found non-tabu neighborhood move is stored and finally
executed. The MREL neighborhood operator is described in pseudo-code in Algo-
rithm 7.3.

7.2.4 Large Neighborhood Search (LNS)

In each iteration of the solution method in which the LNS neighborhood operator is
applied, LNS iteratively evaluates the relocation of a chosen number of n requests
by removing and subsequently reinserting them at their least cost positions in se-
quence of non-increasing objective function contribution values. Since LNS also
uses a sub-procedure of REL for reinserting each individual request, the possibility
7.2 Neighborhood Operators 195

Algorithm 7.3 Neighborhood operator MultiRelocate (MREL)

1: c ← u(MRELexecMin , MRELexecMax )
2: n ← u(MRELreqMin , MRELreqMax )
3: Generate set R w : Select the n · MRELreqSelR worst requests from set RτU+ , i.e.,
the set R w contains the requests that contribute most to the objective function
value
4: for i ← 1 to c do
5: Generate set R ∗ by randomly selecting n requests out of R w
6: Remove each request r ∈ R ∗ from Pτt∗+
7: Reinsert all removed requests into the tour plan at least cost insertion posi-
tions. Perform this reinsertion consecutively in all possible permutations and
store the best non-tabu neighborhood move
8: end for
9: Perform the best stored neighborhood move

of utilizing additional unused vehicles at the depot is also considered in this neigh-
borhood operator. The selection and reinsertion process is repeated c times where c
is randomly chosen from interval [LNSexecMin , LNSexecMax ].
Compared to the MREL operator, the LNS neighborhood operator provides a
substantially higher variety of the number n of requests that are simultaneously
considered in one move. Note that this has been shown to be efficiently appli-
cable to VRPs (cf. Ropke and Pisinger 2006). In each LNS iteration, n is set to
max(2, m · |RτU+ |), where m is randomly drawn from interval [0, LNSreqRemF ],
LNSreqRemF < 1. These n requests are randomly chosen from list R w that was in-
troduced in MREL. Analogous to MREL, the size of the set R w depends on n, in
this case it holds that |R w | = min(|RτU+ |, n · LNSreqSelR ). Consequently, since n may
be much larger than in MREL, it is computationally intractable to evaluate all exist-
ing n! permutations for reinsertion. Therefore, LNS reinserts the selected requests in
one particular sequence as mentioned above, that is, in sequence of non-increasing
objective function contribution values. The pseudo-code of the LNS neighborhood
operator is shown in Algorithm 7.4.

7.2.5 Exchange Between Tours (XBT)

In the neighborhood operator XBT, two requests which are located on different tours
are exchanged so that their positions in both tours are preserved. Due to the fixed
reinsertion positions, diversification is attained. Similar to the WRT neighborhood
operator, this neighborhood operator works exhaustively, i.e., it considers all exist-
ing pairs of requests. After evaluating the possible combinations of two requests, the
best found non-tabu move is implemented. The pseudo-code of this neighborhood
operator is depicted in Algorithm 7.5.
196 7 The Proposed Tabu Search Solution Method

Algorithm 7.4 Neighborhood operator Large Neighborhood Search (LNS)

1: c ← u(LNSexecMin , LNSexecMax )
2: m ← U (0, LNSreqRemF )
3: n ← u(2, m · |RτU+ |)
4: Generate set R w : Select the min(|RτU+ |, n · LNSreqSelR ) worst requests from
set RτU+ , i.e., the set R w contains the requests that contribute most to the ob-
jective function value
5: for i ← 1 to c do
6: Generate set R ∗ by randomly selecting n requests out of R w
7: Remove each request r ∈ R ∗ from Pτt∗+
8: Reinsert all removed requests into the tour plan. Due to complexity reasons,
do this only in one specific permutation. This permutation is determined by
reinserting r ∗ of R ∗ first, i.e., the request that has the biggest contribution to
the objective function out of r ∗ of R ∗ . Store the best non-tabu neighborhood
move
9: end for
10: Perform the best stored neighborhood move

Algorithm 7.5 Neighborhood operator Exchange Between Tours (XBT)

1: for each request r ∈ RτU+ do
2: for each request r located on a tour other than r do
3: Exchange r and r while preserving their tour positions
4: Store best non-tabu neighborhood move
5: end for
6: end for
7: Perform the best stored neighborhood move

7.3 The Implemented Tabu List

As mentioned in Sect. 7.1.1, in order to avoid cyclical computations in the search
process, Tabu Search approaches make use of a tabu list. For this purpose, the defini-
tion of the tabu-active attribute which is used to identify tabu-active solutions is an
important task. Since solutions should be forbidden only if they have already been
examined by the search process, the practicability of a tabu-active attribute depends
on its ability to correctly identify solutions. Due to the exponential number of possi-
ble solutions, a unique identification cannot be reasonably implemented. Therefore,
a reasonable compromise between uniqueness and computational effort has to be
found.
In the literature, many authors use the objective function value of a solution as
the tabu-active attribute (cf., e.g., Badeau et al. 1997 and Gendreau et al. 2006).
However, our empirical studies for the considered RDOPG problems indicate that
significantly different solutions may have the same objective function value. Specifi-
cally, tour plan constellations can occur in which several dummy customer requests i
7.3 The Implemented Tabu List 197

can be reached by vehicles at or before rie . Since dummy customer requests have no
cost contribution to the objective function value before the time point rie , signifi-
cantly different tour plans might not be distinguishable by their objective function
value. Consequently, these different tour plans could be falsely identified as equal
and therefore erroneously considered tabu-active. Furthermore, in our real-time con-
trol approach, another reason speaks against using the objective function value as
a tabu-active attribute. Specifically, since the considered static system state is be-
ing modified during the transportation process, the objective function values of all
solutions also changes over different static problem instances. Therefore, using the
objective function value as a tabu-active attribute does not allow for identification
tour plans over different considered static problem instances.
In order to overcome these limitations, a fingerprint of the tour plan for real-
izing the tabu-active attribute is used. This fingerprint is based on the sequence
of requests in the tour plan and their assignment to vehicles. This method allows
the Tabu Search solution method to identify the same tour plans even between
consecutive static problems. The fingerprint is implemented by a combination of
the two quickly computable checksums CRC-32 (cf. Moon 2005, p. 147 et seq.)
and Adler-32 (cf. Deutsch and Gailly 1996). For a given tour plan, these meth-
ods calculate checksum values by using the tours of the tour plan as input data,
i.e., by considering the sequence of requests and their assignment to the vehicles.
In these checksum calculations, each request is represented by a unique request
ID and a value of the assigned vehicle which depends on its status. Specifically,
if the vehicle has already been utilized, i.e., has left the depot, a unique vehicle
ID is used. Otherwise, the same value for all unused vehicles waiting at the depot
is used. Using this fingerprint, two tour plans P1 , P2 are identified as identical if
both described checksums are equal in the tour plans, i.e., cscrc32 (P1 ) = cscrc32 (P2 )
and csadler32 (P1 ) = csadler32 (P2 ). Note that in Stone (2001, p. 128 et seq.) and Stone
et al. (2002), the authors state that the Adler-32 checksum has some drawbacks when
it is applied to short input data. Specifically, since the Adler-32 checksum is basi-
cally calculated by the summation of individual values, in case of short input data,
only small checksum values are usually generated. Hence, in problem instances with
only a few requests, Adler-32 may have an increased tendency to produce equal
checksums for different tour plans of short length. However, since CRC-32 does not
possess this drawback and the combination of both checksums is always used to
identify the tabu-active state of a solution, no disadvantages arise. The Adler-32 al-
gorithm is used due to its fast computation time which additionally helps to reduce
the false detection rate of different solutions in larger problem instances.
Besides the implementation of the tabu-active attribute, the tabu list requires the
definition of the tabu tenure that determines the number of iterations for which
tabu-active attributes are valid. Since preliminary empirical studies in the consid-
ered RDOPG applications reveal an almost unique identification of solutions by
the applied fingerprint, a tabu-active solution is never withdrawn. This allows for
identification of solutions over static problem instances. Clearly, since each change
regarding the set of pending requests at time τ + defined by RτU+ makes former and
198 7 The Proposed Tabu Search Solution Method

Fig. 7.4 Empty tour plan GFED

@ABC
r1 @ABC
GFED
r2
with two real requests and
three dummy customer
requests (denoted as r3 , r4 , GFED
@ABC
89:;
?>=<
r3 @ABC
GFED
89:;
?>=<
r4 GFED
@ABC
89:;
?>=<
r5
and r5 ) to be inserted by the
construction heuristic GFED
@ABC / @ABC
GFED
D o E

r3 r4 r5

D r2 r1 v1 E

Resulting tour plan with inserted dummy customers:

D r2 r1 r3 v1 r5 r4 v2 E

Fig. 7.5 Real requests inserted by the construction heuristic. If dummy customers are used, these
are inserted afterwards

new solutions incomparable, the tabu list is reset whenever the pending request sit-
uation changes. Moreover, as mentioned in Sect. 5.2.2, the solution method returns
to the first stage and is directly applied to the solution that is currently in execution
in its pre-simulated state at τ + .

7.4 Construction of the Initial Tour Plan

Since the Tabu Search metaheuristic is an improvement heuristic, an initial solution

needs to be provided. Hence, before the real-time control of the considered trans-
portation process starts, an initial tour plan is constructed at the beginning of the
planning horizon.
In our approach, the initial tour plan is generated by an iterative insertion of all
real requests that have arrived before the execution of the transportation process.
These requests are inserted into the arising tour plan at their least cost position in
sequence of non-decreasing arrival times. Note that for these requests it is assumed
that their customer inconvenience starts at the beginning of the planning horizon.
Afterwards, if stochastic knowledge is utilized, the additional dummy customer re-
quests are inserted into the current tour plan at their least cost positions in sequence
of non-decreasing starting times of their time window. In Figs. 7.4 and 7.5, the con-
struction process is illustrated. Clearly, in both steps of the construction heuristic,
the possible utilization of additional unused vehicles in the tour plan is evaluated.
7.5 Stage-Based Neighborhood Operator Selection Scheme 199

Table 7.1 Stage-based selection scheme of the proposed Tabu Search solution method
Stage Operator Switch after number of iterations without new best solution

1 WTI 0 (only accept improvements)

2 REL 10
3 MREL 10
4 LNS 1,000
5 XBT 5

Table 7.2 Applied parameter

values for the neighborhood Parameter Value Parameter Value
operators MREL and LNS
MRELreqMin 2 LNSreqRemF 0.75
MRELreqMax 3 LNSreqSelR 1.5
MRELreqSelR 3 LNSexecMin 1
MRELexecMin 10 LNSexecMax 2
MRELexecMax 20

7.5 Stage-Based Neighborhood Operator Selection Scheme

In each iteration of the Tabu Search procedure, the neighborhood operator to be

applied has to be chosen. In the proposed solution method, the neighborhood oper-
ator is chosen using a deterministic neighborhood operator selection scheme. This
scheme has different stages where each stage consists of one of the neighborhood
operators that have been introduced above. In each iteration, the neighborhood op-
erator of the current stage is applied and the best found non-tabu move is executed.
After a predefined number of performed iterations in which no new best known so-
lution has been found, the stage is switched to the next one. If a new best known
solution is found or if the last stage of the scheme has been completed by the search
process, the search returns to the first stage and the current number of iterations
without improvement is reset.
Preliminary tests have been conducted in order to determine reasonable param-
eter settings for the stages as well as for the neighborhood operators. Table 7.1
provides a detailed definition of the implemented stage-based selection scheme.
Additionally, Table 7.2 shows information about chosen parameter values for the
neighborhood operators MREL and LNS.
In order to combine intensification and diversification aspects in the enumeration
process of the implemented Tabu Search procedure, each stage can be interpreted
as representing a specific assumed distance of promising, still-unexplored solutions
to the currently considered solution. Specifically, due to a large number of recent
consecutively failed attempts at improving the currently best known solution, this
distance is assumed to be significantly larger in higher stages than in lower stages.
Hence, by switching to higher stages, the search process increasingly applies more
200 7 The Proposed Tabu Search Solution Method

globally-acting neighborhood operators that provide a higher diversification. More-

over, whenever a move leads to a new overall best known solution, the search pro-
cess is intensified in its direct vicinity by switching back to the first stage.

7.6 Integration of the Solution Method into the Real-Time

Control Approaches
As described in Sect. 5.2.2, the proposed Tabu Search solution method is invoked in
both proposed real-time control approaches, i.e., in the deterministic and in the pro-
active approach, after the theoretical tour plan has been adapted to the anticipated
system situation at τ + . Furthermore, the current execution status of the solution
method, i.e., the current tabu list, the current stage, and the current number of it-
erations without improvement are stored at the end of the execution of the solution
method in the current anticipation horizon so that this information can be used in the
adaptation performed in the consecutive anticipation horizon. In so doing, the solu-
tion method is enabled to continue the search process in the same state over consec-
utive problem instances. If changes in the pending request situation have occurred,
the tabu list is reset and the solution method starts on the first stage of the neigh-
borhood operator selection scheme. This is performed since the changed request
situation in the system makes it likely that the application of intensifying neighbor-
hood operators quickly leads to improvements of the tour plan. After the available
computational time for the solution method has elapsed, the solution method returns
the best feasible tour plan that has been found during the application of the solution
method as well as the last evaluated theoretical tour plan to the real-time control
approach. In the update handling performed in the next anticipation horizon, the
real-time control approach utilizes these generated tour plans in order to select the
new relevant tour plan as described in Sect. 5.2.2.

7.7 Summary
This chapter introduced the solution method which is used to efficiently solve the
individual static problem instances generated by the deterministic and pro-active
real-time control approach during the execution of the transportation process. In or-
der to cope with the tight computational time restrictions imposed by the real-time
nature of the considered RDOPG applications, an elaborated Tabu Search meta-
heuristic solution method was proposed. It utilizes different neighborhood operators
which are selected by a deterministic stage-based selection scheme. After describ-
ing the Tabu Search metaheuristic in general, the utilized neighborhood operators
were presented in detail. Moreover, the applied tabu-active attribute in the tabu list
was explained. In the proposed approach, a fingerprint consisting of two quickly
computable checksums was used as the tabu-active attribute in order to allow an
almost unique identification of individual solutions. This fingerprint also allows the
7.7 Summary 201

identification of solutions across consecutive static problem instances. Furthermore,

the generation of the initial tour plan and the stage-based selection scheme was pre-
sented. As mentioned above, this scheme is used to select the neighborhood operator
which is to be applied in each iteration of the Tabu Search solution method. After
a predefined number of iterations without finding a new best global solution, the
current stage is switched to the next higher one. In doing so, neighborhood oper-
ators which provide a higher level of diversification are applied in the subsequent
iterations. When a new best global solution is found, the stage is reset to the first
one so that an intensifying neighborhood operator is applied in the next iterations.
Moreover, the integration of the proposed Tabu Search solution method into both
proposed real-time control approaches was described.
Chapter 8
Computational Results

In this chapter, we present the results of our detailed computational experiments

conducted on the proposed real-time control approaches. The experiments are car-
ried out in order to answer our three research questions posed in Sect. 1.3.1. Since
no benchmark instances are available for the considered RDOPG applications, ap-
propriate test instances were generated. Each test instance consists of a problem
instance and a setting. A problem instance defines given problem characteristics
such as the existing request data, customer inconvenience function, and the avail-
able vehicle fleet size. Hence, the problem instance represents the unchangeable
system environment. In contrast, a setting comprises selectable parameters. Besides
choosing the real-time control approach to be utilized for adapting the transporta-
tion process during its execution, this also includes defining appropriate parameters
for it. Furthermore, if the pro-active real-time control approach is used, parameters
for generating and integrating stochastic knowledge are defined in the setting.
Before presenting the results of the computational experiments, we first intro-
duce the discrete event-based simulator which simulates the system environment of
RDOPG applications including vehicle activities and dynamically arriving requests.
This is followed by the description of different settings evaluated in our compu-
tational experiments consisting of test environment, parameter values, and request
data classes. After that, the computational experiments are conducted in eight steps:
1. The performance of the Tabu Search solution method that is applied for solving
individual static problem instances which occur during the transportation process
is analyzed by comparing its performance to a lower bound of an exact solution
procedure.
2. A detailed performance evaluation of deterministic real-time control approaches
is provided. Specifically, the proposed deterministic real-time control approach
is compared to a standardized least cost insertion procedure that integrates newly
incoming requests. Moreover, a suitable time limit for the application of the Tabu
Search metaheuristic in the deterministic real-time control approach is derived.
3. A detailed performance evaluation of allowing vehicle en-route diversion is pro-
vided. In this comparison, the deterministic real-time control approach is com-
pared to a variant of this approach in which vehicle en-route diversion is for-

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 203

DOI 10.1007/978-3-642-33472-6_8, © Springer-Verlag Berlin Heidelberg 2013
204 8 Computational Results

bidden. Test results illustrating the positive effects of allowing vehicle en-route
diversion on the solution quality are provided.
4. The efficiency of measuring structural distortion using the Goodness-of-Fit test
used in our forecasting methods is evaluated. This evaluation is conducted on
request data sets that have different levels of structural distortion. Moreover, a
suitable value for the minimum demanded Poisson quality parameter for gener-
ating stochastic knowledge is derived.
5. Using the derived Poisson quality parameter value, results of generating stochas-
tic knowledge for S REAL and S GEN , which is used in our main computational
experiments, are presented.
6. Based on the results of the previous five evaluations, a comprehensive compu-
tational study regarding the performance evaluation of the pro-active real-time
control approach is provided in the main computational experiments of this book
in order to answer the first and second research question. Specifically, a detailed
analysis of the conditions under which the generated stochastic knowledge can
be reasonably utilized in the pro-active real-time control approach is carried out.
This analysis is conducted by assessing the performance of the proposed pro-
active real-time control approach in comparison to its deterministic counterpart,
i.e., the deterministic real-time control approach which does not make any use of
stochastic knowledge. In addition to an existing real-world request data set, test
studies on designed request data sets are conducted.
In order to analyze under which conditions the use of the proposed pro-active
real-time control approach is recommendable, we identify crucial characteristics
of the evaluated test instances. From the results of this study we conclude that
the structural quality of the existing request data set is one of these crucial char-
acteristics. Specifically, we propose that the structural quality of a request data
set depends on its structural distortion and structural diversity:
• Structural distortion: In the proposed forecasting approach it is assumed that
request arrivals can be modeled by non-homogeneous time-space Poisson pro-
cesses whose rate parameters remain constant from day to day (see Sect. 6.2).
If this assumption is not fulfilled in the existing request data set, future requests
are likely to occur with significant differences from what was forecasted. In
this case, utilizing the stochastic knowledge generated by the proposed fore-
casting method is likely to lead to a worse solution quality. Hence, we state that
it is important that the available request data has a low structural distortion.
In our proposed forecasting approach the Chi-Square Goodness-of-Fit test is
applied in order to measure the Poisson quality of a given request data set (see
Sect. 6.2.1). We evaluate the efficiency of this test for identifying the structural
distortion on request data sets with different levels of structural distortion in
Sect. 8.6.
• Structural diversity: In addition to requiring a low structural distortion, we
classify the structural quality of request data sets according to the changes
which occur in the request arrival rates in individual subregions during the con-
sidered planning horizon. We define that request data sets with nearly constant
request arrival rates comprise a low structural diversity while request data sets
8.1 The Discrete Event-Based Simulator 205

with significantly changing request arrival rates in the subregions have a high
structural diversity. In order to analyze the effects of structural diversity on
the attainable improvements, we generate test instances with designed levels
of structural diversity (see Sect. 8.2.4.2) and evaluate the attainable improve-
ments depending on the existing level of structural diversity. Moreover, we
propose two methods for determining the structural diversity of given request
data sets in Sects. 8.8.1 and 8.9.
According to the above criteria, a request data set is considered to have a high
structural quality if it has a low structural distortion and a high structural diver-
sity. Therefore, the structural quality of a request data set determines the value
of the stochastic knowledge which can be generated out of this request data set.
Consequently, a high structural quality indicates that high improvements of the
solution quality can be achieved if the proposed pro-active real-time control ap-
proach is used compared with the results attained by its deterministic counterpart.
7. In the previous computational experiments, it was shown that high levels of
structural diversity significantly increase the improvements attainable by the pro-
active real-time control approach. Based on these insights, we propose a measure
for determining the structural diversity of arbitrary request data sets. This mea-
sure is denoted as the degree of structural diversity and its significance is statis-
tically proven.
8. In the third evaluation, the advantages of allowing vehicle en-route diversion in
the considered RDOPG applications are shown. However, allowing vehicle en-
route diversion may increase driver inconvenience since drivers are obliged to
quickly react to occurring changes in their tour while they are traveling to their
next assigned request. In order to deal with this issue, we additionally consider
driver inconvenience aspects in the real-time control approaches. For this pur-
pose, a method for limiting the number of performed vehicle en-route diversion
activities is developed. The effects of this method on the solution quality and
resulting driver inconvenience is assessed. Recommended parameter settings for
achieving a suitable trade-off between the contradicting aims of driver inconve-
nience and customer inconvenience as well as parameter settings for prioritizing
each of the aims are provided.
After these computational experiments, we conclude our evaluations by giving a
summary and answers to the three central research questions which were posed in
Sect. 1.3.1.

8.1 The Discrete Event-Based Simulator

In order to analyze the real-time control approaches under real-time conditions, all
transportation processes of the considered RDOPG applications are carried out by
an implemented discrete event-based simulator. In discrete event-based simulators,
the simulation time is always set from the current time to the time point when the
206 8 Computational Results

Execution of the transportation process

Process 2

Process 1

Waiting time 2 Activity 2

(eq. to response time 2)

Waiting time 1 Activity 1

(eq. to response time 1)

Time point 0: Event 1 Event 2 Event 3 Event 4 Event 5 Event 6 End of the time
Beginning of the execution of the
execution of the Arrival of Arrival of Start of Start of End of End of transportation
transportation request 1 request 2 service of service of service of service of process:
process request 1 request 2 request 1 request 2 All requests have
been serviced,
Request 1 is Request 2 is all vehicles are
completed, completed, back at the depot
leaves the leaves the
system system

Fig. 8.1 Interrelation of event, process, and operation using the example of servicing two re-
quests (cf. Biethahn et al. 1999, p. 52)

next event in the system takes place as mentioned in Sect. 4.7.3. In addition to con-
ducting the computational experiments, the implemented simulator is also used for
pre-simulating the system state at the beginning of each anticipation horizon. This
pre-simulation concept was introduced in Sect. 5.2.1 in order to generate the antic-
ipated system state at the end of the anticipation horizon so that the concurrency of
the execution and adaptation of the transportation process is considered. By utilizing
this pre-simulated system state, the static problem instance for the current anticipa-
tion horizon is generated (see Sect. 5.2.2). In the individual anticipation horizons
conducted by the real-time control concept presented in Sect. 5.2.1, it is assumed
that the real vehicle activities are in line with the activities that have been antici-
pated.
As mentioned in Sect. 4.7.3, a system consists of entities that act or interact to-
gether towards the accomplishment of some logical end (cf. Schmidt and Taylor
1970). Hence, in order to model a complex dynamic system such as RDOPG ap-
plications, it is necessary to identify all entities of the system. In the considered
system, the entities consist of requests and vehicles. Dynamically arriving requests
trigger so-called exogenous events that unexpectedly change the relevant informa-
tion of the transportation process from outside the system since new requests arrive
with uncertainty. In contrast, vehicle activities trigger endogenous events that mod-
ify the relevant information in ways that are generated by the system itself and can
be expected with high certainty, such as request servicing activities performed by
vehicles.
As depicted in Fig. 8.1, the executed transportation process can be subdivided
into several individual processes. A process is defined by mutually dependent events
and intermediate operations. A new process starts when a request arrives in the sys-
tem. It consists of exactly two operations: Waiting time and activity. The waiting
8.1 The Discrete Event-Based Simulator 207

time operation directly starts upon the arrival of a new request. In the RDOPG ap-
plications, the length of the waiting time determines the response time and hence
the resulting customer inconvenience. The activity operation starts when a vehicle
reaches a customer request location. This operation is defined as an activity since the
vehicle that services the request cannot perform other tasks during this activity. The
length of this activity is defined by the service time of the request. In Fig. 8.1, the
interrelation between the participating entities in the system is illustrated. In what
follows, we describe the simulation of requests as well as the simulation of vehicles.

8.1.1 Simulation of the Requests

As mentioned in Sect. 3.1, dynamism in the considered RDOPG applications arises

from newly arriving requests. The requests that arrive during the transportation pro-
cess are read from a corresponding request data set file that contains all relevant re-
quest attributes, i.e., their arrival time and location. These request arrivals are stored
in a list called future event set (FES). Since all request-related attributes are speci-
fied by the request data set, it is ensured that the same problem instance can be run
several times using different settings so that different variants of the real-time con-
trol approaches can be compared with each other under the same conditions. During
the execution of the transportation process, the request information is revealed to
the simulator over time at the arrival times of the individual requests. The integra-
tion of the newly arrived requests is performed by the real-time control approach as
described in Sect. 5.2.1.
In the computational experiments, relevant performance statistics observed dur-
ing the simulation are recorded for each test instance in order to allow a perfor-
mance analysis of the utilized setting under the given problem instance. Among
these statistics, the request response times and the corresponding objective function
values are determined. According to the objective function values of all requests,
the solution quality for each test instance is determined by calculating the average
customer inconvenience per request. This value is used as the main quality criterion
in all computational experiments presented later. Since satisfaction of a maximum
response time for all requests is also pursued, the number of late requests that vi-
olate this maximum response time is considered as another important performance
indicator.

8.1.2 Simulation of the Vehicles

In each anticipation horizon, the vehicle activities are simulated by executing the
current relevant tour plan. Specifically, at time point τ , after the relevant tour plan Pτr
has been determined, each vehicle executes its tour according to Pτr for t a time units.
Depending on the current vehicle’s status (see also Sect. 5.3.2.2), this execution
comprises the following activities:
208 8 Computational Results

• Traveling: The vehicle travels in the road network to the next assigned request of
its tour.
• Servicing: The vehicle services a request at the corresponding request location.
• Waiting at a dummy customer request: The vehicle waits at a dummy customer
request location.
• Waiting at the last request location: The vehicle waits at the location of its last
serviced request if the vehicle is currently idle since it has no further requests
assigned.
• Waiting at the depot: The vehicle waits at the depot location if it is unused.
Within the time period [τ, τ + ), it is possible that a vehicle performs more than one
of these activities if the current activity requires less time than is available in the
current anticipation horizon.

8.2 Test Environment, Parameter Values, and Request Data

Classes
In this section, we describe settings and parameters which are evaluated in the com-
putational experiments. First, we describe the system environment and real-time
control parameter values under which the computational experiments are evaluated.
Furthermore, utilized tabu search solution method parameter values are presented.
Afterwards, appropriate dummy customer request parameter values that have been
obtained in preliminary tests are described. This is followed by presenting the re-
quest parameter values and request data classes which are evaluated in the com-
putational studies. Moreover, the number of conducted computational experiments
is presented. Finally, the performance evaluation using the additionally attainable
improvement is described.

8.2.1 System Environment and Real-Time Control Parameter

Values

In order to analyze the performance of the different evaluated real-time control

approach variants, all generated transportation processes are executed utilizing a
detailed road network of Dortmund, a German medium-sized city. The road net-
work is made available for our research by DDS Digital Data Services GmbH in
Karlsruhe, Germany. The considered service region in this city comprises an area
of 22.5 kilometers × 20 kilometers = 450 square kilometers. Since only the ur-
ban area is considered, the utilized road network is quite small comprising about
17,300 nodes and 42,000 arcs. Hence, shortest path calculations can be performed
within a short time, i.e., within a few milliseconds. In all experiments, the length of
the anticipation horizon is set to t a = 20 seconds. Moreover, before each transporta-
tion process the initial generated tour plan is optimized by the Tabu Search solution
method for 120 seconds.
8.2 Test Environment, Parameter Values, and Request Data Classes 209

The experiments are conducted with fleet sizes of 8, 10, and 12 vehicles. These
fleet sizes have been empirically chosen from preliminary tests. They represent in-
teresting constellations for the number of requests in the considered request sets.
Specifically, while a fleet size of 8 vehicles represents a scenario with a very re-
stricted set of available resources, the 12 vehicle setting provides significantly more
flexibility for vehicle dispatching and pro-active strategies.
All proposed real-time control approaches and the segment-based forecasting ap-
proach were implemented using Borland Delphi while specific request generation
and evaluation tools were developed using Microsoft Excel. The implemented appli-
cations comprise in total about 50,000 lines of code. The computational experiments
were run on Intel Pentium D 2.8 GHz CPU (2.5 GB RAM) computers.

8.2.2 Tabu Search Solution Method Parameter Values

As illustrated in Sect. 7.2, some neighborhood operators in the proposed Tabu

Search metaheuristic utilize random numbers during their application. For gener-
ating these random numbers, we use an extended version1 of the random number
generator proposed by Park and Miller (1988). In this extended version, an im-
proved multiplicator (cf. Park et al. 1993) is applied and an individual stream of
random numbers is provided for each randomly drawn parameter in all neighbor-
hood operators of the Tabu Search method.
As mentioned in Sect. 5.2.2, static problem instances are adapted by the Tabu
Search solution method which is applied for a given time period. Note that due to
this time-limit based optimization, slightly imprecise operating system based mea-
surements of the used CPU time may occur which may lead to a variation of the
number of performed iterations in different replications of the same evaluated test
instance. Consequently, this takes considerable effect on the random number genera-
tion process applied in some of the neighborhood operators of the Tabu Search solu-
tion method. Specifically, in different replications, those stochastic features caused
by the real-time nature of the process can lead to varying performance results of the
Tabu Search procedure (cf. Gendreau et al. 2006).
In order to reduce these unwanted effects, we use a variance reduction tech-
nique denoted as Common Random Numbers (CRN, cf. Law and Kelton 2000,
p. 582 et seq.). By using this technique, all seeds of the random number streams
used in a test instance are generated before the start of the simulation in an of-
fline step. Therefore, in this offline step, a suitable upper bound on occurring static
problem instances during the transportation process is determined. In order to de-
rive suitable seed values for solving each static problem instance, we consecutively
generate 10,000 random numbers of each utilized random number generator stream
between two seed calculations. After each of these 10,000 random numbers, the re-
sulting set of seed values is stored as initial seeds of the streams of random numbers
in the corresponding static problem instance.

1 Available at http://www.cs.wm.edu/~va/software/park/.
210 8 Computational Results

8.2.3 Dummy Customer Request Parameter Values

In order to obtain useful parameter values for generating and integrating dummy
customer requests, extensive preliminary tests have been conducted. The obtained
results indicate that the following parameter values for dummy customer generation
and integration are reasonable:
• Spatial and temporal precision: According to results of preliminary compu-
tational experiments, the extension of each segment used in the segment-based
clustering approach is set to DCse × DCte = (2.5 kilometers × 2.5 kilometers) ×
1 minute. Therefore, the considered service area of size 22.5 kilometers ×
20 kilometers is divided into 9 × 8 subareas. Since the considered transporta-
tion process is controlled for four hours, we obtain the set S with altogether
|S| = 9 · 8 · 60 · 4 = 17,280 segments.
• Range of past request information: The number of past working days utilized
in the generation of the stochastic knowledge is set to nf = 60.
• Maximum cluster extension: The maximum spatial cluster extension DCmse is
set to 2 by 2 segments (i.e., 5 kilometers × 5 kilometers) while the maximum
allowed temporal cluster height (DCmte ) is set to 15 segments (i.e., 15 minutes).
• Dummy customer location: For determining a reasonable location for the
dummy customer, the maximum predefined travel time radius DCradiusTT is set
to 300 seconds whereas the maximum allowed average travel time within a clus-
ter DCmaxAvgTT is set to 650 seconds.
• Dummy customer prediction quality: The following five values are evaluated
for the minimum dummy customer quality DCminλ : DCminλ ∈ {1.0, 1.2, 1.5, 1.8,
2.0}. Consequently, in these cases, the expected occurrence probability of at least
one request in an accepted cluster amounts to at least 63.21 %, 69.88 %, 77.69 %,
83.47 %, or 86.47 %, respectively.
• Dummy customer removal: Two different parameter values are evaluated
for DCλrem , in particular DCλrem ∈ {0.50, 0.25}. Hence, a dummy customer is
removed when the arrival probability of at least one more request in the assigned
cluster becomes lower than 39.45 % or 22.12 %, respectively.

8.2.4 Request Parameter Values and Request Data Classes

The computational experiments are conducted on the following two request data
classes: Real-world request data class S REAL and generated request data class S GEN .
In addition to the equal request service time defined in Sect. 3.1 by R st = 60 sec-
onds, two further parameters are used for real requests as well as for dummy cus-
tomer requests in both request data classes. Specifically, in all experiments, the max-
imum allowed request response time R mrt is set to 3,600 seconds while the lateness
penalty costs R pen , that occur when R mrt is exceeded, are set to 100. As mentioned
in Sect. 5.3.2.4, since a request has a cost contribution of only 1.0 to the objective
8.2 Test Environment, Parameter Values, and Request Data Classes 211

function value at a response time equal to R mrt , the significantly higher lateness
penalty costs R pen are considered prohibitive costs. Clearly, due to the real-time na-
ture of the considered problem and depending on the available fleet size, the arrival
of additional dynamic requests may result in a scenario where some requests can
only be serviced after the maximum allowed request response time R mrt .
In what follows, we describe both evaluated request data classes in detail. First,
the real-world request data class S REAL is described followed by the generated re-
quest data class S GEN . Note that all request data sets of both request data classes
consist of 30 days. Consequently, in the computational experiments, each evalua-
tion of a setting results in 30 test instances.

8.2.4.1 Real-World Request Data Class S REAL

The request data class S REAL comprises one request data set of 30 randomly cho-
sen days out of over 400 weekdays and results from a case study of the subsequent
delivery process carried out by a German newspaper publishing company. In this re-
quest data set, 150 requests per day have to be processed on average. The subsequent
delivery of newspapers starts at 7 am. Since the majority of subsequent delivery re-
quests occur before 11 am, the computational experiments consider all requests that
arrive up to this point in time. Therefore, each simulated day consists of four hours
during which new requests may occur. Clearly, for generating stochastic knowledge,
for each of the 30 randomly chosen days of S REAL , the nf = 60 past working days
of the available real-world data are utilized.
Regarding request arrivals, an analysis of the provided data has shown that only
a few requests arrive before 7 am so that more than 90 % of all requests arrive after
7 am (see Fig. 8.2). Therefore, according to Larsen et al. (2002) and as mentioned
in Sect. 2.6.5, a high degree of dynamism can be identified for this request data set.
Note that, due to the real-world nature of this request data set, underlying request
distributions as well as other possibly existing non-deterministic structures are un-
known in S REAL . Hence, neither the structural distortion nor the structural diversity
of S REAL is known. Both criteria are evaluated by our proposed methods in this
chapter.

8.2.4.2 Generated Request Data Class S GEN

In contrast to S REAL , request arrivals in the request data class S GEN are designed
according to the following characteristics. This request data class comprises re-
quests that have been generated by time-space Poisson processes with time- and
space-dependent rate parameters which remain constant from day to day. The gen-
eration is carried out in a manner similar to the procedure described in Ichoua
et al. (2006). Specifically, the considered service area is subdivided into a set
of P quadratic disjunct regions of equal size a1 , . . . , aP while the considered time
212 8 Computational Results

Request arrivals

before 7:00- 7:30- 8:00- 8:30- 9:00- 9:30- 10:00- 10:30- 11:00- 11:30- 12:00- 12:30- 1:00- 1:30- 2:00- 2:30- after
begin 7:30 8:00 8:30 9:00 9:30 10:00 10:30 11:00 11:30 12:00 12:30 1:00 1:30 2:00 2:30 3:00 3:00

Time

Fig. 8.2 Average request arrival distribution of the request data class S REAL

period for new request arrivals from 6:45 am to 11 am is divided into Q differ-
ent disjunct time slices t1 , . . . , tQ which do not overlap each other. The parame-
ter λ(ai , tj ) determines the request arrival rate of a time-space Poisson process in
region ai during time slice tj . Moreover, tjs defines the start time and tje defines the
end time of time slice tj , j ∈ {1, . . . , Q}. Given the current time slice tj , the arrival
time of the next request is determined using a Poisson process with rate parame-

ter λ(tj ) = P i=1 λ(ai , tj ). The region in which this new request will occur is deter-
λ(a ,t )
mined according to the region probability values p(ai , tj ) = λ(ti j )j . Finally, within
the chosen region ai , the spatial position of the new request is uniformly drawn.
Hence, as suggested by Larsen (2000, p. 45), we utilize time-dependent Poisson
processes for generating request arrivals in the request data class S GEN . As in the
solution method, parameter values are generated by using an individual stream of
random numbers, which are provided by the extended version of the random num-
ber generator of Park and Miller (1988) and Park et al. (1993), for each type of
parameter.
Based on the rate parameters λ(ai , tj ) used in the applied setting, different levels
of spatial and temporal structural diversity, i.e., levels of variance in request arrivals,
occur. We introduce two dimensions of structural diversity: While RegionDiversity
(RD) measures the regional variance of request arrival rates in the service area within
one time slice, TimeDiversity (TD) gives the variance of request arrival rates in each
considered region ai of the service area during the considered planning horizon.
Moreover, TD also defines variations in λ(tj ) from one time slice to the other. In
order to evaluate the impact of different levels of structural diversity on the struc-
tural quality of request data sets, different request data sets comprising various value
settings of both structural diversity dimensions are designed in S GEN . All settings,
i.e., request data sets, are iteratively generated by starting from an initial setting
with a defined maximum level of structural diversity denoted as RD = TD = 1.00.
The structural diversity of this initial setting results from individually defined rate
parameters λ(ai , tj ). In subsequently generated request data sets with lower levels
of structural diversity, the rate parameter values are iteratively reduced by apply-
ing a linear averaging that balances the rate parameters λ(ai , tj ) accordingly. In
8.2 Test Environment, Parameter Values, and Request Data Classes 213

0.40
Request arrival probability per region
0.35 a4

0.30 a3
0.25

0.20 a2
0.15

0.10
a6
0.05 a5
a1
0.00
1.00 0.75 0.50 0.25 0.00

RegionDiversity (RD )

Fig. 8.3 Linear averaging at the example of RegionDiversity (RD)

Fig. 8.3, an example of this linear averaging process for resulting request arrival
probabilities is illustrated using P = 6 regions a1 , . . . , a6 and one time slice, i.e.,
Q = 1. As can be seen, setting RD = 1.00 results in a request data set in which
the defined probabilities are fully present in each region while RD = 0.00 results in
the same request arrival probability in each region. With regard to S GEN , by using
a step size of 0.25 and completing with a final setting RD = TD = 0.00, we ob-
tain 25 request data sets altogether. Apart from negligible distortion effects caused
by the real-world character of the road network, setting RD = TD = 0.00 produces
a complete uniform distribution of request arrivals during the considered planning
horizon. Moreover, the setting RD = 1.00, TD = 0.00 results in a request data set
with significantly different request arrival rates in the regions. However, request
arrival rates in the regions and the total request arrival rate per time slice remain
constant during the planning horizon. In contrast, the setting RD = 0.00, TD = 1.00
generates a request data set in which request arrivals have the same probability in
all regions but the total request arrival rate in a time slice changes between time
slices according to the initially defined values for RD = TD = 1.00. Algorithm 8.1
illustrates the generation of a request data set of the request data class S GEN with
defined levels of structural diversity RD and TD. In lines 1–13 the mentioned linear
averaging of the request arrival rates and request arrival probabilities is performed.
Specifically, in line 5, for each time slice tj the request arrival rate λ (tj ) is cal-
culated according to the chosen TD setting. In line 8, the average request arrival
probability over all time slices for region ai denoted as p̄ (ai ) is calculated depend-
ing on the initially defined request arrival rates and the selected RD setting. Based
on this value, the resulting request arrival probability of region ai in time slice tj is
calculated in line 10 and 11. According to these calculated values, in lines 14–26,
the request data set is generated as described above.
In all 25 generated request data sets of S GEN , the number of regions is set
to P = 18 and the number of time slices to Q = 5. The request arrival probabil-
ity parameters utilized for generating the setting RD = TD = 1.00 are illustrated in
Fig. 8.4. Note that this figure shows a simplified illustration of the road network,
214 8 Computational Results

Algorithm 8.1 Generation of a request data set of S GEN with RD ∈ [0, 1] and
TD ∈ [0, 1]
1: {Generate request arrival rate parameters based on given structural diversity
levels}
Q
2: nall ← j =1 λ(tj )
3: for j ← 1 to Q do
tje −tjs
4: n ← nall · tQe −t s
1
5: λ (tj ) ← n + (λ(tj ) − n) · TD
6: end for
7: for i ← 1 to P do
1 Q
8: p̄ (ai ) ← Q · j =1 ( P1 + (p(ai , tj ) − P1 ) · RD)
9: for j ← 1 to Q do
10: x ← P1 + (p(ai , tj ) − P1 ) · RD
11: p (ai , tj ) ← p̄ (ai ) + (x − p̄ (ai )) · TD
12: end for
13: end for
14: {Generate request data set}
15: for d ← 1 to number of days to be generated do
16: for j ← 1 to Q do
17: t ← tjs
18: while t < tje do
19: t ← Next request arrival time defined by a Poisson process with rate
parameter λ (tj )
20: if t < tje then
21: Determine region k ∈ {1, . . . , P} in which the request will occur ac-
cording to p (ai , tj ) values, i ∈ {1, . . . , P}
22: Generate new request i at time t with a uniformly distributed location
in region k. Determine node ñ ∈ Ñ closest to this location in the road
p
network R̃ and set ri = ñ. Furthermore, set request arrival time ria = t
23: end if
24: end while
25: end for
26: end for

i.e., only highways and major roads are included in this illustration. Moreover, in
order to generate all request data sets of S GEN with a degree of dynamism that is
comparable to the one observed in S REAL , only a few requests arrive before the
considered start of the transportation process at 7 am. Specifically, over 90 % of
all requests occur after the start of the execution of the transportation process in
all generated settings. In contrast to S REAL , in request data sets of S GEN with high
levels of structural diversity, request arrival rates increase more rapidly during the
planning horizon. Moreover, the peak of request arrivals is moved to the center of
the considered planning horizon and the number of request arrivals is maintained
8.2 Test Environment, Parameter Values, and Request Data Classes 215

0.01 0.01 0.01 0.16

0.01 0.01 0.01 0.01
0.01 0.01 0.26 0.01
0.01 0.01 0.01 0.01
0.01 0.01 0.10 0.01
0.20 0.01 0.01 0.01 0.01
0.01 0.01 0.01 0.30 0.40
0.01 0.01 0.01 0.01 0.01
0.20 0.30 0.01 0.01 0.01
0.01 0.01 0.20 0.01 0.01
0.01 0.01 0.01 0.01 0.20
0.01 0.10 0.01 0.01 0.06
0.01 0.01 0.01 0.01 0.01
0.01 0.01 0.01 0.01 0.01
0.01 0.07 0.01 0.01 0.30
0.30 0.01 0.01
0.01 0.01 0.01
0.20 0.10 0.30
0.35 0.01 0.01
0.20 0.01 0.01
0.01
0.01
0.01
0.01
0.01

Fig. 8.4 Simplified illustration of the road network with request arrival probabilities p(ai , tj ) for
the structural diversity setting S GEN RD = TD = 1.00

on a high level after the peak. This is performed in order to also generate a high
system utilization after the request arrival peak. Despite these slight modifications,
S GEN possesses characteristics which are similar to S REAL , i.e., a low request arrival
rate at the beginning of the planning horizon and a high increase of request arrival
rates before the peak. Moreover, the subsequent decrease is significantly lower than
its increase before the peak of request arrival rates. The time slices and number of
expected request arrivals are illustrated in Table 8.1. Figure 8.5 visualizes the cor-
responding total request arrival rates of three TD settings of S GEN . Note that since
regional differences of request arrivals are not illustrated in this diagram, the de-
picted total request arrival rates are independent of RD.
According to the described parameters, Fig. 8.6 visualizes three request data sets
of S GEN . As can be seen in this figure, higher levels of structural diversity result in
the desired significant differences of temporal and spatial request arrivals. All 25 re-
quest data sets of S GEN are illustrated in the Appendix in Fig. B.1. Analogous to
S REAL , each of the 25 request data sets of S GEN consists of 30 days. Since request
arrivals in the request data sets of S GEN are generated by time-space Poisson pro-
cesses that are constant from day to day, the proposed forecasting methods are able
to generate stochastic knowledge that can reliably forecast future request arrivals.
Therefore, we conclude that the request data sets of the request data class S GEN have
a low structural distortion.
216 8 Computational Results

Table 8.1 Parameter values for the time slices utilized in generating S GEN
Time slice tj Start time tjs End time tje Request arrival rate λ(tj )

0 06:45 08:00 15
1 08:00 09:00 40
2 09:00 09:30 30
3 09:30 10:30 35
4 10:30 11:00 10

35
TD = 1.00
30
TD = 0.50
25 TD = 0.00
Request arrivals

0
before begin 7:00-7:30 7:30-8:00 8:00-8:30 8:30-9:00 9:00-9:30 9:30-10:00 10:00-10:30 10:30-11:00
Time

Fig. 8.5 Average request arrival distribution of three TD settings of S GEN

Fig. 8.6 Three request data sets of S GEN with different levels of structural diversity

8.2.5 Number of Conducted Computational Experiments

All computational experiments of both request data classes are separately conducted
with both customer inconvenience functions, i.e., linear2X as well as quadratic. As
mentioned in Sect. 8.2.1, each available day in both tested request data sets is con-
ducted using 8, 10, and 12 vehicles in order to analyze the impact of the number
8.2 Test Environment, Parameter Values, and Request Data Classes 217

of available resources on the solution quality. In our experiments, 30 days of each

of the 26 request data sets which consist of one request data set for S REAL and
25 request data sets for S GEN are tested. By evaluating the described dummy cus-
tomer parameter settings as well as the deterministic real-time control approaches,
the total number of experiments that are presented in the following computa-
tional studies amounts to more than 38,000 test instances. Additionally, prelim-
inary computational experiments and their results are presented in the following
sections.

8.2.6 Performance Evaluation Using the Additionally Attainable

Improvement

In order to evaluate the performance of different real-time control approaches,

we utilize a measure which is based on the value of information presented in
Sect. 2.6.6.2. Specifically, instead of evaluating the solution quality of a solution
approach on an instance in comparison to its performance on the corresponding of-
fline instance, we compare the results of two solution approaches A1 and A2 on
the same problem instance denoted by I . In doing so, the additionally attainable
improvement of A2 in comparison to the improvement2 attained by A1 on I can
be calculated. With zA (I ) denoting the objective function value of the solution on
instance I generated by algorithm A, the additionally attainable improvement is
defined as:
zA (I ) − zA2 (I )
AAI(A1 , A2 , I ) = 1
zA1 (I )
As mentioned above, each request data set consists of 30 days which are used in
each computational study in order to compare two real-time control approaches A1
and A2 . Specifically, using a specific request data set combined with a chosen ob-
jective function and a defined fleet size results in 30 problem instances denoted as
the set I˜. In order to calculate the additionally attainable improvement of A2 in
comparison to A1 on the instance set I˜, we define the AAI as follows:

I ∈I˜ zA1 (I )
− I ∈I˜ zA2 (I )
|I˜| |I˜| I ∈I˜ zA1 (I ) − I ∈I˜ zA2 (I )
AAI(A1 , A2 , I˜) = =
I ∈I˜ zA1 (I ) I ∈I˜ zA1 (I )
|I˜|

For example, in order to determine the performance of the pro-active real-time con-
trol approach by comparing its results to the ones obtained by the deterministic

2 We assume that the application of the real-time control approach A1 has already led to an im-
provement in comparison to carrying out the coordination of the transportation process without
using any real-time control approaches.
218 8 Computational Results

real-time control approach, we calculate AAI(A1 , A2 , I˜) with A1 as the determinis-

tic real-time control approach and A2 as the pro-active real-time control approach
and with I˜ as the set of conducted problem instances. Whenever the additional im-
provement of a real-time control approach in comparison to another real-time con-
trol approach is described in the computational studies which are presented in what
follows, the AAI(A1 , A2 , I˜) is utilized. As mentioned in Sect. 5.3.2.4, the primary
objective is to minimize the number of late requests. Due to the prohibitive character
of the lateness penalty costs R pen caused by late requests, the number of late requests
is separately considered in the computational results in order be able to separately
determine the improvements attained in variable customer inconvenience. As shown
in the computational results, in almost all scenarios tested with a fleet size of 10 and
12 vehicles, no late requests occur at all so that the number of late requests is mainly
of interest in scenarios where a fleet size of 8 vehicles is present.

8.3 Performance of the Tabu Search Solution Method

In this section, we present a performance evaluation of the Tabu Search metaheuris-

tic that has been introduced in Chap. 7. Since this solution method is iteratively ap-
plied in the transportation process in order to adapt the current tour plan according
to newly incoming requests, its efficiency is of significant importance. Particularly,
due to the strong time restrictions of the real-time control, attaining near-optimal to
optimal solutions within given limits becomes an essential requirement. Therefore,
in order to assess the applicability of the proposed Tabu Search metaheuristic in the
real-time control approach, its solution quality is evaluated by comparing the ob-
jective function value of attained solutions with the lower bound generated by an
exact procedure which is based on the column generation approach of Krumke et al.
(2002). In order to allow for comparing the solutions generated by this approach
with the ones generated by our approach, we modified this exact approach according
to the objective functions of the considered RDOPG applications and replaced the
Euclidean-based travel time calculations with the road network based calculations of
our proposed real-time control approach. In order to avoid possible distorting effects
caused by the real-time environment, we directly compare the solution quality of in-
dividual static problem instances that occur during the transportation process of two
selected days of the request data class S REAL . Each tested static problem instance is
evaluated using a fleet size of 8 and 10 vehicles. The selected days are different with
respect to the number of dynamically arriving requests. Consequently, two groups
of static problem instances that can be classified into large-sized and moderate-sized
problems arise. The first group comprises 137 static problem instances with up to
41 pending requests for a fleet of 10 vehicles and up to 57 pending requests for 8 ve-
hicles. Additionally, there are 96 static problem instances in the second group that
consist of up to 26 pending requests for a fleet size of 10 vehicles and 34 pending
requests when a fleet of 8 vehicles is used. Note that the use of an increased fleet
size results in accelerated request processing which leads to a smaller number of
8.3 Performance of the Tabu Search Solution Method 219

unserviced pending requests in the 10 vehicle case compared to the 8 vehicle sce-
nario. As mentioned in Sect. 4.6, this column generation approach is well suited
for solving routing problem instances in which an optimal solution consists of tours
of short length. Since this is likely to be the case in the problem instances which
are evaluated, as can be seen from the low request to vehicle ratio described above,
this exact approach is an adequate method for assessing the performance of the pro-
posed Tabu Search metaheuristic. In the conducted experiments, the computational
time available for the Tabu Search procedure is restricted to 10 seconds while a time
limit of one hour is given for computing the lower bound. Clearly, if the exact so-
lution approach is able to compute an optimal solution within the given time limit,
the value of the lower bound is equal to the objective function value of the found
optimal solution.
The evaluation using 10 vehicles indicates that the Tabu Search metaheuristic
attains an almost optimal solution for all static problem instances. Specifically, the
average solution quality of the Tabu Search procedure is only 0.017 % above the
lower bound values which were computed by using the approach of Westphal and
Krumke (2008). By restricting the fleet of vehicles to only 8 vehicles, which results
in more complex static problem instances, this gap increases to 1.26 %. Note that
this increase is solely caused by a few specific static problem instances in which
distorting effects for single requests, whose maximum allowed response time is vi-
olated, occur.
The provided results indicate that the proposed real-time control approach attains
a good solution adaptation performance that is necessary for a real-time application.
Consequently, the application of the implemented Tabu Search metaheuristic for
solving static problem instances allows a sufficiently fast and efficient tour plan
adaptation. Note that, as mentioned in Sect. 2.5.3.2.2, optimally solving individ-
ual static problem instances does not guarantee an optimal result for the overall
problem considered during the planing horizon. Therefore, by integrating stochas-
tic knowledge into the static problem instances, the proposed Tabu Search meta-
heuristic supports the generation of solutions that provide an increased flexibility
for efficiently integrating expected future requests. If high-quality stochastic knowl-
edge is available, optimally solving individual static problem instances during the
transportation process is likely to also result in a high quality solution of the overall
problem considered. Note that since the best solution can be found in 10 seconds in
most of the cases, the utilization of an anticipation horizon with the aforementioned
length of 20 seconds ensures that a time period of 10 seconds is available to the
solution method, even if pre-simulation activities such as road network calculations
consume a time period tacalc of several seconds. Since requests arrive minute-wise
in our request data sets, no drawback in the attained solution quality occurs in the
computational results. However, since the average request response time is between
800 and 1,400 seconds in the evaluated scenarios, an anticipation horizon with a
length of 20 seconds represents a small time period by which newly arriving re-
quests would be additionally buffered if second-wise request arrival times were con-
sidered. Hence, the length of the anticipation horizon is not evaluated any further in
this book. Nevertheless, note that the choice of this parameter generally depends on
220 8 Computational Results

ROLLING20

TIMELIMIT10

ZEROTIME

Initial tour plan time

generation

Arrival of a new request Length of the anticipation horizon

Zero-time integration of new requests Realization of tour plan adaptations

Generation of a Tour plan adaptation

static problem instance utilizing the solution method

Fig. 8.7 Adaptation handling of the evaluated deterministic real-time control approaches

the considered application. For example, in applications with more pending requests
and hence larger static problem instances, more computational time is required in
order to attain solutions comprising a high solution quality. In such cases, a trade-off
which requires an in-depth evaluation of different lengths of the anticipation hori-
zon may arise. Research work which evaluates different settings for determining the
length of the anticipation horizon can be found in, e.g., Ichoua et al. (2000), Bock
(2004), and Bock (2010).

8.4 Performance Evaluation of Deterministic Real-Time Control

Approaches
In what follows, the deterministic real-time control approach presented in Chap. 5
is designated as ROLLING20 since it applies the solution method in a rolling fash-
ion for the complete length of each anticipation horizon. In the evaluation presented
in this section, this approach is tested in direct comparison with two other deter-
ministic real-time control approaches. The first one is called ZEROTIME. It repre-
sents a zero time adaptation that integrates newly arrived requests into the tour plan
by applying a least cost insertion heuristic, i.e., by using steps 1 and 2 of the up-
date concept presented in Sect. 5.2.2. The second approach is designated as TIME-
LIMIT10 and extends ZEROTIME by additionally applying the Tabu Search proce-
a
dure for t2 = 10 seconds whenever new requests have arrived during the previous
anticipation horizon. Therefore, TIMELIMIT10 utilizes the same update handling
as ROLLING20 (while applying the solution method for half of the time) whenever
dynamic events have recently arrived. All three real-time control approaches are il-
lustrated in Fig. 8.7. This preliminary performance evaluation is carried out on the
30 days of S REAL as well as on 30 systematically chosen days from all structural
diversity settings of S GEN which are denoted as S GENall .
The results that are depicted in Tables 8.2 and 8.3 show that both TIME-
LIMIT10 and ROLLING20 significantly outperform the ZEROTIME variant for
8.4 Performance Evaluation of Deterministic Real-Time Control Approaches 221

Table 8.2 Performance of the evaluated deterministic real-time control approaches on S REAL
Vehicles Percentage improvements attained by
TIMELIMIT10 vs. ZEROTIME ROLLING20 vs. TIMELIMIT10

linear2X customer inconvenience

8 32.25 %, pen: 12 vs. 186 0.30 %, pen: 9 vs. 12
10 18.46 %, pen: 0 vs. 8 0.13 %, pen: 0 vs. 0
12 11.18 %, pen: 0 vs. 0 −0.09 %, pen: 0 vs. 0

quadratic customer inconvenience

8 54.02 %, pen: 6 vs. 185 0.37 %, pen: 8 vs. 6
10 33.62 %, pen: 0 vs. 4 −0.50 %, pen: 0 vs. 0
12 22.04 %, pen: 0 vs. 0 −0.39 %, pen: 0 vs. 0

Note: Listed is the average improvement, pen denotes the total number of late requests

Table 8.3 Performance of different deterministic real-time control approaches on S GENall

Vehicles Percentage improvements attained by
TIMELIMIT10 vs. ZEROTIME ROLLING20 vs. TIMELIMIT10

linear2X customer inconvenience

8 32.45 %, pen: 11 vs. 175 0.22 %, pen: 13 vs. 11
10 14.33 %, pen: 0 vs. 9 −0.09 %, pen: 0 vs. 0
12 9.21 %, pen: 0 vs. 0 −0.15 %, pen: 0 vs. 0

quadratic customer inconvenience

8 51.96 %, pen: 14 vs. 158 −0.79 %, pen: 15 vs. 14
10 30.65 %, pen: 0 vs. 5 −0.53 %, pen: 0 vs. 0
12 19.28 %, pen: 0 vs. 1 −0.56 %, pen: 0 vs. 0

Note: Listed is the average improvement, pen denotes the total number of late requests

both tested objective functions. Specifically, most improvements are achieved in the
8 vehicle scenarios where resources are strongly limited. This is due to the fact
that in this reduced fleet size setting, a very efficient usage of the limited resources
which is attained by TIMELIMIT10 and ROLLING20 is of high relevance. This
can also be observed in Figs. 8.8 and 8.9 in which the results on S REAL are pre-
sented in detail. Here it can be clearly observed that in problem instances with more
requests resulting in more complex transportation situations, an efficient usage of
vehicle resources yields high improvements. As mentioned above, these attainable
improvements are highest when the available fleet size is restricted. Therefore, in ac-
cordance with the results achieved in former studies (cf., e.g., Gendreau et al. 1999a
and Bock 2010), it can be concluded that the application of elaborated real-time
control approaches is advantageous if an efficient execution of complex transporta-
222 8 Computational Results

60%

50%
8 vehicles
Improvement

40% 10 vehicles
12 vehicles
30%
Linear (8 vehicles)
20% Linear (10 vehicles)
Linear (12 vehicles)
10%

0%
120 130 140 150 160 170 180 190

Number of requests

Fig. 8.8 Percentage improvements of TIMELIMIT10 vs. ZEROTIME, linear2X customer incon-
venience on S REAL

80%
70%
60% 8 vehicles
Improvement

50% 10 vehicles
40% 12 vehicles
30% Linear (8 vehicles)
20% Linear (10 vehicles)
10% Linear (12 vehicles)

0%
-10%
120 130 140 150 160 170 180 190

Number of requests

Fig. 8.9 Percentage improvements of TIMELIMIT10 vs. ZEROTIME, quadratic customer incon-
venience on S REAL

tion processes is pursued. The corresponding results for S GENall are illustrated in
Figs. 8.10 and 8.11.
Furthermore, by directly comparing the results of the real-time control ap-
proaches TIMELIMIT10 and ROLLING20 (Tables 8.2 and 8.3), it can be observed
that ROLLING20 does not produce any significant improvements in the considered
RDOPG applications. Thus, it can be concluded that ROLLING20 does not use the
increased computational times efficiently. This result can be explained by analyzing
the Tabu Search evaluation in Sect. 8.3 once again. Specifically, the Tabu Search so-
lution approach is utilized in this evaluation in the same way as in TIMELIMIT10.
Since solutions in this evaluation that are very close to an optimal solution were
attained, no further significant improvements can be achieved by ROLLING20 in
this scenario. In what follows, all experiments are carried out using TIMELIMIT10
since this significantly reduces the time required to perform all test instances (see
8.4 Performance Evaluation of Deterministic Real-Time Control Approaches 223

60%

50%
8 vehicles
40% 10 vehicles
Improvement

12 vehicles
30%
Linear (8 vehicles)

20% Linear (10 vehicles)

Linear (12 vehicles)
10%

0%
100 110 120 130 140 150 160

Number of requests

Fig. 8.10 Percentage improvements of TIMELIMIT10 vs. ZEROTIME, linear2X customer incon-
venience on S GENall

80%
70%
60% 8 vehicles
Improvement

50% 10 vehicles
40% 12 vehicles
30% Linear (8 vehicles)
20% Linear (10 vehicles)
10% Linear (12 vehicles)

0%
-10%
100 110 120 130 140 150 160

Number of requests

Fig. 8.11 Percentage improvements of TIMELIMIT10 vs. ZEROTIME, quadratic customer in-
convenience on S GENall

Sect. 8.2.5). Nevertheless, note that other scenarios exist in which ROLLING20 can
achieve further improvements compared to TIMELIMIT10 but these are not con-
sidered in this book. Specifically, if scenarios with a significantly higher number
of vehicles and requests are present, it is unlikely that TIMELIMIT10 will be able
to find optimal solutions within the given time limit. Therefore, in such scenar-
ios, the continuous application of the real-time control approach as proposed in
ROLLING20 is likely to result in further significant improvements. Note that in
the scenarios considered in Sects. 8.8 and 8.9, due to its achieved solution qual-
ity TIMELIMIT10 represents a challenging reference setting for attaining further
improvements by additionally utilizing stochastic knowledge in the pro-active real-
time control approach.
224 8 Computational Results

Table 8.4 Performance of TIMELIMIT10 vs. TIMELIMIT10NODIV on S REAL

Vehicles Objective function value Travel time Travel distance #Reassignments

linear2X customer inconvenience

8 17.18 %, pen: 12 vs. 23 6.36 % 8.89 % 45.87 (5.73)
10 17.56 %, pen: 0 vs. 0 7.68 % 9.87 % 54.07 (5.41)
12 14.24 %, pen: 0 vs. 0 8.79 % 11.16 % 46.93 (3.91)

quadratic customer inconvenience

8 20.67 %, pen: 6 vs. 20 5.07 % 7.23 % 35.07 (4.38)
10 24.60 %, pen: 0 vs. 0 7.27 % 9.67 % 41.77 (4.18)
12 18.81 %, pen: 0 vs. 0 7.18 % 8.54 % 39.87 (3.32)

Note: Listed is the average improvement of the objective function value, total number of late re-
quests (pen), average improvement of travel time and travel distance, and the average number of
first request reassignments per test instance and per vehicle in brackets observed in TIMELIMIT10

8.5 Performance Evaluation of Allowing Vehicle En-Route

Diversion

In this section we evaluate the impact of allowing vehicle en-route diversion on

the solution quality in TIMELIMIT10. This is evaluated by disallowing all vehicle
en-route diversion activities resulting in a variant of TIMELIMIT10 designated as
TIMELIMIT10NODIV. The evaluation of both variants of the deterministic real-
time control approach is carried out on all 30 days of S REAL .
The results that are depicted in Table 8.4 show that the application of vehicle
en-route diversion significantly improves the solution quality in the tested scenarios.
Specifically, depending on the number of available vehicles and the customer incon-
venience function, improvements between 14.24 % and 24.60 % can be achieved
while the highest improvements can be attained if a quadratic customer inconve-
nience function is used. Furthermore, in scenarios with a strongly limited fleet size
of 8 vehicles, allowing vehicle en-route diversion leads to a significantly lower num-
ber of late requests that are serviced after the maximum allowed response time de-
fined by R mrt . These results are in line with the ones obtained by Ichoua et al. (2000)
and Richter (2005). Moreover, although the performed travel time as well as travel
distance are not considered in the RDOPG objective function, both of these values
are also reduced when vehicle en-route diversion is allowed. In Figs. 8.12, 8.13, and
8.14, the attained request response times for all three tested fleet sizes are illustrated.
In these figures it can be observed that by allowing vehicle en-route diversion ac-
tivities, many requests can be serviced earlier in comparison to the setting in which
vehicle en-route diversion is forbidden.
Following these results, we conclude that it is advisable to allow the use of the
vehicle en-route diversion activities in the considered RDOPG applications. How-
ever, allowing vehicle en-route diversion requires more responsiveness from drivers
during travel activities since they have to be able to quickly react to changes in
8.5 Performance Evaluation of Allowing Vehicle En-Route Diversion 225

linear2X, TIMELIMIT10NODIV
linear2X, TIMELIMIT10
serviced requests

quadratic, TIMELIMIT10NODIV
quadratic, TIMELIMIT10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.12 Impacts of allowing vehicle en-route diversion on request response times using S REAL
and 8 vehicles grouped by intervals of 50 seconds

linear2X, TIMELIMIT10NODIV
linear2X, TIMELIMIT10
serviced requests

quadratic, TIMELIMIT10NODIV
quadratic, TIMELIMIT10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.13 Impacts of allowing vehicle en-route diversion on request response times using S REAL
and 10 vehicles grouped by intervals of 50 seconds

linear2X, TIMELIMIT10NODIV
linear2X, TIMELIMIT10
serviced requests

quadratic, TIMELIMIT10NODIV
quadratic, TIMELIMIT10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.14 Impacts of allowing vehicle en-route diversion on request response times using S REAL
and 12 vehicles grouped by intervals of 50 seconds

their tour that affect the next assigned request. The last column of Table 8.4 shows
the number of first request reassignments, i.e., vehicle en-route diversion activities,
that have occurred on average per test instance in TIMELIMIT10. Clearly, these
values are 0 for TIMELIMIT10NODIV. Note that a first request reassignment is
recorded only when the first request of a vehicle is changed during a travel activity.
Specifically, a change of the first request is not accounted for if a vehicle is cur-
rently servicing a request or waiting idly for new requests. Moreover, note that if
the first request of a vehicle is reassigned to another vehicle as first request, this
226 8 Computational Results

is considered as two reassignment activities since the first request of both vehicles
is changed. According to the values depicted in Table 8.4 it can be concluded that
vehicle en-route diversion activities occur quite frequently in these computational
experiments. When considering the drivers, this large number of en-route diversion
activities may lead to a high driver inconvenience which is a rarely addressed issue
in dynamic routing applications. Hence, in Sect. 8.10, we continue to discuss this
issue and introduce a method for reducing the number of reassignment activities and
thus accounting for driver inconvenience aspects in the proposed real-time control
approach.
Moreover, note that a higher number of available vehicles leads to a slight in-
crease of the performed vehicle en-route diversion activities. This can be explained
by considering the fact that when more vehicles are available, they are likely to
be closely located to each other. When new requests arrive, reassigning requests to
those vehicles that are nearby becomes more attractive with regard to the overall
solution quality.

8.6 Measuring Structural Distortion Using the Goodness-of-Fit

Test
As mentioned in Chap. 6, our proposed forecasting methods as well as the pro-active
real-time control approach assume that request arrivals can be modeled by a set of
unknown time-space Poisson processes whose rate parameters are also unknown but
remain constant from day to day. Since the expected number of request arrivals per
day of generated clusters is determined by calculating the average request arrival
rate per cluster on the past nf working days (see Sect. 6.2), the application of the
presented pro-active real-time control approach on request data sets which do not
fulfill this assumption is likely to result in a poor solution quality. For example, con-
sider a request data set in which requests arrive according to two different patterns
which change every day. In this case, forecasting expected future requests by calcu-
lating the average of all request arrivals observed on the past nf working days may
result in the generation of clusters which predict requests according to a request ar-
rival pattern that is a mixture of both request arrival patterns. However, since this
generated average day does not exist, it is likely that the pro-active real-time con-
trol approach produces a poor solution quality on both types of days. Hence, it is
crucial that the request data fulfills the mentioned assumption to a certain extent.
For this reason, our proposed forecasting approach examines the existing request
data set. Specifically, the Chi-Square Goodness-of-Fit test described in Sect. 6.2.1
checks the likelihood that request arrivals observed over the past nf working days
in potential clusters are generated by Poisson processes and discards them if this is
too unlikely. We denote this likelihood as the Poisson quality of a cluster. According
to our definition, a high Poisson quality defines a low structural distortion and vice
versa.
In order to assess the effectiveness of our approach in identifying structural dis-
tortion in request data sets, we generate different request data sets that, according
8.6 Measuring Structural Distortion Using the Goodness-of-Fit Test 227

y y y

e e e
m m m
Ti Ti Ti
x x x
SMIXa SMIXb SMIXc

First day request arrival pattern

Second day request arrival pattern

Fig. 8.15 Three request data sets with a high structural distortion

to the assumptions of the proposed forecasting approach, have a high structural dis-
tortion. Specifically, we generate four request data sets each having two individual
request arrival patterns that alternate each working day. The generated request data
sets are:
1. Request data set S MIXa : The first pattern is defined as S GEN RD = TD = 1.00
whereas no request arrivals are defined in the second pattern. Hence, on each
second working day, no requests arrive.
2. Request data set S MIXa2x : The first pattern is defined as S GEN RD = TD =
1.00 with doubled request arrival rates, i.e., doubled λ(tj ) values from Table 8.1
whereas no request arrivals are defined in the second pattern. These doubled
request arrival rates are used in order to generate average request arrival rates
which are similar to S GEN RD = TD = 1.00.
3. Request data set S MIXb : In this request data set, the first pattern is defined as a
variation of S GEN RD = TD = 1.00 in which all requests only arrive in subareas
located on the left side of the considered service region. The second pattern is
a variation of S GEN RD = TD = 1.00 in which all requests arrive in subareas
located on the right side of the considered service region.
4. Request data set S MIXc : The first pattern is defined as S GEN RD = TD = 1.00
whereas the second pattern is defined as S GEN RD = TD = 0.00. This request
data set is used to assess the performance of the used Chi-Square Goodness-of-
Fit test on a request data set in which requests arrive every day at all locations
but according to two different request arrival patterns.
The request data sets S MIXa , S MIXb , and S MIXc are illustrated in Fig. 8.15. In ad-
dition to the four request data sets above, we evaluate the structural distortion of
S GEN RD = TD = 1.00, RD = TD = 0.75, and of S REAL . For the minimum de-
manded Poisson quality of a valid cluster denoted by DCminPQ , we evaluate values
of DCminPQ ∈ {0, 0.05, 0.10, 0.20, 0.40}. For this purpose, in the applied Chi-Square
Goodness-of-Fit test the type I error parameter is set to α = DCminPQ . Moreover, in
order to generate clusters which, according to the assumptions of our forecasting ap-
228 8 Computational Results

proach, have a high probability that real customers will arrive, we apply a minimum
demanded dummy customer quality of DCminλ = 1.8.
The results of the structural distortion evaluation are illustrated in Table 8.5. As
can be seen in this table, a significant number of dummy customers can be gen-
erated in all request data sets in the setting DCminPQ = 0 in which no minimum
Poisson quality is demanded for the generated dummy customers. Therefore, one
could falsely presume a high suitability of the four request data sets (S MIXa to
S MIXc ) to generate well-usable stochastic knowledge using the proposed forecast-
ing approach. For example, by comparing S MIXa2x and S GEN RD = TD = 1.00,
the number of generated dummy customers is similar, i.e., 35 to 37. However, as
the applied Chi-Square Goodness-of-Fit test reveals, the generated average Poisson
quality of these four request data sets is significantly lower than that of the request
data sets S GEN and S REAL . Specifically, generated dummy customers of the three
request data sets S MIXa , S MIXa2x , and S MIXb have a significantly low Poisson qual-
ity of less than 0.01 % which indicates a high structural distortion for these request
data sets. Moreover, each of these three request data sets have a request realiza-
tion rate of at most 50 %. Specifically, in each of the corresponding clusters of the
generated dummy customers, request arrivals are observed on, at most, 50 % of the
nf working days. This can be explained by their request arrival patterns mentioned
above. These observed request realization rates which are significantly smaller than
the expected request realization rate defined by PS1.8 (X ≥ 1) = 83.5 % also indicate
that these three request data sets have a high structural distortion. Consequently, if
a minimum Poisson quality of DCminPQ = 0.05 is demanded, no dummy customers
can be generated at all in these three request data sets. In contrast to the three previ-
ously evaluated request data sets, the average Poisson quality of generated dummy
customers in S MIXc using DCminPQ = 0 is higher, i.e., about 5 % compared to an
average Poisson quality of less than 0.01 %. Nevertheless, this Poisson quality is
still about an order of magnitude smaller than in S GEN and S REAL . Moreover, an-
other indicator that S MIXc has an inferior Poisson quality is that the request real-
ization rate of corresponding clusters is on average 71 %, which is significantly
smaller than the expected request realization rate of 83.5 %. Hence, if a minimum
Poisson quality of DCminPQ = 0.05 is demanded, almost half of the 13 clusters are
discarded. Note that this is a significant difference to S GEN and S REAL in which
almost the same number of dummy customers are generated in DCminPQ = 0.00
and DCminPQ = 0.05. By further increasing the minimum demanded Poisson qual-
ity to DCminPQ = 0.40, only one dummy customer is generated in S MIXc while in the
evaluated request data sets of S GEN and S REAL only a slight decrease of the num-
ber of generated dummy customers occurs. Preliminary computational experiments
on S GEN and S REAL evaluating the previously mentioned values of DCminPQ showed
that the setting DCminPQ = 0.40 yields the best solution quality among the tested
values. Consequently, DCminPQ = 0.40 guarantees a low number of erroneously
erased clusters while ensuring that inappropriate clusters with a significantly inferior
Goodness-of-Fit quality are removed. Note that we refrained from evaluating values
for DCminPQ higher than 0.40 for two reasons. First, as mentioned above, this value
efficiently removes almost all dummy customers with a low Poisson quality in the
Table 8.5 Evaluating the structural distortion on different generated request data sets using different DCminPQ settings and DCminλ = 1.8
Request data set Minimum Poisson quality DCminPQ
0 0.05 0.10 0.20 0.40

S MIXa 8 0 0 0 0
<0.0001 – – – –
47 % 49 % 50 % – – – –
S MIXa2x 35 0 0 0 0
<0.0001 – – – –
47 % 49 % 50 % – – – –
S MIXb 19 0 0 0 0
<0.0001 – – – –
47 % 49 % 50 % – – – –
S MIXc 13 7 5 2 1
0.0495 0.1295 0.167 0.3162 0.5206
62 % 71 % 78 % 73 % 76 % 78 % 75 % 77 % 78 % 78 % 78 % 78 % 80 % 80 % 80 %
S GEN 15 15 15 15 15
RD = TD = 0.75 0.426 0.4314 0.5243 0.5747 0.6386
77 % 84 % 93 % 77 % 84 % 93 % 77 % 85 % 93 % 80 % 86 % 95 % 80 % 86 % 95 %
S GEN 37 37 37 37 35
8.6 Measuring Structural Distortion Using the Goodness-of-Fit Test

RD = TD = 1.00 0.5124 0.5139 0.4925 0.581 0.7048

80 % 86 % 93 % 80 % 86 % 93 % 78 % 86 % 93 % 78 % 86 % 93 % 82 % 86 % 92 %
S REAL 15.3 15 14.9 14.1 13.1
0.4047 0.4238 0.4537 0.5148 0.6261
68 % 81 % 92 % 73 % 81 % 92 % 73 % 81 % 93 % 75 % 82 % 93 % 77 % 82 % 92 %

Note: Listed is the average number of generated dummy customers, their average Poisson quality, and the minimum, average, and maximum request realization
rate, i.e., the percentage of the nf working days on which at least one request arrives in one corresponding cluster of the generated dummy customers
229
230 8 Computational Results

Average number of dummy customers 40

35
30

25
S GEN 1.00
20 S GEN 0.75
15 S REAL

5
0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Minimum demanded Poisson quality

Fig. 8.16 Average number of generated dummy customers on S REAL , S GEN RD = TD = 0.75,
and S GEN RD = TD = 1.00 using DCminλ = 1.8

1.0
Average Poisson quality

0.8

0.6 S GEN 1.00

S GEN 0.75
0.4 S REAL

0.2

0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Minimum demanded Poisson quality

Fig. 8.17 Average Poisson quality on S REAL , S GEN RD = TD = 0.75, and S GEN RD = TD = 1.00
using DCminλ = 1.8

tested request data sets that have a high structural distortion. Second, as illustrated
in Fig. 8.16, choosing higher DCminPQ values increasingly reduces the number of
dummy customers that can be generated, especially in S REAL so that a trade-off
arises between the resulting Poisson quality and the number of generated dummy
customers. Since setting DCminPQ to 0.40 results in an average Poisson quality of
more than 0.6 (see Fig. 8.17), we conclude that this setting allows for the generation
of clusters which sufficiently fulfill the demanded assumptions. Note that we deviate
from setting type I error to α = 0.05 on purpose (cf. Israel 2008, p. xxii, Sapsford
and Jupp 2006, p. 217) in order to only generate stochastic knowledge which has a
sufficient Poisson quality.
From the results presented above we conclude that the proposed approach using
the Chi-Square Goodness-of-Fit test to determine the Poisson quality of request data
sets is an efficient method for determining the structural distortion of an existing
request data set. Specifically, it allows for identifying request data sets with a high
8.7 Results of Generating Stochastic Knowledge for S REAL and S GEN 231

structural distortion in which a reasonable number of dummy customers can be gen-

erated without applying this test. Note that the structural distortion of a request data
set depends on the applied forecasting approach. Specifically, by extending the pro-
posed forecasting approach, it becomes possible to efficiently generate and integrate
stochastic knowledge out of specific request data sets which have a high structural
distortion in our forecasting approach. Nevertheless, for each forecasting approach,
there are specific request data sets which have a high structural distortion with re-
gard to the respective forecasting approach. Hence, we recommend determining the
structural distortion of given request data sets for all forecasting approaches.

Structural Distortion of S REAL The attained results also provide information

on the structural distortion of the real-world request data set S REAL . Specifically,
since S REAL possesses a Poisson quality which is similar to both evaluated S GEN
request data sets that possess a low structural distortion with certainty, we conclude
that S REAL has also a low structural distortion. Note that this makes it reasonable to
evaluate S REAL in our main computational experiments in Sect. 8.8.

8.7 Results of Generating Stochastic Knowledge for S REAL

and S GEN

In this section, results of generating the dummy customers for the request data
classes S REAL and S GEN using the exact clustering approach proposed in Chap. 6
are presented. For generating this stochastic knowledge the previously derived min-
imum Poisson quality DCminPQ = 0.40 is used.
After all valid clusters are generated, a Mixed Integer Programming (MIP) prob-
lem is solved in order to select the largest number of valid clusters which start ear-
liest so that no segments overlap as described in Sect. 6.2.2. All of these formulated
MIP problems are solved using ILOG CPLEX 11.0. With regard to the required
computational time, MIP problems with higher DCminλ values can be solved within
seconds while lower DCminλ values result in significantly longer computational
times due to the high number of valid clusters. Specifically, for the lowest evaluated
minimum demanded dummy customer quality DCminλ = 1.0, up to 2,362 valid clus-
ters were generated for S REAL and 3,272 valid clusters for S GEN RD = TD = 1.00.
Note that this cluster selection process can often be significantly accelerated. Specif-
ically, by using characteristics of the problem structure (see objective function de-
fined in Eq. (6.1) in Sect. 6.2.2), for the objective function value z of the opti-
mal MIP solution
with n dummy customers it holds that n · M − U ≤ z ≤ n · M
with U = i∈Cnlatest cistartTL where Cnlatest ⊆ C represents the set of the n clusters
of C which start latest. By using this property and consulting the currently known
upper bound during the solution process of an MIP problem, a new value of the
maximum number of selectable clusters defined by C max can be dynamically de-
termined. For example, assume M = 100,000, U = 18,000, and a currently upper
bound of 8,532,102, e.g., the objective function value of the currently best known
232 8 Computational Results

Table 8.6 Number of generated dummy customers on S REAL depending on the minimum required
dummy customer quality DCminλ using DCminPQ = 0.40
Quality of stochastic knowledge (DCminλ )
1.0 1.2 1.5 1.8 2.0

67.7 (13.4 %) 44.6 (12 %) 24.4 (10.4 %) 13.1 (10.1 %) 9.4 (7.7 %)

Note: Listed is the number of dummy customers generated by the proposed exact clustering
method. The values in brackets denote the additional percentage of dummy customers that were
generated compared to a heuristic greedy variant of the exact clustering approach

LP relaxation. Assuming further that the currently best known MIP solution has an
objective function value of 8,489,125, the current gap is 8,532,102−8,489,125
8,532,102 ≈ 5 %.
Using the described technique, it can be derived that the number of clusters that can
be selected is at most 85 since 8,532,102 < 86 · 100,000 − 18,000 = 8,582,000.
Hence, the objective function value of the known upper bound is at least reduced
to 8,500,000 so that the gap is reduced to at most 8,500,000−8,489,125
8,532,102 ≈ 0.12 %. We
utilize this technique in the conducted MIP problems by examining the currently
best known upper bound during the solution process and adjusting the value of pa-
rameter C max accordingly. In doing so, all conducted MIP problems can be solved to
optimality within two hours of computational time while most of the MIP problems
are solved within a few minutes.
In Tables 8.6 and 8.7, the number of dummy customers which were generated us-
ing our proposed exact cluster selection approach are presented. Note that in S REAL ,
for each of the 30 conducted days, the last corresponding nf = 60 working days
out of the available real-word request information were used. Hence, for each of
the 30 days of S REAL , five MIP problems were generated for the different DCminλ
values so that in total 150 MIP problems were solved for S REAL . Consequently,
Table 8.6 illustrates the average number of dummy customers generated per sce-
nario and DCminλ value. Since all days of S GEN are generated according to the
procedure described in Sect. 8.2.4.2, dummy customers are generated once for all
25 structural diversity settings and DCminλ value. These dummy customers are then
used in each of the 30 days. As can be seen in Table 8.7, specific DCminλ values
require a minimum level of structural diversity so that in total 87 MIP problems
were solved for S GEN . Moreover, the number of generated dummy customers is
compared to a heuristic clustering approach which generates clusters in the same
manner as the exact approach but greedily selects dummy customers whenever a
valid cluster has been found. Although the heuristic clustering approach only re-
quires about one minute for solving each dummy customer selection problem, the
results in Tables 8.6 and 8.7 show that it is reasonable to utilize the exact cluster-
ing approach since it yields on average 10.7 % more generated dummy customers
on S REAL and 6.4 % more on S GEN .
With regard to the base sizes which are used in the selected clusters, in S GEN it
can be observed that in request data sets with higher levels of structural diversity
the average number of base segments in a selected cluster becomes lower. Specif-
8.7 Results of Generating Stochastic Knowledge for S REAL and S GEN 233

Table 8.7 Number of generated dummy customers on S GEN depending on the level of structural
diversity and the minimum required dummy customer quality DCminλ using DCminPQ = 0.40
Structural Quality of stochastic knowledge (DCminλ )
diversity
RD TD 1.0 1.2 1.5 1.8 2.0

0.00 0.00 2 (0 %)
0.00 0.25 11 (10 %)
0.00 0.50 21 (23.5 %) 5 (0 %)
0.00 0.75 36 (9.1 %) 12 (9.1 %)
0.00 1.00 46 (15 %) 19 (11.8 %) 7 (40 %) 2 (0 %)
0.25 0.00 5 (0 %)
0.25 0.25 17 (6.3 %) 3 (0 %)
0.25 0.50 26 (8.3 %) 8 (14.3 %)
0.25 0.75 36 (9.1 %) 21 (5 %) 8 (0 %)
0.25 1.00 45 (9.8 %) 26 (13 %) 10 (11.1 %) 6 (0 %) 3 (0 %)
0.50 0.00 20 (0 %) 5 (0 %)
0.50 0.25 31 (3.3 %) 10 (0 %) 2 (0 %)
0.50 0.50 36 (9.1 %) 20 (0 %) 7 (0 %)
0.50 0.75 44 (7.3 %) 29 (7.4 %) 16 (6.7 %) 8 (0 %) 4 (0 %)
0.50 1.00 54 (17.4 %) 37 (23.3 %) 24 (20 %) 14 (0 %) 11 (0 %)
0.75 0.00 32 (0 %) 15 (0 %) 2 (0 %)
0.75 0.25 38 (8.6 %) 18 (5.9 %) 6 (0 %)
0.75 0.50 40 (11.1 %) 28 (7.7 %) 16 (0 %) 6 (0 %) 2 (0 %)
0.75 0.75 54 (10.2 %) 38 (11.8 %) 24 (4.3 %) 15 (0 %) 11 (0 %)
0.75 1.00 70 (9.4 %) 50 (16.3 %) 33 (3.1 %) 27 (8 %) 20 (5.3 %)
1.00 0.00 37 (2.8 %) 22 (4.8 %) 14 (0 %) 1 (0 %) 1 (0 %)
1.00 0.25 45 (4.7 %) 28 (0 %) 12 (0 %) 5 (0 %) 2 (0 %)
1.00 0.50 61 (15.1 %) 42 (10.5 %) 25 (8.7 %) 16 (14.3 %) 9 (0 %)
1.00 0.75 74 (12.1 %) 54 (12.5 %) 34 (3 %) 26 (13 %) 20 (11.1 %)
1.00 1.00 80 (14.3 %) 65 (14 %) 46 (9.5 %) 35 (12.9 %) 29 (3.6 %)

ically, the maximum allowed spatial extension of a cluster of DCmse = (2 × 2) =

4 segments (see Sect. 8.2.3) is always used when scenarios with low levels of struc-
tural diversity are present. In contrast, the average number of base segments declines
in RD = TD = 1.00, depending on the minimum quality of stochastic knowledge,
to values from 3.153 segments (DCminλ = 1.0) to 3.438 segments (DCminλ = 2.0).
Hence, higher levels of structural diversity in S GEN do not only allow for generating
234 8 Computational Results

more dummy clusters but also those which have a higher precision. This increases
the value of the generated stochastic knowledge. On S REAL , the average number of
base segments in a selected cluster ranges from 3.729 segments (DCminλ = 1.0) to
4 segments (DCminλ = 2.0).
In Table 8.8, the request realization rates of the generated stochastic knowledge
in all 26 request data sets are illustrated. Moreover, for each DCminλ setting, we
simulated 100 clusters. For each of these clusters, we generated 60 replications, i.e.,
60 days, of request arrivals using a Poisson process with λ = DCminλ and measured
the request realization rates. The obtained values as well as the arrival probability
of at least one request according to the Poisson distribution are also illustrated. The
results show that the request realization rates mainly depend on DCminλ but not
on the level of structural diversity. The similarity of the results of the 26 request
data sets to the simulated as well as estimated values underscores the fact that all
request data sets have a low structural distortion. Hence, we conclude that in both
evaluated request data classes request arrivals occur according to Poisson processes
to an extent which is sufficient to conduct the main computational experiments on
these request data classes in what follows.

8.8 Performance Evaluation of the Pro-Active Real-Time

Control Approach

In this section, we present the main results of our computational experiments. In

these experiments, the performance of the pro-active real-time control approach that
utilizes the generated stochastic knowledge is evaluated. As mentioned above, since
no benchmark instances for the considered RDOPG applications are available in the
literature, this evaluation is carried out by comparing the results obtained by the pro-
active real-time control approach with the ones of the deterministic real-time control
approach TIMELIMIT10. In order to select an appropriate setting of the pro-active
real-time control approach, a preliminary study comparable to the one performed
for the deterministic real-time control approach in Sect. 8.4 has been conducted.
It reveals that, similar to the results obtained for the deterministic real-time con-
trol approach, the continuous application and the time-limit variant of the pro-active
real-time control approach produce a comparable solution quality. Hence, in what
follows, the computational experiments are carried out by comparing the pro-active
real-time control approach applied for 10 seconds when dynamic events have oc-
curred in the previous anticipation horizon (PROACTIVE10) with its deterministic
variant TIMELIMIT10. Consequently, both approaches only differ from each other
in the stochastic knowledge additionally utilized by PROACTIVE10.

Computational Results on Dummy Customer Removal Decisions Regarding

the two tested values of the parameter DCλrem , which defines the time point at
which each dummy customer will be individually removed from the tour plan (see
Sect. 8.2.3), preliminary tests indicate that an earlier removal of dummy customers
8.8 Performance Evaluation of the Pro-Active Real-Time Control Approach 235

Table 8.8 Request realization rates (in percents) in S GEN and S REAL depending on the parame-
ter DCminλ using DCminPQ = 0.40
Structural Quality of stochastic knowledge (DCminλ ) in percents
diversity
RD TD 1.0 1.2 1.5 1.8 2.0

0.00 0.00 60 61 62
0.00 0.25 58 62 67
0.00 0.50 60 64 72 65 70 73
0.00 0.75 58 65 77 67 71 77
0.00 1.00 60 65 75 67 72 77 72 77 83 83 86 88
0.25 0.00 60 63 67
0.25 0.25 60 64 68 68 69 70
0.25 0.50 58 64 72 68 72 75
0.25 0.75 58 64 77 67 71 78 75 79 83
0.25 1.00 60 65 78 65 71 77 75 79 83 82 85 88 87 88 88
0.50 0.00 60 64 70 67 68 72
0.50 0.25 58 64 72 68 72 78 77 78 80
0.50 0.50 60 65 75 67 72 78 73 78 83
0.50 0.75 60 65 78 65 72 78 75 80 85 82 87 92 82 87 92
0.50 1.00 60 66 77 67 73 78 73 80 85 80 85 92 85 89 93
0.75 0.00 58 65 75 67 71 75 78 79 80
0.75 0.25 60 65 70 65 72 78 72 78 85
0.75 0.50 60 66 77 63 71 78 75 79 87 78 83 87 85 89 93
0.75 0.75 57 65 73 63 72 78 75 80 85 80 86 95 85 89 92
0.75 1.00 58 65 77 65 71 83 73 80 87 77 85 92 83 89 93
1.00 0.00 57 65 70 63 72 78 72 78 85 87 87 87 87 87 87
1.00 0.25 58 65 70 65 71 80 73 79 83 83 86 88 90 90 90
1.00 0.50 60 66 77 65 72 87 75 80 87 80 85 90 85 89 93
1.00 0.75 58 65 80 65 72 82 72 79 88 80 85 90 82 89 95
1.00 1.00 58 66 80 63 72 80 73 80 90 82 86 92 83 88 95
S REAL 55 63 78 62 70 83 70 77 87 77 82 92 78 85 93

Simulated 48 64 80 58 71 87 62 77 88 72 83 93 75 86 97
PSDCminλ (X ≥ 1) 63.2 69.9 77.7 83.5 86.5

Note: Listed is the minimum, average, and maximum request realization rate, i.e., the percentage
of the nf working days on which at least one request arrives in one corresponding cluster of the
generated dummy customers

is more promising in the evaluated RDOPG application. Specifically, setting the pa-
rameter DCλrem to 0.50 instead of 0.25 allows a more flexible tour plan adaptation.
This results from the fact that when a smaller value for DCλrem is used, the probabil-
236 8 Computational Results

Table 8.9 Percentage improvements achieved by PROACTIVE10 vs. TIMELIMIT10 (i.e., im-
provements attained by utilizing stochastic knowledge ) on S REAL
Vehicles Quality of stochastic knowledge (DCminλ )
1.0 1.2 1.5 1.8 2.0

linear2X customer inconvenience

8 ↑ −11.50 % (28) ↑ −7.66 % (21) ↑ −3.12 % (18) ↑ −0.27 % (18) ↑ 1.60 % (10)
10 −4.86 % (22) −1.42 % (20) −0.68 % (17) 1.00 % (16) 0.01 % (16)
12 −0.67 % (15) −0.14 % (16) 1.70 % (13) 0.64 % (11) 0.88 % (11)

quadratic customer inconvenience

8 ↑ −19.91 % (27) ↑ −7.52 % (21) ↑ 0.21 % (14) ↑ 1.17 % (15) ↑ 1.13 % (13)
10 −3.99 % (20) 0.70 % (12) 2.76 % (9) 2.79 % (10) 0.71 % (14)
12 6.59 % (9) 7.93 % (5) 6.78 % (9) 7.31 % (8) 6.08 % (8)

Note: Illustrated is the average improvement. The values in brackets denote the number of worse
test instances. Bold: best DCminλ . ↑: more late requests than in TIMELIMIT10

ity of a future request arrival in the corresponding cluster becomes too low to justify
keeping a vehicle there any longer. Hence, by using higher values for DCλrem , ve-
hicles are available at an earlier time point so that pending requests can be fulfilled
faster. In what follows, DCλrem is set to 0.50 for all experiments carried out using
PROACTIVE10.

8.8.1 Results on S REAL

The results that are attained by evaluating the real-world request data class are il-
lustrated in Table 8.9. Specifically, the table shows the additional improvements
attained by PROACTIVE10 compared to the results obtained by TIMELIMIT10.
Moreover, the number of worse test instances out of the 30 evaluated ones is illus-
trated. Specifically, according to the objective function described in Sect. 5.3.2.4,
solutions generated by PROACTIVE10 and TIMELIMIT10 are compared accord-
ing to the number of late requests (primary objective) and the total customer incon-
venience (secondary objective). Therefore, a solution generated by PROACTIVE10
is denoted as worse if either (a) it comprises more late requests or (b) it comprises
an equal number of late requests but a larger total variable customer inconvenience
than the one generated by TIMELIMIT10.
As indicated by the boldly printed results, under specific conditions, significant
improvements can be achieved by the additional use of the stochastic knowledge
although, as mentioned above, the underlying request distributions in S REAL are
completely unknown. Particularly, in scenarios with a larger fleet size (i.e., 10 and
12 vehicles), the application of the quadratic customer inconvenience function en-
ables a successful integration of the generated stochastic knowledge. This is due to
the fact that the over-linear increase of the quadratic inconvenience function leads
8.8 Performance Evaluation of the Pro-Active Real-Time Control Approach 237

to a lower number of requests that are significantly postponed in favor of other re-
quests as can be observed in the linear2X inconvenience function. Consequently,
generated tour plans are more balanced according to the resulting request response
times so that violations of the maximum allowed request response time can be less
observed in case the quadratic customer inconvenience is used. In contrast to this,
the application of the linear2X inconvenience function more frequently leads to dis-
torting effects, where some requests suddenly become extremely urgent due to an
imminent violation of the maximum allowed response time. Since in these scenar-
ios additional delays by further arriving requests may cause considerable deteri-
orations, tour plans generated if the quadratic inconvenience function is used are
more robust when delays caused by incorrectly forecasted requests or unexpected
additional requests occur. Moreover, reductions of resulting response times due to
correctly forecasted requests have more significant effects in the quadratic customer
inconvenience case.
In situations where more vehicles are available, the negative effect of incorrectly
forecasted request arrivals that becomes more likely with lower DCminλ values has
less impact on both tested inconvenience functions. Hence, also more uncertain in-
formation can be gainfully utilized so that the best results are achieved by using
smaller values of parameter DCminλ . In contrast, in scenarios where resource bot-
tlenecks can be observed (8 vehicles), the mentioned robustness of generated tour
plans when using a quadratic inconvenience function diminishes. Specifically, since
the limited resources are very busy with servicing the pending requests, the resulting
request response times increase towards the maximum allowed response time for a
large number of requests. Therefore, as depicted above for the linear2X customer
inconvenience case, an additional delay caused by incorrectly forecasted requests
often results in a substantial increase of customer inconvenience caused by the high
penalty costs (see Sect. 5.3.2.4). Consequently, no improvement can be attained for
all tested settings with the smallest fleet size. Hence, in case of a small fleet size,
the integration of stochastic knowledge is reasonable only if the stochastic knowl-
edge about future requests possesses a certain value which we consider later in this
section. Otherwise, the attempt to integrate stochastic knowledge often causes dete-
riorating effects which lead to a significant postponement of pending requests and
thus to a higher number of late requests and a reduced solution quality.
The described circumstances can also be observed by analyzing the distribution
of the attained request response times in the experiments. Specifically, when using
a small fleet size, Fig. 8.18 illustrates that the utilization of uncertain stochastic
knowledge (i.e., setting DCminλ = 1.0) further increases the workload of the already
very busy vehicles. Since the uncertainty of this knowledge does not allow for a
faster service of many requests, a high number of late serviced customers occurs
which can be observed by the high peak that occurs shortly before the maximum
allowed request response time of R mrt = 3,600 seconds. This is especially true when
the linear2X customer inconvenience function is used which, as mentioned above,
tends to postpone some requests in order to achieve a better average request response
time which becomes problematic when future requests unexpectedly arrive. As can
be seen in Fig. 8.19, the use of more certain stochastic knowledge generated by
238 8 Computational Results

linear2X, TIMELIMIT10
linear2X, PROACTIVE10
serviced requests

quadratic, TIMELIMIT10
quadratic, PROACTIVE10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.18 Request response times on S REAL using 8 vehicles and DCminλ = 1.0 grouped by inter-
vals of 50 seconds

linear2X, TIMELIMIT10
linear2X, PROACTIVE10
serviced requests

quadratic, TIMELIMIT10
quadratic, PROACTIVE10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.19 Request response times on S REAL using 8 vehicles and DCminλ = 2.0 grouped by inter-
vals of 50 seconds

higher DCminλ values may reduce these negative effects. Nevertheless, due to the
restricted fleet size, the busy vehicle situation remains so that no improvements can
be attained in this situation either.
In contrast, in scenarios where 10 vehicles are available, a high number of re-
quests that are almost late can still be observed when uncertain stochastic knowledge
is utilized, but to a significantly lower extent compared to the 8 vehicle case (see
Figs. 8.20 and 8.21). Hence, if the linear inconvenience function is used, improve-
ments of up to 1.0 % can be achieved whereas improvements of up to 2.79 % can be
achieved if the quadratic inconvenience function is used. When 12 vehicles are used,
the use of more uncertain stochastic knowledge can gain significant improvements.
Especially if the quadratic inconvenience function is used, a meaningful number of
requests can be serviced earlier (see Figs. 8.22 and 8.23). As can be seen in Ta-
ble 8.9, significant improvements in the solution quality can be attained for both
evaluated inconvenience functions. Specifically, if the linear inconvenience func-
tion is used, improvements of up to 1.7 % can be achieved whereas improvements
of up to 7.93 % can be achieved if the quadratic inconvenience function is used. The
complete figures depicting the results of the computational experiments on S REAL
are illustrated in the Appendix in Sect. C.1.

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.23 Request response times on S REAL using 12 vehicles and DCminλ = 2.0 grouped by
intervals of 50 seconds
240 8 Computational Results

Request arrivals per subarea (SMA30 minutes) 1.8

1.6

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0.0
07:30 08:00 08:30 09:00 09:30 10:00 10:30 11:00

Time

Fig. 8.24 Request arrival rates in the 72 subareas during the considered planning horizon in
S REAL . All values are calculated using a simple moving average of 30 minutes

limits the performance attainable by applying PROACTIVE10. Specifically, al-

though the arrival rate of requests varies during the day (see again Fig. 8.2), the
occurrence probability of request arrivals in each subarea of the considered ser-
vice area remains almost constant. This can be derived from Fig. 8.24 in which the
request arrival rates in 72 subareas during the considered planning horizon are illus-
trated using a simple moving average with a length of 30 minutes. Note that these
72 subareas correspond to the ones which were also used for generating stochas-
tic knowledge (see also Fig. 6.1). The most important aspect which can be derived
from this figure is that the relative ordering of request arrival rates in the subareas
in which most of the requests arrive remains almost constant during the day. As a
consequence, vehicles are suitably located to service future arriving requests once
they have spread over the service area. In such cases, by applying the idle vehi-
cle waiting strategy (see Sect. 5.2.4), TIMELIMIT10 already makes efficient use
of this positioning so that attainable improvements of utilizing stochastic knowl-
edge are strongly limited. Therefore, we state that the request data set of S REAL
has a low structural diversity that makes an advantageous application of PROAC-
TIVE10 unlikely. Moreover, the subareas in which the highest request arrivals can
be observed are adjacently located in the middle of the considered service area. This
can be explained by the fact that most of the subsequent delivery requests occur in
the downtown city area where the population density in the considered service area
is highest. Consequently, the generation of stochastic knowledge with high quality,
i.e., using high DCminλ values, results in dummy customers which are only placed
in these subareas. Since the depot is also located there, vehicles are by default well
positioned for servicing future requests so that the additional value of stochastic
knowledge is strongly limited. Further computational experiments on S REAL using
8.8 Performance Evaluation of the Pro-Active Real-Time Control Approach 241

stochastic knowledge solely generated from past request information of the same
weekday did not result in any significant improvements either.
Besides the low structural diversity, individual days of S REAL may possess struc-
tural differences due to its real-world character. Specifically, request arrivals in this
request data set may contain additional request arrival structures so that the request
arrivals cannot entirely be represented by a set of time-space Poisson processes that
are constant from day to day. Therefore, stochastic knowledge generated by using
the last nf = 60 weekdays may represent a rather unreliable prediction of future re-
quests. However, as shown by our analysis in Sect. 8.6, it could not be determined
that S REAL possesses a high structural distortion. This is underscored by another
∗
test study in which a request data set denoted as S REAL is generated based on the
request arrival characteristics observed in S REAL as shown in Fig. 8.24. Therefore,
∗
possible existing day-to-day differences of S REAL are eliminated so that S REAL has
a low level of structural distortion with certainty. Computational results conducted
∗
on S REAL show a comparable solution quality to the one obtained on S REAL . Hence,
we conclude that if there were any day-to-day differences in S REAL , they would
only represent a minor reason for the limited improvements attained by PROAC-
TIVE10 in comparison to those attained by TIMELIMIT10.
With regard to our definition of structural quality depending on structural dis-
tortion and structural diversity, we conclude that a low structural distortion is a
prerequisite in order to efficiently apply the proposed forecasting method and the
pro-active real-time control approach. As shown at the example of S REAL , we fur-
ther state that a low structural diversity generally impedes an efficient utilization
of stochastic knowledge in pro-active real-time control approaches. Accordingly,
we state that the structural quality of request data mainly depends on the structural
diversity. Hence, S REAL also has a low structural quality which limits the value
of attainable stochastic knowledge thereof. Crucial characteristics of S REAL which
lead to the low structural diversity and hence the limited improvements that were
attained are further analyzed in Sect. 8.9.

8.8.2 Results on S GEN

In this section, we present the computational results obtained by applying the deter-
ministic and the pro-active real-time control approach on the 25 generated request
sets of the request data class S GEN .
In Table 8.10, the additional improvements attained by PROACTIVE10 in com-
parison to TIMELIMIT10 for all structural diversity settings of S GEN as well as the
number of worse test instances out of the corresponding 30 test instances are illus-
trated. From the results in this table, it can be derived that the level of structural di-
versity significantly influences the improvement rates attainable by PROACTIVE10.
This can be explained by the fact that if higher levels of structural diversity are
present in a request data set, stochastic knowledge has a higher value. Specifically,
since in this case request arrival rates in individual regions strongly vary during the
242

Table 8.10 Percentage improvements achieved by PROACTIVE10 vs. TIMELIMIT10 (i.e., improvements attained by utilizing stochastic knowledge) on S GEN
Vehicles RegionDiversity TimeDiversity (TD)
(RD) 0.00 0.25 0.50 0.75 1.00

linear2X customer inconvenience

8 0.00 0.08 % (1.0, ∗) ↓ 1.23 % (1.0, ∗) ↑ −0.18 % (1.2, ∗) ↓ 1.95 % (1.2, ∗) ↓ 0.26 % (1.5, ∗)
0.25 ↑ −0.21 % (1.0, ∗) 1.77 % (1.0, ∗) ↑ 0.35 % (1.0, ∗) 2.51 % (1.2, ∗) ↓ 4.48 % (1.5, ∗)
0.50 1.46 % (1.0, 8) ↓ 1.85 % (1.2, ∗) ↓ 3.92 % (1.2, ∗) ↓ 2.92 % (1.5, 8) ↓ 3.85 % (1.5, ∗)
0.75 −0.55 % (1.5, ∗) 1.65 % (1.5, ∗) ↓ 5.06 % (1.5, 7) ↓ 7.94 % (1.5, 6) ↓ 10.83 % (1.5, 1)
1.00 ↑ −0.33 % (1.0, ∗) ↓ 2.98 % (1.0, 7) 6.30 % (1.5, 5) ↓ 8.92 % (1.5, 5) ↓ 19.51 % (2.0, 1)
10 0.00 0.51 % (1.0, ∗) 1.86 % (1.0, 7) 2.30 % (1.2, 8) 3.22 % (1.2, 8) ↓ 2.28 % (1.2, ∗)
0.25 0.79 % (1.0, ∗) 2.27 % (1.0, 7) 4.07 % (1.0, 6) ↓ 5.21 % (1.2, 4) ↑ 4.84 % (1.2, 6)
0.50 4.72 % (1.0, 3) 5.37 % (1.0, 3) 5.19 % (1.0, 4) 6.44 % (1.5, 5) ↓ 10.02 % (1.5, 0)
0.75 4.44 % (1.0, 3) 6.12 % (1.0, 2) 7.84 % (1.2, 4) 11.14 % (1.0, 2) 14.61 % (1.5, 1)
1.00 3.87 % (1.0, 3) 6.75 % (1.0, 2) 9.65 % (1.2, 0) 14.39 % (1.5, 1) 22.95 % (2.0, 0)
12 0.00 1.22 % (1.0, 9) 3.18 % (1.0, 5) 2.04 % (1.0, ∗) 4.52 % (1.0, 3) 5.95 % (1.0, 6)
8

0.25 0.87 % (1.0, ∗) 2.37 % (1.0, ∗) 4.50 % (1.0, 2) 6.11 % (1.0, 4) 6.38 % (1.2, 3)
0.50 5.48 % (1.0, 2) 6.21 % (1.0, 1) 8.05 % (1.0, 0) 9.04 % (1.2, 0) 11.34 % (1.5, 0)
0.75 6.09 % (1.0, 1) 8.46 % (1.0, 0) 9.97 % (1.0, 1) 15.85 % (1.0, 0) 20.27 % (1.2, 0)
1.00 5.81 % (1.0, 3) 9.19 % (1.0, 1) 11.86 % (1.2, 0) 17.99 % (1.2, 0) 26.72 % (1.5, 0)
Computational Results
Table 8.10 (Continued)
Vehicles RegionDiversity TimeDiversity (TD)
(RD) 0.00 0.25 0.50 0.75 1.00

quadratic customer inconvenience

8 0.00 1.48 % (1.0, ∗) 3.02 % (1.0, ∗) ↓ 3.43 % (1.0, 9) ↓ 4.21 % (1.2, 7) ↑ 4.57 % (1.2, ∗)
0.25 ↑ 2.46 % (1.0, ∗) ↓ 5.44 % (1.0, 7) ↓ 6.72 % (1.0, ∗) ↓ 7.17 % (1.2, 9) ↓ 3.00 % (1.5, ∗)
0.50 ↑ 4.14 % (1.2, 8) ↑ 5.66 % (1.0, 9) ↓ 10.45 % (1.0, 6) ↓ 9.20 % (1.2, 3) ↓ 7.06 % (1.5, ∗)
0.75 5.51 % (1.0, ∗) ↓ 5.26 % (1.0, 8) ↑ 9.99 % (1.2, 6) ↓ 14.63 % (1.5, 3) ↓ 16.75 % (1.5, 2)
1.00 8.26 % (1.0, 7) 9.98 % (1.2, 2) ↓ 15.16 % (1.5, 3) ↓ 22.64 % (1.5, 3) ↓ 29.86 % (2.0, 0)
10 0.00 1.37 % (1.0, ∗) 6.40 % (1.0, 6) 7.79 % (1.0, 6) 9.16 % (1.0, 5) 5.08 % (1.2, ∗)
0.25 1.80 % (1.0, ∗) 8.19 % (1.0, 5) 10.13 % (1.0, 2) ↓ 14.49 % (1.2, 0) ↓ 10.05 % (1.5, 3)
0.50 9.08 % (1.0, 3) 11.32 % (1.0, 2) 15.31 % (1.0, 1) 14.05 % (1.0, 4) ↓ 17.89 % (1.5, 1)
0.75 9.59 % (1.0, 0) 13.60 % (1.0, 1) 19.88 % (1.0, 1) 20.08 % (1.0, 1) 27.30 % (1.5, 0)
1.00 12.20 % (1.0, 4) 14.38 % (1.0, 2) 19.56 % (1.2, 2) 25.78 % (1.5, 1) 39.33 % (2.0, 0)
12 0.00 2.34 % (1.0, 8) 5.02 % (1.0, 8) 7.74 % (1.0, 3) 13.16 % (1.0, 1) 15.36 % (1.0, 1)
0.25 0.42 % (1.0, ∗) 8.08 % (1.0, 2) 14.04 % (1.0, 2) 17.05 % (1.0, 0) 15.05 % (1.0, 2)
0.50 8.72 % (1.0, 2) 13.60 % (1.0, 0) 16.70 % (1.0, 1) 18.24 % (1.0, 1) 21.64 % (1.2, 1)
0.75 10.17 % (1.0, 2) 15.68 % (1.0, 1) 21.01 % (1.0, 0) 29.25 % (1.0, 0) 37.15 % (1.2, 0)
8.8 Performance Evaluation of the Pro-Active Real-Time Control Approach

1.00 11.48 % (1.0, 2) 15.40 % (1.0, 0) 24.00 % (1.0, 0) 33.72 % (1.0, 0) 47.24 % (1.2, 0)

Note: Illustrated is the average improvement. The values in brackets denote the best DCminλ value and the number of worse test instances (∗: more than 9 worse).
↑: more total late requests than in TIMELIMIT10, ↓: less total late requests than in TIMELIMIT10
243
244 8 Computational Results

30%

25%
8 vehicles
20%
Improvement

10 vehicles
12 vehicles
15%
Quadratic (8 vehicles)
10% Quadratic (10 vehicles)
Quadratic (12 vehicles)
5%

0%
0.00/0.00 0.25/0.25 0.50/0.50 0.75/0.75 1.00/1.00
RD and TD structural diversity values

Fig. 8.25 Attainable improvements on S GEN at different levels of structural diversity using lin-
ear2X customer inconvenience

50%
45%
40%
8 vehicles
35%
Improvement

10 vehicles
30%
12 vehicles
25%
Quadratic (8 vehicles)
20%
Quadratic (10 vehicles)
15%
10%
Quadratic (12 vehicles)

5%
0%
0.00/0.00 0.25/0.25 0.50/0.50 0.75/0.75 1.00/1.00
RD and TD structural diversity values

Fig. 8.26 Attainable improvements on S GEN at different levels of structural diversity using
quadratic customer inconvenience

day, vehicle positions resulting from last serviced requests do not represent a good
positioning for future request arrivals. Hence, future arriving requests are likely to
be serviced with a high response time and thus a high customer inconvenience. It
is worth mentioning that increasing both levels of structural diversity conjectures
a quadratic dependency between the level of structural diversity and attainable im-
provements, independent of the utilized customer inconvenience function and fleet
size (see Figs. 8.25 and 8.26).
In Figs. 8.27 and 8.28, a visual analysis of two request data sets of S GEN is
performed. In these tables, the impacts of high levels of structural diversity on tem-
poral and spatial changes in request arrival rates can be clearly observed. Specif-
ically, according to the values defined in S GEN (see again Fig. 8.4), the highest
request arrival rates occur in different subareas during the day which are also lo-
cated far away from each other. Therefore, the subarea with the highest request ar-
8.8 Performance Evaluation of the Pro-Active Real-Time Control Approach 245

Request arrivals per subarea (SMA30 minutes) 4.0

3.5

3.0

2.5

2.0

1.5

1.0

0.5

0.0
07:30 08:00 08:30 09:00 09:30 10:00 10:30 11:00

Time

Fig. 8.27 Request arrival rates in the 72 subareas during the considered planning horizon in
S GEN RD = TD = 1.00. All values are calculated using a simple moving average of 30 minutes
Request arrivals per subarea (SMA30 minutes)

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0.0
07:30 08:00 08:30 09:00 09:30 10:00 10:30 11:00

Time

Fig. 8.28 Request arrival rates in the 72 subareas during the considered planning horizon in
S GEN RD = TD = 0.50. All values are calculated using a simple moving average of 30 minutes

rival rates significantly changes in the considered service area during the day. Note
that this significantly distinguishes S GEN from the request data class S REAL (see
again Fig. 8.24). Since the idle vehicle waiting strategy that is applied in TIME-
LIMIT10 relies on the value of these positions, TIMELIMIT10 is significantly
outperformed by PROACTIVE10 which efficiently repositions vehicles to future
request-likely areas in advance if high levels of structural diversity are present. This
has an even higher impact since the highest request arrival rates are about twice as
246 8 Computational Results

high in S GEN RD = TD = 1.00 as in S REAL (i.e., 3.5 compared to 1.7). However,

also in the request data set S GEN RD = TD = 0.50, in which the highest request ar-
rival rates are about 1.4 and therefore lower than in S REAL , improvements between
3.75 % and 18.32 % could be attained (see Table 8.10). These improvements are
significantly higher than the ones achieved for S REAL . Hence, it can be concluded
that the value of generatable stochastic knowledge and thus the structural quality of
a request data set is significantly determined by variations of request arrivals in the
subareas during the planning horizon. As can be seen in Fig. 8.28, the mentioned
changes in request arrival rates in the individual subregions also exist in the request
data set S GEN RD = TD = 0.50. As a result, since S GEN has a low structural distor-
tion, the level of structural diversity present in a request data set of S GEN directly
influences its structural quality. This interrelation is further discussed in Sect. 8.9.
Moreover, in comparison to S REAL , substantial improvements can also be achieved
even with the smallest fleet size of 8 vehicles in S GEN . Hence, the application of the
pro-active real-time control approach is also advantageous in case of a small fleet
size if a specific minimum level of structural diversity is present.
Furthermore, different impacts on the achieved solution quality according to the
chosen DCminλ setting can be observed. Specifically, if low levels of structural di-
versity is present, low DCminλ settings achieve the best results. This can be ex-
plained by considering Table 8.7 which illustrates the number of generated dummy
customers depending on the levels of structural diversity and DCminλ . It becomes
obvious that a minimum level of structural diversity is necessary to generate a suf-
ficient amount of stochastic knowledge of the minimum demanded reliability. For
example, using the setting DCminλ = 2.0, minimum levels of structural diversity
of, e.g., RD = 0.50, TD = 1.00 or RD = TD = 0.75 is necessary. Therefore, low
levels of structural diversity settings do not allow for the generation of stochastic
knowledge which comprises requests expected with a high reliability.
Two further aspects become relevant when a small fleet size of 8 vehicles is
used. First, as shown in Table 8.10, applying PROACTIVE10 reduces the number
of late requests in almost all structural diversity settings. This implies that also in
TIMELIMIT10 the vehicles are very busy in the evaluated test instances of S GEN .
Second, similar aspects as observed in S REAL may become important. Specifically,
if a request data set with low levels of structural diversity exists, the utilization of
stochastic knowledge should be carefully considered. For example, considering the
setting RD = 0.00, TD = 1.00, only stochastic knowledge of lower quality is avail-
able due to the low level of RD in this setting (see Table 8.7). This leads to a reduced
solution quality due to the limited resources as depicted in Fig. 8.29. Note that in this
figure, the reduced number of requests serviced just before the maximum allowed
response time in PROACTIVE10 is caused by additional requests that are serviced
after the maximum allowed response time. Nevertheless, it is worth mentioning that
when a larger fleet size is used (e.g., 10 vehicles), the generated stochastic knowl-
edge can be advantageously utilized even in this request data set (see Fig. 8.30).
In contrast to the discussed effects in scenarios where a low structural diver-
sity is present, in scenarios with a higher structural diversity, a significant number
8.8 Performance Evaluation of the Pro-Active Real-Time Control Approach 247

linear2X, TIMELIMIT10
linear2X, PROACTIVE10
serviced requests

quadratic, TIMELIMIT10
quadratic, PROACTIVE10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.29 Request response times on S GEN using 8 vehicles and DCmin λ = 1.5 for RD = 0.00 and
TD = 1.00 grouped by intervals of 50 seconds

linear2X, TIMELIMIT10
linear2X, PROACTIVE10
serviced requests

quadratic, TIMELIMIT10
quadratic, PROACTIVE10

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.30 Request response times on S GEN using 10 vehicles and DCmin λ = 1.2 for RD = 0.00
and TD = 1.00 grouped by intervals of 50 seconds

of future requests can be reliably forecasted. By efficiently integrating this high-

quality stochastic knowledge into the pro-active real-time control approach, sig-
nificant reductions of customer inconvenience can be attained. Specifically, larger
DCminλ values perform best for fleet sizes of 8 and 10 vehicles. If the fleet size
is enlarged to 12 vehicles and vehicles are therefore less utilized, further improve-
ments are possible by also integrating less reliable stochastic knowledge into the
pro-active real-time control approach. This does not apply for smaller fleet sizes
since in these settings vehicles are busier and a misplacement of vehicles caused by
an incorrect forecast has substantial negative consequences. For the structural diver-
sity setting RD = TD = 1.00, the request response time distributions for the DCminλ
settings of 1.0, 1.5, and 2.0 are illustrated in Figs. 8.31, 8.32, and 8.33 for all three
evaluated fleet sizes. As can be seen in Fig. 8.31, the integration of stochastic knowl-
edge significantly reduces the number of almost late serviced requests when a re-
stricted fleet size is present. Moreover, in all three fleet sizes, a considerably higher
number of requests is quickly serviced when stochastic knowledge is utilized. The
request response times attained on additional request data sets of S GEN are illus-
trated in the Appendix in Sect. C.2. The request response times of all request data
sets of S GEN can be found online.3

3 Available at http://extras.springer.com.
248 8 Computational Results

linear2X, TIMELIMIT10
linear2X, PROACTIVE10, DC minλ = 1.0
linear2X, PROACTIVE10, DC minλ = 1.5
serviced requests

linear2X, PROACTIVE10, DC minλ = 2.0

quadratic, TIMELIMIT10
quadratic, PROACTIVE10, DC minλ = 1.0
quadratic, PROACTIVE10, DC minλ = 1.5
quadratic, PROACTIVE10, DC minλ = 2.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.31 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0, 1.5, and 2.0 for
RD = 1.00 and TD = 1.00 grouped by intervals of 50 seconds

linear2X, TIMELIMIT10
linear2X, PROACTIVE10, DC minλ = 1.0
linear2X, PROACTIVE10, DC minλ = 1.5
serviced requests

linear2X, PROACTIVE10, DC minλ = 2.0

quadratic, TIMELIMIT10
quadratic, PROACTIVE10, DC minλ = 1.0
quadratic, PROACTIVE10, DC minλ = 1.5
quadratic, PROACTIVE10, DC minλ = 2.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.32 Request response times on S GEN using 10 vehicles and DC min λ
= 1.0, 1.5, and 2.0 for
RD = 1.00 and TD = 1.00 grouped by intervals of 50 seconds

linear2X, TIMELIMIT10
linear2X, PROACTIVE10, DC minλ = 1.0
linear2X, PROACTIVE10, DC minλ = 1.5
serviced requests

linear2X, PROACTIVE10, DC minλ = 2.0

quadratic, TIMELIMIT10
quadratic, PROACTIVE10, DC minλ = 1.0
quadratic, PROACTIVE10, DC minλ = 1.5
quadratic, PROACTIVE10, DC minλ = 2.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.33 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0, 1.5, and 2.0 for
RD = 1.00 and TD = 1.00 grouped by intervals of 50 seconds

To sum up the computational results of the pro-active real-time control approach,

two main criteria can be highlighted that are necessary for an efficient application
of our proposed pro-active real-time control approach. First, a sufficient structural
quality is of significant importance. It presupposes a minimum level of structural di-
versity as well as a low level of structural distortion. According to the computational
results, we conclude that the level of structural diversity mainly defines the struc-
tural quality of the individual request data sets of S GEN , i.e., request data sets with
high RD and TD structural diversity values possess a high structural quality, and
8.9 The Degree of Structural Diversity 249

vice versa. Second, the number of available resources (i.e., the fleet size) directly
influences the minimum quality of the stochastic knowledge required for attaining
promising results. Specifically, if an increased fleet size is available, the minimum
required structural quality is decreased since the negative effects of incorrectly fore-
casted requests diminish due to lower utilization of the vehicles.

8.9 The Degree of Structural Diversity

In the previous section, a two-dimensional structural diversity was introduced in

order to generate request data sets with different structural quality. In our computa-
tional results, the impact of the structural diversity present in different request data
sets of S GEN on the attainable stochastic knowledge and on the attainable solution
quality improvements has been evaluated. Particularly, it was shown that an effi-
cient generation and integration of stochastic knowledge requires a minimum level
of RegionDiversity (RD) as well as TimeDiversity (TD). However, although this
provides new significant insights into the importance of structural quality of request
data with regard to the efficiency of pro-active real-time control approaches, the
proposed two-dimensional structural diversity does not provide a general definition
of structural quality. Specifically, by applying it as presented in Sect. 8.2.4.2, dif-
ferent scenarios with individual levels of structural quality are newly generated by
linearly transforming an initial setting of a defined highest structural diversity which
is given by RD = TD = 1.00. Therefore, the generated request data sets neither al-
low for a classification of given request data sets, e.g., real-world data sets such as
S REAL nor do they cover all relevant characteristics. For instance, the initial set-
ting RD = TD = 1.00 is not the maximum structural diversity (and hence structural
quality) that can exist in a request data set of the considered RDOPG application.
Instead, this setting represents a request data set in which the freely chosen maxi-
mum diversity values defined in Fig. 8.4 are fully present. Such a general measure
for structural quality may increase the practicability of the proposed pro-active real-
time control approach since it is likely to significantly support the identification of
scenarios where the integration of stochastic knowledge is reasonable. As mentioned
in Sect. 4.8, this is also motivated by Bent and van Hentenryck (2005) and Ghiani
et al. (2009) who suggest analyzing the impact of the quality of the available past
request information on the performance of pro-active real-time control approaches.
Furthermore, such a measure may help to choose appropriate parameters for the pro-
active real-time control approach according to the considered real-world application.
Consequently, we propose a first approach for determining the structural quality of
arbitrary given request data sets in what follows. We evaluate the applicability of
this approach on S REAL and S GEN .
In order to determine the structural quality of a given request data set, we propose
a measure denoted as the degree of structural diversity (dosd). In contrast to the
previously used two-dimensional structural diversity, the dosd is a one-dimensional
value. It ranges from 0 to 1. While 0 corresponds to a request data set with no
250 8 Computational Results

structural diversity and hence does not allow for efficiently integrating stochastic
knowledge due to the existing low structural quality, the value 1 defines a theoretical
upper bound on the maximum structural diversity which a request data set can have.
Based on the insights of the previous section, the main idea of the dosd is that a
high degree of structural diversity is present if there exists significant spatial vari-
ations in request arrivals in the subareas during the considered planning horizon.
Specifically, if the subareas in which high request arrival rates subsequently occur
are located far away from each other, the value of using stochastic knowledge and
hence the dosd becomes high. Moreover, also the number of requests that occur in
the individual subareas is important. Specifically, the more requests occur in a sub-
area, the higher the impact on the solution quality since these requests are serviced
with a high response time leading to a high customer inconvenience. Furthermore,
changes in the request arrival rates in the subareas change the vehicles’ positions
within the considered service area accordingly. Therefore, in order for a request
data set to obtain a high dosd, changes in the request arrival rates should occur only
after a certain time period so that vehicles will have traveled to another position in
the meantime.
For a given request data set, the dosd is determined as follows. The rate param-
eters λ(ps ) of the time-space Poisson processes attached to each segment s ∈ S
are calculated as described in Sect. 6.2.1 for a specified time frame of the given
request data set, e.g., the last nf working days. The considered planning horizon
of the corresponding real-world application is divided into T time periods with a
y
length of m minutes each. For each time period t ∈ {1, . . . , T }, the tuple (btx , bt )
defines the geographical position of the barycenter of requests that have arrived in
time period t according to the subarea separation depicted in Fig. 6.1. Moreover, nt
gives the number of requests which occurred on average in time period t in the
considered time frame of the request data set. Furthermore, we assume that the con-
sidered service area is divided into x · y subareas. In order to calculate the dosd,
we track how the barycenter of request arrivals changes between consecutive time
periods. Specifically, the Euclidean distance between the barycenters of two consec-
utive time periods t1 , t2 ∈ T is computed and weighted with the number of request
arrivals of t2 given by nt2 . By calculating the weighted distances for all time pe-
riod pairs and normalizing the sumof weighted distances with an upper bound on
the distance defined by maxDist = x 2 + y 2 , we obtain the following definition of
the dosd:
T −1
y y
t=1 nt+1 · (bt − bt+1 ) + (bt − bt+1 )
x x 2 2
dosd = T −1
maxDist · t=1 nt+1
Note that since in our application all vehicles start their tour at an identical depot,
the first time period is used to determine individual starting positions of the vehicles.
However, since in many request data sets considered in our RDOPG applications
only a small portion of all requests arrive in the first time period, n1 is comparably
small and its omission does not cause significant changes to the calculated dosd.
According to this definition, variations in request arrivals of a given request data set
8.9 The Degree of Structural Diversity 251

are quantified by the weighted length of an approximate path that the majority of
the vehicle fleet travels in order to service the requests. Consequently, the larger the
resulting distances between consecutive barycenters and the higher the number of
requests, the more valuable the utilization of stochastic knowledge becomes. This
is due to the fact that in such situations a deterministic real-time control approach
is increasingly forced to relocate a larger number of vehicles in each time period
which is likely to result in higher response times and hence increased customer in-
convenience. Therefore, parameter m is approximately set to the average duration
during which the vehicles have completed their travel activity between two consec-
utive barycenters.
In order to assess the practicability of the proposed dosd, we evaluate it on the
request data classes S REAL and S GEN . Note that all request data sets of these request
data classes have a comparable number of requests. Consequently, if a small number
of vehicles is available, they are highly utilized while a large number of available
vehicles results into the vehicles becoming less utilized in all of these request data
sets. Hence, the values of the dosd calculated for the request data sets of both request
data classes are comparable, i.e., the stochastic knowledge which can be generated
from them has a similar value. Specifically, different request data sets with a similar
dosd value are likely to generate a similar improvement in the solution quality when
the pro-active real-time control approach is used compared to using the determinis-
tic real-time control approach. The advantage that the dosd has over computational
experiments is that it makes an instant estimation of the efficiency of pro-active real-
time control approaches on given request data sets possible. In order to statistically
prove the above statement, we directly compare the respective dosd value calcu-
lated for each request data set of both request data classes with the improvement
rates which have been attained by using PROACTIVE10 in the previously described
computational experiments. In this comparison, we use the best improvement rates
attained among the evaluated parameter settings for DCminλ which are illustrated
in Tables 8.9 and 8.10. As presented in the computational experiments, these best
improvement rates were attained by conducting the five DCminλ settings on each
of the 25 request data sets of S GEN as well as for the request data set of S REAL .
In order to analyze the correlation between both variables, we apply six linear re-
gression models in total—one for each of both customer inconvenience functions
combined with each of the three evaluated fleet sizes on the 25 request data sets of
S GEN . For calculating the dosd, we set m = 30 resulting in eight time periods of
the considered planning horizon which has a duration of four hours as mentioned
above.
The results of this evaluation are illustrated in Figs. 8.34 and 8.35. As can be
seen from these figures, the proposed dosd is an effective measure for determin-
ing the structural quality of request data sets and can therefore identify scenar-
ios in which the application of stochastic knowledge is advantageous. Specifically,
for S GEN , a detailed correlation analysis using the Pearson product-moment cor-
relation coefficient r (cf. Rodgers and Nicewander 1988) provides values between
0.9469 and 0.9647 for the linear2X customer inconvenience function and values
between 0.9161 and 0.9491 for the quadratic customer inconvenience function,
252 8 Computational Results

30%

R2 = 0.9228 SGEN, 8 vehicles

25% SREAL, 8 vehicles
SGEN, 10 vehicles
R2 = 0.9306
Improvement

20% SREAL, 10 vehicles

SGEN, 12 vehicles
R2 = 0.8966
15% SREAL, 12 vehicles
Linear (SGEN, 8 vehicles)
Linear (SGEN, 10 vehicles)
10%
Linear (SGEN, 12 vehicles)

0%
0.00 0.02 0.04 0.06 0.08 0.10 0.12

dosd

Fig. 8.34 Results of the degree of structural diversity using the linear2X customer inconvenience
function

50%
R2 = 0.9007
SGEN, 8 vehicles
40% SREAL, 8 vehicles
R2 = 0.8849 SGEN, 10 vehicles
Improvement

SREAL, 10 vehicles
30% SGEN, 12 vehicles
R2 = 0.8393 SREAL, 12 vehicles
Linear (SGEN, 8 vehicles)
20%
Linear (SGEN, 10 vehicles)
Linear (SGEN, 12 vehicles)

10%

0%
0.00 0.02 0.04 0.06 0.08 0.10 0.12

dosd

Fig. 8.35 Results of the degree of structural diversity using the quadratic customer inconvenience
function

depending on the utilized fleet size. Accordingly, the corresponding coefficient of

determination R 2 (defined by r 2 ) is illustrated in both figures. Consequently, a
strong linear relationship between the dosd and the improvement that is attainable
by the pro-active real-time control approach is statistically proven. This indicates
a good quality for using the dosd on S GEN . The results also reproduce the pre-
vious observation in Sect. 8.8 that the attainable improvement rates increase with
the available fleet size. Moreover, the figures show that the dosd represents a good
measure for estimating attainable improvements on the real-world request data set
S REAL . Specifically, the determined low dosd value for S REAL reflects the obser-
vations in Sect. 8.8 with regard to the low structural data quality of S REAL . Since
the dosd provided good results on S REAL and S GEN , it is likely that the dosd is
8.9 The Degree of Structural Diversity 253

20
linear2X
Number of available vehicles

18 linear2X (estimated)
16 quadratic

14 quadratic (estimated)
PROACTIVE10 is
12 recommended for
10 quadratic
PROACTIVE10 is
8
recommended for
6 linear2X and quadratic
4
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

dosd

Fig. 8.36 Managerial recommendations based on the dosd and S GEN demanding a minimum im-
provement of 10 %

20
linear2X
Number of available vehicles

18
linear2X (estimated)
16 quadratic

14 quadratic (estimated)
PROACTIVE10 is
12 recommended for
10 quadratic
PROACTIVE10 is
8
recommended for
6 linear2X and quadratic

4
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14
dosd

Fig. 8.37 Managerial recommendations based on the dosd and S GEN demanding a minimum im-
provement of 15 %

also a suitable measure for determining the structural quality of other request data
sets. According to the attained results, it can be concluded that a dosd ≥ 0.06 of a
request data set strongly indicates that the existing structural quality enables an ad-
vantageous application of the proposed pro-active real-time control approach on this
request data set. In accordance with the computational results provided in Sect. 8.8,
this is particularly true if a larger fleet size is available. Moreover, since the attain-
able improvements utilized in this evaluation represent the best parameter settings
of the tested ones, appropriate parameter values for generating stochastic knowledge
and for utilizing the proposed pro-active real-time control approach can be chosen
according to the dosd of a given request data set.
Based on the results of the dosd on S GEN , a two-dimensional matrix providing
managerial recommendations can be derived. Specifically, Figs. 8.36, 8.37, and 8.38
illustrate the scenario constellations in which the utilization of the pro-active
254 8 Computational Results

20
linear2X
Number of available vehicles

18 linear2X (estimated)
16 quadratic
quadratic (estimated)
14
PROACTIVE10 is
12 recommended for
10 quadratic
PROACTIVE10 is
8 recommended for
6 linear2X and quadratic

4
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14
dosd

Fig. 8.38 Managerial recommendations based on the dosd and S GEN demanding a minimum im-
provement of 20 %

real-time control approach is recommended, depending on the utilized customer in-

convenience function and on a minimum demanded improvement which can be at-
tained by PROACTIVE10 compared to TIMELIMIT10. Clearly, a lower minimum
demanded improvement rate increases the number of scenarios in which the pro-
active real-time control approach can be efficiently applied. This is also true when
the quadratic customer inconvenience function is used. Moreover, based on the eval-
uated request data sets of S GEN it is conjectured that differences in the applicability
of PROACTIVE10 with regard to the linear2X and the quadratic objective function
decrease with higher degrees of structural diversity. Specifically, in request data sets
with a low structural quality, using the quadratic customer inconvenience function
has a beneficial impact since, as mentioned above, it avoids situations in which many
requests have a high response time close to the R mrt due to incorrectly forecasted
requests. In contrast, if the request data set comprises a high structural quality, the
number of incorrectly forecasted requests is significantly reduced. Hence, since a
high structural diversity reduces the number of incorrectly forecasted requests and
increased the value of the generatable stochastic knowledge, this issue is expected
to become less significant.
To summarize the results obtained from this evaluation, the proposed dosd is
an effective general measure for assessing the structural quality of arbitrary given
request data sets. Utilizing it for a given request data set makes it possible to
estimate the solution quality improvements which are additionally attainable by
utilizing PROACTIVE10 compared to TIMELIMIT10. Moreover, it supports the
identification of parameter settings that can be recommended for efficiently apply-
ing stochastic knowledge in real-time control approaches. These insights signifi-
cantly improve the practicability of the proposed pro-active real-time control ap-
proach.
8.10 Driver Inconvenience Aspects in the Real-Time Control Approaches 255

8.10 Driver Inconvenience Aspects in the Real-Time Control

Approaches
In this section, we continue to analyze the impact on the solution quality when
allowing vehicle en-route diversion activities. Specifically, in the previously pre-
sented computational experiments, vehicle en-route diversion is allowed without
any penalty costs. Hence, vehicle en-route diversion activities are considered by
the solution method as additional decisions that are available without any draw-
backs. However, in practical applications, allowing an arbitrary number of vehicle
en-route diversion activities may expose a huge amount of inconvenience to the ve-
hicle drivers. This may result in a lack of motivation for work so that the positive
effects that can be attained by utilizing the pro-active real-time control approach
cannot be completely realized. On the other hand, as presented in Sect. 8.5, allow-
ing vehicle en-route diversion provides an effective way of significantly increasing
the solution quality. Therefore, although a complete prohibition of vehicle en-route
diversion activities does not seem to be reasonable, it is advisable to include mech-
anisms that limit the number of performed vehicle en-route diversion activities. In
order to do so, it is recommended to evaluate the trade-off of limiting vehicle en-
route diversion activities and the attained solution quality.
In this section, we propose a method for limiting the number of performed vehicle
en-route diversion activities. This limitation operates by only permitting vehicle en-
route diversion activities whose improvement with regard to the objective function
value is above a certain threshold. Specifically, penalty costs for performing vehicle
en-route diversion activities are temporally included in the objective function during
the solution process. Hence, only en-route diversion activities, whose improvement
is larger than these penalty costs, are performed. Therefore, in this computational
study, we do not only take into account inconvenience on the customer’s side, but
also inconvenience aspects of the deployed drivers.
Note that due to the different evolutions of both tested objective functions as de-
scribed in Sect. 5.3.2.4, these penalty costs have to be individually defined for each
customer inconvenience function. Specifically, the previously mentioned threshold
is defined by a total decrease in request response times that is required to be achieved
in order to justify a vehicle en-route diversion activity. This response time-based ap-
proach has been chosen since it is assumed that the effort to save a certain amount of
request response time is equal across scenarios with a different customer inconve-
nience function. For this threshold, several values which are illustrated in Table 8.11
are evaluated. For example, a defined response time threshold of 36 seconds means
that the total improvement of the objective function value by performing the vehicle
en-route diversion activity is required to be larger than 0.01 in the linear2X case and
0.0001 in the quadratic case, respectively.
The computational experiments in this section are carried out using the real-
time control approaches TIMELIMIT10 and PROACTIVE10 on the request data
set S GEN RD = TD = 1.00 with a fleet size of 10 vehicles while evaluating both
customer inconvenience functions. For PROACTIVE10, according to the compu-
tational results depicted in Table 8.10, the best dummy customer quality settings
256 8 Computational Results

Table 8.11 Evaluated response time thresholds and penalty costs included in the objective func-
tion for vehicle en-route diversion activities depending on the applied customer inconvenience
function
Response time threshold [seconds] Objective function penalty costs
linear2X quadratic

(unrestricted) 0 0.0 0.0

36 0.01 0.0001
72 0.02 0.0004
180 0.05 0.0025
360 0.1 0.01
720 0.2 0.04
1800 0.5 0.25
3600 1.0 1.0

represented by DCminλ = 2.0 for the linear2X as well as for the quadratic customer
inconvenience function are used. The upper graph in Fig. 8.39 shows the performed
number of vehicle en-route diversion activities if different threshold parameter set-
tings are applied to linear2X. In the same figure, the respective changes in the
objective function value are presented in the lower graph. Analogously, Fig. 8.40
shows the results that have been achieved using the quadratic customer inconve-
nience function. Note that for the approach TIMELIMIT10, the number of vehicle
en-route diversion activities is recorded according to the rules described in Sect. 8.5.
For the approach PROACTIVE10, a change of a dummy customer to a real request
is not counted as a vehicle en-route diversion activity when a dummy customer is
currently assigned as the first request to a vehicle. Specifically, due to our defined
utilization of dummy customer requests, the purpose of dummy customers is to act
as place-holders for an expected number of real requests arriving in the correspond-
ing request-likely area. When a vehicle is currently traveling to the location of a
dummy customer request and its next request is changed to a real request, it is likely
that this change happens due to the fact that a real customer request which is located
near the dummy customer has recently arrived. Therefore, since it is the purpose of
dummy customers to guide vehicles into request-likely areas without knowing the
exact position of arriving requests, such a change is not considered as a deterioration
of the transportation process. It is assumed that drivers are notified by the on-board
units when they are currently traveling to a dummy customer so that they are aware
of a likely change of this dummy customer request. Similarly, a change of the first
assigned request, due to the removal of a dummy customer from the tour plan (see
Sect. 6.3.1), is not counted as a vehicle en-route diversion activity either. Neverthe-
less, a reassignment of a still active dummy customer to another dummy customer is
considered as a vehicle en-route diversion activity that is penalized in the objective
function and also recorded in the computational results (see Table 8.12).
In Figs. 8.41 and 8.42, the request response times resulting from the application
of particular penalty costs are illustrated. As can be seen, higher penalty costs lead to
8.10 Driver Inconvenience Aspects in the Real-Time Control Approaches 257

6.00
5.34
First request reassignments per vehicle

5.00
4.53

3.82 4.00
4.00
3.53
3.32
TIMELIMIT10
3.00
2.42 2.39
PROACTIVE10

2.00
1.59 1.59

0.79 0.71
1.00

0.13 0.11 0.04 0.01

0.00
0.0 0.01 0.02 0.05 0.1 0.2 0.5 1.0
Penalty value

0.45
0.3811 0.3927
0.40
0.3645
0.3486 0.3486 0.3464 0.3480 0.3546
0.35
Objective function value

0.3089
0.2830 0.2977
0.30 0.2738
0.2686 0.2659 0.2656 0.2692
0.25 TIMELIMIT10
0.20 PROACTIVE10

0.15

0.10

0.05

0.00
0.0 0.01 0.02 0.05 0.1 0.2 0.5 1.0
Penalty value

Fig. 8.39 Results of limiting vehicle en-route diversion activities on S GEN RD = TD = 1.00 with
DCminλ = 2.0 using the linear2X customer inconvenience function and a fleet size of 10 vehicles

longer response times which results in a later service of requests. Note that this is in
accordance with the results obtained in Sect. 8.5. Following the results presented in
both figures and Table 8.12, it can be observed that both proposed real-time control
approaches often make use of vehicle en-route diversion activities when vehicle en-
route diversion activities can be performed at no additional costs. Note that this is
similar to the results observed on S REAL presented in Sect. 8.5.
Furthermore, the results show that our proposed threshold-based approach can
be successfully applied in order to limit the number of performed vehicle en-route
activities. Interestingly, forbidding vehicle en-route diversion activities that would
gain only low objective function improvements significantly reduces the number of
vehicle en-route diversion activities but leads to only slight increases of the attained
average customer inconvenience. Moreover, in scenarios where the linear2X cus-
tomer inconvenience function is used, applying small threshold parameter values of
up to 0.02 does not increase the attained customer inconvenience at all but rather
258 8 Computational Results

6.00
5.26
First request reassignments per vehicle

5.24 5.14
5.00
4.64

3.92 3.89 3.86

4.00
3.61
3.22
3.02 TIMELIMIT10
3.00
PROACTIVE10

1.781.82
2.00

1.00
0.28 0.23
0.02 0.00
0.00
0.0 0.0001 0.0004 0.0025 0.01 0.04 0.25 1.0
Penalty value
0.25

0.1930
0.20
Objective function value

0.1748

0.1533 0.1549 0.1551 0.1576

0.1529 0.1529
0.15
0.1228 TIMELIMIT10
0.1136
0.0998 PROACTIVE10
0.0928 0.0945 0.0946 0.0937 0.0954
0.10

0.05

0.00
0.0 0.0001 0.0004 0.0025 0.01 0.04 0.25 1.0
Penalty value

Fig. 8.40 Results of limiting vehicle en-route diversion activities on S GEN RD = TD = 1.00 with
DCminλ = 2.0 using the quadratic customer inconvenience function and a fleet size of 10 vehicles

even reduces it while significantly reducing the number of vehicle en-route diver-
sion activities. This can be explained by the fact that disallowing low-improving
vehicle en-route diversion activities does not have a significant impact on the so-
lution quality. In contrast, by prohibiting these vehicle en-route diversion activities
vehicles are less disturbed while traveling to their next request so that, in the end,
this may even lead to a slight improvement in the achieved customer inconvenience.
As a positive effect, the number of vehicle en-route diversion activities can be sig-
nificantly reduced at the cost of negligible drawbacks while significantly reducing
driver inconvenience.
When vehicle en-route diversion activities are allowed without additional penalty
costs, the number of performed vehicle en-route diversion activities is significantly
higher in PROACTIVE10 compared to TIMELIMIT10. This is especially true for
the quadratic customer inconvenience function. This can be explained by the fact
that due to the quadratic increase of customer inconvenience, even many small ve-
8.10 Driver Inconvenience Aspects in the Real-Time Control Approaches 259

Table 8.12 Results for TIMELIMIT10 and PROACTIVE10 under different penalty costs on
S GEN RD = TD = 1.00 using 10 vehicles
Penalty TIMELIMIT10 PROACTIVE10
#ra per vehicle Objective function value #ra per vehicle Objective function value

linear2X customer inconvenience

0.0 4.53 0.3486 5.34 (−17.9 %) 0.2686 (22.9 %)
0.01 3.82 (15.5 %) 0.3486 (0.0 %) 4.00 (11.6 %) 0.2659 (23.7 %)
0.02 3.32 (26.7 %) 0.3464 (0.6 %) 3.53 (22.0 %) 0.2656 (23.8 %)
0.05 2.42 (46.5 %) 0.3480 (0.2 %) 2.39 (47.2 %) 0.2692 (22.8 %)
0.1 1.59 (64.9 %) 0.3546 (−1.7 %) 1.59 (64.8 %) 0.2738 (21.5 %)
0.2 0.79 (82.6 %) 0.3645 (−4.6 %) 0.71 (84.3 %) 0.2830 (18.8 %)
0.5 0.13 (97.1 %) 0.3811 (−9.3 %) 0.11 (97.5 %) 0.2977 (14.6 %)
1.0 0.04 (99.1 %) 0.3927 (−12.6 %) 0.01 (99.7 %) 0.3089 (11.4 %)

quadratic customer inconvenience

0.0 3.92 0.1529 5.26 (−34.3 %) 0.0928 (39.3 %)
0.0001 3.89 (0.8 %) 0.1529 (0.0 %) 5.24 (−33.8 %) 0.0945 (38.2 %)
0.0004 3.86 (1.4 %) 0.1533 (−0.3 %) 5.14 (−31.2 %) 0.0946 (38.1 %)
0.0025 3.61 (7.9 %) 0.1549 (−1.3 %) 4.64 (−18.5 %) 0.0937 (38.7 %)
0.01 3.02 (22.9 %) 0.1551 (−1.4 %) 3.22 (17.8 %) 0.0954 (37.6 %)
0.04 1.78 (54.6 %) 0.1576 (−3.1 %) 1.82 (53.5 %) 0.0998 (34.7 %)
0.25 0.28 (92.9 %) 0.1748 (−14.3 %) 0.23 (94.1 %) 0.1136 (25.7 %)
1.0 0.02 (99.4 %) 0.1930 (−26.2 %) 0.00 (100.0 %) 0.1228 (19.7 %)

Note: Listed is the average number of first request reassignments (#ra) per vehicle and the average
objective function value based on the applied penalty costs for vehicle en-route diversion activities.
For both values, the percentage improvement of the respective penalty threshold parameter setting,
presented in brackets, is compared to TIMELIMIT10 in which all vehicle en-route diversion activ-
ities are allowed

hicle en-route diversion activities lead to an improvement of individual system sit-

uations during the execution of the transportation process. However, when applying
small threshold parameter values of up to 0.02 in the linear2X and 0.0025 in the
quadratic case, this significantly reduces the number of performed vehicle en-route
diversion activities thereby reducing driver inconvenience. At the same time, these
small threshold parameter values provide a quality of customer inconvenience that
is similar to the quality attainable when all vehicle en-route diversion activities are
allowed. In contrast, when higher threshold parameter values are used, this results
in a significant increase in customer inconvenience since vehicle en-route diversion
activities by which high improvements of the solution quality could be achieved are
also increasingly forbidden. Since almost all vehicle en-route diversion activities are
removed if the highest penalty costs of 1.0 are applied, this setting produces an in-
crease in the customer inconvenience that is similar to the already considered case in
which vehicle en-route diversion is completely forbidden (see Sect. 8.5). Note that
260 8 Computational Results

0.0
0.01
0.02
serviced requests

0.05
0.1
0.2
0.5
1.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.41 Impacts of assigning penalty costs to vehicle en-route diversion activities on request
response times attained by PROACTIVE10 on S GEN RD = TD = 1.00 using 10 vehicles and the
linear2X customer inconvenience function grouped by intervals of 50 seconds

0.0
0.0001
0.0004
serviced requests

0.0025
0.01
0.04
0.25
1.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. 8.42 Impacts of assigning penalty costs to vehicle en-route diversion activities on request
response times attained by PROACTIVE10 on S GEN RD = TD = 1.00 using 10 vehicles and the
quadratic customer inconvenience function grouped by intervals of 50 seconds

in this case, in which the highest reduction of driver inconvenience generated by

vehicle en-route diversion activities can be attained, PROACTIVE10 still achieves
a significant improvement of the customer inconvenience of 11.4 % or 19.7 % if the
linear2X or the quadratic customer inconvenience function is applied respectively,
compared to TIMELIMIT10 when all vehicle en-route diversion activities are per-
mitted.
From another perspective, when considering TIMELIMIT10 with a threshold
parameter value of 0 as the reference, PROACTIVE10 achieves a lower number of
vehicle en-route diversion activities for the linear2X customer inconvenience start-
ing from values of 0.01 and for the quadratic inconvenience function also start-
ing from values of 0.01. When this threshold parameter value is used, PROAC-
TIVE10 achieves a better solution quality with regard to customer inconvenience of
23.7 % in the linear2X and 37.6 % in the quadratic customer inconvenience case.
Moreover, when applying the same threshold parameter value to TIMELIMIT10 and
PROACTIVE10, the number of vehicle en-route diversion activities becomes com-
8.10 Driver Inconvenience Aspects in the Real-Time Control Approaches 261

parable at 0.05 in the linear2X case and 0.04 in the quadratic case. From these
threshold parameter values up to the highest tested threshold parameter values,
the attained improvements of customer inconvenience by PROACTIVE10 com-
pared to TIMELIMIT10 remain constant at about 21–22 % in the linear2X case
and 35–38 % in the quadratic case. Hence, for these threshold parameter values,
PROACTIVE10 provides significant improvements with regard to the achievable
customer inconvenience while attaining a level of driver inconvenience reduction
which is at least as high as is achieved by TIMELIMIT10.
To sum up, the application of the pro-active real-time control approach PROAC-
TIVE10 is able to significantly reduce customer inconvenience and driver incon-
venience at the same time compared to TIMELIMIT10. With regard to the effects
of setting the introduced threshold parameter on driver inconvenience, which is the
focus of this section, different recommendations of setting this parameter in ac-
cordance to the pursued aim can be derived from the results of this computational
study:
• If the minimization of customer inconvenience is the only aim and driver in-
convenience aspects can be completely neglected, allowing all vehicle en-route
diversion activities is generally advisable. Furthermore, as mentioned above, if
the linear2X customer inconvenience function is used, small applied threshold
parameter values can slightly improve the solution quality of TIMELIMIT10 as
well as of PROACTIVE10. Furthermore, the use of these small threshold pa-
rameter values also significantly reduces driver inconvenience. In this case, using
PROACTIVE10 achieves improvements of up to 23.8 % or 39.3 % when the lin-
ear2X or the quadratic customer inconvenience function is used, respectively.
• If the minimization of customer inconvenience and driver inconvenience are both
important, moderate values for the threshold parameter are recommended. By
utilizing a threshold parameter value of 0.1 if the linear2X customer inconve-
nience function is applied and 0.04 if the quadratic customer inconvenience func-
tion is used, vehicle en-route diversion activities can be reduced by more than
50 % while customer inconvenience is only increased by 1.9 % (linear2X) or
7.1 % (quadratic). This still results in an attainable improvement by PROAC-
TIVE10 of 21.5 % using the linear2X customer inconvenience function and
34.7 % if the quadratic customer inconvenience function is applied compared
to TIMELIMIT10 where all vehicle en-route diversion activities are allowed.
• If the minimization of driver inconvenience is highly prioritized, forbidding all
vehicle en-route diversion activities is attainable by setting the threshold param-
eter to 1.0 or higher. In this case, the customer inconvenience is significantly
increased in TIMELIMIT10 and PROACTIVE10. Nevertheless, also in this case
PROACTIVE10 provides a significantly reduced customer inconvenience com-
pared to TIMELIMIT10. In addition to this, the results achieved by PROAC-
TIVE10 are still better than the ones attained by TIMELIMIT10 in which all vehi-
cle en-route diversion activities are permitted. In particular, PROACTIVE10 still
achieves a significant improvement of customer inconvenience of 11.4 % or
19.7 % if the linear2X or if the quadratic customer inconvenience function is
applied, respectively.
262 8 Computational Results

8.11 Summary and Answers to the Three Central Research

Questions
In this chapter, extensive computational experiments were conducted in order to
evaluate the deterministic and the pro-active real-time control approach which were
proposed for coordinating the transportation process of RDOPG applications. After
describing the discrete event-based simulator utilized to carry out the experiments
and after presenting the tested environment including evaluated parameters, the two
request data classes on which the computational experiments are conducted were
introduced. Specifically, while the first request data class S REAL consists of a real
request data set derived from a real-world application dealing with the subsequent
delivery of newspapers, the second request data class S GEN consists of 25 different
generated request data sets. These generated request data sets possess individually
designed levels of structural diversity which allows a differentiated analysis of im-
provements attainable by the proposed pro-active real-time control approach when
different levels of structural quality exist.
Three preliminary computational studies were performed. In the first of these
studies, the performance of the proposed Tabu Search solution method was evalu-
ated. Specifically, its attained results on various static problem instances were com-
pared with the ones obtained by a lower bound of an appropriate exact solution ap-
proach known in the literature. It was shown that applying the proposed Tabu Search
approach for 10 seconds yields near-optimal results for individual static problem in-
stances which arise during the considered RDOPG transportation processes. In the
second study, the performance of the deterministic real-time control approach pro-
posed in Chap. 5 was evaluated against two other deterministic real-time control
approaches. This study revealed that the deterministic real-time control approach
denoted as TIMELIMIT10, which invokes the Tabu Search solution method for
10 seconds whenever new requests have arrived in the previous anticipation hori-
zon, achieves a performance that is comparable to the original deterministic real-
time control approach denoted as ROLLING20. This result can be explained by
the fact that, as shown in the first evaluation, TIMELIMIT10 already achieves a
near-optimal solution quality within the given time limit so that no further signifi-
cant improvements can be attained by applying the proposed solution method for a
longer time period as performed in ROLLING20. Hence, TIMELIMIT10 was used
as the reference setting in the further computational studies. Due to its high solu-
tion quality, TIMELIMIT10 represents a challenging reference setting for attaining
additional improvements by utilizing the pro-active real-time control approach. In
the third preliminary study, the impacts of allowing vehicle en-route diversion on
the solution quality were evaluated. A comparison of TIMELIMIT10 to its variant
which prohibits vehicle en-route diversion activities showed that, depending on the
fleet size and the utilized customer inconvenience function, it is possible to attain
improvements between 14 % and 25 % while simultaneously reducing travel time
and travel distance when vehicle en-route diversion is allowed. After this study, an
evaluation of the Poisson quality approach in our forecasting methods, which uses a
Goodness-of-Fit test to determine the structural distortion of request data sets, was
8.11 Summary and Answers to the Three Central Research Questions 263

conducted. This evaluation revealed that this approach allows for efficiently identi-
fying request data sets which have a high structural distortion. Moreover, informa-
tion on the stochastic knowledge generated using S REAL and S GEN were provided.
In the main computational experiments, the results from the proposed exact clus-
tering method with regard to the generated stochastic knowledge were presented.
This generated knowledge was then used to analyze the improvements which can be
additionally attained by utilizing the pro-active real-time control approach PROAC-
TIVE10 compared to TIMELIMIT10 on S REAL and S GEN . This analysis was carried
out in order to answer the first two questions posed in Sect. 1.3.1:

Answer: As shown in the computational experiments in Sect. 8.8, utilizing the gen-
erated stochastic knowledge in the proposed pro-active real-time control approach
makes it possible to significantly reduce customer inconvenience and thus signifi-
cantly improve the solution quality. Specifically, on the request data class S REAL ,
depending on the available fleet size and utilized customer inconvenience func-
tion, improvements of up to 7.93 % were achieved by PROACTIVE10 compared
to TIMELIMIT10. However, in problem instances with a limited fleet size of 8 ve-
hicles, no improvements could be attained at all. On the request data class S GEN , in
almost all tested request data sets comprising different levels of structural diversity,
significant improvements were achieved. In the setting with the highest tested levels
of structural diversity represented by RD = TD = 1.00, 19.51 % to 26.72 % of ad-
ditional improvement could be attained under the linear2X customer inconvenience
function depending on the fleet size. In scenarios in which the quadratic customer in-
convenience function was utilized, improvements of between 29.86 % and 47.24 %
were achieved. According to the observed differences regarding attainable improve-
ments on S REAL and S GEN by PROACTIVE10 we can conclude that the existence
of specific characteristics in the request data sets fosters improvements that can be
attained by utilizing pro-active real-time control approaches. In this book, we de-
noted these characteristics, which are discussed in what follows, as the structural
quality of a request data set.

Answer: In Sects. 8.8 and 8.9, it was shown that certain prerequisites need to be
fulfilled in a request data set in order to enable PROACTIVE10 to attain signif-
icant improvements. Hence, when dealing with the coordination of the operative
264 8 Computational Results

transportation processes of RDOPG applications, it is advantageous to know under

which conditions an integration of stochastic knowledge is useful. According to the
computational results, we identified characteristics of real-world scenarios under
which the application of the proposed pro-active real-time control approach is rec-
ommended. Specifically, two aspects that mainly influence an efficient application
of the proposed pro-active real-time control approach are illustrated:
• Structural quality of the available request data: This criterion describes the
quality of the stochastic knowledge that can be generated out of the available
request data set. Based on our results, it can be subdivided into two criteria:
– Structural distortion: As mentioned above, our proposed forecasting methods
assume that request arrivals can be modeled by a set of unknown time-space
Poisson processes whose rate parameters are also unknown but remain constant
from day to day. Hence, the application of the pro-active real-time control ap-
proach on request data sets which do not fulfill this assumption is likely to
result in a poor solution quality. Consequently, the existence of a low structural
distortion in the request data set is a crucial pre-requisite for deriving stochas-
tic knowledge of high value and hence for efficiently applying the proposed
pro-active real-time control approach.
– Structural diversity: As illustrated by the computational results on S REAL and
S GEN , it has been shown that the value of the stochastic knowledge utilized in
the pro-active real-time control approach increases with higher structural di-
versity. Specifically, when higher levels of structural diversity are present in
the available request data set, an integration of stochastic knowledge into the
deterministic real-time control approach becomes increasingly beneficial. This
can be explained by the fact that higher variations in request arrival rates in dif-
ferent regions of the considered service area during the day increase the value
of information about expected future requests. In these cases, more substan-
tial improvements can be attained due to a higher value which the stochastic
knowledge possesses. In contrast, request data sets with a low structural di-
versity limit the improvements that can be additionally attained by utilizing
stochastic knowledge since, in this case, the deterministic real-time control ap-
proach can take advantage of keeping idle vehicles at their last serviced request
locations using the idle vehicle waiting strategy. This is because in request data
sets with a low structural diversity future requests are likely to occur close to
requests that have already arrived.
• Number of available resources: In addition to the structural quality of avail-
able request data sets, the degree of utilization of the available vehicles has also
been proven as a decisive aspect for an efficient application of the proposed pro-
active real-time control approach. Specifically, the available fleet size directly af-
fects the minimum required structural quality of existing request data which leads
to promising results. In case of a limited fleet size, vehicles are very busy with
servicing the pending requests. Hence, it is recommended to only integrate reli-
ably forecasted stochastic knowledge since the integration of uncertain stochastic
knowledge is likely to lead to postponement of pending requests and therefore
to a reduced solution quality. In contrast, vehicles are less busy with servicing
8.11 Summary and Answers to the Three Central Research Questions 265

the pending requests in scenarios where larger fleet sizes are available. Hence,
the minimum required structural quality of a request data set decreases since an
incorrectly forecasted request does not cause significant deteriorations in the ex-
ecution of the transportation process.
We conclude that the structural distortion of a request data set depends on the
applied forecasting approach while the structural diversity represents a general cri-
terion which significantly determines attainable improvements by using pro-active
real-time control approaches. Therefore, with regard to structural distortion, it is
possible to also efficiently derive and utilize stochastic knowledge for additional
request data sets which are classified as having a high structural distortion in our
forecasting approach by extending the proposed forecasting approach. Moreover,
the structural diversity of the request data set mainly determines its structural qual-
ity if a low structural distortion is present. Hence, structural diversity was further
evaluated in this book. Besides the request data class S GEN for which the level of
structural diversity can be defined during the generation of the individual request
data sets, two methods were proposed for analyzing the structural diversity of given
request data sets. First, according to the qualitative assessment of request data sets
performed by the visual analysis of S REAL in Fig. 8.24 and S GEN in Figs. 8.27
and 8.28, we concluded that request data sets with a high structural quality have
high variations in request arrival rates in the subareas during the day which allows
for an efficient application of the proposed pro-active real-time control approach. In
order to further define structural quality, we introduced the degree of structural di-
versity (dosd) for quantitatively measuring the structural diversity of given request
data sets (see Sect. 8.9). The effectiveness of this measure was evaluated by ana-
lyzing the dosd values for all 25 request data sets of the request data class S GEN .
Specifically, linear regression models analyzing the correlation between the dosd
and the improvement attained in the computational experiments were applied to
both customer inconvenience functions and the three tested fleet sizes. The resulting
correlation coefficients were between 0.9161 and 0.9606 indicating that the dosd
is an effective measure for determining the structural quality of given request data
sets. Moreover, an analysis of the dosd values determined for the real-world request
data class S REAL also yielded good results. Therefore, it was concluded that for the
first time the proposed dosd provides a reasonable starting point for analyzing the
structural quality of given request data sets. As mentioned above, this is an essential
factor for estimating the improvements that can be attained by utilizing pro-active
real-time control approaches. Moreover, it provides decision support for efficiently
customizing the proposed pro-active real-time control approach by choosing appro-
priate DCminλ values for generating stochastic knowledge depending on the avail-
able request data set, customer inconvenience function, and fleet size.
In a further computational study, we continued to investigate the impacts of ve-
hicle en-route diversion activities on the solution quality and the resulting driver
inconvenience:

Question 3: What impact does vehicle en-route diversion have on the solution qual-
ity? Since this increases driver inconvenience, is there a conflict between customer
266 8 Computational Results

inconvenience and driver inconvenience and does a useful trade-off between both
objectives exist?

Answer: As shown in Sect. 8.5, allowing vehicle en-route diversion significantly

improves the solution quality since vehicles can quickly react to newly arriving re-
quests which are close to the current vehicles’ positions. However, as illustrated
in Table 8.5, this imposes a high number of request reassignment activities while
drivers are currently traveling to their next assigned customer request. This increases
driver inconvenience. For this reason, a threshold-based approach for limiting vehi-
cle en-route diversion activities was proposed in order to reduce driver inconve-
nience. By using this approach, vehicle en-route diversion activities are limited to
the ones which achieve objective function value improvements which are higher
than the penalty incurred by the specified threshold parameter value. It was shown
that the application of small threshold parameter values significantly reduces the
number of vehicle en-route diversion activities while only marginally increasing the
resulting customer inconvenience. When using higher threshold parameter values, a
further decrease of performed vehicle en-route diversion activities could be achieved
but with a significant increase in the attained customer inconvenience. Furthermore,
the largest threshold parameter value of 1.0 that we evaluated effectively prohibited
almost all vehicle en-route diversion activities which resulted in the lowest achiev-
able driver inconvenience with regard to the considered vehicle en-route diversion
activities.
Furthermore, this computational study showed that customer inconvenience and
driver inconvenience are conflicting aims. Nevertheless, parameter settings for the
proposed threshold-based approach which achieve useful trade-offs between both
conflicting aims were determined. Specifically, compared to scenarios in which only
customer inconvenience is considered, driver inconvenience can be significantly
reduced while only marginally increasing customer inconvenience by using small
threshold parameter values whose actual value depends on the utilized customer
inconvenience function. Moreover, in order to prioritize customer inconvenience,
the threshold parameter value should be set to 0.0 while a prioritization of driver
inconvenience can be achieved by setting the threshold parameter value to 1.0.
The computational results also showed that for request data sets with a high struc-
tural quality, using the threshold-based approach with PROACTIVE10 can reduce
customer inconvenience and driver inconvenience at the same time compared to the
results achieved by TIMELIMIT10. It was concluded that the proposed threshold-
based approach is an effective method in limiting the number of vehicle en-route
diversion activities to a desired level.
Chapter 9
Summary and Outlook on Future Work

In this chapter, the evaluated RDOPG applications and the achieved results of this
book are presented. First, a summary about the work in this book, attained results,
and contributions is given. Finally, an outlook on future work that provides avenues
for future research is given.

9.1 Summary

In many business relationships, the unexpected occurrence of events may lead to

unfavorable situations which cause a high inconvenience or high costs for affected
customers. Two examples of such situations are the unsuccessful delivery of news-
papers to subscribers and the breakdown of important machines during production
processes. Hence, in these situations, customers demand quick assistance from a
service company assumed to be determined in advance in order to resolve the prob-
lematic situation. What is crucial is that since the occurred situation has already
caused a high amount of inconvenience to the customer, a delayed delivery of these
goods causes additional significant inconvenience. Therefore, distribution processes
of such goods are defined as real-time distribution of perishable goods (RDOPG) ap-
plications in this book. The considered RDOPG applications belong to the class of
dynamic routing problems in which uncertainty arises from newly arriving requests
during the execution of the transportation process. Specifically, in the RDOPG ap-
plications, only a few customer requests are known in advance so that a high degree
of dynamism is present.
Since marketing research has revealed that customer satisfaction relates to how a
company deals with a complaint, it is the topmost priority of the considered service
company to fulfill or exceed the expected service quality in order to generate high
customer satisfaction and to establish long-term customer connectivity. Moreover,
since in the considered RDOPG applications customer inconvenience directly starts
at the point in time when a customer places a request at the service company and
since it depends on the perceived request response time, attained service quality can

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 267

be increased by providing a fast service of pending requests. Hence, RDOPG appli-

cations are characterized by a high level of customer orientation so that service qual-
ity is the topmost aim in these applications. Specifically, a late supply of required
goods to the customers causes inconvenience and penalty costs that are significantly
higher than costs arising from the operative distribution processes. Therefore, the
objective solely consists of minimizing the mentioned customer inconvenience.
Motivated by these challenges, two real-time control approaches for efficiently
coordinating the transportation process of RDOPG applications are proposed in this
book. The deterministic real-time control approach which utilizes a newly devel-
oped solution method for solving individual problem instances arising during the
execution of the transportation process is proposed. Based on this deterministic
real-time control approach, a new pro-active real-time control approach is proposed
whose purpose is to pro-actively position vehicles according to expected future re-
quest arrivals so that further arriving requests can be efficiently serviced with re-
duced customer inconvenience. In order to do so, information about expected future
requests referred to as stochastic knowledge is utilized. In the literature it was al-
ready shown that the integration of stochastic knowledge can significantly increase
the solution quality attained in different dynamic routing problems. However, in
these approaches it is often assumed that stochastic knowledge is directly available
which is not the case in many real-world applications. Therefore, the generation of
stochastic knowledge is an important aspect since in many real-world applications
no stochastic knowledge is given a priori but can be derived from available past
request information. Hence, one of the main contributions of this book is to pro-
pose new methods for generating stochastic knowledge by analyzing past request
information and integrating this stochastic knowledge into the real-time control ap-
proach. Since the proposed approach generates the stochastic knowledge from past
request information and only assumes the existence of time-space Poisson processes
that are constant from day to day, it can be applied to many practical RDOPG appli-
cations.
In order to evaluate the attainable improvements of the solution quality by using
the pro-active real-time control approach, its results are compared in detail in an ex-
tensive study to the ones generated by the deterministic real-time control approach.
The impacts on the results are evaluated using different parameter settings for the
real-time control approaches as well under different request data conditions arising
on real-world and on different generated request data sets. Attained results are uti-
lized in order to derive suitable measures for determining the structural quality of
given request data sets. In what follows, we provide a summary of the individual
chapters of this book.
In Chap. 1, a motivation was given by describing challenges found in real-world
processes that are represented by the considered RDOPG applications. Moreover,
the contributions of this book were described. In Chap. 2, a comprehensive in-
troduction to routing problems was presented. Besides defining the general tasks
found in routing problems, general differences between arc- and node-based prob-
lems were described. Since the considered RDOPG applications in the book belong
to the node-based routing problems, several classic node-based routing problems
9.1 Summary 269

such as the Traveling Salesman Problem (TSP), Vehicle Routing Problem (VRP),
and Pickup and Delivery Problem (PDP) were introduced. Furthermore, since the
considered RDOPG applications can be represented as variants of the VRP, in the
remainder of Chap. 2, characteristics and variants of VRPs were discussed. These
characteristics include solution complexity issues and possible solution methods. In
the main part of this chapter, different variants of the VRP were described. Specif-
ically, variants of the static VRP that differ from each other in terms of problem
attributes such as transportation requests, number of depots, available vehicles, as
well as different pursued objective functions were discussed. As mentioned above,
RDOPG applications belong to the class of dynamic routing problems due to newly
arriving requests during the execution of the transportation process. Therefore, after
describing static routing problems in which the relevant information is constant, dy-
namic routing problems in which the relevant information changes were presented.
For this purpose, different classifications in the literature that deal with differences
between static and dynamic routing problems were described. Since dynamic prob-
lems can be distinguished according to the type of relevant information revelation,
appropriate existing classifications were also presented. In order to aggregate the
results that are relevant for the considered RDOPG applications from these classifi-
cations, a unified approach for classifying routing problems was developed. In this
unified approach, the type of changes that occur during the execution of the trans-
portation process in the relevant information as well as their availability and process-
ing and the type of available stochastic knowledge are considered. According to this
classification, different possible variants of VRPs which are known in the literature
were introduced. Afterwards, in accordance with the considered RDOPG applica-
tions, Dynamic Vehicle Routing Problems (DVRPs) in which a centralized coordi-
nation of the tour plan utilizing real-time control approaches is applied were con-
sidered next. Specifically, characteristics of such routing problems were presented.
Besides typical objectives pursued in dynamic routing problems and a description
of dynamic events considered in dynamic routing problems, a measure proposed by
Lund et al. (1996) and Larsen (2000) for determining the difficulty level of solving
individual problem instances in real-time routing problems was introduced. Based
on this measure denoted as the degree of dynamism, a three-echelon classification
proposed by Larsen (2000) was introduced and various real-world dynamic routing
applications were described according to their degree of dynamism. In order to eval-
uate the performance of dynamic routing approaches, two different measures known
in the literature were presented. Finally, a general classification scheme for charac-
terizing the routing problems was described. This scheme was adopted from the
literature and extended in order to cover additional relevant problem characteristics
which are important for RDOPG applications.
In Chap. 3 the considered problem represented by the RDOPG applications was
described in detail. One of the main characteristics that distinguishes RDOPG ap-
plications from many other dynamic routing applications is the objective function
that consists either of a linear or a quadratic customer inconvenience function. This
customer inconvenience is different from penalty costs which only occur when a
270 9 Summary and Outlook on Future Work

request is serviced late since customer inconvenience directly starts with the re-
quest arrival in the system. If a maximum allowed request response time is not re-
garded, high lateness penalty costs occur. Furthermore, vehicle en-route diversion is
allowed. Specifically, allowing to re-route a vehicle while it is traveling to a request
is likely to result in an improved solution quality since this enables quick service
of a newly incoming request which is closely located to the vehicle. Moreover, past
request information is available and can be used to generate stochastic knowledge
about expected future requests which is important in order to apply the pro-active
real-time control approach. After this problem description, various other character-
istics of the RDOPG applications, such as its problem complexity, were described
and the RDOPG applications were classified according to the general classification
scheme which was presented in Chap. 2.
Based on this problem description, an extensive literature review was conducted
in Chap. 4 considering various known approaches for dynamic routing problems
which cover important aspects of RDOPG applications. After describing the first
approaches for dynamic routing problems that have been published in the litera-
ture, selected research work dealing with dynamic routing problems was presented.
Furthermore, general techniques for providing an increased flexibility in dynamic
routing problems were introduced. These techniques were subdivided according to
whether they require stochastic knowledge or not. This distinction was later on used
when presenting the approaches in the literature that provide an increased flexibil-
ity. In the literature review, different approaches including strategic approaches and
reoptimization-based approaches were considered. From this literature review it can
be observed, as mentioned above, that the majority of the approaches presented in
the literature which utilize stochastic knowledge assume that it is directly avail-
able by prescribed probability distributions. In addition to methods for increasing
flexibility in real-time control approaches, the literature review also presents ap-
proaches in the literature which cover other aspects that play an important role in
the considered RDOPG applications. These aspects include objective functions that
are comparable to the one utilized. Also described were real-time control concepts,
road networks for mapping vehicle movements and travel time as well as travel dis-
tance calculations, and appropriate technical equipment for implementing real-time
control approaches.
After this literature review, the real-time control concept utilized in the RDOPG
applications was introduced in Chap. 5. First, a general categorization of differ-
ent control concepts that are often applied in logistic and production systems was
presented. After this categorization, a general challenge of real-time control ap-
proaches, i.e., the concurrency between tour plan execution and tour plan adap-
tation, was described. This is important for many real-time applications since the
controlled process is often executed simultaneous to its adaptation. Since trans-
portation processes of RDOPG applications also belong to these types of real-time
control processes, the anticipation horizon concept (cf. Bock 2010) which is used
for dealing with this challenge was described. Moreover, the update handling used
for integrating dynamically arriving requests and generating individual problem in-
stances as well as the idle vehicle waiting strategy were introduced. By making use
9.1 Summary 271

of this strategy, idle vehicles wait for further arriving requests at their last serviced
request location instead of directly traveling back to the depot. In this context, the
applied vehicle scheduling strategy was presented. Using this strategy, vehicles di-
rectly drive to their next assigned request location since pending requests can always
be directly serviced when they arrive in the system. Furthermore, the mathematical
model used to represent individual problem instances was presented including all
constraints that need to be fulfilled for the considered RDOPG applications. More-
over, both considered objective functions denoted as linear2X and quadratic were
defined. Finally, the introduced RDOPG real-time control concept was categorized
according to the classification at the beginning of the chapter.
The previously described real-time control approach is deterministic since it only
considers requests that have recently arrived during tour plan adaptation but does not
utilize stochastic knowledge about requests expected in the future. In order to utilize
the available stochastic knowledge, new methods were presented in Chap. 6 for ex-
tending the deterministic real-time control approach to the new proposed pro-active
real-time control approach. This pro-active real-time control approach integrates
the generated stochastic knowledge into the tour plan coordination process by using
dummy customers. In contrast to real requests, dummy customers are not serviced
but are used to guide vehicles into request-likely areas where future requests are
expected with a high probability according to the stochastic knowledge. For this
purpose, new methods for generating stochastic knowledge from real past request
information were described. This generation is performed by partitioning the con-
sidered service area into subareas and segments. For each of these segments, request
arrivals are modeled by a time-space Poisson process whose rate parameter is cal-
culated by analyzing the number of request arrivals in the respective segment on a
specific number of considered past working days. Afterwards, dummy customers
are created by aggregating segments to clusters that fulfill certain quality criteria.
Among others, one criterion to be fulfilled consists of requiring that one or more
requests actually arrives in place of a dummy customer at least with a specified
probability. Moreover, as mentioned above, in the utilized methods for generating
stochastic knowledge it is assumed that requests in the individual segments arrive
according to time-space Poisson processes. In order to check this assumption, for
each dummy customer the likelihood that past request arrivals observed in the cor-
responding cluster are generated by a time-space Poisson processes is evaluated.
If this is too unlikely, the dummy customer is discarded. In order to generate as
many valid dummy customers as possible, an exact two-step solution method was
presented. In the first step, all valid dummy customers are generated while in the sec-
ond step the maximum number of non-overlapping dummy customers is selected by
utilizing an MIP solver. Moreover, modifications of the real-time control approach
required for integrating and handling dummy customers during the execution of the
transportation process were presented. This also included the description of an ex-
tended vehicle scheduling strategy. Using this strategy, a vehicle waits at its current
position if its next assigned request is a dummy customer and if its departure at the
current time in the transportation process would otherwise make the vehicle arrive
earlier than desired at the dummy customer’s location.
272 9 Summary and Outlook on Future Work

In order to efficiently solve the individual static problems generated by the real-
time control approach under the tight computational time restrictions imposed by
the real-time nature of the considered RDOPG applications, a sophisticated heuris-
tic solution method was proposed in Chap. 7. This solution method is based on a
Tabu Search metaheuristic that utilizes different neighborhood operators which are
chosen by a deterministic stage-based selection scheme. After explaining the gen-
eral procedure of the Tabu Search metaheuristic and describing differences between
intensifying and diversifying neighborhood operators, the utilized neighborhood op-
erators were described in detail. After this description, the applied tabu-active at-
tribute in the tabu list that is used to identify tabu-active solutions was explained.
Specifically, in order to allow an almost unique identification of individual solutions,
a fingerprint consisting of two quickly computable checksums was used as the tabu-
active attribute. This also allows the identification of solutions across consecutive
problem instances. After a description of the initial tour plan generation process,
the stage-based selection scheme was presented. This scheme selects the neighbor-
hood operator which is to be applied in each iteration of the solution process. After
a predefined number of iterations without finding a new best global solution, the
current stage is switched to the next higher one. Therefore, neighborhood operators
providing a higher level of diversification are applied in the subsequent iterations.
When a new best global solution is found, the stage is reset to the first one so that
an intensifying neighborhood operator is applied in the next iterations.
In Chap. 8, the extensive computational experiments that are performed to eval-
uate the proposed variants of the real-time control approach and to answer the cen-
tral research questions of this book were presented. This includes a description of
the discrete event-based simulator and the tested environment consisting of the uti-
lized road network as well as certain parameters, i.e., the length of the anticipating
horizon, tested fleet sizes, applied Tabu Search parameters, and evaluated dummy
customer parameters for generating and integrating stochastic knowledge into the
pro-active real-time control approach. Furthermore, the two request data classes on
which the evaluations are conducted were introduced. The first request data class
consists of a real request data set derived from a real-world application dealing with
the subsequent delivery of newspapers. The second request data class consists of
different generated request data sets that possess specific designed structural quality
aspects. Therefore, the second request data class allows a differentiated analysis of
the attainable improvements of the proposed pro-active real-time control approach
under different levels of structural diversity present in the designed request data sets
of this request data class.
Before the main computational study, several preliminary studies were per-
formed. First, the performance of the Tabu Search solution method was evaluated by
comparing its attained results with the ones obtained by a lower bound of an exact
solution approach for a related dynamic vehicle routing problem. In this study, it
was shown that the proposed Tabu Search metaheuristic yields near-optimal results
within 10 seconds for individual problem instances that arise during the transporta-
tion process of the considered RDOPG applications. Hence, we concluded that the
9.1 Summary 273

proposed Tabu Search metaheuristic allows a sufficiently fast and efficient adap-
tation of the individual problem instances. In a second preliminary study, the per-
formance of the deterministic real-time control approach was evaluated against two
related deterministic real-time control approaches. The results revealed that the vari-
ant denoted as TIMELIMIT10, which invokes the solution method for 10 seconds
whenever new requests have arrived in the previous anticipation horizon, achieves a
performance that is comparable to the original approach which continuously applies
the solution method for the complete length of the anticipation horizon. An analysis
showed that this can be explained by the fact that TIMELIMIT10 already achieves
a near-optimal solution quality within the given time limit so that no significant fur-
ther improvements can be attained by applying the proposed solution method for a
longer time period. Hence, TIMELIMIT10 was used as the reference setting which
is challenging due to the utilization of the described solution method. The next pre-
liminary study dealt with the impacts of allowing vehicle en-route diversion on the
attained solution quality in the deterministic real-time control approach. The results
showed that, depending on the fleet size and utilized customer inconvenience func-
tion, improvements in the solution quality between 14 % and 25 % can be attained
compared to the solution quality when vehicle en-route diversion activities are pro-
hibited while reducing travel time and travel distance at the same time. In a further
study, the performance of the Poisson quality approach in our forecasting methods,
which uses a Goodness-of-Fit test to determine the structural distortion of request
data sets, was evaluated. This test was used to check the assumption that request
arrivals can be modeled by a set of unknown time-space Poisson processes whose
rate parameters are also unknown but remain constant from day to day. Hence, a low
structural distortion of the existing request data set is important in order to allow for
an efficient application of the proposed pro-active real-time control approach. Evalu-
ations conducted on different request data sets that have different levels of structural
distortion showed that the proposed Poisson quality approach is an efficient method
for determining the structural distortion of request data sets. The evaluations also
indicated that S REAL has, similar to S GEN , a low structural distortion. Moreover, a
suitable value for the minimum demanded Poisson quality parameter for generating
stochastic knowledge was derived. Using this parameter value, information on the
stochastic knowledge generated for the request data sets evaluated our main compu-
tational experiments was given. This also included information about the process of
generating this stochastic knowledge.
In the main computational experiments, first the results of the generated stochas-
tic knowledge by utilizing the proposed forecasting method were presented. There-
after, a solution quality comparison of the pro-active real-time control approach and
its deterministic variant was presented. In these experiments, the achieved results
of evaluating the real-world request data class S REAL were provided. The results
revealed that depending on the used customer inconvenience function as well as
on the fleet size, different dummy customer settings with regard to the minimum
demanded quality DCminλ yielded the best results. Specifically, when few vehicles
are present, the utilization of stochastic knowledge should be reconsidered in gen-
eral. This is especially important if the linear2X customer inconvenience is applied
274 9 Summary and Outlook on Future Work

since in this case, the deterministic real-time control approach tends to postpone
the service of requests in order to service other requests earlier. This often causes
problems when further unexpected requests arrive so that some requests cannot be
serviced within the maximum allowed response time. This situation is even wors-
ened when uncertain stochastic knowledge is used in the pro-active real-time con-
trol approach. Therefore, the utilization of stochastic knowledge only comprising
dummy customers, for which at least one request will actually arrive in place of the
dummy customer with a high probability, is recommendable in this situation. When
more vehicles are present, the utilization of generated stochastic knowledge can at-
tain significant improvements. Moreover, the use of forecasted stochastic knowledge
that also comprises less reliable information can lead to additional improvements of
the solution quality. This can be explained by the fact that due to the larger fleet
size, available vehicles are less utilized so that uncertain information can be ben-
eficially used without disturbing the service of pending requests. With the largest
tested fleet size, improvements of up to 1.7 % in the linear customer inconvenience
case and 7.93 % in the quadratic customer inconvenience case could be attained.
Furthermore, a detailed visual analysis of S REAL revealed that the relative re-
quest arrival rates in all subareas of the service area do not significantly vary during
the day. Therefore, since newly arriving requests keep occurring at similar locations
as before, the deterministic real-time control approach already advantageously po-
sitions the vehicles in advance for servicing future requests by making use of the
presented idle vehicle waiting strategy. Therefore, this results in a low structural di-
versity for S REAL so that additional improvements attainable by utilizing stochastic
knowledge in the pro-active real-time control approach are limited in this request
data set. A further potentially problematic issue given by the fact that it may not be
possible to reasonably represent request arrivals in S REAL by unknown time-space
Poisson processes, which are constant from day to day, could not be identified as a
significant problem. Note that this is in line with the previously attained results of
analyzing the structural distortion of S REAL that indicated a low structural distortion
for this request data set. Hence, the low structural diversity of S REAL seems to be
the major reason for the prohibition of further improvements attainable on S REAL
by using the pro-active real-time control approach.
The second evaluated request data class S GEN was generated using time-space
Poisson processes with individual request arrival rates in different subregions. In
this data, these arrival probabilities as well as the total request arrival rate signif-
icantly change during the day. Note that higher arrival rate variations reduce the
probability that vehicles are closely located to future request arrivals by just waiting
at the location of the last serviced request. Therefore, in this case, an a priori reposi-
tioning of vehicles according to expected future request arrivals based on stochastic
knowledge increasingly becomes advantageous. The mentioned variation in request
arrival rates was defined according to two dimensions that are called RegionDiver-
sity (RD) and TimeDiversity (TD). Utilizing these two dimensions, the structural
diversity of this request data class was defined. Overall, 25 request data sets with
different levels of structural diversity were generated. The computational results
showed that in request data sets with higher levels of structural diversity, signifi-
cantly higher solution quality improvements can be attained by using the pro-active
9.1 Summary 275

real-time control approach compared to the results obtained by the deterministic

real-time control approach. Specifically, while in almost all tested structural diver-
sity settings improvements could be achieved, in the highest structural diversity set-
ting, 19.51 % to 26.72 % of additional improvement could be attained under the
linear2X customer inconvenience function, depending on the fleet size. When the
quadratic customer inconvenience function was utilized, improvements of between
29.86 % and 47.24 % were achieved. As performed on S REAL , a visual analysis of
this request data class was provided. In this analysis, significant structural differ-
ences between the evaluated real-world and a generated request data set comprising
a high level of structural diversity were illustrated. Specifically, while S REAL pos-
sesses a low structural diversity, the structural diversity of the designed request data
sets depends on the RD and TD parameter settings used for generating the cor-
responding request data set. By applying this structural diversity classification on
the evaluated request data sets combined with the above described Poisson quality
approach for determining structural distortion, differences in the attained improve-
ment of the solution quality by using the pro-active real-time control approach in
comparison to its deterministic variant became explainable.
Based on the insights of the main computational study with regard to the level of
structural diversity and the improvement achievable by applying generated stochas-
tic knowledge in the pro-active real-time control approach, we extended the previ-
ously described visual analysis. Specifically, a measure called the degree of struc-
tural diversity (dosd) was introduced for determining the structural quality of given
request data sets. This measure is based on the fact that stochastic knowledge pos-
sesses a higher value if the relative request arrival rates in the individual subareas
of the service area significantly change during the considered planning horizon.
The dosd ranges from 0 to 1 representing the lowest or highest structural diversity
which a given request data set can possess according to this measure. The effective-
ness of this measure was evaluated by analyzing the dosd values generated for all
25 designed request data sets of S GEN with regard to the additional improvements
that have been obtained in the main computational experiments by the pro-active
real-time control approach. In total, six different data sets were analyzed resulting
from the two customer inconvenience functions evaluated under the three different
fleet sizes. A linear regression model was applied to each of the six data sets and the
correlation coefficient was calculated. The resulting correlation coefficients were
between 0.9161 and 0.9647 which statistically proved the significance of the dosd.
Moreover, good results were also obtained for S REAL . Therefore, it was concluded
that the proposed dosd provides a reasonable starting point for generally measuring
the structural quality of a given request data set. As shown above, this is an essential
factor for estimating the improvements attainable by utilizing pro-active real-time
control approaches.
Besides the evaluations dealing solely with the minimization of customer incon-
venience, we investigated another computational study which additionally considers
the number of vehicle en-route diversion activities performed during the transporta-
tion process. It was concluded that this number is an important factor in real-world
applications since drivers need to react to assigned vehicle en-route diversion activ-
ities very quickly. Therefore, a high number of vehicle en-route diversion activities
276 9 Summary and Outlook on Future Work

imposes a high inconvenience on the drivers which may result in reduced work mo-
tivation. In order to reduce driver inconvenience by lowering the number of vehicle
en-route diversion activities, a threshold-based approach for limiting these vehicle
en-route diversion activities was proposed. By using this approach, vehicle en-route
diversion activities are limited to the ones which achieve objective function value
improvements which are higher than the penalty incurred by the threshold parameter
value. In computational experiments performed by using the deterministic real-time
control approach, the impact of different values of the proposed threshold parame-
ter on the solution quality was analyzed. It was shown that the application of small
threshold parameter values significantly reduces the number of vehicle en-route di-
version activities while only marginally increasing the resulting customer inconve-
nience. When using higher threshold parameter values, a further decrease of per-
formed vehicle en-route diversion activities could be observed but with a significant
increase in the attained customer inconvenience. Furthermore, the highest investi-
gated threshold parameter value effectively prohibited all vehicle en-route diversion
activities which resulted in the lowest achievable driver inconvenience with regard
to the number of performed vehicle en-route diversion activities.
Moreover, the utilization of the proposed pro-active real-time control approach
was also recommended when considering driver inconvenience aspects. Specifically,
the computational results showed that the pro-active real-time control approach is
able to reduce customer inconvenience and driver inconvenience at the same time
when appropriate threshold parameter values are used. It was concluded that the
proposed threshold-based approach is an effective method in limiting the number
of vehicle en-route diversion activities to a desired level. For the different pursued
aims, i.e., prioritizing customer inconvenience or driver inconvenience or consider-
ing a balance of both contradicting aims, adequate values for the threshold parameter
approach were provided.
The work presented in this book extends the literature of stochastic and dynamic
vehicle routing problems by a new pro-active real-time control approach and an
approach for generating stochastic knowledge out of past request information. All
approaches have been extensively tested under different conditions. Furthermore,
several evaluations which consider the impacts of the structural request data quality
on the solution quality as well as a method for limiting vehicle en-route diversion
activities have been proposed and impacts on the solution quality have been as-
sessed. The achieved results are likely also to provide additional insights into other
variants of dynamic routing problems which are related to the considered RDOPG
applications.

9.2 Outlook on Future Work

The attained results provide promising starting points for avenues of further research
which are presented in what follows. First, extensions of the proposed methods for
generating stochastic knowledge seem to be useful. For instance, the integration of
9.2 Outlook on Future Work 277

geographical information that is available in the road network could have a signif-
icant impact on the performance of the pro-active approach. Specifically, the per-
formed generation of compact clusters on predefined segments might restrict the
generation of stochastic knowledge and its applicability if natural barriers such as
rivers or channels exist. Therefore, an extension of the segment-based clustering
methods used for generating stochastic knowledge should be examined. Another ex-
tension may improve the performance of the pro-active real-time control approach
if the existing request data does not consist of request arrival characteristics which
are constant from day to day but instead consists of different recurring types of days
in which request arrivals significantly vary and the type of the upcoming day cannot
be determined in advance. For example, this is the case when the requests arrive
on different days according to individual sets of time-space Poisson processes. In
such cases, further improvements are likely to be attained by integrating sophisti-
cated pattern generation and recognition methods into the cluster generation and
the pro-active real-time control approach. Specifically, as an extension to the pro-
posed forecasting approach which assumes one pattern that is constant from day to
day, different day patterns can be learned based on past request information. During
the execution of the transportation process, an appropriate day pattern is identi-
fied according to requests that have already arrived during the current day. After
this identification, adequate stochastic knowledge is dynamically integrated into the
pro-active real-time control approach during the execution of the transportation pro-
cess. As mentioned in Sect. 8.6, an extension of the proposed forecasting methods
requires further developing methods for the purpose of identifying structural distor-
tion. Furthermore, using methods for an ex-post comparison of predicted requests
and those that have occurred during the transportation process can be useful in fine-
tuning the stochastic knowledge.
Moreover, as mentioned in the literature review, other pro-active approaches ex-
ist in the literature for utilizing stochastic knowledge. For example, sampling-based
approaches and approximate dynamic programming approaches have also been suc-
cessfully applied to dynamic routing problems. In this context, ideas inspired from
these types of approaches for generating and utilizing stochastic knowledge can be
integrated into the proposed pro-active real-time control approach. In doing so, an
additional increase in solution quality improvements is likely to be attained. Fur-
thermore, as can be seen in the computational experiments, the impact of different
parameters used in the generation of the stochastic knowledge, for example the im-
pacts of a minimum demanded rate parameter of each cluster on the solution quality
have been evaluated in detail. In addition, there exist other parameters whose impact
on the solution quality could be analyzed in order to improve the methods used for
generating stochastic knowledge. For example, a more detailed analysis of chosen
parameter values such as the minimum demanded Poisson quality or the number of
working days nf used could be investigated. Moreover, the development of methods
which are suitable for assessing whether practical request data sets fulfill require-
ment (c) of Poisson processes (see Sect. 6.2.1) could be researched.
Furthermore, the integration of other sources of dynamism into the pro-active
real-time control approach can increase its practicability. For example, this includes
278 9 Summary and Outlook on Future Work

the consideration of traffic jams, time-dependent travel times, or vehicle break-

downs. Moreover, as described in Sect. 8.1, in the computational experiments all
vehicle activities are performed in line with the anticipated activities. Hence, the
real-time control approach as well as the simulator could be extended in order to
deal with uncertainties in anticipated and actually performed vehicle activities. For
this purpose, the development of methods which synchronize the anticipated tour
plan with the current system situation at the end of each anticipation horizon is
advantageous. Furthermore, as mentioned in the considered RDOPG application
dealing with the subsequent delivery of newspapers, the service time of requests
is assumed to be known and constant for all requests (see Sect. 3.1). By analyzing
past request information according to service times required for servicing requests
and by including this additional stochastic knowledge into the real-time control ap-
proach, the proposed approaches can be also be applied to other types of RDOPG
applications. This is especially useful if the request data has specific characteris-
tics, e.g., if requests located in the same subregion of the service area have a ho-
mogeneous service time which significantly varies between subregions or within a
subregion over different time periods during the considered planning horizon. Other
scenarios that can take advantage of this extension are those in which different types
of requests for which estimations about expected service times can be individually
derived exist.
The previously mentioned integration of time-dependent travel times into the
real-time control approach is also an important aspect in the considered urban trans-
portation system environment. Adequate methods for generating and integrating
suitable stochastic knowledge for these time-dependent travel times could be de-
veloped, e.g., for predicting required travel times according to weekday-dependent
speed profiles observed in the past. This information could be directly considered in
the real-time control approach for determining the fastest routes that vehicles should
travel between requests depending on the current time of the day. The relevance of
such a system could be evaluated by comparing its attainable results to the results
of the current version of the real-time control approach that does not make use of
time-dependent travel time information.
Moreover, the proposed pro-active real-time control approach can be extended
to other types of dynamic routing problems, such as Dynamic Pickup and Delivery
Problems (DPDP). In doing so, other real-world applications which also consider
customer inconvenience and service quality aspects, such as urban express courier
services, could be considered. Due to differences in the problem structure, each
request consists of two request locations instead of one as is the case in the cur-
rently considered RDOPG applications. Therefore, the methods used for generating
stochastic knowledge are required to be extended according to different aspects.
For example, the expected pickup-delivery pairs need to be identified in order to
anticipate future requests. This information is also required in order to calculate
the corresponding expected total request service time which results from the sum of
service times at both request locations and the expected travel time required for trav-
eling between both locations. Since in other problem variants such as the DPDP the
neighborhoods of the applied neighborhood operators may be more complex and
9.2 Outlook on Future Work 279

hence more time-consuming to evaluate, the integration of further techniques for

time-efficiently calculating neighborhood move values as proposed by Badeau et al.
(1997) can be advantageous in order to still be able to quickly generate solutions
of high quality. Moreover, other variants of VRPs mentioned in Sect. 2.4 provide
additional ideas for extending the considered problem.
Furthermore, the considered objective function can be extended. Specifically, be-
sides only considering customer inconvenience, one extension has been proposed
that evaluates the possibility of reducing driver inconvenience represented by the
number of performed vehicle en-route diversion activities and the impact on cus-
tomer inconvenience. Other objectives, for example the minimization of travel time
or travel distance, could also be integrated. Computational experiments which eval-
uate the impact of also considering other objectives on the customer inconvenience
could lead to new insights. Additionally, apart from the evaluated linear2X and
quadratic objective function for modeling the evolution of customer inconvenience,
other functions could also be investigated.
In the computational experiments, a first approach for calculating the structural
diversity for given request data sets referred to as the degree of structural diver-
sity (dosd) was presented. Since the existence of a high structural quality has been
identified as a generally important characteristic for pro-active real-time control ap-
proaches which pursue the aim of improving the system performance by utilizing
stochastic knowledge, it could be interesting to develop this measure further. Specif-
ically, the integration and extension of the Poisson quality approach proposed for de-
termining the structural distortion into the dosd could be investigated. Furthermore,
the impact of different other performance characteristics on the attainable solution
quality, such as the request to vehicle ratio or road network characteristics present
in a scenario, or the ratio of generated dummy customers to the total number of cus-
tomer requests could be evaluated and also integrated into this measure. In doing so,
the dosd can be extended in ways which make it possible to enable a comparison
even of request data sets or scenarios with significantly different characteristics.
As mentioned by Stenfors et al. (2007), “Tools which support creativity are
needed”. As a consequence, a further enhancement of the proposed approaches to-
wards more interactive real-time control approaches could be advantageous in in-
creasing the acceptance and practicability of a decision support system which uti-
lizes the developed forecasting methods and the pro-active real-time control ap-
proach in practice. This especially includes interactive control possibilities in the
decision support system for the dispatcher. By enabling such control possibilities,
transportation process situations which require individual decisions that cannot be
easily included in the real-time control approach in advance could be manually han-
dled by the dispatcher. In doing so, for example, the possibility that a customer
may want to be serviced by a specific service worker could be considered. Further-
more, decision support of the approach could be improved by directly integrating
the dispatcher’s knowledge into the real-time control of the transportation process.
Specifically, the proposed approach could be extended so that the dispatcher can ad-
ditionally provide stochastic knowledge to the system or review stochastic knowl-
edge which has been automatically generated by the proposed approach. Moreover,
280 9 Summary and Outlook on Future Work

instead of choosing a tour plan directly, the real-time control approach could pro-
vide several alternative tour plans to the dispatcher who selects a desired tour plan
and may also apply further modifications to it. Another possibility of increasing
the interactivity can be realized by integrating additional actions into the on-board
units of the vehicles so that service workers are able to inform the central dispatch-
ing center about the progress of request activities. For instance, a service worker
can inform the dispatching center when the current customer request requires more
time for service than planned. Furthermore, a new estimated completion time of the
current service activity could be provided.
Moreover, the validation of attainable solution quality improvements by applying
the proposed pro-active real-time control in real-world applications is another field
of research which might be very interesting for theory and practice.
Appendix A
Additional Facts Related to Chap. 1

90%
Market share (% ton-kilometers)

80%

70%

60%

50%

40%

30%
Domestic trucks
20%
Foreign trucks
10%
1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

Year

Fig. A.1 Market share of domestic and foreign trucks in Germany for domestic and cross-border
transportation from 1991 to 2010 (cf. BGL e.V. 2011)

Table A.1 Development of transportation performance in billion ton-kilometers in the EU-27

from 1995 to 2008 (cf. European Commission 2010, p. 108)
Year Road Sea Rail Inland waterways Pipelines Air Total

1995 1,289 1,146 386 122 115 2.0 3,060

1996 1,303 1,160 392 120 119 2.1 3,095
1997 1,352 1,193 410 128 118 2.2 3,202
1998 1,414 1,232 393 131 125 2.3 3,297
1999 1,470 1,268 384 129 124 2.3 3,376
2000 1,519 1,314 404 134 127 2.5 3,499
2001 1,556 1,334 386 133 133 2.5 3,544
2002 1,606 1,355 384 133 128 2.4 3,608
2003 1,625 1,378 392 124 130 2.4 3,651
2004 1,747 1,427 416 137 132 2.5 3,861
2005 1,800 1,461 414 139 136 2.6 3,952
2006 1,854 1,505 440 138 135 2.7 4,076
2007 1,915 1,532 453 147 127 2.8 4,176
2008 1,878 1,498 443 145 124 2.7 4,090

Increase 1995–2008 45.7 % 30.7 % 14.7 % 19.0 % 8.0 % 35.0 % 33.7 %

Note: Air and Sea: only domestic and intra-EU-27 transport; provisional estimates. Road: national
and international haulage by vehicles registered in the EU-27

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 281

Table A.2 Newspaper advertising expenditures in the United States of America from 1950 to 2010
(cf. Newspaper Association of America 2011)
Year Newspaper print total Newspaper online total Combined newspaper
print and online total
$Millions %Change $Millions %Change $Millions %Change

1950 $2,070 8.3 %

1951 $2,251 8.7 %
1952 $2,464 9.5 %
1953 $2,632 6.8 %
1954 $2,685 2.0 %
1955 $3,077 14.6 %
1956 $3,223 4.7 %
1957 $3,268 1.4 %
1958 $3,176 −2.8 %
1959 $3,526 11.0 %
1960 $3,681 4.4 %
1961 $3,601 −2.2 %
1962 $3,659 1.6 %
1963 $3,780 3.3 %
1964 $4,120 9.0 %
1965 $4,426 7.4 %
1966 $4,865 9.9 %
1967 $4,910 0.9 %
1968 $5,232 6.6 %
1969 $5,714 9.2 %
1970 $5,704 −0.2 %
1971 $6,167 8.1 %
1972 $6,939 12.5 %
1973 $7,481 7.8 %
1974 $7,842 4.8 %
1975 $8,234 5.0 %
1976 $9,618 16.8 %
1977 $10,751 11.8 %
1978 $12,213 13.6 %
1979 $13,863 13.5 %
1980 $14,794 6.7 %
1981 $16,527 11.7 %
1982 $17,694 7.1 %
1983 $20,581 16.3 %
1984 $23,522 14.3 %
1985 $25,170 7.0 %
A Additional Facts Related to Chap. 1 283

Table A.2 (Continued)

Year Newspaper print total Newspaper online total Combined newspaper
print and online total
$Millions %Change $Millions %Change $Millions %Change

1986 $26,990 7.2 %

1987 $29,412 9.0 %
1988 $31,197 6.1 %
1989 $32,368 3.8 %
1990 $32,280 −0.3 %
1991 $30,349 −6.0 %
1992 $30,639 1.0 %
1993 $31,869 4.0 %
1994 $34,109 7.0 %
1995 $36,092 5.8 %
1996 $38,075 5.5 %
1997 $41,330 8.5 %
1998 $43,925 6.3 %
1999 $46,289 5.4 %
2000 $48,670 5.1 %
2001 $44,305 −9.0 %
2002 $44,102 −0.5 %
2003 $44,939 1.9 % $1,216 $46,156
2004 $46,703 3.9 % $1,541 26.7 % $48,244 4.5 %
2005 $47,408 1.5 % $2,027 31.5 % $49,435 2.5 %
2006 $46,611 −1.7 % $2,664 31.4 % $49,275 −0.3 %
2007 $42,209 −9.4 % $3,166 18.8 % $45,375 −7.9 %
2008 $34,740 −17.7 % $3,109 −1.8 % $37,848 −16.6 %
2009 $24,821 −28.6 % $2,743 −11.8 % $27,564 −27.2 %
2010 $22,795 −8.2 % $3,042 10.9 % $25,838 −6.3 %

Note: Numbers are rounded

Appendix B
Additional Information on Request Data Sets
of Request Data Class S GEN

B.1 Parameters for Generating Selected Request Data Sets

of S GEN

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 285

DOI 10.1007/978-3-642-33472-6, © Springer-Verlag Berlin Heidelberg 2013
Table B.1 Resulting request arrival probabilities of selected request data sets in S GEN with different levels of structural diversity
286

j Request arrival probabilities p (ai , tj ) per time slice in each segment ai , i ∈ {1, . . . , 18} λ (tj )

RD = TD = 1.00
1 0.01 0.01 0.01 0.16 0.20 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.20 0.01 0.30 0.01 0.01 0.013 15.00
2 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.30 0.40 0.10 0.01 0.01 0.06 0.01 0.01 0.01 0.01 0.01 40.00
3 0.01 0.01 0.26 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.20 0.10 0.30 0.01 30.00
4 0.01 0.01 0.01 0.01 0.20 0.30 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.35 0.01 0.01 0.01 35.00
5 0.01 0.01 0.10 0.01 0.01 0.01 0.20 0.01 0.01 0.07 0.01 0.01 0.30 0.01 0.20 0.01 0.01 0.01 10.00

RD = TD = 0.75
1 0.0214 0.0214 0.0341 0.1114 0.1425 0.0323 0.0285 0.0323 0.0360 0.0270 0.0214 0.0214 0.1481 0.0214 0.2224 0.0248 0.0323 0.0214 20.8088
2 0.0214 0.0214 0.0341 0.0270 0.0356 0.0323 0.0285 0.1954 0.2554 0.0776 0.0214 0.0214 0.0694 0.0214 0.0593 0.0248 0.0323 0.0214 37.6471
3 0.0214 0.0214 0.1748 0.0270 0.0356 0.0323 0.0285 0.0323 0.0360 0.0270 0.0214 0.0214 0.0413 0.0214 0.1661 0.0754 0.1954 0.0214 26.3235
4 0.0214 0.0214 0.0341 0.0270 0.1425 0.1954 0.0285 0.0323 0.0360 0.0270 0.0214 0.0214 0.0413 0.0214 0.2505 0.0248 0.0323 0.0214 33.8971
5 0.0214 0.0214 0.0848 0.0270 0.0356 0.0323 0.1354 0.0323 0.0360 0.0608 0.0214 0.0214 0.2044 0.0214 0.1661 0.0248 0.0323 0.0214 11.3235

RD = TD = 0.50
1 0.0328 0.0328 0.0498 0.0778 0.0993 0.0473 0.0423 0.0473 0.0523 0.0403 0.0328 0.0328 0.1068 0.0328 0.1558 0.0373 0.0473 0.0328 26.6176
2 0.0328 0.0328 0.0498 0.0403 0.0518 0.0473 0.0423 0.1198 0.1498 0.0628 0.0328 0.0328 0.0718 0.0328 0.0833 0.0373 0.0473 0.0328 35.2941
3 0.0328 0.0328 0.1123 0.0403 0.0518 0.0473 0.0423 0.0473 0.0523 0.0403 0.0328 0.0328 0.0593 0.0328 0.1308 0.0598 0.1198 0.0328 22.6471
4 0.0328 0.0328 0.0498 0.0403 0.0993 0.1198 0.0423 0.0473 0.0523 0.0403 0.0328 0.0328 0.0593 0.0328 0.1683 0.0373 0.0473 0.0328 32.7941
5 0.0328 0.0328 0.0723 0.0403 0.0518 0.0473 0.0898 0.0473 0.0523 0.0553 0.0328 0.0328 0.1318 0.0328 0.1308 0.0373 0.0473 0.0328 12.6471
RD = TD = 0.25
1 0.0442 0.0442 0.0569 0.0592 0.0703 0.0550 0.0513 0.0550 0.0588 0.0498 0.0442 0.0442 0.0759 0.0442 0.1002 0.0475 0.0550 0.0442 32.4265
2 0.0442 0.0442 0.0569 0.0498 0.0584 0.0550 0.0513 0.0732 0.0832 0.0554 0.0442 0.0442 0.0672 0.0442 0.0820 0.0475 0.0550 0.0442 32.9412
3 0.0442 0.0442 0.0725 0.0498 0.0584 0.0550 0.0513 0.0550 0.0588 0.0498 0.0442 0.0442 0.0640 0.0442 0.0939 0.0532 0.0732 0.0442 18.9706
B Additional Information on Request Data Sets of Request Data Class S GEN
Table B.1 (Continued)
j Request arrival probabilities p (ai , tj ) per time slice in each segment ai , i ∈ {1, . . . , 18} λ (tj )

4 0.0442 0.0442 0.0569 0.0498 0.0703 0.0732 0.0513 0.0550 0.0588 0.0498 0.0442 0.0442 0.0640 0.0442 0.1033 0.0475 0.0550 0.0442 31.6912
5 0.0442 0.0442 0.0625 0.0498 0.0584 0.0550 0.0632 0.0550 0.0588 0.0535 0.0442 0.0442 0.0822 0.0442 0.0939 0.0475 0.0550 0.0442 13.9706
RD = TD = 0.00
1 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 38.2353
2 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 30.5882
3 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 15.2941
4 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 30.5882
5 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 15.2941

RD = 1.00, TD = 0.00
1 0.01 0.01 0.078 0.04 0.086 0.068 0.048 0.068 0.088 0.04 0.01 0.01 0.116 0.01 0.212 0.028 0.068 0.01 38.2353
2 0.01 0.01 0.078 0.04 0.086 0.068 0.048 0.068 0.088 0.04 0.01 0.01 0.116 0.01 0.212 0.028 0.068 0.01 30.5882
3 0.01 0.01 0.078 0.04 0.086 0.068 0.048 0.068 0.088 0.04 0.01 0.01 0.116 0.01 0.212 0.028 0.068 0.01 15.2941
4 0.01 0.01 0.078 0.04 0.086 0.068 0.048 0.068 0.088 0.04 0.01 0.01 0.116 0.01 0.212 0.028 0.068 0.01 30.5882
5 0.01 0.01 0.078 0.04 0.086 0.068 0.048 0.068 0.088 0.04 0.01 0.01 0.116 0.01 0.212 0.028 0.068 0.01 15.2941
B.1 Parameters for Generating Selected Request Data Sets of S GEN

RD = 0.00, TD = 1.00
1 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 15.00
2 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 40.00
3 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 30.00
4 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 35.00
5 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 0.0556 10.00

Note: Numbers are rounded

287
288 B Additional Information on Request Data Sets of Request Data Class S GEN

B.2 Visualization of All Request Data Sets of S GEN

y y y y y

e e e e e
m m m m m
Ti Ti Ti Ti Ti
x x x x x
RD = 0.00, TD = 0.00 RD = 0.00, TD = 0.25 RD = 0.00, TD = 0.50 RD = 0.00, TD = 0.75 RD = 0.00, TD = 1.00
y y y y y

e e e e e
m m m m m
Ti Ti Ti Ti Ti
x x x x x
RD = 0.25, TD = 0.00 RD = 0.25, TD = 0.25 RD = 0.25, TD = 0.50 RD = 0.25, TD = 0.75 RD = 0.25, TD = 1.00
y y y y y

e e e e e
m m m m m
Ti Ti Ti Ti Ti
x x x x x
RD = 0.50, TD = 0.00 RD = 0.50, TD = 0.25 RD = 0.50, TD = 0.50 RD = 0.50, TD = 0.75 RD = 0.50, TD = 1.00
y y y y y

e e e e e
m m m m m
Ti Ti Ti Ti Ti
x x x x x
RD = 0.75, TD = 0.00 RD = 0.75, TD = 0.25 RD = 0.75, TD = 0.50 RD = 0.75, TD = 0.75 RD = 0.75, TD = 1.00
y y y y y

e e e e e
m m m m m
Ti Ti Ti Ti Ti
x x x x x
RD = 1.00, TD = 0.00 RD = 1.00, TD = 0.25 RD = 1.00, TD = 0.50 RD = 1.00, TD = 0.75 RD = 1.00, TD = 1.00

Fig. B.1 All request data sets of S GEN with different levels of structural diversity
Appendix C
Request Response Times of the Examined
Request Data Classes

C.1 Request Response Times of S REAL

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
PROACTIVE10, DC minλ = 1.2
serviced requests

PROACTIVE10, DC minλ = 1.5

PROACTIVE10, DC minλ = 1.8
PROACTIVE10, DC minλ = 2.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.1 Request response times on S REAL using 8 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8, and
2.0, linear2X customer inconvenience grouped by intervals of 50 seconds

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
PROACTIVE10, DC minλ = 1.2
serviced requests

PROACTIVE10, DC minλ = 1.5

PROACTIVE10, DC minλ = 1.8
PROACTIVE10, DC minλ = 2.0

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.6 Request response times on S REAL using 12 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8, and
2.0, quadratic customer inconvenience grouped by intervals of 50 seconds
C.2 Request Response Times of S GEN 291

C.2 Request Response Times of S GEN

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
serviced requests

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.7 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8,
and 2.0, linear2X customer inconvenience, RD = 0.00 and TD = 0.00 grouped by intervals of
50 seconds

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
serviced requests

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.8 Request response times on S GEN using 8 vehicles and DC = 1.0, 1.2, 1.5, 1.8,
min λ

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.11 Request response times on S GEN using 8 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8,
and 2.0, linear2X customer inconvenience, RD = 0.50 and TD = 0.50 grouped by intervals of
50 seconds

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
PROACTIVE10, DC minλ = 1.2
serviced requests

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.17 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8,
and 2.0, linear2X customer inconvenience, RD = 0.00 and TD = 0.00 grouped by intervals of
50 seconds

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
serviced requests

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.18 Request response times on S GEN using 10 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8,
and 2.0, quadratic customer inconvenience, RD = 0.00 and TD = 0.00 grouped by intervals of
50 seconds
C.2 Request Response Times of S GEN 295

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
PROACTIVE10, DC minλ = 1.2
serviced requests

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.27 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8,
and 2.0, linear2X customer inconvenience, RD = 0.00 and TD = 0.00 grouped by intervals of
50 seconds
298 C Request Response Times of the Examined Request Data Classes

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
serviced requests

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Fig. C.28 Request response times on S GEN using 12 vehicles and DCmin λ = 1.0, 1.2, 1.5, 1.8,
and 2.0, quadratic customer inconvenience, RD = 0.00 and TD = 0.00 grouped by intervals of
50 seconds

TIMELIMIT10
PROACTIVE10, DC minλ = 1.0
PROACTIVE10, DC minλ = 1.2
serviced requests

response time [s]

500 1000 1500 2000 2500 3000 3500 4000

Francesco Ferrucci was born in 1980 in Herford, Germany. He studied Business

Computing at the University of Paderborn and received his Diploma Degree (with
distinction) in 2006. During his time as a PhD student, he wrote different publi-
cations and visited the CIRRELT in Montreal for three months where he jointly
worked with Professor Michel Gendreau. He received his PhD (summa cum laude)
under the guidance of Professor Stefan Bock in 2012. His research interests include
dynamic routing problems, especially real-time control approaches and forecasting
approaches.

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 301

Aberle, G. (2009). Transportwirtschaft: Einzelwirtschaftliche und gesamtwirtschaftliche Grund-

lagen (Wolls Lehr- und Handbücher der Wirtschafts- und Sozialwissenschaften) (5th ed.).
München: Oldenbourg Wissenschaftsverlag. ISBN 3486579517.
Ahuja, R. K., Magnanti, T. L., & Orlin, J. B. (1993). Network flows: theory, algorithms, and appli-
cations. Englewood Cliffs: Prentice-Hall.
Angelelli, E., Savelsbergh, M. W. P., & Grazia Speranza, M. (2007). Competitive analysis of a dis-
patch policy for a dynamic multi-period routing problem. Operations Research Letters, 35(6),
713–721.
Ascheuer, N., Krumke, S. O., & Rambau, J. (2000). Online dial-a-ride problems: minimizing the
completion time. In H. Reichel & S. Tison (Eds.), Lecture notes in computer science: Vol. 1770.
STACS 2000 (pp. 639–650). Berlin: Springer.
Attanasio, A., Bregman, J., Ghiani, G., & Manni, E. (2007). Real-time fleet management at
Ecourier Ltd. In V. Zeimpekis, G. M. Giaglis, I. Minis & C. D. Tarantilis (Eds.), Operations
research/computer science interfaces series: Vol. 38. Dynamic fleet management (pp. 219–238).
Boston: Springer. ISBN 978-0-387-71722-7.
Badeau, P., Guertin, F., Gendreau, M., Potvin, J.-Y., & Taillard, É. D. (1997). A parallel tabu search
heuristic for the vehicle routing problem with time windows. Transportation Research. Part C,
Emerging Technologies, 5(2), 109–122.
Baldacci, R., Toth, P., & Vigo, D. (2007). Recent advances in vehicle routing exact algorithms.
4OR: A Quarterly Journal of Operations Research, 5(4), 269–298.
Baldacci, R., Bartolini, E., & Laporte, G. (2010). Some applications of the generalized vehicle
routing problem. Journal of the Operational Research Society, 61(7), 1072–1077.
Barceló, J., Grzybowska, H., & Pardo, S. (2007). Vehicle routing and scheduling models, simu-
lation and city logistics. In V. Zeimpekis, G. M. Giaglis, I. Minis & C. D. Tarantilis (Eds.),
Operations research/computer science interfaces series: Vol. 38. Dynamic fleet management
(pp. 163–195). Boston: Springer. ISBN 978-0-387-71722-7.
Barnhart, C., Johnson, E. L., Nemhauser, G. L., Savelsbergh, M. W. P., & Vance, P. H. (1998).
Branch-and-price: column generation for solving huge integer programs. Operations Research,
46(3), 316–329.
Bauer, R., Delling, D., Sanders, P., Schieferdecker, D., Schultes, D., & Wagner, D. (2008). Combin-
ing hierarchical and goal-directed speed-up techniques for Dijkstra’s algorithm. In C. McGeoch
(Ed.), Lecture notes in computer science: Vol. 5038. Experimental algorithms (pp. 303–318).
Berlin: Springer.
Bektas, T. (2006). The multiple traveling salesman problem: an overview of formulations and
solution procedures. Omega, 34(3), 209–219.
Bell, W. J., Dalberto, L. M., Fisher, M. L., Greenfield, A. J., Jaikumar, R., Kedia, P., Mack, R. G.,
& Prutzman, P. J. (1983). Improving the distribution of industrial gases with an on-line comput-

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 303

erized routing and scheduling optimizer. Interfaces, 13(6), 4–23.

Bellman, R. (1954). The theory of dynamic programming. Bulletin, New Series, of the American
Mathematical Society, 60(6), 503–515.
Bellman, R. E. (2003). Dover books on mathematics. Dynamic programming. Mineola: Dover.
ISBN 9780486428093.
Bent, R., & van Hentenryck, P. (2003). Dynamic vehicle routing with stochastic requests. In Pro-
ceedings of the eighteenth international joint conference on artificial intelligence (IJCAI-2003),
Acapulco, Mexico (pp. 1362–1363).
Bent, R., & van Hentenryck, P. (2004a). A two-stage hybrid local search for the vehicle routing
problem with time windows. Transportation Science, 38(4), 515–530.
Bent, R., & van Hentenryck, P. (2004b). Scenario-based planning for partially dynamic vehicle
routing with stochastic customers. Operations Research, 52(6), 977–987.
Bent, R., & van Hentenryck, P. (2004c). Regrets only! online stochastic optimization under time
constraints. In Proceedings of the nineteenth national conference on artificial intelligence
(AAAI-2004), San Jose, California (pp. 501–506).
Bent, R., & van Hentenryck, P. (2005). Online stochastic optimization without distributions. In
Proceedings of the fifteenth international conference on automated planning and scheduling
(ICAPS-2005), Monterey, California (pp. 171–180).
Bent, R., & van Hentenryck, P. (2007). Waiting and relocation strategies in online stochastic vehicle
routing. In Proceedings of the twentieth international joint conference on artificial intelligence
(IJCAI 2007), Hyderabad, India (pp. 1816–1821).
Bent, R., Katriel, I., & van Hentenryck, P. (2005). Sub-optimality approximations. In P. van Beek
(Ed.), Lecture notes in computer science: Vol. 3709. Principles and practice of constraint
programming—CP 2005 (pp. 122–136). Berlin: Springer.
Bent, R., van Hentenryck, P., & Upfal, E. (2010). Online stochastic optimization under time con-
straints. Annals of Operations Research, 177(1), 151–183.
Berbeglia, G., Cordeau, J. F., & Laporte, G. (2010). Dynamic pickup and delivery problems. Eu-
ropean Journal of Operational Research, 202(1), 8–15.
Bertsimas, D. J., & van Ryzin, G. V. (1991). A stochastic and dynamic vehicle routing problem in
the Euclidean plane. Operations Research, 39(4), 601–615.
Bertsimas, D. J., & van Ryzin, G. V. (1993a). Stochastic and dynamic vehicle routing in the Eu-
clidean plane with multiple capacitated vehicles. Operations Research, 41(1), 60–76.
Bertsimas, D. J., & van Ryzin, G. V. (1993b). Stochastic and dynamic vehicle routing with general
demand and interarrival time distributions. Advances in Applied Probability, 25(4), 947–978.
BGL e.V. (2011). Marktanteile deutscher und ausländischer Lkw in Deutschland. http://www.
bgl-ev.de/web/daten/grenz_lkwind.htm. Accessed on 2011-04-25.
Bianchi, L. (2000). Notes on dynamic vehicle routing—the state of the art (Technical report IDSIA-
05-01). Manno-Lugano, Switzerland.
Bieding, T., Görtz, S., & Klose, A. (2009). On-line routing per mobile phone: a case on subsequent
deliveries of newspapers. In J. A. E. E. van Nunen, M. G. Speranza & L. Bertazzi (Eds.), Lecture
notes in economics and mathematical systems: Vol. 619. Innovations in distribution logistics
(pp. 29–51). Berlin: Springer. ISBN 978-3-540-92944-4.
Biethahn, J., Hummeltenberg, W., Schmidt, B., Stähly, P., & Witte, T. (1999). Simulation als be-
triebliche Entscheidungshilfe: State of the Art und neuere Entwicklungen. Heidelberg: Physica-
Verlag. ISBN 3790811785.
Bloos, M., Schönberger, J., & Kopfer, H. (2009). Supporting cooperative demand fulfillment in
supply networks using autonomous control and multi-agent-systems. In S. Fischer, E. Maehle,
& R. Reischuk (Eds.), Informatik 2009: Im Focus das Leben. Proceedings of the 39th annual
conference of the Gesellschaft für Informatik e.V. (GI), 28.9.–2.10.2009, Lübeck (pp. 3590–
3604). ISBN 978-3-88579-248-2.
BMW Group (2011). Annual Report 2010. http://annual-report.bmwgroup.com/2010/gb/files/pdf/
en/BMW_Group_AR2010.pdf. Accessed on 2011-05-12.
Bock, S. (2004). Echtzeitfähige Steuerung von Speditionsnetzwerken: Nutzung moderner
Informations- und Kommunikationstechnologien zur effizienten Durchführung von Transporten:
References 305

Habilitationsschrift Universität Paderborn, 2004 (1st ed.). Wiesbaden: Deutscher Universitäts-

Verlag. Gabler Edition Wissenschaft: Schriften zur quantitativen Betriebswirtschaftlehre. ISBN
3-8244-8214-2.
Bock, S. (2010). Real-time control of freight forwarder transportation networks by integrating
multimodal transport chains. European Journal of Operational Research, 200(3), 733–746.
Bodin, L., & Golden, B. (1981). Classification in vehicle routing and scheduling. Networks, 11(2),
97–108.
Borndörfer, R., Grötschel, M., Klostermeier, F., & Küttner, C. (1997). Telebus Berlin: vehicle
scheduling in a dial-a-ride system (Technical Report SC 97-23). Konrad-Zuse-Zentrum für In-
formationstechnik Berlin.
Brady, M. K., & Cronin Jr., J. J. (2001). Some new thoughts on conceptualizing perceived service
quality: a hierarchical approach. Journal of Marketing, 65(3), 34–49.
Branchini, R. M., Vinícius, A. A., & Løkketangen, A. (2009). Adaptive granular local search
heuristic for a dynamic vehicle routing problem. Computers & Operations Research, 36(11),
2955–2968.
Brandão, J., & Mercer, A. (1997). A tabu search algorithm for the multi-trip vehicle routing and
scheduling problem. European Journal of Operational Research, 100(1), 180–191.
Branke, J., Middendorf, M., Noeth, G., & Dessouky, M. (2005). Waiting strategies for dynamic
vehicle routing. Transportation Science, 39(3), 298–312.
Bräysy, O., & Gendreau, M. (2005a). Vehicle routing problem with time windows, Part I: route
construction and local search algorithms. Transportation Science, 39(1), 104–118.
Bräysy, O., & Gendreau, M. (2005b). Vehicle routing problem with time windows, Part II: meta-
heuristics. Transportation Science, 39(1), 119–139.
Brotcorne, L., Laporte, G., & Semet, F. (2003). Ambulance location and relocation models. Euro-
pean Journal of Operational Research, 147(3), 451–463.
Cathey, F. W., & Dailey, D. J. (2003). A prescription for transit arrival/departure prediction using
automatic vehicle location data. Transportation Research. Part C, Emerging Technologies, 11(3-
4), 241–264.
Chang, H., Givan, R., & Chong, E. (2000). On-line scheduling via sampling. In Artificial intelli-
gence planning and scheduling (AIPS), Breckenridge, Colorado (pp. 62–71).
Christiansen, C. H., & Lysgaard, J. (2007). A branch-and-price algorithm for the capacitated vehi-
cle routing problem with stochastic demands. Operations Research Letters, 35(6), 773–781.
Cordeau, J. F., & Laporte, G. (2003). The dial-a-ride problem (DARP): variants, modeling issues
and algorithms. 4OR: A Quarterly Journal of Operations Research, 1(2), 89–101.
Cordeau, J. F., & Laporte, G. (2007). The dial-a-ride problem: models and algorithms. Annals of
Operations Research, 153(1), 29–46.
Cordeau, J. F., Gendreau, M., Hertz, A., Laporte, G., & Sormany, J. S. (2005). New heuristics for
the vehicle routing problem. In A. Langevin & D. Riopel (Eds.), Logistics systems: design and
optimization (pp. 279–297). New York: Springer. ISBN 978-0-387-24977-3.
Cordeau, J. F., Laporte, G., Potvin, J.-Y., & Savelsbergh, M. W. P. (2007a). Transportation
on demand. In C. Barnhart (Ed.), Handbooks in operations research and management sci-
ence: Vol. 14. Transportation (pp. 429–466). Amsterdam: Elsevier/North-Holland. ISBN
0444513469.
Cordeau, J. F., Laporte, G., Savelsbergh, M. W. P., & Vigo, D. (2007b). Vehicle routing. In C. Barn-
hart (Ed.), Handbooks in operations research and management science: Vol. 14. Transportation
(pp. 367–428). Amsterdam: Elsevier/North-Holland. ISBN 0444513469.
Cortés, C. E., Núñez, A., & Sáez, D. (2008). Hybrid adaptive predictive control for a dynamic
pickup and delivery problem including traffic congestion. International Journal of Adaptive
Control and Signal Processing, 22(2), 103–123.
Cortés, C. E., Sáez, D., Núñez, A., & Muòoz-Carpintero, D. (2009). Hybrid adaptive predictive
control for a dynamic pickup and delivery problem. Transportation Science, 43(1), 27–42.
Crainic, T. G., & Kim, K. H. (2007). Intermodal transportation. In C. Barnhart (Ed.), Handbooks
in operations research and management science: Vol. 14. Transportation (pp. 467–537). Ams-
terdam: Elsevier/North-Holland. ISBN 0444513469.
306 References

Cubillos, C., Crawford, B., & Rodríguez, N. (2007). Distributed planning for the on-line dial-a-
ride problem. In F. Preparata & Q. Fang (Eds.), Lecture notes in computer science: Vol. 4613.
Frontiers in algorithmics (pp. 124–135). Berlin: Springer.
Daimler Group (2011). Annual Report 2010. http://www.daimler.com/Projects/c2c/channel/
documents/1985489_Daimler_Annual_Report_2010.pdf. Accessed on 2011-05-12.
Dakin, R. J. (1965). A tree-search algorithm for mixed integer programming problems. Computer
Journal, 8(3), 250–255.
Dantzig, G., Fulkerson, R., & Johnson, S. (1954). Solution of a large-scale Traveling-Salesman
Problem. Journal of the Operations Research Society of America, 2(4), 393–410.
Dantzig, G. B., & Ramser, J. H. (1959). The truck dispatching problem. Management Science,
6(1), 80–91.
Davis, M. M., & Maggard, M. J. (1990). An analysis of customer satisfaction with waiting times
in a two-stage service process. Journal of Operations Management, 9(3), 324–334.
de Oliveira, H. C. B., Rocha, G. M., de Souza, M. M., Ciscon, L. A., Borges, V. R., & Mateus, G. R.
(2008). A vehicular waiting time heuristic for dynamic vehicle routing problem. In Proceedings
of the 2008 ACM symposium on applied computing (SAC ’08) (pp. 13–17). New York: ACM.
de Paepe, W. E. (2002). Complexity results and competitive analysis for vehicle routing problems.
PhD Thesis, Eindhoven: Technische University of Eindhoven.
Desaulniers, G., Desrosiers, J., & Solomon, M. M. (2005). GERAD 25th anniversary series. Col-
umn generation. Boston: Springer. ISBN 0387254854.
Desrochers, M., & Laporte, G. (1991). Improvements and extensions to the Miller-Tucker-Zemlin
subtour elimination constraints. Operations Research Letters, 10(1), 27–36.
Desrochers, M., Lenstra, J. K., Savelsbergh, M. W. P., & Soumis, F. (1988). Vehicle routing with
time windows: optimization and approximation. In Vehicle routing: Methods and studies (pp.
65–84). Amsterdam: North-Holland.
Desrosiers, J., Dumas, Y., Solomon, M. M., & Soumis, F. (1995). Time constrained routing and
scheduling. In M. O. Ball, T. L. Magnanti, C. L. Monma & G. L. Nemhauser (Eds.), Hand-
books in operations research and management science: Vol. 8. Network routing (pp. 35–139).
Amsterdam: North-Holland. ISBN 978-0444821416.
Deutsch, P., & Gailly, J. L. (1996). RFC1950: ZLIB compressed data format specification version
3.3. http://www.ietf.org/rfc/rfc1950.txt. Accessed on 2009-03-04.
DHL Express Germany GmbH (2010). Export Services. http://www.dhl.de/en/express/
export-services.html. Accessed on 2010-12-06.
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische Mathematik,
1(1), 269–271.
Domschke, W., & Drexl, A. (2011). Einführung in operations research (8th ed.). Berlin: Springer.
ISBN 3642181112.
Domschke, W., & Scholl, A. (2010). Oldenbourgs Lehr- und Handbücher der Wirtschafts- und
Sozialwissenschaften: Vol. 2. Rundreisen und Touren (5th revised and extended ed.). München:
Oldenbourg. ISBN 3486590936.
Dror, M. (1989). Vehicle routing with stochastic demands: properties and solution frameworks.
Transportation Science, 23(3), 166–176.
Dror, M. (2005). Vehicle routing with stochastic demands: models & computational methods. In
M. Dror, P. L’Ecuyer & F. Szidarovszky (Eds.), International series in operations research &
management science: Vol. 46. Modeling uncertainty (pp. 625–649). Boston: Springer. ISBN
978-0-306-48102-4.
Dror, M., Laporte, G., & Louveaux, F. V. (1993). Vehicle routing with stochastic demands and
restricted failures. Mathematical Methods of Operations Research, 37(3), 273–283.
Dumas, Y., Soumis, F., & Desrosiers, J. (1990). Technical note—optimizing the schedule for a
fixed vehicle path with convex inconvenience costs. Transportation Science, 24(2), 145–152.
Dumas, Y., Desrosiers, J., Gelinas, E., & Solomon, M. M. (1995). An optimal algorithm for the
Traveling Salesman Problem with time windows. Operations Research, 43(2), 367–371.
References 307

Edmonds, R. (2010). NAA Report: newspaper inserts “Under Siege”. http://www.poynter.org/

latest-news/buiness-news/the-biz-blog/100805/naa-report-newspaper-inserts-under-siege/. Ac-
cessed on 2011-05-18.
Ehmke, J. F., Meisel, S., Engelmann, S., & Mattfeld, D. C. (2009). Data chain management for
planning in city logistics. International Journal of Data Mining, Modelling and Management,
1(4), 335–356.
Ehmke, J. F., Meisel, S., & Mattfeld, D. C. (2010). Floating car data based analysis of urban travel
times for the provision of traffic quality. In J. Barceló & M. Kuwahara (Eds.), International
series in operations research & management science: Vol. 144. Traffic data collection and its
standardization (pp. 129–149). New York: Springer. ISBN 978-1-4419-6070-2.
Eiselt, H. A., Gendreau, M., & Laporte, G. (1995). Arc routing problems, part I: the Chinese
postman problem. Operations Research, 43(2), 231–242.
Eksioglu, B., Vural, A. V., & Reisman, A. (2009). The vehicle routing problem: a taxonomic re-
view. Computers & Industrial Engineering, 57(4), 1472–1483.
El-Sherbeny, N. A. (2010). Vehicle routing with time windows: an overview of exact, heuristic and
metaheuristic methods. Journal of King Saud University: Computer and Information Sciences,
22(3), 123–131.
European Commission (Ed.) (2010). European union energy & transport in figures: Statistical
pocketbook 2010. Luxembourg: Publications office of the European communities. ISBN 978-
92-79-13815-7.
Fagerholt, K. (2000). Evaluating the trade-off between the level of customer service and transporta-
tion costs in a ship scheduling problem. Maritime Policy and Management, 27(2), 145–153.
Fagerholt, K. (2001). Ship scheduling with soft time windows: an optimisation based approach.
European Journal of Operational Research, 131(3), 559–571.
Faria, J. A., Nunes, E., & Matos, M. A. (2010). Cost and quality of service analysis of produc-
tion systems based on the cumulative downtime. International Journal of Production Research,
48(6), 1653–1684.
Ferrucci, F. (2006). Entwicklung und Validierung einer Echtzeitsteuerung für realitätsnahe Pickup-
and-Delivery-Probleme. Diploma Thesis, University of Paderborn.
Ferrucci, F., & Bock, S. (2008). Statische Analyse des Optimierungspotenzials einer Nachliefer-
ungssteuerung für Tageszeitungen. In D. C. Mattfeld, S. Voß, L. Suhl & H.-O. Günther (Eds.),
DSOR Beiträge zur Wirtschaftsinformatik: Vol. 5. Informations- und Kommunikationssysteme
in Supply Chain Management, Logistik und Transport. Norderstedt: Books on Demand. ISBN
978-3-8334-8931-0.
Ferrucci, F., Bock, S., & Gendreau, M. (2013). A pro-active real-time control approach for dy-
namic vehicle routing problems dealing with the delivery of urgent goods. European Journal of
Operational Research, 225(1), 130–141.
Fischer, K., Müller, J. P., & Pischel, M. (1996). Cooperative transportation scheduling: an applica-
tion domain for DAI. Applied Artificial Intelligence, 10(1), 1–34.
Fisher, M. L. (2004). The Lagrangian relaxation method for solving integer programming prob-
lems. Management Science, 50(12), 1861–1871.
Flatberg, T., Hasle, G., Kloster, O., Nilssen, E. J., & Riise, A. (2005). Dynamic and stochastic
aspects in vehicle routing: a literature survey (Technical report STF90A05413 SINTEF). Oslo,
Norway.
Fleischmann, B., Gietz, M., & Gnutzmann, S. (2004a). Time-varying travel times in vehicle rout-
ing. Transportation Science, 38(2), 160–173.
Fleischmann, B., Gnutzmann, S., & Sandvoß, E. (2004b). Dynamic vehicle routing based on online
traffic information. Transportation Science, 38(4), 420–433.
Fukasawa, R., Longo, H., Lysgaard, J., Aragão, M. P., Reis, M., Uchoa, E., & Werneck, R. F.
(2006). Robust branch-and-cut-and-price for the capacitated vehicle routing problem. Mathe-
matical Programming, 106(3), 491–511.
Garey, M. R., & Johnson, D. S. (1978). “Strong” NP-completeness results: motivation, examples,
and implications. Journal of the ACM, 25(3), 499–508.
308 References

Garey, M. R., & Johnson, D. S. (1979). A series of books in the mathematical sciences. Computers
and intractability: a guide to the theory of NP-completeness (1st ed.). New York: Freeman.
ISBN 0716710455.
Gass, S. I. (2003). Linear programming: methods and applications (5th ed.). New York: Dover.
ISBN 048643284X.
Gendreau, M., & Potvin, J.-Y. (1998). Dynamic vehicle routing and dispatching. In T. G. Crainic
(Ed.), Center for research on transportation 25th anniversary series, 1971–1996. Fleet man-
agement and logistics (pp. 115–126). Heidelberg: Springer. ISBN 0792381610.
Gendreau, M., Laporte, G., & Séguin, R. (1996). Stochastic vehicle routing. European Journal of
Operational Research, 88(1), 3–12.
Gendreau, M., Guertin, F., Potvin, J.-Y., & Séguin, R. (1998). Neighborhood search heuristics for
a dynamic vehicle dispatching problem with pick-ups and deliveries (Technical Report CRT-98-
10). Centre de Recherche sur les Transport (CRT), Universite de Montreal.
Gendreau, M., Guertin, F., Potvin, J.-Y., & Taillard, É. D. (1999a). Parallel tabu search for real-time
vehicle routing and dispatching. Transportation Science, 33(4), 381–390.
Gendreau, M., Laporte, G., Musaraganyi, C., & Taillard, É. D. (1999b). A tabu search heuristic for
the heterogeneous fleet vehicle routing problem. Computers & Operations Research, 26(12),
1153–1173.
Gendreau, M., Laporte, G., & Semet, F. (2001). A dynamic model and parallel tabu search heuristic
for real-time ambulance relocation. Parallel Computing, 27(12), 1641–1653.
Gendreau, M., Guertin, F., Potvin, J.-Y., & Séguin, R. (2006). Neighborhood search heuristics for
a dynamic vehicle dispatching problem with pick-ups and deliveries. Transportation Research.
Part C, Emerging Technologies, 14(3), 157–174.
Gendreau, M., Potvin, J.-Y., Bräumlaysy, O., Hasle, G., & Løkketangen, A. (2008). Metaheuristics
for the vehicle routing problem and its extensions: a categorized bibliography. In B. L. Golden,
S. Raghavan & E. A. Wasil (Eds.), Operations research/computer science interfaces series:
Vol. 43. The vehicle routing problem: latest advances and new challenges (pp. 143–169). Hei-
delberg: Springer. ISBN 0387777776.
Ghiani, G., Guerriero, F., Laporte, G., & Musmanno, R. (2003). Real-time vehicle routing: solu-
tion concepts, algorithms and parallel computing strategies. European Journal of Operational
Research, 151(1), 1–11.
Ghiani, G., Laporte, G., & Musmanno, R. (2004). Wiley-Interscience series in systems and op-
timization. Introduction to logistics systems planning and control. Chichester: Wiley. ISBN
9780470849170.
Ghiani, G., Manni, E., Quaranta, A., & Triki, C. (2009). Anticipatory algorithms for same-day
courier dispatching. Transportation Research. Part E, Logistics and Transportation Review,
45(1), 96–106.
Ghiani, G., Manni, E., & Thomas, B. W. (2012). A comparison of anticipatory algorithms for the
dynamic and stochastic Traveling Salesman Problem. Transportation Science, 46(3), 374–387.
Giaglis, G. M., Minis, I., Tatarakis, A., & Zeimpekis, V. (2004). Minimizing logistics risk through
real-time vehicle routing and mobile technologies. International Journal of Physical Distribu-
tion & Logistics Management, 34(9), 749–764.
Glover, F. (1989). Tabu search—part I. ORSA Journal on Computing, 1(3), 190–206.
Glover, F., & Laguna, M. (1998). Tabu search (4th ed.). Heidelberg: Springer. ISBN 0792381874.
Godfrey, G. A., & Powell, W. B. (2002). An adaptive dynamic programming algorithm for dynamic
fleet management, I: single period travel times. Transportation Science, 36(1), 21–39.
Golden, B. L., & Stewart Jr., W. (1978). Vehicle routing with probabilistic demands. In Computer
science and statistics: tenth annual symposium on the interface: proceedings of the 10th annual
symposium held at the National Bureau of Standards, Gaithersburg, Maryland (pp. 252–259).
Golden, B. L., Raghavan, S., & Wasil, E. A. (Eds.) (2008). Operations research/computer science
interfaces series: Vol. 43. The vehicle routing problem: latest advances and new challenges.
Heidelberg: Springer. ISBN 0387777776.
Hadjiconstantinou, E., & Roberts, D. (2002). Routing under uncertainty: an application in the
scheduling of field service engineers. In P. Toth & D. Vigo (Eds.), SIAM monographs on dis-
References 309

crete mathematics and applications. The vehicle routing problem (pp. 331–352). Philadelphia:
Society for Industrial and Applied Mathematics. ISBN 0898714982.
Haghani, A., & Yang, S. (2007). Real-time emergency response fleet deployment: concepts, sys-
tems, simulation & case studies. In V. Zeimpekis, G. M. Giaglis, I. Minis & C. D. Tarantilis
(Eds.), Operations research/computer science interfaces series: Vol. 38. Dynamic fleet manage-
ment (pp. 133–162). Boston: Springer. ISBN 978-0-387-71722-7.
Haklay, M., & Weber, P. (2008). OpenStreetMap: user-generated street maps. IEEE Pervasive
Computing, 7(4), 12–18.
Halonen, T. (2003). GSM, GPRS, and edge performance: evolution towards 3G/UMTS (2nd ed.).
Chichester: Wiley. ISBN 0470866942.
Hansen, P., & Mladenovic, N. (2001). Variable neighborhood search: principles and applications.
European Journal of Operational Research, 130(3), 449–467.
Held, M., & Karp, R. M. (1970). The Traveling-Salesman Problem and minimum spanning trees.
Operations Research, 18(6), 1138–1162.
Held, M., & Karp, R. M. (1971). The Traveling-Salesman Problem and minimum spanning trees:
Part II. Mathematical Programming, 1(1), 6–25.
Hillier, F. S., & Lieberman, G. J. (2008). Introduction to operations research (8th ed.). New York:
McGraw-Hill. ISBN 007123828X.
Homberger, J., & Gehring, H. (2005). A two-phase hybrid metaheuristic for the vehicle routing
problem with time windows. European Journal of Operational Research, 162(1), 220–238.
Hoos, H. H., & Stützle, T. (2005). Stochastic local search: foundations and applications. San
Francisco: Morgan Kaufmann. ISBN 1558608729.
Horn, M. E. T. (2002a). Fleet scheduling and dispatching for demand-responsive passenger ser-
vices. Transportation Research. Part C, Emerging Technologies, 10(1), 35–63.
Horn, M. E. T. (2002b). Multi-modal and demand-responsive passenger transport systems: a mod-
elling framework with embedded control systems. Transportation Research. Part A, Policy and
Practice, 36(2), 167–188.
Hülsmann, M., Kopfer, H., Cordes, P., & Bloos, M. (2009). Collaborative transportation plan-
ning in complex adaptive logistics systems: a complexity science-based analysis of decision-
making problems of “groupage systems”. In J. Zhou (Ed.), Lecture notes of the institute for
computer sciences, social informatics and telecommunications engineering: Vol. 4. Complex
sciences (pp. 1160–1166). Berlin: Springer. ISBN 978-3-642-02466-5.
Hvattum, L. M., Løkketangen, A., & Laporte, G. (2006). Solving a dynamic and stochastic vehicle
routing problem with a sample scenario hedging heuristic. Transportation Science, 40(4), 421–
438.
Hvattum, L. M., Løkketangen, A., & Laporte, G. (2007). A branch-and-regret heuristic for stochas-
tic and dynamic vehicle routing problems. Networks, 49(4), 330–340.
Ichoua, S. (2001). Problèmes de gestion de flottes de véhicules en temps réel. PhD thesis, Université
de Montréal, Montréal, Canada.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2000). Diversion issues in real-time vehicle dispatching.
Transportation Science, 34(4), 426–438.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2003). Vehicle dispatching with time-dependent travel
times. European Journal of Operational Research, 144(2), 379–396.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2006). Exploiting knowledge about future demands for
real-time vehicle dispatching. Transportation Science, 40(2), 211–225.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2007). Planned route optimization for real-time ve-
hicle routing. In V. Zeimpekis, G. M. Giaglis, I. Minis & C. D. Tarantilis (Eds.), Operations
research/computer science interfaces series: Vol. 38. Dynamic fleet management (pp. 1–18).
Boston: Springer. ISBN 978-0-387-71722-7.
Irani, S., Lu, X., & Regan, A. (2004). On-line algorithms for the dynamic traveling repair problem.
Journal of Scheduling, 7(3), 243–258.
Israel, D. (2008). Data analysis in business research: a step-by-step nonparametric approach. Los
Angeles: SAGE. ISBN 9788178298757.
310 References

Jaillet, P. (1987). Stochastic routing problems. In G. Andreatta, F. Mason, & P. Serafini (Eds.),
Advanced school on stochastics in combinatorial optimization (pp. 197–213). Singapore: World
Scientific.
Jézéquel, A. (1985). Probabilistic vehicle routing problems. M.Sc. Dissertation, Massachusetts
Institute of Technology, Cambridge, MA.
Johnston, R., & Clark, G. (2008). Service operations management: improving service delivery (3rd
ed.). Harlow: Prentice Hall. ISBN 9781405847322.
Jula, H., Dessouky, M., & Ioannou, P. A. (2008). Real-time estimation of travel times along the arcs
and arrival times at the nodes of dynamic stochastic networks. IEEE Transactions on Intelligent
Transportation Systems, 9(1), 97–110.
Kaplan, E. D., & Hegarty, C. J. (2006). Understanding GPS: principles and applications (2nd ed.).
Boston: Artech House. ISBN 1580538940.
Kendall, D. G. (1953). Stochastic processes occurring in the theory of queues and their analysis
by the method of the imbedded Markov chain. The Annals of Mathematical Statistics, 24(3),
338–354.
Kenyon, A. S., & Morton, D. P. (2003). Stochastic vehicle routing with random travel times. Trans-
portation Science, 37(1), 69–82.
Kilby, P., Prosser, P., & Shaw, P. (1998). Dynamic VRPs: a study of scenarios (Technical Report
APES-06-1998). University of Strathclyde.
Klein Haneveld, W. K., & van der Vlerk, M. H. (1999). Stochastic integer programming: general
models and algorithms. Annals of Operations Research, 85, 39–57.
Kleywegt, A. J., Savelsbergh, M., & Uyar, E. (2009). A dynamic stochastic routing problem (Tech-
nical Report). Georgia Institute of Technology, Atlanta, GA.
Kohout, R., & Erol, K. (1999). In-time agent-based vehicle routing with a stochastic improvement
heuristic. In Proceedings of the 16th national conference on artificial intelligence and the 11th
on innovative applications of artificial intelligence, Orlando, USA (pp. 864–869).
Korte, B., & Vygen, J. (2012). Algorithms and combinatorics: Vol. 21. Combinatorial optimiza-
tion: theory and algorithms (5th ed.). Berlin: Springer. ISBN 3642244874.
Kristensen, K., Kanji, G. K., & Dahlgaard, J. J. (1992). On measurement of customer satisfaction.
Total Quality Management & Business Excellence, 3(2), 123–128.
Krumke, S. O., Rambau, J., & Torres, L. M. (2002). Real-time dispatching of guided and unguided
automobile service units with soft time windows. In R. Möhring & R. Raman (Eds.), Lec-
ture notes in computer science: Vol. 2461. Algorithms—ESA 2002 (pp. 417–424). Heidelberg:
Springer.
Kvam, P. H., & Vidakovic, B. (2007). Nonparametric statistics with applications to science and
engineering. Wiley series in probability and statistics. Hoboken: Wiley. ISBN 0470081473.
Kwan, M. K. (1962). Graphic programming using odd or even points. Chinese Mathematics, 1,
273–277.
Lackner, A. (2004). Göttinger Wirtschaftsinformatik: Vol. 47. Dynamische Tourenplanung mit aus-
gewählten Metaheuristiken: Eine Untersuchung am Beispiel des kapazitätsrestriktiven dynamis-
chen Tourenplanungsproblems mit Zeitfenstern: Univ., Diss.–Göttingen, 2004 (1st ed.). Göttin-
gen: Cuvillier. ISBN 3865370845.
Lambert, V., Laporte, G., & Louveaux, F. V. (1993). Designing collection routes through bank
branches. Computers & Operations Research, 20(7), 783–791.
Land, A. H., & Doig, A. G. (1960). An automatic method of solving discrete programming prob-
lems. Econometrica, 28(3), 497–520.
Laporte, G. (1992a). The vehicle routing problem: an overview of exact and approximate algo-
rithms. European Journal of Operational Research, 59(3), 345–358.
Laporte, G. (1992b). The traveling salesman problem: an overview of exact and approximate algo-
rithms. European Journal of Operational Research, 59(2), 231–247.
Laporte, G., & Louveaux, F. (1990). Formulations and bounds for the stochastic capacitated vehicle
routing problem with uncertain supplies. In J. Gabszewicz, J.-F. Richard & L. A. Wolsey (Eds.),
Economic decision-making (pp. 443–455). Amsterdam: North-Holland. ISBN 044488422X.
References 311

Laporte, G., & Nobert, Y. (1980). A cutting planes algorithm for the m-Salesmen Problem. Journal
of the Operational Research Society, 31(11), 1017–1023.
Laporte, G., Louveaux, F., & Mercure, H. (1992). The vehicle routing problem with stochastic
travel times. Transportation Science, 26(3), 161–170.
Laporte, G., Louveaux, F. V., & van Hamme, L. (2002). An integer L-shaped algorithm for the
capacitated vehicle routing problem with stochastic demands. Operations Research, 50(3), 415–
423.
Larsen, A. (2000). The dynamic vehicle routing problem. PhD Thesis, Technical University of
Denmark, Lyngby, Denmark.
Larsen, A., Madsen, O. B. G., & Solomon, M. M. (2002). Partially dynamic vehicle routing—
models and algorithms. Journal of the Operational Research Society, 53(6), 637–646.
Larsen, A., Madsen, O. B. G., & Solomon, M. M. (2004). The a priori dynamic Traveling Salesman
Problem with time windows. Transportation Science, 38(4), 459–472.
Larsen, A., Madsen, O. B. G., & Solomon, M. M. (2007). Classification of dynamic vehicle rout-
ing systems. In V. Zeimpekis, G. M. Giaglis, I. Minis & C. D. Tarantilis (Eds.), Operations
research/computer science interfaces series: Vol. 38. Dynamic fleet management (pp. 19–40).
Boston: Springer. ISBN 978-0-387-71722-7.
Larsen, A., Madsen, O. B. G., & Solomon, M. M. (2008). Recent developments in dynamic ve-
hicle routing systems. In B. L. Golden, S. Raghavan & E. A. Wasil (Eds.), Operations re-
search/computer science interfaces series: Vol. 43. The vehicle routing problem: latest advances
and new challenges (pp. 199–218). Heidelberg: Springer. ISBN 0387777776.
Larson, R. C., & Odoni, A. R. (1981). Urban operations research. Englewood Cliffs: Prentice-Hall.
http://web.mit.edu/urban_or_book/www/book/.
Law, A. M., & Kelton, W. D. (2000). McGraw-Hill series in industrial engineering and man-
agement science. Simulation modeling and analysis (3rd ed.). Boston: McGraw-Hill. ISBN
0070592926.
Lenstra, J. K., & Rinnooy Kan, A. H. G. (1979). Computational complexity of discrete optimization
problems. In P. L. Hammer, E. L. Johnson & B. H. Korte (Eds.), Proceedings of the advanced
research institute on discrete optimization and systems applications of the systems science panel
of NATO and of the discrete optimization symposium: Vol. 4. Annals of discrete mathematics:
discrete optimization I (pp. 121–140). Amsterdam: Elsevier. ISBN 978-0-444-85322-6.
Lenstra, J. K., & Rinnooy Kan, A. H. G. (1981). Complexity of vehicle routing and scheduling
problems. Networks, 11(2), 221–227.
Lew, A., & Mauch, H. (2007). Dynamic programming: a computational tool (1st ed.). Berlin:
Springer. ISBN 3642072003.
Li, J. Q., Mirchandani, P. B., & Borenstein, D. (2009a). Real-time vehicle rerouting problems with
time windows. European Journal of Operational Research, 194(3), 711–727.
Li, J. Q., Mirchandani, P. B., & Borenstein, D. (2009b). A Lagrangian heuristic for the real-time
vehicle rescheduling problem. Transportation Research. Part E, Logistics and Transportation
Review, 45(3), 419–433.
Lin, S. (1965). Computer solutions of the traveling salesman problem. The Bell System Technical
Journal, 44(10), 2245–2269.
Little, J. D. C., Murty, K. G., Sweeney, D. W., & Karel, C. (1963). An algorithm for the Traveling
Salesman Problem. Operations Research, 11(6), 972–989.
Lund, K., Madsen, O. B. G., & Rygaard, J. M. (1996). Vehicle routing with varying degrees of dy-
namism (Technical report IMM-REP-1996-1). Institute of Mathematical Modelling, Technical
University of Denmark, Lyngby, Denmark.
MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations.
In L. M. Le Cam & J. Neyman (Eds.), Proceedings of the fifth Berkeley symposium on math-
ematical statistics and probability (Vol. 1, pp. 281–297). Berkeley: University of California
Press.
Malandraki, C., & Daskin, M. S. (1992). Time dependent vehicle routing problems: formulations,
properties and heuristic algorithms. Transportation Science, 26(3), 185–200.
312 References

Markus, K. G. (2009). Online stochastic vehicle routing. Master’s Thesis, Department of Mathe-
matics and Computer Science (IMADA), University of Southern Denmark.
Mendoza, J. E., Castanier, B., Guéret, C., Medaglia, A. L., & Velasco, N. (2009). Constructive
heuristics for the multi-compartment vehicle routing problem with stochastic demands (Techni-
cal report). Universidad de los Andes.
Mihram, G. A. (1972). Mathematics in science and engineering: Vol. 92. Simulation: statistical
foundations and methodology. New York: Academic Press. ISBN 0124959504.
Miller, C. E., Tucker, A. W., & Zemlin, R. A. (1960). Integer programming formulation of traveling
salesman problems. Journal of the ACM, 7(4), 326–329.
Mintsis, G., Basbas, S., Papaioannou, P., Taxiltaris, C., & Tziavos, I. N. (2004). Applications of
GPS technology in the land transportation system: new technologies in transportation systems.
European Journal of Operational Research, 152(2), 399–409.
Mitchell, M. (1998). An introduction to genetic algorithms. Cambridge: MIT Press. ISBN
0262631857.
Mitrovic-Minic, S., & Laporte, G. (2004). Waiting strategies for the dynamic pickup and delivery
problem with time windows. Transportation Research. Part B: Methodological, 38(7), 635–
655.
Mitrovic-Minic, S., Krishnamurti, R., & Laporte, G. (2004). Double-horizon based heuristics for
the dynamic pickup and delivery problem with time windows. Transportation Research. Part B:
Methodological, 38(8), 669–685.
Moon, T. K. (2005). Error correction coding: mathematical methods and algorithms. Hoboken:
Wiley-Interscience. ISBN 0471648000.
Mu, Q., Fu, Z., Lysgaard, J., & Eglese, R. (2010). Disruption management of the vehicle routing
problem with vehicle breakdown. Journal of the Operational Research Society, 62(4), 742–749.
Muyldermans, L., & Pang, G. (2010). On the benefits of co-collection: experiments with a multi-
compartment vehicle routing algorithm. European Journal of Operational Research, 206(1),
93–103.
Nanry, W. P., & Wesley Barnes, J. (2000). Solving the pickup and delivery problem with time
windows using reactive tabu search. Transportation Research. Part B: Methodological, 34(2),
107–121.
Newspaper Association of America (2011). Annual Advertising Expenditures. http://www.naa.
org/Trends-and-Numbers/Advertising-Expenditures/Annual-All-Categories.aspx. Accessed on
2011-06-05.
Novoa, C., & Storer, R. (2009). An approximate dynamic programming approach for the vehicle
routing problem with stochastic demands. European Journal of Operational Research, 196(2),
509–515.
Novoa, C., Berger, R., Linderoth, J., & Storer, R. (2006). A set-partitioning-based model for the
stochastic vehicle routing problem (Technical Report 06T-008). Industrial and Systems Engi-
neering, Lehigh University, USA.
Núñez, A., Sáez, D., & Cortés, C. E. (2008). Hybrid predictive control for the vehicle dynamic rout-
ing problem based on Evolutionary Multiobjective Optimization (EMO). In 17th IFAC world
congress, Seoul, Korea.
OpenStreetMap (2011). The OpenStreetMap Project. http://www.openstreetmap.org. Accessed on
2011-06-20.
Or, I. (1976). Traveling salesman-type combinatorial problems and their relation to the logistics
of regional blood banking. PhD Thesis, Northwestern University, Evanston.
Pankratz, G. (2002). Speditionelle Transportdisposition: Modell- und Verfahrensentwicklung
unter Berücksichtigung von Dynamik und Fremdvergabe. Diss. FernUniv. Hagen. Wiesbaden:
Deutscher Universitätsverlag. ISBN 3824476681.
Papastavrou, J. D. (1996). A stochastic and dynamic routing policy using branching processes with
state dependent immigration. European Journal of Operational Research, 95(1), 167–177.
Parasuraman, A. (1987). Customer-oriented corporate cultures are crucial to services marketing
success. The Journal of Services Marketing, 1(1), 39–46.
References 313

Parasuraman, A., Zeithaml, V. A., & Berry, L. L. (1985). A conceptual model of service quality
and its implications for future research. The Journal of Marketing, 49(4), 41–50.
Parasuraman, A., Berry, L. L., & Zeithaml, V. A. (1991). Understanding customer expectations of
service. Sloan Management Review, 32(3), 39–48.
Park, S. K., & Miller, K. W. (1988). Random number generators: good ones are hard to find.
Communications of the ACM, 31(10), 1192–1201.
Park, S. K., Miller, K. W., & Stockmeyer, P. K. (1993). Technical correspondence: remarks on
choosing and implementing random number generators. Communications of the ACM, 36(7),
105. doi:10.1145/159544.376068.
Parragh, S., Doerner, K., & Hartl, R. (2008). A survey on pickup and delivery problems: Part I:
transportation between customers and depot. Journal für Betriebswirtschaft, 58(1), 21–51.
Perugini, D., Lambert, D., Sterling, L., & Pearce, A. (2003). A distributed agent approach to global
transportation scheduling. In 2003 IEEE/WIC international conference on intelligent agent tech-
nology (IAT ’03), Halifax, Canada.
Pillac, V., Guéret, A., & Medaglia, A. (2010). Vehicle routing problems: state of the art and
prospects (Technical report, 10/4/AUTO). Ecole des Mines de Nantes, France.
Pollack, B. L. (2009). Linking the hierarchical service quality model to customer satisfaction and
loyalty. The Journal of Services Marketing, 23(1), 42–50.
Potvin, J.-Y., & Rousseau, J. M. (1995). An exchange heuristic for routeing problems with time
windows. Journal of the Operational Research Society, 46(12), 1433–1446.
Powell, W. B. (1988). A comparative review of alternative algorithms for the dynamic vehicle
allocation problem. In B. L. Golden (Ed.), Studies in management science and systems. Vehicle
routing (pp. 249–291). Amsterdam: Elsevier. ISBN 0444704078.
Powell, W. B. (1996). A stochastic formulation of the dynamic assignment problem, with an appli-
cation to truckload motor carriers. Transportation Science, 30(3), 195–219.
Powell, W. B. (2007). Approximate dynamic programming: solving the curses of dimensionality.
Hoboken: Wiley-Interscience. ISBN 0470171553.
Powell, W. B. (2009). What you should know about approximate dynamic programming. Naval
Research Logistics, 56(3), 239–249.
Powell, W. B., & Topaloglu, H. (2005). Fleet management. In S. W. Wallace & W. T. Ziemba (Eds.),
MPS-SIAM series on optimization: Vol. 5. Applications of stochastic programming (pp. 185–
215). Philadelphia: Society for Industrial and Applied Mathematics. ISBN 0898715555.
Powell, W. B., Jaillet, P., & Odoni, A. (1995). Stochastic and dynamic networks and routing. In
Handbooks in operations research and management science (Vol. 8, pp. 141–295).
Psaraftis, H. N. (1980). Dynamic programming solution to the single vehicle many-to-many im-
mediate request dial-a-ride problem. Transportation Science, 14(2), 130–154.
Psaraftis, H. N. (1988). Dynamic vehicle routing problems. In B. L. Golden (Ed.), Studies in
management science and systems. Vehicle routing (pp. 223–248). Amsterdam: Elsevier. ISBN
0444704078.
Psaraftis, H. N. (1995). Dynamic vehicle routing: status and prospects. Annals of Operations Re-
search, 61(1), 143–164.
Pureza, V., & Laporte, G. (2008). Waiting and buffering strategies for the dynamic pickup and
delivery problem with time windows. INFOR. Information Systems and Operational Research,
46(3), 165–176.
Puterman, M. L. (2005). Wiley series in probability and statistics. Markov decision processes: dis-
crete stochastic dynamic programming. Hoboken: Wiley-Interscience. ISBN 9780470316887.
Regan, A. C., Mahmassani, H. S., & Jaillet, P. (1995). Improving efficiency of commercial vehicle
operations using real-time information: potential uses and assignment strategies. Transportation
Research Record, 1493, 188–198.
Regan, A. C., Mahmassani, H. S., & Jaillet, P. (1998). Evaluation of dynamic fleet management
systems: simulation framework. Transportation Research Record, 1645, 176–184.
Rice, J., & Caniato, F. (2003). Building a secure and resilient supply network. Supply Chain Man-
agement Review, 7(5), 22–30.
314 References

Richter, A. (2005). Dynamische Tourenplanung: Modifikation von klassischen Heuristiken für das
Dynamische Rundreiseproblem (DTSP) und das Dynamische Tourenplanungsproblem (DVRP)
mit der Möglichkeit der Änderung des aktuellen Fahrzeugzuges. PhD Thesis, University of
Dresden, Germany.
Rodgers, J. L., & Nicewander, W. A. (1988). Thirteen ways to look at the correlation coefficient.
American Statistician, 42(1), 59–66.
Ropke, S., & Pisinger, D. (2006). An adaptive large neighborhood search heuristic for the pickup
and delivery problem with time windows. Transportation Science, 40(4), 455–472.
Ross, S. M. (2010). Introduction to probability models (10th ed.). Burlington: Academic
Press/Elsevier. ISBN 0123756863.
Sáez, D., Cortés, C. E., & Núñez, A. (2008). Hybrid adaptive predictive control for the multi-
vehicle dynamic pick-up and delivery problem based on genetic algorithms and fuzzy clustering.
Computers & Operations Research, 35(11), 3412–3438.
Sapsford, R., & Jupp, V. (2006). Data collection and analysis (2nd ed.). London: SAGE. ISBN
9780761943631.
Savelsbergh, M. W. P. (1992). The vehicle routing problem with time windows: minimizing route
duration. ORSA Journal on Computing, 4, 146–154.
Savelsbergh, M. W. P., & Sol, M. (1995). The general pickup and delivery problem. Transportation
Science, 29(1), 17–29.
Savelsbergh, M. W. P., & Sol, M. (1998). DRIVE: dynamic routing of independent vehicles. Op-
erations Research, 46(4), 474–490.
Schmid, V. (2012). Solving the dynamic ambulance relocation and dispatching problem using ap-
proximate dynamic programming. European Journal of Operational Research, 219(3), 611–
621.
Schmidt, J. W., & Taylor, R. E. (1970). Simulation and analysis of industrial systems. Homewood:
Richard D. Irwin.
Schönberger, J. (2010). Adaptive demand peak management in online transport process planning.
OR Spektrum, 32(3), 831–859.
Schultes, D. (2008). Route Planning in Road Networks. PhD Thesis, Karlsruhe Institute of Tech-
nology.
Secomandi, N. (2000). Comparing neuro-dynamic programming algorithms for the vehicle routing
problem with stochastic demands. Computers & Operations Research, 27(11-12), 1201–1225.
Secomandi, N., & Margot, F. (2009). Reoptimization approaches for the vehicle-routing problem
with stochastic demands. Operations Research, 57(1), 214–230.
Sedgewick, R. (1983). Addison-Wesley series in computer science. Algorithms. Reading: Addison-
Wesley. ISBN 0201066726.
Séguin, R., Potvin, J.-Y., Gendreau, M., Crainic, T. G., & Marcotte, P. (1997). Real-time decision
problems: an operational research perspective. Journal of the Operational Research Society,
48(2), 162–174.
Sexton, T., & Choi, Y. (1986). Pickup and delivery of partial loads with “soft” time windows.
American Journal of Mathematical and Management Sciences, 6(3-4), 369–398.
Sexton, T. R., & Bodin, L. D. (1985). Optimizing single vehicle many-to-many operations with
desired delivery times: I. scheduling. Transportation Science, 19(4), 378–410.
Sleator, D. D., & Tarjan, R. E. (1985). Amortized efficiency of list update and paging rules. Com-
munications of the ACM, 28(2), 202–208.
Smith, A. K., Bolton, R. N., & Wagner, J. (1999). A model of customer satisfaction with service
encounters involving failure and recovery. Journal of Marketing Research, 36(3), 356–372.
Solomon, M. M. (1987). Algorithms for the vehicle routing and scheduling problems with time
window constraints. Operations Research, 35(2), 254–265.
Stenfors, S., Tanner, L., Syrjänen, M., Seppälä, T., & Haapalinna, I. (2007). Executive views con-
cerning decision support tools. European Journal of Operational Research, 181(2), 929–938.
Stone, J., Stewart, R., & Otis, D. (2002). RFC3309: Stream Control Transmission Protocol (SCTP)
checksum change. http://www.ietf.org/rfc/rfc3309.txt. Accessed on 2008-08-03.
References 315

Stone, J. R. (2001). Checksums and the internet. PhD diss, Department of Computer Science,
Stanford University, Stanford, CA, USA. http://wwwlib.umi.com/dissertations/fullcit/3028177.
Stumpf, P. (1998). GVB: Schriftenreihe: Vol. 39 Tourenplanung im speditionellen Güterfern-
verkehr: Univ., Diss.–Augsburg, 1998 (1st ed.). Nürnberg: GVB. ISBN 3922665349.
Swihart, M. R., & Papastavrou, J. D. (1999). A stochastic and dynamic model for the single-vehicle
pick-up and delivery problem. European Journal of Operational Research, 114(3), 447–464.
Tarantilis, C. D., & Kiranoudis, C. T. (2007). A flexible adaptive memory-based algorithm for real-
life transportation operations: two case studies from dairy and construction sector. European
Journal of Operational Research, 179(3), 806–822.
Tarantilis, C. D., Diakoulaki, D., & Kiranoudis, C. T. (2004). Combination of geographical infor-
mation system and efficient routing algorithms for real life distribution operations: new tech-
nologies in transportation systems. European Journal of Operational Research, 152(2), 437–
453.
Tax, S. S., Brown, S. W., & Chandrashekaran, M. (1998). Customer evaluations of service com-
plaint experiences: implications for relationship marketing. The Journal of Marketing, 62(2),
60–76.
Telefónica Germany GmbH & Co. OHG (2011). Tarif o2 Active Data Card. http://www.
o2online.de/nw/tarife/datentarif/datentarif.html. Accessed on 2011-06-20.
Thangiah, S. R., Osman, I. H., Vinayagamoorthy, R., & Sun, T. (1993). Algorithms for the vehicle
routing problems with time deadlines. American Journal of Mathematical and Management
Sciences, 13(3-4), 323–355.
The Media Audit (2011). Daily Newspaper Websites see 5.6 Percent Growth. http://
www.themediaaudit.com/press/archived-newsletters/the-media-audit-fyi/march-2011/daily-
newspaper-websites-see-56-percent-growth. Accessed on 2011-07-29.
Thomas, B. W. (2007). Waiting strategies for anticipating service requests from known customer
locations. Transportation Science, 41(3), 319–331.
Thomas, B. W., & White III, C. C. (2004). Anticipatory route selection. Transportation Science,
38(4), 473–487.
Tijms, H. C. (2003). A first course in stochastic models (2nd ed.). Chichester: Wiley.
Tillman, F. A. (1969). The multiple terminal delivery problem with probabilistic demands. Trans-
portation Science, 3(3), 192–204.
Toth, P., & Vigo, D. (Eds.) (2002). SIAM monographs on discrete mathematics and applications.
The vehicle routing problem. Philadelphia: Society for Industrial and Applied Mathematics.
ISBN 0898714982.
Uyar, E. (2008). Routing in stochastic environments. PhD Thesis, Georgia Institute of Technology.
van de Klundert, J., & Wormer, L. (2010). ASAP: the After-Salesman Problem. Manufacturing &
Service Operations Management, 12, 627–641.
van Hemert, J. I., & La Poutré, J. A. (2004). Dynamic routing problems with fruitful regions:
models and evolutionary computation. In X. Yao, E. Burke, J. A. Lozano, J. Smith, J. J. Merelo-
Guervós, J. A. Bullinaria, J. Rowe, P. Tino, A. Kabán & H.-P. Schwefel (Eds.), Lecture notes in
computer science: Vol. 3242. Parallel Problem Solving from Nature—PPSN VIII (pp. 692–701).
Berlin: Springer.
van Hentenryck, P., & Bent, R. (2009). Online stochastic combinatorial optimization. Cambridge:
MIT Press. ISBN 0262513471.
Verweij, B., Ahmed, S., Kleywegt, A. J., Nemhauser, G., & Shapiro, A. (2003). The sample av-
erage approximation method applied to stochastic routing problems: a computational study.
Computational Optimization and Applications, 24(2-3), 289–333.
Weiss, G. (2000). Multiagent systems: a modern approach to distributed artificial intelligence (2nd
ed.). Cambridge: MIT Press. ISBN 0262731312.
Westphal, S., & Krumke, S. O. (2008). Pruning in column generation for service vehicle dispatch-
ing. Annals of Operations Research, 159(1), 355–371.
Wilson, N. H. M., & Colvin, N. J. (1977). Computer control of the Rochester dial-a-ride system
(Technical report R-77-30). Department of Civil Engineering, Massachusetts Institute of Tech-
nology, Center for Transportation Studies, Cambridge, MA.
316 References

Wu, C. W., Beck, J. C., & Brown, K. N. (2004). Dynamic vehicle routing with uncertain customer
demand. In Proceedings of Changes ’04 (International workshop on constraint solving under
change and uncertainty at CP2004), Toronto, Canada (pp. 75–76).
Xu, R. (2005). Survey of clustering algorithms. IEEE Transactions on Neural Networks, 16(3),
645–678.
Yang, J., Jaillet, P., & Mahmassani, H. S. (2004). Real-time multivehicle truckload pickup and
delivery problems. Transportation Science, 38(2), 135–148.
Yang, W. H., Mathur, K., & Ballou, R. H. (2000). Stochastic vehicle routing problem with restock-
ing. Transportation Science, 34(1), 99–112.
Yates, D., Moore, D., & McCabe, G. (1999). The practice of statistics (1st ed.). New York: Free-
man. ISBN 0716737744.
Zeimpekis, V., & Giaglis, G. M. (2005). A dynamic real-time vehicle routing system for distri-
bution operations. In G. J. Doukidis & A. P. Vrechopoulos (Eds.), Consumer driven electronic
transformation (pp. 23–37). Berlin: Springer. ISBN 978-3-540-27059-1.
Zeimpekis, V., Giaglis, G. M., & Minis, I. (2005). A dynamic real-time fleet management system
for incident handling in city logistics. In IEEE 61st vehicular technology conference, VTC 2005-
Spring, Stockholm, Sweden (Vol. 5, pp. 2900–2904).
Zeithaml, V. A., Berry, L. L., & Parasuraman, A. (1996). The behavioral consequences of service
quality. The Journal of Marketing, 60(2), 31–46.
Zhan, F. B. (1997). Three fastest shortest path algorithms on real road networks: data structures
and procedures. Journal of Geographic Information and Decision Analysis, 1(1), 69–82.
Zhan, F. B., & Noon, C. E. (1998). Shortest path algorithms: an evaluation using real road net-
works. Transportation Science, 32(1), 65–73.
Zhan, F. B., & Noon, C. E. (2000). A comparison between label-setting and label-correcting algo-
rithms for computing one-to-one shortest paths. Journal of Geographic Information and Deci-
sion Analysis, 4(2), 1–11.
Index

A Digraph, 19, 22, 38, 153, 155–157, 159

ADFSVRPSTW, 85, 152 Dispatcher, 64, 67, 108, 117, 145, 279, 280
Adler-32, 197 Diversification, 155, 185, 192, 195, 199–201,
Anticipation horizon, 143, 152, 153, 155, 272
163–165, 200, 206–208, 219, 220, 234, dosd, x, xiii, 11, 147, 249–254, 265, 275, 279
262, 270, 273, 278
Driver inconvenience, xiii, xv, 9, 11, 140, 146,
205, 226, 255, 258–261, 265, 266, 276,
B 279
Best theoretical tour plan, 152
DSVRP, 56
Dummy customer, ix, x, xii, 10, 114, 121, 122,
C
159, 168–172, 175, 177–183, 196–198,
Centralized coordination, 15, 16, 47, 51, 55,
208, 210, 217, 228–235, 240, 246, 255,
56, 58, 79, 83, 87, 269
256, 271–274, 279
Checksum, 197, 200, 272
Chi-Square Goodness-of-Fit test, 175–177, DVRP, 56, 79, 85, 86, 93, 95, 96, 101, 102,
204, 226–228, 230 116, 120, 121, 126, 129, 133, 137, 143,
Cluster selection, 172, 178, 231, 232 145, 269
Clustering method, 125, 174, 177, 232, 233, Dynamic parameter updating, 179
263
Concurrency, xii, 10, 96, 115, 143, 144, 150,
152, 158, 165, 206, 270 E
CRC-32, 197 En-route diversion, x, xiii, 9–11, 14, 64, 84,
Customer dissatisfaction, 2, 94, 95 91, 94, 96, 98, 101, 104, 106, 107, 114,
Customer inconvenience, xi, xii, 1–3, 6, 7, 117, 121, 127–129, 133, 140–142, 146,
9–11, 23, 37, 38, 59, 67, 71, 72, 78, 159, 203–205, 224–226, 255–262, 265,
81–83, 85–87, 90, 91, 99, 100, 107, 266, 270, 273, 275, 276, 279
110, 126, 137–140, 145, 156, 159, Expected future requests, 3, 10, 11, 57, 101,
161–163, 165, 168, 181, 198, 203, 205, 122, 124, 130, 133, 146, 151, 168–170,
207, 216, 218, 221–224, 236, 237, 172, 173, 181, 219, 226, 264, 268, 270
242–244, 247, 250, 251, 254–263,
265–270, 273–276, 278, 279
Customer loyalty, xi, 2, 6, 10, 72 F
Fingerprint, 197, 200, 272
D Forecasting approach, xii, 10, 13, 167, 169,
Decentralized coordination, 50–52 172, 176, 178, 183, 204, 209, 226–228,
DFSVRP, 56, 86, 87 231, 265, 277

F. Ferrucci, Pro-active Dynamic Vehicle Routing, Contributions to Management Science, 317

G RDOPG, xi, 2–10, 13, 15, 16, 18, 32, 37–39,

Generated request data class, 210, 211 47, 56–59, 67, 73, 75, 79, 81–89, 91,
Goodness-of-Fit test, 204, 226, 262, 273 93, 94, 97–99, 101, 106, 108, 110, 137,
140, 142–146, 149, 151, 152, 156, 157,
H 159–161, 164, 165, 168, 172, 186, 193,
High-quality dummy customers, 169 196, 197, 200, 203, 205–207, 218, 222,
High-quality stochastic knowledge, 219, 247 224, 234, 235, 249, 250, 262, 264,
267–272, 276, 278
I Real-time control approach, xii, xiii, 2, 3, 7,
Increased flexibility, 13, 57, 58, 97, 101, 129, 9–11, 13, 14, 41, 55, 56, 61–64, 83, 84,
137, 145, 167, 168, 170, 219, 270 87, 90, 91, 97, 124, 127, 128, 143, 146,
Intensification, 185, 191, 199 149, 151, 152, 154, 156, 158, 162–165,
167–170, 177, 179, 182, 183, 186, 197,
K 200, 203–205, 207, 208, 217–220, 223,
k-means, 173, 174 224, 226, 234, 241, 246–249, 251–255,
261–265, 268, 270–280
L Real-time control layer, 8, 55, 164, 165
Lateness penalty, 127, 136, 161–163, 210, 211, Real-time distribution of perishable goods, xi,
218, 270 2
Linear averaging, 212, 213 Real-world request data class, 210, 211, 236,
linear2X customer inconvenience function, 265, 273
163, 237, 251, 252, 257, 260, 261, 263, Regional variance of request arrival rates, 212
275 RegionDiversity, 212, 213, 242, 243, 249, 274
LNS, 31, 190, 194–196, 199 REL, 190, 192–194, 199
Relevant information, 38–45, 47–54, 56–59,
M 62, 63, 66, 73, 74, 77, 79, 81, 85–87,
MREL, 190, 194, 195, 199 90, 93, 142–144, 167, 206, 269
Relevant tour plan, 152, 154, 155, 165, 200,
O 207
Operational costs, 7, 35, 52, 61, 78, 95, 125, Request distribution, 10, 113, 171, 173, 211,
126 236
Operative layer, 8, 39, 165 Request-likely area, ix, xii, 10, 83, 84, 170,
171, 173, 178, 179, 181–183, 245, 256,
P 271
Past request information, iv, ix, xii, xv, 3, 10, Request reassignment, 224, 225, 259, 266
11, 13, 56, 78, 83–85, 87, 90, 91, 102, Response time, xi, 2, 3, 6, 10, 37, 38, 57,
108, 146, 147, 167–169, 171, 172, 174, 59–61, 67, 71, 72, 78, 81, 82, 85, 91,
176, 183, 210, 241, 249, 268, 270, 271, 96, 100, 101, 106, 107, 110, 117, 124,
276–278 125, 136–138, 145, 159, 161–163, 167,
Perishable goods, iv, xi, xv, 1–5, 10, 81, 85, 168, 207, 210, 211, 219, 224, 225,
267 237–239, 244, 246–248, 250, 251,
PROACTIVE10, 234, 236, 240–242, 245, 246, 254–257, 260, 267, 270, 274
251, 254–256, 258–261, 263, 266

Q S
Quadratic customer inconvenience function, Segment-based clustering, 10, 13, 168, 171,
161, 163, 224, 236, 251, 252, 254, 256, 210, 277
258, 260, 261, 263, 269, 275 Segments, 172–174, 178, 179, 183, 210,
231–234, 271, 277
R Service level agreement, xi, 6
RD, 212–215, 227, 228, 231, 233, 242, 243, Service quality, 2, 3, 6, 59, 62, 68, 70, 72, 137,
245–249, 255, 259, 263, 275 138, 140, 267, 268, 278
Index 319

Stochastic knowledge, iv, ix–xiii, 3, 9–13, 38, Theoretical tour plan, 152, 155, 188, 200
43–47, 49, 51–53, 55–58, 68, 71, 74, Three-dimensional convex area, 173
78, 79, 83–87, 90, 91, 94, 98, 100–103, Threshold-based approach, 257, 266, 276
108–110, 114–131, 133–137, 145–147, TimeDiversity, 212, 242, 243, 249, 274
151, 158, 165, 167–173, 177–179, 183, TIMELIMIT10, 220–225, 234, 236, 240–243,
198, 203–205, 210, 211, 215, 219, 223, 245, 246, 254–256, 258–263, 266, 273
228, 230–238, 240–242, 246, 247, TIMELIMIT10NODIV, 224, 225
249–251, 253, 254, 263–265, 268–279 Time-space Poisson process, 172, 183, 212,
Strategic layer, 7 271
Structural distortion, x, xii, 11, 13, 146, 147, Tour plan adaptation, xii, 10, 39, 41, 51, 52,
204, 205, 211, 216, 226–231, 234, 241, 55, 57, 62–64, 67, 70, 83, 96, 109, 127,
246, 248, 262–265, 273–275, 277, 279 143, 144, 152, 157, 158, 163, 168, 169,
Structural diversity, x, xii, xiii, 11, 13, 146, 219, 235, 270, 271
147, 204, 205, 211–216, 220, 232–235,
240, 241, 244–250, 252, 254, 262–265, V
272, 274, 275, 279 Vehicle scheduling decisions, 34, 35, 84
Structural quality, xii, xiii, 9, 11, 13, 14, 56, Visual analysis, xiii, 147, 238, 244, 265, 274,
146, 147, 174, 204, 205, 212, 238, 241, 275
246, 248–251, 253, 254, 262–266, 268, VRP, xvi, 15, 16, 18, 20, 21, 24, 27, 29, 30,
272, 275, 279 32–36, 39, 47, 49–51, 67, 79, 85, 86,
Subareas, 172, 173, 183, 210, 227, 240, 93, 97, 98, 103, 121, 142, 195, 269, 279
244–246, 250, 265, 271, 274, 275
Subsequent delivery of newspapers, xi, xii, 3, W
4, 8, 9, 72, 73, 82, 85, 159, 211, 262, WTI, 190–192, 199
272, 278
X
T X 2 test statistic, 176
Tabu list, 185–187, 189, 196–198, 200, 272 XBT, 190, 195, 196, 199
Tactical layer, 8
TD, 212–215, 227, 228, 230, 231, 233, 235, Z
246–249, 255, 257, 258, 263, 275 ZEROTIME, 220–223

E.H. Carr - Notes
71% (28)
E.H. Carr - Notes
6 pages
Free Will in Philosophical Theology Kevin Timpe Instant Download
100% (1)
Free Will in Philosophical Theology Kevin Timpe Instant Download
38 pages
Innovations in Last Mile Delivery System
No ratings yet
Innovations in Last Mile Delivery System
227 pages
Complete Download Reconsidering Identity Economics: Human Well-Being and Governance 1st Edition Laszlo Garai (Auth.) PDF All Chapters
100% (2)
Complete Download Reconsidering Identity Economics: Human Well-Being and Governance 1st Edition Laszlo Garai (Auth.) PDF All Chapters
55 pages
Donald Davendra (Ed.) - Traveling Salesman Problem, Theory and Applications - InTech (2010) PDF
No ratings yet
Donald Davendra (Ed.) - Traveling Salesman Problem, Theory and Applications - InTech (2010) PDF
336 pages
Sociology of The Quran Part 2 (Ayatullah Murtadha Mutahhari) - Siatbook
No ratings yet
Sociology of The Quran Part 2 (Ayatullah Murtadha Mutahhari) - Siatbook
47 pages
Is Freedom Real
No ratings yet
Is Freedom Real
2 pages
Two Senses of Freedom Epictetus On The Will - Bra Then
No ratings yet
Two Senses of Freedom Epictetus On The Will - Bra Then
35 pages
Scholars Cup Guiding Questions Answer Key
No ratings yet
Scholars Cup Guiding Questions Answer Key
16 pages
Esu Yoruba God, Power and TheImaginative Frontiers.
No ratings yet
Esu Yoruba God, Power and TheImaginative Frontiers.
6 pages
(Communications in Cybernetics, Systems Science and Engineering) Tsan-Ming Choi - Fast Fashion Systems - Theories and Applications
No ratings yet
(Communications in Cybernetics, Systems Science and Engineering) Tsan-Ming Choi - Fast Fashion Systems - Theories and Applications
158 pages
Eindhoven University of Technology: Award Date: 2011
No ratings yet
Eindhoven University of Technology: Award Date: 2011
108 pages
Metaheuristics For Enterprise Data - Kaustubh Vaman Sakhare
No ratings yet
Metaheuristics For Enterprise Data - Kaustubh Vaman Sakhare
159 pages
Big Data Analytics in Operation Management
100% (1)
Big Data Analytics in Operation Management
35 pages
Understanding The Digital and AI Transformation
No ratings yet
Understanding The Digital and AI Transformation
240 pages
Supply Chain Coordination Mechanisms - New Approaches For Collaborative Planning (PDFDrive)
No ratings yet
Supply Chain Coordination Mechanisms - New Approaches For Collaborative Planning (PDFDrive)
234 pages
Trust between Cooperating Technical Systems: With an Application on Cognitive Vehicles
From Everand
Trust between Cooperating Technical Systems: With an Application on Cognitive Vehicles
Walter Bamberger
No ratings yet
(Annals of Information Systems 10) Marco Caserta, Stefan Voß (auth.), Vittorio Maniezzo, Thomas Stützle, Stefan Voß (eds.) - Matheuristics_ Hybridizing Metaheuristics and Mathematical Programming-Spri
No ratings yet
(Annals of Information Systems 10) Marco Caserta, Stefan Voß (auth.), Vittorio Maniezzo, Thomas Stützle, Stefan Voß (eds.) - Matheuristics_ Hybridizing Metaheuristics and Mathematical Programming-Spri
283 pages
Why We Should Reject Semiretributivism and Be Skeptics About Basic Desert Moral
No ratings yet
Why We Should Reject Semiretributivism and Be Skeptics About Basic Desert Moral
31 pages
Book
No ratings yet
Book
156 pages
The Hijacked Brain, Is It True
No ratings yet
The Hijacked Brain, Is It True
4 pages
Strategic Thinking in Turbulent Times
No ratings yet
Strategic Thinking in Turbulent Times
7 pages
PDF
No ratings yet
PDF
580 pages
mpc96 62a PDF
No ratings yet
mpc96 62a PDF
102 pages
Simio and Simulation 6e
No ratings yet
Simio and Simulation 6e
547 pages
Sharda Dss11e ch01
100% (2)
Sharda Dss11e ch01
60 pages
Vehicle Routing Problem PDF
100% (1)
Vehicle Routing Problem PDF
152 pages
Supply Chain Performance Evaluation: Farhad Hosseinzadeh Lotfi Tofigh Allahviranloo Morteza Shafiee Hilda Saleh
No ratings yet
Supply Chain Performance Evaluation: Farhad Hosseinzadeh Lotfi Tofigh Allahviranloo Morteza Shafiee Hilda Saleh
453 pages
OperationsResearch 2
No ratings yet
OperationsResearch 2
294 pages
DLMBAEOIM01 Course Book
No ratings yet
DLMBAEOIM01 Course Book
124 pages
A Comprehensive Survey On Pretrained Foundation Models
No ratings yet
A Comprehensive Survey On Pretrained Foundation Models
97 pages
Agm Thesis Signed
No ratings yet
Agm Thesis Signed
57 pages
Swords 2019
No ratings yet
Swords 2019
78 pages
Collingwood
No ratings yet
Collingwood
21 pages
Free Will - Internet Encyclopedia of Philosophy
No ratings yet
Free Will - Internet Encyclopedia of Philosophy
28 pages
Thesis - Heuristic and Exact Algorithms For Vehicle Routing Problems
No ratings yet
Thesis - Heuristic and Exact Algorithms For Vehicle Routing Problems
256 pages
Tutorial10 Multi Vehicle Routing With Time Windows PDF
100% (1)
Tutorial10 Multi Vehicle Routing With Time Windows PDF
114 pages
Research Paper
No ratings yet
Research Paper
26 pages
2014 Book SupplyChainManagementUnderFuzz
No ratings yet
2014 Book SupplyChainManagementUnderFuzz
671 pages
MCQ On Block Chain 5eea6a0d39140f30f369e30d
No ratings yet
MCQ On Block Chain 5eea6a0d39140f30f369e30d
20 pages
Bricmont, J. (2017) - Quantum Sense and Nonsense. Basel Springer (c31)
No ratings yet
Bricmont, J. (2017) - Quantum Sense and Nonsense. Basel Springer (c31)
287 pages
Arne Schuldt - Multiagent Coordination Enabling Autonomous Logistics
No ratings yet
Arne Schuldt - Multiagent Coordination Enabling Autonomous Logistics
284 pages
AnyLogistix Book Ivanov
No ratings yet
AnyLogistix Book Ivanov
157 pages
Xian Petri Valk
No ratings yet
Xian Petri Valk
66 pages
Tibor R. Machan-A Primer On Ethics (1997)
100% (1)
Tibor R. Machan-A Primer On Ethics (1997)
201 pages
Supply Chain Analytics
No ratings yet
Supply Chain Analytics
25 pages
Lecture 6 - Improving Local Search
No ratings yet
Lecture 6 - Improving Local Search
28 pages
CHAPTER 8 Designing Distribution Networks and Applications To E Business WHN 2020
No ratings yet
CHAPTER 8 Designing Distribution Networks and Applications To E Business WHN 2020
23 pages
The Pollution-Routing Problem
No ratings yet
The Pollution-Routing Problem
19 pages
A Guide To Vehicle Routing Heuristics PDF
No ratings yet
A Guide To Vehicle Routing Heuristics PDF
12 pages
A Five-State Financial Distress Prediction Model
No ratings yet
A Five-State Financial Distress Prediction Model
13 pages
The Real-Time Vehicle Routing Problem
No ratings yet
The Real-Time Vehicle Routing Problem
20 pages
A New Capacitated Vehicle Routing Proble PDF
No ratings yet
A New Capacitated Vehicle Routing Proble PDF
20 pages
Ivanov Supply Chain Simulation and Optimization With Alx
No ratings yet
Ivanov Supply Chain Simulation and Optimization With Alx
210 pages
Post Tes
40% (5)
Post Tes
13 pages
Introduction To Operations and Supply Chain Management
100% (1)
Introduction To Operations and Supply Chain Management
46 pages
GHSCPSM Say1q2 Final
No ratings yet
GHSCPSM Say1q2 Final
48 pages
Real Estate Report in India
No ratings yet
Real Estate Report in India
62 pages
Digital Twin For Logistics
100% (1)
Digital Twin For Logistics
33 pages
Advanced Operation Research
No ratings yet
Advanced Operation Research
6 pages
Materials and Financial
100% (1)
Materials and Financial
267 pages
Smart and Sustainable Supply Chain and Logistics - Challenges, Methods and Best Practices
No ratings yet
Smart and Sustainable Supply Chain and Logistics - Challenges, Methods and Best Practices
282 pages
Robust Ordinal Regression
No ratings yet
Robust Ordinal Regression
429 pages
Karma in Buddhism
No ratings yet
Karma in Buddhism
15 pages
1.1: Overview of Automobile Industry in India: Chapter 1: Industrial Profile
No ratings yet
1.1: Overview of Automobile Industry in India: Chapter 1: Industrial Profile
93 pages
Why Do We Find Randomness in Nature?: Paul Almond
No ratings yet
Why Do We Find Randomness in Nature?: Paul Almond
9 pages
10.1007@978 3 030 35102 1
No ratings yet
10.1007@978 3 030 35102 1
293 pages
Deep Learning Approaches For Network Int
No ratings yet
Deep Learning Approaches For Network Int
116 pages
015.search Formulation Problems Basic Strategies
No ratings yet
015.search Formulation Problems Basic Strategies
121 pages
MngtScience Week1 2
No ratings yet
MngtScience Week1 2
12 pages
A Multi-Size Compartment Vehicle Routing Problem For Multi-Product Distribution: Models and Solution Procedures
No ratings yet
A Multi-Size Compartment Vehicle Routing Problem For Multi-Product Distribution: Models and Solution Procedures
21 pages
Effects of Using Internet of Things (I.o.T) Systems in Transportation On Customer Service Delivery
No ratings yet
Effects of Using Internet of Things (I.o.T) Systems in Transportation On Customer Service Delivery
6 pages
Fundamentals of Management: Foundations of Decision Making
100% (1)
Fundamentals of Management: Foundations of Decision Making
39 pages
Decision-Making: The Essence of The Manager's Job
100% (1)
Decision-Making: The Essence of The Manager's Job
53 pages
Probabilistic Subsurface Forecasting - What Do We Really Know?
No ratings yet
Probabilistic Subsurface Forecasting - What Do We Really Know?
6 pages
Final Paper On Notes From The Underground
No ratings yet
Final Paper On Notes From The Underground
5 pages
Lecture Notes in Logistics
No ratings yet
Lecture Notes in Logistics
144 pages
Naturalism
No ratings yet
Naturalism
15 pages
ICT Business Models - 0
No ratings yet
ICT Business Models - 0
16 pages
Realism and Objectivism in Quantum Mechanics Vassilios Karakostas
No ratings yet
Realism and Objectivism in Quantum Mechanics Vassilios Karakostas
20 pages
CHP 2 Making Decisions
100% (1)
CHP 2 Making Decisions
9 pages
Demand Analysis & Forecasting - The Car Industry: Presented by
No ratings yet
Demand Analysis & Forecasting - The Car Industry: Presented by
31 pages
Continuous and Discrete Model
No ratings yet
Continuous and Discrete Model
19 pages
Introduction To The Philosophy of The Human Person
No ratings yet
Introduction To The Philosophy of The Human Person
23 pages
Undergraduate Text
No ratings yet
Undergraduate Text
351 pages
Simplifying Complexity-A Review of Complexity Theory
No ratings yet
Simplifying Complexity-A Review of Complexity Theory
10 pages
Process Scheduling Algorithms
No ratings yet
Process Scheduling Algorithms
3 pages
An Improved Clarke and Wright Algorithm To Solve The Capacitated Vehicle Routing Problem
No ratings yet
An Improved Clarke and Wright Algorithm To Solve The Capacitated Vehicle Routing Problem
3 pages
Essay On Freedom and Determinism
100% (3)
Essay On Freedom and Determinism
3 pages
Sap Case Study Divekar November 2011
No ratings yet
Sap Case Study Divekar November 2011
2 pages