Copyright Undertaking

This thesis is protected by copyright, with all rights reserved.

By reading and using the thesis, the reader understands and agrees to the following terms:

1. The reader will abide by the rules and legal ordinances governing copyright regarding the
use of the thesis.

2. The reader will use the thesis for the purpose of research or private study only and not for
distribution or further reproduction or any other purpose.

3. The reader agrees to indemnify and hold the University harmless from and against any loss,
damage, cost, liability or expenses arising from copyright infringement or unauthorized
usage.

IMPORTANT
If you have reasons to believe that any materials in this thesis are deemed not suitable to be
distributed in this form, or a copyright owner having difficulty with the material being included in
our database, please contact [email protected] providing details. The Library will look into
your claim and consider taking remedial action upon receipt of the written requests.

Pao Yue-kong Library, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong

http://www.lib.polyu.edu.hk
 OPTIMIZATION APPLICATIONS IN
MARITIME LOGISTICS AND OPERATIONS

WANG KAI

PhD

The Hong Kong Polytechnic University

2019
The Hong Kong Polytechnic University

Department of Logistics and Maritime Studies

Optimization Applications in Maritime Logistics and

Operations

Wang Kai

A thesis submitted in partial fulfillment of the

requirements for the degree of
Doctor of Philosophy

July 2019
 Certificate of Originality

I hereby declare that this thesis is my own work and that, to the best of my knowledge
and belief, it reproduces no material previously published or written, nor material that
has been accepted for the award of any other degree or diploma, except where due
acknowledgment has been made in the text.

_________________________ (Signed)

_________________________ (Name of student)

Optimization Applications in Maritime Logistics and Operations i

 Abstract

This thesis investigates four optimization problems in maritime logistics and

operations, where the first two problems are related to container ships that transport
cargo and the other two problems are related to cruise ships that transport passengers.
The first problem concerns the container ship type decision. It aims to determine the
ship types deployed on shipping routes while taking the possible empty container
repositioning and the usage of novel foldable containers into account. The second
problem addresses the optimal reefer slot conversion for container freight
transportation. It optimizes the number of reefer slots in a fleet of container ships
deployed on a shipping route and re-optimizes the sequence of these ships to maximize
the revenue. The third problem investigates the cruise itinerary schedule design for a
cruise ship. It determines the visiting sequence of several ports of call and the
corresponding arrival and departure times at the ports, so as to maximize the monetary
value of cruise passengers’ utility minus operations costs. The fourth problem focuses
on cruise service planning. It proposes a solution approach to schedule available cruise
services for a cruise ship over a planning horizon while considering berth availability
at ports of call and decreasing marginal profit for each cruise service. To solve the four
problems, different operations research methods are proposed, such as network flow
modeling, mixed integer linear programming, simulation algorithms, dynamic
programming, model linearization techniques, and heuristic algorithms. By referring
to real-world data, extensive numerical experiments are conducted to validate the
effectiveness of the proposed methods. Some potential managerial insights behind the
problems are also revealed.

Optimization Applications in Maritime Logistics and Operations ii

 Acknowledgments

I would like to thank my supervisor Dr Shuaian Wang for his full support all the
time and thank my thesis board members, Professor Miao Song, Professor Xiangtong
Qi, Professor Yanzhi Li, and Dr Shuaian Wang for their time and valuable feedback.
I would also like to thank the financial support from the Hong Kong Ph.D. Fellowship
Scheme 2016/2017 and thank the research funding support from the Department of
Logistics and Maritime Studies, The Hong Kong Polytechnic University.

Optimization Applications in Maritime Logistics and Operations iii

 Table of Contents

Certificate of Originality ........................................................................................................... i

Abstract .................................................................................................................................... ii
Acknowledgments................................................................................................................... iii
Table of Contents .................................................................................................................... iv
List of Figures ......................................................................................................................... vi
List of Tables ........................................................................................................................ viii
List of Abbreviations ................................................................................................................x
Chapter 1: Introduction ...................................................................................... 1
1.1 Background .....................................................................................................................1
1.2 Thesis outline ..................................................................................................................2
Chapter 2: Container Ship Type Decision ......................................................... 3
2.1 Introduction ....................................................................................................................3
2.2 Literature review.............................................................................................................7
2.3 Problem description ........................................................................................................9
2.4 Model formulation ........................................................................................................12
2.5 Solution approach .........................................................................................................22
2.6 Computational experiment............................................................................................31
2.7 Conclusion ....................................................................................................................43
2.8 References ....................................................................................................................44
Chapter 3: Container Reefer Slot Conversion ................................................ 47
3.1 Introduction ..................................................................................................................47
3.2 Estimating the profit for a given string .........................................................................50
3.3 Optimizing the sequence of ships in a string ................................................................54
3.4 Optimizing reefer slot conversion ................................................................................62
3.5 Conclusion ....................................................................................................................67
3.6 References ....................................................................................................................68
Chapter 4: Cruise Itinerary Schedule Design ................................................. 71
4.1 Introduction ..................................................................................................................71
4.2 Literature review...........................................................................................................74
4.3 Problem description ......................................................................................................76
4.4 Solution approach for CISD .........................................................................................83
4.5 Computational experiment............................................................................................88
4.6 Utility estimation by marketing techniques ..................................................................99
4.7 Conclusion ..................................................................................................................102

Optimization Applications in Maritime Logistics and Operations iv

4.8 References ..................................................................................................................103
Chapter 5: Cruise Ship Service Planning ...................................................... 109
5.1 Introduction ................................................................................................................109
5.2 Literature review.........................................................................................................111
5.3 Mathematical model ...................................................................................................113
5.4 Complexity analysis and comparison with heuristics .................................................118
5.5 Computational experiment..........................................................................................125
5.6 Conclusion ..................................................................................................................135
5.7 References ..................................................................................................................136
Chapter 6: Conclusions ................................................................................... 139
Appendices .............................................................................................................. 141
Appendix A : Specification of standard containers and foldable containers ........................141
Appendix B : Proof for the results of Figure 2.10.................................................................141
References ............................................................................................................... 143

Optimization Applications in Maritime Logistics and Operations v

 List of Figures

Figure 2.1: Four foldable containers and a standard container (Shintani et al.,
2010) .............................................................................................................. 4
Figure 2.2: A transpacific shipping service route operated by CMA CGM
(CMA CGM, 2017) ........................................................................................ 9
Figure 2.3: A sub-network for standard containers without considering the
short-term container leasing ......................................................................... 15
Figure 2.4: The sub-network defined for the preliminary NF model ......................... 18
Figure 2.5: A sub-network for foldable containers .................................................... 20
Figure 2.6: Connections between standard containers and foldable containers in
TNF model ................................................................................................... 20
Figure 2.7: An example of the pivot cycle W ............................................................ 24
Figure 2.8: Adding dummy nodes to the network ..................................................... 26
Figure 2.9: Three selected shipping services routes operated by CMA CGM
(CMA CGM, 2017) ...................................................................................... 32
Figure 2.10: Effect of the long-term leasing cost on the foldable container
usage ............................................................................................................ 38
Figure 2.11: Effect of the folding and unfolding cost on foldable container
usage ............................................................................................................ 39
Figure 2.12: Long-term leasing cost dependent sensitivity to the folding and
unfolding cost............................................................................................... 41
Figure 3.1: A sequence of container ships in a string ................................................ 50
Figure 3.2: The shipping routes involved in three case studies (CMA CGM,
2017) ............................................................................................................ 60
Figure 4.1: Itinerary of 7 Day Western Caribbean of Carnival (Carnival Cruise
Line, 2016) ................................................................................................... 77
Figure 4.2: Utility distribution for one day ................................................................ 78
Figure 4.3: Utility distributions and time spent at two ports ..................................... 79
Figure 4.4: Statistic on trips for the Carnival Vista (Cruise Ship Schedule,
2016) ............................................................................................................ 84
Figure 4.5: City ports in “14 Night Singapore to Fremantle Cruise” (Google
Map, 2018a) ................................................................................................. 92
Figure 4.6: City ports around the Caribbean Sea (Google Map, 2018b) ................... 95
Figure 4.7: Comparisons between the designed and actual itineraries ...................... 96
Figure 5.1: Concavity of the profit function ............................................................ 115
Figure 5.2: Locations of the port cities .................................................................... 129
Figure 5.3: The robustness test on 𝒑 ........................................................................ 131

Optimization Applications in Maritime Logistics and Operations vi

Figure 5.4: The robustness test on 𝒂 ........................................................................ 131
Figure 5.5: Upper bound 𝒙𝒓 obtained by the greedy algorithm and Eq. (5.9) ......... 133
Figure 5.6: Number of repeats of cruise services under different berth
availability scenarios .................................................................................. 134

Optimization Applications in Maritime Logistics and Operations vii

 List of Tables

Table 2.1: Data for relevant container costs............................................................... 33

Table 2.2: Candidate ship type and fixed operation cost ........................................... 33
Table 2.3: Comparing the proposed method with the MILP model by CPLEX
solver ............................................................................................................ 34
Table 2.4: Comparing the proposed model and the traditional model without
empty container allocation ........................................................................... 35
Table 2.5: Comparison between the proposed model and the model that does
not use the foldable containers ..................................................................... 37
Table 2.6: Total cost saving after using foldable containers under different
long-term leasing cost .................................................................................. 39
Table 2.7: Sensitivity analysis on the number of weeks allowed for returning
empty containers .......................................................................................... 42
Table 3.1: The capacities of the ships in the fleet used for the Yangtze Service
route ............................................................................................................. 48
Table 3.2: Comparison of computation times for different values of the
parameter 𝜷 .................................................................................................. 58
Table 3.3: The relevant input costs used in the case studies ...................................... 59
Table 3.4: Information on the shipping routes and ships deployed ........................... 60
Table 3.5: Ships deployed in the three shipping routes ............................................. 61
Table 3.6: Results of the string optimization process ................................................ 62
Table 3.7: Reefer slot conversion results for the three shipping routes ..................... 65
Table 3.8: Sequence re-optimization and slot conversion ......................................... 66
Table 3.9: Basic information on the two shipping routes .......................................... 67
Table 3.10: Results obtained using string optimization and slot conversion ............. 67
Table 4.1: Comparison between different settings of the time unit ........................... 90
Table 4.2: Computational efficiency with and without using enumeration
improving ..................................................................................................... 91
Table 4.3: Information of selected ports in a real case .............................................. 92
Table 4.4: Sensitivity analysis on fuel price .............................................................. 93
Table 4.5: Sensitivity analysis on minimum staying hours........................................ 93
Table 4.6: Sensitivity analysis on opening hours ....................................................... 94
Table 4.7: Further analysis of fuel price .................................................................... 96
Table 4.8: Sensitivity analysis on the utility distributions ......................................... 97
Table 4.9: Comparisons between the proposed method and two heuristics .............. 98
Table 4.10: Attributes and levels used in the conjoint analysis ............................... 100

Optimization Applications in Maritime Logistics and Operations viii

Table 5.1: The itinerary for the cruise service ......................................................... 114
Table 5.2: Comparison between the model 𝑭𝑴𝟐′ with and without Constraints
(14) ............................................................................................................. 127
Table 5.3: Comparison between the two linear models ........................................... 127
Table 5.4: Comparison between the model 𝑭𝑴𝟐′′ and two myopic rules .............. 128
Table 5.5: Information on the cruise services .......................................................... 129
Table 5.6: Outputs of the model under different berth availability scenarios .......... 134

Optimization Applications in Maritime Logistics and Operations ix

 List of Abbreviations

 Twenty-foot equivalent units (TEUs)

 Origin-destination (O–D)
 Network flow (NF)
 Mathematical programming (MP)
 Mixed integer linear programming (MILP)
 Cruise itinerary schedule design (CISD)
 Revenue management (RM)
 Vehicle routing problems with time windows (VRPTW)
 Dynamic programming (DP)
 Cruise service planning (CSP)

Optimization Applications in Maritime Logistics and Operations x

Chapter 1: Introduction

This thesis includes four essays with a focus on maritime logistics and operations. It
studies four optimization application problems in the field, where the first two problems are
related to container ships that transport cargo and the other two problems are related to cruise
ships that ship cruise passengers. Different operations research methods are adapted to provide
effective solution approaches for optimization problems. Extensive experiments are conducted
to draw managerial insights on the problems.

1.1 BACKGROUND

In the container shipping industry, container ships are used to transport cargo among
seaports in the world, which contributes to the majority of shipping volume in the global trade.
The cost-effectiveness of sea transportation is the major competitiveness compared with other
transportation modes, such as road and air transportation. Thus, shipping companies endeavor
to optimize their operations such that they can reduce their operations costs. This is especially
important now since the industry is experiencing a transformation to an environmentally
friendly industry (Xia et al., 2019), which request more investments from the shipping
companies.
In the cruise shipping industry, cruise ships serve the purposes of providing pleasure
voyages. Once cruise passengers are on aboard, the voyages and the activities by the cruise
ships bring utility experience to the passengers, for which the cruise ships earn profits. The
transportation is not the major purpose of the cruise shipping since the cruise passengers
normally will return to the same port where they embark. The cruise industry is a very
promising industry and it keeps booming in recent years, especially for the Asia market (Wang
et al., 2016; Wang et al., 2017). However, the studies on this industry are rare in the literature,
but the operations of the industry are very urgent to be investigated.
The container shipping industry and the cruise shipping industry share many similarities
since they all belong to the maritime logistics and operations. They all follow some pre-
determined schedules to traverse a set of fixed ports of call, and they need to dwell at berths
of ports to conduct some on-shore activities. The major operational cost of the container and
cruise ships is the fuel consumption cost. However, they have several remarkable differences.
For instance, the container shipping aims to transport the cargo from its origin port to its
destination port, but the cruise shipping aims to provide voyage service for the cruise
passengers on aboard.

Chapter 1: Introduction 1
 In this thesis, four operation optimization problems will be addressed and solved for the
container shipping and the cruise shipping, from which some similarities and differences
between the two industries can be sensed. The first two problems aim to optimize the
operations for container ships such that the operational costs can be reduced. The other two
problems aim to improve the operations for cruise ships such that the operating profits can be
increased. Different operations research methods are adopted to solve the problems, and
extensive numerical experiments are conducted to explore the hidden insights behind the
problems.

1.2 THESIS OUTLINE

The remainder of this thesis is organized as follows. Chapter 2 and 3 are related to the
operations of container ships, and Chapter 4 and 5 are related to the operations of cruise ships.
In specific, Chapter 2 presents a study of ship type decision considering empty container
repositioning and foldable containers. Chapter 3 studies the optimal reefer slot conversion for
container freight transportation. Chapter 4 considers the problem of cruise itinerary schedule
design for a cruise ship. Chapter 5 addresses the cruise service planning considering berth
availability and decreasing marginal profit. Chapter 6 concludes the thesis.

Chapter 1: Introduction 2
Chapter 2: Container Ship Type Decision

This chapter addresses a ship fleet type decision problem with considering empty
container repositioning and foldable containers. This decision problem determines the capacity
of container ships deployed in a given shipping route. In reality, the ship fleet deployment
decision is usually made empirically according to laden container transportation and does not
consider the possible empty container repositioning. Meanwhile, a novel mode ‘foldable
container’ has shown its economic and logistical viability in recent years. This study hence
also considers the use of foldable containers and aims to find under what conditions, a shipping
liner needs to use the foldable containers in its liner shipping services. To solve the problem,
we formulate a network flow model with a revised network simplex algorithm, based on which
an exact solution approach is designed to determine the optimal ship type. A mixed-integer
programming model is also formulated for the problem. Numerical experiments based on real-
world voyages are conducted to find some managerial implications on the ship fleet
deployment and the foldable container usage.

2.1 INTRODUCTION

A shipping liner normally operates weekly-serviced ship routes with fixed schedules to
transport containers. Given a shipping route, a shipping liner deploys a fleet of container ships
for the operation over a planning horizon, e.g., six months. One of the critical decisions for the
shipping liner on the fleet is ship type decision, which determines the capacity of container
ships of the fleet deployed on the shipping route. Empirically, the shipping liner deploys a
suitable fleet type of ships on each route based on the laden container transportation over the
planning horizon, which guarantees that the deployed ships have the capacity to accommodate
all the laden containers in all the voyages. Under this circumstance, the shipping liner would
not deploy a ship fleet with a larger capacity as it increases the fixed operating cost for
maintaining the fleet. It is reasonable for the shipping liner to make such decision only
considering the laden container transportation. However, if we further consider the empty
container repositioning on the route, the ship type decision can be more complicated.
The empty container repositioning originates from the imbalance of container flow between
different regions in liner shipping routes. Take the trans-Pacific trade lane for example:
according to UNCTAD (2016), in 2015, the annual container flow from Asia to North America
(i.e., the eastbound) was around 15.8 million twenty-foot equivalent units (TEUs), and the
container flow in the opposite westbound direction was 7.4 million TEUs, which generated

Chapter 2: Container Ship Type Decision 3

the imbalance of container flow for 8.4 million TEUs. This imbalance contributes to
tremendous empty container accumulation in import-dominant areas (North America) and the
serious empty container shortage in export-dominant areas (Asia). This leads to a critical
problem on the empty container availability for the laden container transportation consignment
in those export-dominant areas or ports. In those export-dominant ports, the arriving laden
containers from incoming ships become empty and are stored in depots after devanning, which
can only fulfill part of the empty container requirement for the sake of their export-dominant
characters. As a result, the empty container repositioning from the surplus ports (i.e., import-
dominant ports) to the deficit ports (i.e., export-dominant ports) becomes necessary. However,
the empty containers repositioned between the ports occupy the capacity of the container ships
traversing the corresponding voyages in the shipping route. Henceforth, the ship type decision
is no longer straightforward when considering the empty container repositioning.

Figure 2.1: Four foldable containers and a standard container (Shintani et al., 2010)

Storing empty containers in the depots and repositioning empty containers among the ports
inevitably incur storage cost and repositioning cost for the shipping liner, respectively (Lee
and Yu, 2012). To reduce costs, the usage of foldable containers is an effective method. The
idea of foldable containers is not so new, and several container companies have developed
foldable containers, such as Fallpac AB and Holland Container Innovation. Those foldable
containers have equivalent storage capacity and size as standard containers and a foldable
container only occupies one-quarter storage space of a standard container in folded status, as
shown in Figure 2.1 (See Appendix A for the specification comparison between a standard
container and a foldable container). After becoming empty, the foldable containers will be in
folded status for the storage in the depots or for the repositioning to other ports. As four
foldable empty containers in the folded status equal one standard empty container, it saves
75% storage space by using foldable containers, which leads to significant decrease in the
repositioning cost and the storage cost. However, using the foldable containers could incur

Chapter 2: Container Ship Type Decision 4

additional costs for the shipping liner. Firstly, the purchasing fee or long-term leasing cost of
the foldable containers is higher than that of the standard containers. Secondly, folding and
unfolding processes involve labor cost in the ports for the empty container repositioning.
Therefore, there is a trade-off by using foldable containers between reducing the storage cost
and the repositioning cost and incurring the additional costs.
Currently, foldable containers are not widely used in the liner shipping industry and
stakeholders of the industry are trying to make the foldable containers prevalent. Here, we
summary two practical concerns that may impede the usage of foldable containers at present.
Firstly, maintaining a foldable container fleet needs a considerable investment at the first
phrase, as the building cost of a foldable container is double as that of a standard container
(Goh et al., 2016). Considering the shipping market is experiencing a depression (UNCTAD,
2016), the majority of shipping lines may not have enough funding to replace the standard
containers in their container fleet with foldable containers. Secondly, folding and unfolding
activities in container terminals incur additional labor operations. Henceforth, container
terminals and shipping lines need to negotiate a comprehensive agreement on maintaining the
operations and training technicians, which may not be achieved at the moment. Although these
concerns can exist in practices, the usage of foldable containers is still promising in the near
future. We take Holland Container Innovations (HCI), the manufacturer of the 4FOLD
foldable container, as a typical example to illustrate industry trends of using foldable
containers. Holland Container Innovations (2017a) reported that some major shipping lines
(e.g., APL, Samudera Indonesia and Seatrade) have used 4FOLD foldable containers in their
shipping routes, and an increasing number of shipping lines have signed the contracts with
HCI to promote the usage of the foldable containers, such as Emirates Shipping Line (Word
Cargo News, 2017). Meanwhile, HCI is providing the folding training programs for some
container terminals around the world in the preparation for using foldable containers, such as
Tetris Container Terminal in Moscow, Ljubljana Container Terminal in Slovenia and Qingdao
Shitengkeyun Depot.
Motivated by the above problem justifications and industry trends, our study aims to solve
a problem of ship type decision considering the empty container repositioning and foldable
containers, in order to minimize the total cost that occurs in a given planning horizon for the
shipping route. The problem focuses on related decisions in both tactical and operational
levels. In the tactical level, it first determines the ship type of the container ship fleet (denoted
as ship type decision), which decides the capacity (in TEUs) of container ships deployed in
the shipping route. Then, the problem determines the number of foldable and standard

Chapter 2: Container Ship Type Decision 5

containers leased (or kept) in the ports initially for the usage of the planning horizon, which is
a container fleet sizing in essential (denoted as long-term container leasing). In the operational
level, upon each weekly service, if there are empty containers surplus in some ports, the
problem decides the number of empty containers that the visiting ship should reposition to
other deficit ports. In case of the empty container deficit, the shipping liner can lease empty
containers in origin ports and return them in destination ports (denoted as short-term container
leasing) to fulfill the transportation consignments. Here, we summarize the empty container
repositioning and the container fleet sizing as the empty container allocation.
If the empty container repositioning is not involved, the ship type decision is to guarantee
that the deployed ships have the capacity to accommodate all the laden container transportation
and a ship fleet with a larger capacity will not be an option. However, involving empty
container repositioning complicates the ship fleet deployment. The empty container
repositioning provides the shipping liner with the motivation to deploy a ship fleet with a larger
capacity. Although it will raise the fixed operation cost, it gives the shipping liner more
flexibility and capacity to reposition empty containers among ports. Meanwhile, the container
fleet sizing intertwines with the ship fleet deployment and the empty container repositioning.
The container fleet sizing determines the total number of containers flowing in the planning
horizon. As those containers would be either in the depots or on the ships, the effects on the
ship type decision are inescapable. Normally, the larger the container fleet, the more the empty
containers repositioned.
Based on the above analysis, this paper presents an explorative study on the problem of ship
type decision considering the empty container repositioning and foldable containers. In the
study, we find that given the ship type with a specific capacity, the problem transfers to a
nonstandard minimum cost flow problem, which solves the container fleet sizing and the
empty container repositioning. For the nonstandard minimum cost flow problem, we build a
network flow model by constructing a network for the flow of empty containers. Due to the
usage of standard and foldable containers, the network inevitably has some parallel arcs
sharing the same capacity restriction such that one cannot apply some standard network
algorithms (e.g., the network simplex algorithm (Ahuja et al., 1993)) to solve it. To tackle the
above issue, we propose a revised network simplex algorithm (an exact algorithm) by
introducing dynamic capacity restrictions and revising the pivot operation of the standard
algorithm. Based on the reduced costs derived from the revised network simplex algorithm,
we propose a solution approach to determine the optimal ship type. By applying the solution

Chapter 2: Container Ship Type Decision 6

approach, we conduct extensive experiments to find insights on the ship type decision and the
foldable container usage.
The remainder of this chapter is as follows: Section 2.2 reviews some related works. Section
2.3 elaborates the background information and decisions on the problem. Section 2.4 presents
a network flow model given the capacity of a ship. Section 2.5 elaborates the developed
solution approach that embeds a revised network simplex algorithm. Section 2.6 shows some
experiments for finding insights on the ship type decision and the foldable container usage.
The last section presents some conclusions.

2.2 LITERATURE REVIEW

There have been numerous studies related to the ship fleet deployment and empty container
repositioning problems. For the ship fleet deployment, to the best of our knowledge, Perakis
and Jaramillo (1991) was the first study to address it, in which they built integer linear
programming models for the problem. Thereafter, there were generally two types of studies
on the problem. One type assumes that the container shipment demand is deterministic
(Gelareh and Meng, 2010; Brouer et al., 2013; Plum et al., 2013). For instance, Gelareh and
Meng (2010) developed a mixed integer nonlinear programming model for a short-term fleet
deployment problem, in which the optimal vessel speeds for different vessel types on different
routes are considered. The other type of studies relaxed the deterministic demand assumption
and treated the container shipment demand in a stochastic manner (Meng and Wang, 2012;
Ng, 2014; Ng, 2015). Meng and Wang (2012) addressed a practical ship fleet problem under
the background of week-dependent container shipment demand. Their study generated
practical container routes considering transit time constraints by using space-time network
approach. For more works on the ship fleet deployment, one can refer to Ng (2016), in which
it elaborated a class of fleet deployment models.
For the empty container repositioning or empty container allocation, Crainic et al. (1993)
introduced two dynamic formulations for empty container allocation in single and multi-
commodity cases, which provided a general framework to formulate this class of problems.
Cheung and Chen (1998) considered a dynamic empty container allocation problem, which
helped to determine the number of containers leased to fulfill the demands of customers over
time. The management of importing and exporting empty containers in a port was analyzed by
Li et al. (2004) based on the multi-stage inventory theory, and Markov decision processes were
proposed for the problem. Li et al. (2007) extended the previous study to a multi-port
application with a proposed heuristic algorithm for the problem. The empty container
repositioning problem for general shipping service routes was formulated by Song and Dong

Chapter 2: Container Ship Type Decision 7

(2010) based on container flow balancing mechanism. Menh and Wang (2011) embedded the
empty container repositioning into the liner shipping service network design. They verified
that the network considering empty container could cut down the network cost significantly.
Song and Dong (2011) combined the laden container routing problem and empty container
repositioning problem. With fixed vessel schedules and shipping service network, their study
attempted to minimize the sum of all container related costs in routing and repositioning
processes.
However, the majority of these related works treated the ship fleet deployment and the
empty container repositioning as individual parts, i.e. they did not consider the two decisions
in their studies simultaneously. As it shows in the introduction that the two decisions interact
with each other in real-world operations, one may obtain a local optimum rather than the global
optimum for the shipping service only considering a single part of the two decisions.
On the other hand, few kinds of literature have studied the empty container repositioning
considering the usage of foldable containers. Konings (2005) has addressed the economic and
logistical viability of using foldable containers, which enhanced the confidence to use foldable
containers in sea transport. Shintani et al. (2010) investigated the impact of using foldable
containers in hinterland transport, which showed that the foldable containers substantially save
on the repositioning cost compared to the standard containers. Moon et al. (2013) was almost
the first to model the usage of standard containers and foldable containers in empty container
repositioning and proposed a heuristic algorithm to solve their problem. Based on Moon et al.
(2013), Myung and Moon (2014) found that the previous problem transfers to a minimum cost
flow if they do not consider the capacity restrictions when repositioning empty containers. In
our study, we recognize the trouble that when considering standard containers and foldable
containers, after involving the capacity restrictions (which is compulsory, as our problem
needs to determine the ship capacity), we will obtain a nonstandard minimum cost flow.
Fortunately, we design a revised network simplex algorithm that can easily tackle the trouble.
Compared with the above literature, our study incorporates the empty container allocation
(including the container fleet sizing and the empty container repositioning) into the ship type
decision in order to obtain the global optimal solution for a shipping route over a planning
horizon. It further considers the usage of foldable containers, which aims to see whether the
shipping liner should apply the foldable containers in the shipping service. Given a ship type
for considering capacity restrictions, it overcomes the trouble faced by Myung and Moon
(2014) by designing a revised network simplex algorithm. In all, we can tackle all the above
decisions by an integrated solution approach.

Chapter 2: Container Ship Type Decision 8

2.3 PROBLEM DESCRIPTION

Our problem focuses on ship type decision considering the empty container repositioning
and foldable containers for a given shipping service route. A fleet of container ships with
certain capacity is to be determined for weekly serving the ports along the shipping route. The
incoming container ship transports weekly laden containers originating from the ports to
destination ports. In each port, the empty container availability is critical for the shipping liner
to meet the laden container transportation. Generally, there are three empty container supplies
to fulfill laden container transportation. The economic supply is the arriving containers with
fully loaded goods from the incoming ship. Once these fully loaded containers arrive at the
ports and are delivered to consignees for unloading, the containers become empty and are
stored in depots for the laden consignment. The second supply is to reposition the empty
containers from surplus ports to deficit ports along the shipping route, which occupy the
capacity of the container ships among voyage legs. However, if the stored and repositioned
empty containers cannot meet the weekly laden container transportation, the shipping liner
must lease empty containers from container companies by short-term container leasing, which
is the third supply for the empty containers. The short-term container leasing is on an O-D pair
basis (See Section 2.3.2), which is different from the long-term container leasing that is on a
planning horizon basis (See Section 2.3.1).

Figure 2.2: A transpacific shipping service route operated by CMA CGM (CMA CGM,

2017)

In this study, the given shipping route has a fixed port rotation, shown by an illustrative
example in Figure 2.2. The itinerary of this route forms a loop: we can arbitrarily deem that
this itinerary starts at Shanghai and ends at Shanghai. Let 𝑝 ∈ 𝑃 represent the index of the
ports on a round trip for the route. Then, for this route, we can define Shanghai as Port 1,
Ningbo as Port 2, Pusan as Port 3, Los Angeles as Port 4 and Oakland as Port 5. Based on the
given route, the shipping liner has a set of O–D (Origin-Destination) port pairs 𝐷. The laden

Chapter 2: Container Ship Type Decision 9

container transportation arises on those pairs in each week. We represent (𝑜, 𝑑) as the index
for O–D port pairs, where 𝑜 ∈ 𝑃, 𝑑 ∈ 𝑃. Note that we study the shipping routes that have no
butterfly ports in the routes, i.e., the route can visit each port at most once.
For the given shipping service route, the shipping liner normally offers services for the ports
on a weekly basis. In other words, there is a week interval for each port to be visited by one
round trip and its next round trip. Let 𝑒 represent the index of round trips, and 𝐸 as the set of
round trips for one planning horizon. The weekly laden container transportation consignments
for each port accumulate between the visiting times of two adjacent round trips and are fulfilled
by the latter round trip. Here, we denote 𝑑𝑜𝑑,𝑒 as the number of laden containers accumulated
in Port 𝑜 between the time that the port is visited by the (𝑒 − 1)𝑠𝑡 round trip and the time that
the port is visited by the 𝑒 𝑡ℎ round trip, and will be transported to Port 𝑑 by the 𝑒 𝑡ℎ round trip.
Here, note that we use the number of round trips rather than the number of weeks to represent
the planning horizon, and the time interval between two round trips for a port is one week.
The number of ships deployed in the route is pre-determined. The number of ships
corresponds to the number of weeks needed for the round trip in the weekly shipping service,
i.e., if a round trip for the route needs 𝑁 weeks, there must be 𝑁 ships deployed in the route
such that the weekly shipping services can be guaranteed. Here notice that the 𝑛𝑡ℎ ship (𝑛 ∈
{1,2, … , 𝑁}) is assigned for the {𝑛𝑡ℎ , (𝑛 + 𝑁)𝑡ℎ , (𝑛 + 2𝑁)𝑡ℎ , (𝑛 + 3𝑁)𝑡ℎ , … } round trips.

2.3.1 Tactical level planning

In the tactical level, the problem involves two decisions. The first one is the ship type
decision. The ship type decision is mainly on choosing the ship type to deploy with a certain
capacity, measured by TEUs. Fulfilling the laden container transportation on the shipping
service route is the basic criteria for the deployment, which means the deployed ship type must
have the capacity to carry all the laden containers from origin ports to destination ports.
However, there is a trade-off on whether to deploy the ships with a larger capacity or not. If
the deployed ship type has a larger capacity, the fixed operating cost for the ship fleet is higher,
but the larger capacity means more empty containers can be repositioned from surplus ports
to deficit ports, which could save the high cost spent on short-term container leasing in those
deficit ports.
The second decision at the tactical level is the container fleet sizing (or we say the long-
term container leasing), which aims to determine the number of foldable empty containers and
standard empty containers that are leased in the ports along the shipping route initially. All
those leased empty containers construct a container fleet for the usage of the planning horizon

Chapter 2: Container Ship Type Decision 10

to serve the laden container demands. Both foldable containers and standard containers are
available for long-term leasing.

2.3.2 Operational level planning

Once a port is visited by a round trip, the shipping liner should make some operational level
decisions. To fulfill the weekly laden container transportations, the numbers of foldable
containers and standard containers in the depot need to be allocated for the laden container
transportation. However, if the stored empty containers are insufficient for the transportation
consignment, the shipping liner must lease empty containers from container leasing companies
by short-term leasing. The leased containers return to the container leasing companies at
destination ports, and the repositioning of the containers, as well as container repairs and
maintenance, are the duties of the container leasing companies. The short-term leasing follows
the so-called Master Lease Agreement for shipping containers (Wolff et al., 2007; Container
Auction, 2017). If there are empty containers surplus after fulfilling the transportation
consignment, the shipping liner needs to decide on whether to reposition those empty
containers to other deficit ports and how many empty containers to be repositioned by
considering the remaining capacity of the ship when visiting the port by the round trip.
In summary, our problem aims to solve the ship type decision and the container fleet sizing
at the tactical level and allocate the empty containers among ports at the operational level. In
the tactical level, the ship type decision determines the capacity (in TEUs) of the container
ships deployed for the fleet. The container fleet sizing determines the number of long-term
leasing foldable containers and the number of long-term leasing standard containers in each
port. In the operation level, foldable empty containers and standard empty containers are
allocated in each port to fulfill the laden container transportation consignment, the short-term
leasing containers are involved if there is empty container deficit, and the empty container
repositioning is to be determined if there is empty container surplus.

2.3.3 Assumptions
Before addressing the model for our problem, we clarify some underlying assumptions:
(i) In the beginning, all the ships depart from the first port in the shipping route, and they
depart one by one with one-week interval (i.e., the first ship departs in the first week; the
second ship departs in the second week, and so on).
Assumption (i) indicates that when a shipping liner starts to operate a weekly shipping
service, it will mass a ship fleet at a homeport (i.e., the first port), and dispatch ships with a
one-week interval to form the weekly service pattern.

Chapter 2: Container Ship Type Decision 11

 (ii) For each port, the weekly laden container transportation consignments that accumulate
between two adjacent round trips must be fulfilled by the latter round trip, i.e., those laden
containers must be loaded into the incoming ship by the latter round trip.
Under assumption (ii), the accumulated consignments must be transported to destination
ports by the incoming ship; otherwise, they will not arrive at customers on time.
(iii) When laden containers arrive at destination ports, it will take several weeks for
devanning (i.e., the process that the containers are delivered to consignees for unpacking and
returned to the ports). After the container devanning process, the laden containers become
empty containers for the next transportation consignments.
Assumption (iii) shows when laden containers arrive at destination ports, the containers
cannot be used immediately for next consignments, as they carry cargo inside. It takes time
for delivering the laden containers to customers and unloading the cargo. It is worthwhile to
mention that shipping liner companies normally would pose a required time for the devanning
process, upon which the consignees should return empty containers (Hanh, 2003; Shipping
and Freight Resource, 2017). The required time for returning empty containers (i.e., the
devanning process) vary among shipping liner companies, such as COSCO requires ten days
for returning (COSCO, 2017), OOCL allows five days for returning (OOCL, 2017). Supposing
a shipping liner company sets two weeks for the required time, in reality, it may happen that
consignees return empty containers earlier than two weeks (e.g., five days or ten days) in a
stochastic manner. Here, to facilitate the exposition of our study, we assume all those empty
containers are available for next transportation consignments after the required time for the
devanning process (e.g., two weeks in the above case), even though some empty containers
may be returned earlier.
The objective of our problem is to minimize the total cost over one planning horizon,
including the fixed operation cost, the long-term container leasing cost, the short-term
container leasing cost, the repositioning cost, and the storage cost.

2.4 MODEL FORMULATION

In essence, our problem has a latent network structure when the capacity of the ship fleet is
determined. Given the ship capacity, we can transfer the problem to a nonstandard minimum
cost flow problem by formulating a network flow (NF) model, rather than a mathematical
programming (MP) model. An MP model normally needs some commercial solvers to
optimize, such as CPLEX and Gurobi, which are not desirable for many shipping liner
companies. However, several network algorithms can easily solve an NF model to optimality,
for example, the network simplex algorithm. Therefore, in this section, we construct an NF

Chapter 2: Container Ship Type Decision 12

model by building a network for the problem given the ship capacity. In the next section, we
will design a solution approach with a revised network simplex algorithm to solve the NF
model and optimize the ship capacity.
In the section, we first introduce some notations for sets and input parameters. Then, to
avoid confusion, we will build a sub-network for the problem when only considering using
standard containers, which leads to a preliminary NF model (denoted as 𝑷𝑵𝑭). In the next, we
further build a sub-network for foldable containers. By incorporating it to the sub-network for
standard containers, we obtain a whole network for the NF model (denoted as 𝑻𝑵𝑭) that solves
the problem given a ship capacity. However, the incorporation would impose a trouble that
two parallel arcs share a specific capacity restriction, simultaneously. We will tackle this
trouble when designing the revised network simplex algorithm in the next section.

2.4.1 Notations
Indices and sets:
𝑝: Index of ports,
𝑃: Set of all the ports in the shipping service route,
(𝑜, 𝑑): Index of O–D port pairs, where 𝑜 ∈ 𝑃, 𝑑 ∈ 𝑃,
𝐷: Set of O–D port pairs and 𝐷 ≔ {(𝑜, 𝑑) ∈ 𝑃 × 𝑃|𝑜 ≠ 𝑑},
𝑣: Index of ship types,
𝑉: Set of ship types for the shipping service route,
𝑒: Index of round trips,
𝐸: Set of all the round trips in the planning horizon,
Input parameters:
𝑁: Number of ships deployed in the shipping service route,
𝑀: Number of foldable containers that can be folded into one standard container,
𝐶 𝑣 : Fixed operation cost in the planning horizon when the ships in Type 𝑣 are deployed in
the shipping service route, 𝑣 ∈ 𝑉,
𝐾 𝑣 : Container capacity of the ships in Type 𝑣 with the unit of TEUs, 𝑣 ∈ 𝑉,
𝑤𝑝 : Number of weeks that are needed for the devanning process in Port 𝑝, 𝑝 ∈ 𝑃,
𝑑𝑜𝑑,𝑒 : Number of laden container transportation consignments from Port 𝑜 to Port 𝑑 that
should be transported by the 𝑒 𝑡ℎ round trip, (𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸,
𝑠𝑝𝑆 : Unit weekly storage cost of a standard container in Port 𝑝, 𝑝 ∈ 𝑃,

𝑠𝑝𝐹 : Unit weekly storage cost of a foldable container in Port 𝑝, 𝑝 ∈ 𝑃,

Chapter 2: Container Ship Type Decision 13

 𝑆
𝑟𝑜𝑑 : Unit repositioning cost of a standard container from Port 𝑜 to Port 𝑑 , including
container loading cost (𝑎𝑜𝑆 ) and unloading cost (𝑏𝑑𝑆 ), (𝑜, 𝑑) ∈ 𝐷,
𝐹
𝑟𝑜𝑑 : Unit repositioning cost of a foldable container from Port 𝑜 to Port 𝑑 , including
container loading cost (𝑎𝑜𝐹 ) and unloading cost (𝑏𝑑𝐹 ), folding cost (𝐴𝑜 ) and unfolding cost (𝐵𝑑 ),
(𝑜, 𝑑) ∈ 𝐷,
𝑙𝑜𝑑 : Unit short-term leasing cost of a standard container from Port 𝑜 to Port 𝑑, (𝑜, 𝑑) ∈ 𝐷,
𝐿𝑆𝑝 : Long-term leasing cost of a standard container for the planning horizon usage in Port 𝑝,
𝑝 ∈ 𝑃,
𝐿𝐹𝑝 : Long-term leasing cost of a foldable container for the planning horizon usage in Port 𝑝,
𝑝 ∈ 𝑃.
Here, notice that “from Port 𝑜 to Port 𝑑 by the 𝑒 𝑡ℎ round trip” means that the containers are
loaded from Port 𝑜 to the ship when the port is visited by the 𝑒 𝑡ℎ round trip, and those
containers will be transported to Port 𝑑. If 𝑜 < 𝑑, the containers will arrive at Port 𝑑 by the
𝑒 𝑡ℎ round trip; If 𝑜 > 𝑑, the containers will arrive at Port 𝑑 by the (𝑒 + 𝑁)𝑡ℎ round trip, as the
ship would return to Port 1 and restart a new round trip.

2.4.2 A preliminary NF model for standard containers

Given the type of ships deployed has the capacity in 𝐾 TEUs, we only allow using standard
containers to transport goods in this subsection. The sub-network built for standard containers
will firstly embed the decisions on the container fleet sizing (i.e., the long-term container
leasing) and the empty container repositioning, as the decisions are also the same for foldable
containers. Then, we extend the sub-network to consider the short-term container leasing,
which only involves the standard containers.

Long-term containers leasing and empty container repositioning

𝑆
Before constructing the sub-network for standard containers, we divide 𝑟𝑜𝑑 (i.e., unit
repositioning cost of a standard container from Port 𝑜 to Port 𝑑) into two parts, i.e., the loading
cost in Port 𝑜 (denoted as 𝑎𝑜𝑆 ) and the unloading cost in Port 𝑑 (denoted as 𝑏𝑑𝑆 ), where 𝑟𝑜𝑑
𝑆
=
𝑎𝑜𝑆 + 𝑏𝑑𝑆 as suggested in the parameter definition.
Notice that the imbalance of laden container flow leads to the empty container repositioning
among the ports. As all laden container transportation consignments (i.e., 𝑑𝑜𝑑,𝑒 ) have to be
fulfilled by the 𝑒 𝑡ℎ round trip, we could analyze the empty containers deficit or surplus in each
port on each round trip caused by the imbalance of laden container flow. Here, we denote 𝑄𝑝,𝑒

as the number of empty containers deficit or surplus in Port 𝑝 when visited by the 𝑒 𝑡ℎ round
trip, calculated by Eq. (2.1) and Eq. (2.2).

Chapter 2: Container Ship Type Decision 14

 𝑄𝑝,𝑒 = − ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷 𝑑𝑜𝑑,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 = 1, (2.1)

𝑄𝑝,𝑒 = ∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝑑𝑜𝑑,𝑒−𝑤𝑝 + ∑𝑑=𝑝,𝑜>𝑑,(𝑜,𝑑)∈𝐷 𝑑𝑜𝑑,𝑒−𝑁−𝑤𝑝 − ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷 𝑑𝑜𝑑,𝑒

∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸/{1}, (2.2)
If 𝑄𝑝,𝑒 > 0 (𝑄𝑝,𝑒 ≤ 0), it indicates that there are 𝑄𝑝,𝑒 (−𝑄𝑝,𝑒 ) number of empty containers

surplus (deficit) in Port 𝑝 when visited by the 𝑒 𝑡ℎ round trip. The long-term containers leasing
generates empty containers, and the empty container repositioning induces the empty container
flow between the deficit ports (i.e., the demand ports) and the surplus ports (i.e., the supply
ports). Henceforth, we can construct a flow network for the empty containers.

Figure 2.3: A sub-network for standard containers without considering the short-term

container leasing

The sub-network contains |𝑃| + 2 ∙ |𝐸| ∙ |𝑃| + 2 nodes. Among them, the |𝑃| nodes define
the flow conversion for the long-term container leasing in Port 𝑝 at the beginning of the
planning horizon, labelled as 𝐿𝑡𝑒𝑟𝑚 𝑆 _𝑝. The 2 ∙ |𝐸| ∙ |𝑃| nodes are categorized into |𝐸| ∙ |𝑃|
groups of two kinds of nodes. One kind of nodes in a group denotes the flow conservation in
the ship for the 𝑒 𝑡ℎ round trip after visiting Port 𝑝, labeled as 𝑆ℎ𝑖𝑝 𝑆 _𝑝_𝑒. The other kind of
nodes denotes the flow conservation in Port 𝑝 after visited by the 𝑒 𝑡ℎ round trip, labeled as
𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒. Two additional dummy nodes represent the source node (labelled as 𝑆𝑜𝑢𝑟𝑐𝑒) and
the sink node (labelled as 𝑆𝑖𝑛𝑘) respectively. To facilitate the understanding of the sub-

Chapter 2: Container Ship Type Decision 15

network construction, we give a shipping service route for the example through this section.
The route is: Singapore (1) → Hong Kong (2) → Xiamen (3) → Singapore. A round trip for
the route needs two weeks, which means two liner ships are deployed with the capacity as 𝐾.
A planning horizon includes four round trips. Then, Figure 2.3 illustrates an example for the
sub-network construction.
In the sub-network network for standard containers, the arcs between nodes correspond to
the decision variables of the problem.
 The arc from 𝐿𝑡𝑒𝑟𝑚 𝑆 _𝑝 to 𝑃𝑜𝑟𝑡 𝑆 _𝑝_1 represents the long-term container leasing activity in
Port 𝑝. The amount of flow on the arc denotes the number of empty containers leased for
the planning horizon usage. The cost for the arc is 𝐿𝑆𝑝 . The capacity on the arc is ∞.

 The arc from Node 𝑆ℎ𝑖𝑝 𝑆 _𝑝_𝑒 to 𝑆ℎ𝑖𝑝 𝑆 _(𝑝 + 1)_𝑒 represents the voyage leg from Port 𝑝 to
Port 𝑝 + 1 by the ship for the 𝑒 𝑡ℎ round trip. The amount of the flow on the arc denotes the
number of empty containers carried on the ship in the voyage leg. Here note that: (i) The
voyage leg from Port |𝑃| to Port 1 by the same ship for the 𝑒 𝑡ℎ round trip is the arc from
Node 𝑆ℎ𝑖𝑝_|𝑃|_𝑒 to 𝑆ℎ𝑖𝑝_1_(𝑒 + 𝑁) . (ii) The cost for the arc is zero, as we have
decomposed the repositioning cost. (iii) As the flows of laden containers are pre-determined,
the remaining capacity in each voyage leg (on each arc) is confirmed, denoted as 𝐾𝑝,𝑒 (the

remaining capacity in the voyage leg from Port 𝑝 to Port 𝑝 + 1 by the ship for the 𝑒 𝑡ℎ round
trip), which can be calculated as follows.
𝐾𝑝,𝑒 = 𝐾 − ∑𝑜≤𝑝,𝑑>𝑝,(𝑜,𝑑)∈𝐷 𝑑𝑜𝑑,𝑒 − ∑𝑑≥𝑝+1,𝑜>𝑑,(𝑜,𝑑)∈𝐷 𝑑𝑜𝑑,𝑒−𝑁 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸, (2.3)

 The arc from Node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝑆ℎ𝑖𝑝 𝑆 __𝑝_𝑒 (resp., Node 𝑆ℎ𝑖𝑝 𝑆 _𝑝_𝑒 to 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 )
represents the empty container loading process from Port 𝑝 to the ship (resp., unloading
process from the ship to Port 𝑝) for the 𝑒 𝑡ℎ round trip. The amount of the flow on the arc
denotes the number of empty containers loaded to the ship (resp., unloaded to the port), and
the cost for the arc is 𝑎𝑝𝑆 (resp., 𝑏𝑝𝑆 ), i.e., the container loading cost at Port 𝑝 (resp., the
container unloading cost at Port 𝑝). The capacity on the arc is ∞.
 The arc from Node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝑃𝑜𝑟𝑡 𝑆 _𝑝_(𝑒 + 1) represents the empty container
inventory (i.e., the empty containers left) in Port 𝑝 after visited by the 𝑒 𝑡ℎ round trip. The
amount of the flow on the arc denotes the number of empty containers left in the port after
visited by the 𝑒 𝑡ℎ round trip (i.e., the inventory level 𝛿𝑝,𝑒
𝑆
). The cost for the arc is 𝑠𝑝 . The
capacity on the arc is ∞.
With the two additional dummy nodes, i.e., Node 𝑆𝑜𝑢𝑟𝑐𝑒 and Node 𝑆𝑖𝑛𝑘 , we firstly
construct the dummy arcs from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to all the nodes 𝐿𝑡𝑒𝑟𝑚 𝑆 _𝑝 with zero cost

Chapter 2: Container Ship Type Decision 16

coefficient and infinity capacity. Then, we build the dummy arcs from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to all the
nodes 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 such that 𝑄𝑝,𝑒 > 0 (resp., from all the nodes 𝑃𝑜𝑟𝑡_𝑝_𝑒 such that 𝑄𝑝,𝑒 ≤ 0 to
Node 𝑆𝑖𝑛𝑘). The costs for all the arcs are zero. The capacity on the arc from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to
Node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 is 𝑄𝑝,𝑒 (resp., from Node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝑆𝑖𝑛𝑘 is −𝑄𝑝,𝑒 ).

Short-term container leasing

We extend the previous sub-network to consider short-term container leasing. Referring to
the concept of short-term container leasing in Section 2.3, if empty containers in origin ports
are scarce, the shipping liner has to lease empty containers in the ports for laden container
transportation consignments and return those containers at destination ports after unpacking
the containers. In essence, the short-term container leasing has no effects on laden containers,
as all transportation consignments must be fulfilled. However, it affects the empty container
demand in the origin ports and the empty container supply in the destination ports.
Specifically, assuming that when Port 𝑜 is visited by the 𝑒 𝑡ℎ round trip, due to the dearth of
empty containers in the port, the shipping liner has to lease some empty containers (say 𝛾𝑜𝑑,𝑒 )
to transport the goods in laden containers to Port 𝑑. The short-term container leasing fulfills
𝛾𝑜𝑑,𝑒 empty container demand in Port 𝑜, but 𝛾𝑜𝑑,𝑒 empty containers are excluded in the empty
container supply in Port 𝑑 as those leased empty containers have to be returned to container
leasing companies. Therefore, the empty container demand in Port 𝑜 by the 𝑒 𝑡ℎ round trip
decreases by 𝛾𝑜𝑑,𝑒 , and the empty container supply in Port 𝑑 by the (𝑒 + 𝑤𝑑 )𝑡ℎ (if 𝑑 > 𝑜) or
(𝑒 + 𝑁 + 𝑤𝑑 )𝑡ℎ (if 𝑑 < 𝑜) round trip decreases by 𝛾𝑜𝑑,𝑒 , virtually.
Based on the above analysis, if we further consider the short-term container leasing, the
previous sub-network without considering the short-term container leasing (i.e., Figure 2.3)
should be modified as follows (see Figure 2.4 for the following modifications): (i) We add a
pairwise node for each node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒, denoted as 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒. Here, Node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒
(resp., Node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒) shows the empty container flow before (resp., after) using the short-
term container leasing. (ii) We disconnect the dummy arcs from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to all the nodes
𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 and from all the nodes 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to Node 𝑆𝑖𝑛𝑘. (iii) We construct the dummy
arcs from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to all the nodes 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 such that 𝑄𝑝,𝑒 > 0 (resp., from all the

nodes 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to Node 𝑆𝑖𝑛𝑘 such that 𝑄𝑝,𝑒 ≤ 0). The costs for all the arcs are zero. The

capacity on the arc from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 is 𝑄𝑝,𝑒 (resp., from Node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒
to 𝑆𝑖𝑛𝑘 is −𝑄𝑝,𝑒 ). (iv) If 𝑄𝑝,𝑒 > 0 ( 𝑄𝑝,𝑒 ≤ 0), we construct the dummy arc from Node

𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 (resp., from Node 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒) with arc capacity as
∞ and arc cost as zero. (v) For two nodes 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒, there exist an origin node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑜_𝑒

Chapter 2: Container Ship Type Decision 17

and a destination node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑑_𝑒 1 (if 𝑑 > 𝑜, 𝑒 1 = 𝑒 + 𝑤𝑑 ; if 𝑑 < 𝑜, 𝑒 1 = 𝑒 + 𝑁 + 𝑤𝑑 ).
When there is empty container deficit in Node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑜_𝑒 (i.e., 𝑄𝑜,𝑒 ≤ 0), the short-term
container leasing is possible in the node, and the arc from Node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑑_𝑒 1 to 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑜_𝑒
for the short-term container leasing is constructed. Notice that for the voyage direction, the
laden containers flow from Node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑜_𝑒 to 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑑_𝑒 1, but here, the empty containers
flow in the opposite direction. The capacity for the arc is 𝑑𝑜𝑑,𝑒 as the maximum empty
container demand from Port 𝑜 to 𝑑 is the laden container demand 𝑑𝑜𝑑,𝑒 . The cost on the arc is
𝑙𝑜𝑑 (i.e., the short-term leasing cost).

Figure 2.4: The sub-network defined for the preliminary NF model

It is worthwhile to mention that in this subsection, we refer the short-term container leasing
to the Master Lease Agreement mentioned in Section 2.3.2, in which empty containers are
leased and the cost is charged on an O-D port pair basis. In real-world operations, the shipping
liner company and container leasing company could have a lease agreement that is on a round
trip basis. Under the agreement, the shipping liner company can lease a certain number of
empty containers at Port 𝑝, and this number of empty containers must be returned at Port 𝑝
after several weeks (or round trips). If this is the agreement applied between some shipping
liner companies and container leasing containers, we can make the following modifications to
the network. Firstly, we define 𝑔𝑝 as the number of weeks that the shipping liner company
needs to return leased empty containers at Port 𝑝, and define 𝑞𝑝 as the unit cost of leasing
empty containers for 𝑔𝑝 weeks at Port 𝑝. Then, for modifying the network, we connect the

Chapter 2: Container Ship Type Decision 18

arcs from 𝑃𝑜𝑟𝑡 𝑆 _𝑝_(𝑒 + 𝑞𝑝 ) to 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 with arc capacity as ∞ and arc cost as 𝑔𝑝 . Note

that although some empty containers (say 𝜉 empty containers) are leased by the 𝑒 𝑡ℎ round trip
and are returned by the (𝑒 + 𝑞𝑝 )𝑡ℎ round trip at Port 𝑝, the empty container flow in the
network is in the opposite direction. This is due to that the empty container supply by the
(𝑒 + 𝑞𝑝 )𝑡ℎ round trip (resp., the 𝑒 𝑡ℎ round trip) at Port 𝑝 will decrease by (resp., increase by)
𝜉 empty containers for the returning process (resp., the leasing process).
Until now, we have constructed a whole sub-network for the problem when only considering
using standard containers, which leads to the preliminary NF model (𝑷𝑵𝑭 model). For 𝑷𝑵𝑭
model, we need to fulfill Node 𝑆𝑖𝑛𝑘 with total empty container demand ∑𝑝∈𝑃,𝑒∈𝐸(−𝑄𝑝,𝑒 )+ by
originating from Node 𝑆𝑜𝑢𝑟𝑐𝑒. The goal is to minimize the total cost through the sub-network,
which is a standard minimum cost flow problem.

2.4.3 The NF model for both standard containers and foldable containers
In this subsection, we incorporate the foldable containers into 𝑷𝑵𝑭 model for the NF model
that solves our problem given the ship capacity 𝐾 (𝑻𝑵𝑭 model). Firstly, we construct a sub-
network for foldable containers. Except from the short-term container leasing, the foldable
containers also have the long-term container leasing and the empty container repositioning.
Henceforth, the sub-network for foldable containers is similar to the sub-network shown for
the standard containers in Figure 2.3. Figure 2.5 shows an example for the sub-network for
foldable containers, in which the nodes 𝑆ℎ𝑖𝑝𝐹 _𝑝_𝑒, 𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒 and 𝐿𝑡𝑒𝑟𝑚𝐹 _𝑝 correspond to
the nodes 𝑆ℎ𝑖𝑝 𝑆 _𝑝_𝑒, 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 and 𝐿𝑡𝑒𝑟𝑚 𝑆 _𝑝 in Figure 2.3, respectively. Note that the
costs on the arcs from Node 𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒 to 𝑆ℎ𝑖𝑝𝐹 __𝑝_𝑒 and from Node 𝑆ℎ𝑖𝑝𝐹 _𝑝_𝑒 to
𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒 are slightly different from that of standard containers, as there are additional
folding and unfolding costs (𝐴𝑜 and 𝐵𝑑 ) for foldable containers.
To embed the sub-network for the foldable containers into 𝑷𝑵𝑭 model defined in Section
2.4.2, we need to build some connections as shown in Figure 2.6: (i) we add a pairwise dummy
node for nodes 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 and 𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒, denoted as 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒. Here, Node 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒
accumulates the empty container surplus or deficit for standard containers and foldable
containers; (ii) we disconnect all dummy arcs from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to all other nodes or from
all other nodes to Node 𝑆𝑖𝑛𝑘; (iii) we construct the dummy arcs from Node 𝑆𝑜𝑢𝑟𝑐𝑒 to all the
nodes 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 such that 𝑄𝑝,𝑒 > 0 (resp., from all the nodes 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 to Node 𝑆𝑖𝑛𝑘 such
that 𝑄𝑝,𝑒 ≤ 0). The costs for all the arcs are zero. The capacity on the arc from Node 𝑆𝑜𝑢𝑟𝑐𝑒

to 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 is 𝑄𝑝,𝑒 (resp., from Node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝑆𝑖𝑛𝑘 is −𝑄𝑝,𝑒 ); (iv) if 𝑄𝑝,𝑒 > 0, we

construct the dummy arcs from Node 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 to 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 and Node 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 to

Chapter 2: Container Ship Type Decision 19

𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒, to split empty container flow to the sub-network of standard containers and
foldable containers, respectively. (v) if 𝑄𝑝,𝑒 < 0, we construct the dummy arcs from Node

𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 to 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 and Node 𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒 to 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 to accumulate empty
container flow from the sub-network of standard containers and foldable containers,
respectively. Note that 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 serve as the node to allocate the empty container demand
to standard containers by node 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 and foldable containers by node 𝑃𝑜𝑟𝑡𝐹 _𝑝_𝑒.

Figure 2.5: A sub-network for foldable containers

Figure 2.6: Connections between standard containers and foldable containers in TNF model

As we have constructed the whole flow network for our problem, it is worthwhile to show
the total number of nodes involved in the network by referring Figure 2.3 to Figure 2.6. Figure

Chapter 2: Container Ship Type Decision 20

2.3 defines the subnetwork for standard containers and it contains |𝑃| + 2|𝑃||𝐸| nodes; Figure
2.4 introduces 𝐵𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒 nodes for considering the short-term container leasing and it adds
|𝑃||𝐸| nodes to the network; Figure 2.5 defines the subnetwork for foldable containers and it
adds |𝑃| + 2|𝑃||𝐸| nodes to the network; Figure 2.6 introduces 𝑇𝑃𝑜𝑟𝑡_𝑝_𝑒 nodes for
combining two subnetworks and it adds |𝑃||𝐸| nodes to the network. In all, the network has
|𝑃| + 6|𝑃||𝐸| + 2 after adding the dummy 𝑆𝑜𝑢𝑟𝑐𝑒 and 𝑆𝑖𝑛𝑘 nodes, which is a polynomial
function to the input parameters |𝑃| and |𝐸|.
Capacity restriction sharing
The above incorporation process of the two sub-networks for 𝑻𝑵𝑭 model induces a trouble
that makes our problem a nonstandard minimum cost flow problem. That is two parallel arcs
sharing a specific capacity restriction, simultaneously. In Figure 2.3, the arc from Node
𝑆ℎ𝑖𝑝 𝑆 _𝑝_𝑒 to 𝑆ℎ𝑖𝑝 𝑆 _(𝑝 + 1)_𝑒 carries the flow of standard empty containers on the voyage
leg from Port 𝑝 to Port 𝑝 + 1 by the ship for the 𝑒 𝑡ℎ round trip. In Figure 2.5, the arc from
Node 𝑆ℎ𝑖𝑝𝐹 _𝑝_𝑒 to 𝑆ℎ𝑖𝑝𝐹 _(𝑝 + 1)_𝑒 carries the flow of foldable empty containers on the
same voyage leg. Henceforth, the two parallel arcs share the same capacity restriction 𝐾𝑝,𝑒 , as
they require repositioning empty containers on the same ship at the same voyage. More
importantly, the flows on the parallel arcs occupy different units of the capacity, due to the
𝑆 𝐹
feature of foldable containers. By supposing 𝑥𝑝,𝑒 and 𝑥𝑝,𝑒 are the flow on the two parallel arcs,
we need to enforce that:
𝐹
𝑥𝑝,𝑒
𝑆
0 ≤ 𝑥𝑝,𝑒 + 𝑀
≤ 𝐾𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸, (2.4)

for the capacity restriction sharing. This characteristic leads to the trouble that we cannot
directly transfer our problem given the ship capacity to a standard minimum cost flow problem.
Alternatively, to avoid the trouble, we introduce two dynamic capacity restrictions for the two
𝑆 𝐹 𝑆 𝐹
parallel arcs, denoted as 𝜇𝑝,𝑒 and 𝜇𝑝,𝑒 , respectively. Based on 𝜇𝑝,𝑒 and 𝜇𝑝,𝑒 , we separate the
sharing capacity restriction 𝐾𝑝,𝑒 to two individual ones such that:
𝑆 𝑆
0 ≤ 𝑥𝑝,𝑒 ≤ 𝜇𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸, (2.5)
𝐹 𝐹
0 ≤ 𝑥𝑝,𝑒 ≤ 𝜇𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸. (2.6)
Meanwhile, the two individual capacity restrictions must hold that:
𝐹
𝜇𝑝,𝑒
𝑆
𝜇𝑝,𝑒 + = 𝐾𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸. (2.7)
𝑀
𝑆 𝐹
Note 𝜇𝑝,𝑒 and 𝜇𝑝,𝑒 keep dynamically changed in the algorithm developed in the next section,
𝑆 𝐹
so long as Eq. (2.7) holds. Here, we can initialize any values for 𝜇𝑝,𝑒 and 𝜇𝑝,𝑒 without
considering that the values will lead to no feasible solutions for the 𝑻𝑵𝑭 model. This is due
to that the short-term container leasing of standard containers can always treat the empty

Chapter 2: Container Ship Type Decision 21

container imbalance between deficit nodes and surplus nodes. With the separated capacity
restrictions, we can solve the 𝑻𝑵𝑭 model by several network algorithms.

2.5 SOLUTION APPROACH

The solution approach for our problem is an iterative procedure. In this section, we firstly
design a revised network simplex algorithm to solve the 𝑻𝑵𝑭 model given a ship capacity.
Then, for the solution approach, we initialize at 𝐾 = 𝐾 1 that is the capacity of Type 1 ship
(assuming to be the smallest ship type) for running the algorithm. Based on the information
from the previous running of the algorithm, we select another ship capacity 𝐾 𝑣 , given which
it again invokes the algorithm to derive the minimum flow cost for empty containers, denoted
as 𝜎 𝑣 . The summation of the fixed operation cost 𝐶 𝑣 and 𝜎 𝑣 is the total minimum cost when
using ships in Type 𝑣 for the fleet. By comparing those costs, we can obtain the optimal
solution for our problem as well as the total optimal cost. In this section, we also formulate a
mixed-integer linear programming model (MILP) for the problem, which can be solved
directly by CPLEX solver for verifying the optimality of the proposed solution approach.

2.5.1 A revised network simplex algorithm

The network simplex algorithm is perhaps the most powerful algorithm to solve the
minimum cost flow problem. The Cycle Free Property, Spanning Tree Property and Minimum
Cost Flow Optimality Conditions are the major principles supporting the algorithm. One can
refer to Chapter 11 of the book by Ahuja et al. (1993) for detailed descriptions of those
properties and the algorithm. Here, we will briefly introduce the properties and the algorithm,
and propose a revised network simplex algorithm for solving our 𝑻𝑵𝑭 model. Note that to
present the algorithm in a general way and avoid confusions, we use the set 𝒜 to denote the
set of all arcs in the network constructed in Section 2.4. Arc (𝑖, 𝑗) ∈ 𝒜 denotes a specific arc
from node 𝑖 to node 𝑗 in the network with a cost coefficient 𝑐𝑖𝑗 and a capacity restriction 𝜇𝑖𝑗 .
(i) Spanning Tree Property: If a minimum cost flow problem is bounded from below by a
feasible region, the problem will always have an optimal spanning tree solution. A spanning
tree solution splits the arc set 𝒜 of the network into three parts. (a) 𝑻, the spanning tree arc
set, in which the flow on the arcs are unbounding. (b) 𝑳, a non-tree arc set, in which the flow
on the arcs equals zero. (c) 𝑼, a non-tree arc set, in which the flow on the arcs equals the
capacity restrictions. The triple (𝑻, 𝑳, 𝑼) is the so-called spanning tree structure.
(ii) Minimum Cost Flow Optimality Conditions: A feasible spanning tree (𝑻, 𝑳, 𝑼) is the
𝜋
optimal solution if the arc reduced costs 𝑐𝑖𝑗 satisfy the following conditions. (a) For all (𝑖, 𝑗) ∈
𝜋 𝜋 𝜋
𝑻, 𝑐𝑖𝑗 = 0; (b) For all (𝑖, 𝑗) ∈ 𝑳, 𝑐𝑖𝑗 ≥ 0; (c) For all (𝑖, 𝑗) ∈ 𝑼, 𝑐𝑖𝑗 ≤ 0. Given a spanning tree

Chapter 2: Container Ship Type Decision 22

 𝜋
structure, 𝑐𝑖𝑗 is derived by using the cost coefficients on arcs and defined node potentials for

nodes: it first assigns a node potential 𝜋(1) = 0 for the root node 1. Then by holding,
𝜋
𝑐𝑖𝑗 = 𝑐𝑖𝑗 − 𝜋(𝑖) + 𝜋(𝑗) = 0 ∀(𝑖, 𝑗) ∈ 𝑻, (2.8)
𝜋
it obtains all the nodes potentials 𝜋(𝑖) in the network, based on which 𝑐𝑖𝑗 of the non-tree arcs
𝜋 𝜋
is derived by 𝑐𝑖𝑗 = 𝑐𝑖𝑗 − 𝜋(𝑖) + 𝜋(𝑗), ∀(𝑖, 𝑗) ∈ 𝑳, (𝑖, 𝑗) ∈ 𝑼 . Note that 𝑐𝑖𝑗 has a similar

meaning with the reduced cost in the primal simplex algorithm for the linear programming
problem, which indicates the cost changed in the objective if we increase one more unit flow
on the arc (𝑖, 𝑗).
(iii) The network simplex algorithm generally maintains a feasible spanning tree structure
in each iteration and moves from one structure to another one until reaching the optimality. (a)
Given a (𝑻, 𝑳, 𝑼), the algorithm checks the optimality conditions. If all the conditions are
satisfied, the algorithm stops, otherwise selects the most violation arc (by finding the
𝜋 𝜋
maximum |𝑐𝑖𝑗 | ) from the arc set 𝒱 = {(𝑖, 𝑗) ∈ 𝑳 ∪ 𝑼: 𝑖𝑓 (𝑖, 𝑗) ∈ 𝑳, 𝑐𝑖𝑗 < 0; 𝑖𝑓 (𝑖, 𝑗) ∈
𝜋
𝑼, 𝑐𝑖𝑗 > 0}. (b) It adds the violation arc (called the entering arc) into the current spanning tree
structure, by which it obtains a negative cost cycle. That means increasing the flow on the
forward direction of the cycle will decrease the current objective of the total flow cost. (c) It
augments the maximum possible flow on the negative cost cycle until the flow of one arc in
the cycle reaches zero or its capacity restriction (called the leaving arc). (d) It replaces the
leaving arc with the entering arc for a new feasible spanning tree structure (𝑻′, 𝑳′, 𝑼′). The
algorithm repeats the above procedure (a)-(d) until finding the optimal spanning tree structure
(𝑻∗ , 𝑳∗ , 𝑼∗ ).

Revisions in the pivot operation

The Part (c) in the above procedures of the network simplex algorithm is the pivot
operation. Considering that our network constructed in Section 2.4 has dynamic capacity
restrictions on parallel arcs, we revise the pivot operation of the standard network simplex
algorithm for a revised network simplex algorithm to solve our 𝑻𝑵𝑭 model. Here, we will
firstly elaborate the standard pivot operation and then propose our revisions.
Supposing that the entering arc is (𝑘, 𝑙), the combination of (𝑘, 𝑙) and the spanning tree 𝑻
forms a negative cost cycle 𝑾, known as the pivot cycle. The cycle 𝑾 has a direction that is
the same as (𝑘, 𝑙) if (𝑘, 𝑙) ∈ 𝑳, and is opposite to (𝑘, 𝑙) if (𝑘, 𝑙) ∈ 𝑼. For all arcs in the cycle,

the arcs following the cycle direction belong to a forward arc set 𝑾, and the arcs following
the opposite cycle direction belong to a backward arc set 𝑾. Figure 2.7 shows an example for

the cycle, in which arc (3, 4) is the entering arc, arcs (5, 0), (0, 1), (2, 3) belong to the set 𝑾,
and arcs (5, 4), (2, 1) belong to the set 𝑾.

Chapter 2: Container Ship Type Decision 23

 Figure 2.7: An example of the pivot cycle W

Supposing that 𝑥𝑖𝑗 is the existing flow on arc (𝑖, 𝑗), the maximum flow change on each arc
in the cycle is denoted by 𝛿𝑖𝑗 , calculated by:

𝜇𝑖𝑗 − 𝑥𝑖𝑗 , 𝑖𝑓 (𝑖, 𝑗) ∈ 𝑾

𝛿𝑖𝑗 = { (2.9)
𝑥𝑖𝑗 , 𝑖𝑓 (𝑖, 𝑗) ∈ 𝑾
In the standard pivot operation, it will augment 𝛿 ∗ = min{𝛿𝑖𝑗 : (𝑖, 𝑗) ∈ 𝑾} amount of flow
to guarantee the feasibility. Here, we start to revise the pivot operation to capture the dynamic
capacity restrictions under two principles: (i) we need to make 𝛿 ∗ as large as possible. This
is due to 𝑾 is a negative cost cycle such that the more flow augmented on the cycle, the
larger decreasing on the total flow cost. (ii) We must guarantee the feasibility when
augmenting flow.
𝑆 𝐹
To facilitate our description of revisions for the pivot operation, we define 𝜖𝑝,𝑒 and 𝜖𝑝,𝑒 as
the indices for two parallel arcs mentioned in Section 2.4.3 rather than still use (𝑖, 𝑗) to denote
the arcs. Then, we have two parallel arc sets 𝒢 𝑆 and 𝒢 𝐹 , where 𝜖𝑝,𝑒
𝑆
∈ 𝒢 𝑆 ⊂ 𝒜 and 𝜖𝑝,𝑒
𝐹
∈

𝒢 𝐹 ⊂ 𝒜. Supposing that the current allocated capacity for two parallel arcs are 𝜇𝑝,𝑒
𝑆 𝐹
and 𝜇𝑝,𝑒
satisfying Eq. (2.5)-(2.7), we will dynamically change the capacity allocation in the revised
pivot operation to reach the optimality and guarantee the feasibility. Here, we suppose that
𝑆 𝐹
𝑥𝑝,𝑒 and 𝑥𝑝,𝑒 are existing flow on the two parallel arcs.
In the revised pivot operation, we need to check each pair of parallel arcs on the cycle 𝑾.
Basically, we need to distinguish three situations for one of two parallel arcs, i.e., (i) the arc

belongs to the forward arc set 𝑾; (ii) the arc belongs to the backward arc set 𝑾; (iii) the arc
does not belong to the cycle 𝑾. Then, for an integration of two parallel arcs, there are nine
scenarios, for each of which we have different ways to change the capacity restrictions for
the parallel arcs.
𝑆 𝐹
Scenario 1: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, this is the scenario that needs the most attention, as
we need to care about their capacity restrictions simultaneously based on Eq. (2.8). To make

Chapter 2: Container Ship Type Decision 24

𝛿 ∗ larger, we need to change the capacity restrictions by solving the following optimization
𝑆 𝐹
problem, where 𝜇𝑝,𝑒 and 𝜇𝑝,𝑒 are decision variables.
𝑆 𝑆 𝐹 𝐹
[𝑴𝟏] max{min{𝜇𝑝,𝑒 − 𝑥𝑝,𝑒 , 𝜇𝑝,𝑒 − 𝑥𝑝,𝑒 }} (2.10)
subject to:
𝐹
𝜇𝑝,𝑒
𝑆
𝜇𝑝,𝑒 + 𝑀
= 𝐾𝑝,𝑒 (2.11)
𝑆 𝑆
𝜇𝑝,𝑒 ≥ 𝑥𝑝,𝑒 (2.12)
𝐹 𝐹
𝜇𝑝,𝑒 ≥ 𝑥𝑝,𝑒 (2.13)
𝑆 𝑆 𝐹 𝐹
where objective (2.10) is equivalent to max{(𝜇𝑝,𝑒 − 𝑥𝑝,𝑒 ) × (𝜇𝑝,𝑒 − 𝑥𝑝,𝑒 )} holding constraints
(2.12) and (2.13). Then, we can easily obtain closed-form solutions for the 𝑴𝟏 such that
𝑆 −𝑥 𝐹
𝑀𝐾𝑝,𝑒 +𝑀𝑥𝑝,𝑒 𝑆 +𝑥 𝐹
𝑀𝐾𝑝,𝑒 −𝑀𝑥𝑝,𝑒
𝑆 ∗ 𝑝,𝑒 𝐹 ∗ 𝑝,𝑒
𝜇𝑝,𝑒 = and 𝜇𝑝,𝑒 = .
2𝑀 2
𝑆 𝐹
Scenario 2: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, as the capacity restriction is not important for arc
𝐹 ∗ 𝐹
𝐹 𝐹 𝐹 ∗ 𝑆 ∗ 𝜇𝑝,𝑒 𝑥𝑝,𝑒
𝜖𝑝,𝑒 based on Eq. (2.9), we adjust 𝜇𝑝,𝑒 = 𝑥𝑝,𝑒 and 𝜇𝑝,𝑒 = 𝐾𝑝,𝑒 − = 𝐾𝑝,𝑒 − such that
𝑀 𝑀
𝑆
the capacity restriction on arc 𝜖𝑝,𝑒 is maximum. Note that in this scenario, we need to pay
more attention on the flow after the revised pivot operation. Supposing that we argument 𝛿 ∗
𝑆 𝑆 𝑆 ∗
on the cycle, and 𝜖𝑝,𝑒 is the leaving arc such that 𝜇𝑝,𝑒 − 𝑥𝑝,𝑒 − 𝛿 ∗ = 0, we can further
𝐹
argument more flow after the revised pivot operation. This is due to the flow on arc 𝜖𝑝,𝑒 will
∗
decrease by 𝛿 ∗ , under which condition we can further decrease its capacity to 𝜇𝑝,𝑒
𝐹 𝐹
= 𝑥𝑝,𝑒 −
𝐹 ∗ 𝐹 −𝛿 ∗
∗ 𝜇𝑝,𝑒 𝑥𝑝,𝑒
𝛿 ∗ and increase the capacity of arc 𝜖𝑝,𝑒
𝑠 𝑆
to 𝜇𝑝,𝑒 = 𝐾𝑝,𝑒 − 𝑀
= 𝐾𝑝,𝑒 − 𝑀
. We repeat the
𝑆
above process until 𝜖𝑝,𝑒 is not the leaving arc. This leads to the following proposition:

𝑆 𝐹 𝑆
Proposition 2.1: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, the arc 𝜖𝑝,𝑒 cannot be the leaving arc.

𝑆 𝐹 𝐹
Scenario 3: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∉ 𝑾, as the arc 𝜖𝑝,𝑒 is not in the cycle, its flow will not
𝐹 𝐹 𝑆 ∗ ∗
change during the revised pivot operation. Then, we adjust 𝜇𝑝,𝑒 = 𝑥𝑝,𝑒 and 𝜇𝑝,𝑒 = 𝐾𝑝,𝑒 −
𝐹 ∗ 𝐹
𝜇𝑝,𝑒 𝑥𝑝,𝑒 𝑆
= 𝐾𝑝,𝑒 − such that the capacity restriction on arc 𝜖𝑝,𝑒 is maximum.
𝑀 𝑀
𝑆 𝐹
Scenario 4: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, the operation is similar to Scenario 2, which is not
repeated here. We can also obtain a proposition that is similar to Proposition 2.1:
𝑆 𝐹 𝐹
Proposition 2.2: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, the arc 𝜖𝑝,𝑒 cannot be the leaving arc.
𝑆 𝐹
Scenario 5: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, we do not change the capacity restrictions.
𝑆 𝐹
Scenario 6: if 𝜖𝑝,𝑒 ∈ 𝑾 and 𝜖𝑝,𝑒 ∉ 𝑾, we do not change the capacity restrictions.

Chapter 2: Container Ship Type Decision 25

 𝑆 𝐹
Scenario 7: if 𝜖𝑝,𝑒 ∉ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, the operation is similar to Scenario 3, which is not
repeated here.
𝑆 𝐹
Scenario 8: if 𝜖𝑝,𝑒 ∉ 𝑾 and 𝜖𝑝,𝑒 ∈ 𝑾, we do not change the capacity restrictions.
𝑆 𝐹
Scenario 9: if 𝜖𝑝,𝑒 ∉ 𝑾 and 𝜖𝑝,𝑒 ∉ 𝑾, we do not change the capacity restrictions.
Until now, we obtain a revised network simplex algorithm, which embeds the above-revised
pivot operation. The intuition behind the revised network simplex algorithm is that the capacity
restrictions are only involved in the pivot operation when determining a leaving arc. Thus, we
can dynamically change the capacity allocation for two parallel arcs in the pivot operation, by
𝐹
𝜇𝑝,𝑒
𝑆
holding 𝜇𝑝,𝑒 + 𝑀
= 𝐾𝑝,𝑒 to guarantee the feasibility and by maximizing 𝛿 ∗ to reach the

optimality. Note that the revised network simplex algorithm is applicable to any minimum cost
flow problems with sharing a capacity restriction among arcs.

A further revision of the network

In the network constructed in Section 2.4, there are some pairs of nodes having two arcs in
the opposite directions, such as the two opposite arcs between node 𝑆ℎ𝑖𝑝 𝑆 _𝑝_𝑒 and 𝑃𝑜𝑟𝑡 𝑆 _𝑝_𝑒
shown in Figure 2.3. When applying the revised network simplex algorithm, it may cause
trouble in the revised pivot operation on increasing/decreasing flow between two nodes. To
avoid the trouble, we introduce some dummy nodes to ensure that between any two nodes,
there is at most one connected arc. Figure 2.8 shows an example when there are two opposite
arcs between node 𝑖 and node 𝑗. By adding two dummy node 𝑖 1 and node 𝑗 1 and reconnecting
arcs, we can deal with the above trouble.

Figure 2.8: Adding dummy nodes to the network

2.5.2 Ship type decision by reduced costs

Given a ship type 𝑣 with the capacity 𝐾 𝑣 and fixed operation cost 𝐶 𝑣 , we can invoke the
above revised network simplex algorithm to derive the minimum flow cost 𝜎 𝑣 for the empty
containers. A straightforward way to find the optimal ship type is to invoke the algorithm to
derive 𝜎 𝑣 for all 𝑣 ∈ 𝑉. Then, we can get the optimal ship type by min{𝐶 𝑣 + 𝜎 𝑣 : ∀𝑣 ∈ 𝑉}.

Chapter 2: Container Ship Type Decision 26

This way is applicable to the practice, as the candidate set 𝑉 for ship types is normally limited
(in computational experiments, there are four ship types for selection). However, in this
subsection, we introduce a reduced cost based bound to exclude some ship types from the
optimal one.
We rank the set of ship types with an increasing container capacity order such that 𝐾 1 <
𝐾 2 < ⋯ < 𝐾 |𝑉| and 𝐶 1 < 𝐶 2 < ⋯ < 𝐶 |𝑉| . Here, 𝐾 1 is the smallest capacity that should be
able to carry all laden containers. We initialize the ship capacity at 𝐾 = 𝐾 1 and run the revised
network simplex algorithm. Supposing that it obtains the optimal spanning tree solution
𝜋
(𝑻𝟏 , 𝑳𝟏 , 𝑼𝟏 ), the solution has reduced costs 𝑐𝑖𝑗 satisfying the Minimum Cost Flow Optimality

Conditions. Starting from the ship capacity 𝐾 1 , if we increase the ship capacity to 𝐾 2 , the
fixed operation cost increases by 𝐶 2 − 𝐶 1 . However, the minimum flow cost may decrease as
we relax the ship capacity from 𝐾 1 to 𝐾 2 . Therefore, when increasing ship capacity, there is a
trade-off between the increasing of fixed operation cost and the decreasing of minimum flow
cost. As the increasing of the fixed operation cost is pre-determined (𝐶 2 − 𝐶 1 ), we need to
judge whether the decreasing of minimum flow cost can compensate for the increasing
operation cost 𝐶 2 − 𝐶 1 .
For notational convenience, we denote ∆1,2 = 𝜎 1 − 𝜎 2 as the decreasing of minimum flow
𝜋
cost by changing Type 1 ship to Type 2 ship. Here, we can use the reduced costs 𝑐𝑖𝑗 at ship
𝜋
capacity 𝐾 = 𝐾 1 to derive an upper bound for ∆1,2 . Recall that 𝑐𝑖𝑗 ≤ 0, ∀(𝑖, 𝑗) ∈ 𝑼 actually

means if the capacity of arc (𝑖, 𝑗) increases by one unit, the minimum flow cost has the
𝜋
potential to decrease by |𝑐𝑖𝑗 |. Considering a minimum cost flow problem is equivalent to
𝜋 𝜋
min{∑(𝑖,𝑗)∈𝑳 𝑐𝑖𝑗 𝑥𝑖𝑗 − ∑(𝑖,𝑗)∈𝑼 |𝑐𝑖𝑗 | 𝑥𝑖𝑗 } (Ahuja et al., 1993), ∆1,2 has the upper bound
𝜋
∑(𝑖,𝑗)∈𝑼|𝑐𝑖𝑗 | (𝐾 2 − 𝐾 1 ) if we do not have the parallel arcs and all arcs (𝑖, 𝑗) ∈ 𝑼 are restricted
by the ship capacity. However, under the situation that our problem has parallel arcs that
sharing capacity restrictions, the increased ship capacity 𝐾 2 − 𝐾 1 may be used by
repositioning standard containers or foldable containers. Thus, we propose Procedure 2.1 to
𝑈
derive the upper bound for ∆1,2 (denoted by ∆1,2 ), in which 𝑐𝜖𝜋𝑆 and 𝑐𝜖𝜋𝐹 denote the reduced
𝑝,𝑒 𝑝,𝑒

costs for the parallel arcs. Note that if we increase one unit ship capacity, the capacity on the
parallel arcs for foldable containers can increase by 𝑀 units. After obtaining the upper bound
𝑈 𝑈
∆1,2 , we can compare it with 𝐶 2 − 𝐶 1 . If ∆1,2 ≤ 𝐶 2 − 𝐶 1 , it is unnecessary to further invoke

the revised network simplex algorithm to derive 𝜎 2 for Type 2 ship, as the ∆1,2 is impossible
to compensate 𝐶 2 − 𝐶 1 . We will repeat the Procedure 1 by increasing ship capacity to 𝐾 𝑣 until
𝑈
we find a ship type 𝑣 such that ∆1,𝑣 ≥ 𝐶 𝑣 − 𝐶 1 , and then invoke the revised network simplex

Chapter 2: Container Ship Type Decision 27

to derive the optimal spanning tree solution (𝑻𝒗 , 𝑳𝒗 , 𝑼𝒗 ). Starting from the ship capacity 𝐾 𝑣 ,
we also use the above method to exclude some ship types, until exploring the ship capacity
𝐾 |𝑉| .

Procedure 2.1. Deriving the upper bound for ∆1,2

𝜋
Input: The reduced cost 𝑐𝑖𝑗 for the spanning tree (𝑇 1 , 𝐿1 , 𝑈1 )
𝑈
Output: The upper bound ∆1,2
𝑈
Initialize ∆1,2 ← 0
𝑆
for all parallel arc 𝜖𝑝,𝑒 ∈ 𝒢 𝑆 and 𝜖𝑝,𝑒
𝐹
∈ 𝒢 𝐹 do
𝑆 𝐹
if 𝜖𝑝,𝑒 ∈ 𝑼 and 𝜖𝑝,𝑒 ∉ 𝑼 then
∆1,2 ← ∆1,2 + (𝐾 − 𝐾 1 )|𝑐𝜖𝜋𝑆 |
𝑈 𝑈 2
𝑝,𝑒
𝑆 𝐹
else if 𝜖𝑝,𝑒 ∉ 𝑼 and 𝜖𝑝,𝑒 ∈ 𝑼 then
∆1,2 ← ∆1,2 + (𝐾 − 𝐾 )𝑀|𝑐𝜖𝜋𝐹 |
𝑈 𝑈 2 1
𝑝,𝑒
𝑆 𝐹
else if 𝜖𝑝,𝑒 ∈ 𝑼 and 𝜖𝑝,𝑒 ∈ 𝑼 then
∆1,2 ← ∆1,2 + (𝐾 − 𝐾 )max{|𝑐𝜖𝜋𝑆 |, 𝑀|𝑐𝜖𝜋𝐹 |}
𝑈 𝑈 2 1
𝑝,𝑒 𝑝,𝑒
end if
end for

To facilitate the understanding, we show an example to exclude ship types. Supposing that
we have invoked the algorithm to derive the reduced costs for the ship capacity at 𝐾 = 𝐾 1, we
𝑈 𝑈
derive ∆1,2 = 120 by Procedure 2.1. Then, if 𝐶 2 − 𝐶 1 = 200 ≥ ∆1,2 , it is unnecessary to

invoke the algorithm for the ship capacity at 𝐾 = 𝐾 2. In the next, we replace 𝐾 2 with 𝐾 3 in
𝑈
Procedure 2.1. If we have ∆1,3 = 300 and 𝐶 3 − 𝐶 1 = 250, we need to invoke the algorithm

for the ship capacity at 𝐾 = 𝐾 3 .

Based on the upper bound derived by reduced costs, we can skip some ship types when
invoking the revised network simplex algorithm. In the end, we compare the total costs of all
those ship types explored by the algorithm for finding the optimal ship type.
In summary for the proposed solution approach, we have made the following improvements
compared with the traditional network simplex algorithm. (i) Our network flow model has
some parallel arcs sharing the same capacity restrictions, which are inevitable due to the co-
existence of foldable and standard containers in the network (cf. Section 2.4.3). As the capacity
restrictions tackled in the traditional network simplex algorithm must be unchanged with given
values, we propose a tailored (or revised) network simplex algorithm for the problem. In the
revised one, we introduce dynamic capacity restrictions for the parallel arcs and revise the
pivot operations to guarantee the optimality (cf. Section 2.5.1). (ii) Our problem needs to
determine the ship capacity, which further decides capacity restrictions on the arcs of the
network. To achieve the goal, we use the reduced costs from our tailored algorithm to make
the ship type decision (cf. Section 2.5.2). (iii) Using a network simplex algorithm to solve a

Chapter 2: Container Ship Type Decision 28

minimum cost flow problem requires the construction of a network flow model as the prior
task. Our solution approach elaborates a procedure to construct a flow network (cf. Section
2.4) for the empty container allocation.

2.5.3 A MILP model

The proposed solution approach is built on the network flow model constructed in Section
2.4. In this subsection, we formulate a MILP model for our research problem that can be solved
by CPLEX solver directly. In the computational experiment section, we will use the results of
the MILP model to verify the optimality of the proposed solution approach. Here, we first
introduce some decision variables and then provide the model formulation.

Decision variables:
𝜀𝑣 : binary, set to one if the ships in Type 𝑣 are deployed, otherwise zero, 𝑣 ∈ 𝑉;
𝜁𝑝𝑆 : integer, number of long-term leasing standard containers in Port 𝑝, 𝑝 ∈ 𝑃;

𝜁𝑝𝐹 : integer, number of long-term leasing foldable containers in Port 𝑝, 𝑝 ∈ 𝑃;

𝑆
𝛼𝑜𝑑,𝑒 : integer, number of empty standard containers used to satisfy the laden container

transportation consignments from Port 𝑜 to Port 𝑑 by the 𝑒 𝑡ℎ round trip, (𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸;

𝐹
𝛼𝑜𝑑,𝑒 : integer, number of empty foldable containers used to satisfy the laden container

transportation consignments from Port 𝑜 to Port 𝑑 by the 𝑒 𝑡ℎ round trip, (𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸;

𝑆
𝛽𝑜𝑑,𝑒 : integer, number of empty standard containers repositioned from Port 𝑜 to Port 𝑑 by

the 𝑒 𝑡ℎ round trip, (𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸;

𝐹
𝛽𝑜𝑑,𝑒 : integer, number of empty foldable containers repositioned from Port 𝑜 to Port 𝑑 by

the 𝑒 𝑡ℎ round trip, (𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸;

𝛾𝑜𝑑,𝑒 : integer, number of short-term empty containers leased in Port 𝑜 for the 𝑒 𝑡ℎ round trip
and will returned in Port 𝑑, (𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸;
𝑆
𝛿𝑝,𝑒 : integer, inventory level of empty standard containers (i.e., the empty standard

containers left) at Port 𝑝 after visited by the 𝑒 𝑡ℎ round trip, 𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸;

𝐹
𝛿𝑝,𝑒 : integer, inventory level of empty foldable containers (i.e., the empty foldable

containers left) at Port 𝑝 after visited by the 𝑒 𝑡ℎ round trip, 𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸;

𝜂𝑝,𝑒 : integer, number of containers carried on the container ship of the 𝑒 𝑡ℎ round trip after
it visits Port 𝑝, 𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸;

Mathematical model:
[𝑴𝟐] min ∑𝑣∈𝑉 𝐶𝑣 𝜀𝑣 + ∑𝑝∈𝑃(𝐿𝑆𝑝 𝜁𝑝𝑆 + 𝐿𝐹𝑝 𝜁𝑝𝐹 ) + ∑𝑒∈𝐸 ∑(𝑜,𝑑)∈𝐷 𝑙𝑜𝑑 𝛾𝑜𝑑,𝑒 +
𝑆 𝑆 𝐹 𝐹
∑𝑒∈𝐸 ∑(𝑜,𝑑)∈𝐷(𝑟𝑜𝑑 𝛽𝑜𝑑,𝑒 + 𝑟𝑜𝑑 𝛽𝑜𝑑,𝑒 ) + ∑𝑒∈𝐸 ∑𝑝∈𝑃(𝑠𝑝𝑆 𝛿𝑝,𝑒
𝑆
+ 𝑠𝑝𝐹 𝛿𝑝,𝑒
𝐹
) (2.14)

Chapter 2: Container Ship Type Decision 29

subject to:
∑𝑣∈𝑉 𝜀𝑣 = 1 (2.15)
𝑆 𝐹
𝛼𝑜𝑑,𝑒 + 𝛼𝑜𝑑,𝑒 + 𝛾𝑜𝑑,𝑒 = 𝑑𝑜𝑑,𝑒 ∀(𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸, (2.16)
𝑆 𝑆 𝑆
𝜁𝑝𝑆 + ∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒 = ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷(𝛼𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 𝑆
) + 𝛿𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 = 1,
(2.17)
𝐹 𝐹 𝐹
𝜁𝑝𝐹 + ∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒 = ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷(𝛼𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 𝐹
) + 𝛿𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 = 1,
(2.18)
𝑆 𝑆 𝑆
𝛿𝑝,𝑒−1 + ∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒 + ∑𝑑=𝑝,𝑜>𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒−𝑁 +
𝑆
∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛼𝑜𝑑,𝑒−𝑤 𝑆 𝑆 𝑆
𝑝
+ ∑𝑑=𝑝,𝑜>𝑑,(𝑜,𝑑)∈𝐷 𝛼𝑜𝑑,𝑒−𝑁−𝑤𝑝
= ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷(𝛼𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 )+
𝑆
𝛿𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸\{1}, (2.19)
𝐹 𝐹 𝐹
𝛿𝑝,𝑒−1 + ∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒 + ∑𝑑=𝑝,𝑜>𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒−𝑁 +
𝐹
∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛼𝑜𝑑,𝑒−𝑤 𝐹 𝐹 𝐹
𝑝
+ ∑𝑑=𝑝,𝑜>𝑑,(𝑜,𝑑)∈𝐷 𝛼𝑜𝑑,𝑒−𝑁−𝑤𝑝
= ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷(𝛼𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 )+
𝐹
𝛿𝑝,𝑒 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸\{1}, (2.20)
𝑆 𝐹
𝜂𝑝,𝑒 = ∑𝑜≤𝑝,𝑑>𝑝,(𝑜,𝑑)∈𝐷(𝑑𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 /𝑀) + ∑𝑑≥𝑝+1,𝑜>𝑑,(𝑜,𝑑)∈𝐷(𝑑𝑜𝑑,𝑒−𝑁 +
𝑆 𝐹
𝛽𝑜𝑑,𝑒−𝑁 + 𝛽𝑜𝑑,𝑒−𝑁 /𝑀) ∀𝑝 ∈ 𝑃/{𝑃}, 𝑒 ∈ 𝐸, (2.21)
𝑆 𝐹
𝜂𝑝,𝑒 = ∑𝑜>𝑑,(𝑜,𝑑)∈𝐷(𝑑𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 /𝑀) ∀𝑝 = |𝑃|, 𝑒 ∈ 𝐸 (2.22)
𝜂𝑝,𝑒 ≤ ∑𝑣∈𝑉 𝐾𝑣 𝜀𝑣 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸, (2.23)
𝜀𝑣 ∈ {0,1} ∀𝑣 ∈ 𝑉, (2.24)
𝜁𝑝𝑆 , 𝜁𝑝𝐹 ≥ 0 ∀𝑝 ∈ 𝑃, (2.25)
𝑆 𝐹 𝑆 𝐹
𝛼𝑜𝑑,𝑒 , 𝛼𝑜𝑑,𝑒 , 𝛽𝑜𝑑,𝑒 , 𝛽𝑜𝑑,𝑒 , 𝛾𝑜𝑑,𝑒 ≥ 0 ∀(𝑜, 𝑑) ∈ 𝐷, 𝑒 ∈ 𝐸, (2.26)
𝑆 𝐹
𝛿𝑝,𝑒 , 𝛿𝑝,𝑒 , 𝜂𝑝,𝑒 ≥ 0 ∀𝑝 ∈ 𝑃, 𝑒 ∈ 𝐸. (2.27)
𝑆 𝑆 𝐹 𝐹
Here, note that 𝛼𝑜𝑑,𝑒 : = 0, 𝛽𝑜𝑑,𝑒 : = 0, 𝛼𝑜𝑑,𝑒 : = 0 and 𝛽𝑜𝑑,𝑒 : = 0 when 𝑒 < 1.
In the above model, Objective (2.14) minimizes the total cost, including the fixed operation
cost, the long-term leasing cost, the short-term leasing cost, the repositioning cost, and the
storage cost. Constraint (2.15) guarantees that only one type of ships can be selected.
Constraints (2.16) enforce that the laden container transportation consignments in each port
by each round trip must be fulfilled by using available empty standard containers, foldable
containers or short-term leasing containers in the port. Constraints (2.17) provide the inventory
equations for empty standard containers in each port after visited by the 1𝑠𝑡 round trip. The
left sides of the equations list the number of leased empty containers in each port at the
beginning (i.e., 𝜁𝑝𝑆 ) and the empty containers arrived in each port by the 1𝑠𝑡 round trip (i.e.,
𝑆
∑𝑑=𝑝,𝑜<𝑑,(𝑜,𝑑)∈𝐷 𝛽𝑜𝑑,𝑒 ). The right sides of the equations show the inventory level of empty

Chapter 2: Container Ship Type Decision 30

containers in each port after the 1𝑠𝑡 round trip (i.e., 𝛿𝑝,𝑒
𝑆
) and the number of empty containers
𝑆 𝑆
flowing out of the port by the 1𝑠𝑡 round trip (i.e., ∑𝑜=𝑝,(𝑜,𝑑)∈𝐷(𝛼𝑜𝑑,𝑒 + 𝛽𝑜𝑑,𝑒 )). Constraints
(2.18) are similar as Constraints (2.17), which are the inventory equations for empty foldable
containers by the 1𝑠𝑡 round trip. Constraints (2.19) and Constraints (2.20) list the inventory
equations in Port 𝑝 after visited by the 𝑒 𝑡ℎ round trip. The left-sides of the constraints also
show the empty container supply, including the empty containers left in the port after visited
by (𝑒 − 1)𝑡ℎ round trip (i.e., the inventory level in Port 𝑝 after visited by (𝑒 − 1)𝑡ℎ round
trip), the arrived repositioning empty containers, and the arrived devanning laden containers.
Here, notice that: (i) if Port 𝑝 is the destination port 𝑑, and the origin port 𝑜 < 𝑑 (the origin
port 𝑜 > 𝑑), the containers transported from Port 𝑜 to Port 𝑑 by the 𝑒 𝑡ℎ round trip will arrive
at the Port 𝑑 by the 𝑒 𝑡ℎ round trip (the (𝑒 + 𝑁)𝑡ℎ round trip). (ii) When laden containers arrive
at the destination ports, it will take 𝑤𝑝 weeks for the devanning of the containers and become
available empty containers for the next consignment in the ports. The right sides of Constraints
(2.19) and Constraints (2.20) are the same as that of Constraints (2.17) and Constraints (2.18)
respectively. Constraints (2.21) and Constraints (2.22) calculate the number of containers
carried in the deployed ship on the 𝑒 𝑡ℎ round trip after visiting Port 𝑝. Constraints (2.23)
enforce that the number of containers carried in the ships cannot exceed the capacity of the
type of the deployed ships. Constraints (2.24-2.27) define the decision variables.

2.6 COMPUTATIONAL EXPERIMENT

In this section, based on three real-world shipping service routes, we conduct extensive
computational experiments to find insights on the ship type decision and the foldable container
usage by using our proposed solution approach. We run the experiments by a PC equipped
with 3.30GHz of Intel Core i5 CPU and 16GB of RAM. For all the test instances in the
experiments, the planning horizon is half a year, i.e., 26 weeks. As the shipping service is on
a weekly basis, the total round trips are 26 (i.e., |𝐸|=26).

2.6.1 Test instances on three real-world shipping service routes

To generate test instances for the experiments, we select three real-world shipping service
routes operated by CMA CGM shipping liner, which is labeled as BOHAI, LIBERTY2, and
AANAANLCMA. Figure 2.9 depicts the three real-world shipping service routes. (i) The port
rotation of the route BOHAI is Lianyungang (1) → Shanghai (2) → Ningbo (3) → Los Angeles
(4) → Oakland (5) → Lianyungang; the rotation time is 42 days, and 6 ships are deployed. (ii)
The port rotation of the route LIBERTY2 is Antwerp (1) → Bremerhaven (2) → Rotterdam

Chapter 2: Container Ship Type Decision 31

(3) → Le Havre (4) → New York (5) → Norfolk (6) → Charleston (7) → Antwerp; the rotation
time is 28 days, and 4 ships are deployed. (iii) The port rotation of the route AANAANLCMA
is Yokohama (1) → Osaka (2) → Pusan (3) → Shanghai (4) → Ningbo (5) → Kaohsiung (6)
→ Melbourne (7) → Sydney (8) → Brisbane (9) → Yokohama; the rotation time is 42 days,
and 6 ships are deployed. In fact, the three shipping service routes are the representatives of
three trade routes: Asia-North America trade route, North Europe-North America trade route,
and Australia-Far East trade route. The three routes have different degrees of the imbalance of
laden container flow.

Figure 2.9: Three selected shipping services routes operated by CMA CGM (CMA CGM,

2017)

According to Word Shipping Council (2013), we estimate that the imbalance ratio of laden
container flow for Asia-North America trade route is about 1.99 (i.e., Eastbound / Westbound),

Chapter 2: Container Ship Type Decision 32

the imbalance ratio for North Europe-North America trade route is about 1.27 (i.e., Westbound
/ Eastbound), and the imbalance ratio for Australia-Far East trade route is about 1.73 (i.e.,
Southbound / Northbound). In fact, different imbalance ratios lead to different degrees of the
necessity for the empty container repositioning. Based on the imbalance ratios, the weekly
laden container transportation consignments are randomly generated for each selected
shipping service route as follows: (i) For the route BOHAI, the transportation consignments
from an Asian port to a North American port follow the uniform distribution 𝑈(600, 800),
and the transportation consignments from a North American port to an Asian port follow the
uniform distribution 𝑈(300, 400) . (ii) For the route LIBERTY2, the transportation
consignments from a North Europe port to a North American port follow the uniform
distribution 𝑈(400,500), and the transportation consignments from a North American port to
a North Europe port follow the uniform distribution 𝑈(300, 400) . (iii) For the route
AANAANLCMA, the transportation consignments from an Asian port to an Australian port
follow the uniform distribution 𝑈(500, 700), and the transportation consignments from an
Australian port to an Asian port follow the uniform distribution 𝑈(300, 400). For all the three
routes, there are no transportation consignments within the same region.

Table 2.1: Data for relevant container costs

Relevant container costs Per foldable container Per standard container

Weekly storage cost US$10 US$40
Loading or unloading cost US$13 US$50
Long-term leasing cost US$960 US$480
Folding and unfolding cost US$20

Table 2.2: Candidate ship type and fixed operation cost

Ship type (TEUs) and fixed operating costs (million US$)

Ship route
4,400 5,000 5,400 5,800 6,200 11,000 11,400 11,800 12,000
BOHAI 14.07 15.00 15.21 15.32
LIBERTY2 8.62 9.53 10.20 10.48
AANAANLCMA 22.23 23.00 23.44 23.52

By referring to Moon and Hong (2016), and Konings (2005), we set all the relevant costs
for foldable and standard empty containers, which are shown in Table 2.1. Here, we assume
that all the costs are the same for different ports, and four foldable containers can be folded as
one standard container. The short-term leasing cost is charged based on the travel time between
origin ports and destination ports. According to Moon and Hong (2016), the unit short-term
leasing cost is set as US$170/week, for example, if the travel time between an origin port and
a destination port is two weeks, the short-term leasing cost is US$340. For the three selected

Chapter 2: Container Ship Type Decision 33

shipping service routes, the candidate ship types and the fixed operating cost for the whole
planning horizon by each ship type are presented in Table 2.2 (Meng and Wang, 2012).

2.6.2 Optimality check for the proposed solution approach

In Section 2.5, we propose the solution approach and formulate the MILP model for the
problem. Here, we apply the two methods to solve problem instances of three shipping routes.
The test instances are randomly generated by using the parameter settings in Section 2.6.1.
The results of using the proposed solution approach and using the MILP model by CPLEX
solver are given in Table 2.3. As can be seen, both methods derive the same optimal solution
for the problem, which verifies the optimality of the proposed solution approach. With respect
to the computation time, both methods outperform each other in some instances and on
average, using the proposed solution is slightly faster than using the MILP model by CPLEX
solver. Since CPLEX solver is a commercial solver and does not outperform the proposed
method, the proposed solution embedded with the revised network simplex algorithm is more
desirable for shipping liner companies, as it does not invoke any MILP solvers.

Table 2.3: Comparing the proposed method with the MILP model by CPLEX solver

Instance The solution approach The MILP model

Time
Instance Z(million CPU time Z(million CPU time ratio
Shipping route
ID US$) (seconds) US$) (seconds)
B_3_1 36.32 3 36.32 3 1.00
B_3_2 37.45 4 37.45 6 0.67
BOHAI B_3_3 38.64 2 38.64 3 0.67
B_3_4 37.71 3 37.71 6 0.50
B_3_5 38.06 3 38.06 3 1.00
L_3_6 33.78 11 33.78 15 0.72
L_3_7 33.45 24 33.45 21 1.14
LIBERTY2 L_3_8 32.97 12 32.97 15 0.77
L_3_9 33.63 24 33.63 20 1.21
L_3_10 33.85 21 33.85 17 1.24
A_3_11 83.74 58 83.74 59 0.99
A_3_12 82.67 55 82.67 53 1.04
AANAANLCMA A_3_13 83.31 70 83.31 74 0.94
A_3_14 82.78 97 82.78 83 1.16
A_3_15 82.46 83 82.46 95 0.87
Average: 0.93
Note: (i) “𝑍(million US$)” represents the objective values derived by the two methods with the unit of
million US$. (ii) “CPU time” shows the computation time with the unit of seconds. (iii) “Time ratio” is
the CPU time of the solution approach divided by the CPU time of the MILP model by CPLEX solver.

2.6.3 Comparing a model that only considers laden container transportation

One of the major motivations of this paper is to incorporate the empty container allocation
into the ship type decision for the ship fleet deployment. Here, we conduct experiments to

Chapter 2: Container Ship Type Decision 34

compare our proposed model with the model that does not consider the empty containers on
the problem. For the traditional model, we decide the ship type by only considering the laden
containers, by which we can calculate the container flows for all the voyages. Then, the ship
type that has the minimum capacity to accommodate all the container flows is determined as
the selected ship type. Assuming that the capacity and the fixed operating cost for the selected
ship type is 𝐾 1 and 𝐶 1 , we run the revised network simplex algorithm only to find the
minimum flow cost for empty containers, which determines the total cost 𝐶 1 + 𝜎 1 for the
traditional model.

Table 2.4: Comparing the proposed model and the traditional model without empty

container allocation

Instance The proposed model The traditional model

Instance 𝑍1(million 𝑍2(million Gap
Shipping route Ship fleet Ship fleet
ID US$) US$)
B_4_1 38.13 5,400 38.39 5,000 0.68%
B_4_2 37.68 5,000 38.00 4,400 0.86%
BOHAI B_4_3 37.10 5,000 37.37 4,400 0.74%
B_4_4 37.72 5,400 37.72 5,400 0.00%
B_4_5 38.38 5,800 38.74 5,400 0.93%
Imbalance ratio of laden container flow: 1.99; # of ports: 5 Average: 0.64%
L_4_6 33.13 5,800 33.28 5,400 0.46%
L_4_7 34.04 6,200 34.21 5,800 0.51%
LIBERTY2 L_4_8 33.61 5,800 33.77 5,400 0.49%
L_4_9 33.26 5,400 33.26 5,400 0.00%
L_4_10 33.13 5,400 33.13 5,400 0.00%
Imbalance ratio of laden container flow: 1.27; # of ports: 7 Average: 0.29%
A_4_11 84.26 12,000 85.53 11,400 1.51%
A_4_12 83.14 11,400 84.37 11,000 1.48%
AANAANLCMA A_4_13 82.93 11,400 83.93 11,000 1.21%
A_4_14 83.96 11,800 85.09 11,400 1.35%
A_4_15 83.16 11,400 84.08 11,000 1.11%
Imbalance ratio of laden container flow: 1.73; # of ports: 9 Average: 1.33%

Note: (i) 𝑍1(million US$) and 𝑍2(million US$) represent the objective values of the two models with
the unit of million US$. (ii) “Ship fleet” shows the capacity (in TEUs) of the selected ship type. (iii)
“Gap” shows the difference on the total costs by the two models, which is calculated by (𝑍2 − 𝑍1)/𝑍1.

The comparison between the proposed model and the traditional model is listed in Table
2.4, where “Ship fleet” shows the capacity of the selected ship fleet types by the two models,
and “Gap” shows the difference in the total costs by the two models. In the majority of the
instances in Table 2.4, the selected ship fleet types are different by the two models. More
importantly, the ship fleet capacity selected by the proposed model is no less than the ship
fleet capacity selected by the traditional model, which is consistent with our previous

Chapter 2: Container Ship Type Decision 35

discussion, i.e., considering the empty containers gives the shipping liner motivation to deploy
larger container ships. However, the impact of considering the empty container allocation
varies among the three shipping routes. For the route AANAANLCMA, which does not
consider the empty container allocation, the total cost rises by near 1.33% on average; but for
the route LIBERTY2, the total cost rises by near 0.29% on average. Here, note that if the
selected ship fleet types by the two models are the same, the total costs are the same, because
we also optimize empty container related decisions in the traditional model.
The above-mentioned phenomenon attributes to the different trade routes that the three
shipping routes belong. The route AANAANLCMA is one shipping route of Australia-Far
East trade route, and the route LIBERTY2 is one shipping route of North Europe-North
America trade route. As a result, the route AANAANLCMA has a higher imbalance ratio of
laden container flow than the route LIBERTY2, which makes the empty container
repositioning more necessary for the route AANAANLCMA. Thus, for the route
AANAANLCMA, it is more important to consider empty containers in the ship type decision,
which is verified by the high total cost gaps in “Gap”. For the route BOHAI, although the
imbalance of laden container flow for the route is significant, the total cost gap is not so
obvious compared with the route AANAANLCMA, which is due to that only five ports of call
are involved in the route BOHAI. Based on the above analysis, we can conclude that
considering the empty container allocation is critical for the ship fleet deployment, especially
for the shipping routes that have high imbalance ratios of laden container flow, and the
shipping routes that traverse many ports of call.

2.6.4 Performance evaluation of using foldable containers

Although researchers have proved that the economic and logistical viability of using
foldable container, the foldable containers still are not prevalent among the shipping services.
In this section, we aim to investigate how much cost the shipping liner can save if the foldable
containers are truly used in shipping services. In this subsection, we conduct some experiments
on comparing the proposed model (i.e., 𝑻𝑵𝑭 model and ship type decision) with the model
that does not use the foldable containers (i.e., 𝑷𝑵𝑭 model and ship type decision). Note that
we still use the solution approach proposed in Section 2.5 to derive the optimal container flow
and the optimal ship type for the two cases. The results are reported in Table 2.5, where
“Container fleet” shows the total number of containers used for the complete planning horizon,
and “Gap” shows the difference on the total costs by the two models.

Chapter 2: Container Ship Type Decision 36

 Table 2.5: Comparison between the proposed model and the model that does not use the

foldable containers

Instance With foldable containers Without foldable containers

Instance Z(million Ship Container Z(million Ship Container Gap
Shipping route
ID US$) fleet fleet US$) fleet fleet
B_5_1 37.88 5,000 38,474 37.98 5,000 38,451 0.27%
B_5_2 38.12 5,400 38,825 38.27 5,800 38,367 0.40%
BOHAI B_5_3 37.88 5,000 38,367 38.01 5,400 38,290 0.34%
B_5_4 37.96 5,400 38,605 38.06 5,400 38,460 0.26%
B_5_5 38.16 5,800 38,880 38.26 5,800 38,429 0.28%
Average: 0.31%
L_5_6 33.13 5,400 42,492 33.15 5,400 42,485 0.05%
L_5_7 33.23 5,800 42,555 33.25 5,800 42,555 0.03%
LIBERTY2 L_5_8 33.24 5,800 42,494 33.25 5,800 42,494 0.04%
L_5_9 33.22 5,800 42,621 33.25 5,800 42,603 0.08%
L_5_10 33.05 5,400 42,348 33.06 5,400 42,348 0.04%
Average: 0.05%
A_5_11 82.84 11,400 99,516 83.30 11,800 99,316 0.56%
A_5_12 83.12 11,800 100,110 83.63 12,000 100,013 0.62%
AANAANLCMA A_5_13 82.80 11,400 99,597 83.27 11,400 99,325 0.56%
A_5_14 82.71 11,000 99,548 83.23 11,400 99,361 0.63%
A_5_15 83.25 11,800 99,733 83.69 12,000 99,645 0.53%
Average: 0.58%

Note: (i) 𝑍1(million US$) and 𝑍2(million US$) represent the objective values with the unit of million
US$. (ii) “Ship fleet” shows the capacity (in TEUs) of the selected ship type. (ii) “Container fleet”
shows the total number of empty containers used for the whole planning horizon, including the total
number of the empty containers owned initially and the long-term leasing empty containers. (iii) “Gap”
shows the difference on the total costs by the two models, which is calculated by (𝑍2 − 𝑍1)/𝑍1.

In Table 2.5, the container fleet (i.e., the total number of containers) using foldable
containers is no less than the container fleet that does not use foldable containers. This result
suggests that using foldable containers motivates the shipping liner to enlarge its container
fleet, which makes it more powerful to handle the laden container demands. Meanwhile, there
is nearly no advantage on cost reduction for the route LIBERTY2 by using foldable containers
as the total cost gaps are less than 0.10%, and using foldable containers has the biggest impact
on the route AANAANLCMA among the three routes. The results are in line with the results
of Table 2.4. However, the impacts of using foldable containers on the total cost for all the
three routes are not so significant as all the total cost gaps are less than 0.70%, which implies
that under the current cost settings, the shipping liner does not have a strong incentive on cost
reduction to use the foldable containers. This may be the reason why the foldable containers
are still not prevalent among the shipping services. To find strong incentives for using the
foldable containers for the shipping liner, we will analyze the effects of cost settings on the
foldable container usage in Section 2.6.5. Meanwhile, Table 2.5 shows the foldable container
usage could affect the ship type decision. In two instances of the route BOHAI (Instance

Chapter 2: Container Ship Type Decision 37

B_6_2 and B_6_3) and all instances of the route AANAANLCMA except Instance A_6_13,
after using foldable containers, it will choose a smaller ship fleet. It may attribute to that after
using foldable containers the empty container repositioning will occupy less storage space on
ships when repositioning foldable containers. Thus, a smaller ship fleet may be enough for
carrying all laden and empty containers.

2.6.5 Effects of cost settings on the foldable container usage

In this section, our goal is to find under which conditions, the shipping liner would use
foldable containers on large scale. Compared with standard containers, foldable containers
have higher long-term leasing cost and extra folding and unfolding cost. Therefore, we test the
effects of the long-term leasing cost and the folding and unfolding cost on the foldable
container usage in the section. We define a ratio 𝜌 = ∑𝑝∈𝑃 𝜁𝑝𝐹 /(∑𝑝∈𝑃 𝜁𝑝𝐹 + ∑𝑝∈𝑃 𝜁𝑝𝑆 ) to show

the usage of foldable containers, where 𝜁𝑝𝑆 and 𝜁𝑝𝐹 are the number of standard containers and
foldable containers used for the long-term container leasing. The bigger the ratio is the more
foldable containers are leased for the usage of the planning horizon.

Figure 2.10: Effect of the long-term leasing cost on the foldable container usage

Figure 2.10 illustrates the effect of the long-term leasing cost on the foldable container
usage, where the y-axis shows the ratio 𝜌, and x-axis indicates the long-term leasing cost of a
foldable container. In Figure 2.10, we keep the long-term leasing cost of a standard container
unchanged (i.e., US$480), and increase the long-term leasing cost of a foldable container from
US528 to US$1200. As can be seen, all three curves for the three shipping routes descend fast
when the long-term leasing cost increases, and a formal proof (See Appendix B) can verify the
non-increasing trend of using foldable containers. Under the current cost setting, i.e., the long-

Chapter 2: Container Ship Type Decision 38

term leasing cost of a foldable container is US$960, the foldable container usage is in low
ratio. If the cost reduces to US$768, the shipping liner can have equal usage on both standard
containers and foldable containers. However, if the cost is beyond US$1056, there is no need
to consider the usage of foldable containers. Therefore, we can summarize that the foldable
container usage is highly dependent on the long-term leasing cost, and reducing the long-term
leasing cost could be an effective way to make the foldable containers become prevalent.
Under the different long-term leasing cost of foldable containers, we also compare the
average total cost gap between the case with using foldable containers and the case without
using foldable containers, the results of which are listed in Table 2.6. As can be seen, when
the long-term leasing cost decreases, using foldable containers saves the total cost more
significantly. If the long-term leasing cost drops to US$528, the total saving reaches 27.51%.
However, if the long-term leasing cost is beyond US$1104, it makes no difference between
the case by using foldable containers and the case without using foldable containers.

Table 2.6: Total cost saving after using foldable containers under different long-term leasing

cost

long-term leasing
528 576 624 672 720 768 816
cost
Total cost saving 27.51% 22.26% 17.04% 12.49% 8.56% 5.44% 3.19%
long-term leasing
864 912 960 1008 1056 1104 1152
cost
Total cost saving 1.83% 0.86% 0.26% 0.11% 0.02% 0.00% 0.00%

Figure 2.11: Effect of the folding and unfolding cost on foldable container usage

Chapter 2: Container Ship Type Decision 39

 Figure 2.11 shows the effect of the folding and unfolding cost on the foldable container
usage, where the y-axis shows the ratio 𝜌, and the x-axis indicates the folding and unfolding
cost of a foldable container (US$20 in the current cost setting). Here, we increase the folding
and unfolding cost of a foldable container from US10 to US$30. In the figure, the three curves
of three shipping service routes are almost flat, and the foldable container usage maintains at
around 8.5% for the route AANAANLCMA, around 6.4% of the route BOHAI, and around
3.0% for the route LIBERTY2. Thus, we can conclude that the foldable container usage is not
sensitive to the folding and unfolding cost, and the reduction of the folding and unfolding cost
will not spur the foldable container usage significantly.
In summary, the foldable container usage is highly dependent on the long-term leasing cost,
but it is not sensitive to the folding and unfolding cost in the current cost settings. The reason
why the usage of foldable containers shows an insensitive reaction to the folding and unfolding
cost may attribute to the dominant role of the long-term leasing cost of using foldable
containers. According to Figure 2.10, when the long-term leasing cost of foldable containers
is US$960, the percentage of the foldable container usage is low for the three shipping routes
(below 10%), indicating that using standard containers is much more cost-effective than using
foldable containers. Under the case, the decreasing of the folding and unfolding cost may not
be a comparatively strong incentive for the shipping lines to use foldable containers.

2.6.6 Analysis of cost-dependent sensitivity

The previous subsection shows that when the long-term leasing cost is high, the foldable
container usage has a low sensitivity in response to the folding and unfolding cost. In this
subsection, we aim to explore the relationship between the foldable container usage’s
sensitivity to the folding and unfolding cost and the long-term leasing cost. To investigate
whether the long-term leasing cost affects the sensitivity to the folding and unfolding, we focus
on the route AANAANLCMA and conduct sensitivity analysis for the folding and unfolding
under different long-term leasing cost. Given a long-term leasing cost, in order to measure the
𝜌1 −𝜌0
sensitivity in the same metric, we define a sensitivity ratio 𝜎 = , where 𝜌0 is the
𝜌0

percentage of foldable containers used under US$20 folding and unfolding cost (the baseline
setting), and 𝜌1 is the percentage under other costs, such as US$10 and US$30. In the
experiments, we set the long-term leasing cost to US$960, US$860, US$760, US$660 and
US$560, under each of which, we change the folding and unfolding cost to detect the
sensitivity.
Figure 2.12 shows the relationship between the sensitivity to the folding and unfolding cost
and the long-term leasing cost, where the y-axis shows the sensitivity ratio 𝜎, and x-axis

Chapter 2: Container Ship Type Decision 40

indicates the folding and unfolding cost. See the highest “blue” point for an example to depict
the figure, which shows that, given the long-term leasing cost as US$860, when the folding
and unfolding cost decreases from US$20 to US$5, the percentage of foldable container usage
will increase by 54%. Each line in the figure corresponds to the sensitivity under each long-
term leasing cost, and a steeper line means the foldable container usage is more sensitive to
the folding and unfolding cost. In general, Figure 2.12 illustrates that the sensitivity to the
folding and unfolding cost depends on the long-term leasing cost. More specifically, when the
long-term leasing cost decreases from US$960 to US$860 (resp. decreases from US$860 to
US$560), the line becomes steeper (resp. smoother) such that the foldable container usage
becomes more sensitive (resp. less sensitive) to the folding and unfolding cost.

Figure 2.12: Long-term leasing cost dependent sensitivity to the folding and unfolding cost

An intuitive explanation for the phenomenon is that when the long-term leasing cost drops
to a certain level (say US$860), using foldable containers becomes nearly the same cost-
effective as using standard containers. At that level, the changes in the folding and unfolding
cost may have evident effectiveness towards the foldable container usage. However, if the
long-term leasing cost further reduces to a low level (say US$560), using foldable containers
is much more cost-effective than using standard containers (cf. Figure 2.10, approximately
97.8% foldable container usage). Henceforth, decreasing the folding and unfolding cost may
have a limited incentive to increase the foldable container usage. Based on the phenomenon,
we can obtain some managerial insights for shipping lines when popularizing the usage of
foldable containers. (i) If container leasing companies charge a high price for the long-term
leasing of foldable containers, it may not be economic to use foldable containers considering

Chapter 2: Container Ship Type Decision 41

some fixed operation costs incurred for using the foldable containers. (ii) If container leasing
companies charge a moderate price, the shipping lines may make efforts to cut down the
folding and unfolding cost by negotiating with port terminals, which can lead to a profitable
result. (iii) If container leasing companies charge a low price, it is cost-effective to use foldable
containers and the bargaining motivation of shipping lines on the folding and unfolding cost
with port terminals might not be strong, as the cost reduction leads to a tiny benefit.

2.6.7 Sensitivity analysis on the number of weeks for the devanning process
In Section 2.4.1, there is an input parameter 𝑤𝑝 showing the number of weeks allowed for
the devanning process. As discussed in Section 2.3.3, this parameter indicates the required
time for consignees to return empty containers. In essential, the parameter constructs a trade-
off between shipping liner companies and customers (or consignees). If 𝑤𝑝 becomes larger,
the customers will have more flexibility to deal with the cargo carried in containers, but this
would increase the opportunity cost for shipping liner companies, as they need empty
containers as soon as possible to fulfill next transportation consignments. Here, to investigate
the opportunity cost, we conduct a sensitivity analysis on the parameter 𝑤𝑝 , the results of
which are given in Table 2.7.

Table 2.7: Sensitivity analysis on the number of weeks allowed for returning empty

containers

No. of weeks 0 1 2 3 4 5 6 7
Total cost 70.93 75.53 85.48 96.13 107.35 118.80 129.98 140.81
Slope 4.60 9.95 10.65 11.22 11.45 11.18 10.83 NaN
Note: (i) “Total cost” with a unit of million US$ denotes the total cost on average of ten randomly
generated instances. (ii) “Slope” shows the increasing rate at a specific number of weeks, for
example, at “0” weeks, the rate is (75.53 − 70.93)/(1 − 0) = 4.60. (iii) “0” week for the devanning
process is unrealistic in the real-world operations. The “0” week setting only serves as a benchmark in
the sensitivity analysis.

In the table, when the number of weeks allowed for the devanning process increases, the
total cost grows significantly, which suggests the high opportunity cost for allowing more
devanning time. Meanwhile, the slope of total cost increases before reaching “4” weeks and
decreases after “4” weeks, which reveals that the total cost is increasing convex in the number
of weeks first and then becomes increasing concave. More importantly, given “0” weeks as
the benchmark, we can see that from “0” week to “1” week, the total cost increases by 4.60
million US$, which is far less than other increasing rates (e.g., 9.95 million US$ from “1”
week to “2” weeks). Here, we can derive a managerial insight for shipping liner companies:
Allowing one week for the devanning process is a better choice for shipping companies, which

Chapter 2: Container Ship Type Decision 42

is the choice by OOCL (OOCL, 2017). Giving more weeks for the devanning process offers
more flexibility for customers such that shipping liner companies may charge higher freight
fee for compensating the opportunity cost. However, the loss may outweigh the benefit, as the
total cost increases in a convex manner at the beginning based on the sensitivity analysis.

2.7 CONCLUSION

This paper makes an explorative study on the ship type decision considering the empty
container repositioning and the foldable containers. Different from traditional research works
on the ship fleet deployment, our study incorporates both the laden container transportation
and the empty container repositioning into ship type decision in order to achieve the global
optimum for a shipping service route over a whole planning horizon. Meanwhile, as
researchers have shown the economic and logistical viability of foldable containers, the
problem also considers the use of foldable containers, which aims to find under what
conditions, the shipping liner needs to use the foldable containers in its liner shipping services.
In this study, we find that given the ship type with a certain capacity, the problem transfers
to a nonstandard minimum cost flow. Henceforth, we build a network flow model for the
problem by constructing a network. When considering standard containers and foldable
containers, trouble arises in the network construction that is some parallel arcs share the same
capacity restriction. To overcome this trouble, we design a revised network simplex algorithm
that changes the standard pivot operation. The algorithm is applicable to any minimum cost
flow problem with sharing capacity restrictions. Based on the algorithm, we develop a solution
approach by using reduced costs for excluding some ship type, which can find the optimal ship
type in the end. By using the solution approach, we conducted extensive numerical
experiments to find some managerial implications on the ship fleet deployment and the
foldable container usage.
Some useful managerial implications of this study are summarized from three perspectives.
(i) Ship type decision: when deciding the ship type deployed in a ship fleet, only involving
laden container transportation leads to sub-optimal solutions, as the possible empty container
repositioning affects the decision. After including foldable containers, a smaller ship type can
be expected to deploy, because foldable containers have the storage space advantage in the
empty container repositioning. (ii) Foldable container usage: under the current cost setting, it
is not cost-effective for shipping lines to use foldable containers, as the long-term leasing cost
is high. However, if the long-term leasing cost cuts down, using foldable containers is
encouraged, as foldable container usage is highly dependent on the long-term leasing cost. The
foldable container usage’s sensitivity to the folding and unfolding cost depends on the long-

Chapter 2: Container Ship Type Decision 43

term leasing cost. With different long-term leasing costs, different efforts can be made to
reduce the folding and unfolding cost. For example, if container leasing companies charge a
moderate price for long-term leasing, the shipping lines may devote much efforts to cut down
folding and unfolding cost, which can lead to a profitable result. (iii) Container devanning
time: It is better to allow one-week time for the container devanning process. Although
allowing more weeks for the devanning process brings more flexibility for customers such that
shipping lines may obtain some benefits, the opportunity cost can outweigh the benefits, as the
opportunity cost increases significantly after the one-week setting.

2.8 REFERENCES

Ahuja, R. K., Magnanti, T. L., Orlin, J. B., 1993. Network Flows: Theory, Algorithms, and
Applications, New Jersey, USA.
APL, 2017. The perfect fit for all your different needs: your APL equipment guide,
https://www.apl.com/wps/wcm/connect/659ca9b6-7eb3-4c16-82de-
8de04abab7e9/Equipment+guide.pdf?MOD=AJPERES. Accessed on 10 August 2017.

Bell, M.G.H., Liu, X., Rioult, J., Angeloudis, P., 2013. A cost-based maritime container
assignment model. Transportation Research Part B 58, 58–70.
Brouer, B.D., Alvarez, J.F., Plum, C.E., Pisinger, D., Sigurd, M.M., 2013. A base integer
programming model and benchmark suite for liner-shipping network design.
Transportation Science 48(2), 281–312.
Cheung, R.K., Chen, C.Y., 1998. A two-stage stochastic network model and solution methods
for the dynamic empty container allocation problem. Transportation Science 32(2), 142–
162.
Container Auction, 2017. Shipping container lease agreements.
https://containerauction.com/read-news/shipping-container-lease-agreements. Accessed
on 29 June 2016.
COSCO, 2017. Aust port charges: import container detention,
http://www.fivestarshipping.com.au/australian-port-charges/container-detention/.
Accessed on 12 August 2017.
CMA CGM, 2017. Lines and Services, https://www.cma-cgm.com/products-services/line-
services. Accessed on 12 August 2017.
Crainic, T.G., Gendreau, M., Dejax, P., 1993. Dynamic and stochastic models for the
allocation of empty containers. Operations Research 41, 102–126.
Gelareh, S., Meng, Q., 2010. A novel modeling approach for the fleet deployment problem
within a short-term planning horizon. Transportation Research Part E 46(1), 76–89.

Chapter 2: Container Ship Type Decision 44

Goh, S. H., & Lee, J. L., 2016. Commercial viability of foldable ocean containers. Proceedings
- International Conference on Industrial Engineering and Operations Management,
Kuala Lumpur, Malaysia, March 8-10, 2016.

Hanh, L. D. 2003. The logistics of empty cargo containers in the Southern California region.
Final Report, Long Beach, CA.

Holland Container Innovations, 2017a. 4FOLD customer cases. http://hcinnovations.nl/4fold-

customer-cases/. Accessed on 10 August 2017.

Holland Container Innovations, 2017b. 4FOLD foldable container.

http://hcinnovations.nl/4fold-foldable-container/. Accessed on 10 August 2017.

Konings, R., 2005. Foldable containers to reduce the costs of empty transport? A cost-benefit
analysis from a chain and multi-actor perspective. Maritime Economics and Logistics
7(3), 223–249.
Lee, C.-Y., Yu, M.Z., 2012. Inbound container storage price competition between the
container terminal and a remote container yard. Flexible Services and Manufacturing
24(3), 320–348.
Li, J.A., Liu, K., Leung, S.C.H., Lai, K.K., 2004. Empty container management in a port with
long-run average criterion. Mathematical and Computer Modelling 40, 85–100.
Li, J.A., Leung, S.C.H., Wu, Y., Liu, K., 2007. Allocation of empty containers between multi-
ports. European Journal of Operational Research 182, 400–412.
Meng, Q., Wang, T.S., Wang, S., 2012. Short-term liner ship fleet planning with container
transshipment and uncertain container shipment demand. European Journal of
Operational Research 223, 96–105.
Meng, Q., Wang, S., 2011. Liner shipping service network design with empty container
repositioning. Transportation Research Part E 47(5), 695–708.
Meng, Q., Wang, S., 2012. Liner ship fleet deployment with week-dependent container
shipment demand. European Journal of Operational Research 222, 241–252.
Moon, I. K., Do, Ngoc A. D., Konings, R., 2013. Foldable and standard containers in empty
container repositioning. Transportation Research Part E 49, 107-124.
Moon, I. K., Hong, H., 2016. Repositioning of empty containers using both standard and
foldable containers. Maritime Economics & Logistics 18, 61-77.
Myung, Y. S., Moon, I. K., 2014. A network flow model for the optimal allocation of both
foldable and standard containers. Operations Research Letters 42(6-7), 484-488.
Ng, M.W., 2014. Distribution-free vessel deployment for liner shipping. European Journal of
Operational Research 238(3), 858–862.

Chapter 2: Container Ship Type Decision 45

Ng, M.W., 2015. Container vessel fleet deployment for liner shipping with stochastic
dependencies in shipping demand. Transportation Research Part B 74, 79–87.
Ng, M.W., 2016. Revisiting a class of liner fleet deployment models. European Journal of
Operational Research 257(3), 773–776.
OOCL, 2017. Demurrage & detention free time and charges,
http://www.oocl.com/hongkong/eng/localinformation/ddfreetime/Pages/default.aspx?sit
e=hongkong&lang=eng. Accessed on 12 August 2017.
Perakis, A.N., Jaramillo, D.I., 1991. Fleet deployment optimization for liner shipping Part 1.
Background, problem formulation and solution approaches. Maritime Policy &
Management 18(3), 183–200.
Plum, C.E., Pisinger, D., Sigurd, M.M., 2013. A service flow model for the liner shipping
network design problem. European Journal of Operational Research 235(2), 378–386.
Shintani, K., Konings, R., Imai, A., 2010. The impact of foldable containers on container fleet
management costs in hinterland transport. Transportation Research Part E 46, 750–763.
Shipping and Freight Resource, 2017. Difference between demurrage and detention,
https://shippingandfreightresource.com/difference-between-demurrage-and-detention/.
Accessed on 12 August 2017.
Song, D.P., Dong, J.X., 2011. Flow balancing-based empty container repositioning in typical
shipping service routes. Maritime Economics and Logistics 13(1), 61–77.
Song, D.P., Dong, J.X., 2012. Cargo routing and empty container repositioning in multiple
shipping service routes. Transportation Research Part B 46(10), 1556–1575.
UNCTAD, 2016. Review of maritime transport.
http://unctad.org/en/PublicationsLibrary/rmt2016_en.pdf. Accessed on 10 August 2017.
Wolff, J., Herz, N., & Flamig, H., 2007. Report on case study: Empty container logistics:
Hamburg-Baltic Sea region. Hamburg University of Technology, The Baltic Sea region
programme, 2013.
Word Shipping Council, 2013. About the industry: major trade routes in global trade.
http://www.worldshipping.org/about-the-industry/global-trade/trade-routes. Accessed
on 28 March 2016.
Word Cargo News, 2017. Container industry news: Folding boxes stack up.
http://hcinnovations.nl/wp-content/uploads/2017/05/Press-
20170401_World_Cargo_News.pdf. Accessed on 10 August 2017.

Chapter 2: Container Ship Type Decision 46

Chapter 3: Container Reefer Slot Conversion

This chapter addresses the reefer slot conversion problem for container freight transportation.
Given a fleet of container ships of varying capacity, a cost-efficient approach for improving
fleet utilization and reducing the number of delayed containers is to optimize the sequence of
container ships in a given string, a problem which belongs to the ship-deployment class. A
string sequence with ‘uniformly’ distributed ship capacity is more likely to accommodate a
random container shipment demand. The number of one’s total ship slots acts as a gauge of
the capacity of the container ships. Meanwhile, in reality, there are actually two types of ship
slots: dry slots and reefer slots. A dry slot only accommodates a dry container, while a reefer
slot can accommodate either a dry or a reefer container. The numbers of dry and reefer slots
for ships in a string are different. Therefore, in this study we propose a model that considers
both dry and reefer slots and use it to elucidate the optimal ship-deployment sequence. The
objective is to minimize the delay of dry and reefer containers when the demand is uncertain.
Furthermore, based on the optimal sequence deduced, the study also investigates the need to
convert some dry slots to reefer slots for the container ships.

3.1 INTRODUCTION

In a liner shipping network, the liner shipping company normally operates weekly-serviced
ship routes with fixed schedules. That is, for a given shipping route, there is a fleet of container
ships deployed such that the ports along the shipping route are visited on a weekly basis. In
order to provide weekly services for all the ports, the number of ships deployed in the fleet
should be the number of weeks needed to complete the shipping route (Bell et al., 2011; Meng
and Wang, 2012; Lin and Tsai, 2014; Wang, 2017). For example, if traversing a route takes
two weeks, the number of ships deployed should be two. When a container ship visits a port,
the containers that have arrived at the port during the past week would then be loaded onto the
ship to send them onwards to their destination ports. However, the ship’s capacity, as measured
by the number of ‘twenty-foot equivalent units’ (TEUs), is limited. Thus, some containers
cannot be loaded onto the ship, and will therefore be delayed for one week.
In a container ship, ‘‘slots’’ are used to accommodate containers. The number of slots in
a ship reflects the ship’s capacity. Generally, there are two types of slots for a container ship,
dry slots, and reefer slots. Reefer slots are equipped with an electrical outlet so that reefer
containers can be accommodated which have an integrated cooling unit. Reefer containers
cannot be placed in dry slots (due to lack of electricity supply) while dry containers can be

Chapter 3: Container Reefer Slot Conversion 47

placed on the reefer slots if there are any still available. For a ‘‘cold chain’’, these reefer
containers and slots are critical in order to keep goods fresh (Cheaitou and Cariou, 2012;
Rodrigue and Notteboom, 2015).
Among the world’s ship fleets, reefer container slots have been expected to rise 22%
between 2013 and 2018 driven by the growth in demand for reefer cargo transportation
(Sowinski, 2015). Nowadays, reefer slot capacity has already become a critical measure of the
competitiveness between shipping liners. The largest-ever ships in the Hamburg Süd Group
have more than 2,000 reefer slots on board and are the vessels with the largest reefer slot
capacities in the world (Hamburg Süd, 2013). Hamburg Süd (2013) attributed the fact that they
were performing well under difficult business conditions to their large reefer slot capacity. In
comparison, Hanjin (which was the 7th largest shipping line in the world) only maintained
several hundred reefer slots in their ships (Chen and Yahalom, 2013). Currently, Hanjin has
already filed for bankruptcy due to a crisis in its financial affairs. Cool Logistics (2014) even
ascribed the Hanjin crisis to their reefer cargo transportation arrangements. Empirically, in
light of the underlying fierce competition between shipping liners, slot configurations, and slot
conversion are, in practice, topics of significant importance among shipping companies (Lin
et al., 2017).
Table 3.1: The capacities of the ships in the fleet used for the Yangtze Service route

No. Ship name Capacity (TEU)

1 ARCHIMIDIS 7,943
2 CMA CGM NABUCCO 8,488
3 CSCL EAST CHINA SEA 10,036
4 CSCL SOUTH CHINA SEA 10,036
5 NAVARINO 8,533
6 XIN DA YANG ZHOU 8,533

In maritime studies, a great deal of research effort has been devoted to container ship fleet
deployment problems. This effort is mainly focused on determining the ship type for a given
shipping service route (Gelareh and Meng, 2010; Meng et al., 2012; Song and Dong, 2013;
Ng, 2014). In such studies, it is assumed that the container ships are categorized into different
types. Moreover, all the ships belonging to a particular category are homogenous, i.e. they
have the same capacity. In addition, the ship fleet deployed for a given route only consists of
one type of ship. However, in reality, the ships in a fleet may not all have the same capacity
(Lin and Liu, 2011; Du et al., 2017). For instance, consider the example listed in Table 3.1
(relating to the China–USA Yangtze Service route operated by France’s CMA CGM Group,

Chapter 3: Container Reefer Slot Conversion 48

the third largest shipping liner in the world). The table lists the six ships deployed in the fleet
which consists of four different ship types, each with different capacity (CMA CGM, 2016).
Corresponding to the data in Table 3.1, there is a novel optimization problem which has
been addressed by Wang (2016). Put briefly, how should we arrange the sequence of the ships
for a given fleet in order to maximize the ship utilization of the fleet and reduce the number of
delayed containers? Here, the sequence in which the ships visit each port in the shipping route
is indicated using a ‘‘string’’ (Figure 3.1). If the sequence is 1→2→3→4→5→6, for example,
then each port will be visited by the ARCHIMIDIS first, CMA CGM NABUCCO second, and
so on. Optimization of the sequence depends on the weekly-dependent demand for container
shipment (Meng and Wang, 2012). That is, shipment demand does not remain constant and
will vary from week to week. Let us suppose that the demand for the Yangtze Service route is
random and varies between 8,000 and 10,000 TEU. Intuitively, the sequence of 3→5→1→4
→6→2 might be expected to outperform the sequence of 1→2→3→4→5→6, as the former
sequence has a more uniformly distributed ship capacity (which is more likely to be able to
handle an uncertain demand pattern).
This study aims to optimize the ship sequence considering the availability of dry slots and
reefer slots when the ship visits each port in a given route. Note that reefer containers have a
higher priority than dry containers when loaded onto visiting ships in each port. Here, to
account for the randomness in the container shipment demand, we assume that a predetermined
demand probability distribution can be found. Based on the optimized ship sequence found for
a string subject to uncertainty, this study further optimizes the associated problem of reefer
slot conversion. That is, we aim to determine whether or not a shipping line should convert
some of the dry slots in the ships to reefer slots (and how many dry slots should be converted)
considering that reefer slots are more flexible (i.e. can carry either dry or reefer containers).
To improve the utilization of container ships and the efficiency of container handling,
much effort has been devoted to studying various different aspects, e.g. optimization of
shipping networks and container terminal operation. Whether or not reefer slot conversion
provides a cost-efficient approach is still an open question. Meanwhile, the open question is
also meaningful as Arduino et al. (2015) have proposed many technical and operational
advantages of using reefer containers and slots. Although the study conducted by Wang (2016)
provided some useful rules to improve the sequencing of container ships, it does not consider
the availability of reefer slots in the container ships, nor does it judge whether the slot
configurations in the ships can be improved. As ships have both reefer and dry slots to include
to measure the ship’s capacity, the consideration of both types of slots in sequence
optimization is inevitable.

Chapter 3: Container Reefer Slot Conversion 49

 Based on the discussion above, this paper presents a practical study of optimizing reefer
slot conversion for container ships in a string. First, we derive the relevant equations required
to estimate the profit of a certain string/ship sequence. Then, we propose a simulation-based
approach to optimizing the sequence with the objective of maximizing the profit. We solve the
slot-conservation problem using a slot-conversion algorithm that embeds a simulation-based
approach for sequence optimization. Furthermore, to validate the effectiveness of the proposed
approach, we consider several case studies based on real shipping routes operated by CMA
CGM.

3.2 ESTIMATING THE PROFIT FOR A GIVEN STRING

Figure 3.1 depicts the sequence of container ships corresponding to a given string of the
weekly-serviced Yangtze Service shipping route. The sequence shown is 1→4→6→5→2→3
(which is equivalent to 4→6→5→2→3→1, and the other sequences obtained by cyclically
permuting the original string, as the string of ships forms a loop). Without losing generality,
we assume that the sequence is written so that the first ship in the sequence is the ship with the
smallest capacity (e.g. ship 1 for the Yangtze Service route). Given such a string, we derive in
this section some equations that can be used to calculate the weekly profit of the shipping route.
Note that all the equations derived here are considered to be on a weekly basis. In the following
subsections, the variables and parameters used in the equations will be elaborated upon in
detail. We also outline some practical container delay and rejection rules to be applied when
the container ships visit the ports.

Figure 3.1: A sequence of container ships in a string

Chapter 3: Container Reefer Slot Conversion 50

3.2.1 Decision variables and parameters

When a container ship visits a port, the containers accumulated in the past week need to be
loaded into the available slots in the container ship for shipment to their destination ports. If
the slots in the container ship are insufficient to stow all the accumulated containers, then some
containers m be delayed for one or more weeks. Such a delay incurs additional costs for the
shipping line. The main costs are: (i) the cost of storing the delayed container in the yard space
of the port, (ii) the cost to customer satisfaction (who, presumably, will not be happy about the
delay), and (iii) the cost incurred supplying electricity to a delayed container if it is of the
reefer variety.
In the following week, when the next container ship visits the port, the containers that
have been delayed will have priority when it comes to being loaded onto the ship. If some
delayed containers still cannot be transported due to the limited availability of slots in the ship,
then those delayed containers must be transported using slots from other shipping lines or by
transshipment (Hasheminia and Jiang, 2017; Jiang et al., 2017). Such containers subsequently
transported by other shipping lines are referred to as ‘‘rejected containers’’. Container
rejection is the last thing that the shipping liner wants as it leads to two consequences for the
shipping line: (i) it has to pay the freight rates for the shipment provided by the other shipping
line, and (ii) it loses the goodwill of the customers affected. Here, one-week delays reflect the
service level offered by the shipping liner with respect to guaranteeing the transit time for
transportation of the container. We can adjust it to two or more weeks delay depending on the
policy of the shipping liner. Meanwhile, container rejection does not mean the shipping line
will lose profit. The marginal profit associated with the container transportation will decrease,
but can still be positive after some additional incurred costs are deducted.
Based on the above discussion, the decision variables and parameters required are as
follows.

Decision variables:

Dvd1  Dvr1  : The number of dry (reefer) containers that were delayed from ship v – 1 in

the previous week.

Dvd  Dvr  : The number of dry (reefer) containers that are delayed from ship v in the

current week.

D
d
D 
r
: The average number of dry (reefer) containers that are delayed.

Rvd  Rvr  : The number of dry (reefer) containers that are rejected from ship v in the

current week.

Chapter 3: Container Reefer Slot Conversion 51

Input parameters:

c d ,1  c r ,1  : The loss of goodwill if a dry (reefer) container is rejected.

c r ,2 : The cost of electricity/fuel to transport a reefer container.

c d ,3  c r ,3  : The cost of storing a dry (reefer) container in the yard.

c r ,4 : The extra electricity cost when storing a delayed reefer container.

c d ,5  c r ,5  : The cost of customer dissatisfaction.

gd gr  : The freight rates for a dry (reefer) container.

pnd  pnr  : The mass probability that n dry (reefer) containers need to be transported in the

current week.

qd  qr  : The expected number of dry (reefer) containers that need to be transported in the

current week.

 d  r  : The new dry (reefer) container demand in the current week.

 
d
  r
: The realization of new dry (reefer) container demand in the current week.

Evd  Evr  : The dry (reefer) container capacity of ship v.

Evdr : The capacity of reefer slots that are available to dry containers in the current week.

Nd Nr  : The maximum number of dry (reefer) containers in the current week.

V : The number of ships in a string.

Every week, there is a demand for dry containers, which is a random variable denoted by
d
 . We assume that  is known to support integers between 0 and N d . The probability
d

d
mass function of the random variable  is assumed to be known based on historical data:
d d
Pr(  n)  pnd , n  0,1,..., N d . The expectation value (E) of  is q d : E( )   n 0 npnd .
d Nd

r
The symbols  , N , pn and q are correspondingly defined for reefer containers. Notice that
r r r

we consider the container demands on the ‘‘hit-haul’’ leg (i.e. the leg on which the highest
number of containers is carried) of the long-haul liner service routes rather than all the legs in
the shipping route. For instance, for Asia–Europe service routes, the leg after the last port in
Asia is normally the hit-haul leg, as the container demands from Asian ports to European ports
d r
are much higher than in the opposite direction. Thus, we only have  and  to denote the
demands on the hit-haul leg for dry and reefer containers, respectively.

3.2.2 Container delay and rejection rules

Suppose that the V ships are in the sequence given by the string 1  2 v  V . The dry
(reefer) container capacity of the ship v is Evd ( Evr ). In a particular week, when the ship v

Chapter 3: Container Reefer Slot Conversion 52

arrives at its destination port, the number of dry (reefer) containers that are at the port because
d r
they were delayed from the previous week is Dv 1 ( Dv 1 ). Moreover, the new dry (reefer)
d r
container demand in the current week is  (  ), which is the realization of  (  ). Based
d r

on the values of Dv 1 and Dv 1 (which were determined one week ago) and the values of 
d r d

and  (which are just observed), the shipping line needs to determine the number of dry
r

d r
(reefer) containers to postpone. Denote this quantity by Dv ( Dv ) and the number they need
d r
to reject by Rv ( Rv ). We analyze the decisions as follows:

(i) If Dv 1    Ev and Dv 1    Ev , then all of the containers will be transported and

d d d r r r

Dvd  Dvr  Rvd  Rvr  0 .

(ii) As a reefer slot can also be used to transport a dry container, if Dv 1    Ev ,

d d d

Dvr1   r  Evr , and Dvd1   d  Dvr1   r  Evd  Evr , then all of the containers will be

transported (some dry containers are stored in reefer slots) and Dv  Dv  Rv  Rv  0 .

d r d r

(iii) We assume that reefer containers have higher priority because they bring in more profit
than dry containers. If Dv 1    Ev , then not all reefer containers can be transported
r r r

immediately, and we allow some reefers to be postponed to the next week. Note that in
r
the following week, ship v  1 with reefer capacity of Ev 1 will arrive. If
Dvr1   r  Evr  Evr1 and if all of the Dvr1   r  Evr reefers are postponed, then some of
them still cannot be transported in the following week. In this study, we assume that a
container (dry or reefer) can only be postponed by one week. Hence, if Dv 1    Ev
r r r

and Dv 1    Ev  Ev 1 , then Dv 1    Ev reefers are postponed, i.e.

r r r r r r r

Dvr  Dvr1   r  Evr and Rvr  0 ; if Dvr1   r  Evr and Dvr1   r  Evr  Evr1 , then Evr1

reefers are postponed, i.e. Dv  Ev 1 and Rv  Dv 1    Ev  Ev 1 .

r r r r r r r

(iv) We now analyze the dry containers. We allow dry containers to occupy reefer slots that
are not used in the current week, but we do not allow dry containers to reserve reefer slots
in the following week. The available reefer slots in the current week are
Evdr : max 0, Evr  ( Dvr1   r )
. If Dv 1    Ev  Ev and Dv 1    Ev  Ev  Ev 1 ,
d d d dr d d d dr d

then Dv 1    Ev  Ev dry containers are delayed, i.e. Dv  Dv 1    Ev  Ev and

d d d dr d d d d dr

Rvd  0 ; if Dvd1   d  Evd  Evdr and Dvd1   d  Evd  Evdr  Evd1 , then Evd1 dry

containers are delayed, i.e. Dv  Ev 1 and Rv  Dv 1    Ev  Ev  Ev 1 .

d d d d d d dr d

3.2.3 Calculation of the weekly profit

The shipping line aims to maximize their profit per week. To this end, they try to transport as
many containers as possible. However, due to the randomness of the demand, it may occur
that in some weeks the demand is very high and not all the containers can be transported. In

Chapter 3: Container Reefer Slot Conversion 53

this subsection, we derive an equation for the weekly profit based on a detailed analysis of the
relevant cost parameters.
Let the freight rates for a dry container and reefer container be g d and g r , respectively,
d r
and the average number of rejected dry and reefer containers per week be R and R ,
respectively. Then, the company’s expected revenue per week is:


g d qd  R
d
  g q  R  .
r r r
(3.1)

The container shipment also incurs some costs. First, if a dry (reefer) container is rejected,
the cost associated with loss of goodwill will be denoted by c d ,1 ( c r ,1 ). Second, reefer
containers need electricity to maintain the desired temperature during transportation, and the
cost of the electricity/fuel when transporting a reefer container is c r ,2 . Third, if a dry (reefer)
container is postponed, it has to be stored in the container yard for one week, and so we
represent the corresponding storage cost of using the yard space by c d ,3 ( c r ,3 ). Fourth, the
additional electricity cost when storing a delayed reefer container is denoted by c r ,4 . Fifth, if
a dry (reefer) container is delayed, the cost of customer dissatisfaction is denoted by c d ,5 ( c r ,5 ).
d r
Denote the average number of delayed dry (reefer) containers per week by D ( D ).
Then, the expected cost per week is given by the expression
d r
c d ,1 R  c r ,1 R  c r ,2 q r  R  r
  c d ,3
 c d ,5  D   c r ,3  c r ,4  c r ,5  D . (3.2)
d r

Consequently, the expected profit per week is

  
 g d qd  Rd  g r qr  Rr  
  
      
 c d ,1 R d  c r ,1 R r  c r ,2 q r  R r  c d ,3  c d ,5 D d  c r ,3  c r ,4  c r ,5 D r 
 (3.3)
  g q   g  c
d d r r ,2
q r


 g d  c d ,1  R d   g r  c r ,1  c r ,2  R r   c d ,3  c d ,5  D d   c r ,3  c r ,4  c r ,5  D r  .
 

Note that the first term in the right-hand side of Eq. (3.3) is a constant and is independent of
the sequence of the ships in the string.

3.3 OPTIMIZING THE SEQUENCE OF SHIPS IN A STRING

In this section, we use the profit equation derived above to optimize the sequence of ships in
the string in order to maximize the weekly profit. Note that sequence optimization is a tactical
level decision, which is unchanged over the weeks. There are two major reasons why the
sequence may need to be re-optimized:

Chapter 3: Container Reefer Slot Conversion 54

(i) The capacities of the ships change. For example, some old ships in the fleet used for the
shipping route may be replaced by new ships of different capacity; or the rotation time of
the shipping service might be adjusted so that new ships can be added to the fleet.

(ii) The demand probability changes. In other words, the probability mass function changes
to a different one. For example, new competitors may enter the shipping market which
will affect the demand pattern for existing shipping liners.

Given the capacities of the ships in the fleet and demand probability mass function, we
design a simulation-based approach for the optimization process. Given V ships, any two of
which have different capacity, there will be S  (V  1)! different possible ship sequences (as
the sequence forms a loop, it does not matter which ship is chosen to be the first one). Let the
sequence be denoted by s  (1, 2, , S ) and the corresponding weekly profit be P ( s ) , as
d r d r
determined by Eq. (3.3). (With the obvious notation that D ( s ) , D ( s ) , R ( s ) , and R ( s ) are
d r d r
used in place of D , D , R , and R , respectively.) Mathematically, therefore, the weekly
profit of sequence s, P ( s ) , is given by:

P ( s )   g d q d   g r  c r ,2  q r  
(3.4)
 g d  c d ,1  R d ( s )   g r  c r ,1  c r ,2  R r ( s )   c d ,3  c d ,5  D d ( s )   c r ,3  c r ,4  c r ,5  D r ( s )  .
 
d r d r
We use Monte-Carlo simulations to calculate D ( s ) , D ( s ) , R ( s ) , and R ( s ) in
Eq. (3.4). The essence of the approach is as follows. (i) Define the number of weeks T we wish
to simulate (e.g. T  100, 000 weeks). (ii) Randomly generate the demand for each week using
the probability mass function. The demand for dry (reefer) containers in the week t are
denoted by td ( tr ). (iii) Simulate the decision-making process from weeks 1 to T. For each
week t, record the number of dry (reefer) containers postponed Dtd ( s) ( Dtr ( s) ) and rejected
d r d r
Rtd ( s) ( Rtr ( s ) ). (iv) The quantities D ( s ) , D ( s ) , R ( s ) , and R ( s ) can be subsequently
estimated using
1 T d 1 T 1 T 1 T
 Dt ( s ), D ( s )   Dtr ( s ), R ( s )   Rtd ( s ), R ( s )   Rtr ( s ) . (3.5)
d r d r
D (s) 
T t 1 T t 1 T t 1 T t 1

By substituting the results from Eq. (3.5) into Eq. (3.4), we can calculate P ( s ) , and
therefore choose the sequence with the maximum profit:
s *  arg max P ( s ) . (3.6)
s 1,2 S

3.3.1 A two-stage simulation approach

To find the optimal sequence, we apply the two-stage simulation method proposed by Nelson
et al. (2001). In the first stage, a given number of weeks of the process are simulated for each

Chapter 3: Container Reefer Slot Conversion 55

sequence. Some sequences, those whose average profits are much smaller than that with the
largest profit, are removed. The remaining sequences are evaluated in the next stage. Moreover,
the variance of the weekly profit for each remaining sequence can be estimated. In the second
stage, additional weeks of the process are simulated for each of the remaining sequences. In
particular, sequences that had smaller variances can be simulated for a smaller number of
weeks. The details of the algorithm are elaborated below.

Two-stage simulation algorithm

Step 0: (i) Select a value for  to represent a practically significant difference. For instance,

 could be set at 1000 US$/week, meaning that we are indifferent to two solutions if

their expected difference in weekly profit is less than 1000 US$. This also implies that

if we find a sequence that is not the optimal one, but has a profit less than that of the

optimal one by at most  , then we also consider it to be an optimal solution. (ii) Select

the overall confidence level 1   . For example, if  is chosen to be 10%, it means

that the chance that the found solution is the optimal one is at least 90% (as mentioned

before, a solution is considered to be optimal if its profit is within a gap of  relative

to the optimal one).

Step 1: Select a confidence level 1   0 for the first stage. For example, if  0 is 5%, it means

that the probability that the optimal sequence is not removed in the first stage is at

least 95%.

Step 2: Select a value for the grouping parameter U, e.g., 30. To appreciate the significance

of this parameter, consider the following discussion of the pertinent random variable.

The average profit of sequence s over U V weeks is given by PUV ( s ) . The central

limit theorem implies that PUV ( s ) is approximately normally distributed. Select a

parameter T1 that is related to the sample size of the first stage (e.g. T1 might be 50).

Simulate the process for each of the S sequences for TUV

1 weeks. We thus obtain T1

realizations of the random variable PUV ( s ) for each s , which we denote by

PUV ,1 ( s), PUV ,2 ( s), , PUV ,T1 ( s) . Now compute the sample mean
1 T1
PUV ( s ) (1)   PUV ,l ( s ) ,
T1 l 1
(3.7)

where the superscript ‘‘(1)’’ means the first stage and sample variance

Chapter 3: Container Reefer Slot Conversion 56

 1 T1
  
2
Var PUV ( s )   PUV ,l ( s )  PUV ( s ) (1)  . (3.8)
T1  1 l 1

Step 3: Define a value  by

 : t 1 1 ( S 1) (3.9)
 0  ,T1 1

which is the 1   0 
1 ( S 1)
quantile of the t distribution with T1  1 degrees of

freedom. Let
1
1 2
   
1 (3.10)
Ws , s ' :    Var PUV ( s )  Var PUV ( s ')  , s  1, 2, , S, s '  1, 2, , S, s  s'
 T1 T1 

Remove all of the s '  1, 2, , S if there exists an s  1, 2, , S with s  s ' such that

PUV ( s ')(1)  PUV ( s ) (1)  max 0,Ws , s '    . (3.11)

The remaining sequences comprise a set denoted by I . The probability that I

contains the optimal sequence is at least 1   0 . If I is a singleton, then stop and

return the sequence in I ; otherwise, go to the next step.

Step 4: The second-stage confidence level is 1  1 where 1     0 . For example, if

 1  5% it means that, if the set I contains the optimal sequence, then the chance

that the optimal sequence will be identified in the second stage is at least 97.5%. For

each s  I , compute the second-stage sample size:

  h  2 
T2 s  max T1 ,   Var PUV ( s )     (3.12)
     

where h : h(1  1 , T1 , S ) is Rinott’s constant and  rounds up to the next largest

integer.

Step 5: Simulate the process for each sequence s  I for more (T2 s  T1 )UV weeks. Then,

compute the overall sample mean of PUV ( s ) :

T2 s
1
PUV ( s ) ( 2) 
T2 s
P UV ,l ( s ), sI (3.13)
l 1

Select the sequence with the largest PUV ( s )(2) .

Our exploratory experiments show that the above algorithm is time-consuming and not
able to find the optimal sequence using a ‘‘reasonable’’ amount of computation time. (All

Chapter 3: Container Reefer Slot Conversion 57

experiments in this study were conducted using MATLAB R2016b installed on a standard PC
built around an Intel Core i5 processor running at 2.83GHz and with 8GB of RAM.) The
problem can be attributed to the fact that T2s (in Step 4) is extremely large, ranging from
1  108 to 3  108 using the baseline settings given above. Thus, in the second stage, simulating
the process for each remaining sequence for other (T2 s  T1 )UV weeks takes a considerable
amount of computation time. Essentially, based on Eqs. (3.8) and (3.12), the order of
magnitude of T2s is determined by Var  PUV ( s )  , whose order of magnitude is, in turn,
determined by PUV ,l ( s ) (i.e. the weekly profit). Using some real-world input parameters (vide
infra), the weekly profits PUV ,l ( s ) are of the order of a few million dollars, and this is the
primary cause of the problem.
To reduce computation time, we have to ‘fold’ the weekly profit, PUV ,l ( s ) , by scaling it
to reduce the order of magnitude of T2s . This is especially important when determining the
number of dry slots to convert into reefer ones. To do this, we change the unit used for the
weekly profit by dividing PUV ,l ( s ) by a scaling parameter  so that:
1
PUV ,l ( s ) PUV ,l ( s ) . (3.14)


For instance, if PUV ,l ( s ) is 5,343,000 US$ and   1000 , the weekly profit becomes 5,343
where the units used are 1,000 US$. Such a step allows us to deal with the above problem and
accelerate the algorithm. Table 3.2 shows the computation times required for different values
of .
Table 3.2: Comparison of computation times for different values of the parameter 𝜷

a Computation time (s) Selected sequence index

1 – –
5 20,346 46th
10 1,425 46th
25 306 29th
50 98 10th
100 21 113rd
1000 6 113rd
a The case  =1 corresponds to the original two-stage simulation algorithm.

As can be seen from Table 3.2, increasing the value of  accelerates the algorithm.
However, we cannot increase  as much as we like in order to simply accelerate the algorithm,
as there are repercussions. When   1000 , we can only claim that the selected sequence is
the optimal one to an accuracy of one unit, i.e. 1,000 US$. However, this will not necessarily
be the optimal sequence to the desired accuracy of 1 US$. As shown in the table, different 
values lead to different sequences being selected. Therefore, we need to make a comprise
between speed and accuracy. In this study, we select   10 . This should be sufficiently close

Chapter 3: Container Reefer Slot Conversion 58

to the desired accuracy (so we can be confident we have selected the correct sequence) and
allow us to obtain the solution in a reasonable amount of time. In the following sections, we
refer to the algorithm wherein   10 is used for scale reduction (i.e. via Eq. (3.14)) as the
‘‘revised two-stage simulation algorithm’’.

3.3.2 Case studies

In this section, we use the revised two-stage simulation algorithm to optimize the sequence of
a string of some particular cases of interest. First of all, according to EPRI (2010) and Ting
and Tzing (2004), we have to fine-tune the input parameters required. To this end, the cost
coefficients which we shall use were collated and are presented in Table 3.3. Here, we take
c d ,3  c r ,3  , the storage cost incurred to keep a dry (reefer) container in the yard for one week,
as an example of how we fine-tuned the costs. According to EPRI (2010), the storage space in
a yard is charged at a rate of US$0.21 per square foot per day. Thus, a twenty-foot standard
container corresponds to a weekly storage cost of US$30.
Table 3.3: The relevant input costs used in the case studies

Parameter Value (US$)

𝑑,1
𝑐 150
𝑐 𝑟,1 250
𝑐 𝑟,2 230
𝑐 𝑑,3 30
𝑐 𝑟,3 30
𝑐 𝑟,4 50
𝑐 𝑑,5 100
𝑐 𝑟,5 200
𝑔𝑑 640
𝑔𝑟 960

We focus in this work on three shipping routes operated by CMA CGM: ‘‘Northwest
Express’’, ‘‘South Atlantic Express’’, and ‘‘Europe Pakistan India Consortium 1’’. In the
interest of brevity, we denote the three shipping routes as ‘‘1N’’, ‘‘2S’’, and ‘‘3E’’,
respectively. Figure 3.2 indicates the port rotations of these three shipping routes.

1N

Chapter 3: Container Reefer Slot Conversion 59

 2S

3E

Figure 3.2: The shipping routes involved in three case studies (CMA CGM, 2017)

Some key information about the shipping routes and ships deployed are given in
Table 3.4 — further details can be found by referring to CMA CGM (2017). During string
optimization, the number of ships deployed is another key element that can affect the scale of
the problem. This is because the number of ships determines the number of possible sequences
to consider. Therefore, we chose three shipping routes that deploy different numbers of ships
(determined by the rotation time, considering the weekly service frequency). We assume that
the dry- and reefer-container demands follow a uniform distribution with their ranges as given
in Table 3.4. The last column in this table shows the variance of the total container demand
for both dry and reefer containers.
Table 3.4: Information on the shipping routes and ships deployed

Rotation Container demand (TEUs) Total

Ship
Route Name time demand
No. Dry Reefer
(days) variance
Northwest
1N 6 42 [7800, 9220] [0, 800] 4.11×105
Express
South Atlantic
2S 7 49 [4500, 6100] [50, 900] 5.00×105
Express
Europe Pakistan
3E India Consortium 8 56 [7250, 9100] [100, 1300] 7.75×105
1

Chapter 3: Container Reefer Slot Conversion 60

 To facilitate discussion of the experiment results, we list in Table 3.5 some important
information about the ships deployed in each of the three shipping routes. As can be seen, in
each shipping route, the deployed ships have various capacities which help address the
importance of string optimization.
Table 3.5: Ships deployed in the three shipping routes

Route Ship no. Ship name Capacity (TEU) Dry slots Reefer slots
COSCO
1 9,469 8,769 700
GUANGZHOU
2 COSCO INDONESIA 8,501 8,501 0
1N 3 COSCO JAPAN 8,501 7,801 700
4 COSCO PACFIC 10,020 9,220 800
5 COSCO PHLIPPINES 8,501 7,801 700
6 COSCO PUSAN 9,572 8,872 700
1 E.R. LONDON 6,008 5,208 800
2 MSC KATYAYNI 5,711 5,177 534
3 MSC KRYSTAL 5,782 5,222 560
2S 4 MSC MARGARITA 5,770 5,138 632
5 MSC ORIANE 5,782 5,082 700
NORTHERN
6 6,732 6,232 500
MAJESTIC
7 RIO BARROW 5,551 5,001 550
1 APL CHARLESTON 9,336 8,322 1,014
CMA CGM
2 9,130 8,530 600
AMAZON
CMA CGM
3 9,130 7,630 1,500
URUGUAY
3E 4 MSC ALBANY 8,762 7,762 1,000
5 MSC ALGHERO 8,827 8,827 0
6 MSC SILVANA 8,400 7,700 700
7 MSC TOMOKO 8,400 7,700 700
8 UASC AL KHOR 9,400 8,900 500

Based on the information presented on the shipping routes, we implemented the revised
two-stage simulation algorithm to find the optimal sequence to employ. The results are shown
in Table 3.6. The table shows that the string optimization process changes all of the original
sequences used in the shipping routes. The last column shows the difference or gap in profit
between the original and optimized sequences (which are all positive). As tactical-level
decisions, these increases in profit can be claimed to be significant considering that the string
optimization procedure is inexpensive, i.e. the approach is cost effective. The results prove the
effectiveness of using the proposed algorithm for string optimization. One remarkable feature
of the table is that the string optimization procedure produced a much larger profit increase for

Chapter 3: Container Reefer Slot Conversion 61

shipping route 3E than for 1N and 2S. This can be attributed to the fact that 3E involves more
ships and a greater demand variance than the other two shipping routes. Therefore, it is
apparent that string optimization becomes more significant when the size of the deployed fleet
is large and there is a greater variance in demand.
Table 3.6: Results of the string optimization process

Original sequence Optimized sequence Gap

Route Profit Profit
Sequence a Sequence a
(106 US$) (106 US$)
1N 1→2→3→4→5→6 5.601 1→5→2→4→3→6 5.643 0.74%
2S 1→2→3→4→5→6→7 3.576 1→5→2→7→6→4→3 3.634 1.63%
3E 1→2→3→4→5→6→7→8 5.343 1→6→5→7→3→8→4→2 5.530 3.50%
a
A sequence can always be written with the first ship in the first place as the ships form a loop (so
sequences such as 1→2→3→4→5→6 and 2→3→4→5→6→1 are equivalent).

3.4 OPTIMIZING REEFER SLOT CONVERSION

The next natural question to ask is whether a shipping line should convert some of the dry slots
of a ship into reefer slots. Before addressing this question, we first consider the feasibility of
carrying out such conversions.
Reefer containers have an integral refrigeration unit with a water-cooling system to keep
the cargo cold. This refrigeration system needs an external power supply when the container
is stored in a ship. Therefore, a reefer slot has to be equipped with an electrical outlet which is
connected to the power system of the ship. Thus, an electrical outlet or plug needs to be
installed to convert a dry slot to a reefer slot. Technically, the conversion process is not much
trouble. Container ships usually have independent power sub-distribution panels that supply
power connections for refrigerated containers. When there is a need for more electrical outlets
to create a reefer slot, therefore, a ship can simply group several electrical outlets and supply
them with electricity using one power cable connected to a power sub-distribution panel (DNV
GL SE, 2015).
The slot conversion problem is addressed in the following way. Suppose that the cost of
converting one dry to one reefer slot is ĉ (US$/week). We need to determine how many dry
slots we should convert. Note that ĉ is an average cost per week. The cost of slot conversion
is mainly that associated with installing the electrical outlet or electrical plug which is used to
power the reefer container. According to EPRI (2010), the installation cost is approximately
1,250 US$ per electrical outlet. As we measure the profit on a weekly basis, we transfer the
installation cost to a depreciation cost of 48 US$ per week (i.e. cˆ  48 ). This assumes that
such a tactical level decision (i.e. decision to implement string optimization and slot

Chapter 3: Container Reefer Slot Conversion 62

conversion) lasts for half a year (26 weeks). Then, we solve the problem using the following
algorithm.

Slot conversion algorithm

Input: V ships whose identities (IDs) are labeled 1, 2, ,V . The initial number of reefer slots

(i.e. that prior to conversion) of the ship with ID v is Evr ( v  1, 2,...,V ).

Step 0: Find the optimal sequence using the revised two-stage simulation algorithm. The

optimal sequence is denoted by  (1)   (2)   (V ) , where  (u ) is the ID of the

u th ship in the sequence ( u  1, 2,...,V ). Thus, we have found PUV ( ) , that is, the

optimal weekly profit derived using the revised two-stage simulation algorithm.

Step 1: If the reefer slots of all of the V ships have reached the limit R, then stop the process.

Step 2: Set u '  1 .

Step 3: If the reefer slots on the ship with ID  (u ') have reached the limit R, then set

u '  u ' 1 and go to Step 5.

Step 4: Temporarily convert one dry slot on the ship with ID  (u ') into a reefer slot. The

resulting new sequence is denoted by  ' , which is the same as sequence  except

that ship  '(u ') has one more reefer slot (and, of course, one fewer dry slot) than the

ship  (u ') . Calculate PUV ( ') :

(i) If PUV ( ')  PUV ( )  cˆ , permanently convert the dry slot into a reefer one and

set    ' (i.e. permanently convert a dry slot into a reefer slot on the ship

 (u ') ). Go to Step 1.

(ii) Otherwise, set u '  u ' 1 and go to Step 5.

Step 5: If u '  V , go to Step 4. Otherwise, this means that just converting dry slots without

changing the sequence is not economically viable. Hence, we need to check what

happens if we change the sequence. To this end, go to Step 6.

Step 6: If the reefer slots of all of the V ships have reached the limit R , stop.

Step 7: Set v '  1 .

Step 8: If the reefer slots of all of the V ships have reached the limit R , set v '  v ' 1 and

go to Step 10.

Step 9: Temporarily, convert one dry slot on the ship with ID v ' into a reefer slot. Use the

revised two-stage simulation algorithm to find the optimal sequence  '' :

Chapter 3: Container Reefer Slot Conversion 63

 (i) If PUV ( '')  PUV ( )  cˆ , permanently convert the dry slot into a reefer one, set

   '' (i.e. permanently convert a dry slot into a reefer one on the ship whose

ID is v ' and adjust the sequence of the ships). Go to Step 1.

(ii) Otherwise, set v '  v ' 1 and go to Step 10.

Step 10: If v '  V , go to Step 9. Otherwise, we can no longer improve the solution, so stop.

Considering the limited availability of electricity on a container ship, we assume in the

above algorithm that the maximum number of reefer slots on a ship cannot exceed the limit
R  1500 . The other input parameters are the same as in the previous section. In practice, a

better limit to slot conversion can be derived by using information about the electrical loads
in the container ships and electricity usage of the reefer containers. For instance, EPRI (2010)
estimates that an electric reefer container needs 2.8875 kW per hour on average. The container
ship Hanjin Paris has a generator installed with a capacity of 7,600 kW and a load factor (%
of capacity) of 63%. This indicates that the ship has 2,812 kW of electrical power available
(Khersonsky et al., 2007). Some detailed information on electricity demand in container ships
can be found in the work of Zis et al. (2014). In the Hanjin Paris case, the container ship can
be equipped with an additional 973 reefer slots (at most) considering the generator’s
limitations (i.e. 2812/2.8875 slots can be supplied). Normally, the larger the capacity of the
ship, the greater the engine power available from the ship to support reefer slots (Zis et al.,
2013).

3.4.1 Slot conversion case study

String optimization is the first step carried out in our research problem. We now want to use
the slot conversion algorithm (which embeds the revised two-stage optimization algorithm to
optimize strings) to conduct a further investigation of the three case studies given in Section
3.3.2 (i.e. using the shipping routes and information given in Tables 3.4 and 3.5).
Using the same input parameters, we ran the slot conversion algorithm to obtain the
number of reefer slots to convert. The results are shown in Table 3.7. As can be seen, all the
ships involved have some dry slots that are converted to reefer slots, apart from one (namely,
the CMA CGM URUGUAY in shipping route 3E which already has the maximum number of
reefer slots permitted in the conversion algorithm). This verifies that the benefits gained by
the greater flexibility of reefer slots outweigh the conversion costs incurred to change the
existing slot configurations of the ships.
The last column in Table 3.7 shows the ‘‘optimal increment’’ that the optimal change in
a number of reefer slots represents. This quantity corresponds to the ratio of the optimal

Chapter 3: Container Reefer Slot Conversion 64

number of converted slots to the ship’s capacity (in TEU) expressed as a percentage. It is
important to note that the mean optimal increments for the shipping routes, as a whole, increase
from 1N to 2S to 3E. This reflects the higher significance of slot conversion for those shipping
routes with larger shipping fleets and higher demand variances (as is the case for the shipping
route 3E).
Table 3.7: Reefer slot conversion results for the three shipping routes

Number of reefer slots

Optimal
Route Ship Name After
Original Converted increment a
conversion
COSCO
1 700 834 134 1.42%
GUANGZHOU
2 COSCO INDONESIA 0 443 443 5.21%
3 COSCO JAPAN 700 861 161 1.89%
1N
4 COSCO PACFIC 800 940 140 1.40%
COSCO
5 700 874 174 2.05%
PHLIPPINES
6 COSCO PUSAN 700 945 245 2.56%
Mean: 2.42%
1 E.R. LONDON 800 1,024 224 3.73%
2 MSC KATYAYNI 534 761 227 3.97%
3 MSC KRYSTAL 560 800 240 4.15%
2S 4 MSC MARGARITA 632 752 120 2.08%
5 MSC ORIANE 700 874 174 3.01%
NORTHERN
6 500 906 406 6.03%
MAJESTIC
7 RIO BARROW 550 650 100 1.80%
Mean: 3.54%
1 APL CHARLESTON 1,014 1,500 486 5.21%
CMA CGM
2 600 954 354 3.88%
AMAZON
CMA CGM
3 1,500 1,500 0 —b
URUGUAY
3E 4 MSC ALBANY 1,000 1,263 263 3.00%
5 MSC ALGHERO 0 640 640 7.25%
6 MSC SILVANA 700 876 176 2.10%
7 MSC TOMOKO 700 1,075 375 4.46%
8 UASC AL KHOR 500 881 381 4.05%
Mean: 4.28%
a
The ratio of the number of converted slots to capacity (in TEU), as given in Table 5.
b
We cannot obtain a value here as the ship has already reached the slot conversion limit (1,500).
Overall, we have to emphasize the importance of having greater numbers of reefer slots.
This is not just based on the Hamburg Süd and Hanjin cases mentioned in the Introduction in
the context of competitiveness, but also because of the experimental results shown in Table 3.7.

Chapter 3: Container Reefer Slot Conversion 65

 The slot conversion algorithm is an iterative algorithm. When some dry slots have been
converted to reefer slots, the previous optimal sequence may no longer be optimal as the
number of reefer slots has changed. Thus, the slot conversion algorithm will not terminate
until both the slot conversion and sequence rearrangement processes can no longer be further
optimized. For the current examples, Table 3.8 illustrates the resulting sequences produced as
a result of re-optimization during slot conversion. From the table, we can see that the optimal
sequence after slot conversion is different from that obtained by just implementing string
optimization. The ‘gap’ column in Table 3.8 corresponds to the loss of profit suffered if we
insist on maintaining the previous optimal sequence. The numbers suggest that string
optimization and slot conversion should always be carried out at the same time as an integrated
optimization approach, exactly like our slot conversion algorithm does. Combining the gaps
or differences given in Tables 3.8 and 3.6, we can say that the optimal sequence with slot
conversion significantly outperforms the original sequence with the current slot configuration
(as the weekly profit increases substantially).
Table 3.8: Sequence re-optimization and slot conversion

Before slot conversion a After slot conversion

Rout
Profit Profit Gap
e Sequence Sequence
(106 US$) (106 US$)
1N 1→5→2→4→3→6 5.643 1→2→4→3→6→5 5.689 0.82%
2E 1→5→2→7→6→4→3 3.634 1→4→3→5→7→6→2 3.671 1.01%
1→6→5→7→3→8→4 1→5→3→4→6→7→8
3E 5.530 5.620 1.64%
→2 →2
a
Optimized results obtained using the revised two-stage simulation algorithm with the original (non-
optimized) slot allocations.

3.4.2 Shipping routes with large fleets

In the above cases, we compare weekly profits after slot conversion and after string
optimization. The profit improvement is not so clear cut, as we optimized the ship fleet using
string optimization in the first place. In this subsection, we conduct two further case studies
based on two shipping routes with very large shipping fleets and extremely high demand
variances. Our intention is to provide further motivation for integrating string optimization
with slot conversion.
The two routes selected are the French Asia Line 1 and Columbus JAX (both belonging
to CMA CGM) which involve 12 and 17 ships, respectively. CMA CGM (2017) gives detailed
information on the two shipping routes but some of the basic information of interest is shown
in Table 3.9. Compared with the previously considered routes (Table 3.4), these routes have
many more ships involved and their variance in demand is much larger.

Chapter 3: Container Reefer Slot Conversion 66

 Table 3.9: Basic information on the two shipping routes

Rotation time Container demand (TEU) Total demand

Route Ships
(days) Dry Reefer variance
French Asia Line 1 12 84 [12000, 18000] [650, 1900] 4.38×106
Columbus JAX 17 119 [5000, 10000] [700, 1600] 2.90×106

We first calculated the weekly profits of the original sequences used in the shipping
routes (Table 3.10). Then, we used the slot conversion algorithm (with string optimization) to
optimize the sequence of ships and convert some dry slots to reefer slots. Note that the slot
conversion limit was increased to 2,500 in these two case studies as the deployed ships have
larger capacities. Table 3.10 shows the effectiveness of the proposed approach. Compared to
the original sequences, the weekly profits of the optimized fleets were improved by 9.06%
(French Asia Line 1) and 7.90% (Columbus JAX). These profit increases are very significant
considering the cost-efficiency of the methods used. Note also that the two routes have larger
dry-to-reefer conversion rates (8.63% and 7.12%) compared to those found in the three
previous cases. This result verifies the previous finding that slot conversion is more critical for
shipping routes with large fleets of ships and high demand variances.
Table 3.10: Results obtained using string optimization and slot conversion

Weekly profit (106 US$)

Average reefer-slot
Route Optimized Gap
Original sequence conversion ratio
sequence
French Asia Line 1 10.27 11.20 9.06% 8.63%
Columbus JAX 5.444 5.874 7.90% 7.12%

3.5 CONCLUSION

This paper presents an improved algorithm to search for the optimal number of reefer slots to
have on a container ship. It is assumed that all the relevant parameters (e.g. freight rates,
storage costs, etc.) are already known. We first used a revised two-stage simulation approach
to optimize the sequence of ships deployed. Based on this, we then formulated a slot
conversion algorithm to determine the optimal slot configurations of the ships, which embeds
the two-stage simulation algorithm for string optimization.
In this study, we used several real shipping routes operated by CMA CGM to highlight
the effectiveness of our approach. Our results reveal that the algorithm is highly efficient and
can help shipping liners to significantly improve their profits. However, there are also several
issues that are worth studying further in future work:
(i) When converting the dry slots to reefer slots, draft and load capacities are not taken into
consideration. This would be of great use when making ship stowage plans.

Chapter 3: Container Reefer Slot Conversion 67

(ii) The use of power packs as a method of supplying electricity could be incorporated into the
analysis. (A self-contained power pack in a standard twenty- or forty-foot container could
act as a power source for multiple reefer boxes.) They are currently used to serve as a
standby or prime power source for intermodal applications including rail, port, ship, and
barge.

3.6 REFERENCES
Arduino, G., Carrillo Murillo, D., & Parola, F., 2015. Refrigerated container versus bulk:
evidence from the banana cold chain. Maritime Policy & Management 42(3), 228-245.
Bell, M. G., Liu, X., Angeloudis, P., Fonzone, A., & Hosseinloo, S. H., 2011. A frequency-
based maritime container assignment model. Transportation Research Part B:
Methodological 45(8), 1152-1161.
Cheaitou, A., & Cariou, P., 2012. Liner shipping service optimisation with reefer containers
capacity: an application to northern Europe–South America trade. Maritime Policy &
Management 39(6), 589-602.
Chen, J., & Yahalom, S., 2013. Container slot co-allocation planning with joint fleet agreement
in a round voyage for liner shipping. The Journal of Navigation 66(4), 589-603.
Cool Logistics, 2014. The Hanjin crisis: what about the reefer cargoes?
http://coollogisticsresources.com/global/the-hanjin-crisis-what-about-the-reefer-cargoes.
Accessed 18 January 2017.
CMA CGM, 2016. Line details: Yangtze Service. https://www.cma-cgm.com/products-
services/line-services/flyer/YANGTZE. Accessed 30 June 2016.
CMA CGM, 2017. Products & Services: Search Line Services. http://www.cma-
cgm.com/products-services/line-services. Accessed 20 January 2017.
DNV GL SE, 2015. Rules for Classification and Construction: Chapter 3 Electrical
Installations. Hamburg, Germany. http://rules.dnvgl.com/docs/pdf/gl/maritimerules/gl_i-
1-3_e.pdf. Accessed 29 April 2017.
Du, Y., Meng, Q., & Wang, S., 2017. Mathematically calculating the transit time of cargo
through a liner shipping network with various trans-shipment policies. Maritime Policy
& Management 44(2), 248-270.
EPRI., 2010. Electric Refrigerated Container Racks: Technical Analysis.
http://www.epri.com/abstracts/Pages/ProductAbstract.aspx?ProductId=0000000000010
19926. Accessed 18 January 2017.

Chapter 3: Container Reefer Slot Conversion 68

Gelareh, S., & Meng, Q., 2010. A novel modeling approach for the fleet deployment problem
within a short-term planning horizon. Transportation Research Part E: Logistics and
Transportation Review 46(1), 76-89.
Hasheminia, H., & Jiang, C., 2017. Strategic trade-off between vessel delay and schedule
recovery: an empirical analysis of container liner shipping. Maritime Policy &
Management 44(4), 458-473.
Hamburg Süd., 2013. Hamburg Süd 2013: Doing well under difficult business conditions.
http://www.hamburgsud-
line.com/hsdg/en/hsdg/regionalinformation/regional_news/regionalenewsdetails_40952
8.html. Accessed 18 January 2017.
Jiang, C., Wan, Y., & Zhang, A., 2017. Internalization of port congestion: strategic effect
behind shipping line delays and implications for terminal charges and investment.
Maritime Policy & Management 44(1), 112-130.
Khersonsky, Y., Islam, M., & Peterson, K., 2007. Challenges of connecting shipboard marine
systems to medium voltage shoreside electrical power. IEEE Transactions on Industry
Applications 43(3), 838-844.
Lin, D. Y., Huang, C. C., & Ng, M., 2017. The coopetition game in international liner shipping.
Maritime Policy & Management 44(4), 474-495.
Lin, D. Y., & Liu, H. Y., 2011. Combined ship allocation, routing and freight assignment in
tramp shipping. Transportation Research Part E: Logistics and Transportation Review
47(4), 414-431.
Lin, D. Y., & Tsai, Y. Y., 2014. The ship routing and freight assignment problem for daily
frequency operation of maritime liner shipping. Transportation Research Part E:
Logistics and Transportation Review 67, 52-70.
Wang, S., & Meng, Q., 2013. Reversing port rotation directions in a container liner shipping
network. Transportation Research Part B: Methodological 50, 61-73.
Meng, Q., & Wang, S., 2012. Liner ship fleet deployment with week-dependent container
shipment demand. European Journal of Operational Research 222(2), 241-252.
Meng, Q., Wang, T., & Wang, S., 2012. Short-term liner ship fleet planning with container
transshipment and uncertain container shipment demand. European Journal of
Operational Research 223(1), 96-105.
Nelson, B. L., Swann, J., Goldsman, D., & Song, W., 2001. Simple procedures for selecting
the best simulated system when the number of alternatives is large. Operations Research
49(6), 950-963.

Chapter 3: Container Reefer Slot Conversion 69

Ng, M., 2014. Distribution-free vessel deployment for liner shipping. European Journal of
Operational Research 238(3), 858-862.
Ng, M., 2015. Container vessel fleet deployment for liner shipping with stochastic
dependencies in shipping demand. Transportation Research Part B: Methodological 74,
79-87.
Rodrigue, J. P., & Notteboom, T., 2015. Looking inside the box: Evidence from the
containerization of commodities and the cold chain. Maritime Policy & Management
42(3), 207-227.
Song, D. P., & Dong, J. X., 2013. Long-haul liner service route design with ship deployment
and empty container repositioning. Transportation Research Part B: Methodological 55,
188-211.
Sowinski, L. L., 2015. Container, specialized carriers boost reefer investments.
http://www.joc.com/international-logistics/cool-cargoes/container-specialized-carriers-
boost-reefer-investments_20150301.html, Accessed 18 January 2017.
Ting, S. C., & Tzeng, G. H., 2004. An optimal containership slot allocation for liner shipping
revenue management. Maritime Policy & Management 31(3), 199-211.
Wang, S., 2015. Optimal sequence of container ships in a string. European Journal of
Operational Research 246(3), 850-857.
Wang, S., 2017. Formulating cargo inventory costs for liner shipping network design.
Maritime Policy & Management 44(1), 62-80.
Zis, T., North, R. J., Angeloudis, P., & Bell, M. G., 2013. A Systematic Evaluation of
Alternative Options for the Reduction of Vessel Emissions in Ports. In Proceedings of
Transportation Research Board 92nd Annual Meeting (No. 13-5346).
Zis, T., North, R. J., Angeloudis, P., Ochieng, W. Y., & Bell, M. G. H., 2014. Evaluation of
cold ironing and speed reduction policies to reduce ship emissions near and at ports.
Maritime Economics & Logistics 16(4), 371-398.

Chapter 3: Container Reefer Slot Conversion 70

Chapter 4: Cruise Itinerary Schedule Design

This chapter addresses the cruise itinerary schedule design problem for a cruise ship.
This problem determines the optimal sequence of a given set of ports of call (a port of call is
an intermediate stop in a cruise itinerary) and the arrival and departure times at each port of
call for maximizing the monetary value of the utility at ports of call minus the fuel cost. To
solve the problem, in view of the practical observations that most cruise itineraries do not have
many ports of call, we first enumerate all sequences of ports of call and then optimize the
arrival and departure times at each port of call by developing a dynamic programming
approach. To improve the computational efficiency, we propose effective bounds on the
monetary value of each sequence of ports of call, eliminating non-optimal sequences without
invoking the dynamic programming algorithm. Extensive computational experiments are
conducted and the results show that, first, using the bounds on the profit of each sequence of
ports of call considerably improves the computational efficiency; second, the total profit of the
cruise itinerary is sensitive to the fuel price and hence an accurate estimation of the fuel price
is highly desirable; third, the optimal sequence of ports of call is not necessarily the sequence
with the shortest voyage distance, especially when the ports do not have a naturally
geographical sequence.

4.1 INTRODUCTION

A cruise itinerary is a cruise route operated by a cruise company: A cruise ship picks up
cruise passengers at an embarkation port, calls at a set of ports of call for cruise passengers to
visit the port cities, and returns to a disembarkation port where cruise passengers get off the
cruise ship. The ship that is deployed on the itinerary, the embarkation port, the sequence of
ports of call, the disembarkation port, and the time schedule are all pre-determined in a cruise
itinerary. Cruise ships are different from other ships such as tankers, bulk carriers and
containerships in that transportation is not the purpose of cruise ships.
The cruising industry has maintained a steady increase in supply for the past 20 years. In
2014, the number of cruise passengers reached a total of 22.04 million and the global cruise
industry generated revenues of 37.1 billion U.S. dollars (Statista, 2015). Meantime, the world
cruise fleet had 296 ships (Cruise Industry News, 2015) with a total of 482,000 lower berths 1
(Statista, 2015). The Caribbean and the Mediterranean areas are the most important cruising

1It is often considered that one cabin has two beds (two lower berths) when calculating the capacity of cruise ships.
Any extra beds in a cabin are referred to as “upper berths”. The actual average number of beds per cabin in a
cruise ship is usually higher than two.

Chapter 4: Cruise Itinerary Schedule Design 71

destinations and hence they are also where most ship capacity is deployed. Cruise passengers
are mainly from developed countries. Among the 22.04 million cruise passengers in 2014,
12.16 million (55%) were from North America, 6.39 million (29%) were from Europe and
3.49 million (16%) were from the rest of the world. However, Cruising is an oligopolistic
industry: Carnival, Royal Caribbean, and Norwegian Cruise Lines are the three largest
companies with market shares of 41.8%, 21.8%, and 8.2%, respectively (Statista, 2015).
A few strategic decisions have a long-lasting effect on the profitability of a cruise company
(Veronneau and Roy, 2009). The first one is the cruise fleet planning. A large cruise ship has
over 5000 lower berths and may cost over one billion US dollars to construct. Companies book
new ships in order to replace the scrapped, damaged, or lost ships, fulfill the rising trend of the
cruising market, and provide extra capacity to block potential entrants to the market. The
second one is ship deployment. Some cruise ships are repositioned from the Caribbean to
Alaska in summer, or from the Mediterranean to the Caribbean in winter. Recently, a number
of mass-market cruise ships were relocated to Asia to gain profit from the fast-growing Asian
market. The third one is the itinerary planning. A cruise itinerary is similar to a container liner
service (Fransoo and Lee, 2013; Pang and Liu, 2014): both have fixed sequences of ports of
call and fixed schedules (arrival and departure time at each port of call); the itineraries are
announced in advance to attract bookings and cruise ships have to adhere to the announced
itineraries irrespective of whether they are full or not. Moreover, both industries are markedly
capital-intensive and characterized by high fixed costs for operators, who seek a high volume
of bookings to fill their capacity (Wang et al., 2015).
Most itineraries are loops with a home port: the itinerary starts and ends at the home port
and most cruise passengers embark and disembark at the home port. The choice of home ports
by cruise companies depends on the passenger market, the air-lift capacity of the port city, and
the infrastructure and services of the port. Typical examples of home ports include Miami and
Barcelona. Some itineraries are one-way in that they start and end at different home ports: for
instance, trans-Atlantic itineraries. A cruise company also needs to determine which ports of
call to include into an itinerary for a cruise ship. Ports of call are chosen based on the attractions
of the port cities, the infrastructure and services of the ports, and the proximity to other ports
in the itinerary. Under the background, the sequence of visiting the ports of call and the arrival
and departure times at the ports of call needs to be determined.
This study assumes that the home port (or home ports in case of one-way itinerary) and the
ports of call have been chosen in advance and addresses the Cruise Itinerary Schedule Design
(CISD) problem that determines the optimal sequence of the ports of call to visit and the arrival
and departure times at the ports of call. The optimal sequence of the ports of call to visit is
mainly determined by geographical distances. In general, a shorter overall itinerary distance
means less fuel consumption and thereby significant fuel cost savings. As reported by Statista

Chapter 4: Cruise Itinerary Schedule Design 72

(2015), the fuel cost was 220 US dollars per cruise passenger on average, which is 15% of the
cruise expenses. Therefore, one percent reduction in the fuel cost is translated to savings of 48
million US dollars (220 US dollars per cruise passenger times 1% and then times 22.04 million
cruise passengers in 2014) for the industry.
Determining the sequence of ports of call simply based on the overall itinerary distance may
not be optimal. For example, quite often the ports of call are close to each other, as is the case
for the Caribbean and Mediterranean areas, and different sequences may not have much effect
on the overall distance. Moreover, some itineraries are along the coast of a continent (e.g.,
from Sydney to the north along the east coast of Australia) and whether a port of call is
included in the direction away from the home port or back to the home port does not affect the
overall itinerary distance. These observations motivate the development of more sophisticated
models that formulate factors beyond the port distances to determine the sequence of visiting
the ports of call. In particular, we take into account the arrival and departure times at each port
of call. A cruise ship generally visits a port of call in the early morning and departs in the late
afternoon so that cruise passengers can go onshore to have a tour to the port city. In extreme
cases, a cruise ship may stay at a port of call for as short as two hours or as long as two days.
In reality, it may not be possible for a cruise ship to visit all of the ports of call at the same
time of a day because it will mean the ship often has to sail very fast or very slowly from the
previous port of call. In other words, although it is preferable for cruise passengers to spend
more hours in the daytime at each port of call, it comes at the cost of reducing the sailing time
at sea, resulting in higher sailing speed and possibly higher fuel consumption.
Based on the above analysis, this paper presents an explorative study on the CISD problem,
in which the optimal sequence of visiting a given set of ports of call and the arrival time and
departure time at each port of call is to be determined. In view of the practical observations
that most cruise itineraries do not have many ports of call, we first enumerate all sequences of
ports of call and then optimize the arrival and departure times at each port of call by developing
a dynamic programming approach. To improve the computational efficiency, we propose
effective bounds on the profit of each sequence of ports of call, eliminating non-optimal
sequences without invoking the dynamic programming algorithm. Extensive computational
experiments are conducted and the results show that, first, using the bounds on the profit of
each sequence of ports of call considerably improves the computational efficiency; second,
the total profit of the cruise itinerary is sensitive to the fuel price and hence an accurate
estimation of the fuel price is highly desirable; third, determining the sequence of ports of call
solely by minimizing the overall voyage distance leads to significant reduction in the total
profit when the ports do not have a naturally geographical sequence.
The existing researches on cruise shipping, such as the above-mentioned works, are mainly
descriptive. There are few quantitative studies on cruise shipping such as Maddah et al. (2010).

Chapter 4: Cruise Itinerary Schedule Design 73

Given that cruise itineraries have fixed sequences of ports of call and fixed schedules,
optimization-based service planning tools should be able to increase the profit or save the cost
for cruise shipping companies and improve the service quality for cruise passengers. Such
tools are urgent in view of the fast-growing cruise market and the gigantism of cruise ships.
This paper develops quantitative models on the CISD problem and thus contributes to the
state-of-the-art research and practice by developing such a tool.

4.2 LITERATURE REVIEW

We first review academic literature on cruise shipping. As the work is also related to
maritime freight transportation and land transportation, we also relate our work to these studies
after introducing some research works on cruise operations.

4.2.1 Cruise operations

There is not much research about cruise shipping in academic literature. This might be
attributed to the reason that tourism researchers have not paid much attention as worldwide
cruise ship tourism accounts for just about 2% of world tourism (Gui and Russo, 2011), and
maritime researchers mainly focus on freight transportation.
Soriani et al. (2009) investigated the structural aspects and evolutionary trends for cruising
in the Mediterranean region. They identified three crucial issues for the future development of
the region. Gui and Russo (2011) introduced a global value chain framework that connects the
global structure of cruise value chains to the regionally land-based cruise services and reflects
some strategies that local agents can adapt to enhance the generation of value at the local level.
Veronneau and Roy (2009) were amongst the first to lay a descriptive theoretical foundation
of a cruise ship supply chain. They pointed out that in the strategic plan, what is unique to the
cruise industry is that the itinerary planning will affect the supply chain design, demand
forecasts, and product mix. Rodrigue and Notteboom (2013) conducted a deep market
investigation in the cruise industry. They mentioned that vessel deployment strategies and
itinerary design are primordial.
Revenue management (RM) in the cruise industry was a hot topic among researchers.
According to Kimies (1989), cruise lines, just like hotels and airlines, can be deemed as
traditional RM industries. Talluri and van Ryzin (2004) stated that the cruise ships are nothing
more than floating hotels. However, Biehn (2006) strongly disagreed with the common idea
that running a cruise ship is identical to managing a hotel, and claimed that hotel management
guidelines should not be directly used for cruise lines in terms of RM strategies. Meanwhile,
He proposed deterministic linear programming to maximize revenue for a cruise ship
considering the capacity limitations on the number of cabins and the number of lifeboat seats.
Based on his study, Maddah et al. (2010) built a discrete-time dynamic capacity control model
to improve the profit of cruise ships, in which the orders from arrival customers follow a

Chapter 4: Cruise Itinerary Schedule Design 74

stochastic process and request one type of cabin combined with one or more lifeboats. Further
review of the cruise operations can be referred to Wang et al. (2016).

4.2.2 Maritime freight transportation

Cruise shipping is akin to container liner services as both of them have fixed port rotation
and schedule. Moreover, the fuel consumption of cruise ships and container ships are both
related to the speeds of the ships. We refer to Meng et al. (2014) for a review of container liner
service operations and planning. Generally, there are two major differences between the
modeling approaches for the two types of operations. First, one liner service alone is usually
not sufficient to transport containers from their origins to their destinations as containers are
often transshipped during their trips (Ng, 2014, 2015). As a result, a liner service cannot be
designed independently without considering other services, and cargo routing among multiple
shipping service routes is critical (Song and Dong, 2012). However, in cruising shipping, only
one cruise ship is deployed on a cruise itinerary and passengers do not transfer between
different itineraries. Therefore, the schedule design for an itinerary can be implemented
separately. Second, the purpose of docking at ports by container ships is to load and unload
containers. Consequently, it is always desirable for a container ship to spend less time at ports
(Song and Dong, 2011; Du et al., 2015). Different from container shipping, the purpose of
docking at ports by cruise ships is for cruise passengers to visit the port city and hence a longer
port time could be advantageous.
Compared with limited research papers on the cruise shipping, tremendous research works
have been devoted to the container liner shipping. Take the research topic of route design and
schedule design in the container liner shipping as an example: Shintani et al. (2007) proposed
a problem for liner shipping networks design, which consists of dozens of shipping routes. Qi
and Song (2012) worked on the problem of designing an optimal schedule in order to minimize
the total fuel consumption. For the uncertainties in port operations, Wang and Meng (2012a)
studied a robust schedule design problem. Meantime, Wang and Meng (2012b) further
considered sea contingency time for the schedule design problem. Song and Dong (2013)
combined both ship deployment and empty container repositioning into a long-haul shipping
route design.

4.2.3 Land transportation

Land transportation, such as the traveling salesman problem (Applegate et al., 2011) and
the vehicle routing problem (Toth and Vigo, 2001), is also relevant to cruise shipping in that
a vehicle/vessel visits several locations. However, land transportation is different from
maritime transportation because the travel speed inland transportation is largely determined
by the traffic conditions and vehicles usually travel at the highest possible safe speed in the
exogenous traffic conditions. On the contrary, ships can sail freely at sea without congestion

Chapter 4: Cruise Itinerary Schedule Design 75

and they do not often sail at their highest speed mainly for economic reasons: a ship burns
more fuel when it sails faster. As the relation between speed and fuel consumption is nonlinear,
optimization models for cruise itinerary are also nonlinear. A more relevant category of
research is vehicle routing problems with time windows (VRPTW) (Cordeau et al., 2001). The
time window at a customer in VRPTW is a time interval, e.g., 9:00 am to 2:00 pm and the
planning horizon in vehicle routing problems is usually one day. However, the cruise ship
schedule design problem covers a planning horizon of many days and the cruise ship can visit
a port in the daytime on any day; hence the “time window” at a port is a set of disconnected
time intervals. Cruise ship schedule design is also relevant to the traveling salesman problem
with profits (Feillet et al., 2005) as passengers gain extra utility by spending time at ports. The
difference is: the amount of extra utility gained by cruise passengers at a port depends on the
time of visit and duration of visit, rather than a fixed value.
The above literature review shows that existing research on cruise shipping is mainly
descriptive with just a few exceptions (e.g. Maddah et al., 2010). Moreover, cruise shipping
modeling is inherently different from other maritime transportation analysis and land
transportation formulations. Given that cruise services have fixed sequences of ports of call
and fixed schedules, optimization-based service planning tools should be able to increase the
profit or save the cost for cruise shipping companies and improve the service quality for cruise
passengers. Such tools are urgent in view of the fast-growing cruise market and the gigantism
of cruise ships. This paper develops a quantitative solution approach to the CISD problem and
thus contributes to the state-of-the-art research and practice by developing such a tool.

4.3 PROBLEM DESCRIPTION

The CISD problem focuses on schedule design for a cruise itinerary with a given home port
and a set of given ports of call (two home ports will be given in case of one-way itineraries).
The optimal sequence of the ports of call to visit and the arrival and departure time at the ports
of call are to be determined. In the CISD, the deployed cruise ship departs from the home port,
denoted by Port 1, visits a set of given ports of call, denoted by Ports 2, ⋯ , 𝑁 − 1, the sequence
of which is to be determined, and finally returns to the home port, denoted by Port 𝑁, which
is the same port as Port 1 in looped itineraries or is a different port in one-way itineraries. We
use 𝑃𝑐 ≔ {2, ⋯ , 𝑁 − 1} to represent the set of ports of call, and 𝑃 ≔ {1, ⋯ , 𝑁} to represent
the set of all of the ports (both the homeports and ports of call). For the example of a cruise
itinerary in Figure 4.1, Miami is the home port (i.e., Port 1 and Port 𝑁 = 6), at which the cruise
itinerary starts and terminates; there are four selected ports of call for the cruise itinerary and
we can define Cozumel as Port 2, Belize as Port 3, Mahogany Bay as Port 4, and Grand
Cayman as Port 5. Figure 4.1 shows that the cruise ship on the cruise itinerary starts from the
home port at 4:00 pm (Day 1), visits Cozumel at 8:00 am (Day 3), spends nine hours at

Chapter 4: Cruise Itinerary Schedule Design 76

Cozumel, departs at 5:00 pm (Day 3), visits Belize (Day 4), Mahogany Bay (Day 5), Grand
Cayman (Day 6), and returns to the home port (Day 8). As shown in Figure 4.1, an instance of
the solution for the CISD problem is presented, in which the sequence of the ports of call and
the times when the cruise ship arrives at and departs from each port of call are displayed.

Figure 4.1: Itinerary of 7 Day Western Caribbean of Carnival (Carnival Cruise Line, 2016)

In the CISD problem, the following information is required as inputs: (i) The departure time
when a cruise ship leaves the home port and its arrival (return) time at the home port for
termination; (ii) The utility distribution at each port of call for the cruise passengers to
experience (for instance, the utility for the cruise passengers to spend time at a port city at 3:00
am is marginal); (iii) The relationship between bunker consumption and speed on each leg (a
leg is the voyage from one port to the next port). Then, based on these inputs, we make two
critical decisions: (i) The sequence of ports of call for the cruise ship to visit; and (ii) The
arrival and departure time at each port of call. The sequence displays the order list of ports, in
which ports of call must be visited by the cruise ship one by one. The arrival and departure
times confirm the staying time that the ship spends at each port of call and the voyage time on
each leg. The objective of the CISD problem is to maximize the total monetary value from the
utilities brought to cruise passengers at port cities minus the bunker cost of the cruise ship.

4.3.1 Departure time from the home port and the return time
We assume that the cruise ship departs from the home port and returns to the home port at
a pre-determined time. This assumption does not restrict the model but simply aims to simplify
the notation. Without loss of generality, we define that the cruise ship departs at Time 0 and
returns at Time 𝑇. Hence, there are a total of 𝑇 + 1 time points to complete the cruise, denoted
by set 𝕋 = {0, ⋯ , 𝑇}. Here, one time period could be set as one hour, as using one hour in the
schedule for cruise itineraries is precise enough (our model can also handle other time periods,
e.g., half an hour). Note that when we mention “at time 𝑡 ∈ 𝕋” we refer to the time at the end
of the 𝑡th time period (or equivalently, at the beginning of the (𝑡 + 1)th time period).

4.3.2 Utility distribution at ports

Regarding the time spent at port cities, we notice that cruise ships generally visit a port in
the morning and leave in the evening so that cruise passengers can have a tour in the port city.

Chapter 4: Cruise Itinerary Schedule Design 77

Evidently, if a cruise ship visits a port at e.g. 3:00 am, then there is no transport available for
the cruise passengers and there is no place for the cruise passengers to visit.
To capture the impact of arrival and departure times on the cruise itinerary, we need to know
the utility of a port in different hours of a day. One example of the utility distribution at a port
is showed in Figure 4.2. In the daytime hours, the utility is positive. In the night hours when
the port is closed, the utility is zero. The utility for each hour can be estimated by expert
judgment or by analyzing existing cruise itineraries, details of which will be discussed in
Section 4.6.1 and Section 4.7.

Figure 4.2: Utility distribution for one day

We make three comments on the utility shown in Figure 4.2. (i) The utility mentioned here
actually refers to the extra utility by spending time at port cities compared with spending time
at sea on cruise ships (as cruise passengers also have a lot of fun at sea). (ii) Different ports
have different utility distributions. For instance, a world-renowned city like Rome should have
high utilities; cities in which people tend to go to bed and wake up early, have different profiles
from those in which people stay late at nights in bars. (iii) It is possible that the ship stays at a
port when the utility is zero. For instance, in Figure 4.3, when two ports are very close, e.g,
two hours’ sailing, it is possible that the ship stays in Port 𝑗 overnight, when there is no utility,
and leaves the port at 8:00 am on the next day.
To facilitate the solution approach development, we define 𝑔𝑖 (𝑡) as the utility at Port 𝑖 in
Time period 𝑡 ∈ {1,2, ⋯ , 𝑇} . 𝑔𝑖 (𝑡) can be derived based on the daily distribution of the utility.
Considering the example in Figure 4.2, if the cruise ship leaves the home port at 4:00 pm
(which, by our definition, is Time 0), then 6:00-7:00 am of the next day corresponds to Time
period 15 and we thus have 𝑔𝑖 (15) = 2 as the utility for the hour from 6:00 am to 7:00 am is
two. Similarly, we have 𝑔𝑖 (15 + 24) = 2 and 𝑔𝑖 (15 + 48) = 2 in which 24 means one day
and 48 means two days.
To synthesize the total amount of utility that the cruise passengers will experience with the
bunker cost in the objective function, we denote 𝑝𝑢 as the monetary value for the cruise

Chapter 4: Cruise Itinerary Schedule Design 78

company from one unit of utility. The product of the unit monetary value (i.e., 𝑝𝑢 ) and the
total utility that the cruise passengers would experience at ports of call is the total monetary
value for visiting the ports of call during the cruise.

Figure 4.3: Utility distributions and time spent at two ports

4.3.3 Fuel consumption

Different arrival and departure times at ports affect the sailing speed, which impacts the fuel
consumption by the main engine of the cruise ship. According to the research conducted by
Du et al. (2011), the fuel consumption rate of a ship is determined by the speed and can be
estimated as follows.
𝑘 + 𝑘 ′ ∙ (𝑣)𝑠 (4.1)
′
Here 𝑘 and 𝑘 are regression coefficients, 𝑣 is the sailing speed, and 𝑠 ∈ {3.5, 4, 4.5}. For
feeders, 𝑠 = 3.5 ; for medium-sized vessels, 𝑠 = 4 ; and for jumbo vessels, 𝑠 = 4.5 . To
calculate the fuel consumption of a leg, let 𝑑 be the distance of the leg and 𝜏 be the sailing
time on the voyage, implying that the speed is 𝑣 = 𝑑/𝜏. Therefore, the bunker consumption
on the leg, denoted by function 𝐹̃ (𝑑, 𝜏), is
𝐹̃ (𝑑, 𝜏) = [𝑘 + 𝑘 ′ ∙ (𝑣)𝑠 ] ∙ 𝜏 = 𝑘𝜏 + 𝑘 ′ 𝑑 𝑠 𝜏 1−𝑠 (4.2)
There exists an optimal sailing speed, denoted by 𝑣 ∗, to save the fuel, which can be derived
by minimizing the consumption in Eq. (4.2). The calculation for the optimal sailing speed is:
𝑘 1/𝑠
𝑣 ∗ = (𝑘 ′ ∙(𝑠−1)) (4.3)

The cruise ship can decelerate (or accelerate) if its sailing speed exceeds (or is lower than) the
optimal speed in order to save fuel.
Given 𝑑 and 𝜏, the average speed is 𝑑/𝜏. If 𝑑/𝜏 ≥ 𝑣 ∗ , the ship should sail at a constant
speed that is equal to 𝑑/𝜏 for saving fuel consumption. Otherwise, the ship should sail at its
optimal speed to the destination and then wait at the destination (Wang et al., 2013). Thus,
given 𝑑 and 𝜏, the minimum fuel consumption, denoted by function 𝐹(𝑑, 𝜏), is denoted as:
𝐹̃ (𝑑, 𝜏) , 𝑖𝑓 𝑑/𝜏 ≥ 𝑣 ∗
𝐹(𝑑, 𝜏) = { (4.4)
[𝑘 + 𝑘 ′ ∙ (𝑣 ∗ )𝑠 ] ∙ (𝑑⁄𝑣 ∗ ) , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Chapter 4: Cruise Itinerary Schedule Design 79

 In the CISD, we define 𝑑𝑖𝑗 and 𝜏𝑖𝑗 as the voyage distance and voyage time between Port 𝑖
and Port 𝑗. 𝑑𝑖𝑗 is an input data, which can be easily obtained from a geographical database. 𝜏𝑖𝑗
is meaningful only if the cruise ship visits Port 𝑗 directly after Port 𝑖; if this is the case, 𝜏𝑖𝑗 is
the time interval between the departure from Port 𝑖 and the arrival at Port 𝑗, and hence is a
decision variable. Given 𝑑𝑖𝑗 and 𝜏𝑖𝑗 , the minimum bunker consumption between Port 𝑖 and
Port 𝑗 can be calculated by 𝐹(𝑑𝑖𝑗 , 𝜏𝑖𝑗 ) in Eq. (4.4). To convert the fuel consumption into the
fuel cost, the unit price of fuel, denoted by 𝑐 𝐹 , is needed as an input.

4.3.4 Sequence of ports of call

Determining the sequence of ports of call is a crucial decision that we should make for the
CISD problem. To represent the sequence in the manner of mathematical models, here, we use
the same way to define it as many vehicle routing problems (VRPs) do: Set 𝑥𝑖𝑗 to one if the
cruise ship visits Port 𝑗 immediately after visiting Port 𝑖, and zero otherwise.
It is worthwhile to mention that visa restrictions of cruise passengers should be considered
when designing the sequence of ports of call. Specifically, some ports of call belong to the
same country and must be visited without interruption. For instance, if the cruise itinerary is
Shanghai (China) Nagoya (Japan) Busan (Korea) Kobe (Japan) Shanghai (China),
then the cruise passengers from China must obtain a tourist visa for Japan that allows multiple
entries to Japan. If this is difficult for the cruise passengers, the two Japanese ports should be
visited without interruption, for example, Shanghai (China) Busan (Korea) Nagoya
(Japan) Kobe (Japan) Shanghai (China). To capture this practical consideration, we define
ℍ as the set of countries that can only be entered once, 𝐻𝑟 as the set of ports belonging to
Country 𝑟 ∈ ℍ, and 𝑁𝑟 as the number of ports in Country 𝑟, 𝑁𝑟 ≔ |𝐻𝑟 |.

4.3.5 Arrival and departure times at ports of call

For arrival time and departure times at ports of call, it does not make sense for a cruise ship
to arrive at a time when the port is closed, for instance, at 3:00 am or leave too late. Therefore,
we define sets of possible arrival and departure times based on realistic situations as follows.
First, we note that different ports may be located in different time zones and ignoring the
difference in time zones will lead to incorrect decisions. Second, given the opening hours in a
day for a port of call (evidently, the opening hours refer to the local time zone), the time zones
of the home port and the port of call, we can define the time windows of the port of call in the
planning horizon (i.e., from Time 0 to Time 𝑇). For instance, if a cruise ship departs from the
home port (time zone: UTC+8) at 8:00 pm (i.e., Time 0 in our model), Port of call 𝑖 is in time
zone UTC+9 and opens every day from 7:00 am to 3:00 pm, then, the time windows of Port
of call 𝑖 is 𝒯𝑖 = [10, 18] ∪ [(10 + 24), (18 + 24)] ∪ [(10 + 48), (18 + 48)] ⋯. The arrival

Chapter 4: Cruise Itinerary Schedule Design 80

and departure times of the cruise ship at Port of call 𝑖, denoted by 𝑎𝑖 and 𝑏𝑖 , respectively, must
be in the set, 𝑎𝑖 ∈ 𝒯𝑖 , 𝑏𝑖 ∈ 𝒯𝑖 , 𝑖 ∈ 𝑃𝑐 .
We define 𝑚𝑖 as the minimum time (e.g., five consecutive hours) that the cruise ship should
stay at Port 𝑖 before the port closes when it arrives at a port during its opening hours 𝒯𝑖 . During
the 𝑚𝑖 hours, the cruise ship could replenish consumables or fuel and the cruise passengers
could have a tour around the city. Given minimal staying hours in ports, we can refine the set
of arrival time windows at Port of call 𝑖 , for instance, if 𝑚𝑖 = 6 and 𝒯𝑖 = [10, 18] ∪
[(10 + 24), (18 + 24)] ∪ [(10 + 48), (18 + 48)] ⋯, the set of possible arrival times (i.e., 𝑎𝑖 )
is 𝒯𝑖 ′ = [10, 12] ∪ [(10 + 24), (12 + 24)] ∪ [(10 + 48), (12 + 48)] ⋯.

4.3.6 Model for Cruise Itinerary Schedule Design (CISD)

This section formulates a mathematical model for the general CISD problem, denoted by
𝐹0. Then, to reduce the amount of input data required, we make some modifications on the
model to solve a special case, denoted by 𝐹0′. Before presenting the models for the CISD
problem, we list the notation below.

Indices and sets:

𝑖: index of a port
𝑡: index of a time period
𝑃: set of all ports of call and home ports, 𝑃 = {1,2, ⋯ , 𝑁 − 1, 𝑁}, where 1 and 𝑁 represent
home ports
𝑃𝑐 : set of all ports of call, 𝑃𝑐 = {2,3, ⋯ , 𝑁 − 2, 𝑁 − 1}, excluding home ports
𝕋 : set of all time periods in one cruise, 𝕋 = {0,1,2, ⋯ , 𝑇 − 1, 𝑇}
Decision variables:
𝑎𝑖 : time when the cruise ship arrives at Port 𝑖
𝑏𝑖 : time when the cruise ship departs from Port 𝑖
𝑥𝑖𝑗 : binary, set to one if Port 𝑖 is immediately followed by Port 𝑖 in the voyage of the cruise
ship
𝜃𝑖𝑗 : sailing time on the leg from Port 𝑖 to Port 𝑗
Input data:
𝐹(𝑑, 𝑡): fuel consumption (tones) of the cruise ship if the voyage distance is 𝑑 and the
sailing time is 𝑡
𝑐 𝐹 : unit fuel price of the cruise ship
𝑑𝑖𝑗 : voyage distance between Port 𝑖 and Port 𝑗
𝑔𝑖 (𝑡) : number of the utility at Port 𝑖 for Time period 𝑡
𝑚𝑖 : minimum time that the cruise ship should stay in Port of call i
𝑝𝑢 : unit monetary value of the utility

Chapter 4: Cruise Itinerary Schedule Design 81

 𝑣 𝑚𝑎𝑥 : maximum speed of the cruise ship
𝒯𝑖 : set of all possible visiting time windows of Port 𝑖
𝒯𝑖 ′ : set of all possible arrival time windows of Port 𝑖
𝐻𝑟 : set of ports that belong to Country 𝑟 ∈ ℍ
ℍ: set of countries that can only be entered once
𝑁𝑟 : number of ports in 𝐻𝑟
𝑀: a sufficiently large number
Mathematical model:
𝑏 −1
[𝑭𝟎] Maximize 𝑍 = 𝑝𝑢 ∙ ∑𝑖∈𝑃𝑐 ∑𝑡=𝑎
𝑖
𝑔 (𝑡) − 𝑐 𝐹 ⋅ ∑𝑖∈𝑃 ∑𝑗∈𝑃 𝑥𝑖𝑗 ∙ 𝐹(𝑑𝑖𝑗 , 𝜃𝑖𝑗 )
𝑖 𝑖
(4.5)
s.t. ∑𝑗∈𝑃 𝑥𝑖𝑗 = 1 ∀𝑖 ∈ 𝑃𝑐 (4.6)
∑𝑖∈𝑃 𝑥𝑖𝑘 − ∑𝑗∈𝑃 𝑥𝑘𝑗 = 0 ∀𝑘 ∈ 𝑃𝑐 (4.7)
∑𝑗∈𝑃𝑐 𝑥1,𝑗 = 1 (4.8)
∑𝑖∈𝑃𝑐 𝑥𝑖,𝑁 = 1 (4.9)
∑𝑖∈𝐻𝑟 ∑𝑗∈𝐻𝑟 𝑥𝑖𝑗 = 𝑁𝑟 − 1 ∀𝐻𝑟 ∈ ℍ (4.10)
𝑏𝑖 + 𝜃𝑖𝑗 − 𝑀 ⋅ (1 − 𝑥𝑖𝑗 ) ≤ 𝑎𝑗 ∀𝑖 ∈ 𝑃\{𝑁}, 𝑗 ∈ 𝑃\{1}, 𝑖 ≠ 𝑗 (4.11)
𝑎𝑖 + 𝑚𝑖 ≤ 𝑏𝑖 ∀𝑖 ∈ 𝑃𝑐 (4.12)
𝜃𝑖𝑗 ≥ ⌈𝑑𝑖𝑗 /𝑣 𝑚𝑎𝑥 ⌉ ∀𝑖, 𝑗 ∈ 𝑃, 𝑖 ≠ 𝑗 (4.13)
𝑏1 ≔ 0 (4.14)
𝑎𝑁 ≔ 𝑇 (4.15)
𝑥𝑖𝑗 ∈ {0,1} ∀𝑖, 𝑗 ∈ 𝑃 (4.16)
𝑎𝑖 ∈ 𝒯𝑖 ′ ∀𝑖 ∈ 𝑃𝑐 (4.17)
𝑏𝑖 ∈ 𝒯𝑖 ∀𝑖 ∈ 𝑃𝑐 (4.18)

In the above model 𝐹0, Objective (4.5) maximizes the monetary value of the total utilities
minus the bunker costs. Constraints (4.6) and (4.7) guarantee that the cruise ship visits each
port of call exactly once. Constraints (4.8) and (4.9) ensure that each cruise starts at the home
port and goes back to the home port finally. Constraints (4.10) states that ports of call in the
same country must be visited without interruption for the sake of the visa restrictions.
Constraints (4.11) show the relationship between the departure time at one port and the arrival
time at the next visited port. Constraints (4.12) ensures that the cruise ship dwells in each port
for at least a certain period of time (i.e., 𝑚𝑖 ). Constraints (4.13) ensure that the real voyage
time between two ports should be more than the minimal voyage time, in which the cruise ship
sails at the maximum speed. Constraint (4.14) and Constraint (4.15) guarantee that the cruise
ship departs from the home port and returns to it at specified times. Constraints (4.16) indicate
that there is no partial connection between two ports. Constraints (4.17) and Constraints (4.18)
ensure that the cruise ship must arrive at one port and leave from it during its opening hours.

Chapter 4: Cruise Itinerary Schedule Design 82

 In reality, it may be difficult for a cruise company to define the utility distribution of each
port (i.e., 𝑔𝑖 (𝑡)), then we consider a new special case of the model, in which the staying time
in each Port 𝑖 (i.e., 𝑏𝑖 − 𝑎𝑖 ) is not influenced by the utility profiles. Instead, 𝑏𝑖 − 𝑎𝑖 just needs
to exceed the minimal staying hours in port 𝑖 (i.e., 𝑚𝑖 ). The model, denoted by 𝐹0′ for this
special case is formulated as follow:
[𝑭𝟎′]: Minimize 𝑍 = 𝑐 𝐹 ⋅ ∑𝑖∈𝑃 ∑𝑗∈𝑃 𝑥𝑖𝑗 ∙ 𝐹(𝑑𝑖𝑗 , 𝜃𝑖𝑗 ) (4.19)
s.t. Constraints (4.6-4.18)

4.4 SOLUTION APPROACH FOR CISD

In this section, we develop an efficient solution algorithm to obtain optimal solutions by

analyzing some special features of the problem.

4.4.1 Complexity of the CISD problem

Proposition 4.1: The CISD problem is NP-hard.
Proof: Suppose that all of the utilities 𝑔𝑖 (𝑡) are zero, all of the minimum staying times 𝑚𝑖
are zero, all of the time windows 𝒯𝑖 ′ = 𝒯𝑖 = 𝕋, and the fuel consumption function 𝐹(𝑑, 𝑡) is
proportional to the distance 𝑑 and there is no visa restriction. Then the CISD problem becomes
the one that finds the shortest distance to visit all of the ports of call from the homeport and
then returns to the homeport, which is exactly the travelling salesman problem (TSP). Since
the TSP is NP-hard (Cormen et al., 2009), the general version of the CISD problem is also NP-
hard. ∎

4.4.2 Dynamic programming for the model

Despite the NP-hardness of the problem in nature, we find that in realistic cases the number
of ports of call is not large. Figure 4.4 shows the statistics on the trips in 2016 of the biggest
cruise ship—The Carnival Vista—owned by Carnival Cruise Line. From the figure, we note
that the number of ports of call and cycle time in a cruise itinerary is not large, ranging from
two to nine and 5 days to 13 days, respectively. Thus, we could enumerate all of the sequences
of visiting the ports of call for one specific cruise itinerary. Meanwhile, as can be seen in the
third part of the figure, the ports of call among the trips are normally located in several
countries. When considering the visa restrictions, some infeasible sequences for the ports of
call can be deleted directly without exploring.
The total number of possible sequences for (𝑁 − 2) ports of call is (𝑁 − 2)!. Let 𝑠 denote
the index of a sequence, 𝑠 ∈ 𝑆 = {1,2, ⋯ , (𝑁 − 2)! − 1, (𝑁 − 2)!}. Then, for each sequence 𝑠,
we develop a dynamic programming (DP) approach to find the optimal arrival and departure
times for each port of call. As mentioned in Section 4.4.3, the sequence (denoted by 𝑥𝑖𝑗 , ∀𝑖 ∈
𝑃\{𝑁}, 𝑗 ∈ 𝑃\{1}, 𝑖 ≠ 𝑗) and the arrival and departure times of each port of call (i.e.,𝑎𝑖 and

Chapter 4: Cruise Itinerary Schedule Design 83

𝑏𝑖 , ∀𝑖 ∈ 𝑃𝑐 ) are two critical decisions in the CISD. The combined enumeration and DP
approach could hence obtain the optimal solution for the CISD problem.

Figure 4.4: Statistic on trips for the Carnival Vista (Cruise Ship Schedule, 2016)

We now consider a special sequence 𝑠 that visits the ports of call according to their IDs, i.e.,
1, ⋯ 𝑗, 𝑗 + 1 ⋯ 𝑁, in which 1 and 𝑁 refer to the home ports and 𝑗 is the 𝑗 𝑡ℎ port visited on the
cruise itinerary. The cruise ship leaves Port 1 at Time 0 (i.e., 𝑏1 = 0), and returns to Port 𝑁 at
Time 𝑇 (i.e., 𝑎𝑁 = 𝑇). We need to determine the optimal arrival and departure times at each
Port 2, ⋯ 𝑗, 𝑗 + 1 ⋯ 𝑁 − 1 . The purpose of examining this special case is for notational
convenience; any sequence can be addressed using the same DP approach.
To apply DP, we firstly define 𝑈𝑗 (𝑡) as the maximum possible profit (to be determined)
from Time 𝑡 to 𝑇, if the cruise ship arrives at Port 𝑗 at Time 𝑡, and 𝑉𝑗 (𝑡) as the maximum
possible profit (to be determined) from Time 𝑡 to 𝑇, if the cruise ship leaves Port 𝑗 at Time 𝑡.
𝑝
Then, we define 𝜏𝑗𝑣 as the voyage time from the 𝑗 𝑡ℎ port to the (𝑗 + 1)𝑡ℎ port, and 𝜏𝑗 as the
𝑝
staying time at the 𝑗 𝑡ℎ port. The 𝜏𝑗𝑣 and 𝜏𝑗 are two decision variables in the DP algorithm.
Third, since the cruise ship has a maximum sailing speed 𝑣 𝑚𝑎𝑥 and the minimum time spent
at Port 𝑗 is 𝑚𝑗 , the earliest possible arrival time at Port 𝑗 can be computed by assuming the
cruise ship sails at its highest speed and spends the least time at previous ports of call:
𝑗−1 𝑑
∑𝑘=1 ⌈ 𝑘,𝑘+1
𝑚𝑎𝑥 ⌉ ,𝑗 = 2
𝑎𝑗𝑚𝑖𝑛 = { 𝑗−1 𝑑𝑣 𝑗−1
(4.20)
∑𝑘=1 ⌈ 𝑣𝑘,𝑘+1
𝑚𝑎𝑥 ⌉ + ∑𝑘=2 𝑚𝑘 , 𝑗 = {3,4, ⋯ 𝑁 − 1}

The latest departure time is:

𝑑
𝑇 − (∑𝑁−1 𝑘,𝑘+1 𝑁−1
𝑘=𝑗 ⌈ 𝑣 𝑚𝑎𝑥 ⌉ + ∑𝑘=𝑗+1 𝑚𝑘 ) , 𝑗 = {2,3, ⋯ 𝑁 − 2}
𝑏𝑗𝑚𝑎𝑥 ={ 𝑑𝑘,𝑘+1
(4.21)
𝑇 − ∑𝑁−1
𝑘=𝑗 ⌈ 𝑣 𝑚𝑎𝑥 ⌉ ,𝑗 = 𝑁 − 1

Define
𝑎𝑗𝑚𝑎𝑥 : = 𝑏𝑗𝑚𝑎𝑥 − 𝑚𝑗 , 𝑏𝑗𝑚𝑖𝑛 ≔ 𝑎𝑗𝑚𝑖𝑛 + 𝑚𝑗 ∀𝑗 ∈ 𝑃𝑐 (4.22)

Chapter 4: Cruise Itinerary Schedule Design 84

To be feasible, the arrival time at Port 𝑗 must be in the interval [𝑎𝑗𝑚𝑖𝑛 , 𝑎𝑗𝑚𝑎𝑥 ] and the departure

time must be in [𝑏𝑗𝑚𝑖𝑛 , 𝑏𝑗𝑚𝑎𝑥 ].

In the DP approach, we have the boundary conditions:
−∞, 𝑡 ≠ 𝑇
𝑈𝑁 (𝑡) = { ∀𝑡 ∈ 𝕋 (4.23)
0 ,𝑡 = 𝑇
and the recursive relations between 𝑉𝑗 (𝑡) and 𝑈𝑗 (𝑡) for all 𝑡 ∈ 𝕋, 𝑗 ∈ {1, 2, ⋯ 𝑁 − 1} are:
𝑚𝑎𝑥⌈𝑑𝑗,𝑗+1/𝑣 𝑚𝑎𝑥 ⌉≤𝜏𝑗𝑣 ≤𝑇−𝑡 {−𝑐 𝐹 ⋅ 𝐹(𝑑𝑗,𝑗+1 , 𝜏𝑗𝑣 ) + 𝑈𝑗+1 (𝑡 + 𝜏𝑗𝑣 )} , 𝑡 ∈ 𝒯𝑗 ∩ [𝑏𝑗𝑚𝑖𝑛 , 𝑏𝑗𝑚𝑎𝑥 ]
𝑉𝑗 (𝑡) = { (4.24)
−∞ , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑝
𝑡+𝜏 −1 𝑝
max𝑚𝑗 ≤𝜏𝑝 ≤𝑇−𝑡 {𝑝𝑢 ∙ ∑ℎ=𝑡𝑗 𝑔𝑗 (ℎ) + 𝑉𝑗 (𝑡 + 𝜏𝑗 )} , 𝑡 ∈ 𝒯𝑗 ′ ∩ [𝑎𝑗𝑚𝑖𝑛 , 𝑎𝑗𝑚𝑎𝑥 ]
𝑈𝑗 (𝑡) = { 𝑗 (4.25)
−∞ , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Eq. (4.24) is used to maximize the profit from Time 𝑡 to 𝑇 given that the cruise ship leaves
Port 𝑗 at Time 𝑡 . Here, the possible departure times at Port 𝑗 should be restricted in the
intersection of 𝒯𝑗 and [𝑏𝑗𝑚𝑖𝑛 , 𝑏𝑗𝑚𝑎𝑥 ]. The 𝒯𝑗 is the set of opening hours of Port 𝑗. To enforce
that the ship must leave Port 𝑗 within its opening hours (i.e., time windows), we also need to
include 𝒯𝑗 in Eq. (4.24). The [𝑏𝑗𝑚𝑖𝑛 , 𝑏𝑗𝑚𝑎𝑥 ] is the possible departure time range defined by Eq.
(4.20), (4.21) and (4.22) based on the speed of the ship and minimal staying hours at the ports
of call. It is impossible for the cruise ship to depart from Port 𝑗 at Time 𝑡 if the 𝑡 is out of the
range (i.e., [𝑏𝑗𝑚𝑖𝑛 , 𝑏𝑗𝑚𝑎𝑥 ]). The intersection of these two parts offers a strict limit on the
possible departure time at Port 𝑗 . This is crucial for Eq. (4.24), because the 𝑉𝑗 (𝑡) with
impossible Time 𝑡 should be set to −∞ in order to avoid unexpected problems in processing
DP. For all possible Time 𝑡, the 𝑉𝑗 (𝑡) is calculated by 𝑈𝑗+1 (𝑡 + 𝜏𝑗𝑣 ) and −𝑐 𝐹 ⋅ 𝐹(𝑑𝑗,𝑗+1 , 𝜏𝑗𝑣 ).
The former is the maximal profit from Time 𝑡 + 𝜏𝑗𝑣 to 𝑇, if the cruise ship arrives at Port 𝑗 + 1
at time 𝑡 + 𝜏𝑗𝑣 . As this maximal profit of Port 𝑗 + 1 has been obtained beforehand in the DP
process, the 𝑉𝑗 (𝑡) can be easily achieved by adding the bunker fuel profit (i.e., −𝑐 𝐹 ⋅
𝐹(𝑑𝑗,𝑗+1 , 𝜏𝑗𝑣 )) between Port 𝑗 and Port 𝑗 + 1. For the maximum 𝑉𝑗 (𝑡), we need to determine
the optimal voyage time (i.e., 𝜏𝑗𝑣 ) between two ports, which is selected within the possible

range (i.e., ⌈𝑑𝑗,𝑗+1 /𝑣 𝑚𝑎𝑥 ⌉ ≤ 𝜏𝑗𝑣 ≤ 𝑇 − 𝑡). Here, we notice that the range may be infeasible
when time 𝑡 is large. For instance, when 𝑡 = 𝑇, the upper bound of the range is zero, which is
less than the lower bound of the range. To avoid these infeasible cases, we set 𝑉𝑗 (𝑡) to −∞
directly without calculation when the upper bound is less than the lower bound.
Eq. (4.25) is used to maximize the profit from Time 𝑡 to 𝑇 given that the cruise ship arrives
at Port 𝑗 at Time 𝑡. Similar to Eq. (4.24), the possible arrival time set at port 𝑗 is also to be
defined before calculation, which is the intersection of 𝒯𝑗 ′ and [𝑎𝑗𝑚𝑖𝑛 , 𝑎𝑗𝑚𝑎𝑥 ]. The possible

arrival time range (i.e., [𝑎𝑗𝑚𝑖𝑛 , 𝑎𝑗𝑚𝑎𝑥 ]) based on the speed of the ship and minimal staying hours
at the ports of call is also achieved by Eq. (4.20), (4.21) and (4.22). For all possible arrival

Chapter 4: Cruise Itinerary Schedule Design 85

 𝑝
𝑝 𝑡+𝜏𝑗 −1
𝑢
Time 𝑡, the 𝑈𝑗 (𝑡) is determined by 𝑉𝑗 (𝑡 + 𝜏𝑗 ) and 𝑝 ∙ ∑ℎ=𝑡 𝑔𝑗 (ℎ) . The former is the
𝑝 𝑝
maximal profit from time 𝑡 + 𝜏𝑗 to 𝑇, if the cruise ship leaves Port 𝑗 at time 𝑡 + 𝜏𝑗 . As this
maximal profit of Port 𝑗 in terms of departure times has been obtained previously, the 𝑈𝑗 (𝑡)
𝑝
𝑡+𝜏 −1
can be easily obtained by adding the utility profit (i.e., 𝑝𝑢 ∙ ∑ℎ=𝑡𝑗 𝑔𝑗 (ℎ)) at Port 𝑗. For the
𝑝
maximum 𝑈𝑗 (𝑡), we need to determine the optimal staying time (i.e., 𝜏𝑗 ) at Port 𝑗, which is
𝑝
chose within the possible range (i.e., 𝑚𝑗 ≤ 𝜏𝑗 ≤ 𝑇 − 𝑡). Similar to Eq. (4.24), the range could
also be feasible. Thus, we set 𝑈𝑗 (𝑡) to minus infinity when the upper bound of the range is less
than the lower bound of the range (i.e., 𝑇 − 𝑡 < 𝑚𝑗 ).
𝑉1 (0) represents the maximum possible profit from Time 0 to 𝑇 if the cruise ship departs
from Port 1 at Time 0 and the value of 𝑉1 (0) is obtained by executing the above algorithm. In
essence, the value of the 𝑉1 (0) is the optimal objective value for the sequence. The optimal
arrival and departure times at each port of call under this sequence can be calculated by using
𝑝
the optimal decision variables (𝜏𝑗𝑣 )∗ and (𝜏𝑗 )∗ in DP, which is shown as follow.
𝑗−1
∗
∑𝑘=1 (𝜏𝑗𝑣 )∗ ,𝑗 = 2
𝑎𝑗 = { 𝑗−1 𝑣 ∗ 𝑗−1 𝑝 (4.26)
∑𝑘=1 (𝜏𝑗 ) + ∑𝑘=2 (𝜏𝑗 )∗ , 𝑗 = 3, 4, ⋯ 𝑁 − 1
𝑝
𝑏𝑗 ∗ = 𝑎𝑗 ∗ + (𝜏𝑗 )∗ , 𝑗 = 2,3, ⋯ 𝑁 − 1. (4.27)
The detailed algorithm for the above-mentioned DP is given as follows:
Algorithm 4.1: Dynamic programming
Initialize 𝑈𝑁 (𝑡), ∀𝑡 ∈ 𝕋 according to Eq. (4.23);
For 𝑗 = 𝑁 − 1 𝑡𝑜 𝑗 = 1
If 𝑗 ≠ 1 Then
For 𝑡 = 0 𝑡𝑜 𝑡 = 𝑇
If 𝑇 − 𝑡 ≥ ⌈𝑑𝑗,𝑗+1 /𝑣 𝑚𝑎𝑥 ⌉ Then
Calculate 𝑉𝑗 (𝑡) by using Eq. (4.24);
Else
𝑉𝑗 (𝑡) = −∞ ;
End If
End For
For 𝑡 = 0 𝑡𝑜 𝑡 = 𝑇
If 𝑇 − 𝑡 ≥ 𝑚𝑗 Then
Calculate 𝑈𝑗 (𝑡) by using Eq. (4.25);
Else
𝑈𝑗 (𝑡) = −∞ ;
End If
End For
Else // 𝑗 = 1
Calculate 𝑉𝑗 (0) by using Eq. (4.24), output the optimal values of the
𝑝
decision variables 𝜏𝑗𝑣 and 𝜏𝑗 , and return.
End If
End For

Chapter 4: Cruise Itinerary Schedule Design 86

4.4.3 Improving the enumeration
When we enumerate all of the sequences, we may stop once we know that this sequence
cannot be better than the incumbent best one. To this end, we develop an efficient method to
find a high-quality upper bound on the profit of a sequence.
To begin with, the total cruise rotation time (i.e., 𝑇) is divided into the sailing time 𝑇̂ (i.e.,
the total voyage time at sea) and port time 𝑇 − 𝑇̂ (i.e., the total staying hours in the ports of
call). The optimal division is to be determined. Then, given a sequence 𝑠, we let 𝑗(𝑠) be the
ID of the physical port of the 𝑗 𝑡ℎ port visited on 𝑠. Here, we need to note that there is an
underlying bound for the total sailing time (i.e., 𝑇̂) when considering the speed restriction on
sail (i.e., the speed cannot exceed 𝑣 𝑚𝑎𝑥 ) and staying time restriction in ports (i.e., the stating
in each port of call must exceed 𝑚𝑖 ). The bound for the total sailing time in each sequence 𝑠
is.
∑𝑁−1
𝑗=1 ⌈𝑑𝑗(𝑠),(𝑗+1)(𝑠) /𝑣
𝑚𝑎𝑥
⌉ ≤ 𝑇̂ ≤ 𝑇 − ∑𝑁−1
𝑖=1 𝑚𝑖 (4.28)
where we notice that for some sequences 𝑠, the upper bound in (4.28) (i.e., 𝑇 − ∑𝑁−1
𝑖=1 𝑚𝑖 )

could be less than the lower bound in (4.28) (i.e., ∑𝑁−1

𝑗=1 ⌈𝑑𝑗(𝑠),(𝑗+1)(𝑠) /𝑣
𝑚𝑎𝑥
⌉), which means
under these sequences, it is impossible for the cruise ship to return to the home port on time
even if it sails at the maximum speed in the whole cruise. To improve the enumeration, we
just drop these sequences and calculate the next one.
An upper bound on the profit from the negative fuel cost when the total sailing time is 𝑇̂,
denoted by 𝑈𝐵𝑠 (𝑇̂, 𝑠), can be calculated as follow.
−𝑐 𝐹 ∙ 𝐹(∑𝑁−1 ̂ ̂
𝑗=1 𝑑𝑗(𝑠),(𝑗+1)(𝑠) , 𝑇 ) , 𝑇 𝑠𝑎𝑡𝑖𝑠𝑓𝑖𝑒𝑠 (4.28)
𝑈𝐵𝑠 (𝑇̂, 𝑠) = { (4.29)
−∞ , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where fuel consumption is the lowest when the cruise ship sails at a constant speed.
An upper bound on the monetary value of the total utilities when the total port time is 𝑇 −
𝑇̂, denoted by 𝑈𝐵𝑝 (𝑇 − 𝑇̂), can be calculated by solving an integer-linear program. First, we
let
𝜏+𝛿 −1
𝐺𝑗 (𝛿𝑗 ) ≔ 𝑚𝑎𝑥0≤𝜏≤23 ∑ℎ=𝜏𝑗 𝑝𝑢 ∙ 𝑔𝑗 (ℎ) (4.30)
That is, 𝐺𝑗 (𝛿𝑗 ) is the maximum monetary value from the utility for spending 𝛿𝑗 consecutive
hours at Port 𝑗. To obtain 𝑈𝐵𝑝 (𝑇 − 𝑇̂), we let binary variable 𝑧𝑗𝛿𝑗 be one if and only if the

time spent at Port 𝑗 is 𝛿𝑗 . Since the minimum time spent at Port 𝑗 is 𝑚𝑗 , to have a feasible port
time solution, the time spent at Port 𝑗 must be between 𝑚𝑗 and 𝑇 − 𝑇̂ − ∑𝑖∈𝑃𝑐 \{𝑗} 𝑚𝑖 . The
model for obtaining 𝑈𝐵𝑝 (𝑇 − 𝑇̂) is:
𝑇−𝑇−∑ 𝑚 ̂
[𝑭𝟏] 𝑈𝐵𝑝 (𝑇 − 𝑇̂) = 𝑚𝑎𝑥 ∑𝑗∈𝑃𝑐 ∑𝛿 =𝑚 𝑖∈𝑃𝑐\{𝑗} 𝑖 𝐺𝑗 (𝛿𝑗 )𝑧𝑗𝛿𝑗 (4.31)
𝑗 𝑗

subject to:

Chapter 4: Cruise Itinerary Schedule Design 87

 𝑇−𝑇̂−∑𝑖∈𝑃𝑐\{𝑗} 𝑚𝑖
∑𝛿 𝑧𝑗𝛿𝑗 =1 ∀𝑗 ∈ 𝑃𝑐 (4.32)
𝑗 =𝑚𝑗

𝑇−𝑇̂−∑𝑖∈𝑃𝑐\{𝑗} 𝑚𝑖
∑𝑗∈𝑃𝑐 ∑𝛿 𝛿𝑗 𝑧𝑗𝛿𝑗 = 𝑇 − 𝑇̂ (4.33)
𝑗 =𝑚𝑗

𝑧𝑗𝛿𝑗 ∈ {0,1} ∀𝑗 ∈ 𝑃𝑐 , 𝛿𝑗 ∈ {𝑚𝑗 , 𝑚𝑗 + 1, ⋯ , 𝑇 − 𝑇̂ − ∑𝑖∈𝑃𝑐 \{𝑗} 𝑚𝑖 } (4.34)

Note that although 𝑭𝟏 is an integer linear program, we only need to solve it once for each
possible value of 𝑇 − 𝑇̂ in one CISD problem.
An upper bound on the profit of a sequence 𝑠 can now be easily obtained:
𝑈𝐵(𝑠) = 𝑚𝑎𝑥∑𝑁−1⌈𝑑 ̂ ≤ 𝑇−∑𝑁−1
𝑚𝑎𝑥 ⌉≤𝑇 [𝑈𝐵𝑠 (𝑇̂, 𝑠) + 𝑈𝐵𝑝 (𝑇 − 𝑇̂)] (4.35)
𝑗=1 𝑗(𝑠),(𝑗+1)(𝑠) /𝑣 𝑖=1 𝑚𝑖

Based on above-mentioned definitions and formulations, the algorithm for improving the
enumeration can be designed (denoted as Algorithm 4.2) and is shown as:
Algorithm 4.2: Improving the enumeration:
Define the incumbent best profit: 𝑈𝐵𝑏𝑒𝑠𝑡 = −∞;
Calculate 𝑈𝐵𝑝 (𝑇 − 𝑇̂) for all possible values of 𝑇 − 𝑇̂;
Enumerate each sequence: sequence 𝑠 ∈ 𝑆;
For 𝑠 = 1 𝑡𝑜 𝑠 = (𝑁 − 2)!
If the sequence 𝑠 violates the visa restriction Then
Continue;
End If
If 𝑇 − ∑𝑁−1 𝑁−1
𝑖=1 𝑚𝑖 < ∑𝑗=1 ⌈𝑑𝑗(𝑠),(𝑗+1)(𝑠) /𝑣
𝑚𝑎𝑥
⌉ Then
Continue;
End If
Calculate the upper bound on the profit of the sequence 𝑠 by Eq. (4.35)
If 𝑈𝐵(𝑠) ≤ 𝑈𝐵𝑏𝑒𝑠𝑡 Then
Continue;
Else
Use DP (i.e., Algorithm 4.1) to find the maximum profit of the sequence 𝑠,
denoted by 𝑈𝐵𝑠𝐷𝑃 .
If 𝑈𝐵𝑏𝑒𝑠𝑡 < 𝑈𝐵𝑠𝐷𝑃 Then
Set 𝑈𝐵𝑏𝑒𝑠𝑡 = 𝑈𝐵𝑠𝐷𝑃 and record the sequence 𝑠;
End If
End If
End For

When the procedure for Algorithm 4.2 is finished, the optimal profit for the CISD problem is
achieved, which is 𝑈𝐵𝑏𝑒𝑠𝑡 , and the best sequence 𝑠 ∗ is recorded. The details of arrival times
and departure times can also be checked in the results of the 𝐷𝑃 for the sequence 𝑠 ∗ .

4.5 COMPUTATIONAL EXPERIMENT

In order to validate the effectiveness of the proposed model 𝐹1 and the efficiency of the
developed solution method, we conduct numerical experiments by using a PC (Intel Core i5,
2.1G Hz; Memory, 4G). The solution method is implemented by Matlab R2013b. The integer
linear program 𝑀1 is solved by CPLEX12.1 with concert technology of C# (VS2008).

Chapter 4: Cruise Itinerary Schedule Design 88

 Before conducting experiments, we need to notice that the objective of our model does not
represent the final profit for a cruise. To calculate the final profit, the incomes (e.g., tickets for
cruise passengers) and the costs (e.g., operating cost of the cruise ship) should be included. To
combine these incomes and costs together, a margin for per cruise passenger per day is
assumed, denoted by 𝑝𝑚 . In practice, this margin can be easily achieved by cruise companies
to analyze the previous profit reports of cruises. Here, we take US$100 as the value of 𝑝𝑚 .
The total number of cruise passengers and the cycle time for one cruise are denoted by 𝜑 and
𝜋 respectively. Thus, the final profit can be calculated by 𝑝𝑟𝑜𝑓𝑖𝑡 = 𝑝𝑚 × 𝜑 × 𝜋 + 𝑍, in which
𝑍 is the optimal result obtained by the proposed method. In the following experiments, the
final profit will be used as the optimal profit to be displayed in tables. Note that the newly
added term 𝑝𝑚 × 𝜑 × 𝜋 does not affect the optimization of our model objective 𝑍 since we
assume the passenger demand 𝜑 is constant in response to different itineraries. We will discuss
in the conclusion a future research to capture that the itinerary will affect the demand.
For the bunker cost function of the cruise ship deployed on the itinerary, according to Du et
al (2011), the coefficients in bunker consumption function (i.e., 𝑘, 𝑘 ′ and 𝑠 in Eq. (4.2)) are
related to the size of cruise ships. Here, we take the Explorer of the Seas (a cruise ship that
belongs to the Royal Caribbean) as the example, which is a jumbo ship with 138,000
deadweight tons. We set the coefficients in Eq. (4.2) as 𝑘 = 698, 𝑘 ′ = 0.000865 and 𝑠 = 4.5.
The fuel price for the ship is assumed to be US$251.50 per metric ton, which is the price of
IFO 380 at the port of Singapore on 8 September 2015 according to Ship & Bunker (2015).

4.5.1 Estimation of utility distribution

When using the above-mentioned method to design itineraries, cruise companies may find
it difficult to evaluate the monetary value of the utilities in ports of call. To facilitate their
designing, we propose a rational idea to help them derive utility profit. This utility profit is
calculated based on analyzing the announced itineraries of their own or other cruise
companies. For example, if we are helping Carnival Cruise Line to design a new itinerary, we
could analyze the itineraries from the Royal Caribbean, which is the biggest competitor for
Carnival Cruise Line, in order to obtain the adopted utility profit.
Our idea to derive the utility profit of ports of call comes from the observation of the
different arrival times at these ports in announced itineraries. Take the “11 Night Middle East
& Asia Cruise” operated by the Royal Caribbean International as an example. In Day 5 of the
itinerary, the cruise ship arrives at the port of Mormugao, India at 6:00 am. In Day 10 of the
itinerary, the ship arrives at the port of Penang, Malaysia at 12:00 pm. Here, we have the
question: why does not the ship arrive at these ports one hour earlier or one hour later? For the
port of Mormugao, the ship cannot arrive at 5:00 am because the port closes for service at that
time, but it is possible to arrive at 7:00 am. However, the cruise ship does not postpone the

Chapter 4: Cruise Itinerary Schedule Design 89

arrival time for one hour even if this means a saving of US$780.65 in the bunker cost, which
is calculated by the bunker cost function. One reasonable explanation for this is that the cruise
ship could earn more than US$780.65 from the utility in the time period from 6:00 am to 7:00
am at the port of Mormugao. For the port of Penang, the cruise ship could arrive at 11:00 am
(or 1:00 pm) if possible, which increases (or decreases) the bunker cost by US$855.02 (or
US$784.45). This means that the utility profit of the port of Penang from 11:00 am to12:00
pm is less than US$855.02 (or from 12:00 pm to 1:00 pm is more than US$784.45).
According to the above analysis, the main idea for the method is: we derive the utility profit
based on the announced itineraries operated by the other cruise companies. This method for
estimating the utility profit might not be the best, but a rational alternative for the estimation
of utility distribution. In Section 4.7, we will also propose a potential marketing method to
estimate the utility.

4.5.2 Impact of different units of time period

In Section 4.3.1, we mentioned that we use one hour as a time period when implementing
the CISD and the method can also handle time periods, e.g., half an hour. Here, we conduct
some experiments on two types of settings of the time unit, including one hour and half an
hour setting. The input data for testing the settings are randomly generated. Based on the
above-mentioned method of the estimation, we randomly generate the utility distribution under
the principles that bigger ports have higher utilities than small ports and noon hours have
higher utilities than other opening hours. The results of the experiments are shown in Table
4.1.

Table 4.1: Comparison between different settings of the time unit

Instance One hour Half an hour Comparison

# of 𝑇ℎ
Cycle Instance (𝑍ℎ −
ports of 𝑍𝑜 𝑇𝑜 𝑍ℎ 𝑇ℎ
time id 𝑍𝑜 ) / 𝑍𝑜 𝑇𝑜
call
6 3_6 1.237 2 1.240 5 0.25% 2.50
7 3_7 2.115 4 2.122 11 0.33% 2.75
3
8 3_8 3.196 5 3.208 16 0.38% 3.20
9 3_9 3.690 8 3.710 26 0.54% 3.25
8 5_8 1.545 6 1.553 29 0.51% 4.83
9 5_9 2.440 10 2.447 41 0.30% 4.10
5
10 5_10 3.374 13 3.392 70 0.53% 5.38
11 5_11 4.070 15 4.079 90 0.22% 6.00
10 7_10 1.934 16 1.939 113 0.23% 7.06
11 7_11 2.646 22 2.653 173 0.24% 7.86
7
12 7_12 3.573 53 3.583 434 0.27% 8.19
13 7_13 4.404 85 4.426 732 0.49% 8.61
Average: 0.36% 5.31
Note: (i) “# of ports of call” column denotes the total number of ports of call involved in one cruise.
This does not include two home ports. (ii) “Cycle time” column indicates the total days for one cruise.
(iii) “ 𝑍𝑜 ” and “ 𝑍ℎ ” columns list the optimal profits in the two settings of time unit with the unit of one
million US dollars. (iv) “ 𝑇𝑜 ” and “ 𝑇ℎ ” columns show the CPU time (seconds) to solve the problem.

Chapter 4: Cruise Itinerary Schedule Design 90

 From Table 4.1, we observe that half an hour setting brings more profit than that of the one-
hour setting. On average, the former increases the profit by 0.36%. However, the smaller time
unit causes trouble in computational time. In the half an hour setting, the time to find the
optimal results is longer than that of the one-hour setting. The average ratio between the two
settings is 5.31. Moreover, the ratio keeps increasing when the scale of instances becomes
larger. For the cruise companies, half an hour setting can still be used to improve the total
profit, as the problem is a strategic decision problem. In the following experiments, in order
to save the CPU time, we will use the one-hour setting rather than half an hour setting to solve
the CISD problem.

4.5.3 Performance of the enumeration improving method

We have proposed two exact solution methods. One enumerates all sequences of ports of
call and applies DP to calculate all these sequences in order to find the optimal one; while the
other one considers the enumeration improving (i.e., Algorithm 4.2). In order to test the
efficiency of the method with the enumeration improving, we conduct experiments under
different instance-scales by using the two methods. The comparisons are listed in Table 4.2.

Table 4.2: Computational efficiency with and without using enumeration improving

Enumeration No enumeration
Instance Time ratio
improving method improving
# of 𝑇1
Cycle Instance
ports of 𝑍0 𝑇0 𝑍1 𝑇1
time id 𝑇0
call
7 3_7 2.115 4 2.115 5 1.25
3 8 3_8 3.196 5 3.196 8 1.60
9 3_9 3.690 8 3.690 20 2.50
9 5_9 2.440 10 2.440 14 1.40
5 10 5_10 3.374 13 3.374 25 1.92
11 5_11 4.096 15 4.096 43 2.87
11 7_11 2.646 22 2.646 62 2.82
7 12 7_12 3.573 53 3.573 198 3.74
13 7_13 4.404 85 4.404 552 6.49
13 9_13 3.803 236 3.803 2035 8.62
9 14 9_14 4.697 374 4.697 3961 10.59
15 9_15 5.413 642 5.413 9753 15.19
Average: 4.92
Note: (i) “ 𝑍0” and “ 𝑍1 ” columns list the optimal profits obtained by two methods with the unit of one
million US dollars. (ii) “ 𝑇0 ” and “ 𝑇1 ” columns show the CPU time (seconds) to solve the problem.
As can be seen in Table 4.2, both two methods obtain optimal results. This is verified by
the same optimal profits shown in the column “ 𝑍0 ” and the column “ 𝑍1 ”. However, the
method without using the enumeration improving is extremely time-consuming. The average
ratio of CPU time between this enumeration method and the method with the enumeration
improving is on average 4.92, which implies the proposed method saves approximately 80%
of the computational time on average. More importantly, the time ratio between two methods

Chapter 4: Cruise Itinerary Schedule Design 91

increases dramatically with the growth of the instance-scale, which is shown in the last column
of the table. This demonstrates that the enumeration improved method is quite efficient to find
the optimal solutions for the problem.

4.5.4 Sensitivity analysis for a real case with a natural geographical sequence
As we have already tested the efficiency of the proposed method, we will use this method
to conduct sensitivity analysis in terms of three inputs, which are fuel price for the ship,
minimal staying hours of ports of call and opening hours of ports of call. Here, we take a
popular cruise line from the Royal Caribbean, “14 Night Singapore to Fremantle Cruise”, as
an example, and suppose that we are helping the Carnival Cruise Line to design a similar one-
way cruise. The nine ports involved in this cruise are shown in Figure 4.5, and it can easily be
seen that this itinerary has a naturally geographical sequence, under which we may design a
good visiting sequence by direct observation.

Figure 4.5: City ports in “14 Night Singapore to Fremantle Cruise” (Google Map, 2018a)

Table 4.3: Information of selected ports in a real case

Port Latitude Longitude Minimal

Port Country
index (N+,S–) (W–,E+) staying hours
1 Singapore Singapore 1.35 103.82 Home Port
2 Phuket Thailand 7.88 98.39 9
3 Langkawi Malaysia 6.35 99.80 6
4 Kuala Lumpur Malaysia 3.14 101.69 6
5 Geraldton Australia –28.77 114.61 6
6 Bali Indonesia –8.41 115.19 9
7 Lombok Indonesia –8.65 116.32 7
8 Broome Australia –17.95 122.24 6
9 Fremantle Australia –31.95 115.86 Home Port

For this cruise, assume that we deploy the Carnival Vista, which is the biggest cruise ship
of the Carnival Cruise Line. It has 133,500 deadweight tons. Same as the “14 Night Singapore
to Fremantle Cruise” operated by the Royal Caribbean International, in this cruise, the cruise
ship departs from the port of Singapore at 5:00 pm on Day 1 and arrives at the port of
Fermantle, Australia at 7:00 am on Day 15. Singapore and Fermantle are the home ports. The

Chapter 4: Cruise Itinerary Schedule Design 92

information about the seven selected ports of call (i.e., Phuket, Langkawi, Kuala Lumpur,
Geraldton, Bali, Lombok, and Broome) and the two home ports are shown in Table 4.3. We
assume that each port opens for the cruise ship at 7:00 am and closes at 7:00 pm on a daily
basis. The fuel price for the cruise ship has been mentioned in Section 4.6.1, which is
US$251.50 per metric ton.
The above-mentioned information is deemed as the baseline setting for the following
sensitivity analysis. The sensitivity analysis of the fuel price is implemented at first, which is
shown in Table 4.4. As it is shown in the table, there are two port sequences for different fuel
prices. For the optimal profit, we notice that with the rising of the fuel price, the profit
decreases evidently. If the fuel price increases by 20%, the profit drops by 0.81%. The number
of voyage hours also changes in different settings of the fuel price. It keeps rising and stays
unchanged for 268 hours when the fuel price increases by more than 40% compared with the
baseline setting. The increasing tendency for the voyage hours in response to the rising fuel
price is reasonable. This is because that lower fuel price may induce the cruise ship to sail
faster in order to gain more utility profit from the ports of call, and the higher fuel price may
impede the cruise ship to sail faster, because the bunker consumption increases significantly
when speeding up, meaning that the extra bunker cost is higher at a higher fuel price.

Table 4.4: Sensitivity analysis on fuel price

Solved optimal Optimal Voyage

Differentiation Deviation
port sequence profit hours
0.40 [1 4 3 2 6 7 8 5 9] 5.969 2.52% 240
0.60 [1 4 3 2 6 7 8 5 9] 5.918 1.64% 263
0.80 [1 4 3 2 6 7 8 5 9] 5.870 0.82% 264
Baseline setting [1 4 3 2 6 7 8 5 9] 5.822 0.00% 265
1.20 [1 4 3 2 6 7 8 5 9] 5.775 –0.81% 267
1.40 [1 2 3 4 6 7 8 5 9] 5.728 –1.62% 268
1.60 [1 2 3 4 6 7 8 5 9] 5.681 –2.43% 268
Note: (i) the coefficients in “Differentiation” column (i.e., {0.40, 0.60, 0.80, 1.20, 1.40, 1.60}) are used
to multiply the baseline setting of the fuel price (i.e., US$251.50 per metric ton) to represent the price
in each instance, e.g., 0.40 × 251.50 = 100.60. (ii) “Deviation” column lists the gap between the
current setting and the baseline setting with respect to the optimal profit. (iii) “Optimal profit” column
list the optimal profits obtained under different settings with the unit of one million US dollars.
Table 4.5: Sensitivity analysis on minimum staying hours

Solved optimal Optimal Staying

Differentiation Deviation
port sequence profit hours
–3 [1 2 3 4 6 7 8 5 9] 5.823 0.02% 55
–2 [1 2 3 4 6 7 8 5 9] 5.823 0.02% 55
–1 [1 2 3 4 6 7 8 5 9] 5.823 0.01% 57
Baseline setting [1 4 3 2 6 7 8 5 9] 5.822 0.00% 60
+1 [1 4 3 2 6 7 8 5 9] 5.821 –0.02% 61
+2 [1 4 3 2 6 7 8 5 9] 5.819 –0.06% 64
+3 [1 4 3 2 6 7 8 5 9] 5.814 –0.14% 70

Chapter 4: Cruise Itinerary Schedule Design 93

Note: (i) the coefficients in “Differentiation” column (i.e., {−3, −2, −1, +1, +2, +3}) are used to add
the baseline setting of the minimal staying hours of each port call (i.e., the last column in Table 4.4 to
obtain the minimal staying hours in each instance, e.g., for Port 2: −3 + 9 = 6.
Table 4.5 shows the results of sensitivity analysis on minimal staying hours for the cruise
ship dwelling in all ports of call. With the decreasing of minimal staying hours, the optimal
profit increases steadily until the minimal staying hours of all ports of call decrease by two
hours compared with the baseline setting. This is the point when the minimal staying hours
lose its effect as the restriction. It is also observed from the table that the cruise ship tends to
reduce its total staying hours in all ports of call with the minimal staying hours decreasing.
However, it stands at 55 hours even when the minimal staying hours decrease further, which
means the cruise ship still needs to stay in the ports of call for enough hours in order to obtain
decent profits from the utility of these ports.
The results of sensitivity analysis on the opening hours of the ports of call are given in Table
4.6. In the baseline setting, all ports of call open to the cruise ship for twelve hours daily, which
starts at 7:00 am and ends at 7:00 pm. Here, in the table, we change the daily opening hours
for all ports of call by one hour each time. As can be seen from the table, the cruise ship tends
to earn more profit as the ports open for a longer time. However, the effect on the profit
increasing from the increase of opening hours is not obvious, which is generally less than
0.05% per hour in the instances shown in the table. Therefore, it is highly recommended that
cruise companies need to be considerate when they want to increase the profit by requiring
more opening hours from the ports of call. In the table, the total staying hours also increase
steadily with more opening hours offered by the ports of call, which is the direct reason for the
increase of the optimal profit as more utility profit is earned from the ports of call.

Table 4.6: Sensitivity analysis on opening hours

Solved optimal Optimal Staying

Differentiation Deviation
port sequence profit hours
–3 [1 4 3 2 6 7 8 5 9] 5.817 –0.09% 55
–2 [1 4 3 2 6 7 8 5 9] 5.820 –0.04% 56
–1 [1 4 3 2 6 7 8 5 9] 5.820 –0.04% 58
Baseline setting [1 4 3 2 6 7 8 5 9] 5.822 0.00% 60
+1 [1 4 3 2 6 7 8 5 9] 5.822 0.00% 60
+2 [1 2 3 4 6 7 8 5 9] 5.823 0.02% 61
+3 [1 2 3 4 6 7 8 5 9] 5.823 0.02% 66
Note: (i) The coefficients in “Differentiation” column (i.e., {−3, −2, −1, +1, +2, +3}) are used to add
the baseline setting of the opening hours of each port call (i.e., twelve hours) to achieve the opening
hours in each instance, e.g., −3 + 12 = 9.
In sum, if the ports on an itinerary have a naturally geographical sequence, then port distance
will dominate the design of the itinerary and other parameters have a marginal effect on the
results. In the next section, we will examine cases in which the ports do not have a naturally
geographical sequence.

Chapter 4: Cruise Itinerary Schedule Design 94

4.5.5 Further analysis for cases without a natural geographical sequence
In some cruise areas with scattered ports and islands, such as the Caribbean Sea area (see
Figure 4.6), the ports of call in some cruise itineraries may not have a natural geographical
sequence, and there are many potentially good sequences given a set of ports of call. In this
section, we examine the value of sophisticated models for the itinerary design in such cruise
areas, in which a geographical sequence may not be easy to derive by direct observation.

Figure 4.6: City ports around the Caribbean Sea (Google Map, 2018b)

As the bunker cost is one of the major concerns for cruise companies, we would like to
further conduct analysis on the fuel price to see how important an optimal port sequence is
needed when the fuel price fluctuates. In order to conduct such analysis, we select 16 real
cruise services operated by Royal Caribbean International in the Caribbean Sea area (Royal
Caribbean International, 2016). Before conducting the analysis, we first redesign the itinerary
schedules for the 16 real cruise service. The comparisons in terms of port sequence and port
staying hours between the 16 designed itineraries and the 16 actual itineraries are listed in
Figure 4.7. Here, the y-axis shows the total port staying hours for each itinerary, and the x-
axis indicates the index of each cruise service. Symbol “Y” (Symbol “N”) above pairs of bars
indicates that the designed itinerary and the actual itinerary have the same port sequence
(different port sequences) for the corresponding cruise service.
For further analysis of the fuel price, based on the above-mentioned 16 cruise services, the
optimal sequence under the baseline setting of the fuel price is obtained by using the proposed
method at first. Then, the profits under other fuel price settings of this optimal sequence (i.e.,
Baseline itinerary) are calculated, which are defined as Baseline itinerary profit. The optimal
sequences and the optimal profits under other fuel price settings are also re-optimized by using
the proposed method, which is compared with the Baseline itinerary and Baseline itinerary
profit. The results of the comparisons are reported in Table 4.7.

Chapter 4: Cruise Itinerary Schedule Design 95

 Figure 4.7: Comparisons between the designed and actual itineraries

From Table 4.7, we can observe that when the fuel price fluctuates, the optimal sequences
for the itineraries are highly likely to change. Especially for large degrees of fluctuation, such
as the 0.4-differentiation and the 1.6-differentiation on the fuel price, the majority of optimal
sequences in the 16 cruise services are different from the optimal sequences obtained under
the baseline setting of the fuel price. However, when the fuel price fluctuates, if the optimal
sequence obtained in the baseline setting is fixed for the itinerary, there is significant profit
loss based on the average deviation in the table. In extreme cases, for instance, with 0.4-
differentiation on the fuel price, the average profit loss is 4.52%. Therefore, it is critical for
the cruise companies to re-optimize the visiting sequence when the fuel price fluctuates
significantly in order to achieve a higher profit.

Table 4.7: Further analysis of fuel price

Number of instances
having the same Average
Differentiation on the Average Baseline Average
optimal port optimal
fuel price itinerary profit deviation
sequence as the profit
baseline setting
0.4 3 5.022 4.805 4.52%
0.6 6 4.894 4.772 2.56%
0.8 8 4.779 4.721 1.23%
Baseline setting 16 4.663 4.663 0.00%
1.2 10 4.538 4.490 1.07%
1.4 6 4.424 4.317 2.48%
1.6 5 4.322 4.159 3.92%
Note: (i) “Average optimal profit” column list the average profit among the 16 cruise services under
each fuel price setting. (ii) “Average optimality gap” column shows the gap between the average
optimal profit and the average Baseline itinerary profit.

Based on the 16 selected cruise services from Royal Caribbean International, we also
conduct some sensitivity analysis on the utility distributions of the ports of call. For each port
of call in one instance, its utility distribution among hours in one day may have different values
of mean and standard deviation (SD). The SD indicates the degree of variation of the utility
among 24 hours. Here, we change the SD of the utility distribution for each port of call by

Chapter 4: Cruise Itinerary Schedule Design 96

multiplying a coefficient, which generates a new case of the instance for the sensitivity
analysis. Given the 16 cruise services, we test each instance in different SDs. The effects on
the optimal sequence and the optimal profit by different SDs (i.e., different variations of the
utility distribution) are shown in Table 4.8. As can be seen from Table 4.8, the differentiation
on the SD of the utility distribution has limited impacts on the optimal sequence and the
optimal profit. For the majority of the 16 cruise services, the optimal sequence under different
SDs is still the same as the optimal sequence under the baseline setting. Meanwhile, if we
insist on using the baseline optimal sequence under different SDs in large differentiations such
as 0.8-differentiation and 1.2 differentiation, the profit loss is around 0.5%.

Table 4.8: Sensitivity analysis on the utility distributions

Number of instances Average

Average
Differentiation on the having the same baseline Average
optimal
standard deviation optimal port sequence itinerary deviation
profit
as the baseline setting profit
0.8 12 4.551 4.526 0.55%
0.9 14 4.615 4.601 0.30%
Baseline setting 16 4.663 4.663 0.00%
1.1 14 4.675 4.667 0.17%
1.2 13 4.717 4.694 0.49%
Note: (i) the coefficients in “Differentiation” column (i.e., {0.80, 0.90, 1.10, 1.20}) are used to multiply
the baseline setting of the standard deviations (SDs) of the utility distributions of all ports of call to
represent the SDs in each instance.

4.5.6 Comparison between the proposed method and two heuristics

To test how important the proposed method is needed for the CISD, we compare our method
with two heuristics, which are the minimum fuel cost heuristic and the shortest path heuristic.
For both heuristics, we first find the sequence with the shortest voyage distance. Then, in the
minimum fuel cost heuristic, we design the itinerary schedule by minimizing the total fuel
consumption among all the voyage legs in the sequence without considering the port staying
profit. In the shortest path heuristic, assuming that we also consider the port staying profit, the
itinerary schedule is derived by using Algorithm 4.1 based on the sequence. The total profits
of the itinerary schedules obtained by the proposed method and the two heuristics under
different problem scales are compared in Table 4.9.
As can be seen in Table 4.9, the deviation in the sense of the average profit between the
proposed method and the minimum fuel cost heuristic is 7.54% on average. Such deviation
suggests that when designing the itinerary schedule, only focusing on the fuel cost
minimization would lead to significant profit loss, as the port staying profit is ignored in the
optimization. Although the port staying profit is considered in the shortest path heuristic, there
is still an average of 3.16% deviation to the optimal profit, which implies that the sequence
with the shortest voyage is not necessarily the optimal sequence, especially when the ports do
not have a naturally geographical sequence.

Chapter 4: Cruise Itinerary Schedule Design 97

 Table 4.9: Comparisons between the proposed method and two heuristics

The
Minimum fuel cost The shortest path
Instance group proposed
heuristic heuristic
method
# of
Average Average Average Average Average
ports Cycle time
profit profit deviation profit deviation
of call
9 2.352 2.225 5.40% 2.325 1.15%
5 10 3.249 3.044 6.31% 3.184 2.00%
11 3.894 3.551 8.81% 3.764 3.34%
11 2.501 2.352 5.96% 2.457 1.76%
7 12 3.297 3.084 6.46% 3.194 3.12%
13 4.235 3.814 9.94% 4.074 3.80%
13 3.712 3.478 6.30% 3.604 2.91%
9 14 4.494 4.121 8.30% 4.284 4.67%
15 5.108 4.578 10.38% 4.818 5.68%
Average: 7.54% 3.16%
Note: (i) In each instance group, there are 10 instances, which are randomly generated based on major
ports or islands in the Caribbean Sea area. (ii) The average profit of 10 instances in one group is
calculated for the proposed method and the two heuristics. (iii) “Average deviation” shows the
deviations between the average profit by the proposed method and the average profits by the heuristics.

4.5.7 Managerial Implications

The above numerical experiments shed lights on the nature of the CISD problem and enable
us to discover a number of useful managerial implications for cruise companies summarized
as follow.
First, when designing itineraries, using smaller time unit causes an increase in
computational time, but it could bring more profits. As the planning of cruise itineraries is a
strategic decision, it is recommended that cruise companies use smaller time unit in order to
obtain more profits. The resulting schedule could be an “internal” schedule and the published
schedule could be the one that rounds the internal schedule to one hour or half an hour. For
instance, if it is calculated that the ship should arrive at a port of call at 6:50 am, then the cruise
company can inform the captain to try to arrive at 6:50 am and can inform the cruise passengers
that the ship arrives at 7:00 am.
Second, when the ports of call on an itinerary have a naturally geographical sequence, then
the distances between ports dominate the design of the schedule. In other words, the sequence
of the ports of call with the shortest total distance is generally the optimal choice, and the
impacts of minimum staying hours at ports and the opening hours at ports are marginal. The
bunker fuel price has a larger effect on the total profit in the following way: when the fuel
price is higher, the voyage time is longer, leading to lower fuel consumption and thereby lower
fuel costs, and vice versa.
Third, when the ports of call on an itinerary do not have a naturally geographical sequence,
as is the case for the largest cruise destination—the Caribbean Sea area, there are many
potentially good schedules. We find that the fuel price has a much larger impact on the design

Chapter 4: Cruise Itinerary Schedule Design 98

of schedule. Specifically, when the fuel price deviates from the estimated price by 60%,
sticking to the optimal schedule based on the estimated price will lead to a profit reduction of
around 4% (cf. Table 4.7). Hence, it is highly desirable for a cruise company to have an
accurate estimation of the fuel price.

4.6 UTILITY ESTIMATION BY MARKETING TECHNIQUES

In this paper, a big challenge to implement the model in practice is the utility distribution
estimation. We have proposed a method for the estimation in Section 4.6.1. Such a method
might not be the best, but a rational alternative. In this section, we further construct a potential
utility estimation approach by using some marketing techniques. Here we design a conjoint
analysis (Green et al., 1996; Ding et al., 2009) for evaluating customers’ preference on cruise
itinerary schedules. A choice-based conjoint experiment (Toubia et al., 2007; Gilbride et al.,
2008) is conducted to obtain conjoint data. Then, the conjoint data is analyzed by a basic
multinomial logit model (Hongmin and Woonghee, 2011; Li, 2011; Paat and Huseyin, 2012),
and henceforth the utility distribution can be obtained. Note that this approach will not be
implemented in this paper, as it involves tremendous research efforts in collecting conjoint
data by interviewing many potential cruise passengers. However, this approach will be
explored and developed in our future study.

4.6.1 Choice data collection

Before illustrating our analysis procedure, we would introduce the background of the
analysis. Our conjoint experiment is conducted for the cruise itinerary schedules in a specific
region (e.g., Asia region) rather than the global. It is due to that: firstly, loop cruise itineraries
only traverse a set of ports in the same region, and secondly, the customers from different
regions have different preferences on cruise itinerary schedules. For instance, the China-Japan-
Korea cruise services are popular in China. In those services, cruise passengers in China would
leave from the home port in China (such as Shanghai and Tianjin), visit some ports of call in
Japan and Korea (such as Nagasaki (Japan), Fukuoka (Japan) and Busan (Korea)), and return
to the home port finally. In the background of China-Japan-Korea area, many cruise itinerary
schedules can be designed based on the cruise passengers’ preference in China.
The first step in the conjoint analysis is to define the attributes (or factors) of a cruise service
(or a cruise itinerary schedule) that have effects on customers’ preference on cruise itinerary
schedules. A straightforward attribute is the ticket price for the cruise service, which is an
explanatory variable on how the attribute motivates the customers to buy the service. Each
attribute has different levels, i.e., the possible values for the attributes. For example, for the
ticket price attribute, there are possible levels at $800, $850, $900 and so on. Xie et al. (2012)

Chapter 4: Cruise Itinerary Schedule Design 99

have proposed several attributes of a cruise ship that affect the customers’ preference. In our
analysis, we focus on some attributes related to the itinerary schedule design.
Table 4.10 lists the attributes and levels used in our analysis under the background of China-
Japan-Korea area. Apart from some regular variables (such as price, home port, and rotation
time), there are some important hourly dummy binary variables (i.e., time-of-day variables),
which are defined to show whether the cruise ship stays at a certain port of call during a certain
hourly time period (e.g., 6:00 – 7:00 am) or not. The purpose of defining the hourly dummy
binary variables is to estimate the utility distribution in each port of call (see Koppelman et al.
(2008) for the application of time-of-day variables). Here, we denote 𝐸 as the set of all the
regular variables, and 𝑅 as the set of all the hourly dummy binary variables. Assume that there
is a mock-up cruise schedule: the ticket price is $850, the rotation time is 6 days, the home
port is Shanghai, and it only has one port of call (Fukuoka (Japan), arrival time: 7:00 am,
departure time: 5:00 pm). Then based on Table 4.10, the cruise service can be depicted in:
“Price” with value 850; “Rotation time” with value 6; “Home port” with value 1; the binary
variables in “Staying hours in Fukuoka(Japan)” corresponding to the staying hours from 7:00
am to 5:00 pm with value 1, and all other variables with value 0.

Table 4.10: Attributes and levels used in the conjoint analysis

Attributes Levels

Price A continuous variable (such as $800, $850 and $900)

Rotation time A discrete variable (such as 5 days, 6 ports and 7 ports)

Homeport A discrete variable (1 for Shanghai, and 2 for Tianjin)

Staying hours at Nagasaki (Japan):
6:00 – 7:00 am A dummy binary variable for each hourly time period (for
7:00 – 8:00 am example, 6:00 – 7:00 am in Nagasaki (Japan): 1 for the cruise
8:00 – 9:00 am ship staying at the port in 6:00 – 7:00 am; otherwise 0)
and so on
Staying hours at Fukuoka (Japan):
6:00 – 7:00 am A dummy binary variable for each hourly time period (for
7:00 – 8:00 am example, 6:00 – 7:00 am in Fukuoka (Japan): 1 for the cruise
8:00 – 9:00 am ship staying at the port in 6:00 – 7:00 am; otherwise 0)
and so on
Staying hours at Busan (Korea):
6:00 – 7:00 am A dummy binary variable for each hourly time period (for
7:00 – 8:00 am example, 6:00 – 7:00 am in Busan (Korea): 1 for the cruise
8:00 – 9:00 am ship staying at the port in 6:00 – 7:00 am; otherwise 0)
and so on
Note: (i) In “Staying hours at Nagasaki (Japan)” attributes, each hourly time period (e.g., 6:00 – 7:00
am in Nagasaki (Japan)) has a corresponding attribute. (ii) For “Staying hours in Nagasaki (Japan)”
attribute, if the cruise ship for a cruise schedule arrives at Nagasaki(Japan) at 7:00 am and departs
from the port at 11:00 am. Then, this schedule has the dummy binary variables (corresponding to 7:00
– 8:00 am, 8:00 – 9:00 am, 9:00 – 10:00 am and 10:00 – 11:00 am) equalling to one, and all other
dummy binary variables equalling to zero.

Chapter 4: Cruise Itinerary Schedule Design 100

 As we have decomposed the cruise schedule into such attributes and levels in Table 4.10.
The next step is to generate some mock-up cruise schedules, which will be used in the
interview with potential cruise passengers. 𝕊 denotes the set for all generated cruise schedules.
Those generated mock-up cruise schedules will be clustered into some choice sets. The
schedule generation and cluster process can be done by Efficient Factorial Design in the
statistical package of SAS (Kuhfeld, 2010). Each choice set contains a certain number of
mock-up cruise schedules (denoted as 𝐽). Thereafter, the conjoint choice data is collected by
interviewing potential cruise passengers (i.e., respondents), and asking them to choose one
preferred cruise schedule (i.e., alternative) form each choice set. Here, we denote 𝑤 and 𝑞 as
the index for the choice set and respondent, respectively, where 𝑤 ∈ 𝕎 and 𝑞 ∈ ℚ. For
instance, we have generated 20 (i.e., |𝕊| = 20) unique mock-up cruise schedules, which are
clustered into 12 choice sets (i.e., |𝕎| = 12) with 3 alternatives in each set (𝐽 = 3). Note that
a mock-up cruise schedule can exist in several choice sets. Assuming that the choice set 1
contains the alternative 1, 2 and 4, each respondent 𝑞 will be asked to choose one of the three
alternatives for the choice set.

4.6.2 Choice data analysis

After collecting the conjoint choice data, the following step is to analyze the conjoint data.
The most popular model to analyze the choice data is the multinomial logit model (Vermeulen
et al., 2008). By using aggregate logit share techniques (Koppelman et al., 2008), we can derive
the utility experienced by the respondent 𝑞 when facing the 𝑗th alternative (𝑗 ∈ {1, . . , 𝐽}) in the
choice set 𝑤 as follows:
𝕌𝑞𝑤𝑗 = ∑𝑒∈𝐸 𝛼𝑒 𝑋𝑤𝑗𝑒 + ∑𝑟∈𝑅 𝛽𝑟 𝑌𝑤𝑗𝑟 + 𝜀𝑞𝑤𝑗 (4.36)
where 𝛼𝑒 represents the partial utility (i.e., the coefficient or “partworths”) for the 𝑒th regular
variable (e.g., the variable for the price), and 𝑋𝑤𝑠𝑒 is the value of the 𝑒th no time-of-day
variable for the 𝑗th cruise schedule in the choice set 𝑤 (e.g., $850 for the price). Similarly, 𝛽𝑟
and 𝑌𝑤𝑗𝑟 are corresponding to the 𝑟th hourly dummy binary variable. 𝜀𝑞𝑤𝑗 is the error term.
Note that 𝛽𝑟 , ∀𝑟 ∈ 𝑅 indicate the utility distribution for all the ports of call.
Then, all the error terms are assumed to be i.i.d., under which we can calculate the
probability that the cruise passenger 𝑞 will choose the 𝑗th alternative in the choice set 𝑤:
exp(∑𝑒∈𝐸 𝛼𝑒 𝑋𝑤𝑗𝑒 +∑𝑟∈𝑅 𝛽𝑟 𝑌𝑤𝑗𝑟 )
𝑃𝑞𝑤𝑗 = ∑ (4.37)
𝑗′ ∈{1,2,..,𝐽} exp(∑𝑒∈𝐸 𝛼𝑒 𝑋𝑤𝑗′ 𝑒 +∑𝑟∈𝑅 𝛽𝑟 𝑌𝑤𝑗′ 𝑟 )

Based on the probability function, we can derive its log-likelihood function:

ln(𝐹(𝛼, 𝛽)) = ∑𝑞∈ℚ ∑𝑤∈𝕎 ∑𝑗∈{1,2,..,𝐽} 𝑍𝑞𝑤𝑗 ln(𝑃𝑞𝑤𝑗 ) (4.38)
where 𝑍𝑞𝑤𝑗 shows the conjoint choice data collected in Section 4.7.1, which equals one if and
only if the cruise passenger 𝑞 choose the 𝑗th alternative in the choice set 𝑤. In order to estimate
𝛼̂𝑒 , ∀𝑒 ∈ 𝐸 and 𝛽̂𝑟 , ∀𝑟 ∈ 𝑅, we can maximize the above log-likelihood function by using the

Chapter 4: Cruise Itinerary Schedule Design 101

conjoint choice data as the input parameters. This procedure can also be done by SAS
(Kuhfeld, 2000), and some technical issues are discussed in Vermeulen et al. (2008). Till now,
the estimated coefficient 𝛽̂𝑟 , ∀𝑟 ∈ 𝑅 are obtained and indicate the utility distribution for each
hour in each port of call.
Based on the estimated coefficient 𝛼̂𝑒 , ∀𝑒 ∈ 𝐸 and 𝛽̂𝑟 , ∀𝑟 ∈ 𝑅 , we further estimate the
possible demand for a newly designed cruise schedule (denoted as 𝛾): firstly, we investigate
all the existing cruise schedule 𝑠 ∈ 𝕊 in the specific region (e.g., the China-Japan-Korea area)
as well as the total regional market share (denoted as 𝕄). The total regional market share can
be easily obtained from some industry reports, such as Statista (2015). Then, the utility is
calculated for each existing cruise schedule and the newly designed cruise schedule by: 𝕌𝑠 =
∑𝑒∈𝐸 𝛼̂𝑒 𝑋𝑠𝑒 + ∑𝑟∈𝑅 𝛽̂𝑟 𝑌𝑠𝑟 and 𝕌𝛾 = ∑𝑒∈𝐸 𝛼̂𝑒 𝑋𝛾𝑒 + ∑𝑟∈𝑅 𝛽̂𝑟 𝑌𝛾𝑟 ( 𝑋 and 𝑌 parameters here
have the same meanings with that in Eq.(21)). Thereafter, the probability that potential cruise
passengers in the regional market will choose the newly designed cruise schedule is 𝑃(𝛾) =
exp(𝕌𝛾 )
∑𝑠∈𝕊 exp(𝕌𝑠 )+exp(𝕌𝛾 )
. Next, the demand in the regional market for the newly designed cruise

schedule 𝛾 is estimated as: 𝕄 ∙ 𝑃(𝛾).

4.7 CONCLUSION

This paper addresses the cruise itinerary schedule design problem that determines the
optimal sequence of a given set of ports of call and the arrival and departure times at each port
to maximize the monetary value of the utility minus the fuel cost. In view of the practical
observations that most cruise itineraries do not have many ports of call, we first enumerate all
sequences of ports of call and then optimize the arrival and departure times at each port of call
by developing a dynamic programming approach. To improve the computational efficiency,
we propose effective bounds on the profit of each sequence of ports of call, eliminating non-
optimal sequences without invoking the dynamic programming algorithm. The computational
experiments show that, first, the proposed bounds on the profit of each sequence of ports of
call can considerably improve the computational efficiency; second, the total profit of the
cruise itinerary is sensitive to the fuel price and hence, it is acceptable to use the shortest
voyage distance method to design the schedule when the ports of call have a naturally
geographical distance; in contrast, determining the sequence of ports of call solely by
minimizing the overall voyage distance frequently leads to a significant reduction in the total
profit when the ports do not have a naturally geographical sequence.
Given that cruise itineraries have fixed sequences of ports of call and fixed schedules,
optimization-based itineraries planning tools should be able to increase the profit or save the
cost for cruise shipping companies and improve the service quality for cruise passengers.
Compared with other areas of transportation such as a truck, rail, and air (we note that

Chapter 4: Cruise Itinerary Schedule Design 102

transportation is not the purpose of cruising, but the cruise shipping is a part of transportation),
there are few quantitative studies on cruise shipping. Nevertheless, cruise shipping has its own
characteristics that need to be explored by industrial engineers/operations researchers.
Moreover, the cruise market has maintained steady growth over the past 20 years despite the
economic crisis in 2008 and cruising companies have ordered a number of large cruise ships
to serve the mass market of cruising. We believe that there is a broad range of research topics
in cruise shipping.
In this study, we have a limitation on assuming that the utility increase for the cruise
passengers is additive over the port staying hours. However, in reality, cruise passengers’
utility may depend on the time period that they want to stay in the ports, and the incremental
utility of an extra port staying hour may decrease over time. Under this circumstance, the
proposed solution can still be applied by analyzing the utility for each possible port staying
period (e.g., 7:00 am to 6:00 pm) rather than each port staying hour. Our future study will
further explore the relationship between the utility increase for cruise passengers and the port
staying hour increase. Another limitation is that we assume the itinerary design has no effect
on the passenger demand. However, if we consider the multinomial logit model in Section 4.6,
given any new schedule, the utility of this schedule will be different and thus the demand will
change accordingly. A future study should also embed the multinomial logit model to measure
the potential demand in response to different itinerary schedules. Meanwhile, we may also
determine the number of days while designing a cruise service rather than follow the fixed
time periods for planning. Henceforth, the pricing strategies should be considered, because the
number of days for a cruise service will affect the pricing, and thus further affect the demand.
For future research topics on the cruise industry, there are some recommendations. (i) The
cruise itinerary design topic: based on the limitation of this study, we do not consider some
practical constraints for the cruise ship to dwell in the ports of call, such as berth availability
and tide effects in ports. Therefore, a more general cruise itinerary design problem can be
studied. (ii) The cruise ship fleet deployment topic: cruise ships are frequently repositioned
from one region to another region. From the perspective of the cruise lines, there are several
decision problems on how cruise ships are repositioned. (iii) The cruise ticket pricing topic: in
airlines, many pricing policies, and strategies have been developed to increase revenue.
However, the pricing in cruise shipping is not so flexible than that of the airlines. Thus, future
studies can be conducted on cruise pricing by borrowing some ideas from the pricing of the
airlines.

4.8 REFERENCES

Applegate, D.L., Bixby, R.E., Chvatal, V., Cook, W.J., 2011. The Traveling Salesman
Problem: A Computational Study. Princeton University Press, New Jersey.

Chapter 4: Cruise Itinerary Schedule Design 103

Biehn, N., 2006. A cruise ship is not a floating hotel. Journal of Revenue and Pricing
Management 5, 135–142.

Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C., 2009. Introduction to Algorithms, Third
Edition. The MIT Press, London, England.

Cordeau, J. F., Laporte, G., Mercier, A., 2001. A unified tabu search heuristic for vehicle
routing problems with time windows, Journal of the Operational Research Society 52(8),
928–936.

Carnival Cruise Line, 2016. 7 day Western Caribbean from Miami, FL.
http://www.carnival.com/itinerary/7-day-western-caribbean-cruise/miami/glory/7-
days/wek. Accessed 4 April 2016.

Cruise Industry News, 2015. Cruise Industry Statistics—Cruise Ship Statistics.

http://www.cruiseindustrynews.com/cruise-industry-stats.html. Accessed 28 Aug 2015.

Cruise Ship Schedule, 2015. Carnival Vista itinerary 2016 cruises.

http://www.cruiseshipschedule.com/carnival-cruise-lines/carnival-vista-schedule/.
Accessed 12 September 2015.

Ding, M., Park, Y.H., Bradlow, E.T., 2009. Barter markets for conjoint analysis, Management
Science 55(6), 1003–1017.

Du, Y., Chen, Q., Quan, X., Long, L., Fung, R.Y.K., 2011. Berth allocation considering fuel
consumption and vessel emissions, Transportation Research Part E 47, 1021–1037.

Du, Y., Chen, Q., Lam, J. S. L., Xu, Y., Cao, J. X., 2015. Modeling the impacts of tides and
the virtual arrival policy in berth allocation, Transportation Science 49(4), 939-956.

Feillet, D., Dejax, P., Gendreau, M., 2005. Traveling salesman problems with profits,
Transportation Science 39(2), 188–205.

Fransoo, J.C., Lee, C.Y., 2013. The critical role of ocean container transport in global supply
chain performance, Production and Operations Management 22(2), 253–268.

Gilbride, T.J., Lenk, P.J., Brazell, J.D., 2008. Market share constraints and the loss function
in choice-based conjoint analysis, Marketing Science 27(6), 995–1011.

Google Map, 2018a. https://www.google.com/maps/@-9.689449,127.0193577,5.17z.

Accessed 4 April 2018.

Google Map, 2018b. https://www.google.com/maps/@15.264912,-75.719404,6.17z.

Accessed 4 April 2018.

Green, P.E., Krieger, A.M., 1996. Individualized hybrid models for conjoint analysis,
Management Science 42(6), 850–867.

Chapter 4: Cruise Itinerary Schedule Design 104

Gui, L., Russo, A.P., 2011. Cruise ports: a strategic nexus between regions and global lines—
evidence from the Mediterranean, Maritime Policy & Management 38(2), 129–150.

Hongmin, L., Woonghee, T.H., 2011. Pricing multiple products with the multinomial logit and
nested logit models: concavity and implications, Manufacturing & Service Operations
Management 13(4), 549–563.

Kimes, S.E., 1989. Yield management: a tool for capacity-considered service firms, Journal
of Operations Management 8, 348–363.

Koppelman, F.S., Goldren, G.M., Parker, R.A., 2008. Schedule delay impacts on air-travel
itinerary demand, Transportation Research Part B: Methodological 42, 263–273.

Kuhfeld, W., 2000. An Introduction to Designing Choice Experiments, and Collecting,

Processing, and Analyzing Choice Data with the SAS System. Cary: SAS Institute Inc..

Kuhfeld, W., 2010. Marketing research methods in SAS. Cary: SAS Institute Inc..

Li, B., 2011. The multinomial logit model revisited: A semi-parametric approach in discrete
choice analysis, Transportation Research Part B: Methodological 45(3), 461–473.

Maddah, B., Moussawi-Haidar, L., El-Taha, M., Rida, H., 2010. Dynamic cruise ship revenue
management, European Journal of Operational Research 207(1), 445–455.

Meng, Q., Wang, S., Andersson, H., Thun, K., 2014. Containership routing and scheduling in
liner shipping: overview and future research directions, Transportation Science 48 (2),
265–280.

Ng, M.W., 2014. Distribution-free vessel deployment for liner shipping, European Journal of
Operational Research 238(3), 858–862.

Ng, M.W., 2015. Container vessel fleet deployment for liner shipping with stochastic
dependencies in shipping demand, Transportation Research Part B 74, 79–87.

Paat, R., Huseyin, T., 2012. Robust assortment optimization in revenue management under the
multinomial logit choice model, Operation Research 60(4), 865–882.

Pang, K.W., Liu, J., 2014. An integrated model for ship routing with transshipment and berth
allocation, IIE Transactions 46(12), 1357–1370.

Qi, X.T., Song, D.P., 2012. Minimizing fuel emissions by optimizing vessel schedules in liner
shipping with uncertain port times, Transportation Research Part E 48(4), 863–880.

Rodrigue, J., Notteboom, T., 2013. The geography of cruises: Itineraries, not destinations,
Applied Geography 38, 31–42.

Royal Caribbean International, 2016. Cruise itineraries sailing to the Caribbean area.
http://www.royalcaribbean.com/cruises?destinationRegionCode_CARIB=true.
Accessed 22 Aug 2016.

Chapter 4: Cruise Itinerary Schedule Design 105

Shintani, K., Imai, A., Nishimura, E., Papadimitriou, S., 2007. The container shipping network
design problem with empty container repositioning, Transportation Research Part E
43(1), 39–59.

Ship & Bunker, 2015. Bunker prices. http://shipandbunker.com/prices#IFO380. Accessed 8

September 2015.

Song, D.P. and Dong, J.X. 2011. Effectiveness of an empty container repositioning policy with
flexible destination ports, Transport Policy 18 (1), 92–101.

Song, D.P. and Dong, J.X. 2012. Cargo routing and empty container repositioning in multiple
shipping service routes, Transportation Research Part B 46(10), 1556–1575.

Song, D.P. and Dong, J.X. 2013. Long-haul liner service route design with ship deployment
and empty container repositioning, Transportation Research Part B 55, 188–211.

Soriani, S., Bertazzon, S., Cesare, F.D., Rech, G., 2009. Cruising in the Mediterranean:
structural aspects and evolutionary trends, Maritime Policy & Management 36(3), 235–
251.

Statista, 2015. Cruise industry—Statista Dossier: Facts and statistics on the cruise industry.
http://www.statista.com/study/11547/cruise-line-industry-statista-dossier/. Accessed 28
Aug 2015.

Talluri, K., van Ryzin, J., 2004. The Theory and Practice of Revenue Management, Kluwer
Academic, Boston.

Toubia, O., Hauser, J., Garcia, R., 2007. Probabilistic polyhedral methods for adaptive choice-
based conjoint analysis: theory and application, Marketing Science 26(5), 596–610.

Toth, P., Vigo, D., 2001. The Vehicle Routing Problem. Society for Industrial and Applied
Mathematics, Philadelphia.

Vermeulen, B., Goos, P., & Vandebroek, M., 2008. Models and optimal designs for conjoint
choice experiments including a no-choice option. International Journal of Research in
Marketing 25(2), 94–103.

Veronneau, S., Roy, J., 2009. Global service supply chains: An empirical study of current
practices and challenges of a cruise line corporation, Tourism Management 30, 128–139.

Veronneau, S., Roy, J., Beaulieu, M., 2015. Cruise ship suppliers: A field study of the supplier
relationship characteristics in a service supply chain, Tourism Management Perspectives
16, 76–84.

Wang, K., Wang, S., Zhen, L., Qu, X., 2016. Cruise shipping review: Operations planning and
research planning. Maritime Business Review 1(2), 133-148.

Chapter 4: Cruise Itinerary Schedule Design 106

Wang, S., Meng, Q., 2012a. Robust schedule design for liner shipping services, Transportation
Research Part E 48(6), 1093–1106

Wang, S., Meng, Q., 2012b. Liner ship route schedule design with sea contingency time and
port time uncertainty, Transportation Research Part B 46(5), 615–633.

Wang. S., Meng. Q., Liu. Z., 2013. Bunker consumption optimization methods in shipping: A
critical review and extensions, Transportation Research Part E: Logistics and
Transportation Review 53, 49–62.

Wang. S., Wang. H., Meng. Q., 2015. Itinerary provision and pricing in container liner
shipping revenue management, Transportation Research Part E: Logistics and
Transportation Review 77, 135–146.

Xie, H., Kerstetter, D.L., Mattila, A.S., 2012. The attributes of a cruise ship that influence the
decision making of cruisers and potential cruisers, International Journal of Hospitality
Management 31, 152–159.

Chapter 4: Cruise Itinerary Schedule Design 107

Chapter 5: Cruise Ship Service Planning

This chapter addresses a decision problem on planning cruise services for a cruise ship
so as to maximize the total profit during a planning horizon. The service is a sequence of ports
(harbor cities) the cruise ship visits. In this decision problem, the constraint about the
availability of berths at each port is taken into account. In reality, if a cruise service is executed
by the ship repeatedly for several times, the profit earned by the cruise service in each time
decreases gradually. This effect of decreasing marginal profit is also considered in this study.
We propose a nonlinear integer programming model to cater to the concavity of the function
for the profit of operating a cruise service repeatedly. To solve the nonlinear model, two
linearization methods are developed, one of which takes advantage of the concavity for a
tailored linearization. Some properties of the problem are also investigated and proved by
using dynamic programming (DP) and two commonly used heuristics. In particular, we prove
that if there is only one candidate cruise service, a greedy algorithm can derive the optimal
solution. Numerical experiments are conducted to validate the effectiveness of the proposed
models and the efficiency of the proposed linearization methods. In case some parameters
needed by the model are estimated inexactly, the proposed decision model demonstrates its
robustness and can still obtain a near-optimal plan, which is verified by experiments based on
extensive real cases.

5.1 INTRODUCTION

In the cruising industry, service planning is often independent among different cruise ships,
and planning problems of different cruise ships can be solved individually. According to
Rodrigue and Notteboom (2013), the cruise ship deployment focuses on a specific cruise ship
rather than a fleet of cruise ships. Cruise ships are often unique, even if the cruise ships have
the same capacity (in passengers), different cruise ships have significantly different onboard
activities, which are part of cruising experience. The same cruise route traversed by two cruise
ships constructs two different cruise services, as the onboard activities are different. Our
research is applicable to this situation. If two ships are very similar and serve the same region
and visit common ports of call, then the competition between the services provided by the
ships have to be accounted for.
This study assumes that when a cruise ship is repositioned to a new region, a home port and
a set of candidate cruise services are chosen in advance and addresses the Cruise Service
Planning (CSP) problem. This problem aims to determine how to plan cruise services for the
cruise ship to operate in the region for a period of time. In other words, over the period of time,

Chapter 5: Cruise Ship Service Planning 109

how to choose cruise services from the candidate cruise services for the cruise ship to operate
in order to maximize total profit. However, the solution to the problem is not as straightforward
as it seems. In a cruise service, there are several ports of call for the cruise ship to visit.
Therefore, the berth availability of the ports of call should be considered when operating a
cruise service. For instance, Wusong Kou terminal is a cruise terminal in Shanghai (China)
with two available berths. Based on the arrival schedule of the terminal for the incoming cruise
ships in the Year 2016, on a specific day, there might be two cruise ships scheduled to moor
at the terminal. If the cruise service operated by the cruise ship also arrives at the terminal on
that day on schedule, the cruise service is unable to be operated due to the lack of berth.
Determining whether to choose a cruise service for the cruise ship is based on the scheduled
rotation time of the service and the marginal profit of operating the service (i.e., the operating
profit). Empirically, operating a cruise service with a high daily operating profit (i.e., the
operating profit divided by the service’s rotation time) is more likely to cater to the preference
of the cruise ship. Whereas, a preferable cruise service might not be always profitable in a
planning horizon: the marginal profit of operating a cruise service is not constant in the real
situation. If a cruise service is repeated several times, its marginal profit decreases gradually.
This phenomenon attributes to that few potential cruise passengers would order a cruise
service if the cruise service has been repeated many times. Therefore, the effect of the
decreasing marginal profit should also be considered in the CSP problem.
Based on the above analysis, this paper presents an explorative study on the CSP problem
considering berth availability and decreasing marginal profit, in which optimal services for a
cruise ship to operate are to be determined. In our study, we first build an integer linear model
assuming that the marginal profit of operating a cruise service is constant. Then, an integer
nonlinear model is formulated as a general problem, and two methods are proposed to linearize
the model. One of the two methods takes advantage of the concavity for a tailored linearization,
and the model linearized by this method is more efficient to be solved based on computational
results. Some properties of the problem are also investigated and proved. By using DP, we
investigate the NP-hardness of the problem under different cases. By analyzing some
commonly used heuristics, we prove some useful theorems, for example, we prove that if there
is only one candidate cruise service, a greedy algorithm can derive the optimal solution. The
effectiveness of the proposed models is verified by extensive numerical experiments. Lastly,
based on extensive real cases, robustness tests are conducted to show that in case some
parameters needed by the model are estimated inexactly, the proposed decision model has its
robustness and can still obtain a near-optimal plan.
The remainder of this chapter is organized as follows. Section 5.2 reviews the related works.
Section 5.3 presents a brief problem description and proposes mathematical models.
Complexity analysis and extensive comparison with heuristics are conducted in Section 5.4.

Chapter 5: Cruise Ship Service Planning 110

The results of numerical experiments are reported in Section 5.5. Conclusions are then outlined
in the last section.

5.2 LITERATURE REVIEW

The cruise shipping related studies could belong to the area of tourism research as cruise
ships provide cruise passengers with tourism service. Meanwhile, it can be also sorted into the
area of maritime research as the cruise services are akin to container liner services. However,
the past research on cruise shipping is limited, the reason of which may include: (i) the
worldwide cruise ship tourism just accounts for about 2% of the world tourism market revenue,
thus the tourism-related researchers have not paid much attention to the cruise shipping related
studies (Gui and Russo, 2011); (ii) the maritime logistic related researchers mainly focus on
freight transportation (e.g., Bell et al., 2011; Meng and Wang, 2012; Song et al., 2015).
The majority of past research on the cruise shipping analyzed the cruising industry as a
tourism service supply chain. Gui and Russo (2011) constructed an analytic framework
connecting the global structure of cruise value chains to the regional land-based cruise
services. The demand side and the supply side of the cruise shipping at the worldwide level
were analyzed by Soriani et al. (2009). They also investigated the main characteristics of the
cruising in the Mediterranean region and examined the main cruising ports in the region. A
field study of a large Florida-based global cruise company’s practices in re-supplying ships
globally was conducted by Veronneau and Roy (2009), which makes them amongst the first
to take a comprehensive study of a specific service supply chain. Veronneau et al. (2015)
investigated the relationships between a major cruise line corporation and its suppliers by a
field study. Rodrigue and Notteboom (2013) focused on capacity deployment and itineraries
in two important markets: the Caribbean and Mediterranean. They found that these two market
areas interact with each other due to seasonal variations in demand.
Although those researchers devoted significant efforts into cruising shipping research, their
works are mainly descriptive and belong to empirical studies. The majority of existing related
works did not provide the cruise industry with quantitative analysis on cruise shipping, which
could be critical for some detailed problems. Maddah et al. (2010) conducted a quantitative
study on cruise shipping, in which a discrete-time dynamic capacity control model was built
to improve the profit of cruise ships. Their model could give cruise ship managers some
suggestions about which requests from customers should be accepted based on the remaining
cabin and lifeboat capacities and the type of requests. The research is adaptable for cruise
companies to improve profit, and it focuses on an operational level problem. As the cruising
industry has developed dramatically, more research efforts should be made on the problems in
strategic level or tactical level.

Chapter 5: Cruise Ship Service Planning 111

 The cruise shipping and the container liner shipping have something in common in the sense
of research. They both follow a designed itinerary to finish service for customers on the sea
and visit selected ports of call in the route. However, in contrast to the cruise shipping, there
are tremendous research works on the container liner shipping. The examples are given as
follows. Meng et al. (2012) proposed a liner ship fleet planning problem considering container
transshipment and uncertain container shipment demand. A liner container seasonal shipping
revenue management problem for a container shipping company was researched by Wang et
al. (2015). Song et al. (2015) addressed a joint tactical planning problem for deciding the
number of ships, the planned maximum sailing speed, and the liner service schedule. Ship
deployment and empty container repositioning related problems were investigated by Song et
al. (2012) and Song et al. (2013). Ng (2014, 2015) studied fleet deployment-related problems
for liner shipping under stochastic environment. A cost-based maritime container assignment
model was formulated by Bell et al. (2013) to assign containers to routes to minimize the total
operational cost. More related works can be referred to Meng et al. (2014) for a review of
container liner service operations and planning. The majority of those works drew attention to
practical problems existing in the container liner shipping and developed useful optimization-
based planning tools.
Although the cruise services are akin to the container liner services, we cannot simply
transfer the methods used in the container liner shipping to the applications for the problems
related to the cruise shipping. There are some essential differences between them. For
example, the problems in the container liner shipping normally did not consider the berth
availability in ports of call. This is due to that the major container transshipment terminals
(e.g., the container terminals in Hong Kong and Singapore) have abundant berth resource, and
useful berth allocation techniques have been proposed by port logistic researchers (e.g., Meisel
and Bierwirth, 2009; Giallombardo et al., 2010; Zhen et al., 2011; Vacca et al., 2012; Iris et
al., 2015; Zhen, 2015), which give more flexibility for liner companies on the berth
availability. However, the berths in major cruise terminals are quite limited. For example,
Wusong Kou cruise terminal (Shanghai) and Kai Tak cruise terminal (Hong Kong) just have
the berth capacity to serve two cruise ships simultaneously. Therefore, the berth availability
should be emphasized in the cruise shipping-related problems. Moreover, the container liner
services are normally designed for repeats on a weekly basis, and the weekly demands from
customers are close to constant (except for some special weeks, such as the Chinese New Year
week and the Christmas week). Thus, the fluctuation of the profit for a container liner service
is little. In comparison, the cruise services appeal to the cruise passengers for their feeling of
freshness (Esteve-Perez and Garcia-Sanchez, 2014), which requires the diversity of the cruise
services. New and interesting cruise services should be provided frequently, and the multiple
repeats on a cruise service would bring significant decreasing on its marginal profit. Based on

Chapter 5: Cruise Ship Service Planning 112

these facts and discussions, the problems arising in the cruise shipping are different from the
problems existing in the container liner shipping inessential.
The decreasing marginal profit phenomenon is widely considered in the research works in
operations management or operations research area (Arthur and Ronald, 2000; Hongmin and
Woonghee, 2011; Li, 2011; Paat and Huseyin, 2012). However, there are limited research
works in maritime transportation area that considers the decreasing marginal profit or
decreasing marginal productivity. Meisel and Bierwirth (2009) and Iris et al. (2015)
investigated the integrated problem of berth allocation and quay crane assignment for
container terminals, in which they considered the decrease of marginal productivity of quay
cranes assigned to the vessels.
In our research, we address a tactical problem in cruise shipping: the cruise service planning
problem considering the berth availability and decreasing marginal profit. In fact, this problem
is a variant of knapsack problem (will be elaborated in Section 5.4). The major structural
difference between our problem and some nonlinear/nonconvex extensions to the knapsack
problem (Bretthauer and Shetty, 2002; Kameshwaran and Narahari, 2009; Poirriez et al., 2009)
is that the berth availability constraints in our setting are not well structured in the past
extensions. We cannot simply take advantage of the existing algorithms for the knapsack
problem and its extensions, all of which exploit the special structure that there are only linear
constraints in the problems. Thus, for our problem, we explore optimization-based service
planning tools to increase the profit by planning cruise services. Some mathematical models
(both linear and nonlinear models) of the problem are formulated in order to contribute to the
state-of-the-art research in a quantitative manner.

5.3 MATHEMATICAL MODEL

In this section, we provide a brief description of the CSP problem for a cruise ship
considering the decreasing marginal profit and the berth availability and formulate it as integer
programming models.

5.3.1 Problem description

The problem focuses on service planning for a specific cruise ship. Given a set of pre-
determined candidate services (denoted by ℝ = {1, 2, … , |ℝ|}) and a set of all days in a
planning horizon with 𝑇 days in total (denoted by 𝕋 = {1, 2, … , 𝑇}), the optimization of the
problem aims at selecting services for the cruise ship to operate in the planning horizon. Each
candidate service 𝑟 ∈ ℝ has a pre-determined rotation time, defined as 𝑠𝑟 (days), which
indicates the number of days needed for the cruise ship to operate such a service. The marginal
profit of operating Service 𝑟 is denoted as 𝑔𝑟 . The objective of the optimization is to maximize

Chapter 5: Cruise Ship Service Planning 113

the total profit by the cruise ship to operate the services selected from the set 𝑅 in the planning
horizon.
In this problem, we consider the berth availability of the visiting ports of call in all the
candidate services. In each day in the planning horizon, each port either has an available berth
or not for the cruise ship, which is known to the cruise ship in advance. A service can be
operated if and only if each port in the service has an available berth for the cruise ship. To
indicate the availability of berths at the ports visited during the planning horizon, we further
define a binary parameter 𝛿𝑟𝑡 , ∀𝑟 ∈ ℝ, 𝑡 ∈ 𝕋. Here, 𝛿𝑟𝑡 equals one if and only if Service 𝑟 can
be operated starting from 0:00 am in Day 𝑡, which means that all the visiting ports in Service
𝑟 have available berths for the cruise ship in future arrival times if the service starts from 0:00
am (Day 𝑡). For instance, suppose there is a cruise service 𝑟 ′ : Shanghai (China)→Cheju
(Korea)→Fukuoka (Japan)→Shanghai (China). The itinerary for the cruise service is given in
Table 5.1. 𝛿𝑟′ 2 is set to one when the cruise ship can operate the cruise service 𝑟 ′ starting from
Day 2 of the planning horizon. Meanwhile, along the cruise route in the planning horizon, it
has all available berths to moor at Shanghai (China) in Day 2, Cheju (Korea) in Day 3, Fukuoka
(Japan) in Day 5 & Day 6, and Shanghai (China) in Day 10. Normally, the cruise ship arrives
at a port around 6:00 am, and leaves from the port around 5:00 pm such that the cruise
passengers could tour around the port city in daytime. Thus, the berth in the port is usually
occupied by the cruise ship for a whole day. However, the berth can be occupied by the cruise
ship for two or three days if more tour time is arranged for onshore activities.

Table 5.1: The itinerary for the cruise service

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9

Shanghai Cheju Sea Fukuoka Fukuoka Sea Sea Sea Shanghai

When planning services for the cruise ship in the planning horizon, we assume that all the
services must be finished before 𝑇. To ensure this assumption, we initially set input data 𝛿𝑟𝑡
in the following way: for each candidate service 𝑟 with rotation time 𝑠𝑟 , 𝛿𝑟𝑡 ≔ 0 if 𝑡 is bigger
than 𝑇 − 𝑠𝑟 + 1 . This setting can guarantee that all the services are finished within the
planning horizon.
For the interest of simplicity, firstly, this study assumes that the marginal profit of operating
Service 𝑟 is constant when the service is repeated several times. Under this assumption, an
integer programming model is formulated, denoted as model 𝐹𝑀1. Then, in order to be close
to the real situation, we change the profit setting to that the marginal profit of operating a
service decreases when the service is repeated several times. Based on this, a nonlinear integer
programming model is formulated, denoted as model 𝐹𝑀2.

Chapter 5: Cruise Ship Service Planning 114

5.3.2 Model with constant marginal profit
As we have defined the rotation times of services 𝑠𝑟 , ∀𝑟 ∈ ℝ, the profits of operating
services 𝑔𝑟 , ∀𝑟 ∈ ℝ , the planning horizon 𝑡 ∈ 𝕋 = {1 , 2, ⋯ , 𝑇 − 1, 𝑇} , and the berth
availability 𝛿𝑟𝑡 , ∀𝑟 ∈ ℝ, 𝑡 ∈ 𝕋, we further define decision variables as 𝑧𝑟𝑡 , ∀𝑟 ∈ ℝ, 𝑡 ∈ 𝕋,
which equals one if and only if Service 𝑟 is operated starting from 0:00 am (Day 𝑡). Then, the
model 𝐹𝑀1 can be formulated as follows:
[𝑭𝑴𝟏] Maximize ∑𝑡∈𝕋 ∑𝑟∈ℝ 𝑔𝑟 𝑧𝑟𝑡 (5.1)
s.t. 𝑧𝑟𝑡 ≤ 𝛿𝑟𝑡 ∀𝑟 ∈ ℝ, 𝑡 ∈ 𝕋 (5.2)
∑𝑟∈ℝ ∑𝑡𝑡 ′ =max(𝑡−𝑠𝑟 +1,1) 𝑧𝑟𝑡 ′ ≤ 1 ∀𝑡 ∈ 𝕋 (5.3)
𝑧𝑟𝑡 ∈ {0, 1} ∀𝑟 ∈ ℝ, 𝑡 ∈ 𝕋. (5.4)
In the above model 𝐹𝑀1, Objective (5.1) maximizes the total profit by selecting cruise
services from a set of pre-determined candidate services in the planning horizon. Constraints
(5.2) ensure that all the designed services are operable considering the availability of berths at
the ports visited. Constraints (5.3) guarantee that in each day, the cruise ship operates at most
one cruise service, and once a cruise service is finished, the next cruise service can be started.
There is no rotation time overlap between the two selected cruise services in the planning
horizon. Constraints (5.4) define the domains of the binary decision variables.

5.3.3 Model with decreasing marginal profit

Figure 5.1: Concavity of the profit function

In the model with decreasing marginal profit, we assume that when a service is repeated for
several times, the marginal profit of the service will decrease gradually. This consideration
makes the problem close to the reality as few potential passengers would choose a cruise
service when the cruise service has been repeated many times. To show the decreasing pattern
of the marginal profit in the model, we define a strictly increasing concave function 𝐺𝑟 (𝑥𝑟 ) as

Chapter 5: Cruise Ship Service Planning 115

the total profit for operating Service 𝑟 ∈ ℝ by a total of 𝑥𝑟 times, which can be designed as
follows:
𝐺𝑟 (𝑥𝑟 ) < 𝐺𝑟 (𝑥𝑟 + 1) (5.5)
𝐺𝑟 (𝑥𝑟 + 2) − 𝐺𝑟 (𝑥𝑟 + 1) ≤ 𝐺𝑟 (𝑥𝑟 + 1) − 𝐺𝑟 (𝑥𝑟 ). (5.6)
Figure 5.1 demonstrates the concavity of the function on the total profit for operating a service
repeatedly.
Based on the above concave function, the model with decreasing marginal profit (i.e., 𝐹𝑀2)
can be formulated as follows:
[𝑭𝑴𝟐] Maximize ∑𝑟∈ℝ 𝐺𝑟 (𝑥𝑟 ) (5.7)
s.t. Constraints (5.2)-(5.4) and;
𝑥𝑟 = ∑𝑇𝑡=1 𝑧𝑟𝑡 ∀𝑟 ∈ ℝ. (5.8)
In this model, Objective (5.7) maximizes the sum of the total profits for operating services.
Constraints (5.8) are used to calculate the number of repeats for each service.
However, 𝐹𝑀2 is a nonlinear model as it involves the concave objective function, which
makes it hard to be solved by some efficient solution methods. Thus, we present two
transformed linear models for 𝐹𝑀2. Before building the two linear models, we define 𝑥𝑟 as
an upper bound of the number of repeats 𝑥𝑟 for Service 𝑟. 𝑥𝑟 can be calculated by:
𝑇
𝑥𝑟 = ⌊ 𝑠 ⌋ ∀𝑟 ∈ ℝ. (5.9)
𝑟

Here, it is worthwhile to mention that there is a better way to derive the upper bound of the
number of repeats for Service 𝑟. We propose a greedy algorithm to find the maximum number
of repeats for Service 𝑟 in the planning horizon (note that the greedy algorithm is applied to
derive optimal 𝑥𝑟 in the numerical experiments in Section 5.5), which is a tighter upper bound
compared with Eq. (5.9). The details of the greedy algorithm and why the greedy algorithm
can find the optimal number of repeats for Service 𝑟 will be given in the proof of Proposition
5 in Section 5.4.3.
The first linear model:
The first linear model is defined as 𝐹𝑀2′ , which does not take advantage of the problem
structure for the linearization. In this linear model, 𝑔𝑟𝑥 ≔ 𝐺𝑟 (𝑥) − 𝐺𝑟 (𝑥 − 1) is used to define
the marginal profit of the 𝑥 𝑡ℎ repeat of Service 𝑟. Meanwhile, the binary decision variable
(i.e., 𝑧𝑟𝑡 ,) in 𝐹𝑀2 is changed to another binary decision variable denoted by 𝑧𝑟𝑥𝑡 , ∀𝑟 ∈ ℝ, 𝑥 ∈
{1, 2 ⋯ 𝑥𝑟 }, 𝑡 ∈ 𝕋. Here, 𝑧𝑟𝑥𝑡 equals one if and only if Service 𝑟 is operated for the 𝑥 𝑡ℎ time
starting from 0:00 am (Day 𝑡 ). Based on the above redefinition of some variables, the
formulation of 𝐹𝑀2′ is:

[𝑭𝑴𝟐′ ] Maximize ∑𝑡∈𝕋 ∑𝑟∈ℝ ∑𝑥𝑥=1

𝑟
𝑔𝑟𝑦 𝑧𝑟𝑥𝑡 (5.10)
s.t. 𝑧𝑟𝑥𝑡 ≤ 𝛿𝑟𝑡 ∀𝑟 ∈ ℝ, 𝑥 ∈ {1, 2 ⋯ , 𝑥𝑟 }, 𝑡 ∈ 𝕋 (5.11)

Chapter 5: Cruise Ship Service Planning 116

 ∑𝑟∈ℝ ∑𝑡𝑡 ′ =max(𝑡−𝑠𝑟 +1,1) ∑𝑥𝑥=1
𝑟
𝑧𝑟𝑥𝑡 ′ ≤ 1 ∀𝑡 ∈ 𝕋 (5.12)
∑𝑇𝑡=1 𝑧𝑟𝑥𝑡 ≤ 1 ∀𝑟 ∈ ℝ, 𝑥 ∈ {1, 2 ⋯ , 𝑥𝑟 } (5.13)
∑𝑇𝑡=1 𝑡𝑧𝑟𝑥𝑡 + 𝑠𝑟 − 1 ≤ ∑𝑇𝑡=1 𝑡𝑧𝑟,𝑥+1,𝑡 + 𝑇(1 − ∑𝑇𝑡=1 𝑧𝑟,𝑥+1,𝑡 ) ∀𝑟 ∈ ℝ, 𝑥 ∈ {1, ⋯ , 𝑥𝑟 −
1} (5.14)
𝑧𝑟𝑥𝑡 ∈ {0, 1} ∀𝑟 ∈ ℝ, 𝑥 ∈ {1, 2 ⋯ , 𝑥𝑟 }, 𝑡 ∈ 𝕋. (5.15)
In the above model 𝐹𝑀2′ , Objective (5.10) maximizes the sum of the total profits in the
planning horizon. Constraints (5.11) guarantee that all the selected services satisfy the
availability constraints of berths at the ports visited. Constraints (5.12) limit that the cruise
ship can only provide one service in one day and once a service 𝑟 starts, the ship cannot
provide other services in 𝑠𝑟 days. Constraints (5.13) guarantee that each repeat of each service
(e.g., the 𝑥 𝑡ℎ repeat of Service 𝑟) can only be operated by the cruise ship at most one time.
Constraints (5.14) enforce that for each service, all the repeats must have a chronological order,
which means a latter repeat cannot be started before a former repeat. Constraints (5.15) define
the domains of the binary decision variables. Note that in the above formulation, for each
service, a former repeat will be selected prior to a latter repeat as the marginal profit of the
former is higher than that of the latter.
Constraints (5.14) could be removed to reduce the computational time for the model 𝐹𝑀2′ .
Without Constraints (5.14), the optimal objective value is still the same, but the optimal
solution for the model might be infeasible for the problem as some latter repeats might be
started before some former repeats in a chronological order. However, the infeasible situation
can be sorted manually by adjusting the chronological order for the repeats of a cruise service.
The above tactic will be tested and verified in the computational experiment section.
The second linear model:
The second linear model is defined as 𝐹𝑀2′′ , which is formulated by taking advantage of
the concavity of 𝐺𝑟 (𝑥𝑟 ) (Premoli, 1986). In order to build this linear model, we introduce
continuous variables 𝑢𝑟 , ∀𝑟 ∈ ℝ that represent the total profit by operating Service 𝑟
repeatedly. With the continuous variables, the formulation of 𝐹𝑀2′ is:
[𝑭𝑴𝟐′′ ] Maximize ∑𝑟∈ℝ 𝑢𝑟 (5.16)
s.t. Constraints (5.2-5.4); (5.8);
𝐺𝑟 (𝑥+1)−𝐺𝑟 (𝑥)
𝑢𝑟 ≤ 𝐺𝑟 (𝑥) + (𝑥+1)−𝑥
(𝑥𝑟 − 𝑥) ∀𝑟 ∈ ℝ, 𝑥 ∈ {0,1, ⋯ , 𝑥𝑟 − 1}. (5.17)

In this model, Objective (5.16) maximizes the sum of the total profits for operating services
by repeating different times. Constraints (5.17) are used to calculate the total profit of each
service in a linear manner. Here, notice that 𝐺𝑟 (0) ≔ 0, ∀𝑟 ∈ ℝ.
Both linear models provide the linearization for the nonlinear model (i.e., 𝐹𝑀2), which will
be further compared in Section 5.5 for the computation efficiency to solve the problem.

Chapter 5: Cruise Ship Service Planning 117

5.4 COMPLEXITY ANALYSIS AND COMPARISON WITH HEURISTICS

In this section, for the complexity analysis, two dynamic programming (DP) based pseudo-
polynomial algorithms are developed for the model 𝐹𝑀1 and the model 𝐹𝑀2 respectively.
Some properties on commonly used heuristic methods for the CSP problem are also
investigated and proved.

5.4.1 Complexity of the problem of 𝑭𝑴𝟏 with constant marginal profit

Proposition 5.1: The problem of 𝐹𝑀1 is NP-hard.
Proof: Suppose all ports always have berths for the cruise ship, which means no matter
which services have been chosen for the planning horizon, the cruise ship always has the
available berths to dwelling whenever it arrives at the ports in the services. Then, the 𝐹𝑀1
becomes a problem of maximizing the total profit by choosing a number of services from the
set of candidate services with different rotation times and profits. This is exactly an unbounded
knapsack (Poirriez et al., 2009). Since the knapsack problem is NP-hard, the general version
of the problem of 𝐹𝑀1 is also NP-hard. ∎
Proposition 5.2: The problem of 𝐹𝑀1 is weakly NP-hard.
Proof: We can propose a DP based pseudo-polynomial algorithm for the problem, which
could demonstrate that the problem of 𝐹𝑀1 is weakly NP-hard. The procedures of the DP
algorithm are elaborated as follows:
To apply the DP for the problem of 𝐹𝑀1, we firstly define 𝑈(𝑡) as the maximum possible
total profit of operating services (to be determined) from 0:00 am (Day 𝑡) to the end of Day 𝑇,
i.e., 0:00 am (Day 𝑇 + 1). Then, we make the decision for each day whether choosing a service
to start or designating the cruise ship to stay in the harbor of the home port. Let 𝑧𝑟𝑡 denote the
binary variable of the decision, which equals one if and only if Service 𝑟 is started in Day 𝑡.
We initially have the boundary conditions as:
𝑈(𝑡) = −∞ 𝑡 = 𝑇 + 2, 𝑇 + 3, … (5.18)
𝑈(𝑡) = 0 𝑡 = 𝑇 + 1. (5.19)
The DP consists of 𝑇 stages (for 𝑡 decreasing from 𝑇 to 1) and computes the total profit at
each stage 𝑡 ∈ 𝕋 based on choosing a service to start or designating the cruise ship to stay in
the harbor of the home port for one day, by using classical Bellman recursion:
𝑈(𝑡) = 𝑚𝑎𝑥𝑟∈ℝ {(1 − 𝑧𝑟𝑡 ) ∙ 𝑈(𝑡 + 1) + 𝑧𝑟𝑡 ∙ (𝑔𝑟 + 𝑈(𝑡 + 𝑠𝑟 )) | 𝑧𝑟𝑡 ≤ 𝛿𝑟𝑡 , 𝑧𝑟𝑡 ∈ {0, 1}}

∀𝑡 ∈ 𝕋. (5.20)
To calculate 𝑈(𝑡) at each stage, we firstly enumerate all the services and select feasible
services by considering the availability of berths at the ports visited (𝛿𝑟𝑡 ). Then, we start the
feasible services one by one to derive the profits for the feasible services, and also calculate

Chapter 5: Cruise Ship Service Planning 118

the profit when no service is started. Those profits are compared to obtain the maximal profit
𝑈(𝑡) at each stage.

Algorithm 5.1. DP-based algorithm for the model FM1

Input: A set of candidate services 𝑟 ∈ ℝ, with operating profit 𝑔𝑟 and rotation time 𝑠𝑟
Output: An optimal schedule to operate cruise services
// initialization
for 𝑡 ← 𝑇 + 1 to 1 do
𝑈(𝑡) ← 0
end for
for 𝑡 ← 𝑇 + 2 to 𝑇 + max{𝑠𝑟 |∀𝑟 ∈ ℝ} do
𝑈(𝑡) ← −∞
end for
// the DP procedure
for 𝑡 ← 𝑇 to 1 do
𝑈(𝑡) ← 𝑈(𝑡 + 1)
for 𝑟 ← 1 to |ℝ| do
if 𝑈(𝑡) < 𝑔𝑟 + 𝑈(𝑡 + 𝑠𝑟 ) and 𝛿𝑟𝑡 = 1 then
𝑈(𝑡) ← 𝑔𝑟 + 𝑈(𝑡 + 𝑠𝑟 )
end if
end for
end for
Finally, we will obtain the value 𝑈(1), the objective value of 𝐹𝑀1, which represents the
maximum total profit of operating services from 0:00 am (Day 𝑡) to the end of Day 𝑇. The
∗
optimal solution can be extracted from the values of 𝑧𝑟𝑡 , ∀𝑟 ∈ ℝ, 𝑡 ∈ 𝕋. The pseudocode of
this DP based algorithm is elaborated in Algorithm 5.1.
In summary, the proposed algorithm runs in 𝑂(𝑇 ∙ |ℝ|) time for the solution. There are 𝑇
stages in the proposed DP. The decision at each stage is which service 𝑟 ∈ ℝ to start or
designating the cruise ship to stay in the harbor of the home port for one day. Therefore, the
computational complexity for the DP is 𝑇 ∙ |ℝ|, which demonstrates that the problem of 𝐹𝑀1
is weakly NP-hard. ∎

5.4.2 Complexity of the problem of 𝑭𝑴𝟐 with decreasing marginal profit

Corollary 5.1: The problem of 𝐹𝑀2 is NP-hard.
Proof: The problem of 𝐹𝑀2 nests the problem of 𝐹𝑀1 as a special case. If we change Eq.
(6) to 𝐺𝑟 (𝑥𝑟 + 2) − 𝐺𝑟 (𝑥𝑟 + 1) = 𝐺𝑟 (𝑥𝑟 + 1) − 𝐺𝑟 (𝑥𝑟 ) , 𝐹𝑀2 becomes 𝐹𝑀1 . As 𝐹𝑀2 is
more general than 𝐹𝑀1, and the problem of 𝐹𝑀1 is NP-hard, the problem of 𝐹𝑀2 is also NP-
hard. ∎
Here, we would like to investigate the problem of 𝐹𝑀2 in a special case, in which we
assume that all the ports in the candidate services have sufficient berth availability at any time
for the cruise ship. The reasons for such investigation are listed as follows: (i) it can be used
as a benchmark for cruise companies in the sense of the total profit. They could assess the
maximal profit that can be earned when the berth availability is in a perfect condition. (ii)

Chapter 5: Cruise Ship Service Planning 119

Some cruise terminals are operated by cruise companies, thus, the investigation on the special
case is meaningful for them to make investment decisions on berth construction. (iii) The
special case is also useful for cruise port policy makers to evaluate whether the berth
availability is the limitation for the local cruise shipping development.
Proposition 5.3: The problem of 𝐹𝑀2 is weakly NP-hard in the special case with sufficient
berth availability.
Proof: We can propose a DP based pseudo-polynomial algorithm for the special case of the
problem, which demonstrates that special case is weekly NP-hard. The procedures of the DP
algorithm are elaborated as follows:
In the special case for 𝐹𝑀2, the availability of berths at the ports visited is sufficient, which
means that all the ports in the services have available berths for the cruise ship at any time.
The application of the DP is significantly different from that for the problem of 𝐹𝑀1 as the
decreasing pattern of the service marginal profit has been considered in the problem of 𝐹𝑀2.
Enlightened by the formulation of 𝐹𝑀2′ , in the DP for the special case of 𝐹𝑀2, we assume
that each combination of (𝑟, 𝑥) is a “detailed service” with the marginal profit 𝑔𝑟𝑥 ≔ 𝐺𝑟 (𝑥) −
𝐺𝑟 (𝑥 − 1), which denotes the 𝑥 𝑡ℎ time for the repeat of Service 𝑟. Each “detailed service” can
only be started for once in the planning horizon. Meanwhile, a latter “detailed service” cannot
be started before a former “detailed service”. For instance, (𝑟, 5) cannot be started if (𝑟, 4) has
not been started. Here, we define an index 𝛽 and a set 𝕊, 𝛽 ∈ 𝕊 for all the possible “detailed
services”; here 𝕊 = {(1,1), (1,2), ⋯ , (1, 𝑥1 ), ⋯ , (𝑟, 1), (𝑟, 2), ⋯ , (𝑟, 𝑥𝑟 )}. The upper bound
for |𝕊| is |ℝ| ∙ 𝑇.
In the case of the problem of 𝐹𝑀2, we can deem the problem as: we are packing |𝕊|
“detailed services” with different profits and rotation times into a period of time 𝑇, which is a
0/1 knapsack problem. To build the DP for the case, We further define 𝑉(𝛽, 𝑡) as the maximum
possible total profit of operating services (to be determined) from 0:00 am (Day 1) to the end
of Day 𝑡 (0:00 am of Day 𝑡 + 1) by choosing the services from first 𝛽 “detailed services”; all
of the operated services must finish by the end of Day 𝑡. Then, we make decisions at each
stage on whether to place “detailed service” 𝛽 to finish at the end of Day 𝑡 or not. Based on
the above information, we initially have the boundary conditions as:
𝑉(0, 𝑡): = 0 ∀ 𝑡 ∈ {0, 1, ⋯ 𝑇} (5.21)
𝑉(𝛽, 0): = 0 ∀ 𝛽 = 1, 2, … , |𝕊|. (5.22)
The DP procedure consists of |𝕊| service stages (for 𝛽 increasing from 1 to |𝕊|) and computes
the total profit at each stage 𝛽 ∈ 𝕊 with the time stage 𝑡 increasing from 1 to 𝑇. The DP
procedure uses the classical Bellman recursion as follows:
𝑉(𝛽 − 1, 𝑡) , 𝑖𝑓 𝑡 < 𝑠𝛽′
𝑉(𝛽, 𝑡) = { ∀𝛽 ∈ 𝕊, 𝑡 ∈ 𝒯 (5.23)
max{𝑉(𝛽 − 1, 𝑡), 𝑉(𝛽 − 1, 𝑡 − 𝑠𝛽′ ) + 𝑔𝛽′ } , 𝑖𝑓 𝑠𝛽′ ≤ 𝑡

Chapter 5: Cruise Ship Service Planning 120

where, 𝑠𝛽′ , 𝑔𝛽′ and 𝛿𝛽𝑡
′
equal to 𝑠𝑟 , 𝑔𝑟𝑥 and 𝛿𝑟𝑡 , respectively, if “detailed service” 𝛽 is the 𝑥 𝑡ℎ
repeat for Service 𝑟 (“detailed service” 𝛽 is the combination of (𝑟, 𝑥)). By conducting the
recursion, we will obtain the value 𝑉(|𝕊|, 𝑇), which represents the maximum total profit
without considering the availability of berths at the ports (i.e., the availability of berths at the
ports visited is sufficient all the time). The pseudocode of this DP based algorithm is elaborated
in Algorithm 5.2.
Algorithm 5.2. DP-based algorithm for the model FM2 in a special case
Input: A set of candidate “detailed service” 𝛽, ∀𝛽 ∈ 𝕊, with operating profit 𝑔𝛽′ and
rotation time 𝑠𝛽′
Output: An optimal schedule to operate “detailed service”
// initialization
for 𝑡 ← 0 to 𝑇 do
𝑉(0, 𝑡) ← 0
end for
for 𝛽 ← 1 to |𝕊| do
𝑉(𝛽, 0) ← 0
end for
// the DP procedure
for 𝛽 ← 1 to |𝕊| do
for 𝑡 ← 0 to 𝑇 do
if 𝑡 < 𝑠𝛽′ then
𝑉(𝛽, 𝑡) ← 𝑉(𝛽 − 1, 𝑡)
else
if 𝑉(𝛽 − 1, 𝑡) < 𝑉(𝛽 − 1, 𝑡 − 𝑠𝛽′ ) + 𝑔𝛽′ then
𝑉(𝛽, 𝑡) ← 𝑉(𝛽 − 1, 𝑡 − 𝑠𝛽′ ) + 𝑔𝛽′
else
𝑉(𝛽, 𝑡) ← 𝑉(𝛽 − 1, 𝑡)
end if
end if
end for
end for

In summary, the model 𝐹𝑀2′ enlightens us to consider each combination of (𝑟, 𝑥) as a

“detailed service”. The problem becomes how to place detailed services into a period of time
𝑇 to maximize the profit. Time complexity of the DP is O(|𝕊| ∙ 𝑇), where |𝕊| ≤ |ℝ| ∙ 𝑇. Thus,
time complexity of the DP is bounded by O(|ℝ| ∙ 𝑇 2 ) and the DP algorithm is a pseudo-
polynomial algorithm, which show that the special case is weekly NP-hard. ∎
For the general case of the problem of 𝐹𝑀2 with considering the berth availability and
decreasing marginal profit, we cannot propose a pseudo-polynomial algorithm using DP.
Therefore, the two linear models 𝐹𝑀2′ and 𝐹𝑀2′′ are solved directly by CPLEX for the
optimal solutions of 𝐹𝑀2.

Chapter 5: Cruise Ship Service Planning 121

5.4.3 Comparison with commonly used heuristics
The models and the solution methods (the DP-based algorithms) solve the CSP problem
optimally under different assumptions. However, in a real situation, there are some commonly
used myopic heuristics to solve the CSP problem. In this section, the properties of the solutions
obtained by those heuristics and the optimal solutions are investigated for the general case
with decreasing marginal profit.
In the knapsack problem and its variants, there are two commonly used heuristics:
operating-profit-first heuristic and unit-profit-first heuristic. For the example of the special
case of the model 𝐹𝑀2 mentioned in the previous section (a variant of the knapsack problem),
the two heuristics are as follows. (i) The operating-profit-first heuristic: according to the
′
sequence of the profits of “detailed services” such that 𝑔1′ ≥ 𝑔2′ ≥ ⋯ ≥ 𝑔|𝕊| , we put the
“detailed services” into the planning horizon sequentially until it is not possible to place more
“detailed services”. (ii) The unit-profit-first heuristic: a service’s unit profit is the service’s
operational profit divided by the service’s rotation time. According to the sequence of the unit
′
𝑔1′ 𝑔2′ 𝑔|𝕊|
profits of the “detailed services” such that ≥ ≥⋯≥ ′ , we put the “detailed services”
𝑠1′ 𝑠2′ 𝑠|𝕊|

into the planning horizon sequentially until it is impossible to place more “detailed services”.
Based on the above two commonly used heuristics, we can design two myopic heuristic
rules to solve the general case of the model 𝐹𝑀2 when considering the berth availability, in
which the decision is made on a daily basis from Day 1 to Day 𝑇.
Myopic Rule_1: For a specified Day 𝑡, we determine all the cruise services that can be
operated with considering the berth availability (i.e., ∀𝑟, 𝛿𝑟𝑡 = 1). Based on those cruise
services, we select the optimal cruise service 𝑟 ∗ with the maximal daily operating profit (i.e.,
𝑔𝑟∗ /𝑠𝑟∗ ). Then, the time is updated to Day 𝑡 + 𝑠𝑟∗ for the next selection. If no cruise service
can be operated, the time is updated to Day 𝑡 + 1 for the next selection.
Myopic Rule_2: For a specified Day 𝑡, we determine all the cruise services that can be
operated with considering the berth availability (i.e., ∀𝑟, 𝛿𝑟𝑡 = 1) at first. Among those cruise
services, we select the optimal cruise service 𝑟 ∗ with the maximal operating profit (i.e., 𝑔𝑟∗ )
to operate. Then, the time is updated to Day 𝑡 + 𝑠𝑟∗ for the next selection. If no cruise service
can be operated, the time is updated to Day 𝑡 + 1 for the next selection. Notice that if two
cruise services have the same daily operating profit or the same operating profit in the rules,
the priority will be given to the cruise service with shorter rotation time as it would occupy
few days in the planning horizon.
Here, Myopic Rule_1 (Myopic Rule_2) is designed by the unit-profit-first heuristic (the
operating-profit-first heuristic) as it selects the optimal cruise service 𝑟 ∗ with the maximal

Chapter 5: Cruise Ship Service Planning 122

daily operating profit 𝑔𝑟∗ /𝑠𝑟∗ (with the maximal operating profit 𝑔𝑟∗ ). The solutions obtained
by two myopic rules will be further compared with the optimal solutions in Section 5.3.
Proposition 5.4: In the worst case, the ratio between the optimal profit obtained by the
model 𝐹𝑀2 and the profit obtained by Myopic Rule_1 or Myopic Rule_2 is close to infinity.
Proof: Let 𝑍 ∗ be the optimal total profit derived by the model 𝐹𝑀2, 𝑍̈ the total profit
derived by the Myopic Rule_1, and 𝑍̂ the total profit derived by the Myopic Rule_2.
Assuming that the number of days in the planning horizon is 𝑇 > 2, and there are two
candidate cruise services with the operating profits and rotation times as follows: 𝑔1 = 𝑘, 𝑠1′ =
2; 𝑔2 = 𝑛 ∙ 𝑘, 𝑠2′ = 𝑇 − 1. Here, 𝑘 is a profit constant with unit of US$, and 𝑛 is a ratio that
is bigger than one. Assume that Cruise service 1 can be operated on Day 1 and Cruise service
2 cannot be operated on Day 1 due to berth unavailability. Then, Cruise service 2 can be in
Day 2 and Cruise service 1 cannot be operated since Day 2 due to the berth availability. In
such situation, the profits derived by two heuristic rules are 𝑍̈ = 𝑍̂ = 𝑔1 = 𝑘, but the optimal
total profit is 𝑍 ∗ = 𝑛 ∙ 𝑘. Thus, the ratio between the optimal total profit and the total profit
obtained by Myopic Rule_1 or Myopic Rule_2 is 𝑛. When 𝑛 → +∞, the ratio is close to
infinity. ∎

For the CSP problem considering berth availability, the two commonly used heuristic rules
do not give the priority on the berth availability, which leads to tremendous profit loss. Thus,
in cruise shipping, the operation managers should keep well informed about the berth
availability from cruise terminals. Based on the information, the managers should make
schedules on the overall picture for a whole period rather than from day to day. In reality, it is
impossible to guarantee the sufficient berth availability in all cruise terminals, but we do
encourage that the cruise lines own some cruise terminals such that they have more flexibility
on berths to operate their cruise services.
Proposition 5.5: When there is only one candidate cruise service 𝑟 ′ , the solution obtained
by Myopic Rule_1 or Myopic Rule_2 is the optimal solution of the model 𝐹𝑀2.
Proof: When there is only one candidate cruise service 𝑟 ′ , the rotation time is constant, then
we can transfer the two myopic heuristic rules to a greedy algorithm for the problem, in which
the decision is made on a daily basis from Day 1 to Day 𝑇: for a specified Day 𝑡, if the cruise
service 𝑟 ′ can be operated with considering the berth availability, this cruise service is settled
for the operation. Then, the time is updated to Day 𝑡 + 𝑠𝑟′ for the next decision. Otherwise,
the time is updated to Day 𝑡 + 1 for the next decision.
As there is only and one candidate cruise service (the cruise service 𝑟 ′ ), the objective of the
model 𝐹𝑀2 aims to maximize the total profit earned by the one cruise service (maximize 𝑍 =
𝐺𝑟′ (𝑥𝑟′ )). Based on the concavity of the function 𝐺𝑟′ (𝑥𝑟′ ), as it is shown in Figure 5.1, the

Chapter 5: Cruise Ship Service Planning 123

objective of the model 𝐹𝑀2 is consistent with aiming to maximize 𝑥𝑟′ (i.e., maximize the
number of repeats of the cruise service 𝑟 ′ ).
Here, we define the number of repeats 𝑥𝑟′ obtained by the greedy algorithm as 𝑁, and the
optimal number of repeats obtained by the model 𝐹𝑀2 as 𝑀∗ , where 𝑀∗ ≥ 𝑁. We denote 𝜑𝑖
( ∀𝑖 ∈ {1,2, … , 𝑁} ) as the start time of 𝑖 𝑡ℎ repeat in the solution obtained by the greedy
algorithm, and denote 𝜙𝑗 (∀𝑗 ∈ {1,2, … , 𝑀∗ }) as the start time of 𝑗 𝑡ℎ repeat in the optimal
solution obtained by the model 𝐹𝑀2.
Firstly, we arbitrarily assume that 𝑀∗ > 𝑁. As the greedy algorithm would start a repeat as
early as possible once it finds a day when the berths are available, we can have a conclusion
that is 𝜑1 ≤ 𝜙1 . As 𝑀∗ > 𝑁, there must exist a 𝑘 such that 𝜑𝑘 > 𝜙𝑘 , where 𝑘 ∈ [2, 𝑁]. If
there is no such 𝑘, there exist 𝜑𝑁 ≤ 𝜙𝑁 , which means the last repeat (the 𝑁 𝑡ℎ repeat) in the
solution obtained by the greedy algorithm starts the repeat earlier than the 𝑁 𝑡ℎ repeat in the
optimal solution, and the former 𝑁 𝑡ℎ repeat ends before the latter 𝑁 𝑡ℎ repeat. This suggests
that the greedy algorithm still have enough residual time space in the planning horizon to
operate the 𝑁 + 1𝑡ℎ repeat, which is in the conflict with the definition. Therefore, there must
exist a 𝑘 such that 𝜑𝑘 > 𝜙𝑘 . As we have proved that 𝜑1 ≤ 𝜙1 , there exist 𝜑𝑖 ≤ 𝜙𝑖 , ∀𝑖 ∈
{1,2, … , 𝑘 − 1} and 𝜑𝑗 > 𝜙𝑗 , ∀𝑗 ∈ {𝑘, 𝑘 + 1, … , 𝑁} . Here, comes another conflict: in the
solution obtained by the greedy algorithm, the (𝑘 − 1)𝑡ℎ repeat starts to be operated earlier
than the (𝑘 − 1)𝑡ℎ repeat in the optimal solution such that 𝜑𝑘−1 ≤ 𝜙𝑘−1 , which implies that
the former (𝑘 − 1)𝑡ℎ repeat ends before the latter (𝑘 − 1)𝑡ℎ repeat. Then, as the greedy
algorithm would start a repeat as early as possible in principle, how could the 𝑘 𝑡ℎ repeat from
the greedy algorithm starts to be operated later than the 𝑘 𝑡ℎ repeat in the optimal solution such
that 𝜑𝑘 > 𝜙𝑘 . This is where the other conflict rises.
In summary, all the above conflicts point out that the initial assumption 𝑀∗ > 𝑁 is wrong.
As we have 𝑀∗ ≥ 𝑁, we could easily conclude that 𝑀∗ = 𝑁, which implies the solution
obtained by the greedy algorithm is the optimal solution obtained by the model 𝐹𝑀2 when
there is only and one candidate cruise service 𝑟 ′ . ∎
Proposition 5.5 shows, when there is only one candidate cruise service, the two myopic rules
work the same as a greedy algorithm to obtain the optimal solution. Such a greedy algorithm
can be applied to derive a better upper bound 𝑥𝑟 for the number of repeats for each candidate
cruise service than that estimated by Eq. (5.9). The greedy algorithm is better than Eq. (5.9)
for the approximation because the former one obtains the optimal number of repeats. The
comparison of the approximation by the greedy algorithm and Eq. (5.9) will be given in
Section 5.5.5.

Chapter 5: Cruise Ship Service Planning 124

5.5 COMPUTATIONAL EXPERIMENT

In this section, in order to validate the effectiveness of the proposed models and efficiency
of solving the models, we conduct extensive numerical experiments by using a PC (Intel Core
i5, 2.3G Hz; Memory, 8G). The integer programs 𝐹𝑀2′ and 𝐹𝑀2′′ are solved by CPLEX12.5
with concert technology of C# (2012).

5.5.1 Generation of test instances

The planning horizon for the problem is 180 days (about half a year). The decisions (𝑧𝑟𝑡 or
𝑧𝑟𝑥𝑡 in the proposed models) are made on each day. The generation of the set of candidate
services is different in the following four subsections of computational experiments. In Section
5.2 and Section 5.3, which aim to test the efficiency and the effectiveness of models, the
candidate services are randomly generated with the rotation time assigned as 𝑠𝑟 ∈ 𝑈[4,11],
where 𝑈 denotes uniformly distributed integer pseudorandom numbers. In Section 5.4 and
Section 5.5, which focus on the robustness test and sensitivity analysis on the model for
Quantum of the Seas (one of cruise ship belonging to Royal Caribbean), the candidate services
for the cruise ship are inputted referring to the published schedule by Royal Caribbean
International (Cruise route: Quantum of the Seas, 2016). The details of those cruise services
will be illustrated in Section 5.5.4.
For the berth availability (𝛿𝑟𝑡 ), we derived the input parameters from the website of cruise
terminals: firstly, we analyzed the arrival times for all the incoming cruise ships in the Year
2016 at Wusong Kou Cruise Terminal (Shanghai) from Arrival Time (2016). Based on the
statistical results, there are 43% days left in the whole year that the terminal has available
berths. Therefore, we assume that the cruise terminal of each port city has randomly 40% to
50% days left for having available berths. Then, the berth availability of each port in a
specified day is randomly generated based on the random percentage obtained, which further
forms a berth availability sheet for each port in the planning horizon. Finally, the berth
availability for each cruise service (𝛿𝑟𝑡 ) can be derived by referring to the berth availability
sheets of the ports that the cruise service will visit. However, for the berth availability applied
in practical applications, the cruise ship managers could contact all the cruise terminals for the
arrival time sheets in advance.
To generate the profits of operating services (𝑔𝑟 or 𝑔𝑟𝑥 ), two input parameters are further
involved, which are the number of possible cruise passengers (denoted as 𝑛) and the average
profit per cruise passenger (denoted as 𝑝) of a cruise service. The profit of operating a service
could be calculated by: 𝑔 = 𝑛 × 𝑝. According to Cruise Industry (2015), the average revenue
per cruise passenger is US$1,728, and the average profit per cruise passenger is US$185,
which suggests that the ratio between the average profit and the average revenue is 0.107.
Meanwhile, according to Cruise Market Watch (2016), the ratio between the ticket price and

Chapter 5: Cruise Ship Service Planning 125

the average revenue per cruise passenger of a cruise service is 0.759. With these two ratios,
we could estimate that around 14% of the ticket price contributing to the average profit per
cruise passenger. The ticket price of a cruise service can be found easily. Thus, for a given
cruise service, the average profit per cruise passenger 𝑝 is also assessable.
Here, notice that in the model 𝐹𝑀2 with decreasing marginal profit setting, we assume 𝑝
(the average profit per cruise passenger) keeps unchanged for a cruise service, but the number
of cruise passenger 𝑛 decreases if the cruise service is repeated many times. We assume that
𝑛 decreases in an equal ratio pattern, which means once a cruise service is repeated one more
time, the number of the cruise passengers for the new repeat is 𝑛 ∙ 𝑎, here 𝑎 is the ratio, and
𝑎 ∈ (0,1). Initially, we randomly set the ratio 𝑎 from 0.80 to 0.90 for each candidate cruise
service.

5.5.2 Efficiency of solving the models

In this section, we compare the model 𝐹𝑀2′ with Constraints (5.14) and without
Constraints (5.14), which are solved by CPLEX in different instance groups. The comparison
results are shown in Table 5.2. As can be seen, in both cases, the optimal solution of each
instance can be obtained. However, in terms of the computational time, CPLEX solves the
model 𝐹𝑀2′ without Constraints (5.14) much faster than the model 𝐹𝑀2′ with Constraints
(5.14). On average, solving the former case only needs around 32% CPU time of the latter
case based on the ratio between 𝑇𝑜 and 𝑇𝑤 . More importantly, the ratio keeps decreasing with
the increase of the problem size. Thus, when using the model 𝐹𝑀2′ to solve the problem,
Constraints (5.14) should be removed for saving the CPU time. The solution obtained by the
model 𝐹𝑀2′ without Constraints (5.14) can be sorted manually for the optimal solution by
adjusting the chronological order for the repeats of each service, as discussed in Section 5.3.3.

In Section 5.3.3, we have proposed two linear models 𝐹𝑀2′ and 𝐹𝑀2′′ for the nonlinear
model 𝐹𝑀2. Here, we test which linear model has a higher efficiency to derive solutions for
the problem. As we have verified that the model 𝐹𝑀2′ without Constraints (5.14) can be
solved faster, the comparison is conducted between this case of the model 𝐹𝑀2′ and the model
𝐹𝑀2′′ . Table 5.3 illustrates the comparison between the model 𝐹𝑀2′ and 𝐹𝑀2′′ . Both linear
models are valid for the linearization of the nonlinear model as the optimal solutions are
obtained in all instance groups. Whereas, the model 𝐹𝑀2′′ can be solved much faster than the
model 𝐹𝑀2′ by CPLEX. The ratio of CPU times between two models is 0.23 on average,
which shows the advantage of using the concavity of 𝐺𝑟 (𝑥𝑟 ) for the linearization. In a
technical perspective of CPLEX, the model 𝐹𝑀2′ spends too much CPU time on pre-solving
the problem, and the nodes explored in CPLEX for two models are more or less the same,
shown by “B&B nodes”.

Chapter 5: Cruise Ship Service Planning 126

 Table 5.2: Comparison between the model 𝑭𝑴𝟐′ with and without Constraints (14)

Without the
Instance With the constraints Comparison
constraints
# of candidate Instance
𝑍𝑤 (US$) 𝑇𝑤 (s) 𝑍𝑜 (US$) 𝑇𝑜 (s) 𝑇𝑜 / 𝑇𝑤
service ID
2_20_1 1.156 35 1.156 16 0.46
2_20_2 1.093 47 1.093 18 0.38
20 2_20_3 1.136 29 1.136 15 0.52
2_20_4 1.084 33 1.084 22 0.67
2_20_5 1.217 24 1.217 11 0.46
2_40_1 1.312 218 1.312 29 0.13
2_40_2 1.311 56 1.311 20 0.36
40 2_40_3 1.313 90 1.313 38 0.42
2_40_4 1.349 47 1.349 21 0.45
2_40_5 1.304 124 1.304 43 0.35
2_80_1 1.392 236 1.392 39 0.17
2_80_2 1.378 504 1.378 57 0.11
80 2_80_3 1.446 378 1.446 38 0.10
2_80_4 1.405 870 1.405 67 0.08
2_80_5 1.407 573 1.407 88 0.15
Average: 0.32
Note: (i) “# of candidate service” column denotes the total number of candidate services. (ii) “ 𝑍𝑤 ”
and “ 𝑍𝑜 ” columns list the optimal profits under two cases with the unit of ten million US dollars. (iii)
“ 𝑇𝑤 ” and “ 𝑇0 ” columns show the CPU time (seconds) to solve the problem.
Table 5.3: Comparison between the two linear models

Compariso
Instance The first model The second model LP-relaxation
n
# of B&B B&B
Instanc 𝑍𝑓 𝑇𝑓 𝑇𝑠
candidat node 𝑍𝑠 ($) node 𝑇𝑠 /𝑇𝑓 𝑍𝑙 Gap
e id ($) (s) (s)
e service s s
3_20_1 1.182 19 1 1.182 4 1 0.21 1.185 0.23%
3_20_2 1.154 20 1 1.154 5 1 0.25 1.158 0.37%
20 3_20_3 1.127 13 1 1.127 5 1 0.38 1.131 0.36%
3_20_4 1.147 12 1 1.147 2 1 0.17 1.150 0.26%
3_20_5 1.167 18 1 1.167 5 1 0.28 1.173 0.55%
3_40_1 1.321 28 162 1.321 6 141 0.21 1.323 0.19%
3_40_2 1.308 21 41 1.308 7 60 0.33 1.312 0.31%
40 3_40_3 1.294 33 453 1.294 10 407 0.30 1.299 0.37%
3_40_4 1.313 28 79 1.313 7 83 0.25 1.317 0.32%
3_40_5 1.301 54 179 1.301 8 303 0.15 1.306 0.39%
3_80_1 1.398 134 593 1.398 17 537 0.13 1.400 0.13%
3_80_2 1.387 64 83 1.387 15 100 0.23 1.389 0.13%
80 3_80_3 1.412 85 154 1.412 14 294 0.16 1.414 0.17%
3_80_4 1.405 101 317 1.405 20 357 0.20 1.407 0.12%
3_80_5 1.394 88 177 1.394 16 326 0.18 1.396 0.16%
Average: 0.23 1.291 0.27%

Chapter 5: Cruise Ship Service Planning 127

Note: (i) “ 𝑍𝑓 ” and “ 𝑍𝑠 ” columns list the optimal profits of two linear models with the unit of ten million
US dollars. (ii) “ 𝑇𝑓 ” and “ 𝑇𝑠 ” columns show the CPU time (seconds) to solve the problem. (iii) “B&B
nodes” shows the number of nodes explored by CPLEX. (iv) “LP-relaxation” shows the objective value
𝑍𝑙 of the LP solution obtained by LP-relaxation of the model and the objective gap (𝑍𝑙 − 𝑍𝑠 )/𝑍𝑠 with
the optimal solution. Two linear models have the same LP solution.

5.5.3 Performance of myopic approaches

In this section, we aim to validate the effectiveness of the model 𝐹𝑀2′′ (the second linear
model for the model 𝐹𝑀2) and investigate the performance of the two myopic approaches
proposed in Section 5.4.3 for the CSP problem. In both rules, the decision is made on a daily
basis from Day 1 to Day 𝑇. Based on different preferences in two heuristic rules and the berth
availability of each day, an optimal cruise service is selected to operate for the day.
The comparisons between the model and two myopic rules are presented in Table 5.4. It
shows Myopic Rule_1 is better than Myopic Rule_2 as more profit can be earned in the
majority of the instances. However, it does not mean Myopic Rule_1 is good enough for the
cruise ship to plan cruise services. There is still 5.23% optimality gap on average between
Myopic Rule_1 and the optimal solution of the model 𝐹𝑀2′′ , which validates the effectiveness
of the model and addresses the importance of having the optimization-based service planning
tool.

Table 5.4: Comparison between the model 𝑭𝑴𝟐′′ and two myopic rules

Instance 𝑭𝑴𝟐′′ Myopic Rule_1 Myopic Rule_2

# of candidate Instance
𝑍𝑚 (US$) 𝑍𝑓 (US$) Gap 𝑍𝑠 (US$) Gap
service ID
4_20_1 1.209 1.125 7.41% 1.124 6.99%
4_20_2 1.123 1.062 5.75% 1.061 5.51%
20 4_20_3 1.156 1.033 11.93% 1.082 6.35%
4_20_4 1.174 1.112 5.59% 1.097 6.58%
4_20_5 1.220 1.168 4.41% 1.124 7.81%
4_40_1 1.327 1.253 5.90% 1.219 8.10%
4_40_2 1.310 1.267 3.45% 1.222 6.74%
40 4_40_3 1.277 1.224 4.36% 1.189 6.88%
4_40_4 1.326 1.272 4.21% 1.179 11.03%
4_40_5 1.305 1.231 6.09% 1.208 7.46%
4_80_1 1.413 1.359 3.94% 1.231 12.83%
4_80_2 1.382 1.335 3.46% 1.217 11.96%
80 4_80_3 1.380 1.322 4.44% 1.244 9.92%
4_80_4 1.395 1.344 3.86% 1.255 10.08%
4_80_5 1.402 1.353 3.62% 1.230 12.26%
Average: 5.23% 8.70%
Note: (i) “ 𝑍𝑚 ” column lists the optimal profit of the model with the unit of ten million US dollars. (ii)
“Gap” columns show the optimality gap between the model and the myopic rule, which are calculated
by (𝑍𝑚 − 𝑍𝑓 )⁄𝑍𝑚 and (𝑍𝑚 − 𝑍𝑠 )⁄𝑍𝑚 respectively.

5.5.4 Robustness tests for a real case

In this section, we take Quantum of the Seas as our targeted cruise ship for some robustness
tests. Quantum of the Seas is a cruise ship for Royal Caribbean International (RCI). As the

Chapter 5: Cruise Ship Service Planning 128

lead ship of the Quantum class of cruise ships, Quantum of the Seas has a large capacity to
carry 4180 cruise passengers for double occupancy and 4905 for maximum occupancy. The
deadweight of this cruise ship is near 168,666 tons. Currently, this cruise ship is designated in
the Asian area with the home port Shanghai (China). According to the schedule published by
Royal Caribbean International (Cruise route: Quantum of the Seas, 2016), there are 13 cruise
services operated by this cruise ship in the Year 2016, and all the cruise services are a loop
with the home port. The information of these 13 cruise services are given in Table 5.5 and the
locations of the port cities visited by those cruise services are shown in Figure 5.2.

Table 5.5: Information on the cruise services

Cruise Ticket Rotation

Cruise route
Index price time
1 Shanghai(1)→Hiroshima(3)→Tokyo(5)→Kobe(6)→Shanghai(9) $1,110 9 days
2 Shanghai(1)→Busan(3)→Fukuoka(4)→Shanghai(6) $745 6 days
3 Shanghai(1)→Nagasaki(3)→Busan(4)→Shanghai(6) $610 6 days
4 Shanghai(1)→Kumamoto(3)→Shanghai(5) $762 5 days
5 Shanghai(1)→Seoul(3)→Shanghai(5) $732 5 days
6 Shanghai(1)→Kumamoto(3)→Miyazaki(4)→Shanghai(6) $1,296 6 days
7 Shanghai(1)→Inchon(3)→Shanghai(5) $561 5 days
8 Shanghai(1)→Busan(3)→Shanghai(5) $671 5 days
9 Shanghai(1)→Busan(3)→Sakaiminato(4)→Shanghai(6) $761 6 days
10 Shanghai(1)→Busan(3)→Nagasaki(4)→Shanghai(6) $595 6 days
11 Shanghai(1)→Busan(3)→Fukuoka(4)→Nagasaki(5)→Shanghai(7) $610 7 days
12 Shanghai(1)→Okinawa(3)→Shanghai(5) $610 5 days
13 Shanghai(1)→Busan(3)→Nagasaki(4)→Shanghai(6) $701 6 days
Note: (i) the numbers inside the brackets indicate the index of the day when the cruise ship moors in
the port cities, for example, Hiroshima(3) suggest that the cruise ship moors in Hiroshima on Day 3.

Figure 5.2: Locations of the port cities

For decision makers of a cruise ship, a challenge of implementing our model is to estimate
the marginal profit of operating a cruise service accurately. Usually, as the money spent by

Chapter 5: Cruise Ship Service Planning 129

cruise passengers during the cruising is uncertain, the operating profit cannot be finalized until
a cruise service is finished. However, we have involved three input parameters for the
estimation of the operating profit of a cruise service, which are the number of possible cruise
passengers 𝑛, the average profit per cruise passenger 𝑝 and the cruise passenger decreasing
ratio for the cruise repeat 𝑎. However, the parameters 𝑝 and 𝑎 could be hard to be estimated
accurately by cruise companies. Thus, we conduct two robustness tests on these two
parameters to see how many profits will be lost compared with the optimal total profit if the
two parameters are estimated inaccurately.
The robustness test for the 𝑝 is conducted in the following ways: firstly, we set the 𝑝 for
each cruise service based on the assumption in Section 5.5.1. By implementing the model, we
can obtain an optimal solution (i.e., optimal cruise service operation plan, denoted as Plan 𝐴)
for the current setting of 𝑝. Then, assuming that after operating cruise services, it turns out that
we estimate the 𝑝 with 𝑒 estimation error (𝑒 is a input parameter ratio, and 𝑒 ∈ (0,1)) for all
cruise services, among which 𝑢 cruise services are underestimated (𝑢 is also a input parameter
ratio, and 𝑢 ∈ (0,1)) and 1 − 𝑢 cruise services are overestimated. Thus, for the cruise services
underestimated, the real average profit per cruise passenger 𝑝̂ = (1 + 𝑒) × 𝑝. For the cruise
services overestimated, the real average profit per cruise passenger 𝑝̂ = (1 − 𝑒) × 𝑝. With all
the 𝑝̂ of the cruise services and Plan 𝐴, we can calculate the total profit (denoted as 𝑍𝑟𝑒𝑎𝑙 ) that
the cruise ship earned in real. Lastly, supposing that we can estimate all the parameters
accurately at the beginning (based on all the 𝑝̂ ), we implement our model again for the optimal
total profit (denoted as 𝑍𝑜𝑝𝑡𝑖𝑚𝑎𝑙 ) that could be earned by the cruise ship. The gap between
𝑍𝑟𝑒𝑎𝑙 and 𝑍𝑜𝑝𝑡𝑖𝑚𝑎𝑙 is the optimality gap calculated by (𝑍𝑜𝑝𝑡𝑖𝑚𝑎𝑙 − 𝑍𝑟𝑒𝑎𝑙 )⁄𝑍𝑜𝑝𝑡𝑖𝑚𝑎𝑙 , which is
also the percentage of the profit lost due to the inaccurate estimation.
The procedure for the robustness test for 𝑎 (the cruise passenger decreasing ratio for the
cruise repeat) is the same as the robustness test for 𝑝. For the robustness test, we have two
testing input parameters, which are 𝑢 (underestimate ratio) and 𝑒 (estimate error). The
underestimate ratio indicates both the percentage of the cruise services underestimated 𝑢 and
the percentage of the cruise services overestimated 1 − 𝑢. The estimate error suggests the
deviation of our estimation from the real situation. For each combination of 𝑢 and 𝑒, we
conduct ten random instances. The average optimality gap obtained from the ten instances is
taken as the output parameter for the two testing input parameters.
The robustness test on the average profit per cruise passenger for cruise services is
illustrated in Figure 5.3. In the test, we set the estimation error 𝑒 from 0.04 to 0.20 with 0.04
interval, and the underestimate ratio 𝑢 from 0.10 to 0.90 with 0.10 interval. In the figure, all
the optimality gaps are less than 2.0% with the estimation error less than 0.16, which implies
that near-optimal solutions can be obtained the estimation error is less than 16%. Figure 5.3

Chapter 5: Cruise Ship Service Planning 130

also shows that the optimality gap would increase when the estimation error increase (see any
five bars with a same underestimate ratio). However, there is an interesting phenomenon: for
the same estimation error (see any five bars with a same colour), 0.50 underestimate ratio (i.e.,
a half cruise services underestimated and a half cruise services overestimated) dominates the
optimality gap. This phenomenon provides the cruise company with a useful hint: when
estimating the marginal profits of cruise services, the cruise company should use the same
method rather than use different methods to estimate the marginal profits of different cruise
services. Using different methods for the cruise services could be more likely to cause the half-
underestimate-half-overestimate result.

Figure 5.3: The robustness test on 𝒑

Figure 5.4: The robustness test on 𝒂

Chapter 5: Cruise Ship Service Planning 131

 The results of the robustness test on the cruise passenger decreasing ratio for the cruise
repeat 𝑎 are consistent with the results of the former robustness test. The robustness test on 𝑎
is shown in Figure 5.4, where we set the estimation error 𝑒 from 0.02 to 0.10 with 0.02 as the
step, and the underestimate ratio 𝑢 from 0.10 to 0.90 with 0.10 as the step. Figure 5.4 shows
that the optimality gaps are less than 2% when the estimation error is less than 0.08, which
shows our model could derive near-optimal solutions (optimality gap less than about 1.5%) as
long as the estimation error on the 𝑎 is less than 8%. Meanwhile, Figure 5.4 also shows that
0.50 underestimate ratio could bring the most profit lost for the cruise ship, which further
emphasizes the importance of the aforementioned hint.
The two above robustness tests demonstrate the robustness of our proposed model. Even if
some input parameters cannot be estimated accurately by the decision makers, our models can
still obtain a near-optimal solution for the CSP problem, so long as the estimation errors can
be controlled in reasonable ranges. The reason why the model has the robustness in the sense
of error estimations on 𝑝 and 𝑎 can be explained as follows: the two error terms actually
determine the error estimation on the marginal profits of services. However, in the
optimization of the model, the berth availability plays the dominant role rather than the
marginal profits. We can image that the daily operating profits 𝑔𝑟 /𝑠𝑟 of cruise services are not
significantly different from each other, as the cruise services with extremely low daily
operating profits cannot be candidate services. However, the berth availability is significantly
different among the cruise services, especially for some cruise services that have many ports
of call. Thus, we can arbitrarily conclude that the prior optimization is to ensure that the cruise
ship is operated for as many as possible days in the planning horizon with considering the
berth availability.

5.5.5 Sensitivity analysis for the berth availability

In this section, based on 13 real candidate cruise services given by Table 5.5, we further
conduct some sensitivity analysis for the berth availability as the berth availability plays the
dominant role in the optimization. In Section 5.5.1, we have assumed that the cruise terminal
of each port city has randomly 40% to 50% days left for having available berths and generated
the berth availability scenario for each port of call accordingly. Here, we define five different
berth availability scenarios, labeled by BA1, BA2, BA3, BA4, and BA5, by changing the
percentages of the days left for having available berths. For BA1, we decrease the percentages
by 20%; for BA2, we decrease the percentages by 10%; for BA3, we keep the percentages
unchanged; for BA4, we increase the percentages by 10%; for BA5, we increase the
percentages by 20%. From BA1 to BA5, the probability that each port of call has available
berths increases. Ten random instances are generated for each berth availability scenario and

Chapter 5: Cruise Ship Service Planning 132

are solved by the proposed model. The average results of the random instances are reported in
Figure 5.5, Figure 5.6, and Table 5.6, and are analyzed below.
Upper bound obtained by the
Upper bound
greedy algorithm in each berth BA1 BA2 BA3 BA4 BA5 Eq.(9) obtained by Eq. (5.9)
availability scenario
40.0

35.0

30.0

25.0

20.0

15.0

10.0

5.0

0.0
C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13

Figure 5.5: Upper bound 𝒙𝒓 obtained by the greedy algorithm and Eq. (5.9)

Figure 5.5 reports 𝑥𝑟 , upper bound of the number of repeats for Service 𝑟 in one planning
horizon. Eq. (5.9) exhibits a way to approximate 𝑥𝑟 . In Proposition 5, we proved that the
greedy algorithm can obtain the optimal 𝑥𝑟 . Here, we report 𝑥𝑟 obtained by the greedy
algorithm under the five berth availability scenarios and by Eq. (5.9) are given in Figure 5.5.
Note that the 𝑥𝑟 obtained by Eq. (5.9) is constant under different berth availability (given by
𝑇
Bar Eq. (5.9)), as it is derived by ⌊𝑠 ⌋, and the 𝑥𝑟 obtained by the greedy algorithm is different
𝑟

under different berth availability scenarios (given by five bars from BA1 to BA5). Therefore,
each candidate cruise service (indexed by C1 to C13) in Figure 5.5 contains six bars. As can
be seen, the upper bound 𝑥𝑟 obtained by Eq. (5.9) is much worse than that of the greedy
algorithm, especially when the berth availability is low (BA1). For example, for Cruise service
13, the 𝑥𝑟 obtained by Eq. (5.9) is more than ten times as large as that of the greedy algorithm
under the berth availability scenario BA1. Thus, to implement our proposed model, the greedy
algorithm should be applied to approximate 𝑥𝑟 .
Figure 5.6 illustrates the average number of repeats of each cruise service under each berth
availability scenario. In general, cruise services 4, 5 and 6 outperform other cruise services
with a higher average number of repeats. This is due to the fact that those cruise services have
comparatively higher marginal profits and shorter rotation times (cf. Table 5.5). Cruise service
11 is the least selected cruise service to be operated, especially when the berth availability is
low. The cruise service with the longest rotation time (cruise service 1) also performs badly in
BA1, but the performance improves when the berth availability increases and around 1.5

Chapter 5: Cruise Ship Service Planning 133

repeats of the cruise service 1 are operated in berth availability BA2 to BA5 (shown by the
last four bars in “C1” of the figure) for the sake of its high marginal profit.

Average number of repeats for a

cruise service under each berth BA1 BA2 BA3 BA4 BA5
availability scenario
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13

Figure 5.6: Number of repeats of cruise services under different berth availability scenarios

Table 5.6 shows the effects of different berth availability scenarios on major outputs of the
model. When the berth availability increases, operation days and total profit rise
simultaneously. However, the increase of operation days or total profit does not keep the pace
with the increase of the berth availability. For instance, from BA3 to BA5, the berth
availability grows by 20%, but the total profit increases by 9.8%. Average profit per day shares
the same trend as the total profit because the number of total days is constant. By comparison,
the average profit per operation day keeps nearly unchanged when the berth availability
fluctuates.

Table 5.6: Outputs of the model under different berth availability scenarios

Average Average
Total Operation Deviation Total Deviation
ID profit per profit per
days days 1 profit 2
day operation day
BA1 180 127.8 −18.1% 8,568,362 −19.1% 47,602 67,043
BA2 180 146.4 −6.2% 9,839,851 −7.1% 54,666 67,186
BA3 180 156.0 0.0% 10,588,128 0.0% 58,823 67,901
BA4 180 168.2 7.8% 11,363,633 7.3% 63,131 67,577
BA5 180 172.8 10.8% 11,622,052 9.8% 64,567 67,266
Note: (i) “Total days” shows the length of one planning horizon. (ii) “Operation days” indicates the
number of days that the cruise ship operates cruise services, i.e., the cruise ship is traveling. (iii)
“Deviation 1” lists the deviation of the operation days between the corresponding berth availability
and BA3. (iv) “Deviation 2” lists the deviation of the total profit between the corresponding berth
availability and BA3.

Chapter 5: Cruise Ship Service Planning 134

5.6 CONCLUSION
This paper addresses a CSP problem that plans cruise services for a cruise ship, in order to
maximize the total profit during a planning horizon. Considering the fact that major cruise
terminals have limited berths, the berth availability is incorporated into the planning. Then,
the problem also considers the phenomenon that the marginal profit of operating a cruise
service would decrease gradually when a cruise service is repeated several times. To solve the
problem, a nonlinear programming model is built, for which two linearization methods are
suggested. By conducting computational experiments, we find that the linearization method
using the concavity of 𝐺𝑟 (𝑥𝑟 ) could improve the efficiency on solving the problem
significantly. Some properties of the problem in different assumptions are also investigated.
In particular, if there is only one candidate cruise service for the problem, a greedy algorithm
can derive the optimal solution. The effectiveness of the proposed models is verified by
extensive numerical experiments. Lastly, based on real-world cases, the robustness tests are
conducted to show that if there are some parameters needed by the model cannot be estimated
accurately, the proposed model has its robustness and can still obtain a near-optimal plan.
Sensitivity analysis is also conducted for the berth availability to see its effects on some outputs
of the model.
This study also contains limitations. For example, this study assumes all the candidate
services’ home port is identical. This assumption holds in the majority of real situations.
However, when a cruise ship is repositioned to a new region, the candidate services for the
ship may have more than one home port. This case may be more common for some cruise
ships that are operated globally. For the cruise service planning problem under the context of
multiple home ports, the models in this study need to be extended. Another challenge
embedded in this extension may lie in that the repositioning cost between two home ports
should be taken into account. Meanwhile, if a set of candidate cruise services are not available
at first hand, the CSP problem is more complicated as the priority is to design profitable
candidate cruise services. In addition, although we have claimed that service planning is
independent among different cruise ships in the first section, two cruise ships can interact with
each other if have some common ports of call in their candidate cruise services or itineraries.
This is due to that the two cruise ships might compete for an available berth of a common port
in a day. Thus, a joint optimization should be designed for such an interaction, especially when
the two cruise ships belong to one cruise line corporation. If we plan to extend our problem to
multiple ships, our model may be revised from a variant of knapsack problem to a variant of
bin packing problem. The main techniques and results in this paper shall be extended and
applied as well, because we can potentially use some decomposition algorithms to decompose
the extended problem into subproblems of individual ships, and each of them corresponds to

Chapter 5: Cruise Ship Service Planning 135

the problem studied in this paper. All of the above issues will be the research directions for
our future studies.

5.7 REFERENCES
Arrival Time, 2016. Arrival Time—Wusong Kou Cruise Terminal.
http://www.wskcruise.com/article/48.html. Accessed 22 February 2016.

Arthur, H., Ronald, T.W., 2000. Stochastic prediction in multinomial logit models.
Management Science 46(8), 1137–1144.

Bell, M.G.H., Liu, X., Rioult, J., Angeloudis, P., 2013. A cost-based maritime container
assignment model. Transportation Research Part B 58, 58–70.

Bell, M.G.H., Liu, X., Angeloudis, P., Fonzone, A., Hosseinloo, S.H., 2011. A frequency-
based maritime container assignment model. Transportation Research Part B 45 (8),
1152–1161.

Bretthauer, K., Shetty, B., 2002. The non-linear knapsack problem - algorithms and
applications. European Journal of Operational Research 138, 459–472.
Cruise Industry, 2015. Cruise industry—Statista Dossier: Facts and statistics on the cruise
industry. http://www.statista.com/study/11547/cruise-line-industry-statista-dossier/.
Accessed 26 December 2015.

Cruise Market Watch, 2016. Growth of the Cruise Line Industry.

http://www.cruisemarketwatch.com/growth/. Accessed 22 February 2016.

Cruise route: Quantum of the Seas, 2016. Royal Caribbean itinerary 2016 cruises.
http://www.wskcruise.com/tours /. Accessed 22 February 2016.

Cruise Ship Schedule, 2016. Carnival Vista itinerary 2016 cruises.

http://www.cruiseshipschedule.com/carnival-cruise-lines/carnival-vista-schedule/.
Accessed 8 January 2016.

Cruise Ship Statistics, 2015. Cruise Industry Statistics—Cruise Ship Statistics.

http://www.cruiseindustrynews.com/cruise-industry-stats.html. Accessed 26 December
2015.

Esteve-Perez, J., Garcia-Sanchez, A., 2014. Cruise market: Stakeholders and the role of ports
and tourist hinterland. Maritime Economics and Logistics 17(3), 1–18.

Fransoo, J.C., Lee, C.Y., 2013. The critical role of ocean container transport in global supply
chain performance. Production and Operations Management 22(2), 253–268.

Giallombardo, G., Moccia L., Salani, M., Vacca, I., 2010. Modeling and solving the tactical
berth allocation problem. Transportation Research Part B 44(2), 232–245.

Chapter 5: Cruise Ship Service Planning 136

Gui, L., Russo, A.P., 2011. Cruise ports: a strategic nexus between regions and global lines—
evidence from the Mediterranean. Maritime Policy & Management 38(2), 129–150.

Hongmin, L., Woonghee, T.H., 2011. Pricing multiple products with the multinomial logit and
nested logit models: concavity and implications. Manufacturing & Service Operations
Management 13(4), 549 –563.

Iris, C., Pacino, D., Ropke, S., and Larsen, A., 2015. Integrated berth allocation and quay crane
assignment problem: Set partitioning models and computational results. Transportation
Research Part E 81, 75–97.
Kameshwaran, S., Narahari, Y., 2009. Nonconvex Piecewise Linear Knapsack Problems.
European Journal of Operational Research 192, 56–68.
Li, B., 2011. The multinomial logit model revisited: A semi-parametric approach in discrete
choice analysis. Transportation Research Part B 45(3), 461 –473.

Maddah, B., Moussawi-Haidar, L., El-Taha, M., Rida, H., 2010. Dynamic cruise ship revenue
management. European Journal of Operational Research 207(1), 445–455.

Meisel, F., Bierwirth. C., 2009. Heuristics for the integration of crane productivity in the berth
allocation problem. Transportation Research Part E 45(1), 196–209.
Meng, Q., Wang, S., 2012. Liner ship fleet deployment with week-dependent container
shipment demand. European Journal of Operational Research 222(2), 241–252.

Meng, Q., Wang, S., Andersson, H., Thun, K., 2014. Containership routing and scheduling in
liner shipping: overview and future research directions. Transportation Science 48 (2),
265–280.

Meng, Q., Wang, T., Wang, S., 2012. Short-term liner ship fleet planning with container
transshipment and uncertain container shipment demand. European Journal of
Operational Research 223(1), 96–105.

Ng, M.W., 2014. Distribution-free vessel deployment for liner shipping. European Journal of
Operational Research 238(3), 858-862.

Ng, M.W., 2015. Container vessel fleet deployment for liner shipping with stochastic
dependencies in shipping demand. Transportation Research Part B 74, 79-87.

Pang, K.W., Liu, J., 2014. An integrated model for ship routing with transshipment and berth
allocation. IIE Transactions 46(12), 1357–1370.

Paat, R., Huseyin, T., 2012. Robust assortment optimization in revenue management under the
multinomial logit choice model. Operation Research 60(4), 865–882.

Poirriez, V., Yanev, N., Andonov, R., 2009. A hybrid algorithm for the unbounded knapsack
problem. Discrete Optimization 6, 110–124.

Chapter 5: Cruise Ship Service Planning 137

Premoli, A., 1986. Piecewise-linear programming: The compact (CPLP)
algorithm. Mathematical Programming 36(2), 210–227.

Rodrigue, J., Notteboom, T., 2013. The geography of cruises: Itineraries, not destinations.
Applied Geography 38, 31–42.

Soriani, S., Bertazzon, S., Cesare, F.D., Rech, G., 2009. Cruising in the Mediterranean:
structural aspects and evolutionary trends. Maritime Policy & Management 36(3), 235–
251.

Song, D.P., Dong, J., 2012. Cargo routing and empty container repositioning in multiple
shipping service routes. Transportation Research Part B 46 (10), 1556–1575.

Song, D.P., Dong, J.-X., 2013. Long-haul liner service route design with ship deployment and
empty container repositioning. Transportation Research Part B 55, 188–211.

Song, D. P., Li, D., Drake, P., 2015. Multi-objective optimization for planning liner shipping
service with uncertain port times. Transportation Research Part E 84, 1–22.

Statista, 2015. Passenger capacity of the global cruise industry from 2007 to 2017.
http://www.statista.com/statistics/301582/passenger-capacity-of-the-global-cruise-
industry/. Accessed 28 Aug 2015.

Tang, L., Sun, D., Liu, J., 2015. Integrated storage space allocation and ship scheduling
problem in bulk cargo terminals. IIE Transactions 48(5), 428-439.

Vacca, I., Salani, M., Bierlaire, M., 2012. An Exact Algorithm for the Integrated Planning of
Berth Allocation and Quay Crane Assignment. Transportation Science 47(2), 148–161.
Veronneau, S., Roy, J., 2009. Global service supply chains: An empirical study of current
practices and challenges of a cruise line corporation. Tourism Management 30, 128–139.

Veronneau, S., Roy, J., Beaulieu, M., 2015. Cruise ship suppliers: A field study of the supplier
relationship characteristics in a service supply chain. Tourism Management Perspectives
16, 76–84.

Wang, Y., Meng, Q., Du, Y., 2015. Liner container seasonal shipping revenue management.
Transportation Research Part B 82, 141–161.

Zhen, L., Chew, E.P., Lee, L.H., 2011. An integrated model for berth template and yard
template planning in transshipment hubs. Transportation Science 45(4), 483–504.
Zhen, L. 2015. Tactical berth allocation under uncertainty. European Journal of Operational
Research 247, 928–944.

Chapter 5: Cruise Ship Service Planning 138

Chapter 6: Conclusions

This thesis investigates four problems in maritime logistics and operations, where the first
two problems are related to container ships that transport cargos and other two problems are
related to cruise ships that transport passengers. The first problem focuses on the ship type
decision considering empty container repositioning and foldable containers. In this study,
given the ship type with a certain capacity, the problem transfers to a nonstandard minimum
cost flow problem. Then, a network flow model for the problem is formulated. When
considering both standard containers and foldable containers, trouble arises in the network
construction that is some parallel arcs share the same capacity restriction. To overcome this
trouble, a revised network simplex algorithm that changes the standard pivot operation is
designed. Based on the algorithm, a solution approach is proposed to solve the optimization
problem. Some useful managerial implications of this study are obtained after conducting real-
case experiments, which mainly includes three perspectives, i.e., ship type decision, foldable
container usage, and container devanning time.
The second problem addresses the optimal reefer slot conversion for container freight
transportation. This study presents an algorithm to search for the optimal number of reefer
slots to have on a container ship. It is assumed that all the relevant parameters (e.g. freight
rates, storage costs, etc.) are already known. To optimize the sequence of ships deployed, a
revised two-stage simulation approach is proposed. Based on this, a slot conversion algorithm
that determines the optimal slot configurations of the ships is formulated, which embeds the
two-stage simulation algorithm for string optimization. In this study, to highlight the
effectiveness of our approach, several real shipping routes operated by CMA CGM are
investigated. The overall results reveal that the algorithm is highly efficient and can help
shipping liners to significantly improve their profits.
The third problem concerns the cruise itinerary schedule design. It aims to determine the
optimal sequence of a given set of ports of call and the arrival and departure times at each port
to maximize the monetary value of the utility minus the fuel cost. To solve the problem, it first
enumerates all sequences of ports of call and then optimizes the arrival and departure times at
each port of call by developing a dynamic programming approach. To improve the
computational efficiency, effective bounds on the profit of each sequence of ports of call are
proposed. The computational experiments show that, first, the proposed bounds on the profit
of each sequence of ports of call can considerably improve the computational efficiency.
Second, the total profit of the cruise itinerary is sensitive to the fuel price and hence, it is
acceptable to use the shortest voyage distance method to design the schedule when the ports

Chapter 6: Conclusions 139

of call have a naturally geographical distance. In contrast, determining the sequence of ports
of call solely by minimizing the overall voyage distance frequently leads to a significant
reduction in the total profit when the ports do not have a naturally geographical sequence.
The last problem investigates the cruise service planning, which plans cruise services for a
cruise ship, in order to maximize the total profit during a planning horizon. Considering the
fact that major cruise terminals have limited berths as like the container terminals (Zhen wet
al., 2016; Wang et al., 2018), the berth availability is incorporated into the planning. The
problem also considers the phenomenon that the marginal profit of operating a cruise service
would decrease gradually when a cruise service is repeated several times. To solve the
problem, a nonlinear programming model is built, for which two linearization methods are
suggested. By conducting computational experiments, it is founded that the linearization
method using the concavity of the objective function could improve computational efficiency
significantly. Some properties of the problem in different assumptions are also investigated.
In particular, if there is only one candidate cruise service for the problem, a greedy algorithm
can derive the optimal solution. Based on real-world cases, the robustness tests are conducted
to show that if there are some parameters needed by the model cannot be estimated accurately,
the proposed model has its robustness and can still obtain a near-optimal plan. Sensitivity
analysis is also conducted for the berth availability to see its effects on outputs of the model.
Based on the above studies, three future research directions are worthwhile to be further
explored. (i) Integrating the optimizations: although the first two problems (container ship type
decision and container reefer slot conversion) or the other two problems (cruise itinerary
schedule design and cruise ship service planning) are treated in isolation, they actually interact
with each other. For instance, in the cruise ship service planning problem, we take the cruise
itinerary schedules exogenously given. However, if we integrate the cruise itinerary schedule
design and cruise ship service planning, we can expect more profits compared with the original
two-stage optimizations since the schedule design can take the inconvenience of the service
planning into account. (ii) Considering the information ambiguity: We now study the problems
by supposing that we have full data information, for example, in the container ship type
decision problem, we assume that the container transportation demand is known and is
deterministic. However, in practice, there may be high uncertainty on the information.
Therefore, we need to obtain robust solutions by designing more solid solution approaches in
response to the information ambiguity. (iii) Extending the problems to generalized cases: The
first two problems of container shipping are based on a single shipping route. It is worthwhile
to extend the problems to consider the whole shipping network of a shipping liner that may
involve many shipping routes. The other two problems of cruise shipping are based on a single
cruise ship. If we consider that, a cruise line owns a fleet of cruise ships. It would be interesting
to extend the problems to the ones that also address the cruise ship fleet management.

Chapter 6: Conclusions 140

Appendices

Appendix A: Specification of standard containers and foldable containers

Table A.1 shows the major specifications of standard containers and foldable containers,
which are almost the same. The specifications of standard containers are collected from APL
(2017) and the specifications of foldable containers are collected from Holland Container
Innovations (2017).

Table A.1: Specifications of foldable and standard containers

Description Standard containers Foldable containers
Cubic capacity 67.7 cubic meters 72.9 cubic meters
Maximum payload 26,760 kg 26,600 kg
Gross weight 30,480 kg 32500 kg
External length 12.192 m 12.192 m
External width 2.438 m 2.438 m
External height 2.591 m 2.896 m
Internal length 12.032 m 12.012 m
Internal width 2.352 m 2.324 m
Internal height 2.392 m 2.615 m
Door opening width 2.340 m 2.172 m
Door opening height 2.280 m 2.508 m
Bundle (4 into 1) height — 2.896 m

Appendix B: Proof for the results of Figure 2.10

Informal, we can prove the non-increasing trend by using a simplified mathematical model.
Assuming 𝑋 represents the vector for the number of standard containers in ports and 𝑌
𝑌𝑒 𝑇
represents the vector for the number of foldable containers in ports. Then, 𝜌 = 𝑋𝑒 𝑇 +𝑌𝑒 𝑇 shows

the percentage of foldable container usage, where 𝑒 = {1, … ,1} and 𝑒 𝑇 is the transposition of
𝑒. Given the defined vector variables 𝑋 and 𝑌, we can use the following simplified standard
model to represent the formulation of our problem.

𝑀𝑖𝑛 𝐶1 𝑋 + 𝐶2 𝑌

𝑠. 𝑡. 𝐴1 𝑋 + 𝐴2 𝑌 = 𝐵

𝑋, 𝑌 ≥ 0

where all coefficient matrixes or vectors (𝐶1 , 𝐶2 , 𝐴1 , 𝐴2 , 𝐵) are positive. In the next, we can
derive:

𝑋 = 𝐴1−1 𝐵 − 𝐴1−1 𝐴2 𝑌

By substituting it to the objective, we have,

Appendices 141
 𝑀𝑖𝑛 𝐶1 𝐴1−1 𝐵 − 𝐶1 𝐴1−1 𝐴2 𝑌 + 𝐶2 𝑌

Based on which, if the cost coefficient 𝐶2 for foldable containers increase, 𝑌 will decrease. As
1 𝑋𝑒 𝑇 +𝑌𝑒 𝑇 𝑋𝑒 𝑇
a result, 𝑋 will increase. As we have 𝜌 = 𝑌𝑒 𝑇
= 𝑌𝑒 𝑇 + 1, the increasing of cost coefficient
1
𝐶2 will lead to the increasing of 𝜌, that is the decreasing of 𝜌. Such a proof verifies the result

shown in Figure 2.10.

Appendices 142
References

APL, 2017. The perfect fit for all your different needs: your APL equipment guide,
https://www.apl.com/wps/wcm/connect/659ca9b6-7eb3-4c16-82de-
8de04abab7e9/Equipment+guide.pdf?MOD=AJPERES. Accessed on 10 August 2017.

Holland Container Innovations, 2017. 4FOLD foldable container.

http://hcinnovations.nl/4fold-foldable-container/. Accessed on 10 August 2017.

Wang, K., Wang, S., Zhen, L., & Qu, X., 2016. Cruise shipping review: operations planning
and research opportunities. Maritime Business Review 1(2), 133-148.

Wang, K., Wang, S., Zhen, L., & Qu, X., 2017. Cruise service planning considering berth
availability and decreasing marginal profit. Transportation Research Part B:
Methodological 95, 1-18.

Wang, K., Zhen, L., Wang, S., & Laporte, G., 2018. Column generation for the integrated
berth allocation, quay crane assignment, and yard assignment problem. Transportation
Science 52(4), 812-834.

Xia, J., Wang, K., & Wang, S., 2019. Drone scheduling to monitor vessels in emission control
areas. Transportation Research Part B: Methodological 119, 174-196.

Zhen, L., Wang, S., & Wang, K., 2016. Terminal allocation problem in a transshipment hub
considering bunker consumption. Naval Research Logistics 63(7), 529-548.

References 143