IT University of Copenhagen

Ph.D. Thesis

Models and Algorithms for

Container Vessel Stowage
Optimization

Author: Supervisor:
Alberto Delgado Assoc.Prof. Rune Møller Jensen

December 22, 2013

 To my wonderful wife, Annemarie

2
 Acknowledgments

The writing of this dissertation and the culmination of my graduate studies have been
possible thanks to the help, guidance, encouragement, and support of several wonderful
people. First, I would like to thank my supervisor Rune Møller Jensen for giving me
the opportunity to pursue a Ph.D. under his supervision. His constant support and
constructive criticizing made the time I spent doing the Ph.D. and writing this thesis
a very enriching experience. He also introduced me to the maritime shipping domain
and showed me the importance of doing applied research, for which I am grateful.
I specially thank Dario Pacino and Kevin Tierney, my friends and colleagues at the
Decision Optimization Lab, with whom I had several insightful discussions during the
past few years. Though I shared my office only with Kevin and Dario, they were not
my only companions on this journey. I would also like to thank Peter Sestoft and all
the other professors at the former SDG (Software Development Group) for insightful
discussions on how to do research and the different ways to present it, and to all the
Ph.D. students at SDG that like me felt the frustrations of missed deadlines, rejected
papers, and bad reviews, but that also felt the satisfaction that doing research brings.
Special thanks to Kosta, Helge, Mohammed, Juan, Joel, Josu, Paolo, and Hugo.
I also owe a big part of the successful culmination of this journey to my family.
Annemarie, my wonderful wife, the one that with her love, patience and support filled
with confidence and determination those moments full of doubts. My parents, my sister,
and relatives that have always support me in the pursuing of my dreams, even when
those would put an ocean between us, and to my family in law, that have helped me
make of Denmark my new home. Special thanks also go to my colleagues from the
Avispa research team in Colombia, where I started tasting the bitter sweetness of doing
research, to Ange Optimization and the Ministry of Science, Innovation and Higher
Education for partially funding my Ph.D., Maersk Line for providing us with valuable
data and the domain knowledge needed to pursue our research goals, and to the IT
University for all the support that provided me to finish my Ph.D.

3
4
 Abstract

Containerized seaborne trade has played a key role in the transformation of the global
economy in the last 50 years. In liner shipping companies, at the heart of this operation,
several planning decisions are made based on the stowage capabilities of container
vessels, from strategic decisions (e.g., selection of vessels to buy that satisfy specific
demands), through to operational decisions (e.g., selection of containers that optimize
revenue, and stowing those containers into a vessel). This thesis addresses the question
of whether it is possible to formulate stowage optimization models, with acceptable
trade-offs between accuracy and scalability, that can be used to assist solving planning
problems in liner shipping companies. We consider two planning contexts to answer
this question: Stowage Planning (SP) and the Cargo Composition Problem (CCP).
SP is the process of deciding where each container of those to be loaded in a port
should be placed in a vessel, i.e., to generate stowage plans. This thesis explores two
different approaches to solve this problem, both follow a 2-phase decomposition that
assigns containers to vessel sections in the first phase, i.e., master planning, and that
stows the containers assigned by the master planning to each vessel section and into
vessel slots in the second phase, i.e., slot planning. The first SP approach automatically
generates representative stowage plans. This dissertation presents an accurate math-
ematical programming model for master planning that includes features of stowage
planning that have not been considered in previous work. For slot planning, a fast and
accurate representative model to optimally stow vessel sections is introduced.
The second SP approach serves as the optimization component of a commercial decision
support tool used for interactive planning of container vessels. Expert’s know-how for-
mulated as user preferences is integrated into the heuristic optimization component and
used to tackle complex constraints and optimize combinatorial objectives. According to
our experimental evaluation, stowage plans computed by our heuristic are competitive
enough with respect to those made by experts under the same conditions.
The CCP evaluates how the stowage characteristics of containers with different
features affect important performance measures used in liner shipping companies, e.g.,
vessel intake and cargo revenue. This dissertation provides the first problem description
and introduces the first mathematical model with variable displacement. Specifically,
a model that assigns containers to vessel slots, and which is actually solvable on real
instances. To increase scalability, a master planning model with variable displacement
to solve the CCP is also presented. Analysis on vessel intake and revenue optimization
are performed on single and multi-port scenarios for a benchmark suite of ten vessels
facilitated by our industrial collaborators.

5
6
Sets and Constants

Sets and constants common to more than one model

Sets
B Bonjean stations, excluding the station closest to stern.
F Frames.
L Locations.
P Scheduled ports to be visited by a vessel.
T Container types.
TR Transports, where each transport hpi , pj i ∈ P 2 is represented by a
load-discharge port pair.
T RpON Transports hpi , pj i with containers on the vessel at departure (pi ≤ p < pj )
from port p.
U Ballast tanks.
Constants
Aα{l,u,b}f Fore or aft distance from frame f to the longitudinal center of gravity
of location l, ballast tank u, and buoyancy section between bonjean stations b and b + 1.
DlL Lcg of location l.
V
Dpl Vcg of location l at port p.
B
Db Distance in meters between bonjean stations b and b + 1 multiplied by
water density.
{T,V,L}
Du Transversal (T ), vertical (V ), and longitudinal (L) center of gravity of
ballast tank u.
α
Kpl Capacity of 20 ’ and 40 ’ containers in location l at port p.
V
Kpl Volume capacity in location l at port p.
TEU
Kpl TEU capacity of location l at port p.
{RF ,RC }
Kpl Reefer plugs (RF ) and reefer cells (RC) available in location l
at port p.
Ku Ballast water capacity of tank u.
LMpO Longitudinal moment of the empty vessel including constant weights and release
at port p.
LDtτ Load-discharge matrix: number of containers of type τ on transport t.
M in{GM } Lower bound for metacentric height (GM ).
{S,B}
M inf Lower bound for shear force (S) and bending moment (B) at frame f .
{S,B}
M axf Upper bound for shear force (S) and bending moment (B) at frame f .
Gα{l,u,b}f Fraction of location l, ballast tank u, and buoyancy section between
bonjean stations b and b + 1 that lies fore or aft of frame f .
Gf Fore-based fraction of frame f , where Gf ∈ [0; 1]. Gf = 1 when f
is the first frame at bow, and Gf = 0 when f is the first frame at stern.
V MpO Vertical moment of the empty vessel including constant weights and release

7
 at port p.
α
Wpl Weight capacity of 20 ’ and 40 ’ containers in location l at port p.
WpO Constant weight of the vessel including release at port p.
WfS,α Constant weights fore or aft of frame f .
WfB,α Bending components of the constant weights fore or aft of frame f .
Wτ Weight in metric tons of type τ .
ΠTEU
τ TEUs needed to stow a container of type τ , i.e., ΠTEU
τ ∈ {1, 2}.
ΠFEU
τ FEUs needed to stow a container of type τ , i.e., ΠFEU
τ ∈ {0.5, 1}.
Φτ Volume required for type τ .

Sets and constants of slot planning

Sets
Classes Stack classes.
Classi Stacks of class i.
M Tuple set of release containers and their corresponding cells indices:
{(j, k, i)| j ∈ J, k ∈ Kj , i ∈ QL }.
J Stacks.
Q Containers to stow.
Q {V ,L,R} Virtual (V ), release (L), and reefer(R) containers.
Q 40 {A,F } Aft40 and Fore40 40 ’ containers. Q40 A and Q40 F have the same cardinality.
Q {20R,40R} 20 ’ and 40 ’ reefer containers.
Q ¬R Non-reefer 20 ’ and 40 ’ containers excluding virtual containers.
Q{20 ,40 } 20 ’ and 40 ’ containers.
S Slots.
S {A,F } Aft and Fore slots.
Sj Slots in stack j.
{A,F }
Sj Aft and Fore slots in stack j.
S {R,¬R} Reefer (R) and non-reefer (¬R) slots.
S ¬RC Slots of location l in cells with no reefer plugs.
S {20 ,40 } Slots of location l with 20 ’ and 40 ’ capacity.
Kj Cells in stack j.
Constants
Aip Indicates whether container i is unloaded at port p. It is 1
if i is unloaded in port p, 0 otherwise.
HiC Height in meters of container i.
HjS Height limit in meters of stack j.
H {DC,HC} Height in meters of a normal (DC) and a high-cube (HC) container.
LC
i Length of container i.
PiC Discharge port of container i.
Q P=p,W =w Containers with discharge port p and weight w to be stowed in location l.
Q {NC ,HC } Normal (NC ) and high-cube (HC ) containers to be stowed in location l.
RiC Indicates whether container i is reefer.
Rjk Number of reefer plugs in cell k of stack j.
WiC Weight in kilograms of container i.
WjS Weight limit of stack j in kilograms.
Cv Penalization for a container overstowing containers below.
Cp Penalization for each different discharge port stowed in a stack.
Cu Penalization for stowing containers in a stack.

8
 Cr Penalization for stowing a non-reefer container in a reefer slot.

Sets and constants of master planning

Constants
AW
{M,T,Bs} Weight coefficient for the linearization of metacenter (M ), trim (T ),
and bonjean at station s (Bs).
AL
{M,T,Bs} Lcg coefficient for the linearization of metacenter (M ), trim (T ),
and bonjean at station s (Bs).
A{M,T,Bs} Constant term for the linearization of metacenter (M ), trim (T ),
and bonjean at station s (Bs).
AW{DA,DF } Weight coefficient for the linearization of aft draft (DA) and fore draft (DF ).
AL{DA,DF } Lcg coefficient for the linearization of aft draft (DA) and fore draft (DF ).
A{DA,DF } Constant term for the linearization of aft draft (DA) and fore draft (DF ).
Eu Initial condition of tank u.
M in{T,DA} Lower bound for trim (T ) and aft draft (DA).
M ax{T,DA,DF } Upper bound for trim (T ), aft draft (DA), and fore draft (DA).
Wp Approximated constant displacement at port p.
Wtτ Weight in metric tons of type τ during transport t.
u Allowed ballast change for tank u.

Sets and constants of heuristic optimization

Sets
LO Locations with odd slots.
Lv Locations in bay v.
Lt Locations specified in the planning scenario to stow containers on transport t.
TR lp Transports leaving from port p with destinations that have not been specified
in the planning scenario as valid for location l.
V Vessel bays.
Constants
A¬R
pl Average volume gained by stowing a single 20 ’ non-reefer container in a
non-reefer odd slot at port p.
ARpl Average volume gained by stowing a single 20 ’ container in a reefer odd slot
at port p.
Clτ Profit for stowing a container of type τ in location l.
C HC Penalization for going over high-cube killing limit.
C OD Penalization for not using odd slots when there are 20 ’ available.
40 V
Kpl Volume capacity for stowing 40 ’ in location l at port p.
¬RO
Kpl Non-reefer odd slots in l at port p.
OR
Kpl Reefer odd slots in l at port p.
O
Kpl Odd slots in location l at port p.
KtO Odd slots in locations specified in the planning scenario to stow containers of transport t (Lt ).
KtO20 Odd slots in Lt that we know in advance cannot P be used
due to lack of 20 ’ containers, i.e., max(0, KtO − τ ∈T 20 LDtτ ).
K
Kpl High-cube killing limit at port p: high-cube containers that can be stowed in l

9
 such that its slot capacity can still be completely utilized.
40 N
Kpl Volume capacity for stowing 40 ’ in stacks with no odd slots of location l at port p.
τ
Kpl Containers of type τ that can be stowed in location l at port p.
u
Wpv Upper weight limit specified in the planning scenario for bay v at port p.

Sets and constants of cargo composition

Sets
C Cells, i.e., C = {(i, j, l)|i ∈ Tlj , j ∈ Jl , l ∈ L}.
Cl Cells in location l ∈ L.
Clj Cells in stack j and location l.
C {¬R,¬20 ,¬40 } Cells with no reefer slots ( ¬R), no 20 ’ capacity (¬20 ), and
no 40 ’ capacity (¬40 ).
Cα¬R Cells with no aft or fore reefer slot capacity.
Cc− Cells below cell c ∈ C.
E Aft and Fore slots of a cell, i.e., E = {A, F }.
Eτ Slots in a cell where containers of type τ can be stowed,
i.e., {e | (τ ∈ T 20 ∧ e ∈ E) ∨ (τ ∈ T 40 ∧ e = A)}.
I Displacement intervals.
Jl Stacks in location l.
LON Locations on deck.
LUl Locations below location l.
Ln Locations in bin n ∈ N .
N Bins of adjacent bays.
{A,OV }
T Rp Transports hpi , pj i with containers loading or discharging (pi = p or pj = p)
at port p (A), or going over (pi < p < pj ) port p (OV).
Constants
AW
{M,T,Bs} (Wi ) Weight coefficient of displacement interval i for the linearization of
metacenter (M ), trim (T ), and bonjean at station s (Bs).
AL
{M,T,Bs} (Wi ) Lcg coefficient of displacement interval i for the linearization of
metacenter (M ), trim (T ), and bonjean at station s (Bs).
A{M,T,Bs} (Wi ) Constant term of displacement interval i for the linearization of
metacenter (M ), trim (T ), and bonjean at station s (Bs).
c+ /c− Cell on top/below cell c ∈ C. Let c = c+ /c− iff c is the top/below cell
of the stack.
CpOP Overstowage probability penalization at port p.
Ctτ Profit for stowing containers of type τ ∈ T and transport t ∈ T R.
{O,C,U }
Cp Overstowage (O), Makespan (C), and ballast water (U ) penalization at
port p ∈ P .
{V,T }
Dc Vertical (V ), Transversal (T ) center of gravity of cell c ∈ C.
Hτ Height in meters of type τ .
Hlj Height limit in meters of stack j in location l.
HC
Kpl High-cube capacity in location l at port p.
McS BigM value for stack overstowage calculation: twice the number of slots
below cell c.
{H,U }
Ml BigM value for hatch overstowage calculations: twice the number of slots in
location l (H) and twice the number of slots in locations below l (U ).
M axV Maximum vertical moment possible for a vessel.
M axL L
i /M ini Maximum and minimum longitudinal center of gravity at displacement interval i.

10
T MpO Transversal moment of the empty vessel including constant weights and release
at port p.
Wli40 40 ’ weight capacity of stack i ∈ Jl in location l ∈ L.
20
Wlie 20 ’ weight capacity of section e ∈ E in stack i ∈ Jl location l ∈ L.
{−,+}
Wi Lower (−) and upper (+) bound of displacement interval i.
Wi Average weight of displacement interval i.

11
12
 CONTENTS

Contents

1 Introduction 15
1.1 Thesis Question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.4 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5 Document Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Container Stowage Planning 25

2.1 Containers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Container Vessels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2.1 Physical Structure of a Container Vessel . . . . . . . . . . . . . 28
2.2.2 Vessel Stability and Stress Limits . . . . . . . . . . . . . . . . . 30
2.3 Maritime Container Terminals . . . . . . . . . . . . . . . . . . . . . . . 34
2.4 Container Stowage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Stowage Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3 Slot Planning 41
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 The Integer Programming (IP) Model . . . . . . . . . . . . . . . . . . . 45
3.5 Global constraint modeling . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6 The Constraint Programming (CP) Model . . . . . . . . . . . . . . . . 49
3.6.1 Symmetry-Breaking and Implied Constraints . . . . . . . . . . . 52
3.6.2 Branching Strategies . . . . . . . . . . . . . . . . . . . . . . . . 53
3.6.3 Lower Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.7 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.7.1 Restowing Location Stowage Plans . . . . . . . . . . . . . . . . 57
3.7.2 Slot Planning Master Plans . . . . . . . . . . . . . . . . . . . . 62
3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

13
CONTENTS

4 Accurate Master Planning 65

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Stability and Stress Model with Ballast Tanks . . . . . . . . . . . . . . 67
4.3.1 A Mixed Integer Programming (MIP) Model . . . . . . . . . . . 69
4.4 Analysis of Model Accuracy . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 Heuristic Optimization for Interactive Stowage Planning 79

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3 Angelstow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.4 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.5 Heuristic Optimization Approach to Stowage Planning . . . . . . . . . 85
5.5.1 LP Model for Master Planning . . . . . . . . . . . . . . . . . . . 85
5.5.2 Greedy Algorithm for Slot Planning . . . . . . . . . . . . . . . . 89
5.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

6 Cargo Composition Problem 97

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.4 Complete Model for the Cargo Composition Problem (CCP) . . . . . . 100
6.5 Decomposed model for the CCP . . . . . . . . . . . . . . . . . . . . . . 107
6.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.6.1 Single Port Experiments . . . . . . . . . . . . . . . . . . . . . . 111
6.6.2 Multi-Port Experiments . . . . . . . . . . . . . . . . . . . . . . 120
6.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

7 Conclusions 125
7.1 Outlook and Future Directions . . . . . . . . . . . . . . . . . . . . . . . 126

Bibliography 129

14
 Chapter 1. Introduction

Chapter 1

Introduction

Modern shipping techniques have become the heart of globalized trade. They have
transformed business models around the world by connecting new suppliers and con-
sumers thousands of kilometers away from each other. The mechanization of sea trans-
port replacing expensive human labor with more efficient equipment, and the devel-
opment of larger and more efficient ships to achieve economy of scales have been key
factors in this transformation. Equally or more importantly, however, has been the
introduction of containerization: the transportation of cargo in standard sized boxes
known as containers. Containerization provided a standard interface between the differ-
ent parties involved in the intermodal freight transport chain, allowing the development
of an integrated transport system.
Seaborne trade accounts for over 80% of the global trade by volume and over 70%
by value. Containerized trade has grown steadily, by almost 14 times its original vol-
ume between 1980 and 2011 (UNCTAD (2012)), reaching 15.8% of all cargo traded by
volume and 52%1 by value in 2011. Though the world economy is still in recovery, the
growth in seaborne container trade is expected to continue, driven mostly by increasing
demand for manufactured goods in emerging economies and by containerization of un-
conventional commodities, e.g., scrap steel and recycled paper (UNCTAD (2012)). The
majority of seaborne containers are transported by container vessels, cellular vessels
specifically designed to carry containers. These vessels are operated by liner shipping
companies 2 , with a fleet capacity that accounts for 97.2% of the total active capacity
deployed on liner trades (Alphaliner (2013)).
Even with the magnitude of their operation, liner companies face a highly competi-
tive market. The liner shipping sector, for instance, is far from being consolidated. By
January 2012, almost 70% of the Twenty-foot Equivalent Unit (TEU)3 capacity was in

1
From the website of the World Shipping Council (WSC) at www.worldshipping.org. Accessed
May 2013.
2
Liner shipping companies, carrier, container carriers, and liner companies will be used interchange-
ably in this thesis.
3
Measure used in the shipping industry for volume capacity. One TEU is equivalent to the dimen-
sions of a standard 20 ’ container.

15
Chapter 1. Introduction

the hands of the top twenty shipping companies, with the three largest companies con-
trolling almost 30% (UNCTAD (2012)). This considerable number of relevant players,
together with the low segmentation of the market and other external factors (e.g., the
financial crisis in 2009), has increased the need for carriers to be more competitive. To
do so, different strategies have been devised. One strategy has been to reduce trans-
portation costs per TEU by increasing vessel capacity (i.e., economies of scale), with
vessels reaching sizes of up to 18,270 TEU4 . This approach, however, is limited by the
demand on the trade routes, since the vessels must be filled to a certain capacity to
achieve savings by economies of scale. Furthermore, the financial crises of 2009 reduced
the demand for freight transport, and with most of the carriers attempting to reduce
costs by acquiring larger vessels, the market was flooded with capacity. By 2013, how-
ever, part of the extra capacity has been absorbed by the implementation of the slow
steaming policy. This policy aims at reducing bunker consumption by having container
vessels sailing at lower speeds. This speed reduction forces the addition of new vessels
to the trade lanes in order to maintain the same service frequency. Though this policy
has provided some relief to the over capacitated market it has not absorbed all the
extra capacity available, keeping freight rates at a very low level.

In another strategy to be more competitive, the liner shipping companies tried to

optimize the different planning problems they face. Traditionally, planning problems
in the liner shipping industry can be classified as strategic, tactical, or operational,
depending on their planning horizon (Christiansen et al. (2007)). Figure 1.1 depicts
the three planning levels and some of the planning problems associated with each of
them. At the strategic level, decisions involving major economical investments and
commitments are made. These decisions cover a medium to long-term horizon that
oscillates between one to five years, but which can be extended by five to ten years,
depending on the decision (designing/buying a vessel, building a container terminal).
The decisions at this level shape the tactical and operational decisions to be made in
the future. Four of the most important strategic decisions are: selecting trade lanes
the carrier wishes to operate, selecting target market share for each trade lane defining
the size of the operation the carrier wants, choosing the fleet mix and size necessary
to satisfy the demand of the trade lanes and market share chosen by the carrier, and
choosing main ports and transshipment hubs to call in the trade lanes previously chosen
(Kjeldsen (2012)).
Tactical decisions cover a medium term horizon, spanning from two to three months
to up to a year. There are three important decisions made at the tactical level: route
design, where the carrier decides sequences of ports to call in order to cover the trade
lanes chosen at the strategic level; vessel deployment, where vessels are assigned to
each route in order to satisfy the forecasted demand; and scheduling, where the carrier
assigns a time to call each port to the vessels deployed in the different routes (Andersen
(2010)).

4
Capacity of the Triple-E class, to be delivered in June 2013

16
 Chapter 1. Introduction

Strategic Tactical Operational

- Selection of trade - Route design - Vessel transition
lanes
- Vessel deployment - Disruption
- Target market management
share - Scheduling
- Sailing route
- Fleet mix and size selection

- Port choice - Revenue

management

- Container stowage

Figure 1.1: Planning levels and their activities in liner shipping companies. The arrows
indicate the direction of influence of each level.

Finally, operational decisions cover a short term horizon, spanning from a few hours
to a few months. Decisions at this level either relate to container vessels or to cargo.
The most common operational decisions related to vessels are: vessel transitions, such
as deciding when and how vessels will move to a new schedule from an old one; dis-
ruption management, where the vessel route or its sailing speed is modified in order to
cope with delays in the schedule; and sailing route selection, where the sailing route and
speed is decided based on weather factors and sea currents in order to reduce the cost
of the trip, (Christiansen et al. (2007)). With respect to operational decisions related
to cargo, the most common decisions are revenue management and container stowage.
In revenue management, route managers decide what kind of cargo is more profitable
to transport on a route based on the available capacity of the vessels currently sailing,
yield analysis, and forecast demand. They provide this information to the sales de-
partment, who will take cargo based on this analysis. Then, capacity managers decide
which containers to load onto the vessel from those taken by the sales department. This
is based on the revenue they generate and their shipping priority. In container stowage,
on the other hand, stowage coordinators decide where to stow the containers that the
capacity manager has decided to take onto the container vessel.

Among many of the decisions made when solving planning problems in a liner ship-
ping company, a common denominator is the need to consider container vessels and
their TEU capacity. For instance, at the strategic level, the carrier purchases the ves-
sels with the capacity and other features needed to satisfy the demand for the trade
lanes they plan to operate. At a tactical level, vessels are deployed onto routes based
on their capacity and other features in order to satisfy the demand of each route. At
the operational level, route managers segment the cargo based on its yield, expected
demand, and on how much of the demand the vessels will be able to transport. More-
over, capacity managers accept and reject containers based on the current capacity of

17
Chapter 1. Introduction

the container vessels.

Most of these decisions are usually made under the assumption that a container
vessel can be filled up with containers of any kind, and that they can be freely stowed
in the vessel. This assumption, however, is far from reality. Stowing containers in a
vessel is a complex problem. High level constraints force containers to be assigned to
vessel sections in such a way that stability and stress forces are kept within limits, while
low level stacking rules and constraints force container stacks to follow valid patterns
and to be within capacity limits. Thus, in a number of planning problems in the
shipping industry, having a clear picture of the viability of stowing cargo compositions
in a container vessel is crucial. Inaccuracies at this level lead to a sub-optimal utilization
of the fleet by over/under-estimating its capabilities, keeping carriers from achieving
economies of scale, and missing opportunities in the market by not satisfying demands.
It is highly relevant for shipping companies to have at hand accurate stowage models
that provide them with stowage footprints of container vessels, allowing them to perform
different kinds of optimizations. Figure 1.2 shows some of the planning problems where
accurate stowage modelling and optimization can be used.

Strategic Tactical
- Fleet mix and size - Vessel deployment

Stowage Modelling

Operational
- Revenue management
- Container stowage

Figure 1.2: Planning decisions where stowage modelling can be used.

Previous work
The academic community has been interested in stowage modelling since the 1970s.
Their work, however, has mainly focused on stowage planning, i.e., the process of gen-
erating stowage plans that assign containers to vessel slots. Academic work can be
divided into two main categories: single-phase and multi-phase approaches. Single-
phase approaches have a plain representation of the stowage planning problem, (e.g.,
stow containers into vessel slots). A common denominator among these approaches
is to sacrifice model accuracy to achieve scalability. For instance, Botter and Brinati
(1992), introduced an IP model of the first accurate but intractable formulation of the
stowage planning problem, and more recently, Li et al. (2008), introduced an IP model
to solve a yet more simplified version of the stowage planning problem. Multi-phase

18
 Chapter 1. Introduction

approaches decompose the problem hierarchically into two or more layers of abstraction
(e.g., first assign containers to vessel sub-sections, then stow them into vessel slots),
and they are currently the most successful in terms of model accuracy and scalability.
Wilson and Roach (1999), introduced a model that includes several major aspects of
stowage planning in a 2-phase approach, resembling the workflow of stowage coordina-
tors. Other approaches inspired by similar decompositions include the recent work of
Ambrosino et al. (2010), and Pacino et al. (2011).
Though on a smaller scale than in stowage planning, there is also academic work
attempting to use stowage modelling and optimization to make planning decisions on
other levels of the shipping companies. Most of the work has focused on revenue
management (e.g., Zurheide and Fischer (2012); Feng and Chang (2008); Ting and
Tzeng (2004)), with models that select which containers to load on container vessels
deployed on a route, based on the revenue they generate. These models, however,
consider very basic volume and weight capacity constraints for containers with a few
different features, disregarding all other aspects of stowage.
Academic work on stowage modelling and its application for liner shipping is con-
siderable; its impact, however, has been limited. The stowage of container vessels is
a problem full of corner cases and subtle details. The lack of a strong industrial col-
laboration has made it difficult for academics to access detailed information on the
vessel designs of key importance for the formulation of accurate models. It has also
been difficult to define representative problems that include the most relevant features
of a stowage model identified by the industry. This has led to most of the published
approaches to have a trade-off between accuracy and scalability that is not practical
for the industrial setup of a stowage model. Either execution times running too high
or problem abstractions miss several key features of the problem.

1.1 Thesis Question

From a long term collaboration with a liner shipping company we have devised a set
of key features that define the level of accuracy and scalability stowage models must
have to be used in an industrial setup. Nonetheless, even when only considering the key
features defined with our industrial collaborator, solving stowage models in a scalable
manner remains a challenging problem. Moreover, as mentioned before, most of the
research on stowage modelling focuses on stowage planning, with almost no academic
work on using accurate stowage modelling for other planning decisions in liner shipping
companies. The few of these decisions for which optimization models have been in-
vestigated, e.g., revenue management, consider over simplified stowage constraints that
reduce their applicability. Thus, the challenges in stowage modelling that this thesis
addresses can be summarized in the question below:
Can stowage optimization models, with an acceptable trade-off between accu-
racy and scalability, be formulated to assist solving different planning prob-

19
Chapter 1. Introduction

lems in liner shipping companies?

1.2 Approach
This thesis addresses the above research question by formulating stowage optimization
models that support two different planning contexts within a liner shipping company:
stowage planning and the cargo composition problem. The acceptable trade-off be-
tween accuracy and scalability is defined by our industrial collaborators and depends
on the industrial setup where the stowage models are meant to be deployed within the
company.
Stowage planning is the most straightforward application of stowage modelling in
liner shipping companies. This thesis investigates two scalable stowage approaches
to generate stowage plans that follow a 2-phase hierarchical decomposition, similar
to that of the most successful approaches (Wilson and Roach (2000); Kang and Kim
(2002); Ambrosino et al. (2010); Guilbert and Paquin (2010)), resembling the workflow
of stowage coordinators (see Figure 1.3). The first phase of this decomposition, master
planning, assigns containers to load in the port to vessel sections. It considers high-
level constraints and objectives such as stability, vessel sections’ capacities, and the
homogeneous distribution of crane workload along the vessel. The second phase, slot
planning, follows the assignment made in the master planning phase and assigns the
containers to load each vessel section to vessel slots. It considers low level stacking
constraints and rules of thumb to generate robust stowage plans.
Master Planning Phase Slot Planning Phase

Containers Master Master Slot Stowage

Vessel Data Planning Planning
Plan Plan
Port Data

Figure 1.3: 2-phase hierarchical decomposition of stowage planning.

Both of the two stowage approaches studied in this thesis have different trade-
offs between accuracy and scalability, based on the industrial setup where they are to
be deployed. The first approach automatically generates stowage plans and includes
all key aspects of stowage planning defined by our industrial collaborators. Since the
approach solves a representative problem, the stowage plans that it generates cannot be
used directly in practice. Stowage coordinators use these plans as initial solutions that
require few modifications to turn into final stowage plans. With respect to scalability,
according to our collaborator these plans must be generated within a time limit of 10
minutes. The reason being that stowage coordinators have only a few hours to generate
stowage plans and within this short time, they must be able to try out several different
forecast scenarios.

20
 Chapter 1. Introduction

During master planning, containers are assigned to vessel sections by a MIP model
(Pacino, Delgado, Jensen, and Bebbington (2012), Delgado et al. (2013)), on the proviso
that the vessel is seaworthy, i.e., stability and stress forces are within limits, high level
objectives to reduce time at port are optimized, and capacity constraints in each vessel
section are fulfilled. Given that we are solving the master planning phase with a MIP
model, a fractional number of containers may be assigned to vessel sections. Though
it is not possible to divide containers, we show, experimentally, that the impact of this
is negligible. During the slot planning phase, the CP model and Local Search (LS) al-
gorithm presented in Delgado et al. (2012) and Pacino and Jensen (2012), respectively,
stow containers into vessel slots in each vessel section, following the assignment made
by the master planning phase. Containers are stowed into container stacks that follow
specific stacking rules and satisfy capacity constraints, while optimizing objectives aim-
ing at producing robust stowage plans.

The second approach for stowage planning is the optimization component of a com-
mercial decision support tool to stow container vessels called Angelstow5 , (Delgado et al.
(2012)). This approach demands a different trade-off between accuracy and scalability
due to the interactive nature of the decision support tool. Accuracy is reduced in order
to provide fast stowage plans, since the time limit defined for the generation of the plans
is two seconds, on average. The master planning phase of this approach integrates ex-
pert know-how provided by the stowage coordinator into a Linear Programmming (LP)
model that considers a reduced set of key aspects of stowage planning. The know-how
is encapsulated into a planning scenario, where the stowage coordinator provides expe-
rience on how to address stability and stress forces, and reduce time at port. The slot
planning phase is solved by a greedy algorithm that builds container stacks, satisfying
stacking constraints and rules.

The second planning context that we consider in this thesis is the cargo composition
problem. Here we evaluate how the stowage characteristics of containers with different
features affect important performance measures used in liner shipping companies, e.g.,
vessel intake and cargo revenue. We introduced the first complete model that stows
containers into vessel slots and solves real instances. To improve scalability, we present
a model based on the master planning phase of the 2-phase decomposition depicted in
Figure 1.3, to generate fast sub-optimal solutions.
Stowage planning and the cargo composition problem consider the same constraints
with respect to stowing containers and vessel seaworthiness. There is, however, a fun-
damental difference. In stowage planning, the vessel displacement is known in advance
and only small changes are expected to happen, since the loadlist of containers has
been already decided. This allows an efficient implementation of accurate stability and
stress forces calculations. In the case of the cargo composition problem, however, part
of the problem is to determine the containers to be stowed, which makes the vessel dis-

5
Ange Optimization ApS, available for purchase at www.http://ange.dk/main/angelstow.

21
Chapter 1. Introduction

placement variable. We address this issue by providing the proposed models for solving
the cargo composition problem, with the capability to efficiently calculate stability and
stress forces in a variable displacement environment, and at a small performance cost.

1.3 Thesis Contributions

To address the research question posed above, this thesis presents the following five
major contributions:

1. Optimal slot planning

This thesis presents an accurate representative model of slot planning of vessel
sections, and introduces (to our knowledge), the fastest approach to generate
optimal slot plans. The model successfully finds optimal solutions for most of the
industrial instances where it has been tested and was successfully deployed in a
planning module for the automatic generation of stowage plans.

2. Optimization module for interactive stowage planning

This thesis introduces the first heuristic optimization module that incorporates
expert’s know-how into the fast generation of stowage plans. Moreover, the mod-
ule has been successfully integrated into a commercial decision support used for
stowage planning tools and its accuracy tested with industrial data.

3. Accurate scalable master planning

This thesis introduces an accurate optimization model for master planning. It
includes key aspects of stability and stress forces, and the modelling of ballast
tanks, (i.e., tanks loading ballast water used to adjust vessel stability). The
model is tested on industrial data and scales to larger instances than previously
tested in the literature.

4. Variable displacement model for master planning

This thesis introduces the first master planning model with accurate stability and
stress calculations that efficiently considers variable displacement of container
vessels.

5. Cargo composition problem

This thesis introduces a new problem into the liner shipping domain and the first
model with variable displacement that assigns containers to slots, that is actually
solvable on real instances. This model is also the first attempt to use accurate
stowage modelling for decision problems other than stowage planning within a
liner shipping company.

22
 Chapter 1. Introduction

1.4 Publications
1. Delgado, Alberto; Møller Jensen, Rune; Janstrup, Kira; Høyer Rose,Trine and
Høj Andersen, Kent. A constraint programming model for fast optimal
stowage of container vessel bays. European Journal of Operational Research
220, no. 1 (2012): 251-261.

2. Delgado, Alberto; Møller Jensen, Rune and Schulte, Christian. Generating

optimal stowage plans for container vessel bays. Principles and Practice of
Constraint Programming-CP 2009, pp. 6-20. Springer Berlin Heidelberg, 2009.

3. Delgado, Alberto; Møller Jensen, Rune and Guilbert, Nicolas. A placement

heuristic for a commercial decision support system for container vessel
stowage. Conferencia Latinoamericana en Informatica (CLEI), 2012 XXXVIII,
pp. 1-9. IEEE, 2012.

4. Dario, Pacino; Delgado, Alberto; Møller Jensen, Rune and Bebbington, Tom. An
accurate model for seaworthy container vessel stowage planning with
ballast tanks. Computational Logistics, pp. 17-32. Springer Berlin Heidelberg,
2012.

5. Delgado, Alberto; Pacino, Dario and Møller Jensen, Rune. Container ves-
sel stowage planning with ballast tanks. Extended abstract to appear in
proceedings of Second INFORMS Transportation Science and Logistics Society
Workshop, (2013 TSL Workshop)

6. Delgado, Alberto, and Rune Møller Jensen. Cargo composition analysis in

container vessels. To be submitted to the European Journal of Operational
Research

7. Delgado, Alberto; Møller Jensen, Rune and Guilbert, Nicolas. AngelStow: A

commercial optimization-based decision support tool for stowage plan-
ning. Presentation at 2nd International Conference on Computational Logistics
(ICCL11), 2011.

8. Pacino, Dario; Delgado, Alberto; Møller Jensen, Rune and Bebbington, Tom.
Fast generation of near-optimal plans for eco-efficient stowage of large
container vessels. In Computational Logistics, pp. 286-301. Springer Berlin
Heidelberg, 2011.

1.5 Document Outline

Chapter 2 Container Stowage Planning. In this chapter we introduce the reader
to the liner shipping domain, providing the terminology and knowledge necessary

23
Chapter 1. Introduction

to understand the contributions of the thesis. The chapter is based on the knowl-
edge gathered from our industrial collaborators, work experience in the domain,
and the book on ship stability by Barrass and Derrett (2011).

Chapter 3 Slot Planning. In this chapter we introduce a representative problem for

the slot planning of vessel sections and two optimization models for solving this
problem: a CP and an IP model. We present a detailed description of both models
and experimental comparisons between them using industrial instances.
The chapter is based on the papers by Delgado et al. (2009) and Delgado et al.
(2012). Note that the defendant is the main contributor of the papers and was in
charge of the implementation of the CP model.

Chapter 4 Accurate Master Planning. In this chapter we introduce a MIP model

for master planning that includes accurate stability and stress calculations, and
ballast tanks. We propose an approach to tackle the non-linearities introduced
by the variable displacement due to the tanks, and experimentally evaluate the
impact of this approach in the vessel stability and stress forces calculations.
The chapter is based on the paper by Pacino, Delgado, Jensen, and Bebbington
(2012). Note that the defendant made significant contributions to the formulation
of the model, experimental section of the paper, and the implementation of the
models.

Chapter 5 Heuristic Optimization for Interactive Stowage Planning. This chap-

ter presents the optimization component of a decision support tool used for the
interactive generation of stowage plans. The heuristic we introduce here generates
stowage plans within seconds following the recommendations from an expert user,
(stowage coordinator).
The chapter is based on the paper by Delgado et al. (2012). Note that the defen-
dant is the main contributor to the paper and is responsible for the implementation
of the methods.

Chapter 6 Cargo Composition Problem. This chapter presents a model that as-
signs containers to vessel slots, and an extension to the model presented in chapter
4, to generate cargo compositions, optimize vessel intake, and cargo revenue. Both
models allow accurate stability and stress calculations when the vessel displace-
ment is variable.
The chapter is based on the paper Delgado and Jensen (2013). Note that the
defendant is the main contributor to the paper and is responsible for the imple-
mentation of the models.

Chapter 7 Conclusions. This chapter presents the conclusions of the thesis and fu-
ture research directions in stowage modelling and, to wit, its applications in the
liner shipping domain.

24
 Chapter 2. Container Stowage Planning

Chapter 2

Container Stowage Planning

Commercial shipping can be divided into three segments according to the type and
volume of the transported commodities: bulk shipping, liner shipping, and specialized
shipping (Stopford (2009)). Bulk shipping transports bulk homogeneous cargo in big vol-
umes, most of the times raw materials, (e.g., crude oil, iron and grain). Liner shipping
transports different types of general cargo in small volumes. It includes containerized
cargo, palletized cargo, and liquid cargo stored in small tanks, among others. Special-
ized shipping deals with bulk and general cargo in big volumes. Vessels are specially
designed for the kind of cargo they transport. It includes the transportation of motor
vehicles, refrigerated foods, and different types of liquid gases that demand specific
tankers.
In the liner shipping domain, general cargo is loaded into standard size boxes called
containers. Containers are mainly transported in container vessels. Container vessels
are assigned to services, fixed routes of ports where the vessels call and where containers
are loaded and unloaded, (similar to a bus service). A vessel visits each port in a service
following a schedule and must sail even if it is only partially full. Liner shipping with
container vessels is the domain assumed in this thesis.

2.1 Containers
A container is a standardized piece of equipment. Containers have been designed to
facilitate their transfer among different transportation modes, e.g., container ship, train,
truck, etc., without intermediate transfer of their content. They are built to support
standard shipping, storing, and handling conditions so they can be used multiple times.
Their lifespan depends on the conditions they have been exposed to and their usage
level. It ranges from ten to fifteen years, in average.
Different types of containers are available to fulfill specific requirements of cargo.
The most common containers are box-shape 20 ’ or 40 ’ long, and 8’ wide. They are
either 8’6” (standard) or 9’6” (high-cube) high, though it is uncommon to see 20 ’ high-
cube containers. Standard 20 ’ and 40 ’ containers, together with 40 ’ high-cube con-

25
Chapter 2. Container Stowage Planning

tainers, account for almost 90% of the world’s container inventory in circulation, with
20 ’ standard and 40 ’ high-cube containers, depicted in Figure 2.1, being the 30% and
45.7% of the total TEU backlog in 2011, respectively, (Drewry Maritime Research
(2011)). Containers of other lengths are also available, e.g., 10 ’, 45 ’, 48 ’, 53 ’, the most

Corner Casting

9'6''
8'6''

40'

20'
8'

8'

Figure 2.1: Dimensions of the most common containers in the fleet, adapted from Pacino
(2012).

common among these lengths being the 45 ’ container. Most of the container vessels,
however, are designed mainly to handle 20 ’, 40 ’, and to some extent 45 ’ containers.
Nowadays, the kind of cargo that container vessels transport go beyond the capabilities
of standard box-shaped containers, thus, containers for cargo with special requirements
are needed. For instance, cargo that needs constant ventilation, e.g., organic prod-
ucts, use ventilated containers that are provided with small openings for air circulation.
Cargo that is temperature sensitive and needs to be kept above or below freezing point,
e.g., perishable goods, is transported in reefer and insulated containers. Reefer contain-
ers are provided with electric cooling engines and need to be stored close to a power
plug, while insulated containers need to be provided with air delivered to the correct
temperature. Though ventilated containers have more capacity than reefer containers
(no cooling engine), their transportation and storage is more demanding since they need
an external unit that provides them with air. Dry bulk cargo (e.g., grain and seeds),
is transported in bulk containers. These containers are provided with three manholes
at the top to facilitate their loading and with discharge openings in each door wing.
Separated tank container fleets are available for the transport of chemical products and
foodstuff. For odd shaped cargo Out-of-gauge (OOG) containers are available. Among
these containers we find the hardtop containers with removable roof, that can be loaded
by crane and are specially designed for heavy and high loads; the open top containers,
specially designed for overheight cargo; the half height open top 20 ’ containers; the
flat containers with only the front, back, and bottom walls that are designed for heavy
loads and over-width cargo; as well as the platform containers with no walls which are
designed specifically for heavy loads an over-sized cargo. Containers carrying hazardous

26
 Chapter 2. Container Stowage Planning

goods are divided into four IMO risk groups based on their content. Each group has a
set of rules specified in regulations (e.g., Hamburg (2008)), that describe how they can
be stowed, i.e., safety distances to other IMO containers, and in which sections of the
container vessel they may be stowed.

2.2 Container Vessels

Container vessels are cellular ships designed and constructed with the sole purpose of
transporting containers. Since the appearance of the first container vessel, the Ideal-
X1 in 1956, container vessel capacity has been continuously increasing to reach an
astonishing 18,270 TEU in 2013, (Triple-E Class). This growth has been driven by the
principle of economy of scale in the shipping industry, the increasing demand of goods,
and the ever growing number of commodities that can be transported in containers.
Figure 2.2 shows the increase in TEU capacity of container vessels in the last fifteen
years. By 2013, the total active capacity deployed on liner trades around the world was

20000
Triple-E
18000

Emma
16000 Mærsk
Marco
Polo

14000
TEU capacity

12000

10000
Soverign
Mærsk Gudrun
Regina Mærsk
8000 Mærsk

6000 NYK
President Altair
Truman

4000
1985 1990 1995 2000 2005 2010
Year

Figure 2.2: Growing TEU capacity of container vessels from 1985 to 2013

16.88 million of TEU carried by 5,929 ships, with 4,951 container vessels accounting
for 97.2% of this capacity. Container vessels can be classified according to their TEU
capacity and hull dimensions in under the following categories (MAN Diesel & Turbo):

• Small Feeder with up to 1,000 TEU capacity and a beam of 23m. They are
usually used for short-sea transportation

1
A transformed tanker carrying fifty-eight 35 ’ containers.

27
Chapter 2. Container Stowage Planning

• Feeder ranging from 1,001 to 2,800 TEU capacity and beam size between 23 to
30.2m. These vessels feed the bigger container vessels and serve areas with low
container demand

• Panamax with capacity from 2,800 to 5,100 TEU, with a max. beam of 32.3m
and a max. draught of 12m. They were the biggest container vessels produced
until 1988, its size matches that of the Panama canal at that time

• Post Panamax with capacity from 5,500 to 10,000 TEU, with bean sizes from
39.8 to 45.6m. These vessels were the first to exceed the 32.3m beam size limit of
the Panama canal in 1988

• New Panamax with capacity from 12,000 to 14500 TEU, max. beam of 48m,
max. draught of 15.2m, and max. length of 365.8m. These are vessels constrained
by the dimension of the new Panama canal, to be operational in 2015

• ULCV (Ultra Large Container Vessels), with capacity from 14,500 and with
beams greater than 48.8m. These vessels are larger than the new Panama Canal,
to go operational in 2015.

Nowadays the world fleet consists of 39.18% of feeder vessels, 20.76% of Panamax,
20.34% of small vessels, 16.42% of Post Panamax, and 3.3% of ULCV. TEU-wise,
however, the distribution is different. Panamax vessels account for the 26.46% of the
fleet TEU capacity, Post Panamax vessels for the 35.5%, Feeder vessels for the 21.45%,
ULCV for the 12.7%, and Small Feeder vessels for the remaining 3.94%.

2.2.1 Physical Structure of a Container Vessel

The area of a container vessel where containers are stowed is longitudinally divided into
sections called bays, (See Figure 2.3). Bays are divided in two sections, on deck and
48 44 40 36 32 28 24 20 16 12 8 4
(49−47)(45−43) (41−39) (37−35) (33−31) (29−27) (25−23) (21−19) (17−15) (13−11) (9−7) (5−3) (1)
Line of sight

Hatch cover Bay Lashing bridge Waterline

Figure 2.3: The arrangement of cargo space in a container vessel.

below deck, by a number of flat leak-proof structures called hatch covers. Figure 2.4 (a)
shows a hatch cover dividing a bay section. Containers are stowed on top of and below
the hatch cover. The below deck section of a bay is also known as cargo hold. Each
sub-section of a bay consists of a row of stacks one container wide, (See Figure 2.5 (a)).

28
 Chapter 2. Container Stowage Planning

(a) (b)
Figure 2.4: (a) Lateral view of a bay. The hatch cover separates the on deck and below deck
section. Containers are stowed in both sections. The picture also shows the cell guides
below deck. (b) Two sections of the lashing bridge with hatch covers in between. The
picture also shows containers on top of the hatch covers, secured to the lashing bridge.

Stacks On Deck Bow Stack

8 6 4 2 1 3 5 7 Aft Fore

86 40’ High−Cube
Tiers
On
Deck 82
84
4 3 4 40’ Reefer

Tiers
Below
8
6 2 1 2 20’ Reefer 20’
Outer
Deck 4 support

2 20’ Reefer

6 4 2 1 3 5 7 Outer Inner
Stern Stacks Below Deck support supports

(a) (b)
Figure 2.5: (a) A front view of a vessel bay. (b) A side view of a partially loaded stack. Each
power plug represents a reefer slot. Reefer containers are drawn with electric cords.

Stacks either on or below deck are grouped into storage areas called locations. A location
is a set of stacks, not necessarily consecutive, that coincide with the hatch covers. Pairs
of storage areas on the starboard and the port side of the vessel are merged into single
locations to ease stability calculation, (see locations 2 and 4 in Figure 2.5 (a)). Note
that locations are not physically defined, but are abstractions used in practice to aid
the process of stowing containers in the vessel. A stack holds vertically arranged cells
indexed by tiers. Cells are divided into two slots, the Aft (closest to the stern), and
Fore, (closest to the bow). A slot can hold a 20 ’ container and two slots in the same
cell can hold a 40 ’ or a 45 ’ container. Due to the shape of the hull some cells in the
vessel have only a single slot available, these cells are known as odd cells, (See Figure 2.5

29
Chapter 2. Container Stowage Planning

(b)). Some slots have extra features that allow the stowing of containers with special
demands, e.g, slots with electric plugs provide refrigerated containers with electricity
for their engines. Figure 2.5 (b) depicts the layout of a stack.
Stacks have height and weight limits. Two weight limits exist for each stack, one
regarding the outer container supports and one regarding the inner supports. Limits
on the inner supports are often the smallest, as the vessel structure in the middle of
a stack is weaker. The inner supports are used only when 20 ’ containers are stowed,
as depicted in Figure 2.5 (b). When 20 ’ and 40 ’ containers are mixed in the same
stack, only half of the 20 ’ weight is considered to be supported by the outer supports,
since the other half sits on the inner supports. Below deck, cell guides secure containers
transversely, (see Figure 2.4 (a)). Containers on deck are secured to each other, to the
deck, or to the lashing bridge - a fixed steel structure running over deck across the ship
that has been designed to provide container stacks with extra support, allowing them
to stack higher. Figure 2.4 (b) depicts two sections of the lashing bridge separated by
hatch covers.
Ballast tanks are carefully located tanks along the vessel that are used to modify
displacement and stability conditions of the vessel by changing the amount of water
they hold. Some of these tanks are situated on the sides of the vessel and act as heeling
tanks when containers are being loaded and unloaded in port.
Containers in a vessel are identified by the bay, stack and tier of the cell where they
are stowed. The industry uses a special indexing system that allows them to quickly
narrow down the area where containers are stowed. Each physical bay is identified by
three consecutive numerical indices, two odd and one even, (See Figure 2.3). The even
index is used to identify the 40 ’ containers stowed in the bay, while the smallest odd
index identifies 20 ’ containers stowed in the fore section of the bay, and the greatest
odd index identifies 20 ’ containers in the aft section. As depicted in Figure 2.5 (b),
tiers indices are even numbers that, on deck, start with the number 82, while below
deck they start with number 2. Stack indices begin from the middle out, with the even
indices towards starboard and the odd indices towards the port.

2.2.2 Vessel Stability and Stress Limits2

For a vessel to leave port it must be deemed seaworthy. All items loaded, i.e., cargo, bal-
last water, fuel, etc., must be distributed along the vessel such that stability values are
acceptable and the stresses on the vessel structure are within limits. The displacement
of a vessel (its total weight), determines the portion of the hull that is under water. The
water line is the line where the water meets the hull of the vessel, and the plane that
cuts the hull at the water line is called water plane. The design waterline of a vessel is
the water line at the point where the vessel is loaded at its full capacity. The vertical
line passing through the interception of the forward surface of the stem (most forward
part of the vessel), with the design waterline is known as the fore perpendicular, while

2
Inspired by Barrass and Derrett (2011)

30
 Chapter 2. Container Stowage Planning

the aft perpendicular is the vertical line drawn where the design waterline intercepts
with the aft side of the rudder post. The mid point between aft and fore perpendicular
is known as midship, (see Figure 2.6).

Aft Fore
Perpend. Midship Perpend.

Design waterline
Waterline
Draft
Keel

Figure 2.6: Position of perpendicular and midship in a container vessel

Draft is the distance between the keel (bottom of the vessel) and the water line, and
can be calculated at any point of the vessel. When calculated at the perpendiculars it
is called Fore and Aft draft, respectively. There are three requirements for the draft
at the perpendiculars. First, the aft draft of a container vessel must be high enough
for the propeller to be submerged under water. Second, when calling into a port, the
maximum of aft and fore draft must be smaller than the allowed draft at the port to
avoid the keel hitting the bottom. Third, the maximum draft must be high enough to
allow port cranes to work properly.
The trim of a vessel is defined as the difference between its aft and fore draft.
When the trim is positive the vessel is trimming by the stern, and when the trim is
negative the vessel is trimming by the bow. Trim and draft play an important role
in fuel consumption. Draft increases as a consequence of an increment in the vessel
displacement. The heavier the vessel gets, the faster the engine must go to keep up
the speed. With respect to trim, the vessel can be trimmed by the bow to achieve its
optimal trim; that is, a trim where, due to the shape of the hull, sailing of the vessel
will be as efficient as possible.
The center of gravity (G), of a container vessel is the point where the gravitational
force may be taken to act. The center of buoyancy (B), is the center of gravity of the
water displaced by the vessel, and it is where the buoyancy force can be taken to act.
When the vessel is at even keel, as depicted in Figure 2.7 (a), G and B are aligned with
the center line, a straight line midway of the vessel running from bow to stern. When
an external force is applied to the vessel, e.g., wind or waves, it heels and the submerged
section of the hull changes, moving the center of buoyancy in the same direction as the
external force. The intersection of the center line of the vessel with a vertical line that
goes through the new center of buoyancy (B1 ), is the metacenter (M ), (see Figure 2.7
(b)). Since G and B1 are misaligned, the two forces act in opposite directions with the
same magnitude (weight w on G and buoyancy force b on B1 ), and generate the moment
of statical stability, that is, the moment to return the vessel to its original position after

31
Chapter 2. Container Stowage Planning

the external force has disappeared. The lever arm of the moment of statical stability is
known as the righting lever, and it is the transversal distance between the two forces.
The moment is calculated from G, and it is equal to b times the righting lever. It is
assumed that there are no free bodies in the vessel, since they force G to move as well.

M M
M

θ θ

Water line
G Z G
G G1
B b B
B B2 B1
w

Center line

(a) (b) (c)

Figure 2.7: Transversal view of (a) Stable vessel. (b) Vessel perturbed by a external force
(heeling). (c) Vessel perturbed by adding new weight (listing)

For heeling angles up to 15 degrees, we may consider that the buoyancy force acts
vertically upwards to the initial metacenter. From figure 2.7 (b), we define Z to be
the projection of G on B1 M with respect to the transverse axis. The righting lever
GZ can then be calculated as GM sin θ, where θ is the heeling angle, and GM the
distance between the vertical coordinate of the metacenter and the center of gravity,
also known as metacentric height. We can then see the direct relation between the
moment of statical stability and the initial GM for small heeling angles, since as the
GM increases, so does the moment. Vessels with large GM are stiff and resistant to
rolling. Once they start rolling, however, they can achieve considerable speed when
coming back to their original position due to the increased moment of stability. This
is particularly inconvenient for crew and passengers, but also for container vessels that
carry container stacks on deck. On the other hand, a decidedly small or negative GM
is not convenient either, since the vessel will tend to capsize before it rolls back to its
original position. Thus, for a vessel to be stable with respect to its statical stability, its
GM must be within a safety range.
Now imagine a container vessel where the resultant forces acting on it are zero, (Fig-
ure 2.7 (a)). When a body already on the vessel is moved transversally, or a new body
is added such that the center of gravity of the vessel moves transversally, the vessel lists
(see figure 2.7 (c)), to bring its center of gravity transversally under the metacenter.
Stability conditions dictate that a vessel must be at even keel to be seaworthy.

Two main forces constantly act on the structure of a vessel in opposite directions.
One is the gravity force acting downwards. The other is acting upwards and is the
buoyancy force, caused by the water displaced by the submerged part of the hull.
These two forces, both with the same magnitude, are not uniformly distributed along
the vessel. For instance, let us assume that a vessel is divided in longitudinal sections
that are independent and can move vertically, as Figure 2.8 shows, and that each section

32
 Chapter 2. Container Stowage Planning

weights differently. Their weight depends on what has been loaded in the section and
the weight of the vessel structure in the section. Sections with downward arrows weigh
more than the buoyancy they provide, and sink deeper into the water until the amount
of displaced water equals the section’s weight. In sections with upward arrows, the
weight is smaller than the buoyancy, and they raise until the amount of water displaced
matches the section’s weight. Though vessels are not divided into independent sections,
Figure 2.8 illustrates how the weight and buoyancy forces are distributed along the
vessel creating stress forces in the structure of the ship. Limits on these stress forces
are given for a set of points along the vessel called frames.

Figure 2.8: Example of the effects of uneven distribution of weights and buoyancy forces along
different sections of a container vessel.

The shear force on a frame can be approximated by the sum of the resulting vertical
forces acting on the vessel section aft or fore of the frame. The weight of a section is
calculated by adding up the individual weights of all the objects in the section, including
the weight of the vessel structure. The calculation of the buoyancy force, nevertheless,
is more challenging. As previously stated, the buoyancy force is the weight of the water
displaced by the submerged section of the hull. Thus, we can calculate the weight of
the water displaced by calculating the volume of the submerged section when knowing
the density of the water. This volume depends on the current draft of the vessel and
the shape of the hull. Due to the irregular shape of hulls, the ship builder provides a
set of reference points along the vessel called stations, where the immersed area of the
transverse section of the hull is given as a function of draft by the bonjean curve of the
station. For any two consecutive stations, we calculate the immersed volume between
the two stations, and approximate the buoyancy force in that section by averaging the
areas calculated at the current draft, times the distance between the two stations. The
bending moment, that is, the moment forcing the vessel to bend at a specific point of the
vessel, must be within limits at each frame. The bending moment can be approximated
by multiplying each vertical force acting on the vessel section aft or fore of the frame,
with the distance from the frame itself to where the force acts.

33
Chapter 2. Container Stowage Planning

2.3 Maritime Container Terminals

Maritime container terminals are port sections where containers are transshipped be-
tween ships, and between ships and land vehicles. They are divided into three main
sections: the quayside, the seaside section of the terminal where the vessels arrive and
containers are loaded and unloaded; the yard, where containers are temporarily piled
up on long rows of container stacks, as can be seen in Figure 2.9 (a); and the land-
side, where containers are picked up and dropped off by trucks and trains. Containers
are moved between the different sections of the terminal by truck trailers, Automated
Guided Vehicles (AGV), and straddlers, among others. Within the yard they are ma-
nipulated by gantry cranes mounted on rubber tiers (RTG), or rails. Figure 2.9 (a)
shows an RTG moving over a row of containers lying in a terminal yard. On the quay-
side, quay cranes with spreaders carry out the loading and unloading of containers from
the vessel, accessing only the top-most containers in stacks. Figure 2.9 (b) depicts a
container being unload from a container vessel by a quay crane and loaded onto a truck.
The figure shows how a spreader holds a container by its top four corners. Quay cranes

(a) (b)
Figure 2.9: (a) Terminal yard with a front view of a gantry crane. (b) Container being
unloaded from a vessel to a truck by a quay crane in the quayside of a terminal.

can perform between 30-37 moves (loading or unloading of a container), per hour. Most
of the modern quay cranes can do twin-lifting, that is, lift two 20 ’ containers at the
same time, increasing their average moves per hour to around 25 for 20 ’, i.e., lifting 25
pairs of 20 ’ containers per hour, and 30 for 40 ’ containers. Quay cranes are also used
to lift hatch covers, an operation that takes longer since spreaders need to be aligned
for the hatch. The same crane can be assigned to attend multiple bays, and a crane-set,
the operation of changing the crane position, can take around two minutes. Cranes run
on rails that are lying parallel to the vessel and cannot pass each other. There must
be a safety distance, e.g., 40, 60, 80 feet, between two adjacent cranes when they are
working in parallel.
The partition of vessel sections into areas handled by a single crane is known as
crane split. A good crane split minimizes crane makespan, which directly impacts port

34
 Chapter 2. Container Stowage Planning

fees since the carrier has to pay not only per container move but also per hour of
equipment usage.

2.4 Container Stowage

Containers must be loaded on board the ship to reach their destination. This process is
not trivial since containers cannot be freely stowed in the vessel. For a start, containers
must be distributed along the vessel such that seaworthiness is not compromised, (see
section 2.2.2). A vessel that is not seaworthy will not be allowed to leave the port.
Moreover, low level stacking constraints, i.e., container stack capacity limits (weight
and height), that are specified in a vessel profile describing the vessel layout, and stack-
ing rules based on container features (e.g., length, IMO types, Pallet-wide types, etc.),
also come into play. Containers are provided with corner castings that are used to
secure them to the vessel and to other containers. Each of the four bottom corner
castings of a container must rest on the corner castings of other containers or sockets
of the vessel. This implies that 20 ’ containers cannot be stowed on top of containers
of greater length, since they do not have castings in the middle. Extra castings and
frames at approximately 75cm from the back and front door have been installed on 45 ’
containers. This allows 45 ’ containers to be stowed on top of 40 ’ containers. It seems
to be possible as well to do the opposite, stow 40 ’ on top of 45 ’ containers, but this
is not allowed. Containers with special features also follow specific stacking rules. For
instance, there is a special type of pallet-wide container (PW6), that cannot be stowed
side by side in consecutive stacks, and another type (PW3), that requires a normal
container be stowed between two stacks stowing this type of pallet-wide container. In
the case of IMO containers, some of the IMO groups have proximity constraints with
respect to other IMOs. IMO-2 requires one container of separation to any other IMO-2
container, while IMO-3 and IMO-4 have more complex rules.

Stacking rules also depend on the section of the vessel. As described in section
2.2.1, container vessels are divided into on deck and below deck sections. The holds are
provided with cell guides at the aft and fore part of the bay. Containers are stowed and
locked on top of each other by stacking cones that mainly resist vertical forces, but are
strong enough to provide adequate longitudinal and transverse support below deck. It
is important to notice that containers cannot hang in the air, they have to be stowed at
the bottom of a bay or on top of other containers. Odd cells at the bottom of a stack
must be filled up with 20 ’ containers before using the cells on top. It is not possible to
stow single 20 ’ containers on top of stacks since the inward end is not secure. Container
vessels of previous generations had special holds for 45 ’ containers, but nowadays all
45 ’ are stowed on deck.
In the on deck section, twistlocks, connectors designed to withhold stronger vertical
and horizontal forces, replace stacking cones to secure containers to each other or to
the deck. Up to the second tier, twistlocks provide enough support for the containers.

35
Chapter 2. Container Stowage Planning

Beyond the second tier, and up to the fifth, lashing rods are needed to provide extra
support to containers stowed in this area. The rods are tied to the deck and need to be
kept tight all times. They are constantly adjusted during the journey. Containers 45 ’
long are stowed on deck. They can be stowed on top of 40 ’ containers but the opposite
is not possible, since the 40 ’ containers on top will not be accessible for lashing. To
go above the fifth tier in stacks on deck containers must be tied to the lashing bridge.
This metallic structure goes a few tiers up from the deck and allow lashing rods to go
further up in the container stack, (up to the bottom corner castings of containers in
the fifth tier). The length between lashing bridges is equivalent to the size of a 40 ’
container, thus, 45 ’ containers must be stowed above the lashing bridge. In the case
of 20 ’ containers, it is only possible to secure one side of the container to the lashing
bridge. The other side, facing inwards, is lashed to the deck up to the fifth tier. After
the fifth tier, the containers are only secured by twistlocks. This makes pure 20 ’ stacks
weaker.
The lashing rods of 20 ’ containers facing inwards must always be accessible during
the journey. Since it is not possible to walk between containers from one lashing bridge
to another, 40 ’ containers in mixed bays (bays that hold 20 ’ and 40 ’ containers), must
be stowed in such a way that there is an access path either from port or starboard, to
the 20 ’ containers. Figure 2.10 (a) shows a stowage pattern in a mixed bay with an
access path to the 20 ’ containers, while Figure 2.10 (b) depicts a pattern where 40 ’
containers block access to the 20 ’ containers. As a consequence of the need for this

(a) (b)

Figure 2.10: Small bay seen from above. (a) Stowage where there is free access to 20 ’ lashing
rods. (b) Stowage 40 ’ containers blocking access to 20 ’ lashing rods.

access path, 20 ’ and 40 ’ containers must rest on different outer sockets (structures on
the hatch covers where containers on the bottom layer are positioned), as depicted in
Figure 2.11. The area occupied by two 20 ’ containers is no longer the same as that of
a single 40 ’ container, thus, Russian stacks (stacks where 20 ’ and 40 ’ containers are
stowed together), are not allowed on deck. In newer vessels, however, it is possible to
do Russian stacks on deck.
Other factors such as the stability of the vessel, neighbor stacks, and visibility must
also be considered while stowing container stacks on deck. When the GM forces the

36
 Chapter 2. Container Stowage Planning

20’ Fore 20’ Aft 40’

20’ Fore 20’ Aft 40’

Outer sockets Inner sockets Outer sockets Inner sockets Outer sockets

(a) (b)

Figure 2.11: The position of the sockets on a bay where lashing rods of 20 ’ containers need
to be accessed. (a) Shows 20 ’ containers using inner and outer sockets. (b) Shows 40 ’
containers using only the outer sockets. Notice that 20 ’ and 40 ’ containers use a different
set of outer sockets.

vessel to have a short rolling period (stiff vessel), containers at the top tiers experience
strong forces that put considerable stress on the twistlocks and the structure of the
containers in the lower tiers of the stack. As a consequence, the size of container
stacks on deck must be reduced to avoid their collapse. The impact of this problem,
however, can be considerably reduced by: 1) Lowering the GM of the vessel and, 2)
Redistributing the weights in the container stacks by stowing the heavy containers at
the bottom tiers and the light ones at the top.
Container stacks on deck must allow full visibility from the pilot house. An imag-
inary line is drawn from the pilot house to a point 500m or twice the length of the
vessel, whichever is the smallest, measured from the foremost point of the vessel (line
of sight), (see Figure 2.3). Essentially, all containers on deck must be below this line. It
is important to notice that the line of sight is influenced by the fore draft of the vessel.
The longer the fore draft, the less restrictive the line of sight becomes, increasing the
on deck capacity of the vessel.

2.5 Stowage Plans

Stowage plans assign containers to vessel slots. These plans are made by stowage coor-
dinators who work for the liner shipping company before the vessel calls into port. In
some liner shipping companies the time a stowage coordinator has to produce a stowage
plan is six hours. When making a stowage plan, stowage coordinators have available a
loadlist of containers (i.e., the list of containers to load in the current port), the vessel
condition when it calls the port, port information, and a forecast based on historical
data. The loadlist of containers includes detailed information about the containers, such
as weight, height, length, type (e.g., standard, reefer, IMO, Pallet-wide), and discharge
port. The release is a list of containers onboard the vessel when it calls into port,
it has the same container information as the loadlist and includes the exact position
where containers are stowed. The status of the ballast, fuel, diesel, and oil tanks when
the vessel calls into port is also provided to the stowage coordinators together with

37
Chapter 2. Container Stowage Planning

information on changes to any of the tanks and non-cargo loads that the vessel carries.
Stowage coordinators are provided with port information for the current and down-
stream ports. This information includes the number of quay cranes assigned to the
vessel and the maximum draft at each port. Stowage plans are made for the current
port but stowage coordinators must take downstream ports in the vessel schedule into
account, since the decisions they make in the current port affect future stowage plans.
Considering that forecasts are not always accurate, stowage coordinators follow rules-
of-thumb to make robust stowage plans, e.g., avoid stowing non-reefer containers into
reefer slots, keep the number of different discharge ports in the same location as low as
possible, and cluster together containers with the same discharge ports that are stowed
in the same vessel sections.

Once a vessel is loaded with a stowage plan the vessel must be seaworthy, container
stacks must satisfy all stacking rules and limits, and containers with special require-
ments must have been handled correctly. Feasible stowage plans, however, should be
economically viable as well. They must save time at port and, if possible, reduce fuel
consumption. There are two main features of a stowage plan that impact the time a
vessel spends at port. The first is the distribution of moves over the vessel. Stowage co-
ordinators attempt to distribute evenly the moves along the vessel to minimize the crane
makespan. Figure 2.12 (a) presents a distribution of containers with a high makespan,
where moves concentrate at the stern of the vessel. A better distribution of the moves
along the vessel, as Figure 2.12 (b) depicts, reduces the makespan.

110 50 200 110 120 130

50 60 15 15 20 50 75 75 50 60 40 60 20 50 50 30

(a) (b)
Figure 2.12: Crane split of three cranes assigned to a container vessel. The numbers in each
vessel section indicates the moves that need to be performed there. (a) Bad distribu-
tion of moves along the vessel with a crane makespan of 200. (b) Improved distribution
distribution of moves with a crane makespan of 130.

The second feature is the number of unnecessary moves performed to unload the
containers on the vessel. Let us assume that we have two different sets of containers,
A and B, stowed in the same bay. The set of A containers is going to Shanghai and is
stowed right on top of a set of B containers going to Yokohama. Since containers can
only be accessed from the top, all containers in A must be unloaded to access containers
in B. If the container vessel calls Shanghai before calling Yokohama, containers in A
will be unloaded in Shanghai, allowing free access to the containers in B at Yokohama.
On the other hand, if the vessel calls Yokohama first, the containers in A will overstow
the containers in B and will have to be shifted, (unloaded in Yokohama and moved
back into the vessel to the same or a different position). Overstowage is measured by
the number of containers that have to be shifted. In our previous example it is the

38
 Chapter 2. Container Stowage Planning

number of containers in A. There are two types of overstowage, stack overstowage,

where containers in sets A and B are stowed in the same stack (see Figure 2.13 (a)),
and hatch overstowage, where containers in sets A and B are stowed above and below
the same hatch cover, respectively, (see Figure 2.13 (b)).

(a) (b)
Figure 2.13: Examples of different kinds of overtoswage in a bay section. The green container
is going to Yokohama, the red containers are going to Shanghai. The vessel calls Yokohama
first. (a) Stack overstowage: Containers on top of the green container must be shifted in
Yokohama. (b) Hatch overstowage: Containers on top of the green container, few from the
same section and all from the section on top, must be shifted in Yokohama.

With respect to fuel saving, stowage coordinators aim at distributing containers

along the vessel such that the trim gets as close to optimal as possible.

39
Chapter 2. Container Stowage Planning

40
 Chapter 3. Slot Planning

Chapter 3

Slot Planning

In chapter 1, we describe a 2-phase decomposition approach for scalable stowage plan-

ning. In this approach, the master planning phase assigns containers to locations along
the vessel, satisfying high-level constraints and objectives. The slot planning phase,
on the other hand, stows containers into vessel slots, satisfying low level stacking con-
straints and objectives, following the assignment made during the master planning
phase. This chapter introduces the definition of a representative problem for slot plan-
ning locations, together with two optimization models to solve this problem efficiently.
This chapter is based on Delgado et al. (2009), and Delgado et al. (2012).

3.1 Introduction
Slot planning is the second phase of the 2-phase heuristic decomposition depicted in
Figure 1.3. In this phase, the containers assigned to locations during master planning
are stowed into vessel slots. Each location is stowed independently such that container
stacks fulfill stacking rules and satisfy capacity limits, and optimize objectives that fo-
cus on robustness. Figure 3.1 depicts the layout of the slot planning phase considered
in this chapter. A typical large container vessel has about 100 locations, which im-

Slot Planning
Location Loc. Slot Slot Plan
1 Planning 1

Master Location Loc. Slot

Stowage
Slot Plan
Plan 2 Planning 2 Plan
...

...
...

Location Loc. Slot Slot Plan

n Planning n

Figure 3.1: Layout of slot planning phase.

plies that the slot planning phase solves about 100 independent slot planning problems.
Thus, when given at most 10 minutes1 to generate complete stowage plans, including

1
Time limit set by our industrial collaborator for the generation of a stowage plan.

41
Chapter 3. Slot Planning

master planning, on hardware that does not support heavy parallelization, we aim at
solving each slot planning problem in less than one second. This is not trivial since
slot planning is NP-hard and each location may hold up to several hundred containers.
Real slot planning problems are full of details, including a wide variety of special vessel
structures and container types. To make their study practical, we introduce the Con-
tainer Stowage Problem for Below Deck Locations (CSPBDL), a representative model
for stowing containers in locations below deck formulated together with our industrial
collaborator. Even though this is a simplified representation of the problem, it is, to
our knowledge the most detailed model published to date. We formulate an Integer
Programming (IP) and Constraint Programming (CP) model for solving the CSPBDL
to optimality.2 The CP model uses state-of-the-art modeling techniques including mul-
tiple viewpoints, specific domain pruning rules, and dynamic lower bounds. The IP
model is a 0-1 formulation where cuts are introduced to strengthen the LP relaxation.
Despite the size and NP-hardness of slot planning problems, our computational
results show that they can often be solved very quickly in practice. We have devised two
experimental setups to evaluate our models. In the first setup, we have derived 236 test
instances from real stowage plans from a system deployed by our industrial collaborator,
(Guilbert and Paquin (2010)). 92% of the instances are solved using a state-of-the-art
constraint solver on our CP model within one second. Similar but slightly worse results
were obtained with the IP model. In the second setup, we incorporate our CP model,
the most successful model based on the first set of experiments, into a module that
computes stowage plans for container vessels, (Pacino et al. (2011)). We used this
module to stow 18 industrial stowage conditions. In this setup, our CP model managed
to generate optimal slot plans for up to 94.8% of the locations with containers to stow.
Thus, somewhat to our surprise, it is possible to define optimal models of slot planning
problems that can be solved fast enough to be used in stowage planning optimization
tools.
This chapter is organized as follows. Section 3.2 defines the problem we address
in this chapter and related work is presented in Section 3.3. In Section 3.4, we give a
detailed description of our IP model. Section 3.5 gives a brief introduction to global
constraint modeling, and in Section 3.6 we present our CP model. Computational
results are presented in Section 3.7 and conclusions and directions for future work are
discussed in Section 3.8.

3.2 Problem Statement

In this chapter we study the slot planning problem, which is to assign a set of containers
to slots in a location that may already hold release containers. Due to the large num-
ber of constraints and objectives involved in real slot planning, we have developed a
representative version of the problem for below deck locations called the CSPBDL. On
2
A column generation model was also formulated for this problem and experimentally tested in
Delgado et al. (2009).

42
 Chapter 3. Slot Planning

deck locations share most constraints and objectives with below deck locations, thus we
expect similar computational results for them. The CSPBDL covers all constraint and
objective classes of the problem, and we anticipate a high correlation with a complete
problem model in terms of solution algorithm performance. Specifically, the CSPBDL
includes stacking rules for 20 ’ and 40 ’ containers, FEU and TEU stack overlapping,
reefer containers, release containers, and weight and height constraints. The objectives
include overstowage and three rules of thumb used by stowage coordinators to ease the
stowing of containers in downstream ports. We limit the containers we consider to be
20 ’ and 40 ’ long, 8’6” and 9’6” high, and reefer and non-reefer. A feasible CSPBDL
must satisfy the following rules.

a) Assigned cells must form stacks, (containers stand on top of each other in the
stacks. They must have support from below)
b) 20 ’ containers cannot be stacked on top of 40 ’ containers
c) A 20 ’ reefer container must be placed in a reefer slot. A 40 ’ reefer container must
be placed in a cell with at least one reefer slot
d ) The length constraint of a cell must be satisfied, (some cells only hold 40 ’ or 20 ’
containers)
e) The sum of the heights and weights of the containers stowed in a stack are within
the stack limits
f ) All release containers must be stowed in their original slots and they cannot be
swapped to any other slots
g) A cell must be either empty or with both slots occupied

Additionally, an optimal CSPBDL minimizes the sum of the following objectives, listed
in order of importance:

h) Minimize overstows. A penalty is paid for each container overstowing any con-
tainers below
i ) Avoid stacks where containers have many different Port of Discharge (POD). A
penalty is paid for each POD included in a stack
j ) Keep stacks empty if possible. A penalty is paid for each stack used
k ) Avoid loading non-reefer containers into reefer slots. A penalty is paid for each
non-reefer container stowed in a reefer slot

The second, third, and fourth objectives are rules of thumb of the shipping industry
when generating slot plans for downstream ports in the route of a vessel. Using as few
stacks as possible increases the available space in a location and reduces the possibility
of overstowage in future ports, so does clustering containers with the same POD. Min-
imizing the reefer objective allows more reefer containers to be loaded in future ports.
The cost units reflect the importance of each objective and has been defined by our
industrial collaborator.

43
Chapter 3. Slot Planning

The CSPBDL is NP-Hard. We show that the bin-packing problem can be reduced
to the CSPBDL. All items in a bin-packing problem are defined as 40 ’, standard height
and non-reefer containers with the same POD, which sizes are the weight of the con-
tainers. Bins are defined as stacks where height limits are set to be sufficiently large
to be non-restrictive, with no reefer slots, and which sizes are the weight capacity of
the stacks. An optimal solution to an arbitrary bin-packing problem can be found by
finding an optimal solution to the CSPBDL. When all containers have the same POD
and are non-reefer, objectives i and j dominate the objective function and the number
of stacks used is minimized, showing that the CSPBDL is NP-Hard.3

3.3 Literature Review

Slot planning optimization algorithms have either been studied as embedded into single-
phase models, or as a part of multi-phase decompositions for generating stowage plans.
In the first category, Avriel et al. (1998), Dubrovsky et al. (2002), and Ambrosino
and Sciomachen (1998), propose simple models to stow complete vessels that address
similar problems to slot planning optimization. Avriel et al. introduce a 0-1 IP model
and a heuristic called the suspensory heuristic, to stow vessels described as a collection
of columns and rows, (a rectangular bay). They consider all containers to have the same
features and focus on minimizing re-shifting of containers, (overstowage). Dubrovsky
et al. presents a genetic algorithm model with the same assumptions as Avriel et al.’s.
They claim, however, that their approach is flexible enough to include new constraints.
Ambrosino and Sciomachen present a CSP model. Though this model is meant to stow
a complete vessel, inter-bay stability constraints can be dropped in order to resemble the
slot planning problem. This approach considers 20 ’ and 40 ’ containers, but lacks reefer
and high-cube containers. Their objective is to minimize overstowage and maximize the
number of containers loaded. Aslidis (1984), Botter and Brinati (1992), Sciomachen
and Tanfani (2003), and Li et al. (2008), present more complex models that also include
several of the constraints considered in the master planning problem. Aslidis introduces
stacking heuristics for minimizing overstowage, while Botter and Brinati, and Li et al.
present 0-1 IP models. Botter and Brinati also present two heuristics to stow containers
since their IP model is not scalable to real-life instances. Sciomachen and Tanfani
introduce a heuristic approach based on the 3D-packing problem. These approaches
consider 20 ’ and 40 ’ containers and overstowage minimization. Sciomachen and Tanfani
also consider high-cube containers.
In the second category, slot planning problems are solved in connection with multi-
phase approaches for complete vessel stowage planning. Wilson and Roach (2000),
briefly describe a tabu search algorithm for solving a version of slot planning that must
have included reefer slots, length restrictions, minimized overstowage, and avoided POD
mixing of stacks. They claim that near optimal solutions could be computed fast, but
3
It can be shown that overstow minimization is also an NP-hard component of the CSPBDL, but
the proof requires a version of the problem with uncapacitated stacks Avriel et al. (2000).

44
 Chapter 3. Slot Planning

only experimental results for generating a complete stowage plan for a single vessel
are described. Kang and Kim (2002), describe an enumeration approach for solving
a very simple version of slot planning, where only overstow minimization and sorting
of 40 ’ containers after weight are considered. As for Wilson and Roach, no indepen-
dent experimental evaluation of the algorithm is provided. Ambrosino et al. (2009),
describes a 0-1 IP model for optimally stowing subsets of vessel bays holding containers
with the same POD. The model minimizes the time for stowing containers. 20 ’ and
40 ’ containers are considered, and containers are sorted according to weight in each
stack. In the experimental section, complete stowage plans for a 198 and 2124 TEUs
container vessels are generated. The maximum bay size is 20 TEUs for the small vessel
and 120 TEUs for the big one. No computational time, however, is provided for solving
these sub-problems. In a later work, Ambrosino et al. (2010), presents a constructive
heuristic to solve the same sub-problem as the one described in Ambrosino et al. (2009),
as part of a complete approach to stow vessels. Using this heuristic, they are able to
stow a vessel of 5632 TEUs. The heuristic uses 11.8 seconds on average to stow all
the bays, but the physical layout of the vessel is not described in detail. Zhang et al.
(2005), and Yoke et al. (2009), present multi-phase approaches where the problems
solved during the slot planning phase are not independent of each other.
A deployed industrial system introduced by Guilbert and Paquin (2010), that pro-
vides data to our experiments, solves slot planning problems as linear assignment prob-
lems with side constraints. Their model considers all containers and most of the con-
straints and objectives present in the CSPBDL. Overstowage is only considered with
respect to loaded containers, and though they minimize mixed stacks with 20 ’ and 40 ’
containers, constraint (b) is not present.

3.4 The IP Model

In this section, we introduce a binary IP model formulated to solve the CSPBDL.
We remind the reader that the sets and constants used in the models presented in this
chapter have been previously defined in the section Sets and Constants, at the beginning
of this document. Below we present the variables used in this model.

Decision Variables
cjki ∈ {0, 1} Container i being stowed in cell k, stack j.

Auxiliary Variables
oi ∈ {0, 1} Container i overstowing.
djp ∈ {0, 1} At least one container in stack j being unloaded at pd.
ej ∈ {0, 1} Stack j being used.
δjkp ∈ {0, 1} Container below cell k, stack j being unloaded before port p.

The sets of variables o, p, and e are used for computing the cost of overstow (h),
clustering (i ), and using stacks (j ) according to the CSPBDL. The variables in c rep-

45
Chapter 3. Slot Planning

resent the slot plan. The set of auxiliary variables delta represents indicator variables
introduced to model the overstowage objective. The IP model is defined as:
X X X X
min C v oi + C p djp + C u ej
i∈Q j∈J p∈P −{1} j∈J
XX X X 1 
+ Cr Rjk cjki (1 − RiC ) + cjki ( Rjk − RiC ) (3.1)
j∈J k∈K 40 20
2
j i∈Q i∈Q

s.t.
1 X X X
cj(k−1)i + cj(k−1)i − cjki ≥ 0 ∀j ∈ J, k ∈ Kj − {1} (3.2)
2 20 40 40
i∈Q i∈Q i∈Q
X X
cjki − cj(k−1)i ≤ 0 ∀j ∈ J, k ∈ Kj − {1} (3.3)
i∈Q20 i∈Q20
1 X X
cjki + cjki ≤ 1 ∀j ∈ J, k ∈ Kj (3.4)
2 20 40
i∈Q i∈Q
XX
cjki = 1 ∀i ∈ Q (3.5)
j∈J k∈Kj
X
cjki0 − 2cjki ≥ 0 ∀j ∈ J, k ∈ Kj , i ∈ Q20 (3.6)
i0 ∈Q20
X
RiC cjki − Rjk ≤ 0 ∀j ∈ J, k ∈ Kj (3.7)
i∈Q
XX
WiC cjki ≤ WjS ∀j ∈ J (3.8)
k∈Kj i∈Q
X 1 X X
( HiC cjki + HiC cjki ) ≤ HjS ∀j ∈ J (3.9)
k∈K
2 20 40
j i∈Q i∈Q
p−1
k−1 X
X X
Aip0 cjk0 i − 2(k − 1)δjkp ≤ 0 ∀j ∈ J, k ∈ Kj , p ∈ P (3.10)
k0 =1 p0 =2 i∈Q

Aip cjki + δjkp − oi ≤ 1 ∀j ∈ J, k ∈ Kj , (3.11)

p ∈ P, i ∈ Q
ej − cjki ≥ 0 ∀j ∈ J, k ∈ Kj , i ∈ Q (3.12)
djp − Aip cjki ≥ 0 ∀j ∈ J , k ∈ Kj , (3.13)
p ∈ P, i ∈ Q
cjki = 1 ∀(j, k, i) ∈ M (3.14)

The objective function (3.1) is a weighted sum of the four objectives as defined in
the CSPBDL. The first three objectives are calculated straightforward since there are
specific auxiliary variables in the model that account for them. The fourth objective

46
 Chapter 3. Slot Planning

is calculated by determining the number of non-reefer containers stowed in slots with

reefer plugs. 40 ’ and 20 ’ containers are considered independently.
Inequality (3.2) ensures that there is either two 20 ’ or one 40 ’ containers below a
cell stowing a 40 ’ container, while inequality (3.3) constrains the containers below a cell
stowing 20 ’ containers to be 20 ’ long (b). Inequality (3.4) requires that all cells stow at
most either two 20 ’ or one 40 ’ container. Containers are forced to be stowed in exactly
one cell by (3.5). Inequality (3.6) forces the number of 20 ’ containers in a cell to be 0
or 2, since the two sides of a stack must be synchronized (g). The reefer capacity of
a cell is constrained by inequality (3.7), covering the fact that all reefer containers in
a cell must be provided with a reefer plug each (c). The weight and height limits of
stacks (e) are enforced by (3.8) and (3.9), respectively. Inequality (3.10) ensures that
the variables δjkd are assigned the correct value according to their semantics. These
variables are then used in inequality (3.11) to assign the overstowage variables oi for
each container. Inequality (3.12) sets a lower bound for the variable related to the
empty stack objective (j ) for each stack, and inequality (3.13) does the same for the
variables related to the clustering objective (i) for each stack at each POD. Release
containers are assigned to their corresponding cell by equality (3.14).

Cuts
We add cuts that focus on removing solutions with non-integer values assigned to vari-
ables δjkp by the Linear Programming (LP) relaxation. First we decompose inequal-
ity (3.10) into several inequalities, one for each variable cjk0 i , that combined together
are semantically equivalent to (3.10):
Aip0 cjk0 i ≤ δjkp ∀j ∈ J, k ∈ Kj , p ∈ P, i ∈ Q, k 0 ∈ K 0 , p0 ∈ P 0 (3.15)
where K 0 = {k 0 |k 0 ∈ {1, ..., k−1}} and P 0 = {p0 |p0 ∈ {2, ..., p−1}}. We then increase the
size of the left hand side term of (3.15) by considering at once all containers unloaded
earlier than port p. Two inequalities are introduced to the model (3.16, 3.17), since 20 ’
and 40 ’ containers need to be treated differently. Additionally, cut (3.18) adds terms
to the left hand side of (3.15) by considering all cells below cell k. The cuts are defined
by:
1 X
cjk0 i ≤ δjkp ∀j ∈ J, k ∈ Kj , p ∈ P, k 0 ∈ K 0 (3.16)
2 0 20
i∈Qp
X
cjk0 i ≤ δjkp ∀j ∈ J, k ∈ Kj , p ∈ P, k 0 ∈ K 0 (3.17)
0
i∈Qp40
k−1
X
Aip0 cjk0 i ≤ δjkp ∀j ∈ J, k ∈ Kj , i ∈ Q, p ∈ P, p0 ∈ P 0 (3.18)
k0 =1
0 0
where Qp20 and Qp40 are the set of 20 ’ and 40 ’ containers with POD earlier than p,
respectively.

47
Chapter 3. Slot Planning

3.5 Global constraint modeling

A Constraint Satisfaction Problem (CSP) is a triple (X, D, C) where X is a set of vari-
ables, D is a mapping of variables to finite sets of integer values, and D(x) representing
the domain of x ∈ X and D(X) = Πx∈X D(x) being the Cartesian product of domains,
and C is a set of constraints. Each c ∈ C is defined over a sequence X 0 ⊆ X as a subset
of allowed combinations of D(X 0 ). A solution to a CSP is a complete assignment that
maps every variable to a value from its domain that satisfies all constraints in C.
Constraint Programming (CP) is a relatively new technique that combines local
consistency algorithms with search. The process of removing inconsistent values from
the domain of the variables is called propagation. A depth-first backtracking search
explores the search space of the problem incrementally. It extends a partial solution by
selecting unassigned variables from X and assigning them to values from their domains.
This selection process is called branching, and a strategy to select variables and values
following a specific criteria is called a branching strategy. Propagation is executed every
time new branching is generated. If the domain of each variable has been reduced to
a single value, the CP solver has found a solution to the CSP. For a partial solution,
we refer to the minimum and maximum value of the domain of variable x as x and x,
respectively. A cost function is defined for a CSP in order to evaluate the quality of its
solutions and branch and bound is used to find optimal solutions.
Constraints in CP share information through the variables in X. Each constraint
has a scope X 0 ⊆ X, that often is relatively small compared to the size of X, limiting its
propagation power. Global constraints have been introduced to overcome this. A global
constraint groups together a set of small constraints capturing tractable structures for
global propagation. Below is a brief description of the global constraints used in our
CP model.
Let x be an integer variable, y a variable with finite domain, and C = {c1 , ..., cn }
a set of constants. The element constraint Van Hentenryck and Carrillon (1988) states
that y is equal to the x-th constant in C.
element(x, y, C) = {(e, f ) | e ∈ D(x), f ∈ D(y), f = ce }.
Let M be a deterministic finite automaton or a regular expression recognizing the
strings in the language L(M ) ⊆ Σ∗ , and let X = {x1 , ..., xn } be a set of variables with
D(xi ) ⊆ Σ for 1 ≤ i ≤ n. Then the regular constraint Pesant (2004), is defined as
regular(X, M ) = {(d1 , . . . , dn ) | ∀i . di ∈ D(xi ), d1 · · · dn ∈ L(M )}.
Let n and v be two integer values, and X = {x1 , ..., xm } a set of finite domain
variables. The exactly constraint ensures that exactly n variables in X are assigned to
value v.
exactly(n, X, v) = {(d1 , . . . , dm ) | ∀i . di ∈ D(xi ), |{di | di = v}| = n}
Let X = {x1 , ..., xn } and Y = {y1 , ..., yn } be two sets of finite domain variables with
domains D(X) = D(Y ) = {1, ..., n}. The channeling constraint states that a value j

48
 Chapter 3. Slot Planning

assigned to a variable xi ∈ X represents the index of the variable yj ∈ Y that has been
assigned value i from its domain. Formally

channeling(X, Y ) = {(e1 , . . . , en , f1 , . . . , fn )|
∀i, j . ei ∈ D(xi ), fj ∈ D(yj ), ei = j ⇔ fj = i}.

A channeling constraint is used to increase the propagation power of the model by

connecting several isomorphic variable sets, (also known as viewpoints Smith (2006)).
If X is a set of variables representing positions with boxes {1, . . . , n} as domains and
Y is a set of variables representing boxes with positions {1, . . . , n} as domains, then
clearly a channeling constraint will link them consistently together. In particular, the
channeling constraint embeds the alldifferent constraint, that in our example ensures
that a position only can hold one box and vice versa.

3.6 The CP Model

In our CP model, the stack in the left most part of the location has the lowest index in
J . Indices in S are assigned to physical slots as follows. For each cell, the aft and fore
slots have consecutive indices. The slot indices in each stack are ordered bottom-up
and the slot indices between stacks are ordered from left to right in the location. We
have Sk = SkF ∪ SkA , and Pic < Pjc iff the vessel calls the POD of container i before the
POD of container j. Below we introduce the variables of our cp model.

Decision Variables
Conts= {c1 , ..., c|Q| } ci ∈ S , slot index of container i.
Slots= {s1 , ..., s|S | } sj ∈ Q, container index of slot j.

Auxiliary Variables
Lengths= {l1 , ..., l|S | } lj ∈ LC , length of container stowed in slot j.
Heights= {h1 , ..., h|S | } hj ∈ H C , height of container stowed in slot j.
W eights= {w1 , ..., w|S | } wj ∈ W C , weight of container stowed in slot j.
P orts= {p1 , ..., p|S | } pj ∈ P C , POD of container stowed in slot j.
HS = {hs 1 , . . . , hs |J | } hs j ∈ {0, ..., HjS }, current height of stack j.
ov ∈ {0, ..., |Q|} Number of overstowing containers.
ou ∈ {1, ..., |J |} Number of used stacks.
p
o ∈ {1, ..., |J ||P |} Number of different POD in each stack.
or ∈ {0, ..., |S R |} Number of non-reefers stowed in reefer cells.
o∈ N Solution cost variable.
V
Conts ⊂ Conts Virtual containers.
SlotsEi ⊂ Slots Slots with the same features in stack i.
π
Lengthsi ⊂ Lengths Length variables of slots in the aft or fore part of stack
i.

49
Chapter 3. Slot Planning

The decision variables represent the slot plan for a set of containers to be stowed.
We use two isomorphic representations. The first one defines a decision variable for
each container in Q to be stowed, and as domain of the variables the slots in S . The
second one defines a decision variable for each slot in S , and as domain of the variables,
the set of containers in Q to be stowed.
The two sets of decision variables mentioned above define two different viewpoints
in our CP model. These two viewpoints are linked with a channeling constraint. The
current formulation of the problem, however, does not allow a straightforward use of
this constraint since in most of the cases the number of slots is larger than the number
of containers, which breaks an important precondition of the channeling constraint. To
tackle this issue, we modify the original definition of the problem by extending the
number of containers with artificial containers to match the number of slots. First,
since a 40 ’ container occupies two slots, all 40 ’ containers are split in two parts, Aft40
and Fore40 , with the size of a single slot. All 40 ’ containers from Q and Q 40 are
replaced by Aft40 and Fore40 containers. Virtual containers, Q V , that will be stowed
in slots meant to remain empty are also added. In the remainder of the chapter, Q will
refer to this extended set of containers. Finally, Q ¬R is the set of non-reefer containers
excluding virtual containers.
In addition to the two sets of decision variables, extra sets of auxiliary variables are
defined to facilitate the modeling of the constraints and objectives. The objectives of
the CP model are given by:
X
ov = ov(i) (3.19)
i∈S A
 
X X
ou =  pi > 0 (3.20)
j∈J i∈Sj
 
X XX
op =  (pi = ρ) > 0 (3.21)
j∈J ρ∈P i∈Sj
X
or = (si ∈ Q ¬R ) (3.22)
i∈S R
o = C o + C p op + C u ou + C r or
v v
(3.23)
Objective (3.19) calculates the total number of overstows. ov (i) is the number of
overstowing containers in a cell represented by its aft slot i. We have:


 2 if si ∈ Q 20 ∧ pi > minP (be(i)) ∧ pi+1 > minP (be(i))
1 if (si ∈ Q 40 ∧ pi > minP (be(i))) ∨

ov (i) =

 (si ∈ Q 20 ∧ (pi > minP (be(i)) ⊕ pi+1 > minP (be(i))))
0 otherwise


where be(i) is the set of slots below slot i in the same stack, minP (S 0 ) is the earliest
POD among the containers assigned to a set of slots S 0 , and ⊕ denotes the exclusive-or

50
 Chapter 3. Slot Planning

boolean operator. The empty stack objective (j ), is represented by (3.20). The smallest
POD index, 0, is assigned to virtual containers. Thus, when a stack i is empty, the
sum of the values assigned to the subset of P orts variables in i is 0, otherwise the stack
is being used. Objective (3.21) calculates the number of different POD of containers
stowed in each stack, and objective (3.22) counts the number of non-reefer containers
stowed in reefer slots. Objective (3.23) defines the cost function of the CSPBDL. The
branch and bound algorithm applied to solve this problem constrains the cost variable
o of the next solution to be lower than that of the solution with the lowest cost found
so far. The constraints of the CP model are given by:

channeling(Conts, Slots) (3.24)

cfore(i) = ci + 1, ∀i ∈ {1, ..., |Q 40A |} (3.25)
element(si , li , LC ) ∀i ∈ S (3.26)
C
element(si , hi , H ) ∀i ∈ S (3.27)
element(si , wi , W C ) ∀i ∈ S (3.28)
element(si , pi , P C ) ∀i ∈ S (3.29)
L
spos(j) = j ∀j ∈ Q (3.30)
regular (Lengths πi , R) ∀π ∈ {A, F }, i ∈ J (3.31)
si 6∈ Q 20R ∀i ∈ S ¬R (3.32)
40R ¬RC
si 6∈ Q ∀i ∈ S (3.33)
si ∈ Q 20 ∀i ∈ S 20 (3.34)
si ∈ Q 40 ∀i ∈ S 40 (3.35)
X
hi ≤ hs j ∀π ∈ {A, F }, i ∈ J (3.36)
i∈Sjπ
X
wi ≤ WjS ∀j ∈ J (3.37)
i∈Sj

Constraint (3.24) connects the two viewpoints such that both sets of variables Conts
and Slots always have the same level of information. fore(i) is the Fore40 container
bound to Aft40 container i. Constraint (3.25) guarantees that the Aft40 and Fore40
part of a 40 ’ container are stowed in the same cell. Element constraints are used to
bind all auxiliary variables introduced in the model to a viewpoint. Constraints (3.26)
, (3.27), (3.28), and (3.29), bind each slot variable to the auxiliary variables representing
the length, height, weight, and POD of the container stowed in the slot. Release
containers are stowed in their pre-defined slots by constraint (3.30), where pos(j) is the
slot occupied by release container j. The valid patterns that containers stowed in stacks
must follow according to their length are defined by (a) and (b). After assigning a length
of 0 to virtual containers, we define a regular expression R = 20∗ 40∗ 0∗ that recognizes
all the valid patterns according to these two constraints. Constraint (3.31) introduces a
regular constraint for each aft and fore stack in order to restrict their stacking patterns

51
Chapter 3. Slot Planning

to follow those defined by R. Constraints (3.32) and (3.33) model the reefer constraint
(c). Constraints (3.34) and (3.35) restrict the domains of slots that have just 20 ’ or
40 ’ container capacity to be the set of 20 ’ and 40 ’ containers, respectively. The height
limit of each stack in the location is constrained by (3.36). All containers stowed in
each side of a stack must be less or equal to the variable representing the height limit
of the stack4 . Constraint (3.37) restricts the weight of all containers stowed in a stack
to be within the limits.

3.6.1 Symmetry-Breaking and Implied Constraints

We introduce a set of constraints to the CP model that aim at reducing the search space
size and increase propagation. These constraints are implied by existing constraints and
break symmetries either already present in the problem or introduced by our model
representation.

exactly(V, 0, |Q V |) ∀ V ∈ {Ports, Weights, Heights} (3.38)

exactly(P orts, p, QP =p ) ∀ p ∈ PC (3.39)
W =w C
exactly(W eights, w, Q ) ∀w ∈ W (3.40)
α α
exactly(Heights, H , Q ) ∀ α ∈ {DC, HC} (3.41)
hi = hk ∀ j ∈ J , i, k ∈
{(i, k)| i ∈ SjA , k ∈ SjF .eqCell (i, k)} (3.42)
V
sort(Conts ) (3.43)
40A F
si 6∈ Q ∀i ∈ S (3.44)
6 Q 40F
si ∈ ∀ i ∈ SA (3.45)
si ≤ sk ∀ j ∈ J , i, k ∈
{(i, k)| i ∈ SjA , k ∈ SjF .eqType(i, k)} (3.46)
sort(SlotsE
j ) ∀j ∈ J (3.47)
i
lex (Class ) ∀ i ∈ Classes (3.48)

Constraints (3.38), (3.39), (3.40), and (3.41) are implied constraints meant to im-
prove the propagation power of the solver with respect to the auxiliary variables P orts,
W eights, and Heights. Each individual auxiliary variable zi is linked to a slot variable
si with an element constraint. This ensures correctness but leads to weak propagation
between the two sets of variables due to a lack of global perspective by the element
constraints. To improve this, we first assign the value zero to the weight, height, and
POD of virtual containers, (these containers are not supposed to affect total height,
weight, or overstowage of each stack). Then, constraint (3.38) limits the number of
4
The HS variables are not necessary to define the height constraint but play an important role in
the height constraint lower bound introduced in Section 3.6.3.

52
 Chapter 3. Slot Planning

variables set to zero from P orts, W eights, and Heights to be the exact number of vir-
tual containers. Additionally, constraints (3.39), (3.40), and (3.41) restrict the number
of variables from P orts, W eights, and Heights assigned to each possible POD, as well
as weight or height to match the total number of containers with such features, respec-
tively. Constraint (3.42) restricts the height of the containers stowed in the aft and fore
slots of a cell to be equal. This is possible since both slots in a cell must be either empty
or occupied at the same time (g), and there are no 20 ’ high-cube containers available
to stow. The function eqCell (j, k) indicates that two slots j and k belong to the same
cell.
The weight of the containers make each of them almost unique, limiting the possibil-
ity of applying symmetry breaking constraints. It is possible, however, to break some of
the symmetries introduced into the problem by our model representation. First, since
all virtual containers have the same features, it is not relevant where each container is
stowed. Constraint (3.43) posts a sorting constraint over the virtual containers, forc-
ing the slots where these containers will be stowed to follow a non-decreasing order.
Second, splitting 40 ’ containers into Aft40 and Fore40 parts also generates symmetri-
cal solutions that are broken by constraint (3.44) and (3.45). Third, constraint (3.46)
limits the possibility of swapping containers between two slots of a cell that have the
same features. The function eqType(j, k) indicates that two slots j and k belong to
the same cell and have the same features, i.e., same reefer plug and length restrictions.
Fourth, when all containers have the same POD, symmetrical solutions are generated
by swapping containers stowed in slots with the same features within the same stack.
Constraint (3.47) sorts in a non-decreasing order, the indices of the containers stowed in
slots with the same features of each stack. Indices are assigned to containers such that
conflicts between constraint (3.47) and valid stacking patterns (3.31) are avoided. 20 ’
containers are assigned a lower index than 40 ’ containers, and virtual containers have
the highest index possible. Finally, symmetries between stacks with identical charac-
teristics are considered. Stacks are classified according to their features: slot capacity,
reefer capacity, height and weight limit. Constraint (3.48) removes symmetrical solu-
tions generated by the containers stowed in similar stacks being swapped with each
other, by requiring a lexicographical ordering on the indices of the containers stowed in
these stacks.

3.6.2 Branching Strategies

Our branching strategy takes advantage of the structure of the model and uses the
sets of different auxiliary variables in order to find high-quality solutions early in the
search. We decompose the branching process into four sub-branchings: the first one
focuses on finding high-quality solutions, the second and third on the feasibility the
of two problematic constraints, and the fourth finds a valid assignment for the deci-
sion variables Slots. A detailed description of our branching strategy can be found in
Delgado et al. (2010). In the case of the first sub-branching, since three of the four ob-
jectives of the CSPBDL rely on the POD of the containers, we start by branching over

53
Chapter 3. Slot Planning

the set of POD variables P orts. Variables bound to slots with containers that favor the
clustering and overstowage objectives among the first free slots bottom-up of all stacks
are preferred. After assigning all variables in P orts, we branch over the height and
weight variables, Heights and W eights. We start by branching over Heights following
a best-fit decreasing approach for selecting a stack, then we assign the smallest height
possible among that of the containers to be stowed into the first free slot bottom-up in
the stack. We use a similar approach for W eights, where the best fit is considered to
be the stack with the greatest amount of free weight. Finally, we branch over Slots in
order to generate a concrete stowage plan. The domain size of variables in P orts are
considerably smaller than any of the viewpoints, making the process of finding valid
assignments for P orts easier. Once a valid slot plan is found, most of the time the
search algorithm backtracks directly to the P orts variables in order to find solutions
with a lower cost. Therefore, a large part of the search process concentrates on a much
smaller sub-problem. Branching is performed over the remaining variables only when
a solution with a lower cost is likely to be found.

3.6.3 Lower Bounds

Five domain pruning rules are defined over partial solutions. Each rule solves a relaxed
version of a sub-problem related to an objective or a constraint, generating lower bounds
for their corresponding objectives and pruning values from the domain of the variables
in the scope of the constraint. The lower bounds are described in more detail in Delgado
et al. (2010).

Overstowage To calculate a lower bound on the overstowage of a partial solution ρ,

we define a new function minP (S 0 ) = mini∈S 0 (pi | pi ∈ P orts) that selects the minimum
upper bound pi among the variables in P orts corresponding to a subset S 0 of slots. Let
ov ρ (i) be a function identical to ov (i) where minP substitutes minP . It is easy to
show that ov ρ (i) is a lower bound of ov (i) for any completion of ρ (Delgado et al.
(2010)).
P The pruning effect of the lower bound is achieved by adding the constraintv
v
o ≥ i∈S A ov ρ (i). An additional pruning rule can be applied when the domain of o
has been reduced to a single value that is equal to the lower bound. In this situation, we
enforce that all containers below non-overstowing containers are discharged at a later
port:
X
|D(ov )| = 1 ∧ ov ρ (i) = ov →
i∈S A
A
∀i ∈ {k ∈ S | ov ρ (k) = 0}, j ∈ be(i) . pi ≤ pj .

Empty stack In the remainder, we refer to container i as unstowed if the domain

of ci has more than one element. For the empty stack lower bound, a relaxation of
the stowage problem is solved. The height capacity of the stacks is the only con-
straint considered and the containers to stow QρN = {i ∈ Q | i 6∈ Q V , |D(ci )| > 1}

54
 Chapter 3. Slot Planning

are accounted as normal height containers. We first consider the used stacks of ρ,
JρU = {j ∈ J | ∃i ∈ Sj , |D(si )| = 1, si 6∈ Q V }, where there are containers already
stowed. The lower bound procedure stows as many containers as possible from QpN in
JρU such that the height capacity constraint is fulfilled. Once the used stacks are com-
pletely filled up, the empty stacks are sorted in decreasing order by height capacity and
filled up with the remaining containers of QρN . The minimum number of used stacks
M inuρ is the sum of used stacks |JρU | plus the empty stacks necessary to stow all remain-
ing containers. M inuρ is a lower bound of the number of used stacks of any completion
of ρ since the approach to solve the relaxed problem uses a minimum number of stacks.
The pruning effect is achieved by adding ou ≥ M inuρ .

Pure stack As with the used stack lower bound, a relaxed assignment problem is
solved considering just the height capacity constraint and all containers not yet stowed
as normal height containers. First, we introduce an alternative definition of the pure
stack objective. Let Φi = |{j ∈ J | ∃k ∈ Sj . pk = i}| be the number of stacks where
at least one
P container with POD i is stowed. We can express the pure stack objective
p
as o = i∈P Φi . For this definition of the pure stack objective, we introduce a lower
bound for a partial solution ρ. Let QρN,P =i be the set of unstowed containers in ρ
with POD i, JρP =i be the set of stacks stowing at least one container with POD i, and
Jρ¬P =i = J \ JρP =i be the set of stacks where no container with POD i is stowed. Our
goal is to generate a lower bound M inpρ (i) independently for each Φi , based on the
approach followed to generate lower P bounds for the used stacks objective. The pruning
effect is achieved by adding op ≥ i∈P M inpρ (i).

Reefer A lower bound M inrρ for the reefer objective of a partial solution ρ can be
deduced from a counting argument. Let Φ¬R R
ρ = |{i ∈ S | si ∈ / Q R , |D(si )| = 1}| denote
the number of reefer slots stowing a non-reefer container in ρ. Clearly, or ≥ Φ¬Rρ for any
completion of ρ. We tighten the lower bound of the reefer objective by considering the
unstowed reefer containers and the reefer slots with more than one container in their
domain that will not stow a virtual container. Let ΨR ρ denote the number of unstowed
UR
reefer containers in ρ. Further, let Φρ be the number of reefer slots where no virtual
container will be stowed. If ΦUR
ρ > ΨR UR R
ρ then at least Φρ − Ψρ extra reefer slots will
stow non-reefer containers. Thus, we can tighten M inrρ as follows
¬R
ΦUR R
: if ΦUR > ΨR

ρ − Ψρ + Φρ
M inrρ = ρ ρ
Φ¬R
ρ : otherwise.

The pruning effect is achieved as usual by adding or ≥ M inrρ .

Height The domains of auxiliary variables from sequences Height and HS are tight-
ened, and some conditions necessary for a partial solution to be viable are checked by
solving three relaxed problems. First, the number of normal and high-cube containers
that can possibly be stowed in the remaining free space of each stack is calculated. A

55
Chapter 3. Slot Planning

stack j of some partial solution ρ has free height hρ (j) = hs j − hsj , where hsj denotes
the height of the stowed containers in stack j. Let MρDC (j) and MρHC (j) denote the
maximum number of normal and high-cube containers that can be placed in stack j,
respectively. We then have:

MρN (j) = bhρ (j)/H DC c,

MρHC (j) = bhρ (j)/H HC c.

Let ΨDCρ and ΨHCρ denote the number of unassigned normal and high-cube containers
of ρ, respectively. Then, all possible slot plans generated from partial solution ρ must
satisfy X X
MρDC (j) ≥ ΨN ρ ∧ MρHC (j) ≥ ΨHC
ρ .
j∈J j∈J

Second, since containers cannot hang in the air, they must be stowed consecutively,
bottom-up in all stacks. Therefore, when the sum of the height of containers stowed
below tier n is equal to hs j , slots above tier n will not stow real containers. We stow
virtual containers in slots of stack j that are above its height upper bound hs j . In the
cases where the height of the container to be stowed is not yet known, it is assumed
that the container will have normal height, since this generates an upper bound in the
number of slots used in stack j. Additionally, the virtual containers are removed from
slots that are below hs j , since these slots must stow real containers. Finally, we update
hs j by applying the bin packing propagation rule suggested in Shaw (2004)
X X
hs j ≥ Hic − hs i , ∀j ∈ S .
i∈Cont i∈J \{j}

3.7 Experiments
We have defined two experimental setups to evaluate the models introduced in this
chapter. In the first setup, we evaluate implementations of the CP and IP models
in 236 slot planning instances, that have been derived from complete stowage plans
provided by our industrial collaborator. Each instance is made by restowing a random
location in one of the stowage plans. Since the plans have been applied in real life,
we can assume that the containers have been assigned to locations according to the
preferences of stowage coordinators. In the second setup, we incorporate the CSPBDL
model that performs better in our first experiments, into the stowage planning module
introduced in Pacino et al. (2011). We use this module to generate stowage plans for
18 industrial stowage conditions and measure the performance of our model.
The penalties for each objective term have been agreed with our industrial partner and
have been set to 100 units for overstowing containers, C v , 20 units for each POD in
a stack, C p , 10 units for using a new stack, C u , and 5 units for stowing a non-reefer
containers in a reefer slot, C r .

56
 Chapter 3. Slot Planning

3.7.1 Restowing Location Stowage Plans

In this experimental setup, the CP and IP models have been implemented in Gecode
3.3 (Gecode Team (2006)) and CPLEX 12.2, respectively. All the experiments were
run on a Linux machine with two Quad Core Opteron processors at 1.7 GHz and 8
GB of memory. In order to characterize hard instances, we have partitioned them
into groups with different features. Table 3.3 presents the features of each group of
instances. Notice that the instances have a low number of POD. This is to expect
from high quality stowage plans that avoid POD mixing. This does not imply that
the instances are easy since instances with a single POD still embeds the bin-packing
problem.

Cap. (TEUs) Cont. (TEUs) #POD

Grp. #inst. 40 ’ 20 ’ R HC
min , max , avg min , max , avg 1 2 ≥3
1 13 16 , 116 , 63 8, 116 , 54 * 13
2 22 8, 168 , 68 8, 136 , 52 * 22
3 13 30 , 124 , 74 8, 124 , 68 * * 13
4 78 6, 208 , 79 2, 202 , 63 * * 78
5 36 38 , 176 , 97 8, 170 , 81 * * * 36
6 15 42 , 172 , 73 16 , 74 , 46 * * 15
7 14 72 , 204 , 147 24 , 202 , 117 * * * * 14
8 14 40 , 148 , 96 40 , 136 , 87 * * * 14
9 17 44 , 220 , 124 36 , 200 , 111 * * * * 15 2
10 8 72 , 176 , 122 10 , 156 , 93 * * * 6 2
11 6 48 , 176 , 101 28 , 148 , 84 * * * * 3 3

Table 3.3: Grouping of instances. The first column is the group id and the second number
of instances in the group. The minimum, maximum, and average capacity of locations
(Cap.), and number of containers to stow (Cont.), in TEUs are presented in the next six
columns. The remaining columns show the features of the containers present in each group
(40 ’, 20 ’, reefer, and high-cube), and the number of instances with 1, 2 or 3 different POD
in the group.

Figure 3.2 shows the space utilization as a function of location size. The space
utilization is the number of TEUs to stow divided by the total TEU capacity of the
location. 172 of the 236 instances have a space utilization above 80%, which is to expect
from real stowage plans.

Impact of CP enhancements
We first analyze the impact of the different enhancements of the CP model introduced
in Section 3.6. We define four CP models. The basic model includes only the core
constraints and objectives of the CSPBDL (3.19 - 3.37). A simple branching strategy
is used in this model, where the stacks are filled bottom-up from left to right, and the
container with the smallest index in the domain of the slot variable to be branched on,
is stowed in the slot. The improved model includes the symmetry-breaking and implied

57
Chapter 3. Slot Planning

100

80

Utilization Containers (%) 60

40

20

0
0 50 100 150 200
Capacity (TEUs)

Figure 3.2: Utilization as a function of location capacity.

constraints from Section 3.6.1. Its branching strategy is similar to the basic model,
however, the containers are assigned indices based on their features to avoid conflicts
with some of the new constraints introduced. Finally, the branching and advanced
models include the tailor-made branching strategy introduced in Section 3.6.2 and the
lower bounds introduced in Section 3.6.3, respectively.
Recall that we aim at solving CSPBDL instances within one second. The solver can
return an optimal solution before that, but after one second it must return its current
solution. The results are summarised in table 3.4, where each row corresponds to one
of the instance groups introduced in table 3.3.

Basic Improved Branching Advanced

Grp. All (%)
sol (%) opt (%) sol (%) opt (%) sol (%) opt (%) sol (%) opt (%)
1 100 100 69 100 85 100 100 100 100
2 82 86 54 82 50 91 90 91 91
3 92 92 61 92 69 100 100 100 100
4 74 82 11 83 61 96 90 96 95
5 36 53 11 69 44 89 75 92 89
6 80 80 7 93 40 100 93 100 93
7 36 71 0 64 14 64 29 64 57
8 50 79 14 79 29 93 93 93 93
9 47 76 18 65 12 76 47 76 76
10 75 62 12 87 37 100 62 100 100
11 67 67 17 83 33 83 67 83 83
Total 66 77 21 80 48 89 80 92 90

Table 3.4: Percentage of instances solved and proven to optimality by the CP models. The first
column provides the id of the group, the second one shows the percentage of instances of
the group solved by all CP models, and the subsequent columns the percentage of instances
solved and proven optimal by each model independently. The last row presents the results
for the complete set of instances.

58
 Chapter 3. Slot Planning

The total number of instances solved and proven optimal increases for each extension
of the basic model. A careful inspection of the table shows that this does not apply
to all groups individually, but overall the impact of the model improvements are quite
similar for each group. The total time needed for processing all instances is also reduced
considerably for each model, from 186.3 seconds for the basic model, to 126.9 seconds
for the improved model, 56.8 seconds for the branching model, and 34 seconds for the
advanced model. Instances not solved were accounted for by one second in the total
time.
To compare the runtimes of individual instances, we analysed the subset of 156
instances (66.1%), solved by all four CP models, (second column of Table 3.4). The
plot in Figure 3.3, shows the runtime of the models for each instance, where label BA
represents the basic model, IM the improved model, and BR and AD the branching
and advance models, respectively. We have sorted the instances such that the expected
runtime dominance between the models is clearly observable.

BA
1 IM
BR
AD

0.8
time (s)

0.6

0.4

0.2

0
0 20 40 60 80 100 120 140
instance

Figure 3.3: Runtime of the four CP models.

The plot in Figure 3.4 shows in the optimality gap of 39 out of the 156 instances,
that at least one model solved sub-optimally. The labels follow the same conventions
as in the left graph. Again, we have sorted the instances to highlight a quite robust
optimality dominance between the models.
Wrt. the hardness of different instance groups, the results in Table 3.4 indicate
that instances with many features (group 8-11), are slightly harder than instances with
few features. In particular, instances only stowing 40 ’ containers (group 1 and all
even groups except 2), are significantly easier than instances stowing both 20 ’ and
40 ’ containers. Interestingly, there is no significant positive correlation between space
utilization and runtime. This is surprising as one would expect the hardness of slot
planning problems to increase with space utilization.

59
Chapter 3. Slot Planning

500
BA
IM
BR
400

optimality gap (%) 300

200

100

0
0 5 10 15 20 25 30 35
instance

Figure 3.4: Optimality dominance of the three CP models with gaps.

Comparing the performance of the IP and CP models

In this section, we compare the performance of the IP and CP model using a one
second and 10 seconds runtime limit. Table 3.5 summarizes the percentage of instances
solved, proven to optimality, and runtime of the models over the instance groups. The
runtime of each instance group is calculated by summing up the time taken for solving
all the instances in the group. Unsolved instances were accounted for by the time limit
of the experiment. The upper table shows the results of the experiments with a one
second time limit. The CP model is clearly dominant in all groups in both, number
of instances solved and instances solved optimally. In total, the CP model solves 28%
more instances than the IP model in just a quarter of the time. When the time limit is
extended to 10 seconds (lower table in Table 3.5), the IP model increases the number
of instances solved by a 22% and the ones solved optimally by 17%. The runtime,
however, increases by a factor of five. The instances solved by the CP model increase
by a modest 1.7%, and the ones solved optimally by 3.8%, while the runtime increases
a factor of six. The total time of the CP model, however, is still around a quarter of
the IP model.
The hardness of instances for the IP model is similar to that of the CP model: in-
stances with many features are slightly harder than instances with few, and instances
stowing 40 ’ containers only are easier than instances stowing both 20 ’ and 40 ’ con-
tainers.
As in the CP experiment, we now focus on the instances solved by both models.
Table 3.6 summarizes the results of the experiments with 1 and 10 seconds time limit,
showing the percentage of instances solved by both models, only the IP model, only the
CP model, and none of them. For the one second time limit experiment, a total of 148
instances (63%), are solved by both models. All solutions produced by the CP model

60
 Chapter 3. Slot Planning

CP 1 sec. IP 1 sec.
Grp.
sol (%) opt (%) time (s) sol (%) opt (%) time (s)
1 100 100 0.1 100 100 1.7
2 91 91 3.6 59 59 11.6
3 100 100 0.5 54 54 8.9
4 96 95 6.0 87 79 27.7
5 92 89 7.1 33 28 31.3
6 100 93 1.2 100 93 4.0
7 64 57 6.8 29 21 11.2
8 93 93 1.5 43 36 10.6
9 76 76 5.2 24 24 14.8
10 100 100 0.7 62 62 4.4
11 83 83 1.3 50 50 3.5
Total 92 90 34 64 59 129.8
CP 10 sec. IP 10 sec.
Grp.
sol (%) opt (%) time (s) sol (%) opt (%) time (s)
1 100 100 0.1 100 100 1.8
2 91 91 21.6 95 91 50.4
3 100 100 0.5 92 85 35.3
4 99 99 19.7 96 94 87.0
5 92 92 39.0 72 56 192.0
6 100 100 5.4 100 93 13.0
7 64 64 53.5 64 29 102.8
8 93 93 10.5 79 64 74.1
9 88 88 36.5 53 41 112.3
10 100 100 0.7 88 62 31.5
11 83 83 10.3 67 50 30.5
Total 94 94 198 86 76 730.9

Table 3.5: Results for the CP and IP models with a 1 second (upper table), and 10 seconds
(lower table), limit.

are optimal. The IP model produced 2 suboptimal solutions with an optimality gap
of 67% and 214%. For the 10 seconds time limit experiment, the number of instances
solved by both models increased by 19% (45 instances). The number of suboptimal
solutions also increased for the IP model (from 2 to 8), with the optimality gap ranging
from 3.1% to 317%. The suboptimal instances are spread over 6 different groups.

1 sec. 10 sec.
Grp.
both (%) IP (%) CP (%) none (%) both (%) IP (%) CP (%) none (%)
1 100 0 0 0 100 0 0 0
2 55 5 35 5 82 14 4 0
3 54 0 46 0 92 0 8 0
4 86 1 10 3 95 1 4 0
5 33 0 58 9 67 6 24 3
6 100 0 0 0 100 0 0 0
7 29 0 36 35 50 14 22 14
8 43 0 50 7 71 8 21 0
9 24 0 53 24 53 0 35 12
10 62 0 38 0 88 0 12 0
11 50 0 33 17 67 0 17 16
Tot 63 1 29 7 82 4 11 3

Table 3.6: Results for the instances solved by any of the two models with 1 and 10 seconds
time limit.

61
Chapter 3. Slot Planning

3.7.2 Slot Planning Master Plans

Given that the CP model outperformed the IP model in the previous experiments, we
move a step forward and incorporate our CP model into a module for the automatic
generation of stowage plans introduced in Pacino et al. (2011). The CP model becomes
part of the slot planning component of this module, stowing the containers assigned to
each location independently. We limit slot planning to the current port, since this is
the only port that stowage coordinators stow. In the cases where the CP model cannot
stow the containers assigned to a location (8% in the previous experiments), a modified
version of the LS algorithm from Pacino and Jensen (2012), that allows rolling out
containers is used to generate the location slot plan. Removing a few containers from
the vessel has little impact on its stability.
For this experiment, we aim at generating stowage plans for 18 industrial stowage
conditions. In Pacino et al. (2011), the master planning phase problem is formulated as
a mathematical programming model. We solve this mathematical model as a MIP and
an IP model, and incorporate an evaluated impact of this decision into the generated
stowage plans. Master plans with a 5% optimality gap and a 3,600s time limit are
acceptable for slot planning. Solving the MIP model is considerably faster than the
IP model, but it has some drawbacks. The master planning IP model assigns integer
numbers of containers to locations. The MIP model, however, may assign fractional
numbers of containers to be stowed in locations, something that is physically impossible.
To deal with this issue, we attempt to stow a unit in the first location with available
capacity, where the largest container fraction has been assigned by the master plan.
If none of the locations have capacity left, the container is rolled out. The quality of
each slot plan is evaluated by comparing it with the best slot plan generated by CP for
the same location within twenty minutes. In the cases where containers are rolled out
by the LS algorithm, the slot plan is evaluated against the best plan generated by CP
within twenty minutes, stowing the same containers as LS, that is, discarding the rolled
out containers. The time limit to stow a location is one second. All experiments were
run on a Linux machine with two Six Core AMD Opteron processors at 2.0 Ghz and
32 GB of memory, with the CP model being implemented using the library Gecode 3.5.
Due to the non-deterministic nature of the LS algorithm, results of slot planning are
reported on average over 10 runs of the algorithm. Table 3.7 summarizes the results of
our experiments.
Vessels are slot planned fast by our approach. For the instances where the master
planning phase was solved, slot plans are generated within 17 seconds in total. When
the slot planning IP master plans (results of the master planning phase when solving it
as an IP model), a reduction in the number of containers rolled out due to fractionality
and by odd numbers of 20 ’ containers (remember that a cell cannot stow a single 20 ’),
in locations is observed in most of the instances. This is, however, a very small (0.4%
max), fraction. The maximum roll-out of an instance is 5.99%, a reasonable number
given the number of containers typically rolled from a loadlist by stowage coordinators.
Only two instances have an average optimality gap over 10%, due to the presence of

62
 Chapter 3. Slot Planning

Slot Planning Current Port

Time(s) Locs. Frac.+Odd Rolled out(%) Gap(%)
ID Conts.
Cont. Int. Cont. Int. Cont. Int. Cont. Int. Cont. Int.
1 990 3.48 4.96 34 34 4 2 0.40 0.20 0.00 0.00
2 2222 13.21 13.10 61 57 15 6 2.58 2.04 3.05 4.60
3 783 4.07 4.94 33 32 3 7 0.38 0.89 0.36 0.00
4 1876 16.44 13.64 69 70 14 5 1.76 1.08 12.77 10.47
5 573 8.94 6.08 29 25 10 5 1.75 0.87 1.41 0.45
6 862 6.83 5.26 27 35 5 6 0.93 0.70 1.25 0.00
7 238 1.52 4.08 12 14 1 1 0.42 0.42 0.00 0.00
8 334 0.99 0.66 14 13 4 4 2.10 1.20 1.57 0.00
9 1064 9.20 11.06 29 32 8 5 4.68 5.99 13.86 16.44
10 314 0.95 - 12 - 1 - 0.32 - 0.00 -
11 757 6.97 7.96 37 27 7 5 0.92 0.66 0.63 0.59
12 894 9.69 9.19 24 23 0 2 0.00 0.22 0.44 0.67
13 2190 13.57 10.73 61 58 7 5 1.30 0.23 0.78 0.03
14 901 5.21 - 30 - 14 - 2.51 - 6.73 -
15 302 1.79 1.46 16 14 2 6 0.66 1.99 0.00 0.00
16 1618 9.73 13.98 50 50 4 6 0.61 2.94 0.38 1.54
17 200 0.94 0.91 12 13 0 0 0.00 0.00 0.00 0.00
18 1261 9,82 7,73 38 37 6 4 0.48 0.40 2,01 9,75

Table 3.7: Slot Planning MIP (Cont.) and IP (Int.) master plans with a 5% gap. The first
and second columns are the id of the instance and the number of containers to stow in the
first port. The next columns show grouped results of slot planning, based on MIP and IP
master plans. The third and fourth columns depict the runtime for the slot plans, the fifth
and sixth columns are the number of locations. The seventh and eighth columns total the
number of rolled out containers by fractionality and by an odd number of 20 ’ types, and
finally the ninth and tenth columns are the percentage of total containers rolled out. The
last two columns show the average gap of the slot plans. A dash indicates that no master
plan was provided.

outliers, and the median of all instances is always 0%. CP generated optimal slot plans,
within one second for 91.8% of the locations of the MIP master plans and 94.8% of the
locations of the IP ones. Moreover, CP was able to prove optimality in 83% of the slot
plans generated for the locations from both, the MIP and IP master plans.

3.8 Conclusions
In this chapter, we have introduced an accurate model called CSPBDL for stowing a set
of containers in a location. The CSPBDL is NP-hard and is an important sub-problem
of the successful multi-phase approaches to stowage planning optimization. We have
developed two CP and IP models to solve the CSPBDL optimally. Our experiments
show that both models can be solved fast and that it is possible to improve the perfor-
mance of the CP model, such that it can produce optimal solutions for 90% of the 236
industrial instances in less than one second, which is well within the time requirements
for practical stowage support tools. We also tested the CP model, the best of the two
models according to our experimental evaluation, as a component of a module for the
automatic generation of stowage plans. In this setup, the CP model behaved similarly

63
Chapter 3. Slot Planning

to the first experiment, generating optimal slot plans for more than 90% of the loca-
tions and proving optimality for 83% of them. Future research includes improving the
performance and stability of our solvers (e.g., diving heuristics and other techniques
may be used to improve the IP model), and extending the CSPBDL to include on deck
locations and special containers such as out-of-gauge, pallet-wide, and containers with
dangerous goods.

64
 Chapter 4. Accurate Master Planning

Chapter 4

Accurate Master Planning

In this chapter, we present an accurate mathematical model for solving the first phase,
the master planning phase, of our 2-phase decomposition approach for scalable stowage
planning. During master planning, containers are assigned to vessel locations rather
than individual slots. The mathematical model formulated in this chapter includes
accurate stability and stress force calculations and ballast tanks. Ballast tanks are a
feature of stowage planning, widely used by stowage coordinators, that have been dis-
regarded in most of the previous academic work. This chapter presents an approach to
address small variations in the vessel displacement, which is a consequence of including
ballast tanks into the mathematical model that affects stability and stress calculations.
This chapter is based on Pacino, Delgado, Jensen, and Bebbington (2012).

4.1 Introduction
All of the mathematical models present in the literature that consider accurate sta-
bility and stress forces are based on the assumption that the displacement of the ship
(the total weight of the loaded vessel), is constant. Hydrostatics, such as buoyancy,
stability, trim, and draft are non-linear functions of the ship’s center of gravity and
vessel displacement. They can, however, easily be linearized when the displacement
is constant and translated to bounds on the position of the center of gravity, which
is also simple to compute when the displacement is constant. Container vessels have
ballast tanks that are used by stowage coordinators to better handle the stability of the
vessel and apply stowage configurations that are otherwise unfeasible. These tanks can
account for approximately 25% of the ship’s total displacement and including them in
a mathematical model brings forth a number of non-linear relations.
In this chapter, we present a mathematical model for the master planning phase
that considers ballast tanks and deals with the variable displacement. This model is
concerned only with the seaworthiness of the vessel, but can easily be extended to
optimize other aspects of the plan such as crane utilization. According to our indus-
trial collaborator, it is possible for stowage coordinators to make an educated guess

65
Chapter 4. Accurate Master Planning

on the amount of ballast water that a vessel might need that is within 15% of the
current amount in the tanks. We use this assumption to define a displacement range
within which we are able to formulate a linearization of the stability constraints with
an acceptable error. We analyse the accuracy of our model experimentally on 10 real
instances provided by our industrial partner. Our analysis suggests that within 5% of
the current displacement of the vessel the model is accurate enough to guarantee the
seaworthiness of the vessel.
The remainder of the chapter is organized as follows. Section 4.2 introduces related
work. Section 4.3 presents our model and sections 4.4 and 4.5 present the experimental
analysis and conclusions.

4.2 Literature Review

Even though several of the publications on stowage planning available in the past years
address stability and stress forces, very few present models that incorporate them as
constraints or objectives. Additionally, none of them consider variable displacement
due to ballast tanks. One of the most complete formulations of stability and stress
moments as part of an IP model is introduced in Botter and Brinati (1992). Though
their model is too complex to solve in practice, it constrains GM, transversal stability
(heel angle), trim, shear forces, and bending moments. Linearizations depending on
the displacement of the vessel (which is variable due to the inclusion of the loading
and unloading sequence of containers into the model), are suggested for the stability
constraints. No evaluation of the impact of these linearizations is presented, probably
due to the fact that their model was not used in practice. Shear forces and bending
moments are addressed, but they disregard the impact of the cargo on the buoyancy
force. The IP formulations introduced in Ambrosino et al. (2004) and Li et al. (2008),
assign containers from a loadlist to individual slots on the vessel. These models handle
transversal stability and trim by constraining the weight difference between transversal
(left and right), and horizontal (bow and stern), sections of the vessel to be within
certain tolerance. The GM is constrained by not allowing heavy containers on top of
light ones, a rule of thumb used in the industry for some vessels, but not necessarily for
all vessels. In Kang and Kim (2002), a model that assigns types of containers to sections
of the vessel is introduced. This model constrains the center of gravity of the vessel
with respect to precomputed constants to satisfy GM, trim, and transversal stability
constraints.
To the best of our knowledge, the only approach available in the literature that
considers the use of ballast tanks is presented in Aslidis (1984). A heuristic uses ballast
water to bring the longitudinal center of gravity within a permissible range defined,
based on the trim desired. Later, a local search is used to fix the GM.

66
 Chapter 4. Accurate Master Planning

4.3 Stability and Stress Model with Ballast Tanks

The introduction of ballast tanks into the optimization model causes the displacement
of the vessel to become variable. This makes the calculation of the center of gravity
non-linear. Thus, it is no longer possible to use the linearization of the hydrostatic data
introduced in Pacino et al. (2011). Two major non-linear relations must be handled
once variable displacement has to be modelled. The first is the calculation of the center
of gravity and the second is the linearization of the hydrostatic data. Consider the
calculation of the longitudinal center of gravity (lcg) without ballast tanks:

LM o + DlL vl
P
l∈L
, (4.1)
W

where LM o is the constant longitudinal moment of the vessel, DlL and vl are theP
lcg and
weight of a location l ∈ L, respectively. The displacement given by W = W o + l∈L vl ,
where W o is the constant weight of the vessel. Since all containers in the loadlist are
loaded, W is constant, which makes the calculation linear. Now consider the same
calculation where we include ballast tanks as a variable:

LM o + l∈L DlL vl + u∈U DuL vu

P P
P , (4.2)
W + u∈U vu

where DuL is the lcg of tank u, and vu is the variable defining the amount of water
to be loaded in tank u ∈ U . Since the amount of water in the tanks is not known a
priori, the displacement of the vessel is no longer constant, and the calculation becomes
non-linear. In order to deal with this, we propose the following approximation:

LM o + DlL vl + u∈U DuL (vu + ∆u )

P P
l∈L
LCG = P (4.3)
W + W T + u∈U ∆u
LM o + l∈L DlL vl + u∈U DuL (vu + ∆u )
P P
≈ , (4.4)
W + WT
where we model the ballast water estimation error of the stowage coordinator with
the variables ∆u . Thus, the total displacement becomes W + W T + u∈U ∆u , where
P
W T represents the weight of the estimated ballast water. We then make a linear
approximation of the lcg of the vessel by removing the allowed changes of ballast water
from the denominator of the fraction, resulting in equation (4.4). Given the total
capacity of the tanks (W BT ), the constant weight of an empty vessel W o ≈ 2W BT ,
and the weight of the cargo W C ≈ 6W BT , we can reasonably assume that the error in
the approximation of the lcg (given that stowage coordinators can estimate the ballast
within 15 % accuracy), is less than 0.15W BT /(2W BT + 6W BT + W BT ) = 1.7%. Note
that the same approximation can be used to calculate the vertical and transversal center
of gravity.

67
Chapter 4. Accurate Master Planning

The assumption that the weight of ballast water lies within a given interval is useful
for the linearization of the hydrostatic calculations. Hydrostatic calculations are in
practice linear approximations of given data points. When the center of gravity and
the displacement of the vessel are known, the linearization is accurate. For the problem
we are going to model, however, this is not the case since both the center of gravity
and the displacement can vary. Figure 4.1 shows a plot of the hydrostatic data for
the trim and metacenter calculation. The functions are clearly non-linear. Notice
that within a small displacement interval it is possible, however, to approximate the
functions accurately with a plane. This is only true for displacement levels that are not
at the extremes of the data tables, but it is reasonable to assume that the displacement
of a stowage plan will be within these extremes. The planes described above can be
defined by the limited ballast water change and thus be used in our model.
The buoyancy of a vessel is the volume of water that the vessel displaces. In order

Meta center (m)

Trim (m)

45

40
5
4 35
3
2
30
1
0
50000
-1
-2 -25 100000
250000 0
-20 150000 -5
200000 -10
-15 Disp.(t) 200000 -15
150000 -10 -20
Disp.(t) 100000 -5
Lcg (m) 250000 -30
-25 Lcg (m)
50000 0

(a) (b)

Figure 4.1: (a) Trim as a function of displacement and lcg (b) Metacenter as a function of
displacement and lcg.

to calculate this volume, it is necessary to know the shape of the vessel hull. For this
purpose, the hydrostatic data tables provide the possibility of calculating the submerged
area of a vessel at a specific point called a station. Figure 4.2 shows an example of such
areas and how stations are distributed along the vessel. Given two adjacent stations, s1
and s2 , the buoyancy of the vessel section between the two stations is approximated by
(As1 +As2 )D(s1 ,s2 )δ W
2
, where As is the underwater area at station s, from now on referred to
as bonjean, D(s1 , s2 ) is the distance between the two stations, and δ W is the density of
the water. As Figure 4.2 shows, stations are not evenly distributed along the vessel. A
greater concentration is found at the vessel’s extremities, where the hull shape changes
the most. Figure 4.3 shows two plots of the hydrostatic data related to the bonjean of
a station at the bow and a station in the middle of a vessel. As expected, the function
describing the hull at the bow is highly non-linear since the hull greatly changes, which
is not the case for stations in the middle of the vessel. Within specific displacement

68
 Chapter 4. Accurate Master Planning

Out of water area

Waterline

Area at station

Stations

Water diplacement

Underwater area

Figure 4.2: Areas for buoyancy calculation and distribution of stations

ranges, it is still possible to approximate the function linearly. Should one want to
model displacement ranges that include the most non-linear parts, piece-wise linear
approximations with a few binary variables can be used.

Bonjean area (m2)

Bonjean area (m2)

700

600
50
500
40
400
30
300
20
200 -30
10 -30
-25
-25
-20 -20
250000 250000 -15
-15 200000
200000 -10 Lcg (m) 150000 -10 Lcg (m)
150000 -5
-5 100000
Disp.(t) 100000
0
Disp.(t) 0
50000
50000

(a) (b)

Figure 4.3: (a) Underwater area as a function of displacement and lcg at a bow station (b)
Underwater area as a function of displacement and lcg at a middle station.

4.3.1 A MIP Model

Following the linear approximation described in the previous section, we propose a MIP
model for the master planning phase that includes ballast tanks modeling. Objectives
that focus on efficiency of the master plan (described in Pacino et al. (2011)) are
not included in the model under analysis as they do not have any influence on the

69
Chapter 4. Accurate Master Planning

seaworthiness of the vessel and are thus irrelevant to this study. Containers are grouped
into mutually exclusive types based on their features. A container type τ ∈ T is a 4-
tuple (l,h,r,w), where l ∈ L is the length of the container in feet, h ∈ H is the height of
the container, r ∈ R is the reefer properties of the container (reefer or non-reefer), and
w ∈ W is the weight class of the container. For the model presented in this chapter we
define T = {(l, h, r, w) | l ∈ L = {20 , 40 }, h ∈ H = {DC}, r ∈ R = {RF, N R}, w ∈
W = {LW, HW }}, where DC refers to dry-cargo containers, RF and N R to reefer and
non-reefer containers, and LW and HW to light and heavy weight classes. Further, let
T α be a subset of these types with an attribute value restricted to be α. For instance,
T RF represents the set of types with reefer capabilities.
When doing master planning it is necessary to consider the current port and the
downstream ports to generate robust stowage plans. We define a transport as a pair
of loading and discharging ports. Thus, the multi-port master planning phase assigns
types of containers to locations during transports. Notice that the container types T
are only a classification and thus are not bound to fixed weight ranges. For each type
τ ∈ T the average weight of the containers during a transport is calculated, causing
a more refined weight categorization than taking the average weight of the containers
in each container class. We remind the reader that the sets and constants used in
the model presented in this chapter have been previously defined in the section Sets
and Constants, at the beginning of this document. Now we define the variables of the
multi-port master planning model.
Decision variables
xtτ
l ∈Z Number of containers of type t to be stowed in location l during transport t.
xpu ∈ R+ Amount of ballast water in tons loaded at port p in tank u.

Auxiliary variables
ypu ∈ R+ Change in ballast water at port p in tank u.
vpW ∈ R+ Vessel displacement.
W
vpl ∈ R+ Weight of containers stowed at port p in location l.
W {20 ,40 } +
vpl ∈R Weight of 20 ’, 40 ’ containers stowed in location l at port p.
vpL ∈ R Longitudinal center of gravity at port p.
vpV ∈ R+ Vertical center of gravity at port p.
vpM ∈ R+ Metacenter at port p.
B
vps ∈ R+ Buoyancy force at port p of section between stations s and s + 1.
Sα
vpf ∈ R+ Shear force at port p fore or aft of frame f .
Bα
vpf ∈ R+ Bending moment at port p fore or aft of frame f .

The objective function of the MIP model for the evaluation of the accuracy of
stability and stress force calculations when modeling ballast tanks is:
minimize
XX
ypu (4.5)
p∈P u∈U

The accuracy of the approximations decreases with the extent of the change in ballast
water, thus, objective (4.5) minimizes this change. Now we introduce the capacity

70
 Chapter 4. Accurate Master Planning

constraints.
s.t.
X X
ΠTEU
τ xtτ TEU
l ≤ Kpl ∀p ∈ P, l ∈ L (4.6)
t∈TR ON
p
τ ∈T
X X
xtτ α
l ≤ Kpl ∀p ∈ P, l ∈ L, α ∈ L (4.7)
τ ∈T α
t∈TR ON
p
X X
xtτ RF
l ≤ Kpl ∀p ∈ P, l ∈ L (4.8)
t∈TR ON
p τ ∈T RF

X X
ΠFEU
τ xtτ RC
l ≤ Kpl ∀p ∈ P, l ∈ L (4.9)
t∈TR ON
p τ ∈T RF

X X
Wtτ xtτ Wα
l = vpl ∀p ∈ P, l ∈ L, α ∈ L (4.10)
t∈TR ON
p
τ ∈T α
W 20
vpl ≤ Wpl20 ∀p ∈ P, l ∈ L (4.11)
W 40
W 20
0.5vpl + vpl ≤ Wpl40 ∀p ∈ P, l ∈ L (4.12)
W 20 W 40 W
vpl + vpl = vpl ∀p ∈ P, l ∈ L (4.13)
All the weight limits and capacities of the model have been reduced to account for the
release. The TEU capacity of each location is enforced by constraint (4.6). Location
specific capacity requirements regarding the length of the containers are enforced using
constraint (4.7). Constraint (4.8), and (4.9) limit, respectively the total number of
reefer TEU and Forty-foot Equivalent Units (FEU) used. With constraint (4.10) we
define the variables vlW 20 and vlW 40 , and use them in constraint (4.11) and (4.12) to
limit the weight of the 20 ’ and 40 ’ containers stowed in a location. Constraint (4.13)
defines the auxiliary variable vlW . Now we introduce the constraints, in order to load
all containers and to limit the changes in ballast water.
X
τ
xtτ
l = LD t ∀τ ∈ T , t ∈ TR (4.14)
l∈L
xpu ≤ Ku ∀u ∈ U (4.15)
(Eu − u ) ≤ xpu ≤ (Eu + u ) ∀p ∈ P, u ∈ U (4.16)
Constraint (4.14) ensures that all containers are loaded. Constraint (4.15) defines the
capacity of the tanks and constraint (4.16) limits their change in ballast water. Rather
than considering the allowed ballast change for all the tanks combined, constraint (4.16)
limits the ballast change for each tank u independently. Though limiting the change of
each ballast tank independently reduces the stowage configurations that can be applied,
this approach minimizes the error in the calculations of the center of gravity of the tanks.
Now we introduce the stability constraints of the model:
X X
W
xpu + vpl + WpO = vpW ∀p ∈ P (4.17)
u∈U l∈L

71
Chapter 4. Accurate Master Planning

W
DlL vpl DuL xpu + LM O
P P
l∈L + u∈U p
= vpL ∀p ∈ P (4.18)
Wp
V W
DuV xpu + VM O
P P
l∈L Dpl vpl + u∈U p
= vpV ∀p ∈ P (4.19)
Wp
M inT ≤ AW W
T vp + ALT vpL + AT ≤ M axT ∀p ∈ P (4.20)
M in ≤ AW
DA W L L
DA vp + ADA vp + ADA ≤ M ax
DA
∀p ∈ P (4.21)
AW W L L
DF vp + ADF vp + ADF ≤ M ax
DF
∀p ∈ P (4.22)
AW W L L
M vp + AM vp + AM = vp
M
∀p ∈ P (4.23)
M V GM
vp − vp ≥ M in ∀p ∈ P (4.24)

The displacement of the vessel is defined by constraint (4.17). Variable v L is defined

by constraint (4.18) using the approximation defined in (4.4). The same approximation
is used in constraint (4.19) for the calculation of the Vertical center of gravity (vcg).
Note that the approximation on the vcg of location l considers the release and therefore
changes at each port p. Constraint (4.20) represents the linearized hydrostatic calcula-
tion of trim. Given the displacement of the vessel v W and its lcg v L , constraint (4.20)
approximates the trim with a plane defined by the coefficients AW L
T , AT , and AT . The
calculated trim is then kept within limits. Changing the coefficients accordingly, con-
straint (4.21) and (4.22) approximate the aft and fore draft of the vessel. Both drafts
are kept within the maximum limits. Due to the propeller it is also necessary to con-
strain the aft draft to be at a minimum depth. The metacenter is also calculated using
the hydrostatic approximation and it is defined in constraint (4.23). The GM is then
calculated in constraint (4.24) and kept above the security limit. Below the calculation
of the buoyancy force and the constraints that limit shear force and bending moment
at different points in the vessel.
X
0.5DbB AW W L L B
Bs vp + ABs vp + ABs = vpb ∀p ∈ P, b ∈ B (4.25)
s∈{b,b+1}
X X X
WfS,α + Gαlf vpl
W
+ Gαuf xpu − Gαbf vpb
B Sα
= vpf
l∈L u∈U b∈B
∀p ∈ P, f ∈ F, α ∈ {A, F } (4.26)
X X X
WfB,α + Aαlf Gαlf vpl
W
+ Aαuf Gαuf xpu − Aαbf Gαbf vpb
B Bα
= vpf
l∈L u∈U b∈B
∀p ∈ P, f ∈ F, α ∈ {A, F } (4.27)
M inSf SF
≤ Gf vpf SA
+ (1 − Gf )vpf ≤ M axSf (4.28)
M inBf
BF
≤ Gf vpf BA
+ (1 − Gf )vpf ≤ M axB f (4.29)

The buoyancy of the section of a vessel between two adjacent stations is approximated
by constraint (4.25). In the shear and bending calculations, we take into account the
forces aft or fore of a frame. Since frames do not always coincide with the starting

72
 Chapter 4. Accurate Master Planning

points of tanks, locations or buoyancy stations, it is necessary for a given frame f ∈ F

(a fixed calculation point), to know the fraction of weight that needs to be taken into
account from a location l, tank u, or station b, which is the purpose of the constants
GA F
{l,u,b}f and G{l,u,b}f . Shear force and bending moment are calculated at each vessel
frame and must be within limits. Since weight and buoyancy components are computed
based on approximations, there is a significant accumulation of error when calculating
forces and moments acting at each frame. To tackle this issue, constraints (4.26) and
(4.27) compute shear forces and bending moments, respectively, both, wrt. resulting
forces acting fore and aft of the frame. Since the fore-based computation is accurate
in the bow and the aft-based computation is accurate in the stern, we base a final
estimate of the shear forces and bending moments on a proportional blend of the aft
and fore computations over the length of the vessel. Thus, the combined shear force is
constrained to be within limits by (4.28), while constraint (4.29) does the same for the
bending moment. Finally, the constraints that compute the change in ballast water:

Eu − xpu ≤ ypu ∀p ∈ P, u ∈ U (4.30)

xpu − Eu ≤ ypu ∀p ∈ P, u ∈ U (4.31)

Constraints (4.30) and (4.31) provide a lower bound for the cost variable yu quantifying
the changes in tank configuration from the initial estimate.

4.4 Analysis of Model Accuracy

The model has been evaluated experimentally in a case study involving 10 industrial
stowage plans for a vessel of approximately 15.000 TEUs. We constrain ourselves,
without loss of generality, to analyse the model for one loading port and several dis-
charge ports, i.e., with p0 being the current port, we define the set of transport as
TR = {< pi , pj > |pi = p0 , pj ∈ P }. Also, without loss of generality, we remove the
integrality constraint from the decision variables xtτ
l and allow the assignment of a con-
tinuous number of containers to a location. In Pacino et al. (2011), it has been shown
that this decision has a reduced impact on the slot planning phase, while considerably
reducing CPU time. The linear approximations used by our now LP model for the
generated solutions are compared with exact manual calculations. Table 4.2 gives an
overview of the characteristics of the instances. We performed experiments allowing
different changes in displacement and observed how the accuracy of the model changed
accordingly. First we consider the linear approximation of the center of gravity of the
vessel. Figure 4.4 shows two graphs describing how the approximations of the longitu-
dinal (a), and vertical (b) center of gravity behave as the displacement changes. For
both graphs, the horizontal axis represents the change of displacement in percent, while
the vertical axis represents the error in meters. Each point in the graph is generated by
forcing changes in the amount of ballast water of the 10 test instances. In Figure 4.4 (a)
it is possible to see, as expected, that when the displacement is unchanged the value of
the lcg is accurate, and the more the displacement moves away from its true value, the

73
 Chapter 4. Accurate Master Planning

Instances Characteristics
TEU (%) Weight (%)
ID Displacement (tons) Tanks (tons)
Total Release Load Total Release Load
1 91.79 39.05 52.74 32.05 13.12 18.93 148520 5160
2 73.83 36.66 37.17 45.43 23.14 22.29 176008 8161
3 59.74 18.45 41.29 42.23 12.07 30.16 168696 6952
4 80.98 28.93 52.06 53.01 19.91 33.10 192286 6962
5 65.65 13.38 52.28 25.76 8.00 17.76 135396 10578
6 48.61 18.01 30.60 24.89 8.83 16.06 133489 10578
7 69.27 28.29 40.98 40.86 17.40 23.46 161492 5199
8 45.81 13.16 32.64 28.84 7.77 21.07 143815 9976
9 58.56 24.89 33.67 30.14 13.78 16.36 141487 5239
10 59.04 20.47 38.57 31.97 9.96 22.01 146120 7341

Table 4.2: Characteristics of the test instances. Starting from the left the first three columns
indicate: the ID of the instance, the total utility percentages in terms of TEU capacity
used, the percentage of containers in the release and in the loadlist. The next three columns
indicate percentages of utilization in terms of weight, in total, for the containers in the
release, and in the loadlist. The initial displacement and the estimated ballast water are
given in the last two columns.

0.9 1.6
0.8 1.4
Vcg approx. error (m)
Lcg approx. error (m)

0.7 1.2
0.6
1
0.5
0.8
0.4
0.6
0.3
0.2 0.4
0.1 0.2
0 0
-8 -6 -4 -2 0 2 4 6 8 -8 -6 -4 -2 0 2 4 6 8
Displ. variation (%) Disp. variation (%)
(a) (b)

Figure 4.4: (a) Error in the lcg. (b) Error in the vcg.

less accurate the approximation becomes. Note that for a displacement range of 5%,
the calculation inaccuracy is at most 0.3 meters and is thus still accurate enough for
practical usage. The calculations for the vcg, however, are not as accurate, (Figure 4.4
(b)). Within the 5% range, the linearized value is at most 0.8 meters from the correct
one. This was an expected result, as it is not possible to precisely estimate the vcg of
locations since we do not know where the containers will be stowed. The accuracy error
of the vcg is, however, still very small.
Now that we have shown that linearizations for the center of gravity are accurate,
we focus on analyzing the accuracy of the hydrostatic data linearization. In Figure 4.5

74
 Chapter 4. Accurate Master Planning

0.4 1.6
0.35 1.4
0.3 1.2
Trim error (m)

GM error (m)
0.25 1
0.2 0.8
0.15 0.6
0.1 0.4
0.05 0.2
0 0
-8 -6 -4 -2 0 2 4 6 8 -8 -6 -4 -2 0 2 4 6 8
Disp. variation (%) Disp. variation (%)
(a) (b)

Figure 4.5: (a) Error for the trim (b) Error for the GM .

we use the same graph as before, with the horizontal axis describing the percentage
displacement changes and the vertical axis the calculation error. Figure 4.5 (a) depicts
the error for the trim which, as it can be seen, is very small. Within a 5% displacement
range the error is, at most, 15 centimeters. Figure 4.5 (b) shows the same analysis for
the calculation of GM . Notice that for both calculations, the trim and GM , an error is
present even at constant displacement. The error we see is due to the linearization of
the hydrostatic functions. For GM it also includes the approximation error of the vcg.
We now move our focus to the linearization of the bonjean. Due to the changes in
the shape of the hull in the extremities of the vessel we expect the bonjean calculations
to be the most inaccurate in the sections close to the bow and stern, while behaving
more smoothly towards midship. Figure 4.6 (a) shows the same analysis we have done
before applied to the bonjean areas. The graph shows the maximum error over all
bonjean areas as a function of the displacement changes. As can be seen, the variation
in displacement is not the main source of error. Most of the inaccuracy is due to
the linearization of the hydrostatic data. Figure 4.6 (b) shows how the bonjean error
is concentrated at the extremities of the vessel where the hull changes most. The
horizontal axis represents the position of the station on the vessel (where 0 is at the
bow), and the vertical axis is the bonjean error. Notice that the largest errors are
found for stations at the bow. This can be explained by the fact that the range of
drafts in our test data forces the linearization of these bonjean areas to be right by the
inflection point of the hydrostatic curve (see figure 4.3 (a)), where the linearization is
most inaccurate. Better approximations can then be expected for larger drafts, as it
is the case for the bonjean of the stations at the stern. The inaccuracy of the bonjean
is, however, still quite small if we consider that in the worst case scenario, there is an
error of only 4 square meters over an area of over 111 square meters.
Shear forces and bending moment calculations depend on the buoyancy of the ves-

75
 Chapter 4. Accurate Master Planning

4.5 4.5
4 4
3.5 3.5
Bonjean error (m2)

Bonjean error (m2)

3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
-8 -6 -4 -2 0 2 4 6 8 0 50 100 150 200 250 300 350 400
Disp. variation (%) Pos. of stations from the stern (m)
(a) (b)

Figure 4.6: (a) Approximation error of the total bonjean as a function of displacement change.
(b) Approximation error of the bonjean areas per station.

sel, which in turn is calculated using the bonjean approximations. Figure 4.7 (a) shows,
in the same way as the other graphs, how the percentage error in the shear calculation
(the vertical axis), behaves as a function of the variation of the displacement, (the hor-
izontal axis). As expected, the dominant error is not the approximation of the center
of gravity of the vessel, since the inaccuracy is more or less the same independent of
how much the displacement changes. A tighter relation can be seen when the shear
calculation is related to the error in the bonjean linearization. Figure 4.7 (b) shows the
percentage error of the shear force calculation as a function of the total error in the
bonjean area from the hydrostatic linearization, with the data points grouped accord-
ing to their percentage of displacement variation. As depicted, the error in the shear
calculation increases with the error in the bonjean linearization for all the displace-
ment variations. Moreover, for the data points with little or no displacement change
(displacement variation between [−0.1, 0.1]), we can see that the tendency remains the
same.
Figure 4.8 (a) shows the error in the bending calculation as a function of the changes
in displacement. As expected, similarly to the shear force calculation, there is no clear
tendency indicating a relation between the error in the bending calculations and the
approximation of the lcg. With respect to the impact of bonjean error in bending
calculations, we expect a similar tendency to that of Figure 4.7 (b), since bending
moment and shear force calculations depend similarly on the vessel buoyancy. Moreover,
due to the fact that the bonjean error is amplified by the multiplication of the arm, it
is interesting to look at the error at each calculation frame in order to see how the error
of the bending moment changes. We do this by forcing the displacement to remain
constant, thus removing the error of the lcg approximation, and analyzing how the
bending error changes at each calculation frame as the draft changes. The result is
shown in Figure 4.8 (b), where the horizontal axis represents the frames of the vessel

76
 Chapter 4. Accurate Master Planning

1.4 3
[-10,-5] [0.1,1]
1.2 [-5,-2.5] [1,2.5]

Shear force error (%)

2.5
Shear force error (%)

[-2.5,-1] [2.5, 5]
1 2 [-1,-0.1] [5,10]
[-0.1,0.1]
0.8 1.5
0.6
1
0.4
0.5
0.2
0
0 2 4 6 8 10 12 14 16 18 20
-8 -6 -4 -2 0 2 4 6 8
Total bonjean error (m2)
Disp. variation (%)
(a) (b)

Figure 4.7: Approximation error of the shear forces as (a) a function of displacement change,
and (b) a function of the total bonjean error with data points grouped by the % of dis-
placement variation.

and the vertical axis is the percentage of error in the bending moment calculation. Each
of the lines plotted in the graph represents one of the 6 test instances selected for this
experiment, each of which has a different draft. The 4 instances not included in the
experiment had a similar draft and followed the same tendency as that illustrated by the
instances plotted in Figure 4.8 (b). According to the bending calculations in our model,
the bending error should be low in frames near the vessel ends since the most accurate
bending calculations are chosen at those points, while the error should increase close
to midship, since at this point the arm length reaches its highest value and amplifies
the bonjean error. This, however, is not the tendency depicted in Figure 4.8 (b), where
the bonjean error seems to even out as it moves away from the vessel ends, even after
being amplified by the arm. This indicates that the impact of the arm in the bending
error is considerably reduced by the bending calculation introduced in this chapter.
One more thing worthy of notice is that there is no direct relation between the bending
error and the draft of the vessel. This is due to the non-linear shape of the hull.
The bending error for constant displacement does not exceed 1.4% (see Figure 4.8
(b)), however, when variable displacement is considered, errors of up to 5% might
be reached within a 5% displacement variation, (see Figure 4.8 (a)). We consider
these approximations acceptable. Higher accuracy can be achieved by reducing the
linearization error of the hydrostatic functions for the bonjean.

4.5 Conclusion
This chapter introduced a mathematical programming model including ballast tanks
for the master planning phase of a 2-phase stowage planning optimization approach.

77
 Chapter 4. Accurate Master Planning

6 1.4
Draft 10.76
Bending approx. error (%)

Bending approx. error (%)

1.2 Draft 11.43
5 Draft 11.52
1 Draft 11.72
4 Draft 12.53
0.8 Draft 14.38
3
0.6
2
0.4
1 0.2
0 0
-8 -6 -4 -2 0 2 4 6 8 0 5 10 15 20 25
Disp. variation (%) Frames from the AP
(a) (b)

Figure 4.8: Approximation error of the bending moment, (a) as a function of displacement
change, and (b) at each bonjean station for constant displacement.

Analysis of 10 real instances show that our model is successful at coping with vari-
able displacement with acceptable error tolerance. Within a 5% band of the current
displacement, the master plans generated were seaworthy, with the error in stability
calculations increasing proportionally to the variability of the displacement. In future
work, we plan to introduce piece-wise linearizations of the bonjean hydrostatic data to
reduce the error in the stress calculations in our model. This must be done carefully
since it might be necessary to include boolean variables that will negatively impact the
performance of the solver.

78
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

Chapter 5

Heuristic Optimization for

Interactive Stowage Planning

Chapter 3 and 4described models for the automatic generation of stowage plans. They
correspond to the second and first phase, respectively, of a 2-phase hierarchical decom-
position described in chapter 1, (see Figure 1.3). In this chapter, we present a heuristic
approach used in Angelstow1 , a commercial decision support tool for the interactive
generation of stowage plans based on stowage coordinators’ know-how. This approach
mimics the 2-phase decomposition mentioned above, incorporating stowage coordina-
tors’ expert knowledge to tackle important objectives (e.g., crane split, overstowage),
and constraints (e.g., stability and stresses forces), into an LP model to solve the master
planning part (first phase), while using a greedy algorithm to solve the slot planning
part (second phase). To guarantee a smooth flow of information with the stowage co-
ordinator, stowage plans are generated within a few seconds. This chapter is based
on Delgado et al. (2012).

5.1 Introduction
Liner shipping companies are in their infancy using optimization-based decision support
tools for stowage planning. Most of the research on stowage planning during the past
twenty years has focused on algorithms to automatically generate stowage plans (e.g.,
Gumus et al. (2008); Ambrosino et al. (2010); Yoke et al. (2009); Pacino et al. (2011)),
limiting the role of expert users to result validation. Though some of these approaches
generate good quality stowage plans, difficulties arise using them in practice. Stowage
planning is full of details disregarded in the models solved by automatic approaches.
Stowage coordinators need to be able to easily adapt automated plans to specific situa-
tions, or be directly involved in their generation for these plans to be successfully used
in practice.
Angelstow is a decision support tool that assists stowage coordinators during the
1
Ange Optimization ApS, available for purchase at www.http://ange.dk/main/angelstow.

79
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

planning process. It provides an interface for defining planning scenarios (i.e., a set
of high and low level preferences that define the backbone of a stowage plan), and the
capability of turning these scenarios into stowage plans. Using these features, stowage
coordinators can stow vessels interactively. A coordinator devises a planning scenario
that Angelstow turns into a stowage plan, providing feedback into the impact of the
coordinators’ preferences. If the resulting stowage plan is not satisfactory, stowage
coordinators can change their scenario and let Angelstow generate a new stowage plan.
To avoid that this interactive process jeopardizes the usability of the tool, fast stowage
planning is a key factor. A desired runtime for a stowage planning algorithm in this
setup is at most two seconds on average.
In this chapter, we introduce a 2-phase heuristic optimization approach that gener-
ates stowage plans for containers to be loaded at a single port. This approach mimics
the decomposition depicted in figure 1.3. In the master planning phase2 the planning
scenario is incorporated into an LP model that assigns containers to locations, while
in the slot planning phase a greedy algorithm stows the containers assigned to each
location into vessel slots. We evaluate our heuristic optimization approach on three
real-life stowage plans. Each stowage plan corresponds to the stowage condition of
a large vessel of approximately 15,000 TEUs after leaving the last of six consecutive
ports, with the vessel being empty in the first port. To carry out our experiments, we
extract the planning scenario that a stowage coordinator devised at each port and then
let our approach generate stowage plans for all ports. In this way, we achieve a fair
comparison between the stowage plans generated by our approach and those generated
by the stowage coordinators.
Our results show that the runtime of our approach is reasonable, (1.27 seconds on
average). Moreover, it rolls out on average 0.58% and at most 1.89% of the containers
to be transported, and utilizes stack heights as efficient as the stowage coordinators.
Additionally, we evaluate the impact of producing master plans with an LP model,
which implies that a continuous number of containers might be assigned to locations,
something physically impossible. The LP model is transformed into an IP model by
adding the corresponding integrality constraints. We solve the IP model with an opti-
mality gap of 1% and 5% and generate stowage plans based on these new master plans.
Then we compare the results with the original plans generated based on the LP master
plans. Our experiments show that for our set of instances there is no dominance from
stowage plans based on the IP model with respect to those based on the LP model,
while the solution time considerably increases. The remainder of the chapter is orga-
nized as follows: Section 5.2 describes the problem. Section 5.3 introduces Angelstow.
Section 5.4 describes related work and section 5.5 presents our heuristic optimization
approach. Finally, Section 5.6 and 5.7 present the experiments and draw conclusions.

2
Note that the master planning phase of the heuristic introduced in this chapter differs from the
one presented in chapter 4. It neither includes specific constraints for stability and stress forces nor
consider ballast tanks.

80
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

5.2 Problem Definition

In this chapter, we focus on the generation of stowage plans based on planning sce-
narios defined by stowage coordinators. A planning scenario encapsulates a set of high
level preferences, i.e., selection of a single discharge port for all containers stowed in a
location and upper bounds for the weight utilization of bays, and low level preferences,
i.e., requiring that certain containers are stowed in certain vessel slots. They embed the
know-how of stowage coordinators with respect to how to tackle complicated combi-
natorial objectives and constraints, and how to stow containers with special demands,
e.g., IMO and Pallet-wide. The high level preferences focus on minimizing hatch over-
stowage to avoid extra crane moves, and on the even distribution of moves over the
length of the vessel to minimize the makespan of the quay cranes. Moreover, they must
ensure seaworthiness of the vessel with respect to shear force and bending moments
limits, as well as line-of-sight, trim, draft and GM requirements.
Master planning assigns containers to locations following the planning scenario de-
vised by the stowage coordinator and satisfying the different capacities of each location
according to the features containers may have, e.g., reefer containers limited by the
available reefer plugs, volume capacity of 20 ’ and 40 ’ containers. A feasible slot plan
must follow the assignment of containers specified in its corresponding master plan and
form container stacks that satisfy stacking rules and constraints. A stowage plan must
satisfy the rules specified in section 3.2 for the slot planning model presented in chap-
ter 3, together with two extra rules: the planning scenario must be respected and no
overstowage is allowed.
In addition, we follow the set of rules of thumb listed below to generate high quality
stowage plans:

• Containers are distributed evenly among stacks on deck

• The number of used stacks is minimized in storage areas below deck

• Wasted volume capacity is minimized

• It is preferred to stow containers below deck, if possible

• Heavy containers should go below deck

• Preferential loading of reefer containers over dry containers

• Heavy containers cannot be stowed on top of light containers on deck

5.3 Angelstow
Angelstow provides an interface that allows stowage coordinators to define planning
scenarios. In these scenarios, a stowage coordinator defines the discharge port of the
containers to stow in a location. Only a single discharge port is allowed for each

81
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

location. This is seldom a limitation in practice since most vessels are split into sufficient
locations to allow only stowing containers of one discharge port. Moreover, due to the
robustness of this approach with respect to avoiding overstowage in downstream ports,
it is considered good practice by stowage coordinators.3
Automatic stowage planning can be activated at any time, but only the locations
for which the stowage coordinator has defined a discharge port will be stowed. The
stowage coordinator can also set upper bounds on the total weight of containers stowed
in a bay. In this way, scenarios only covering vessel sections for which a discharge port
has been specified can be stowed without the stowage coordinator losing control of the
overall weight distribution over the vessel. This is essential for achieving the desired
draft, trim, GM, and stress forces in the final stowage plan.
Recall that a stowage plan is an assignment of the containers to load in a port to
specific vessel slots. To validate a planning scenario, that is, to verify whether the
stowage coordinator’s preferences can be turned into a valid stowage plan, the stowage
coordinator must generate an actual stowage plan that follows the preferences specified
in the scenario. This implies stowing a considerable number of containers for testing
each scenario. Stowage coordinators generate plans under time pressure and uncer-
tainty, since they receive the loadlist of containers a few hours before the vessel calls
port. Due to these conditions, the number of planning scenarios that can be analyzed
is very limited. To address this issue, Angelstow provides stowage coordinators with
the capability of automatically turning these scenarios into stowage plans. The gener-
ation of a stowage plan then becomes the interactive process described next. First, the
stowage coordinator deviss a planning scenario. Next, Angelstow generates a stowage
plan based on the scenario specified by the stowage coordinator. The stowage coordina-
tor can modify the plan that Angelstow generated by manually re-stowing containers.
In case it is not possible to adapt the stowage plan to what the stowage coordinator
considers a reasonable solution, the planning scenario can be changed and the process
starts over again. To avoid this interactive process, to jeopardize the usability of the
tool, fast stowage planning is a key factor, which is a concern addressed by the heuris-
tic optimization approach presented in Section 5.5. We estimate that a time limit of
two seconds for generating stowage plans is necessary to efficiently support interactive
optimization in Angelstow.
A multi-port view of stowage planning is important to reduce the negative impact
of a stowage plan in downstream ports. A bad stowage plan can considerably reduce
the vessel capacity for containers to load in upcoming ports. Angelstow provides a
multi-port view (see Figure 5.1), where stowage coordinators can see the effect of their
decisions in downstream ports. Additionally, even though stowage coordinators focus
their work on the current port, information available in loadlists for downstream ports
(i.e., forecasts), can be used to define planning scenarios for these ports as well.
When a stowage coordinator wants to check the feasibility of the scenarios formu-

3
A disadvantage of following this approach rigidly is a potentially less efficient usage of the volume
and weight capacity of the vessel.

82
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

Figure 5.1: Front view of Angelstow. To the left, a summary of the loadlist of containers
and a view of the current state of the ballast tanks. In the center, the multi-port view of
Angelstow. Each row represents a port where the vessel calls. To the right, the stresses
graphs and weight distribution along the vessel bays.

lated for each of the ports a vessel calls, Angelstow generates stowage plans for each
port independently using the approach introduced in section 5.5. Figure 5.2 depicts the
process that Angelstow follows to generate stowage plans for all the ports the vessel
calls. At each port, the stowage planning component uses as input 1) the planning

Port 1 Port 2 Port n

Automatic Stowage Automatic Automatic Stowage
Vessel Data Stowage Plan Stowage ... Stowage Plan
Planning Port 1 Planning Planning Port n

Planning Planning Planning

Scenario Scenario Scenario
LoadList Loadlist Loadlist

Figure 5.2: Multi-port stowage planning.

scenario defined by the stowage coordinator from Angelstow, and 2) a loadlist of con-
tainers for the port. The containers stowed by Angelstow in the previous ports have
fixed positions and become part of the input for the next port. Thus, the stowage plan
made for the current port translates into a list of release containers for the next down-
stream port. A more sophisticated approach to multi-port stowage planning would be
to represent the problem in a single optimization model, as done in chapters 3 and 4.
This has not been done since 1) the impact is limited because most of the important

83
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

decisions with respect to multi-port planning are part of the planning scenario, and 2)
this may deteriorate the runtime performance.

5.4 Literature Review

Academic work on stowage planning can be divided into two main categories: single-
phase and multi-phase approaches. Single-phase approaches have a plain representation
of the stowage planning problem, (e.g., stow containers into vessel slots). A common
denominator among these approaches is to sacrifice model accuracy to achieve scala-
bility. Early work in this category (e.g., Scott and Chen (1978); Aslidis (1984)), intro-
duced heuristic approaches that solved simplifications of the stowage planning problem.
These approaches sequentially refine solutions by applying placement heuristics, local
search, and solving IP models. Later on, Botter and Brinati (1992), introduced an IP
model of the first accurate but intractable formulation of the stowage planning problem.
Constraint programming (Ambrosino and Sciomachen (1998)), metaheuristics such as
genetic algorithms (e.g., Davidor and Avihail (1996); Dubrovsky et al. (2002)), and sim-
ulated annealing (Flor (1998)), and more recently IP (e.g, Ambrosino and Sciomachen
(2003); Giemsch and Jellinghaus (2004); Li et al. (2008)), have been used to solve still
simplified versions of the stowage planning problem. A case-based reasoning approach
is proposed in Nugroho (2004), where stowage plans are generated by remembering how
similar past plans were made. The system suggests possible solutions to the stowage
coordinators that can be turned into stowage plans.
Multi-phase approaches decompose the problem hierarchically into two or more
layers of abstraction, (e.g., first assign containers to bay sections, then stow them into
vessel slots). These approaches are currently the most successful in terms of model
accuracy and scalability. The first work of this type appeared in the early 1970s, where
Webster and Van Dyke (1970), introduced a 3-phase heuristic. Twenty years later, a
model that includes several major aspects of the problem in a 2-phase approach was
introduced in Wilson and Roach (1999). This approach solves the stowage planning
problem for multiple ports. It uses a branch-and-bound algorithm for solving the first
phase, where containers are assigned to sub-sections of the vessel, and applies a tabu
search for stowing containers into vessel slots in the second phase. Other approaches
that use similar decompositions include solving multi-port stowage planning with an
iterative improvement approach. This is based on the transportation simplex method
(Kang and Kim (2002)), a bin-packing heuristic (Zhang et al. (2005)), and solving a
mixed-integer program for the first phase, and constraint programming and constraint-
based local search models for the second one, (Pacino et al. (2011)). 3-phase approaches
include combinations of constructive heuristics, 0/1 IP, metaheuristics (e.g., Ambrosino
et al. (2010)), and heuristics combined with LS, (e.g., Yoke et al. (2009)).
Multi-phase approaches developed by the industry include a multi-stage placement
heuristic using a number of lower-bounds (Gumus et al. (2008)), and a combination
of sequential LPs for assigning containers to bay sections, and a hierarchy of IPs for

84
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

stowing them into vessel slots (Guilbert and Paquin (2010)).

Several software tools are available on the market for container stowage planning
decision support, (e.g., Navis (2013); INTERSCHALT maritime systems AG (2013);
Muller+Blanck Software GmbH (2013); Autoship Systems Corporation (2013)). All of
them assist stowage coordinators with the generation of stowage plans by providing
feedback on partial stowage planning decisions. To the best of our knowledge, our
heuristic is the first attempt to use optimization techniques for interactive generation
of stowage plans.

5.5 Heuristic Optimization Approach to Stowage

Planning
In this section, we introduce a stowage planning heuristic based on the 2-phase decom-
position presented in chapter 1. This heuristic stows containers following a planning
scenario devised by a stowage coordinator using Angelstow, and fulfills and optimizes
all constraints and objectives specified in section 5.2. In the master planning phase,
containers are grouped together by their features into types and assigned to locations
by an LP model. In the slot planning phase, a greedy algorithm finds specific slots for
all containers following the assignment made by the master planning phase. Currently,
we limit our scope to consider containers that are 20 ’ and 40 ’ long and model reefer
and high-cube characteristics.

5.5.1 LP Model for Master Planning

Here we present the LP model that solves the master planning phase of our heuristic
optimization approach. Containers are grouped into types based on their features.
Recall that a type τ ∈ T is a 4-tuple, (l, h, r, w). For the model presented in this
section we define T = {(l, h, r, w) | l ∈ L = {20 , 40 }, h ∈ H = {DC, HC}, r ∈ R =
{RF, N R}, w ∈ W }, where HC refers to high-cube containers and w ∈ W is the
weight class of the container defined by the weight of the container rounded to the
nearest integer ton. The LP model assigns types of containers loaded on a single port
to locations. Though all containers are loaded in a single port they have multiple
discharge ports. With p0 ∈ P being the loading port (i.e., the port where the containers
are currently being stowed), we define the set of transports in our model as TR = {<
pi , pj > |pi = p0 , pj ∈ P }.
The decision variables xtτ +
l ∈ R represent the number of containers of type τ ∈ T
stowed in location l ∈ L during transport t. In this formulation we have removed the
integrality constraint on the decision variables. Given the number of containers that
can be stowed in a location (up to a couple of hundred), and the small number of valid
types per location seen in practice, the loss of precision in our solution by doing this
is low and can be easily dealt with in the second phase of our heuristic. We test this
assumption in the experimental section.

85
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

We remind the reader that the sets and constants used in the model presented in
this chapter have been previously defined in the section Sets and Constants, at the
beginning of this document. Below we introduce the variables of our model.
Decision Variables
xtτ
l ∈R
+
Containers of type τ in location l during transport t.

Auxiliary Variables
ypHC
0l ∈ R
+
Number of high-cube containers stowed in l above the high-cube killing limit at port p0 .
O
yp0 ∈ R +
Number of misused odd slots at port p0 .
vp¬RO
0l ∈ R+ 40 ’ volume capacity gained by stowing 20 ’ non-reefer containers in the non-reefer
odd slots of location l at port p0 .
OR +
vp0 l ∈ R 40 ’ volume capacity gained by stowing 20 ’ containers in the reefer odd slots of
location l at port p0 .
O
yp0 l ∈ R +
Number of odd slots not used in l at port p0 .
O +
yp0 ∈ R Number of odd slots not used in the vessel due to misplacement of 20 ’ containers
in port p0 .

We propose the LP model below for solving the problem of assigning container types
to locations in the first phase of our heuristic. First we introduce the objective function
of our model.
maximize
X XX X
Clτ xtτ
l −C
HC
ypHC
0l − C
OD O
yp0 (5.1)
t∈TR ON τ ∈T l∈L l∈L
p0

The objective (5.1) maximizes the quality of the stowage plan. Each decision variable
xtτ τ
l has a cost Cl associate to them that reflects preferences defined by our industrial
collaborator with respect to container types and locations. For instance, reefer contain-
ers are more profitable, therefore, we prefer to load reefer over non-reefer containers.
Locations below deck are preferred over locations on deck due to accessibility, and heavy
containers are preferred to be stowed below deck to increase stability of the vessel.
Recall that high-cube containers utilize more volume capacity than dry cargo contain-
ers. When the number of high-cube containers stowed in location l goes beyond KpK0 l ,
the slot capacity of l is reduced in order to provide the extra volume capacity necessary
to stow the additional high-cube containers. The cells involved in this slot capacity
reduction are known as killed cells, i.e., cells that have been partially utilized by high-
cube containers and where it will not be possible to stow any additional container. We
penalize the number of high-cube containers stowed in location l above KpK0 l with cost
C HC . The number of misused odd slots in the vessel is penalized by cost C OD . Odd
slots are at the bottom cells of a stack and it is necessary to fill them up with 20 ’
containers in order to access the full capacity of the stack. Our model penalizes odd
slots that are not being used when there are 20 ’ containers available in the loadlist to
do so. Next, we introduce the constraints representing the planning scenario defined
by the stowage coordinator.
s.t.

86
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

X X
xtτ
l = 0 ∀l ∈ L (5.2)
t∈TR lp0 τ ∈T
X
xtτ τ
l ≤ LDt ∀t ∈ TR, τ ∈ T (5.3)
l∈L
X XX
Wτ xtτ U
l ≤ Wp0 v ∀v ∈ V (5.4)
t∈TR ON l∈Lv τ ∈T
p0

Note that all capacities in the model are reduced by the release. Constraint (5.2)
restricts the containers to stow in location l by removing containers with a different
discharge port to that chosen by the stowage coordinator in the planning scenario.
Constraint (5.3) ensures that we do not stow more containers of type τ on transport
t than those available from the loadlist. Constraint (5.4) restricts the weight of a bay
according to the stowage coordinator’s preferences specified in the planning scenario.
Below we present the constraints necessary to model odd slots and their impact on the
volume capacity of the locations.
X X
A¬R tτ ¬RO
p0 l xl ≥ vp0 l ∀l ∈ LO (5.5)
t∈TR ON τ ∈T 20 ∩T NR
p0

vl¬RO ≤ A¬R ¬RO

p0 l Kp0 l ∀l ∈ LO (5.6)
X X
AR tτ OR
p0 l xl ≥ vp0 l ∀l ∈ LO (5.7)
t∈TR ON τ ∈T 20
p0

vlOR ≤ AR OR
p 0 l Kp 0 l ∀l ∈ LO (5.8)
X X
tτ
vp¬OR
0l vpOR
0l
xl ≥ ¬R + R ∀l ∈ LO (5.9)
ON 20
Ap0 l Ap0 l
t∈TR p0 τ ∈T

When present, odd slots are mainly found at the bottom cells of a stack. For the 40 ’
volume capacity of a stack with odd slots to become available, the odd slots must be
filled in with 20 ’ containers. Thus, in locations with odd slots 40 ’ volume capacity
is increased proportionally to the utilized odd slots. Notice that we must distinguish
between 20 ’ reefer and non-reefer containers since a reefer container can only be placed
¬RO
in reefer odd slots. Constraint (5.5) sets the lower bound of the auxiliary variable vpl
to be the extra 40 ’ volume capacity gained by stowing 20 ’ non-reefer containers in non-
reefer odd slots of location l. An upper bound to this gain is provided by constraint
(5.6). In a similar fashion, constraints (5.7) and (5.8) provide lower and upper bounds,
OR
respectively, for the auxiliary variable vpl . The lower bound is calculated as the extra
40 ’ volume capacity gained by stowing 20 ’ containers (reefer and non-reefer), in the
reefer slots of l, while the upper bound is the maximum 40 ’ volume gain when stowing
20 ’ containers in all reefer odd slots. To avoid counting volume gains produced by the
20 ’ non-reefer containers twice, constraint (5.9), forces the number of 20 ’ containers to
be at least as many as the containers used to calculate the volume gains when stowing

87
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

20 ’ in odd slots. Next, we introduce the constraints that limit volume, weight, and
reefer capacity of the locations.
X X
ΠTEU
τ xtτ TEU
l ≤ Kp 0 l ∀l ∈ L (5.10)
t∈TR ON τ ∈T
p0
X X
xtτ 20
l ≤ Kp 0 l ∀l ∈ L (5.11)
t∈TR ON τ ∈T 20
p0
X X
xtτ RF
l ≤ Kp 0 l ∀l ∈ L (5.12)
t∈TR ON
p0
τ ∈T R

X X
ΠFEU
τ xtτ RC
l ≤ Kp0 l ∀l ∈ L (5.13)
t∈TR ON
p0
τ ∈T R
X X
Wτ xtτ 20
l ≤ Wp0 l ∀l ∈ L (5.14)
20
t∈TR ON
p0
τ ∈T
X X 40
ΠFEU
τ Wτ xtτ
l ≤ Wp0 l ∀l ∈ L (5.15)
t∈TR ON τ ∈T
p0
X X
Φτ xtτ V
l ≤ Kp 0 l ∀l ∈ L (5.16)
t∈TR ON τ ∈T
p0
X X 40 N
Φτ xtτ
l ≤ Kl + vl¬RO + vlOR ∀l ∈ LO (5.17)
t∈TR ON τ ∈T 40
p0
X X 40 V
Φτ xtτ
l ≤ Kp 0 l ∀l ∈ L\LO (5.18)
40
t∈TR ON
p0
τ ∈T
0
X X
xtτ τ
l ≤ Kp 0 l ∀l ∈ L, τ ∈ T (5.19)
t∈TR ON (ΠTEU =ΠTEU )
p0 τ 0 ∈T (Wτ 0 ≤Wτ ) ∩T τ0 τ

Constraints (5.10)-(5.13) and (5.14)-(5.15) are equivalent to constraints (4.6)-(4.9) and

(4.11)-(4.12) from section 4.3.1, respectively, and will not be discussed any further.
Containers consume volume capacity from locations depending on their type. For in-
stance, 20 ’ containers need a single slot, while 40 ’ containers need two. All high-cube
containers are 40 ’ and they consume slightly more than two slots. Constraint (5.16),
restricts the volume of containers stowed in l ∈ L to be within its volume capacity limit
at port p.
We constrain the 40 ’ volume capacity of locations without odd slots by means of
constraint (5.18). Note that this constraint considers volume consumption instead of
TEU consumption since 40 ’ can be dry or high-cube containers. Constraint (5.17),
limits the number of 40 ’ containers in odd locations to be within the 40 ’ volume
capacity. The 40 ’ volume capacity of location l ∈ LO is defined as the volume capacity
of the location stacks with no odd slots plus the increment in 40 ’ volume capacity
given by stowing 20 ’ containers in the odd slots of l. We compute the upper bound

88
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

Kplτ in a pre-processing stage. This upper bound is defined as the maximum number
of containers of type τ that can be stowed in l when relaxing all constraints but those
enforcing weight and TEU capacity. Thus, the number of containers of types with the
same length as τ (i.e., demanding the same TEU capacity), and with greater or equal
weight must be smaller or equal to Kplτ . Constraint (5.19) enforces this. Finally, the
constraints that compute lower bounds for the misused odd slots in the vessel and the
high-cube containers above the high-cube killing limits.
X
xtτ O O
l + yp0 l ≥ Kp0 l ∀l ∈ LO (5.20)
τ ∈T 20
X X
( ypO0 l − KtO20 ) ≤ ypO0 (5.21)
t∈TR l∈Lt
X
xτl t − ypHC K
0 l ≤ Kp0 l ∀l ∈ L (5.22)
τ ∈T HC

Constraint (5.20) provides a lower bound for the number of odd slots not used in
locations with odd slots available. In this constraint we assume that 20 ’ containers will
always be stowed in odd slots if they are available. A lower bound for the number of
odd slots not used in the vessel due to misplacement of 20 ’ containers is provided by
constraint (5.21). This lower bound is the difference between the odd slots not currently
being used, and the odd slots that are known in advance but cannot be used due to the
lack of 20 ’ containers. As mentioned before, when the number of high-cube containers
in a location l surpasses the high-cube killing limit it is necessary to reduce the slot
capacity of the location in order to stow the extra high-cube containers. The relation
between the number of killed cells and the extra number of high-cube containers that
can be allocated in l is not linear, i.e., several high-cube containers can be stowed in l by
killing a single cell. Any linearization attempt on measuring the slot capacity reduction
due to killed cells in l, can be reduced to measure the number of high-cube containers
above KpK0 l stowed in l. Constraint (5.22) provide a lower bound for auxiliary variable
ypHC
0l .

5.5.2 Greedy Algorithm for Slot Planning

We introduce the greedy Algorithm 1 to stow containers in specific slots (i.e., slot
planning) in such a way that the set of stacking constraints and rules of thumb described
in Section 5.2 are satisfied and optimized, respectively. The algorithm works as follows.
Given a set of containers C and a set of stacks S as parameters, for each container c ∈ C
the algorithm finds the best stack s ∈ S where c can be stowed, (lines 5-9). Containers
are stowed bottom-up in stacks, therefore, when looking for a feasible stack for c, the
greedy algorithm only checks the lowest empty cell of each stack. A stack si is preferred
over stack sj for stowing container c if by stowing container c at the lowest empty cell
of stack si more of the rules of thumb described in Section 5.2 are satisfied than by
stowing c at the lowest empty cell of stack sj . Once the best stack has been found, c is

89
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

Algorithm 1 GreedySelection(C, S)
1: uc ← ∅ // Set of unstowed containers from C
2: repeat
3: Select container c from C
4: bs ← NULL // Best stack for container c
5: for si ∈ S do
6: if it is possible to stow c in si then
7: update bs to si if si is better than bs
8: end if
9: end for
10: if there is no feasible stack for c then
11: uc ← uc ∪ {c}
12: else
13: Stow container c in stack bs
14: end if
15: C ← C − {c}
16: until C 6= ∅
17: return uc

stowed in the lowest empty cell of the stack, (line 13). In case it is not possible to find
a feasible stack to stow c, the container is added to the list of unstowed containers, uc,
(line 11). Once all containers have been stowed, or attempted to be stowed, the list of
unstowed containers is returned.
It is important to notice that if the stowing of container c in the lowest empty cell
of stack s fulfills all stacking rules and constraints from Section 5.2 , such stowage
will not make infeasible any of the previous stows of containers in cells of stack s.
Containers are selected from C (line 3), following a specific order to avoid breaking
stacking constraints. To avoid having 20 ’ containers on top of 40 ’ ones, 20 ’ containers
are selected first. Among containers with the same length, reefer containers are selected
before non-reefer since reefer plugs are usually placed at the bottom of the stacks.
Finally, when containers have the same length and reefer capabilities heavier containers
are selected first. Stowing containers bottom-up in stacks eliminates the possibility of
having containers with no support below.
We use the greedy algorithm introduced above to slot plan the vessel following the
assignment made by the corresponding master plan. First, we assign the containers
from the loadlist I to the locations following the solution from the first phase. Let qlτ
be the number of containers of type τ assigned to location l during the first phase of
the heuristic. For each location l ∈ L and type τ ∈ T we select on a first come, first
served basis from the loadlist qlτ containers of type τ to stow on location l. Since qlτ
could be a continuous value, unit containers are stowed in the locations with the largest
fractional part. Once containers are assigned to locations, we stow independently each
l ∈ L with the greedy algorithm.

90
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

Finally, we use a second run of the greedy algorithm to stow some of the rolled out
containers. Let ucl be the set of containers assigned to location l that could not be
stowed by the greedy algorithm, U = {∪l∈L ucl } be the set of all containers the greedy
algorithm could not stow at each location, and S V be the set of stacks of the vessel.
We change to a vessel perspective by considering all the vessel stacks at once, and use
the greedy algorithm to stow the set of containers U in stacks from S V . The containers
that are not stowed at this point are reported as unstowed containers to the stowage
coordinator.

5.6 Experiments
In order to evaluate the heuristic introduced in this chapter we use three stowage plans
made by stowage coordinators on a vessel with approximately 15.000 TEU capacity.
Each stowage plan is a stowage condition of the vessel after visiting the last of six
consecutive ports. For our experiments, we extract the planning scenarios at each port
used to achieve these stowage conditions and assume that the vessel arrives empty to
the first port. The planning scenarios will not include all preferences, as we do not
have information about containers sent internally between the ports. However, stowage
coordinators have actually turned these planning scenarios into stowage plans, making
them suitable for comparing the performance of our heuristic against the stowage co-
ordinators’ work. In the few locations where containers with different discharge ports
are mixed, something our heuristic does not allow, we assign the discharge port of the
majority of containers to the location.
As depicted in Figure 5.2, real-life stowage plans are generated independently for
each port in a schedule. Given that our goal is to evaluate the performance of the
heuristic introduced in Section 5.5, we generate a data set of 18 instances, i.e., one for
each of the six ports in the three real-life stowage plans, and analyse them independently.
Table 5.2 specifies the features of the three real-life stowage plans. All the experiments
were run on an Ubuntu 10.4 system, with Intel Core 2 Duo, 4 GB of RAM, and CPLEX
12.1.
Table 5.3 presents the results of our experiment. We measure three key factors in
our experiments: First, the response time of our heuristic, since one of our goals is
to keep the generation of stowage plans within two seconds on average. Second, the
number of containers that could not be stowed. Third, the number of cells killed in a
stowage plan, which is one of the most common reasons for sub-utilizing the volume
capacity of a container vessel.
Our heuristic managed to generate stowage plans for each instance in 1.27 seconds
on average, a number well within our desired time for the interactive generation of
stowage plans. The longest time was 3.05 seconds (instance 12), while the fastest was
0.17 seconds, (instance 7). There are two important things to notice about the run times
presented in Table 5.3. First, it is easy to see that the heuristic spends most of the time
solving the LP model. Second, we can observe that there is a direct relation between the

91
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

SWP-SC ID P O20 ’ O40 ’ L20 ’ L40 ’ U

1 1 0 0 182 939 2026
2 2 182 939 219 750 3779
3 3 401 1689 563 1439 7220
1 4 4 964 3128 175 412 8219
5 5 1192 3540 228 685 9817
6 6 1420 4225 625 747 11936
7 1 0 0 101 531 1163
8 2 101 531 141 731 2766
9 3 242 1262 403 1416 6001
2 10 4 645 2678 407 614 7636
11 5 1052 3292 199 827 9489
12 6 1251 4119 827 1141 12598
13 1 0 0 102 864 1830
14 2 102 864 242 684 3440
15 3 344 1548 473 1276 6465
3 16 4 817 2824 116 304 7189
17 5 933 3128 144 613 8529
18 6 1077 3741 652 894 10969

Table 5.2: Instance Overview. Column SWP-SC shows the identifier of the stowage plan,
while ID represents the identifier of the instance, and P the port number in the schedule.
O20 ’ and O40 ’ are the number of 20 ’ and 40 ’ release containers on the vessel when it calls
the port, respectively. L20 ’ and L40 ’ represent the number of 20 ’ and 40 ’ containers to
load when the vessel calls the port, and U the TEU utilization of the vessel when leaving
the port.

number of containers to stow and the time used by the heuristic to generate the stowage
plans. For each real-life stowage plan, the ports with the greatest number of containers
to stow, i.e., ports 3 and 6, are also the ports where the stowage planning takes the
longest. It is important to notice as well that the number of release containers does
not seem to have a strong influence on runtime, since otherwise, we would see runtimes
increasing from port 1 to 6 in all stowage plans.
The maximum percentage of rolled out containers is 1.89% of the containers in the
loadlist (instance 6), and 0.58% on average, a considerably low fraction. A difference
between the number of 20 ’ and 40 ’ containers rolled out can be easily spotted. In 4
out of 18 instances 40 ’ containers were rolled out, and with respect to the total rolled
out containers for all instances, only 16 out of 129 are 40 ’ long. There are two main
reasons for this difference. The first reason is that since there are no odd slots in our
test vessel, the number of 20 ’ containers assigned to a location by the LP must be even.
Given that cell capacity must be completely utilized, no cell with an odd number of 20 ’
containers assigned to it is allowed. Thus, our greedy algorithm will roll out at least one
20 ’ container when an odd number of 20 ’ containers has been assigned to a location
by the LP. Moreover, when we use our greedy heuristic for the second time (i.e., vessel
perspective), to attempt to stow all the rolled out containers, most of the stacks are
already full or partially utilized with 40 ’ containers on top, making it impossible to
stow 20 ’ containers. The second reason relates to the fact that reefer plugs are usually
available only at bottom tiers of the vessel locations. In the set of real-life stowage

92
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

SWP-SC ID P C LP(s) G(s) D 20 ’ D 40 ’ KS KH

1 1 1121 0.30 0.06 2 0 8 8
2 2 969 0.74 0.05 3 1 9 10
3 3 2002 1.86 0.09 1 0 6 5
1 4 4 587 1.01 0.03 1 0 8 5
5 5 913 1.34 0.04 4 12 12 14
6 6 1372 2.21 0.06 26 0 7 15
7 1 632 0.14 0.03 1 0 9 9
8 2 872 0.57 0.04 3 2 18 14
9 3 1819 1.47 0.08 3 0 9 16
2 10 4 1021 1.23 0.05 3 0 1 6
11 5 1026 1.23 0.04 3 1 9 14
12 6 1968 2.97 0.08 29 0 7 9
13 1 966 0.23 0.05 0 0 19 23
14 2 926 0.75 0.06 4 0 11 17
15 3 1749 1.66 0.10 3 0 6 16
3 16 4 420 0.58 0.03 2 0 8 3
17 5 757 1.08 0.04 5 0 13 13
18 6 1546 2.57 0.07 18 0 40 20

Table 5.3: Experimental Results. Column SWP-SC presents the identifier of the stowage plan,
ID the identifier of the instances, P the port number on the schedule, and C the number
of containers to stow. LP(s) and G(s) are the time in seconds used by the LP model and
the greedy heuristic, while D20 ’ and D40 ’ present the number of 20 ’ and 40 ’ containers
dropped by the heuristic. Finally, KS and KH are the number of cells killed in the plans
generated by the stowage coordinators and the placement, respectively.

plans, a considerable number of reefer containers, most of them 20 ’ long, are loaded in
port number six. Since more that 50% of the capacity of the vessel is already utilized
at this point, we have already used most of the reefer slots at the bottom tiers of the
vessel, even by stowing non-reefer containers. Moreover, most of the few remaining free
reefer slots are in stacks partially filled with 40 ’ containers.
The number of killed cells in our set of instances average 11.1 per instance for the
stowage plans made by stowage coordinators, and 12.1 for those made by our heuristic.
In 10 out of 18 instances, stowage coordinators managed to kill less cells than our
heuristic, while our heuristic killed less cells in 5 instances. Both, stowage coordinators
and our heuristic, killed the same number of cells in the remaining three instances.
Stowage coordinators killed 20 cells more than our heuristic in the worst case (instance
18), while the heuristic killed 10 more cells than the stowage coordinators, (instance
16). Even though the results of this experiment favours the stowage coordinators, it is
important to notice that stowage coordinators have an accumulated expert knowledge
of years generating stowage plans, and we are getting close to matching that knowledge
(difference on the averaged killed cells is only one), with a heuristic that stows 1,968
containers in a vessel partially full (63%), in three seconds.
It is important to notice when comparing the number of killed cells by the stowage
coordinators and our heuristic that we have two simplifications in the generation of
stowage plans compared to stowage coordinators: we treat 45 ’ containers as 40 ’ and we
do not consider lashing constraints in on deck locations. Even with these simplifications,

93
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

however, we still think that it is relevant to show how our heuristic performs, compared
to stowage coordinators in terms of the number of cells killed, since the impact of
considering 45 ’ containers and lashing constraints will be minimal. Cells are only
killed in locations below deck due to the physical limitations imposed by hatch covers.
Let us consider what will happen if we allow our heuristic to handle 45 ’ containers.
First, all 45 ’ containers will be stowed on deck since on deck locations are the only ones
with 45 ’ capacity. This could affect the number of killed cells compared to the ones
from the current stowage plans if a 45 ’ dry cargo container moved on deck is replaced
by a 40 ’ high-cube container. This is, however, very unlikely since most, if not all, 45 ’
containers are high-cube containers. Any container our heuristic uses to replace the 45 ’
containers going below deck will be of the same height or smaller and will not increase
the number of killed cells.
Roughly speaking, lashing constraints can be seen as constraints that restrict the
weight distribution of containers stowed in on deck locations. When the GM increases,
these constraints force the number of light containers stowed on deck to be increased.
Thus, in cases where all heavy containers are high-cube, we are forced to stow them
below deck and an increment in the number of killed cells would be expected. This
increment, however, will not be dramatic. In the cost values of the variable in the ob-
jective function of our LP model we reward stowing heavy containers in locations below
deck, thus, we are generating stowage plans that already stow as many light containers
on deck as possible.

To measure the impact of generating master plans with an LP model, thus allowing
the assignment of fractional containers into locations, we create a new model for master
planning. To do this we add the integrality constraint over the variables representing
the number of containers to be stowed in all locations (xτl t ). We then proceed to
generate stowage plans using the new IP model. We generate master plans with 1%
and 5% of optimality gap, and evaluate the same criteria as before, i.e., number of rolled
out containers, CPU time, and killed cells. Table 5.4 presents the results of using the
new IP model for the generation of stowage plans. Since we want to compare the new
plans against the ones generated using the LP model, the table presents the difference
between the values associated to the stowage plans based on LP master plans and those
based on the IP master plans. For instance, in the case of the number of 20 ’ containers
rolled out, a positive value indicates that the stowage plan based on the LP model rolls
out more containers than the plan based on the IP model.
From the results in table 5.4 we can see that there is no clear advantage of using
an IP model to generate master plans. The improvement on containers rolled out only
happens in a few instances and it is negligible, (less than 1%). There are even some
cases where the LP based stowage plans roll out less containers, though the number is
negligible as well. With respect to the number of killed cells, the IP model actually
behaves slightly worse than the LP. On average, stowage plans based on the IP model
with a 1% gap killed 0.67 more cells, while the plans based the on IP model with a 5%
gap killed 0.83 more cells.

94
 Chapter 5. Heuristic Optimization for Interactive Stowage Planning

IP(s) D 20 ’ D 40 ’ KC
SWP-SC ID P
1% 5% 1% 5% 1% 5% 1% 5%
1 1 0.24 0.21 0 0 0 0 1 1
2 2 1.71 1.56 0 0 0 0 2 2
3 3 7.68 3.26 0 0 0 0 3 3
1 4 4 138.75 5.67 0 0 0 0 -2 -1
5 5 6.69 1.40 2 2 0 0 -1 1
6 6 21.67 3.18 0 6 0 0 -3 -7
7 1 4.51 4.26 0 0 0 0 -1 -1
8 2 2.67 2.46 0 0 0 2 -2 -2
9 3 78.28 49.12 0 0 0 0 2 0
2 10 4 36.16 0.56 0 0 0 0 3 3
11 5 15.04 1.52 0 0 0 1 -4 -2
12 6 7.77 6.23 6 2 0 0 -4 -7
13 1 3.80 3.79 0 0 0 0 -1 -1
14 2 3.15 3.19 0 0 0 0 0 0
15 3 48.44 42.79 0 0 0 0 -1 -1
3 16 4 11.32 10.55 0 0 0 0 1 1
17 5 1.90 1.82 0 0 0 -4 -3 -3
18 6 11.72 11.15 0 0 0 -4 -2 -2

Table 5.4: Comparing with Integer Results. The column SWP-SC represents the identifier
of the real-life stowage plan, ID the identifier of the instances, P the port number on the
schedule, and the third and fourth columns show the time for the IP model with 1% and 5%
gaps, respectively. The next six columns are grouped in pairs and represent the difference
of a specific criterion between the results from the LP master plans and the IP master
plans with 1% and 5% optimality gap. Columns six and seven present and columns eight
and nine show the difference in 20 ’ and 40 ’ containers rolled out, respectively. Columns
ten and eleven present the difference in number of cells killed.

For six of the instances stowage plans generated based on the LP model killed more
cells that those generated based on both IP models. On the other hand, stowage plans
based on the LP model killed less cells in eleven of the instances compared to the plans
based on the IP model with 1% gap, and in nine of the instances for the IP model with
a 5% gap. One of the reasons for this behaviour is related to the objective minimizing
the number of killed cells. This objective is based on the premise that in order to
avoid the killing of cells the number of HC containers in a location must stay below a
precomputed safety limit. Once the number of HC containers in a location goes over
this limit, the extra HC are penalized linearly. The relation between HC containers
and killed cells, however, is not linear. The slot planning heuristic follows the master
plan as closely as possible, thus any inaccuracy in the master plan will impact the slot
plan. In the case of the HC containers in IP master plans, the slot planning heuristic
attempts to stow the exact number of HC in the locations, even if that number of HC
end up killing more cells than anticipated. In the case of LP master plans, the fractional
number of HC containers gives the slot planner the flexibility to decide where better to
place some of the HC that kill cells.

95
Chapter 5. Heuristic Optimization for Interactive Stowage Planning

5.7 Conclusion
This chapter introduced a heuristic that serves as the optimization component of a
decision support system to interactively generate container stowage plans. Our heuristic
uses the know-how of stowage coordinators represented in a planning scenario to stow
containers from a loadlist into vessel slots. Our experiments showed that with 1.27s of
runtime on average, our heuristic can be used for interactive optimization. Additionally,
the stowage plans produced are competitive with expert users. For a container vessel
of approximately 15,000 TEUs, our heuristic managed to only roll out, at most 1.89%
of the loadlist and 0.58% on average, and only killed 1 cell more on average than the
expert user. As future work we plan to make our stowage plans more accurate by
handling 45 ’ containers and including lashing constraints. We are also interested in
studying the impact of reducing the granularity of the weight classes in the LP model’s
runtime and in the percentage of containers dropped by the greedy heuristic.

96
 Chapter 6. Cargo Composition Problem

Chapter 6

Cargo Composition Problem

In this chapter, we introduce the Cargo Composition Problem (CCP), the problem of
evaluating how the stowage characteristics of containers with different features affect
performance measures used in the liner shipping domain. To solve the CCP, we intro-
duce the first model with variable displacement that assigns containers to vessel slots
and is actually solvable on real instances. Additionally, to increase scalability we have
extended the master planning model presented in chapter 4 with the capability of deal-
ing with variable displacement and use it to solve the CCP. This chapter is based on
Delgado and Jensen (2013).

6.1 Introduction
Several planning decisions in liner shipping companies depend on the stowage capabili-
ties of container vessels. From strategic decisions involving long term investments (e.g.,
designing/buying container vessels to fulfill specific demands), to medium term tactical
decisions (e.g., vessel deployment selecting the vessels that will cover demands), and
daily operational decisions (e.g., selecting containers to transport to increase revenue).
What they all have in common is that they often are made under the assumption that
a container vessel can be filled up with all kinds of containers, and that these containers
can be freely stowed in the vessel. This assumption, however, is far from reality. Thus,
having access to an accurate footprint of the stowage capabilities of a container vessel
is valuable for liner shipping companies.
In this chapter, we introduce the Cargo Composition Problem (CCP), the problem
of evaluating how the stowage characteristics of containers with different features affect
important performance measures in liner shipping companies, (e.g., vessel intake and
cargo revenue). To solve the CCP, we introduce the first complete model that stows
containers into vessel slots that is able to solve real instances. This model can represent
variable displacement, a feature of the CCP that has not been considered in previous
stowage models. To increase scalability, we present an extension to the master planning
model presented in chapter 4 and use it to generate sub-optimal solutions for the CCP.

97
Chapter 6. Cargo Composition Problem

We evaluated our two models on a benchmark suite with 10 container vessels provided
by our industrial collaborator in single and multi-port setups. In the single port setup,
we optimized vessel intake and cargo revenue, and stressed both models by optimizing
intake of containers with specific features such as weight, length, and height. Our
experiments showed that the extended master planning model finds solutions of similar
quality to those found by the complete model in a fraction of the time. Moreover,
we showed experimentally the advantage of using accurate stowage planning to avoid
overestimating the stowage capabilities of a vessel, when doing revenue analysis. In
the multi-port setup, we optimized cargo revenue on a set of 360 instances, where the
extended master planning model successfully scaled and managed to solve 91.7% of the
instances.
The remainder of the chapter is organized as follows. Section 6.2 defines the CCP
and section 6.3 presents related work. In section 6.4, we introduce the complete model
as an IP that solves the CCP by assigning containers to vessel slots. In section 6.5,
we present an extension of the master planning model, introduced in chapter 4, that
assigns containers to locations, and we use it to generate sub-optimal solutions for the
CCP . Our experiments and results are presented in section 6.6, and conclusions are
drawn in section 6.7.

6.2 Problem Definition

The CCP evaluates how the stowage characteristics of containers with different features
affect important performance measures used in liner shipping companies, (e.g., vessel
intake and cargo revenue). We perform this evaluation by considering only contain-
ers matching the features that we want to evaluate and generating seaworthy stowage
plans using these containers. During the generation of the stowage plans, the contain-
ers to stow are selected according to a gain defined based on their features (i.e, length,
height, weight, and reefer capability), and on the performance measure being consid-
ered. That is, we compute seaworthy stowage plans that optimize the gain generated
by the containers stowed.
When optimizing revenue, additionally to the gain associated with each container,
we must also consider costs related to how containers are stowed in the vessel. Con-
tainer ships in liner shipping companies transport containers between ports on a fixed
cyclic route. Containers must be stowed in a vessel such that time at port and fuel con-
sumption are minimized. Time at port is minimized by reducing overstowage, and by
efficiently distributing crane moves over the length of the vessel such that the makespan
of the quay cranes is reduced. To minimize fuel consumption we reduce the amount of
ballast water loaded in the vessel. Ballast water helps fixing stability problems but it
increases the displacement of the vessel, forcing the engines to burn more fuel to achieve
the desired speed.
All major aspects of stowage must be considered in order for our results to be
valuable for the industry. In the CCP, we consider standard ISO containers 20 ’ and

98
 Chapter 6. Cargo Composition Problem

40 ’ long, with reefer capabilities, and high-cube and normal height. We also consider
stacking rules between 20 ’ and 40 ’ containers, weight and height stack limits, and we
force cells with capacity in both slots to be either empty or with both slots occupied.
Additionally, containers are sorted by weight in on deck stacks, a rule of thumb used in
the industry to reduce lashing issues. A more accurate modeling of lashing can be done
with specialized capacity constraints on stacks and locations. Finally, main stability
and stress force calculations are also considered.
In this study we have not considered the line-of-sight constraint and containers
with special requirements, (e.g., IMO, pallet-wide, etc). Line-of-sight can be modeled
by specialized constraints that reduce on deck stacks capacity according to how the fore
draft changes. We do not expect the scalability of our models to be affected by the
inclusion of these constraints. With respect to containers with special requirements,
they are always placed in specific areas of the vessel, and abstractions of their stowing
rules can be modeled with specialized capacity constraints without affecting the overall
results produced by the model. Moreover, these containers are underrepresented in the
container fleet and a precise modeling of their stowing rules might considerably impact
the scalability of the models unnecessarily.

6.3 Literature Review

Previous work on optimization of cargo compositions has been limited to a few pub-
lications on revenue management. In Ting and Tzeng (2004), a revenue management
model that includes the repositioning of empty containers is introduced. Their stowage
model considers volume, 40 ’, reefer, and weight capacity for the complete vessel. It
disregards, however, the vessel’s physical layout and any kind of stability constraints.
Likewise, Feng and Chang (2008), introduces a revenue model for short-sea shipping,
where total volume and weight capacity of the vessel are the only considerations with
respect to stowage of the containers.
Given the similarities between cargo composition optimization and stowage plan-
ning, we consider relevant work from this more active field of research that relates
to the contributions presented in this chapter. Models that generate stowage plans
by assigning containers to vessel slots have mainly been formulated as mathematical
programs. One of the most accurate formulations is introduced in Botter and Brinati
(1992). This IP model includes several stability and stress force constraints and pro-
poses an approach to perform stability calculations when the displacement of the vessel
is variable. The model, however, does not solve in practice. Two more IP formulations
are introduced in Ambrosino et al. (2004) and Li et al. (2008), where containers are
assigned to slots such that heuristic rules modeling stability constraints are satisfied.
Both models consider only standard 20 ’ and 40 ’ containers. Ambrosino et al. report
that their the IP model do no scale well in practice, while Li et al. test their IP model
with random instances on a 800 TEUs vessel, but no evaluation on the scalability of
the model is presented.

99
Chapter 6. Cargo Composition Problem

We refer the reader to sections 4.2 and 5.4 for a review of the academic work on
other approaches for stowage planning.

6.4 Complete Model for the CCP

In this section we formulate the CCP as a boolean IP model. This model assigns
container types to specific vessel slots satisfying constraints and optimizing objectives
described in section (6.2). Recall that a container type τ is a 4-tuple (l, h, r, w), where
l ∈ L is the length of the container in feet, h ∈ H is the height of the container, r ∈ R
is the reefer properties of the container (reefer or non-reefer), and w ∈ W is the weight
class of the container, expressed as the average weight of the containers in the class in
metric tons. We assume that all 20 ’ container types are of the same height, since 20 ’
high-cube containers are rare. Below we define decision and auxiliary variables of the IP
model. We remind the reader that the sets and constants used in the models presented
in this chapter have been previously defined in the section Sets and Constants, at the
beginning of this document.
Decision variables
xtτ
ce ∈ {0, 1} Whether a container of type τ is stowed in slot e of cell c during transport t.
xpu ∈ R+ Amount of ballast water in tons loaded at port p in tank u.

Auxiliary variables
{OH,O}
vpl Hatch (OH), stack (O) overstowing containers at port p in location l.
vpC Crane makespan of port p.
vpW ∈ R+ Vessel displacement.
W
vpl ∈ R+ Weight of containers stowed at port p in location l.
V ML
vp ∈ R+ Vertical moment of all locations at port p.
vpL ∈ R Longitudinal center of gravity at displacement interval i and port p.
vpV M ∈ R+ Vertical moment at port p.
vpV ∈ R+ Vertical center of gravity at port p.
W
vpi ∈ R+ Displacement at port p in displacement interval i.
vpM ∈ R+ Metacenter at port p.
B +
vps ∈ R Buoyancy force at port p of section between stations s and s + 1.
Sα
vpf ∈ R+ Shear force at port p fore or aft of frame f .
Sα
vpf ∈ R+ Bending moment at port p fore or aft of frame f .
φpc ∈ {0, 1} Whether there are containers below cell c going to be discharged at port p.
δpl ∈ {0, 1} Whether there are containers to be discharged at port p in locations below location l.
ψpi ∈ {0, 1} Indicator variable for displacement interval i ∈ I at port p.

The objective function of the IP formulation of the cargo composition problem is

X XX X
M ax : Ctτ xtτ
ce
t∈T R c∈C τ ∈T e∈E τ
X X X X
− (CpO ( OH
vpl + O
vpc ) + CpC vpC + CpU xpu ) (6.1)
p∈P l∈L c∈C u∈U

100
 Chapter 6. Cargo Composition Problem

This objective function maximizes a weighted sum of the number of stowed containers
based on the individual gain of their type. This gain depends on the kind of analysis
being done, (e.g, revenue or intake optimization). The negative impact of hatch and
stack overstowage that the stowed containers produce, the crane makespan, and the
ballast water necessary to make the vessel seaworthy is also considered. The costs of
these latter aspects depend on the scenario the solver is optimizing and will be explained
in the experimental section. Next we introduce the stacking constraints that ensure that
container stacks in the vessel satisfy stacking rules and capacity limits.
X X
xtτ
cA ≤ 1 ∀p ∈ P, c ∈ C (6.2)
t∈T RpON τ ∈T
X X
xtτ
cF ≤ 1 ∀p ∈ P, c ∈ C (6.3)
t∈T RpON τ ∈T 20

X X
( xtτ tτ
cA − xcF ) = 0 ∀p ∈ P, c ∈ C (6.4)
t∈T RpON τ ∈T 20
X X X X 40
( ΠFEU
τ Wτ xtτ
ce ) ≤ Wlj ∀p ∈ P, l ∈ L, j ∈ Jl (6.5)
t∈T RpON c∈Clj τ ∈T e∈E τ
X X X
Wτ xtτ 20
ce ≤ Wlje ∀p ∈ P, l ∈ L, j ∈ Jl , e ∈ E (6.6)
t∈T RpON c∈Clj τ ∈T 20
X XX
Hτ xtτ
cA ≤ Hlj ∀p ∈ P, l ∈ L, j ∈ Jl (6.7)
t∈T RpON c∈Clj τ ∈T
X X X X
xtτ
ce = 0 (6.8)
t∈T R c∈C ¬R τ ∈T R e∈E τ
X X X
xtτ
cα = 0 ∀α ∈ E (6.9)
¬R τ ∈T R ∩T 20
t∈T R c∈Cα
X XX
(xtτ tτ
ce − xc+ e ) ≥ 0 ∀p ∈ P, l ∈ L, j ∈ Jl , c ∈ Clj (6.10)
t∈T RpON τ ∈T e∈E τ
X XX
(Wτ xtτ tτ
ce − Wτ xc+ e ) ≥ 0 ∀p ∈ P, l ∈ LON , j ∈ Jl , c ∈ Clj (6.11)
t∈T RpON τ ∈T e∈E τ
X X X X
xtτ
ce = 0 ∀α ∈ L (6.12)
t∈T R c∈C ¬α τ ∈T α e∈E τ

Constraints (6.2) and (6.3), limit the number of containers to be stowed on the aft
and fore slot of all cells at each port p. The former enforces stowing of, at most, a
single 20 ’ or 40 ’ container in the aft slot. The later forces stowing of, at most, a single
20 ’ container in the fore slot, since the stowage of a 40 ’ container is modeled by only
occupying the aft slot. Constraint (6.4) ensures that either both slots in a cell hold
a 20 ’ container or none does. Stack weight capacity at each port p for 40 ’ and 20 ’
containers is enforced by constraints (6.5) and (6.6), where the former uses the 40 ’

101
Chapter 6. Cargo Composition Problem

weight capacity limit to constrain the combined weight of containers in the stack, while
the latter constrains the 20 ’ capacity in the aft and fore slots of stack i. Containers
stowed in a cell either occupy both slots (40 ’ containers), or have the same height
(20 ’ high-cube containers are not considered in our representative problem since they
are rare). Thus, considering that in our model it is not possible to stow a single 20 ’
container in a cell, the aft and fore height will never be out of synchronization. Stack
height limits are enforced by constraint (6.7), where only the height of the containers
stowed in the aft slots is constrained. Constraint (6.8) removes reefer types from cells
with no reefer slots, while constraint (6.9) complements the previous constraint by
removing 20 ’ reefer types from non-reefer slots in cells with a single reefer slot. We do
not remove 40 ’ reefer containers from these slots since they only need a single reefer slot
to provide them with electricity. Stacking patterns are enforced by constraints (6.10)
and (6.11). The former constraint guarantees that containers have support below,
either by stowing them at the bottom of the stack or atop other container(s), and
that 20 ’ containers are not stowed on top of 40 ’ containers. The latter ensures that
containers stowed in stacks on deck are sorted by their weight bottom-up. Slots can
be restricted to stow containers of specific lengths, thus, constraint (6.12) removes 20 ’
and 40 ’ containers from slots where they cannot be stowed.
Below we define two sets of auxiliary variables used to ease the formulation of
stability and stress force requirements.
X XX X
W
vpl = Wτ xtτ
ce ∀p ∈ P, ∀l ∈ L (6.13)
t∈T RpON c∈Cl τ ∈T e∈E τ
X X XX
vpV M L = DcV Wτ xtτ
ce ∀p ∈ P (6.14)
t∈T RpON c∈C τ ∈T e∈E τ

Auxiliary variables representing the weight of each location are defined by constraint
(6.13), while the auxiliary variable representing the vertical moment of all locations at
port p is defined by constraint (6.14). Next we present constraints to derive the amount
of hatch and stack overstowage and estimate the crane makespan.
X X XX
xtτ S
c0 e ≤ Mc φpc ∀p ∈ P, l ∈ L, j ∈ Jl , c ∈ Clj (6.15)
t∈T RpA c0 ∈Cc− τ ∈T e∈E τ
X XX
( xtτ O
ce ) + 2(1 − φpc ) ≤ vc ∀p ∈ P, c ∈ C (6.16)
t∈T RpOV e∈E τ τ ∈T
X X X XXX
( xtτ H
ce ) ≤ Ml δpl ∀p ∈ P, l ∈ LON (6.17)
l0 ∈LU j∈Jl0 t∈T RA c∈Clj τ ∈T e∈E τ
l
X X X XX
( xtτ U OH
ce ) − Ml (1 − δpl ) ≤ vpl ∀p ∈ P, l ∈ LON (6.18)
j∈Jl t∈T RpOV c∈Clj τ ∈T e∈E τ
X X XX X
( xtτ C
cd ) ≤ vp ∀p ∈ P, n ∈ N (6.19)
t∈T RpA l∈Ln c∈Cl τ ∈T e∈E τ

102
 Chapter 6. Cargo Composition Problem

Stack overstowage is calculated by constraints (6.15) and (6.16). Constraint (6.15) sets
the indicator variable φpc to be active when there are containers below cell c ∈ C to
O
be discharged at port p, while constraint (6.16) defines a lower bound of variable vpc .
When the indicator variable φpc is active, this bound becomes the number of containers
stowed in cell c that are to be discharged in downstream ports, i.e., containers in cell c
that are overstowing one or more containers stowed in cells below.
We consider an alternative formulation inspired by cuts (3.16) and (3.17), introduced
in section 3.4 to improve the linear relaxation of stack overstowage, where constraint
(6.15) is reformulated as the two constraints listed below:

X X
xtτ
c0 A ≤ φpc ∀p ∈ P, l ∈ L, j ∈ Jl , c ∈ Clj , c0 ∈ Cc− (6.20)
t∈T RpA τ ∈T
X X
xtτ
c0 F ≤ φpc ∀p ∈ P, l ∈ L, j ∈ Jl , c ∈ Clj , c0 ∈ Cc− (6.21)
t∈T RpA τ ∈T 20

This new formulation, however, demands a considerably larger number of constraints.

Let us assume n to be the number of cells in a stack i. The current formulation uses n−1
constraints to set the indicator variables needed to calculate the stack overstowage in
i, while the alternative formulation requires (n − 1)(n). Early experimental evaluation
of both formulations showed that the alternative formulation did not have a positive
impact on the performance of the solver. With respect to hatch overstowage, constraint
(6.17) sets the indicator variable δpl to be active when there are containers being loaded
or unloaded at port p, stowed in the locations below location l. Note that the BigM
value MlH must be twice the capacity of locations below l, since these locations could
be emptied and filled up with containers again at the same port p. When the indicator
OH
variable δpl is active, constraint (6.18) defines a lower bound vpl of the number of hatch
overstowing containers in location l at port p. Finally, we calculate a lower bound for
crane utilization by minimizing the makespan of all the cranes that are assigned to
the vessel. This calculation is a lower bound since the makespan of the cranes might
increase due to constraints limiting operability of the cranes, e.g., cranes cannot work
at the same time in adjacent bays. Constraint (6.19) calculates the crane makespan at
each port, and defines a lower bound of the continuous auxiliary variable vpC .

In the CCP, the displacement of the containers considered in stowage plans used to
evaluate a cargo composition is not known in advance. This, combined with the fact that
ballast tanks can be freely filled with water makes the vessel displacement variable. An
approach to tackle the variability in vessel displacement, introduced by the modeling of
ballast tanks, was proposed in section 4.3. In this chapter, we present a generalization of
this approach that allows a linear calculation of the center of gravity with an acceptable
error, and the use of accurate linear approximations of the hydrostatic data when the
variability on vessel displacement depends on cargo and ballast tanks. Let us consider

103
Chapter 6. Cargo Composition Problem

the calculation of the lcg:

LM o + l∈L DlL vl + u∈U DuU vu

P P
P P (6.22)
W o + l∈L vl + u∈U vu
P
This
P expression is non-linear since two of the terms in the denominator, l∈L vl and
u∈U vu , are variable. We aim at turning the lcg calculation into a linear expression
by making the denominator a constant value.
Assume that we can accurately estimate an interval of size W ∆ , that contains the
vessel displacement of the desired cargo composition, and that W is the middle point
of such an interval. We can transform the lcg calculation into the expression:

LM o + l∈L DlL vl + u∈U DuU vu

P P
, (6.23)
W + ∆W
∆ ∆
where ∆W ∈ [− W2 ; W2 ] represents the displacement variation from W . Note that in
∆
this lcg calculation the actual vessel displacement must be within W2 tons of W (i.e.,
∆ ∆
W − W2 ≤ W o + l∈L vl + u∈U vu ≤ W + W2 ). Based on this, we can approximate the
P P
lcg calculation by removing ∆W from the denominator of (6.23), turning this non-linear
expression into the linear expression:

LM o + l∈L DlL vl + u∈U DuU vu

P P
. (6.24)
W
∆
We can reasonably assume that the relative error of this approximation is 2WWM ax , with
W M ax being the maximum vessel displacement allowed. Note that the same approxi-
mation can be used for the vertical and transversal center of gravity.
To complement our approach for addressing variable displacement and for an easier
integration with a mathematical model, we propose to define a fixed number of displace-
ment intervals as follows. With W M in being the minimum vessel displacement, we use
the value W ∆ to split the displacement range (i.e., [W M in ; W M ax ]) into n intervals of size
M ax M in M in ∆ M in ∆
W ∆ , where n = d W W−W ∆ e. Additionally, we define Wi = W +(i−1)W2 +W +iW ,
for i ∈ I, to be the middle point of displacement interval i.
The assumption that the vessel displacement lies within a given interval is useful
for the formulation of accurate linear approximations of the hydrostatic data when
the displacement is variable. Figure 4.1 depicts a plot of the hydrostatic data for
the trim and metacenter calculation. It is easy to see that planes representing linear
approximations of the trim and metacenter calculations are inaccurate in this case.
A closer look, however, shows that for a fixed displacement a linear approximation
of the hydrostatic data with respect to lcg could be used to approximate the trim
and metacenter. Moreover, within small displacement intervals, linear approximations
with respect to lcg and vessel displacement can be used to accurately approximate
trim and metacenter. Thus, we define linear approximations for each displacement
interval i ∈ I and use them to approximate trim and metacenter when the vessel

104
 Chapter 6. Cargo Composition Problem

displacement is within the corresponding interval. As mentioned in section 4.3, care

must be taken since the behaviour of the approximation provided by the linearization
is inaccurate for extreme vessel displacements. It is reasonable to assume, however,
that the displacement of the cargo compositions will be away from these extremes.
For the calculation of the longitudinal buoyancy forces we approximate linearly the
hydrostatics of the bonjean at each station point, following the same principle as with
trim and metacenter. For a deeper discussion on the buoyancy calculations and the
bonjean linearizations we refer the reader to section 4.3.
In chapter 4, we showed through experimental evaluation that the errors introduced
by approximating the center of gravity calculations and the linearizations of stability
and bonjean data are within an acceptable error. For the generalization presented in
this section, it is easy to see that the accuracy of the calculations remains the same.
Now we introduce the constraints modeling variable displacement and the calculation
of the longitudinal and vertical center of gravity:
X X
W
xpu + vlp + WpO = vpW ∀p ∈ P (6.25)
u∈U l∈L
X X
Wi− ψpi ≤ vpW ≤ Wi+ ψpi ∀p ∈ P (6.26)
i∈I i∈I
X
ψpi = 1 ∀p ∈ P (6.27)
i∈I

M inLi ψpi ≤ vpi

L
≤ M axLi ψpi ∀p ∈ P, i ∈ I (6.28)
X X X
DlL vpl
W
+ DuL xpu + LMpO = L
Wi vpi ∀p ∈ P (6.29)
l∈L u∈U i∈I

vpV M ≤ vpV Wi + (1 − ψpi )M axV ∀p ∈ P, i ∈ I (6.30)

vpV M≥ vpV Wi − (1 − ψpi )M axV ∀p ∈ P, ∀i ∈ I (6.31)
X
V ML
vlp + DuV xup + V MpO = vpV M ∀p ∈ P (6.32)
u∈U
X
W
vpi = vpW ∀p ∈ P (6.33)
i∈I

Wi− ψpi W
≤ vpi ≤ Wi+ ψpi ∀p ∈ P, i ∈ I (6.34)

Constraint (6.25) calculates the vessel displacement by considering the weight of the
tanks, the cargo loaded in the locations, and the weight of the empty vessel and all con-
stant weights; while constraint (6.26) sets the indicator variable corresponding to the
displacement interval that contains the current vessel displacement. Constraint (6.27)
restricts the number of active displacement intervals at the same port to be one, and
constraints (6.28) and (6.29) approximate the longitudinal center of gravity of the con-
tainer vessel. Constraint (6.28) sets to zero the variables representing the longitudinal
center of gravity of the inactive displacement intervals, while constraint (6.29) calcu-
lates the longitudinal center of gravity of the vessel based on the longitudinal moment

105
Chapter 6. Cargo Composition Problem

of the cargo in the locations, the ballast tanks, the empty vessel including constant
weights, and the displacement at each interval. Note that having an auxiliary variable
representing the longitudinal center of gravity of each displacement interval is unnec-
essary to approximate the longitudinal center of gravity of the vessel. These auxiliary
variables, however, play an important role in the stability and stress force calculations
based on linearizations of hydrostatic data.
The vcg of the vessel is approximated by constraints (6.30)-(6.32). Constraints
(6.30) and (6.31) set bounds for the vertical moment at each displacement interval.
For the active displacement interval, these two constraints turn into an equality where
the vertical moment is equal to the vcg of the vessel multiplied by the representative
weight of the active displacement interval. For the displacement intervals that are not
active these two constraints have no effect. The vertical moment of the container vessel
is calculated by constraint (6.32). Note that the vcg DuV of ballast tank u has been
approximated to be the vcg of the tank when being 75% full. It can be shown that the
integral of the quadratic error for a boxed formed tank is smallest if we use a lineariza-
tion corresponding to the vcg of the tank being 75% full. Last, constraints (6.33) and
(6.34), set the displacement variable corresponding to the active displacement interval
to be equal to the vessel displacement, and set the displacement variables of the inactive
displacement intervals to zero. Finally, the constraints that restrict stability and stress
forces to be within limits are presented below.
X X XX X
DcT Wτ xτcdt + DuT xpu + T MpO = 0 ∀p ∈ P (6.35)
t∈T RpON c∈C τ ∈T e∈E τ u∈U
X
(AW W L L M
M (Wi )vpi + AM (Wi )vpi + AM (Wi )ψpi ) = vp ∀p ∈ P (6.36)
i∈I
X
(AW W L L
T (Wi )vpi + AT (Wi )vpi + AT (Wi )ψpi ) = 0 ∀p ∈ P (6.37)
i∈I

vpM − vpV ≥ M inGM

p ∀p ∈ P (6.38)
X X
0.5DbB (AW W L L B
Bs (Wi )vpi + ABs (Wi )vpi + ABs (Wi )ψpi ) = vpn
s∈{b,b+1} i∈I

∀p ∈ P, b ∈ B (6.39)
X X X
Sα
vpf = WfS,α + Gαlf vlp
W
+ Gαuf xup − Gαbf vpb
B

l∈L u∈U b∈B

∀p ∈ P, f ∈ F, α ∈ {A, F } (6.40)
X X X
Bα
vpf = WfB,α + Aαlf Gαlf vpl
W
+ Aαuf Gαuf xpu − Aαbf Gαbf vbp
B

l∈L u∈U s∈S

∀p ∈ P, f ∈ F, α ∈ {A, F } (6.41)
SF ore S
M inSf ≤ Gf vpf + (1 − Gf )vpfAf t ≤ M axSf ∀p ∈ P, f ∈ F (6.42)
BF ore B
M inB
f ≤ Gf vpf + (1 − Gf )vpfAf t ≤ M axB
f ∀p ∈ P, f ∈ F (6.43)
The transversal center of gravity is required to be 0 (heeling angle 0) by constraint (6.35).

106
 Chapter 6. Cargo Composition Problem

Remember that constraints (6.28) and (6.34) force the variables corresponding to vessel
W L
displacement, (vpi ), and longitudinal center of gravity, (vpi ), of displacement interval
i in port p to be the actual displacement and longitudinal center of gravity of the
vessel, respectively, when displacement interval i is active, and zero otherwise. Thus,
W L
constraint (6.36) uses variables vpi , vpi , and ψpi to select the linearization coefficients
corresponding to the active displacement and approximate the plane representing the
metacenter. Likewise, constraint (6.37) approximates the plane representing the trim
of the active displacement. This constraint forces the displacement and longitudinal
center of gravity of the vessel to match the coordinates in the plane corresponding to
trim being equal to zero. The vessel’s metacentric high is restricted to be greater than
its lower bound by constraint (6.38), and the buoyancy force of the section between two
consecutive stations is approximated by constraint (6.39). The approximated plane
representing the bonjean at each station is formulated similarly to that of the trim
and metacenter. Constraints (6.40)-(6.43) compute shear force and bending moment at
each vessel frame. These constraints equivalent to constraints (4.26)-(4.29) presented
in section 4.3.1 and will not be discussed any further.
In this section we have introduced the first model with variable displacement, that
assigns containers to vessel slots, that is actually solvable in real instances. However,
since this model does not scale well in some scenarios, in the next section we present a
decomposition that improves scalability with a small impact on model accuracy.

6.5 Decomposed model for the CCP

In this section we present a heuristic decomposition approach for the CCP based on
the 2-phase decomposition depicted in Figure 1.3. This decomposition has been used
in previous chapters to stow containers into vessel slots. Considering that our interest
is not to generate stowage plans but to verify whether cargo compositions that match
specific criteria can be generated, we extend the mathematical programming model for
master planning introduced in chapter 4, to generate feasible cargo compositions and
disregard the slot planning phase. It is important to notice that due to the heuristic
nature of the decomposition, cargo compositions generated by optimizing the math-
ematical model of the master planning phase are not necessarily optimal. Below we
introduce the extra constants and variables used in the mathematical programming
model.
Decision Variables
xtτ
l ∈Z Number of containers of type t to be stowed in location l during transport t.

Auxiliary Variables
{O,OP }
vpl Number of hatch overstowing containers (O), overstowage probability (OP ) in
location l at port p.
W {20 ,40 }
vpl Weight of 20 ’, 40 ’ containers stowed in location l at port p.
γpl Whether there are 20 ’ containers stowed in location l at port p.

107
Chapter 6. Cargo Composition Problem

Below we introduce the objective function of the decomposed model used to solve
the cargo composition problem.
X XX XX X
M ax : Ctτ xtτ
l − ( (C O O
v
p pl + C OP OP
p vpl ) + C C C
v
p p + C U
p xup ) (6.44)
t∈T R l∈L τ ∈T p∈P l∈L u∈U

The objective function is a weighted sum corresponding to that of the IP model. There
is, however, a minor difference. Stack overstowage cannot be computed in this model
since the level of abstraction of the decision variables disallow it. To address this issue,
we replace stack overstowage with a new objective that estimates the probability of 20 ’
containers being overstowed by 40 ’ containers within the same location. Though this
objective does not measure overstowage within a location as accurately as the stack
overstowage does, it provides a good estimate. Next we present the set of location
capacity constraints, which are an abstraction of the stacking constraints presented
in the IP model. These constraints make it likely to stow the containers assigned to
locations in the second phase of the decomposition.
X X
xtτ
l ≤ Kpl
HC
∀p ∈ P, l ∈ L (6.45)
t∈T RpON τ ∈T HC
X X
Φτ xtτ V
l ≤ Kpl ∀p ∈ P, l ∈ L (6.46)
t∈T RpON τ ∈T
X X
ΠTEU
τ xtτ TEU
l ≤ Kpl ∀p ∈ P, l ∈ L (6.47)
t∈T RpON τ ∈T
X X
xtτ α
l ≤ Kpl ∀p ∈ P, α ∈ L, l ∈ L (6.48)
t∈T RpON τ ∈T α
X X
xtτ RF
l ≤ Kpl ∀p ∈ P, l ∈ L (6.49)
t∈T RpON τ ∈T RF
X X
ΠTEU
τ xtτ RC
l ≤ Kpl ∀p ∈ P, l ∈ L (6.50)
t∈T RpON τ ∈T RF
X X
Wα
vpl = Wτ xtτ
l ∀p ∈ P, l ∈ L, α ∈ L (6.51)
t∈T RpON τ ∈T α
W 20
vpl ≤ Wpl20 ∀p ∈ P, l ∈ L (6.52)
W 20 W 40 40
0.5vpl + vpl ≤ Wpl ∀p ∈ P, l ∈ L (6.53)
Constraint (6.45) restricts the number of high-cubes to be within the capacity of the
location. Recall that high-cube containers utilize more volume capacity than dry cargo
containers, thus, constraint (6.46) limits the volume of the containers stowed in location
l to be within limits. Constraints (6.47)-(6.50) and (6.52)-(6.53) are equivalent to
constraints (4.6)-(4.9) and (4.11)-(4.12) presented in section 4.3.1, respectively, and
W 20
will not be discussed any further. Constraint (6.51) defines auxiliary variables vpl
W 40
and vpl . Though this constraint resembles constraint (4.10), they differ with respect

108
 Chapter 6. Cargo Composition Problem

to the coefficients of the variables. In the model introduced in section 4.3.1 the weight
of a container type changes according to the transports, while in the model presented
in this section, container types have fixed weight. We now present two constraints to
calculate auxiliary variables used in the stability and stress force computations.
W W 20 W 40
vpl = vpl + vpl ∀p ∈ P, l ∈ L (6.54)
X X X
vpV M L = V
Dpl Wτ xtτ
l (6.55)
t∈T RON l∈L τ ∈T

Constraint (6.54) calculates the weight of the cargo stowed in a location at a specific
port, while auxiliary variable vpV M L (the vertical moment of the cargo) is calculated by
constraint (6.55). Constraints on the stability and stress forces of the vessel are listed
below.
X
DuT xpu + T Mp0 = 0 ∀p ∈ P (6.56)
u∈U
(6.25) − (6.34)
(6.36) − (6.43)

Constraint (6.56) calculates the transversal center of gravity of the vessel. Notice that
locations are not included in this calculation since due to their construction, their tcg is
assumed to be in the center of the vessel. Figure 2.5 shows the locations of a bay. For
the locations in the center, we can assume that during the slot planning phase weights
will be evenly distributed within the location, displacing its tcg towards the center of
the vessel. With respect to the outer locations, we can see in Figure 2.5 that storage
areas on the starboard and port side form a single location. Since these two storage
areas are usually symmetric, we also assume an even distribution of the weights during
slot planning. Thus, the tcg of the outer locations is displaced towards the center of
the vessel as well. The remaining stability and stress force calculations for this model
are the same as for the IP model. To derive the objective values we introduce the
constraints below:
X X X
( xtτ U
l ) ≤ Ml δpl ∀p ∈ P, l ∈ LON (6.57)
k∈LU
l
t∈T RA τ ∈T
X X
xtτ H O
l − Ml (1 − δpl ) ≤ vpl ∀p ∈ P, l ∈ LON (6.58)
t∈T RpOV τ ∈T
X XX
xtτ C
l ≤ vp ∀p ∈ P, n ∈ N (6.59)
t∈T RpA l∈Ln τ ∈T
X X
xtτ H
l ≤ Ml γpl ∀p ∈ P, l ∈ L (6.60)
t∈T RpA τ ∈T 20
X X
xtτ H OP
l − 0.5Ml (1 − γpl ) ≤ vpl ∀p ∈ P, l ∈ L (6.61)
t∈T RpOV τ ∈T 40

109
Chapter 6. Cargo Composition Problem

Constraints (6.57) and (6.58) model hatch overstowage similar to the IP model, where
indicator variables δpl is activated when there are containers being loaded or unloaded
at port p in the locations below l. Then, constraint (6.58) uses the indicator variables
to calculate a lower bound for hatch overstowage. Also, similar to the computations
in the IP model, constraint (6.59) estimates the crane makespan. Finally, since 20 ’
containers have to be stowed under 40 ’ containers within a location, when there are
40 ’ containers stowed in location l to be discharged after port p there is a probability
that these containers overstow 20 ’ containers to be loaded or discharged at port p in
location l. Constraint (6.60) and (6.61) model the probability of overstowage within
a location. Indicator variable φpl is activated by constraint (6.60) when there are 20 ’
containers stowed in l to be loaded or discharged at port p. Constraint (6.61) then
provides a lower bound for the number of 40 ’ containers being discharged after p, that
will potentially overstow 20 ’ containers to be loaded or discharged at port p.

6.6 Experiments
We test the two models presented in this chapter on a benchmark suite of ten con-
tainer vessels provided by our industrial collaborator. Table 6.5 presents a summary of
the vessels, sorted by TEU capacity. These vessels represent most of the main vessel
categories with respect to TEU capacity, including Feeder, Panamax, Post Panamax,
and Ultra Large container vessels. Note that the TEU capacity only reflects the vessel
size for the IP model introduced in section 6.4. In the model introduced in section 6.5,
containers are assigned to vessel locations instead of stowing them in vessel slots, thus,
a more accurate measure of vessel size for this model is the number of vessel locations.

ID Capacity (TEU) Locations

1 15000 92
2 10000 87
3 10000 100
4 9000 100
5 8000 80
6 7000 90
7 7000 65
8 5000 71
9 4000 61
10 3000 40

Table 6.5: Vessel features. Container vessels features. The first column represents the vessel
ID. The second column shows the TEU capacity of the vessels rounded to the nearest
thousand due to confidentiality issues, while the third column provides the number of
vessel locations.

110
 Chapter 6. Cargo Composition Problem

Recall that the decision variables xtτ

l of the decomposition model presented in sec-
tion 6.5 represent the number of containers of type τ assigned to location l during
transport t. We evaluate experimentally the impact on CPU time and solution quality
when the integrality constraint on variables xtτ
l is removed. The result of this relaxation
is that the solutions produced by the decomposition will include continuous numbers
of containers to be stowed in locations. Though it is physically impossible to do so,
the impact of fractionality is eased by the fact that the TEU capacity of the locations
(from 62 to up to 171 TEUs in general, though a few locations can be smaller), allows
rounding to be performed without having a significant impact, (Pacino et al. (2011)).
For the remainder of the chapter, we denote the decomposition model with integrality
constraints as the IP decomposition, and the model where these constraints are removed
as the MIP decomposition.

We have devised two main sets of experiments to evaluate the performance of the
approaches. In the first set, we focus on evaluating the behaviour of the models in a
single port scenario, where we stress the models by optimizing TEU intake, intake of
containers with specific features (e.g., reefer, high-cube, 20 ’, 40 ’, and different weight
classes), while requiring maximum intake, and cargo revenue. In the second set, we
define a multi-port scenario to analyze the impact of different features (e.g., number of
discharge ports, transport density, and number of types considered), on the complexity
of finding cargo compositions that optimize revenue. In both sets of experiments we split
the displacement range in intervals of size 1% of the maximum displacement to keep the
approximations of the center of gravity and the linearization of hydrostatic data with an
acceptable error. We constrain the displacement to be greater than 50% of maximum
displacement, a reasonable assumption given the type of experiments performed in this
section, forcing our models to only consider a total of 50 displacement intervals. Also
note that for some of the vessels from our benchmark suite, the linearization of the
hydrostatic data was problematic with displacements below 50%. All the experiments
were run on as AMD Opteron processor at 2.8 Ghz, with 4GB of memory. The models
have been implemented using the modeling language OPL and executed using CPLEX
12.5. The time limit for all the experiments was 3,600s.

6.6.1 Single Port Experiments

In this section we analyse the CPU time and solution quality of the results produced
by the IP model, the IP decomposition, and the MIP decomposition mentioned above
in a single port setup. For these experiments, the objective function of each model
is limited to consider the stowed containers and their related coefficients, disregarding
overstowage, crane split, and ballast water penalization.

111
Chapter 6. Cargo Composition Problem

T 32 T 20 T 40
ID
IP(%) DMIP(%) DIP(%) IP(%) DMIP(%) DIP(%) IP(%) DMIP(%) DIP(%)

1 100.00 100.00 99.99 90.40 90.40 90.39 100.00 100.00 100.00

2 100.00 100.00 100.00 85.90 85.90 85.90 99.62 99.62 99.62
3 99.98 100.00 99.98 85.59 85.59 85.59 99.98 100.00 99.98
4 96.73 96.75 96.73 78.12 78.12 78.12 96.73 96.75 96.73
5 100.00 100.00 100.00 97.19 97.19 97.19 100.00 100.00 100.00
6 96.03 96.03 96.03 77.07 77.07 77.07 95.82 95.82 95.82
7 100.00 100.00 100.00 96.22 96.22 96.22 99.82 99.82 99.82
8 100.00 100.00 100.00 92.79 92.79 92.79 100.00 100.00 100.00
9 100.00 100.00 100.00 95.58 95.58 95.58 96.53 96.53 96.53
10 100.00 100.00 100.00 89.14 89.14 89.14 100.00 100.00 100.00

Table 6.6: Intake optimization. First column is the ID of the vessel. Next, each set of three
consecutive columns is the percentage of vessel capacity that could be stowed in the vessels
(maximum intake), considering only container types of the sets T 32 , T 20 , T 40 , by the IP
(IP) model, the MIP decomposition (DMIP), and the IP decomposition (DIP), respectively.

Intake optimization

The first analysis we perform is intake optimization. We attempt to find a cargo

composition that maximizes intake for each container vessel, selecting containers from
three different type-sets. These type-sets mainly differ from each other in two ways,
their size and the length feature of the types they consider. Thus, since the size of the
type-sets is related to the number of variables of the models and a considerable number
of constraints depend on the length of the container types, in this experiment we study
the impact of these two features in CPU time and quality of the solutions generated
by our models. Remember that a type τ is a 4-tuple (l, h, r, w). For the first type-
set, T 32 , we define l ∈ L = {20 , 40 }, h ∈ H = {DC , HC } (dry cargo or high cube),
r ∈ R = {RF , NR}, w ∈ W = {3, 6, 9, 14, 21, 27}, and formally restrict T 32 to be:
T 32 = {(20 , DC , r, w) | r ∈ R, w ∈ W \ {6}} ∪ {(40 , h, RF , w) | h ∈ H, w ∈ W \ {6}} ∪
{(40 , h, NR, w) | h ∈ H, w ∈ W }. For the second set, T 20 , we redefine l ∈ L = {20 },
h ∈ H = {DC }, r ∈ R = {NR}, w ∈ W = {3, 9, 14, 21, 27}, and for the third set, T 40 ,
we only redefine l ∈ L = {40 } and w ∈ W = {3, 6, 9, 14, 21, 27} with respect to the
values in the second set. Table 6.6 presents the optimal intake for the three different
type-sets defined above, solved by the complete model (IP), the IP decomposition (DIP),
and the MIP decomposition (DMIP).
For each type-set, the three approaches were almost able to stow the same number
of containers for all vessels, with a maximum variation of 0.02% (type-sets T 40 and
T 32 , vessels 3 and 4, and type-sets T 32 and T 20 , vessel 1), showing little impact of
using the IP and MIP decompositions for solving the cargo composition problem in this
case. With respect to the impact on maximum intake when considering type-sets that
include different lengths, we take as a reference point the maximum intake reached with
type-set T 32 for each vessel, since this type-set includes types with both lengths, (20 ’
and 40 ’). The most drastic intake reduction comes from the type-set with only 20 ’

112
 Chapter 6. Cargo Composition Problem

containers (T 20 ), with loses starting at 2.81% (vessel 5), and going all the way down
to 18.96% (vessel 6), with respect to T 32 intake. There are two reasons for this intake
loss. The first is that there are sections of the vessel where it is physically impossible to
stow 20 ’ containers, and without 40 ’ containers available, these sections remain empty.
The second reason is the 20 ’ weight capacity. This capacity is usually smaller than the
40 ’ weight capacity on deck, further reducing the number of 20 ’ containers that can be
loaded in the container vessels. For the third type-set, T 40 , none of the three models
was able to reach the T 32 intake for four of the vessels (vessels 2, 6, 7, and 9), though
the capacity loss, ranging from 0.18% (vessel 7), to 3.47% (vessel 9), was less than that
of type-set T 20 . The main reason for the reduction in this case is the fact that there
are also some vessel sections where it is impossible to stow 40 ’ containers (e.g., odd
slots).
In conclusion, the length of the types included in the type-set has a considerable
impact on the maximum intake achieved for a vessel. In our results, vessels are more
suitable for transporting 40 ’ than 20 ’ containers, since we can observe a reduction in
maximum intake in all the vessels when using type-set T 20 , while only in four (vessels
2, 6, 7, and 9), when using type-set T 40 . Moreover, the difference in the reductions
0
is considerable, with a maximum reduction of 18.96% for type-set T 20 , and 3.47% for
type-set T 40 . This impact, however, depends on the vessel’s physical layout.

It took just a few seconds for the IP and MIP decompositions to maximize intake for
all the vessels when considering all three type-sets , with the MIP decomposition finding
optimal solutions for all instances in less than a second. The left plot in Figure 6.1
depicts the CPU time of the IP decomposition, indicating a tendency towards the type-
set T 32 being the hardest to optimize. It took at most 12.03s (vessel 1), and 4.03s on
average for the IP decomposition to find the maximum intake for type-set T 32 , while it
took 0.705s and 0.6s in average to find the maximum intake for type-sets T 20 and T 40 ,
respectively.
With respect to the IP model, the right plot in Figure 6.1 depicts the time it took to
find the maximum intake for each vessel and each type-set. A similar tendency to that
of the IP decomposition is observed, where instances considering type-set T 32 are the
hardest to optimize. In this case, when optimizing type-set T 32 vessel 2 took the longest
time (1,052.01s), and the average was 228.26s. Optimizing intake when considering
type-sets T 20 and T 40 took 98.3s and 32.55s on average, respectively. Thus, from the
two plots depicted in Figure 6.1 we can conclude that there is a tendency towards larger
type-sets taking more CPU time to find optimal solutions.

Type intake optimization

Next, we attempt to stress our models by optimizing specific container types using
type-set T 32 . We constrain the intake in this experiment to be within one percent of
the maximum intake from table 6.6. We start by optimizing intake of container types
with specific length, height, and reefer capabilities. Table 6.7 presents the results for

113
Chapter 6. Cargo Composition Problem

25 1200
32 32
20 20
T40 T40
T T
1000
20

800
15
Time(s)

Time(s)
600

10
400

5
200

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Vessel ID Vessel ID

(a) (b)
Figure 6.1: CPU time of intake optimization of the, (a) IP decomposition. (b) IP model.

reefer and high-cube containers, while table 6.8 provides the results for 20 ’, and 40 ’
containers.
R HC
ID
IP(%) DMIP(%) DIP(%) IP(%) DMIP(%) DIP(%)

1 100 100 100 - 85.32 -

2 - 100 100 - 87.87 -
3 100 100 - 69.52 70.24 -
4 - 100 100 66.40 67.15 67.12
5 - 100 100 72.04 74.62 -
6 100 100 100 - 64.40 64.30
7 100 100 100 - 72.24 72.11
8 100 100 100 74.31 78.00 77.76
9 100 100 100 83.22 96.53 96.53
10 100 100 100 80.11 85.86 85.30

Table 6.7: Type intake optimization for reefer and high-cube containers. First column is
the ID of the vessel. The first set of three consecutive columns (second, third and fourth
columns) are the percentages of reefer slot capacity loaded with reefer containers, while the
second set of three consecutive columns are the percentage of TEU capacity loaded with
high-cube (HC) containers by the IP (IP) model, the MIP decomposition (DMIP), and the
IP decomposition (DIP), respectively, while keeping intake within 1% of the optimal. The
dash represents instances that could not be solved within the 3,600s time limit.

Given the different levels of accuracy between the IP model and the decompositions,
we expect optimal solutions to be slightly different when we stress the models. As seen
in tables 6.7 and 6.8, most of the discrepancies involving a considerable number of con-
tainers took place when optimizing high-cubes. In all solutions found by the IP model
in this case, less high-cube containers were stowed than in their decomposition counter-
parts. One reason for this is the abstraction of stack capacities into location capacity
made in the decompositions and the accumulated error that this abstraction generates.
The volume capacity of a stack fits a specific mix of high-cube and normal contain-

114
 Chapter 6. Cargo Composition Problem

20’ 40’
ID
IP(%) DMIP(%) DIP(%) IP(%) DMIP(%) DIP(%)

1 90.40 90.40 90.39 100.00 100.00 100.00

2 85.90 85.90 85.90 99.62 99.62 99.62
3 85.59 85.59 85.59 99.98 99.99 99.98
4 - 78.12 78.12 96.73 96.74 96.73
5 97.19 97.19 97.19 100.00 100.00 100.00
6 77.07 77.07 77.07 95.82 95.82 95.82
7 96.22 96.22 96.22 99.82 99.82 99.82
8 92.79 92.79 92.79 100.00 100.00 100.00
9 95.58 95.58 95.58 83.22 96.53 96.53
10 89.14 89.14 89.14 100.00 100.00 100.00

Table 6.8: Type intake optimization for 20 ’ and 40 ’ containers. First column represents
the ID of the vessel. The first set of three consecutive columns (second, third and fourth
columns), presents the percentage of 20’ (TF), and 40’ (FF) containers loaded by the IP
(IP) model, the MIP decomposition (DMIP), and the IP decomposition (DIP), respectively,
while keeping intake within 1% of the optimal. The dash represents instances that could
not be solved within the 3,600s time limit.

ers. Stowing mixes with more normal or high-cube containers force a sub-utilization of
the stack’s volume capacity, generating a residual capacity where no container can be
stowed. When abstracting volume stack capacities into volume location capacity, the
residual capacity of the stacks is aggregated such that more high-cube containers can be
assigned to a location than actually can be stowed in the stacks. Thus, when optimiz-
ing high-cube containers the decomposition erroneously uses the aggregated capacity
to stow containers that the IP model will not.
Another discrepancy between optimal values generated by the IP model and the
decompositions occurs when optimizing 40 ’ containers intake for vessel 9 (13.31%). We
can see from table 6.6 that when optimizing intake with type-set T 40 the results are
the same for the IP model and the decompositions. Since T 40 ⊂ T 32 , we would expect
the similar results in table 6.8. As we can see, however, the constraint forcing intake to
be within 1% of the optimal impacts the number of 40 ’ containers that can be stowed
in vessel 9. To achieve optimal intake in this specific case, the IP model must load more
20 ’ containers than the decompositions, indicating an imprecision in the abstraction of
stacking constraints related to the interaction of 20 ’ and 40 ’ containers.

With respect to CPU time and complexity of finding optimal solutions, the MIP
decomposition easily finds optimal solutions for all vessels in 0.43 seconds, on average,
and a maximum time of 0.51 seconds. The IP decomposition managed to easily find
solutions optimizing 20 ’ and 40 ’ containers, but struggled with reefer and high-cube
containers. Optimizing high-cube containers was the hardest, finding solutions only for
six out of ten vessels. Among the vessels that were solved, no optimal solution was
found for four of them (vessels 6-8, and 10), though their final gap was less that 0.7%.
With respect to the cases where reefer containers were optimized, there was only one

115
Chapter 6. Cargo Composition Problem

vessel for which no solution was found, while optimal solutions were found for the rest.
The IP model also struggled when optimizing high-cube and reefer containers, and
slightly when optimizing 20 ’ and 40 ’ containers. With respect to the optimization of
40 ’ containers, no optimal solution was found for vessel 9, though the final gap was
less than 0.1%. In the case of the 20 ’ container optimization, optimal solutions were
found within the time limit for all the vessels solved. For two of the vessels for which
solutions were found when optimizing high-cube containers, optimal solutions could
not be generated within the time limit (vessels 4 and 9), though their gap was less
than 0.2%. When optimizing reefer containers, optimal solutions were found for the six
vessels solved.
There are two important remarks to be made. First, we can see that in some specific
cases (high-cube optimization, optimizing 40 ’ for vessel 9 ), there is an impact of using
the decomposition model on the accuracy of the solutions. Though the inaccuracies
can be considerable (up to 13%), they can be addressed by adding a more refined
set of capacity constraints into the decomposition model. For instance, in the case
of the high-cube optimization, the volume capacity limits could be modified based on
the combination of high-cube and normal containers stowed in the location, instead of
being fixed as they are in the current model. Second, the MIP and IP decompositions
generate almost identical solutions in all cases, with the MIP decomposition using a
small fraction of the CPU time used by the IP decomposition.
Now we proceed to maximize the loading of containers with different weights classes.
Table 6.9 presents the results of optimizing containers with specific weights from the
type-set T 32 . As with the previous experiments, we constrain the intake to be within
W9 W21 W27
ID
IP DMIP DIP IP DMIP DIP IP DMIP DIP

1 12759 12838.97 12838 5093 5147.15 5147 3860 3860.36 3860

2 8612 8541.15 8540 4854 4551.63 4547 3751 3694.39 3693
3 7670 7365.58 7361 4929 4587.53 - 3871 3805.65 3800
4 3320 7341.29 7337 4570 4363.19 - 3506 3489.69 3489
5 8217 8217.00 8217 4836 4835.09 4834 3694 3706.77 -
6 6206 6218.86 6216 - 3856.97 - 2957 2959.43 -
7 6292 6202.45 6202 3663 3599.03 3596 2831 2795.60 2795
8 4352 4078.44 4078 2788 2612.54 2606 2319 2178.49 2167
9 3578 3465.54 3465 1996 1903.63 1901 1635 1582.25 1577
10 2282 2253.93 2253 1375 1276.78 1276 1062 1054.81 1054

Table 6.9: Weight class optimizations. First column is the ID of the vessel. Next, each set
of three consecutive columns are the maximum number of containers weighing 9 (W9), 21
(W21), and 27 (W27) tons stowed by the IP (IP) model, the MIP decomposition (DMIP),
and the IP decomposition (DIP), respectively, while reaching intake within 1% of the
optimal. The dash represents instances that could not be solved within the 3,600s time
limit

1% of the one from table 6.6. At first glance, we can easily see from table 6.9 that
as we increase the weight of the containers to optimize, their intake is reduced. Time

116
 Chapter 6. Cargo Composition Problem

wise, the MIP decomposition managed to compute the optimizations within, at most, 3
seconds. The IP decomposition managed to optimize containers weighing 9 tons within
at most 75 seconds, but struggled with optimizing containers weighing 21 and 27 tons.
When optimizing 21 ton containers, the IP decomposition did not manage to find an
optimal solution for vessel 5, though the remaining six vessels for which a solution was
found were solved optimally within 7 seconds, including the two largest vessels. The
three vessels that were not solved in this case are among the ones with the largest
number of locations. When optimizing 27 ton containers, the IP decomposition found
optimal solutions for seven of the solved vessels within 9 seconds, while the solution
for vessel 10 was found in 119 seconds. The IP decomposition approach does not seem
to be affected by the vessel size when optimizing weight classes, though there is clear
tendency of the problem becoming harder as container weight increases.
As with the IP decomposition, the IP model also becomes harder to solve when we
increase the container weight we want to optimize. When optimizing 9 tons containers,
we found optimal solutions for all but one vessel (vessel 4, 55% gap), in an average time
of 1,011.14s. For the optimization of containers weighing 21 and 27 tons, the number
of optimal solutions kept decreasing, with the optimization of 27 ton containers being
the most affected, (only the last three vessels solved optimally). The average time for
optimizing 21 ton containers was 1,191.88s, and 2,526s for optimizing 27 tons. In the
case of the IP model there also seems to be a tendency that relates vessel size with CPU
time, since the vessels that cannot be solved optimally are usually the largest ones.
In table 6.9 we can also see differences between optimal solutions produced by the
IP model and the decompositions. In two of these situations, when optimizing 9 ton
containers for vessel 4 and when optimizing 21 ton containers for the vessel 1, there
was an optimality gap in the IP results of 55% and 1%, respectively, explaining the
difference in the results. In the cases where the difference favours the IP model, such
as when optimizing 21 and 27 ton containers for most of the vessels, this difference
can be explained by how the vcg of the vessel is calculated in the IP model and the
decompositions. Recall that for the decompositions we pre-compute a fixed vcg for each
location under the assumption that the weight of the containers stowed in them will be
uniformly distributed. The IP model allows a more accurate computation of the vcg
in the locations, since it can use the exact position of the containers to do so. In the
cases where containers of the same weight are stowed in a location this does not have
any impact. However, when containers of different weights are stowed in a location the
vcg’s location can be lowered by stowing heavy containers at the bottom cells and light
containers at the top. This reduction allows the stowing of heavier containers than
those stowed when the vcg is considered fixed, as in the decompositions.
In conclusion, the MIP decomposition scales well when optimizing containers with
specific features, solving all the instances in less than 5s and producing solutions with
negligible differences with respect to those produced by IP decomposition. The IP
model and IP decomposition, however, have problems coping when the models were
stressed with these kind of optimizations, leaving several instances unsolved within the
time limit. We also observed small inaccuracies between the intakes in the solutions

117
Chapter 6. Cargo Composition Problem

produced by the decompositions and the IP models. Though there was a case where
the discrepancies reached 13.31% (vessel 9, 40 ’ intake optimization), in the rest of the
cases they did not go over 5%. Moreover, these discrepancies can properly be tackled
with a more refined set of capacity constraints and a more accurate approximation of
the vcg of the locations.

Revenue optimization

Now we introduce a simplified revenue model for loading containers in a single port to
evaluate the impact of the accuracy of the stowage models when doing revenue analysis.
We assume that the revenue of a dry 40 ’ container is US$1,000. Based on interviews
with our industrial collaborators and brokers, we calculate the revenue for containers
with other features as follows: 20 ’ containers generate 62,5% of the revenue of a 40 ’
container. Reefer containers are 75% more profitable than normal containers and high-
cube containers are US$100 more profitable than normal containers. 20 ’ containers
weighing between 18 and 21 tons are US$150 more profitable than lighter containers,
while those weighing between 21 and 24 tons are US$350 more profitable. Containers
weighing more than 24 tons are US$500 more profitable. Finally, we assign a profit
of US$100 for empty containers. Though these containers do not generate a profit
for the liner shipping company, they need to be constantly repositioned due to trade
imbalance between commercial zones. We assign them a small profit margin so that
they are loaded in the vessel after more profitable containers have been loaded.
Figure 6.2 depicts the results of optimizing revenue for a single port based on the
container revenue model described above for the IP model and the IP and MIP de-
compositions. With respect to the revenue reported for each model, there is a small
difference between the decompositions and the IP model, favouring the IP model for
vessels 2, 3, 4, and 8. As we mentioned before, the IP model has a more accurate
definition of the vcg of the locations than that of the decompositions. This accuracy
allows the IP model to generate stowage plans loading heavier containers, the most prof-
itable containers. Time-wise, the IP and MIP decompositions find optimal solutions in
at most 98.1s and 0.31s, respectively, though in the case of the IP decomposition the
second largest time is 9.33s. With respect to the IP model, no optimal solution was
found for the first three vessels (though the gap was less than 0.03%), and the optimal
solution for the smallest vessel was found in 22.86s. For this model, it is clear that the
vessel size impacts the time it takes to find optimal solutions.
To test the impact of modeling stowage accurately, we have implemented a simplified
model that only constrains the TEU capacity of the vessel and the cargo weight to be
within limits, similar to the model described in Feng and Chang (2008), and compare
the results with the models introduced in this chapter. Figure 6.2 depicts the results
of this experiment. There is a clear overestimation of the revenue of all vessels. On
average, the revenue is overestimated by 7.84%, going as low as 4%, and reaching 11.43%
in the worst case.

118
 Chapter 6. Cargo Composition Problem

1.4e+07
IP
DMIP
DIP
Simplified
1.2e+07

1e+07
Revenue(US$)

8e+06

6e+06

4e+06

2e+06
1 2 3 4 5 6 7 8 9 10
Vessel ID

Figure 6.2: Revenue model for single port with accurate stowage planning models and a
simplified model.

Variable displacement relaxation

The models introduced in this chapter used an extra set of auxiliary boolean variables
to model variable displacement. As mentioned above, for this experimental setup we
have used 50 intervals, thus introducing 50 auxiliary boolean variables. In this section,
we study the behaviour of these variables in the linear relaxation of our models. For
each model, we solve 170 instances optimizing intake, type intake, and revenue, and
determine the number of boolean variables representing displacement intervals that are
greater than zero. Note that we only consider a single decomposition model since the
linear relaxation of the IP and MIP decompositions are the same. Table 6.10 presents
the results of the IP model and the decomposition.
As shown in table 6.10 that a considerable part of the instances have two boolean
variables or less that are different from zero in the linear relaxation, (87.06% and
78.82% for the IP model and the decomposition, respectively). Moreover, there are
only 4 instances out of 340 with more than three boolean variables different to zero.
Additionally, the linear relaxation seems to be tighter in the IP model than in the
decomposition since in 18.24% of the IP model instances the solution was integer,
while this was the case only for 7.65% of the instances solved by the decomposition.
Thus, our experiments show a strong linear relaxation of the model with respect to
the auxiliary boolean variables representing the displacement intervals, reducing the
overhead of branching on these variables.

119
Chapter 6. Cargo Composition Problem

IP D
Variables
Instances Percentage Instances Percentage
1 31 18.24 13 7.65
2 117 68.82 121 71.18
3 21 12.36 33 19.41
4 1 0.59 2 1.18
5 0 0.00 1 0.59

Table 6.10: Linear relaxation of auxiliary boolean variables representing the displacement
intervals. First column is the number of boolean variables greater than 0; second and
third columns are the number and percentage of instances for the IP model; and fourth
and fifth columns the number and percentage of instances for the decomposition.

6.6.2 Multi-Port Experiments

In this section we evaluate the IP model and the IP and MIP decompositions in a
multi-port setup. We study the impact that the number of ports in the schedule,
the size of the type-set, and the transport density have on the revenue, the response
time, and the integrality gap (if no optimal solution is found within the time limit).
We define transport density as the percentage of selected transports leaving a port.
For instance, in a schedule with three ports, the total number of transports leaving
the first port is two, e.g., h1, 2i and h1, 3i. A transport density of 100% includes both
transports, while one of 50% includes one transport, either h1, 2i or h1, 3i. The schedule
we study describes a journey from Asia to Europe, (See table 6.11). We generate
artificial demands by adding randomly generated lower bounds on some of the container
types traveling between ports. We do this such that once the vessel is leaving the Asian
trade zone, the lower bounds represent at most, 70% of the total volume capacity. The

Port Code D. (km) Over. (US$) C.P. (h) C.R. (US$) Zone
Qingdao CNTAO 439 25 25 260 A
Ningbo CNNGB 87 36 25 250 A
Shanghai CNSHA 968 62 30 277 A
Shenzhen CNYTN 1872 78 21 283 A
Tanjung Pelepas MYTPP 9030 59 22 256 A
Le Havre FRLEH 250 148 25 219 E
Rotterdam NLRTM 256 148 20 243 E
Bremerhaven DEBRV 109 121 25 293 E
Hamburg DEHAM 386 224 28 277 E
Antwerp BEANR 0 122 22 270 E

Table 6.11: Port information of the schedule used in the multi-port setup. First column is the
port name, while the second one represents the port code. Third column is the distance in
kilometers from the current port to the next one in the schedule. The fourth column is the
price of a overstowing move, while the fifth column is the crane productivity, i.e., number
of moves performed by a crane per hour. Seventh column is the price of a crane per hour,
and the last column is the trade zone where the port belongs, i.e., A= Asia, E= Europe.

120
 Chapter 6. Cargo Composition Problem

revenue for each container type is calculated based on the revenue model from the
single port experiments, where the revenue of stowing a 40 ’ container (US$ 1,000), is
replaced by the corresponding value in table 6.12. The information presented in these
two tables is partially based on the benchmark introduced in Brouer et al. (2013),
and interviews with container brokers. The coefficients of the remaining objectives are
calculated based on the values from table 6.12. For the IP model, the coefficient for
hatch and stack overstowage is read from the table, while the crane makespan coefficient
is calculated based on the crane productivity, the hourly crane rate, and the number
of cranes assigned to the vessel. Crane productivity and hourly crane rate have been
randomly generated with values between 20 and 30 moves, and US$200 and US$300,
respectively. The number of cranes assigned to a vessel in each port is six. The ballast
water penalization is calculated for each port based on the distance between the current
port and the next port. We read this information from table 6.11 and multiply it by a
scaling constant (0.01). For the decomposition models, the probability of overstowage is
30% of the cost of the overstowage objective. For this experimental setup we consider

M
PP
B

R
A

RV
H
G

N
SH

T
LE
T
N

EH
EB

EA
LR
Y
N

FR
M

D
N
C

B
CNTAO 300 300 300 810 2990 3050 2970 3040 3020
CNNGB - 270 280 790 2850 3000 2800 3010 3000
CNSHA - - 740 530 3160 2960 3070 3140 3120
CNYTN - - - 410 3230 3270 3440 3180 2940
MYTPP - - - - 2820 2750 2700 2760 3350
FRLEH - - - - - 1200 1200 1200 1200
NLRTM - - - - - - 390 400 1200
DEBRV - - - - - - - 400 410
DEHAM - - - - - - - - 400
BEANR - - - - - - - - -

Table 6.12: Transportation revenue matrix for a 40 ’ container in US$.

schedules visiting 4 different number of ports (3, 5, 7, and 10), where the last port in
the Asian trade zone, Tanjuang Pelepas, is always in the middle of the journey. We
define two extra type-sets, T 24 = T 32 − {(l, h, r, w)|l ∈ L, h ∈ H, r ∈ R, w ∈ {6, 14}},
and T 16 = T 24 − {(l, h, r, 27)|l ∈ L, h ∈ H, r ∈ R} ∪ {(40, HC, r, 3)|r ∈ R}, and three
different transport densities: 50%, 75%, and 100%. Thus, for each of the ten container
vessels from table 6.5, we generate 36 multi-port instances and solve them with the IP
model and the IP and MIP decompositions. Table 6.13 summarizes the scalability of
the approaches.
It is evident that the only approach that scales is the MIP decomposition, solving
91.7% of all instances. The IP decomposition managed to solve only 3% of the instances,
and the IP model 0.55%. Memory seems to be a major issue for the IP model, since it
affects 47.2% of the instances. This is somehow expected since the number of variables
and constraints used in this model is considerably greater than that of the IP and MIP

121
Chapter 6. Cargo Composition Problem

Solver Instances Solved Time out Out of mem. infeasible

IP 360 2 183 170 5
DMIP 360 330 29 0 1
DIP 360 13 344 3 0
Table 6.13: Summary of the experimental results in the multi-port setup. The first column
represents the solver used to solve the instances, the second column the total number of
instances, the third column the number of instances solved within 3,600s, the fourth the
number of instances that timed out and, the fifth column the number of instances that ran
out of memory. Finally, the last column represents the number of unfeasible instances.

decompositions and increases proportionally to the number of ports and transports.

With respect to the IP decomposition, all solved instances have a schedule of 3 or 5
ports, with a tendency towards low transport density and size of type-set. All instances
but one are within the last 6 vessels (smaller vessels), with 38% being instances of vessel
10. This tendency seems to indicate a possible influence of the number of variables and
constraints in the problem’s hardness, though it is not the only factor.
With respect to the MIP decomposition, all instances that were not solved have a
schedule visiting the maximum number of ports. The 29 unsolved instances are spread
over the first nine vessels, with the greatest number (7 instances), belonging to vessel
3. With respect to transport density, 82.7% of the unsolved instances have a density of
75% or more, and 55% a density of 100%. The number of types follows a similar pattern,
with unsolved instances with 24 types or more representing 82.7%, and instances with
32 types representing 51.7%. Few conclusions can be drawn with respect to the IP
model and IP decomposition since they did not solve many instances. Not surprising,
there is tendency relating the number of variables (e.g., size of type-set, number of
transports, etc.), and number of constraints (depends on the number of ports in the
schedule), with the hardness of the problem for all three models. Moreover, there is
a significant impact of the decomposition on the memory consumption of the solvers,
since only 3 instances of the IP decomposition run out of memory.
Now we analyse the instances solved by the MIP decomposition. Figure 6.3 depicts
the revenue of all solved instances, grouped by vessel and number of ports in their
schedule. As expected, the tendency observed in the single-port experiments indicating
the impact of the vessel size on revenue generated can also be seen in the multi-port
experiments. Moreover, it is obvious that the number of ports has a direct impact on
revenue. Grouping instances that are similar with respect to transport density and
number of types did not show any tendencies. The left plot in Figure 6.4 depicts the
impact of the number of ports in the schedule on the solution time of the instances.
Though with a few outliers, we can see that for each vessel there is a tendency for
the instances with the greatest number of ports to take the longest time to solve. All
instances with three and five ports solved within 600s, while instances with ten ports
go from 685s, up to 3,600s. Most of the instances have a gap between 0% and 0.1%,

122
 Chapter 6. Cargo Composition Problem

7e+07
3
5
7
6e+07 10

5e+07
Revenue(US$)

4e+07

3e+07

2e+07

1e+07

0
1 2 3 4 5 6 7 8 9 10
Vessel ID

Figure 6.3: Revenue of the instances solved by the MIP decomposition, grouped by number
of ports in their schedule.

a very low percentage, (See right plot in Figure 6.4). Few outliers go over 0.2%, all of
them being instances with ten ports in their schedule. As with the revenue analysis, the
instances were also grouped based on transport density and the size of the type-sets,
but no clear tendencies were observed. In conclusion, the MIP decomposition model
scales well for finding multi-port cargo compositions that optimize revenue. Moreover,
there is a stronger correlation between the hardness of the problem and the number of
ports in the schedule than with the vessel size (number of locations in this case), since
for all the vessels most of the hard instances have 10 ports, followed by the ones with
7 ports.

6.7 Conclusions
In this chapter, we have introduced the definition of the Cargo Composition Problem,
a new application of accurate stowage planning in the liner shipping domain. We have
also introduced the first mathematical model with variable displacement that assigns
containers to vessel slots, and that is solvable on real instances. To increase scalability,
we extended the master planning model presented in chapter 4 to cope with variable
displacement and use it to generate sub-optimal solutions for the CCP.
We tested our models in a benchmark suite of ten real container vessels, in a single
port and multi-port scenario. In the single port setup we showed the flexibility of the
models by optimizing TEU intake and intake of containers with specific features. The

123
Chapter 6. Cargo Composition Problem

5000 1.4
3 3
5 5
7 7
10 1.2 10
4000

Integrallity gap(%)
3000
0.8
Time(s)

0.6
2000

0.4

1000
0.2

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Vessel ID Vessel ID

(a) (b)
Figure 6.4: Instances solved by the MIP decomposition grouped by vessel and number of
ports in the schedule. (a) CPU time of the instances. (b) Integrality gap.

experiments showed that for intake optimization the MIP decomposition produced very
similar results to those of the IP model but only used a small fraction of computation
time. When optimizing intake of containers with specific features, however, we observed
small discrepancies in the solutions due to inaccuracies in the decomposition, though
they can be addressed by improving the capacity constraints and increasing the accuracy
of pre-computed values. Moreover, we optimized revenue and compared our results with
a model from the literature, showing the impact of accurate stowage modeling. Our
experiments also showed the strength of the approach, introduced to cope with variable
displacement when evaluating the linear relaxations of the models. In the multi-port
scenario, we have tested the impact of features such as number of ports in the schedule
and number of different container types. The results showed the advantage of the
decomposition with respect to the complete model in terms of memory consumption
and scalability. Our decomposition successfully scaled to optimize revenue for instances
up to 10 ports and solving 91.7% of all tested instances.
As future work, we plan to incorporate lashing forces to our models, a feature of
stowage models that limits the cargo compositions that can be stowed in sections on
deck of a vessel. Additionally, we plan to study models combining the CCP with other
important problems in the liner shipping industry, (e.g., network design and vessel
deployment).

124
 Chapter 7. Conclusions

Chapter 7

Conclusions

In this thesis we have studied the problem of stowing containers into container vessels
and its application to different operational and commercial decisions of liner shipping
companies. Specifically, we have formulated stowage optimization models with accept-
able trade-offs between accuracy and scalability to assist solving stowage dependant
planning problems in the liner shipping domain. Hence, the answer to the thesis ques-
tion is affirmative.
For stowage planning, we presented two approaches with different trade-offs between
accuracy and scalability for automatic generation of stowage plans. Both approaches
are based on a 2-phase heuristic decomposition that in the first phase (master planning),
assigns containers to vessel sections such that stability and stress force constraints are
satisfied and time at port is minimized; while in the second phase (slot planning),
containers assigned to vessel sections are stowed into slots, satisfying stacking patterns
and capacities, and minimizing time at port as well as robustness of the stowage plans.
The first approach models a representative problem that stows 20 ’ and 40 ’, reefer,
and normal containers in a container vessel with release, such that stability and stress
force constraints were within limits and all stacking constraints were satisfied. We
presented a mathematical model for master planning that accurately represents stability
and stress force constraints based on hydrostatics and includes ballast tanks, a feature
of stowage planning widely disregarded in previous academic work. We focused on
studying the impact of small displacement variations introduced by the ballast tanks
on the accuracy on stability and stress force constraints and the center of gravity. The
accuracy of the model was experimentally evaluated with industrial data and the results
showed that it can successfully use ballast water to alter the stability and stress forces
of a container vessel, allowing the possibility of applying stowage configurations that
are otherwise infeasible.
For slot planning, we introduced an accurate representative model to stow vessel
sections. We developed a CP model that, to our knowledge, is the fastest approach to
generate optimal slot plans. Our CP model produced optimal solutions for most of the
industrial instances where it was tested. Moreover, we successfully evaluated the CP
model as a component of a module for automatic generation of stowage plans, where it

125
Chapter 7. Conclusions

also found optimal solutions for most of the vessel sections in less than a 1s.
The second stowage planning approach was a heuristic that served as the optimiza-
tion component in a commercial decision support tool used to stow container vessels
interactively. For this heuristic, we reduce the model accuracy to cope with the need
for fast generation of stowage plans in an interactive environment. To compensate for
the reduced accuracy, our heuristic integrates the know-how of stowage coordinators,
represented as user preferences, to address stability and stress force constraints and
combinatorial objectives that minimize time at port. Our experiments showed that
this heuristic is able to produce plans fast enough to be used in an interactive decision
support tool, and that these plans are competitive with those produced by expert users.

For the cargo composition problem, we have introduced the first mathematical model
with variable displacement that assigns containers to vessel slots, and is solvable on real
instances. To increase scalability, we extended our accurate master planning model
with the capability of dealing with variable displacement. We tested our models in a
benchmark suit of real container vessels and performed analysis on intake and revenue
optimization in single port and multi-port scenarios. Our experiments showed that the
extended master planning model produced similar solutions to those produced by the
complete model using less CPU time and hardware resources. Moreover, we optimized
cargo revenue and showed the impact of accurate stowage modeling on this kind of
optimization. To our knowledge, this is the first application of stowage modeling to
another problem than stowage planning.

7.1 Outlook and Future Directions

With several planning problems in the liner shipping domain demanding a clear picture
of the viability of stowing cargo compositions in container vessels, the necessity for
developing accurate and scalable stowage optimization models is greater than ever.
The stowage optimization models presented in this dissertation are a first step towards
a widespread use of stowage modeling across liner shipping companies. As future work,
we suggest two directions. The first relates to the accuracy and scalability of the models.
We have, with our industrial collaborator, defined representative problems that consider
the most relevant features of container stowage, such that the solutions produced by
our models can be evaluated by expert users. To incorporate models based on these
representative problems into daily operations, however, they need to be extended with
additional features of the real problem, (e.g., line-of-sight and lashing forces). Thus,
an interesting line of future work is to efficiently formulate the new constraints added
to the representative problems and study their impact on performance and solution
quality of the models presented in this thesis.
With respect to scalability, an interesting topic to investigate is the use of mathe-
matical decompositions that can overcome the scalability problem of models assigning
individual containers to slots. Though it is unlikely that these models scale for practical

126
 Chapter 7. Conclusions

use, they can be used for finding optimal solutions for large instances and allow us to
evaluate the accuracy of heuristics and decompositions formulated to perform stowage
optimization. Additionally, given that the models introduced in this dissertation some-
times took long time to solve, we consider anytime algorithms for stowage optimization
an important research topic.
The second direction of future work is to study models combining accurate stowage
modeling with other important problems in the liner shipping industry ,(e.g., network
design (Brouer et al. (2013)), vessel deployment (Alvarez (2009)), fleet repositioning
(Tierney (2013)), etc.). Models to tackle these problems are built with oversimplified
stowage representations that assume that containers of any kind can be transported by
a vessel without affecting its maximum intake, time at port, or fuel consumption. This
assumption is far from reality and may reduce their applicability. Stowage models with
a trade-off between accuracy and scalability, adapted to each of these problems, can
improve the applicability of their solutions without significantly affecting CPU time.

127
Chapter 7. Conclusions

128
 BIBLIOGRAPHY

Bibliography

Alphaliner (2013, 03). Cellular fleet at 1st march 2013.

Alvarez, J. F. (2009). Joint routing and deployment of a fleet of container vessels.

Maritime Economics & Logistics 11 (2), 186–208.

Ambrosino, D., D. Anghinolfi, M. Paolucci, and A. Sciomachen (2010). An experimental

comparison of different heuristics for the master bay plan problem. In Proceedings of
the 9th Int. Symposium on Experimental Algorithms, pp. 314–325.

Ambrosino, D. and A. Sciomachen (1998). A constraint satisfaction approach for master

bay plans. Maritime Engineering and Ports 36, 175–184.

Ambrosino, D. and A. Sciomachen (2003). Impact of yard organization on the master

bay planning problem. Maritime Economics and Logistics (5), 285–300.

Ambrosino, D., A. Sciomachen, D. Anghinolfi, and M. Paolucci (2009). A new three-step

heuristic for the master bay plan problem. Maritime Economics and Logistics 11 (1),
98–120.

Ambrosino, D., A. Sciomachen, and E. Tanfani (2004). Stowing a containership: the

master bay plan problem. Transportation Research Part A: Policy and Practice 38 (2),
81–99.

Andersen, M. W. (2010). Service Network Design and Management in Liner Container

Shipping Applications. Ph. D. thesis.

Aslidis, A. H. (1984). Optimal container loading. Master’s thesis, Massachusetts Insti-

tute of Technology.

Autoship Systems Corporation (2013). Autoload.

http://cargomanagement.autoship.com.

Avriel, M., M. Penn, and N. Shpirer (2000). Container ship stowage problem: complex-
ity and connection to the coloring of circle graphs. Discrete Applied Mathematics 103,
271–279.

129
BIBLIOGRAPHY

Avriel, M., M. Penn, N. Shpirer, and S. Witteboon (1998). Stowage planning for
container ships to reduce the number of shifts. Annals of Operations Research 76,
55–71.

Barrass, B. and C. D. Derrett (2011). Ship stability for masters and mates. Butterworth-
Heinemann.

Botter, R. and M. Brinati (1992). Stowage container planning: A model for getting an
optimal solution. In Proceedings of the 7th Int. Conference on Computer Applications
in the Automation of Shipyard Operation and Ship Design, pp. 217–229.

Brouer, B. D., J. F. Alvarez, C. E. Plum, D. Pisinger, and M. M. Sigurd (2013). A base

integer programming model and benchmark suite for linear shipping network design.
Transportation Science.

Christiansen, M., K. Fagerholt, B. Nygreen, and D. Ronen (2007). Maritime trans-

portation. Handbooks in Operations Research and Management Science 14, 189–284.

Davidor, Y. and M. Avihail (1996). A method for determining a vessel stowage plan,
Patent Publication WO9735266.

Delgado, A. and R. M. Jensen (2013). Cargo composition analysis in container vessels.

To be submitted to European Journal of Operational Research.

Delgado, A., R. M. Jensen, and N. Guilbert (2012). A placement heuristic for a com-
mercial decision support system for container vessel stowage. In Informatica (CLEI),
2012 XXXVIII Conferencia Latinoamericana En, pp. 1–9. IEEE.

Delgado, A., R. M. Jensen, K. Janstrup, T. H. Rose, and K. H. Andersen (2010, Octo-

ber). A constraint programming model for fast optimal stowage of container vessel
bays. Technical Report TR-2010-133, IT University of Copenhagen.

Delgado, A., R. M. Jensen, K. Janstrup, T. H. Rose, and K. H. Andersen (2012).

A constraint programming model for fast optimal stowage of container vessel bays.
European Journal of Operational Research 220 (1), 251 – 261.

Delgado, A., R. M. Jensen, and C. Schulte (2009). Generating optimal stowage plans
for container vessel bays. In Proceedings of the 15th Int. Conference on Principles
and Practice of Constraint Programming (CP-09), Volume 5732 of LNCS Series, pp.
6–20.

Delgado, A., D. Pacino, and R. Jensen (2013). Container vessel stowage planning
with ballast tanks. In Proceedings of the 2nd INFORMS Transportation Science and
Logistics Society Workshop.

Drewry Maritime Research (2011). Container census 2011: Survey and forecast of
global container units.

130
 BIBLIOGRAPHY

Dubrovsky, O., G. Levitin, and M. Penn (2002). A genetic algorithm with a compact
solution encoding for the container ship stowage problem. Journal of Heuristics 8,
585–599.

Feng, C.-M. and C.-H. Chang (2008). Optimal slot allocation in intra-asia service for
liner shipping companies. Maritime Economics & Logistics 10 (3), 295–309.

Flor, M. (1998). Heuristic algorithms for solving the container ship stowage problem.
Master’s thesis, Technion, Haifa, Isreal.

Gecode Team (2006). Gecode: Generic constraint development environment. Available

from http://www.gecode.org.

Giemsch, P. and A. Jellinghaus (2004). Optimization models for the container ship
stowage problem. In Operations research proceedings 2003: selected papers of the
International Conference on Operations Research (OR 2003), Heidleberg, September
3-5, 2003, pp. 347. Springer Verlag.

Guilbert, N. and B. Paquin (2010). Container vessel stowage planning, Patent Publi-
cation US2010/0145501.

Gumus, M., P. Kaminsky, E. Tiemroth, and M. Ayik (2008). A multi-stage decomposi-

tion heuristic for the container stowage problem. In Proceedings of the Manufacturing
and Service Operations Management (MSOM) conference.

Hamburg, S. V. (2008). Storck Guide: Stowage & Segregation to IMDG Code Including
Amendment 34-08. Storck guide. Storck.

INTERSCHALT maritime systems AG (2013). Stowman.

http://www.interschalt.de/en/is-seacos-stowage-planning.htm.

Kang, J. and Y. Kim (2002). Stowage planning in maritime container transportation.

Journal of the Operations Research Society 53 (4), 415–426.

Kjeldsen, K. H. (2012). Liner shipping network design, routing and scheduling. Ph. D.
thesis.

Li, F., C. Tian, R. Cao, and W. Ding (2008). An integer linear programming for con-
tainer stowage problem. In M. Bubak, G. van Albada, J. Dongarra, and P. Sloot
(Eds.), Computational Science ICCS 2008, Volume 5101 of Lecture Notes in Com-
puter Science, pp. 853–862. Springer Berlin / Heidelberg.

Muller+Blanck Software GmbH (2013). Capstan3.

http://www.capstan3.com/capstan3.html.

Navis (2013). Navis PowerStow. http://www.navis.com/solutions/shipping.

131
BIBLIOGRAPHY

Nugroho, S. (2004). Case-based stowage planning for container ships. In International

Logistics Congress, Number December 2004, pp. 2–3.

Pacino, D. (June 2012). Fast Generation of Container Vessel Stowage Plans - Using
mixed integer programming for optimal master planning and constraint based local
search for slot planning. Ph. D. thesis, IT University of Copenhagen.

Pacino, D., A. Delgado, R. Jensen, and T. Bebbington (2012). An accurate model for
seaworthy stowage planning with ballast tanks. pp. 17–32. Springer. Proceedings of
the Third International Conference on Computational Logistics, (ICCL12).

Pacino, D., A. Delgado, R. M. Jensen, and T. Bebbington (2011). Fast generation

of near-optimal plans for eco-efficient stowage of large container vessels. In J. Bse,
H. Hu, C. Jahn, X. Shi, R. Stahlbock, and S. Vo (Eds.), Computational Logistics,
Volume 6971 of Lecture Notes in Computer Science, pp. 286–301. Springer Berlin /
Heidelberg.

Pacino, D. and R. M. Jensen (2012). Constraint-based local search for container stowage
slot planning. In Proceedings of the International Multi-Conference of Engineers and
Computer Scientists, pp. 1467–1472.

Pesant, G. (2004). A regular language membership constraint for finite sequences of

variables. In Proceeding of Principles and practice of constraint programming, Volume
3258 of Lecture Notes in Computer Science, pp. 482–495. Springer.

Sciomachen, A. and A. Tanfani (2003). The master bay plan problem: a solution
method based on its connection to the three-dimensional bin packing problem. IMA
Journal of Management Mathematics 14, 251–269.

Scott, D. and D. Chen (1978). A loading model for a container ship. Technical report,
Matson Navigation Company, Los Angeles.

Shaw, P. (2004). A constraint for bin packing. In Proceeding of Principles and practice
of constraint programming, Volume 3258 of Lecture Notes in Computer Science, pp.
648–662. Springer.

Smith, B. (2006). Modelling. In F. Rossi, P. van Beek, and T. Walsh (Eds.), Handbook
of Constraint Programming, Chapter 11. Elsevier.

Stopford, M. (2009). Maritime Economics (Third ed.). Routledge.

Tierney, K. B. (2013). Optimizing Liner Shipping Fleet Repositioning Plans. Ph. D.

thesis, IT University of Copenhagen.

Ting, S.-C. and G.-H. Tzeng (2004). An optimal containership slot allocation for liner
shipping revenue management. Maritime Policy & Management 31 (3), 199–211.

132
 BIBLIOGRAPHY

UNCTAD (2012). Review of maritime transport.

Van Hentenryck, P. and J. P. Carrillon (1988). Generality vs. specificity: an experience

with AI and OR techniques. In Proceedings of the National Conference on Artificial
Intelligence (AAAI), pp. 660–664. ACM press.

Webster, W. C. and P. Van Dyke (1970). Container loading. A container allocation

model: I - introduction background, II - strategy, conclusions. In Proceedings of
Computer-Aided Ship Design Engineering Summer Conference. University of Michi-
gan.

Wilson, I. and P. Roach (1999). Principles of combinatorial optimization applied to

container-ship stowage planning. Journal of Heuristics 5, 403–418.

Wilson, I. D. and P. Roach (2000). Container stowage planning: A methodology for gen-
erating computerised solutions. Journal of the Operational Research Society 51 (11),
248–255.

Yoke, M., H. Low, X. Xiao, F. Liu, S. Y. Huang, W. J. Hsu, and Z. Li (2009). An

automated stowage planning system for large containerships. In Proceedings of the
4th Virtual Int. Conference on Intelligent Production Machines and Systems.

Zhang, W., Y. Lin, and Z. Ji (2005). Model and algorithm for container ship stowage
planning based on bin-packing problem. Journal of Marine Science and Applica-
tion 4 (3).

Zurheide, S. and K. Fischer (2012). A revenue management slot allocation model for
liner shipping networks. Maritime Economics & Logistics 14 (3), 334–361.

133