Accepted Manuscript

Title: Solving the 3D Stowage Planning Problem integrated

with the Quay Crane Scheduling Problem by Representation
by Rules and Genetic Algorithm

Author: Anibal Tavares de Azevedo Luiz Leduino de Salles

Neto Antônio Augusto Chaves Antônio Carlos Moretti

PII: S1568-4946(18)30012-7
DOI: https://doi.org/doi:10.1016/j.asoc.2018.01.006
Reference: ASOC 4653

To appear in: Applied Soft Computing

Received date: 22-11-2015

Revised date: 23-12-2017
Accepted date: 9-1-2018

Please cite this article as: Anibal Tavares de Azevedo, Luiz Leduino de Salles Neto,
Antônio Augusto Chaves, Antônio Carlos Moretti, Solving the 3D Stowage Planning
Problem integrated with the Quay Crane Scheduling Problem by Representation by
Rules and Genetic Algorithm, <![CDATA[Applied Soft Computing Journal]]> (2018),
https://doi.org/10.1016/j.asoc.2018.01.006

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
*Graphical abstract (for review)

t
ip
cr
us
an
M
d
p te
ce
Ac

Page 1 of 60
*Highlights (for review)

HIGHLIGHTS

This paper showed a new approach to solve the 3D Stowage Planning for Container ship (3D
SP) problem integrated with the Scheduling of quay cranes (SQC) problem. It has three
modeling advantages:

• It shows how the 3D SP will affect other problems related to port operation, like the SQC
problem.

t
ip
• It provides better accuracy in estimating total time, which can save the charterer charges for

cr
the extra use of the vessel. In the instances studied, the solution from integrated approach
provided solutions with a 45.82% higher total time spent, on average, and prevented an

us
underestimation of necessary time for ship travel.

an
• In 40% of the instances, the Integrated 3D SP and SQP problem helped to avoid a misleading
analysis, where the adoption of good practices for 3D SP produce a worse total time to unload
and load the ship.
M
• The developed approach enables a series of analysis that partially explains the advantages on
ed

use bigger container ships and justifies a long term tendency of continuous increasing on
container ship size.
pt
ce
Ac

Page 2 of 60
*Manuscript
Click here to view linked References

Solving the 3D Stowage Planning Problem integrated

with the Quay Crane Scheduling Problem by

t
Representation by Rules and Genetic Algorithm

ip
Anibal Tavares de Azevedoa , Luiz Leduino de Salles Netob , Antônio

cr
Augusto Chavesb , Antônio Carlos Morettia
a
University of Campinas, R. Pedro Zaccaria, 1300 - Caixa Postal 1068 - CEP 13484-350

us
- Limeira - São Paulo, Brazil, [email protected], [email protected],
++55 19 3701-6662.
b
Federal University of São Paulo, Rua Talim, 330 - São José dos Campos - São Paulo
-CEP 12231-280, Brazil, [email protected], [email protected], +55 12

an
3309-9500.
M
Abstract
The operational efficiency of a port depends on proper container movement
planning, called “stowage planning”, especially because unloading and load-
d

ing container ships demands time, and this has a cost. Thus, the optimization
of operations through stages is important to avoid blockage activities. This
te

paper proposes a framework for solving the 3D Stowage Planning (3D SP)
problem for Container ships integrated with the Scheduling of Quay Cranes
(SQC) problem. 3D SP and SQC problems are interrelated and combina-
p

torial, justifying the applications of meta-heuristics like a genetic algorithm

ce

combined with Simulation and Representation by Rules. The robustness of

the developed approach is attested in problems with 30 ports, 1500 TEUs
ship or 15 ports and 22, 000 TEUs ship and two quay cranes. These studies
showed that the addition of the SQC problem leads to a 45.82% increase in
Ac

load/unload time for the 3D SP problem solution, on average. This could

help the charterer to avoid paying charges to the shipowner due to its an
extra unplanned use of the vessel. Additionally, the developed methodol-
ogy also helps to explain a long term phenomena of continuous increasing in
container ship capacity since 1950's.
Keywords: Stowage Planning, Quay Crane Scheduling, Genetic Algorithm,
Representation by Rules

Preprint submitted to Applied Soft Computing December 23, 2017

Page 3 of 60
1. Introduction
According to [1], over 60% of the world’s deep-sea general cargo is trans-

t
ported in containers, and the routes between some countries are containerized

ip
up to 90%. The improvement of the operational efficiency of container termi-
nals is essential to handle the increasing flow of containers that has occurred
over the last years. The problem of optimizing seaport operations should

cr
be defined and methods of solution proposed for such tasks. In this sense,
the classification made by [1, 2] is useful, since seaport container terminal

us
operations are divided into five main problems.

1. Berth allocation: the output of this problem is the scheduling of the

ships in each berth by considering the minimum safety distance between

an
two ships, given that two ships cannot share the same berth at the same
period of time.

2. Stowage planning: this problem consists of determining how to organize

M
the containers in a ship in order to minimize the number of movements
necessary to unload and load the container ship.

3. Crane Split: the container ship time for unloading and loading depends
d

on the scheduling of Quay cranes for each vessel section. One important
te

constraint is to not allow Quay cranes to pass each other, since their
movements are limited to a common rail.
p

4. Quayside transport: this problem defines which machines will be used

and their trajectory from ship to port yard to transport containers from
ce

ship to landside or vice-versa.

5. Land-side transport: This problem deals with unloading and loading

containers efficiently off from/onto trucks or trains to maximize the
Ac

flow through the port yard.

The division of port operations into five problems may be helpful to de-
velop mathematical models, but the final objective is to find a solution that
encompasses all operations. For example; from problem (2), it could be said
that the number of movements in Stowage Planning is just an estimation of
necessary time to unload or load the ship, because it depends on the avail-
ability of cranes and quayside transport. So, a more precise model should
couple Stowage Planning and Quay Crane Scheduling problems.

Page 4 of 60
 Recent literature shows a tendency to integrate the five aforementioned
problems. For example, some articles integrate berth allocation and quay
crane scheduling [3, 4, 5, 6, 7]; others integrate the allocation of berths and

t
yard operation planning [8]; and some integrate empty container allocation

ip
in the yard with vehicle routing [9].
This recent tendency is justified by the obvious fact that the optimiza-

cr
tion of operations for just one stage does not increase overall port efficiency,
because further and non optimized stages behave as blockage.
The integration of Berth allocation and Crane Split is also a problem as

us
mentioned by [10], who analyzed and classified 120 articles. In the following
commentary, they concluded: “Nevertheless, to obtain more reliable esti-
mates, BAP, QCAP, and QSCP must be solved jointly, which is referred to

an
as integrated seaside operations planning. A deep integration means to solve
a monolithic model where the interdependencies of the involved problem-
individual decisions are considered on the background of the merged set of
constraints. Solving a monolithic model can deliver the best overall solution
M
but is usually extremely difficult due to the huge complexity of the merged
problem.”
Indeed, the integration of port problems is very difficult due to the com-
d
putational complexity of the mathematical model related to each stage which
is NP-Hard. In fact, a port may be seen as a complex system as observed by
te

[11]: “Many complex systems such as manufacturing, supply chain, and con-
tainer terminals are too complex to be modeled analytically. Discrete event
simulation has been a useful tool for evaluating the performance of such sys-
p

tems. However, simulation can only evaluate a given design, not provide
more optimization functions. Therefore, the integration of simulation and
ce

optimization is needed.”
Stowage Planning studies have totally ignored quay crane scheduling or
have dealt superficially with it by making simple assumptions such as the
Ac

use of only one quay crane [12, 13, 14, 15, 16, 17, 18, 19]. In fact, none of
the Stowage Planning papers in the literature have considered the important
features of quay crane scheduling described in [6].
Our study employs a simulation to represent logical and physical con-
straints instead of using a monolithic mathematical model with binary vari-
ables and algebraic equations. This means the use of an alternative approach
called Simulation-Optimization. A very detailed review about it is given by
[20] and a taxonomy is given on [21]. The paper [22] gives a detailed practical
introduction to this approach. Simulation-Optimization is simply the com-

Page 5 of 60
bination of simulation, that describes the complicated logical operations of
one system, with optimization, that cares about how to choose the decision
variables values in order to minimize costs. According to the classification

t
proposed by [21], the approach developed in this paper could be classified as

ip
MH, i.e., an optimization algorithm that employs a simulation to evaluate
each solution. Additionally, our study applies a new innovative approach on

cr
representing decision variables that substantially reduces the search space.
The decision variable in the developed Simulation-Optimization approach
describes the action that could be made in a specific activity in generic terms.

us
For example, instead of creating binary variables and corresponding equa-
tions that control the position of each container in a ship, it is better to
create a generic procedure of how to unload or load containers into a ship

an
and couple the procedure with a simulation. This manner of representing
decisions has been successfully applied in 3D SP and is described in [17],
which we will extend to SQC in this paper.
The paper contributions are:
M
• Solve for the first time the integrated problem of Stowage planning
with a detailed Quay Crane Scheduling which actual monolithic math-
ematical models with binary variables and algebraic equations are not
d

capable to solve or even model;

te

• This was done by extending and employing a new simulation - opti-

mization methodology which was successful to obtain quality solutions
for the Stowage planning problem;
p

• The solution for integrated problem avoids underestimation on time

ce

necessary to do unloading and loading operations which could leads to

solutions for the Stowage Planning Problem with a higher operation
costs;
Ac

• Finally, the developed approach enables a series of analysis that par-

tially explains the advantages on use bigger container ships and justifies
a long term tendency of continuous increasing on container ship size.
This paper is organized as follows. Section 2 presents the model features
for problems 3D SP and SQC. Section 3 explains the representation by rules
and how it is combined with a Genetic Algorithm. Sections 4, 5, and 5
presents and discusses the computational results. Finally, Section 7 presents
the conclusions and Section 8 possible future work.

Page 6 of 60
2. A Definition of the Problem
2.1. The Stowage Planning Model

t
A solution for 3D SP should produce a stowage plan that is strongly

ip
related to the cellular structure of container ships, as shown in Figure 1.

cr
us
an
M
Fig. 1. The cellular structure of a container ship: top view and a cross-sectional view of
a bay.
d

This structure means containers may only be reached by removing any

containers stacked on top of them in a column. There are, thus, two unloading
te

cases:

• Containers to be unloaded at a given port are at the top of the stack.

p

• Containers to be unloaded at a given port are blocked by one or more

ce

containers that are to remain aboard the container ship. These must be
unloaded and reloaded after all containers in the column for that port
have been unloaded. This movement of unloading and loading blocking
Ac

containers is called re-handling and must be minimized to improve port

efficiency.

The more such blockage occurs, the longer the ship must stay in the port.
So, Stowage Planning (SP) is the key to minimizing re-handling movements.
In [23] it was proved that the 2D stowage plan is NP-Hard, which justified
the development of a series of heuristics and meta-heuristics to obtain good
solutions for this problem, as may be seen in approaches from the literature
[24, 12, 16, 19, 15, 14, 13, 25]. The container’s position in the stack affects

Page 7 of 60
ship and stack stability. In the literature, there are two different manners to
deal with ship stability in the mathematical model: as an objective function
[26, 14], or as a constraint [27, 16, 15, 18, 25]. Stability concerns are related

t
to ship features, like metacentric height and trim [14], or the limits of height

ip
and weight in a stack [16, 25], and vertical equilibrium [28]. A recent review
provided a classification of previous articles [18].

cr
The mathematical model assumptions adopted in [17] has been adopted
here and the following assumptions have been made for the sake of simplicity,
without compromising the solution’s general application.

us
(a) The container ship has a rectangular format and can be represented by a
matrix with rows (r = 1, 2, · · · , R), columns (c = 1, 2, · · · , C) and bays

an
(d = 1, 2, · · · , D) with maximum capacity of R × C × D containers. An
irregular format may be achieved by simply adding constraints which
represent imaginary containers that occupy the same spaces during the
whole voyage [19].
M
(b) All containers have the same size and weight (equal to one).

(c) The ship starts to be loaded in Port 1, where it arrives empty;

d

(d) The ship visits ports 2, 3, · · · , N such that the container ship will be
te

empty at the last Port, because the ship performs a circular route where
Port N, in fact, represents Port 1.
p

(e) For each Port i, i = 1, 2, · · · , N, such that the container ship can be
loaded with containers whose destination are ports i + 1, · · · , N.
ce

(f) The container ship can always carry all the containers available in each
port and this will never exceed its capacity.
Ac

The Appendix D presents the corresponding mathematical model formu-

lation. The Appendix G presents a small numerical example to illustrate
some model features related with number of movements and stability.

2.2. Quay Crane Scheduling Model

Quay Crane Scheduling is directly related to stowage planning, detailed
in Figure 2.

Page 8 of 60
 t
ip
cr
us
Fig. 2. Quay Crane Scheduling Problem.

an
The Scheduling of Quay Cranes (SQC) problem should consider con-
straints related to equipment characteristics. For example, quay crane as-
signment through time must not allow the crossing of cranes, since they
M
are mounted on the same rail. One formulation of this problem, with non-
crossing constraints, was shown as NP-Hard in paper [29]. This motivated
the development of a genetic algorithm by the same authors. A lot of au-
d

thors in the literature adopted the same procedure of modeling the quay
crane problem to solve the models by meta-heuristics [2, 30, 31, 32, 33].
te

The SQC problem resembles that of parallel machines, but it has ad-
ditional features that make it more complex as observed by [34]. In this
study, the following assumptions were made in terms of objective function
p

and constraints.
ce

(a) Instead of dividing the container ship subject into arbitrary sections as
done in [31], each container ship bay was considered. This enabled the
direct use of the stowage planning solution as information of the total
Ac

workload for each bay in the container ship.

(b) It was assumed that the total workload was the total number of contain-
ers and each container stored in position (r, c, d) in Port i would take
ψ1i (r, c, d) units of time to be unloaded and ψ2i (r, c, d) to be loaded into
the container ship.

(c) Each bay could only hold one crane at a time.

Page 9 of 60
(d) Once one quay crane started, the service of loading or unloading a bay
would not stop until it was finished (this is no preemption assumption).

t
(e) When a crane changed its operation to another bay, it would take a

ip
constant time to reach its destination which was equal to move three
containers. Thus, it was assumed that this repositioning movement de-
manded three units of time.

cr
(f) All cranes had the same constant service operation rate of one movement
per time unit.

us
(g) One crane could not cross another, since they were mounted on the same
rail.

an
(h) A minimal distance between cranes had to be observed. This meant that
sometimes one crane would not be allowed to move or start its service
until a second crane had finished its work and moved to a new position
M
that ensured a minimal distance of one bay between two cranes.

The last two assumptions are what make the problem more complex than
the parallel machines problem [3].
d

The Appendix E presents the corresponding mathematical model formu-

lation. The Appendix H presents a small numerical example to illustrate
te

some model features related with the impact of non-crossing constraint im-
pact on quay cranes movement and the necessary total time to perform tasks
p

along a container ship.

ce

2.3. Integrating Stowage Planning with Quay Crane Scheduling

The problems modeled in subsections 2.1 and 2.2 are not independent.
Once a container ship arrives at or leaves a port it is necessary to perform
Ac

unloading and loading container movements according to previous Stowage

Plan. The 3D SP model tries to produce container ship arrangement that
minimizes the number of movements. As shown in Figure 3, this is a hard
problem since the arrangement in one port could greatly affect the arrange-
ment in future ports. The element i, j in Figure 3a determines the quantity
of containers that should be transported from origin Port i (row Oi ) to others
j destination ports (Dj s columns on row Oi ). Figure 3b shows the container
ship arrangement in each port after unloading (U) and loading (L) opera-
tions.

Page 10 of 60
 t
ip
cr
D1 D2 D3 D4

us
O1 0 3 3 2
O2 0 0 2 2
O3 0 0 0 2 (b) Container ship arrangement after unloading
and loading operations.

an
O4 0 0 0 0
(a) Information about the num-
ber of containers that must be
transported through ports.
M
Fig. 3. One solution for the Stowage Planning Problem (not optimal).
d

The real problem, however, is more complex since the number of move-
ments is just an estimation of the time necessary to unload or load ship
te

using Quay Cranes. This is done by coupling the 3D SP and SQC prob-
lems through transforming the value of the variable that controls container
position into total workload for unloading and loading operations. In mathe-
p

matical terms, it leads to the introduction of non-linear constraints that can

ce

be linearized with proper techniques. This, however, makes the search for an
optimal solution a challenge for exact methods. Figure 4 shows the physical
meaning of coupling both models.
In Figure 4, once a container ship arrangement is given, it is possible to
Ac

translate the total number of movements in each bay into total workload per
bay. For example, suppose one solution suggests loading the container ship
with containers 3, 3, 2 in Bay 1, containers 3, 4, 4 in Bay 2, and containers
2, 2 in Bay 3 as shown in Figure 4. This information comes from the first
model which is responsible for planning the best arrangement considering
for loading and unloading through all ports. This first phase could be seen
as a planning phase (P). In a second phase it is necessary to put in details
about the necessary port operations for such planning. This means using

Page 11 of 60
 t
ip
cr
us
an
Fig. 4. Coupling Stowage Planning and Quay Cranes Operation problems.
M
the information of how many containers there are in each bay, from the first
phase, to compute the total workload per bay. Once this is done, the second
model will try to determine the best policy of quay crane operation without
d
making any prohibited movement. This phase may be seen as an operation
phase (O) and will use the following values per bay: 3, 3, 2.
te

The concepts described in Figure 4 can be applied to the planning model

given in Figure 3b to produce Figure 5, that illustrates how planning (P)
and operation (O) phases are interrelated.
p

In Figure 5, the estimation of movements to unload and load in the

stowage planning phase (P) will be replaced by total workload for quay crane
ce

operation simulation (O). For such simulation, the workload for each bay in
each port should be computed from the container ship arrangement, en-
abling a more accurate estimation of time to perform a proposed stowage
Ac

plan. This means, however, that a 3D SP problem with 30 ports must solve
SQC 60 times.
The Appendix F presents the corresponding mathematical model formu-
lation. The Appendix I presents a small numerical example to illustrate the
problems in trying to find optimal solution through the integrated model.

10

Page 12 of 60
 t
ip
cr
us
an
Fig. 5. The interrelation of Planning (P) and Operation (O) for the problem presented in
3b.
M
3. Solution Technique
This problem requires more than the integration of simulation and opti-
d

mization methods in order to deal with decision-making in complex systems,

such as ports. For such a purpose, the Zadeh’s incompatibility Principle is
te

very useful [35]: “Stated informally, the essence of the principle is that as the
complexity of a system increases, our ability to make precise and yet signif-
p

icant statements about its behavior diminishes until a threshold is reached

beyond which precision and significance (or relevance) become almost mutu-
ce

ally exclusive characteristics.”

The approach adopted presents generic procedures, as had been success-
fully applied in 3D SP (detailed in [17]). These procedures will be extended
to SQC and 3D SP integrated with SQC as detailed in subsections 3.1 and
Ac

3.2, respectively.

3.1. Representation by rules for Quay Crane Scheduling

After determining the container ship arrangement by solving 3D SP,
through the use of unloading and loading rules detailed in Appendix A,
the next step is to determine how much time is necessary to perform the
operation. For such a task, a new second set of four rules related to quay
crane scheduling are considered. To illustrate this second set of rules, Figure

11

Page 13 of 60
6 shows one of the rules that was employed in this paper. The time to move
cranes was not counted in Figure 6 and the movement of one container was
set at one unit of time.

t
ip
cr
us
(a) Beginning (b) Distance

an
M
(c) Synchronism (d) Distance
d
p te

(e) End Crane 1 (f) Total Time

Fig. 6. A detailed description of operation via Rule 1 for 2 quay cranes.

ce

The first rule for quay crane operations assigns one crane to the container
ship bow and another one to the middle of the ship as shown in Figure 6a. All
Ac

rules must lead to coordinated crane operations which respect the minimum
security distance as shown in 6d. Sometimes this constraint implies in an
increment in total time necessary to perform unloading or loading operations
as shown in Figures 6b and 6c. Figures 6e and 6f show that the total time to
process all the workload is the maximum time between crane one and crane
two to perform all operations.
The second rule for quay crane operations is similar to the first rule,
except that it assigns the first quay crane to the container ship stern and the
second to the middle of the ship. The difference between the crane operations

12

Page 14 of 60
for Rules 1 and 2 can be seen by comparing Figures 6a-6f with Figures 7a-7d,
respectively.
The first and second rules always allocate quay cranes at the same begin-

t
ning position without observing the total container ship workload. The third

ip
and fourth rules differ from the first and second ones by doing the initial allo-
cation of the second quay crane according to an equation that considers the

cr
“center of mass” of the total workload for the container ship. This difference
is illustrated in Figures 7e-7f and 7g-7h.
Although, all four rules for quay cranes used in this study were developed

us
for the operation of two quay cranes, the approach could be generalized to
any number of quay cranes as detailed in Appendix B.

3.2. Combining Stowage Planning rules and Quay Crane Scheduling rules

an
The efficiency of combining rules for loading and unloading container
ships and for quay crane operation can be measured through a two-stage
simulation. The planning stage (P) consists of solving 3D SP via the appli-
M
cation of a first set of rules that produces container ship arrangement and
two values: the total number of movements, and the instability measure as
described in [17]. Then, the next stage, the operation phase (O), converts
d
the container ship arrangement into workload for each container ship bay
and then in total time for quay cranes to perform all operations according
te

to a second set of rules: quay crane scheduling rules. Finally, the shipś total
mooring time depends on the selection of rules to be applied in each stage
from a set of 64 possible combinations. Table 1 describes all possible combi-
p

nations. The symbol # can be 1, 2, 3, or 4. Each rule number represents the

combination of three decisions among the two unloading (UR), eight loading
ce

(LR) and four quay crane (QC) rules. This means a complete solution should
determine which rule will be applied in each port from a set of 64 possible
combinations.
Ac

3.3. The Genetic Algorithm

The genetic algorithm, parameters, and the same strategies of [17] were
adopted in this work with changes made with the respect to the integration
of 3D SP and SQC, as described in Figure 4. A detailed description of these
parameters and its values is given on Appendix J. Additionally, the strategy
of using simulation instead of binary variables with algebraic equations was
adopted and modified to solve SQC. The simulation, as shown in Algorithm
1, is the evaluation scheme used for each genetic algorithm individual.

13

Page 15 of 60
 t
ip
cr
Algorithm 1 The evaluation algorithm encloses functions that emulate
the simulation of 3D SP and SQC.
1: procedure Evaluate

us
2: p ← 1, tmov ← 0, instab ← 0.0
3: Initialize(B, T )
4: while p < N do

an
5: [rc, rd, rq] = extractRules(s(p)) ⊲ Translation using Table 1.
6:
7: if (p > 1) then
[B, T ] = unloading(rd, B, p); ⊲ Planning Phase(P)
M
8:
9: instab ← instab+ calcDXDZ(B);
10: tmov ← tmov + TotalTime(rq, B); ⊲ Operation Phase(O)
11: B ← B2;
d

12: end if
13:
te

14: if (p < N − 1) then

15: [B, T ] = loading(rc, B, T, p); ⊲ Planning Phase(P)
instab ← instab+ calcDXDZ(B)
p

16:
17: tmov ← tmov + TotalTime(rq, B); ⊲ Operation Phase(O)
ce

18: B ← B2;
19: end if
20: p ← p + 1;
21: end while
Ac

22: return α × tmov + β × instab;

23: end procedure

14

Page 16 of 60
 t
ip
(a) Beginning of quay crane - Rule 1 (b) Total Time

cr
us
an
(c) Beginning of quay crane - Rule 2 (d) Total Time
M
d

(e) Beginning of quay crane - Rule (f) Total Time

3
p te
ce

(g) Beginning of quay crane - Rule (h) Total Time

Ac

Fig. 7. A detailed description of operation via rules 1, 2, 3, and 4 for 2 quay cranes.

The symbols used in Algorithm 1 have the following meaning: tmov is

a variable that represents the total time to perform quay crane operations
(unloading or loading) according to the rule described in function name rq;
and container distribution in ship spaces is represented by B matrix. The
instab variable measures the instability of the container ship arrangement

15

Page 17 of 60
Table 1. All rules produced by the combination of loading, unloading, and quay cranes
movement rules.

t
Quay Rule QC#

ip
Loading Rule LR1 LR2 LR3 LR4
Unloading Rule UR1 UR2 UR1 UR2 UR1 UR2 UR1 UR2

cr
Rule 1 2 3 4 5 6 7 8
Quay Rule QC#
Loading Rule LR5 LR6 LR7 LR8

us
Unloading Rule UR1 UR2 UR1 UR2 UR1 UR2 UR1 UR2
Rule 9 10 11 12 13 14 15 16

an
after unloading and loading operations. The unloading and loading functions
(using rd and rc rules, respectively) modify the container ship arrangement
M
after unloading and loading operations, respectively. All eight loading rules
and two unloading rules are detailed in Appendix A. The function calcDXDZ
computes the instability measure as a distance between geometric center and
the center of mass of the ship in accordance with the computation details
d

given by the small example from [17]. The function TotalTime computes the
total time necessary for a quay crane to perform an operation of unloading
te

or loading the ship, considering the minimal distance between quay cranes
as one bay as recommended by [36, 37], employing one rule rq of the four
p

quay crane rules (see Appendix B for quay crane Rule 1 details). Finally,
the symbols α and β are real numbers in interval [0, 1] and represent the
ce

importance given to minimize the total time to unload/load ship or to the

container ship arrangement instability measure, respectively.
Each genetic algorithm individual is a vector of integers for which each
Ac

element is an integer number whose range varies from 1 to 64. Each vec-
tor element represents which rule will be applied in each port. Algorithm
1 is the evaluation of each individual. The genetic operators adopted are:
OX crossover operator, and a mutation operator that modifies a certain per-
centage of bits from all uniform randomly chosen individuals for a given
population. The operators and parameters adopted are detailed in [17].

16

Page 18 of 60
4. Computational Results
The same set of instances used in [17] had been employed here in order

t
to verify the performance of the developed framework. The ship dimensions

ip
adopted for the instances presented in this article are D = 5 (number of
bays), R = 6, C = 50. All instances are available at the following site:
https://drive.google.com/open?id=0B4zUGKjaO9uEVndXZWsteG81S1k.

cr
In Tables 2 and 3, the column index I corresponds to instance number;
N corresponds to how many ports the container ship has to pass through;

us
the column index M refers to the type of transportation matrix (1 - Mixed,
2 - Long, 3- Short; there are more details on Appendix C); F O1 is the
total time to perform all crane movements (see subsection 2.2 for constraints
and parameter details); F O2 is the total measure of instability; T (s) is the

an
computational time spent in seconds to obtain the solution; F O1 − SP is an
estimation of the total time to perform all crane movements by considering
only Stowage Planning constraints; one container movement takes one unit
M
of time and time is simply divided by the number of cranes, which in this case
is two. This estimation merely illustrates what happens with the solution
from an optimization model that basically ignores, or that makes a simple
assumption about quay crane operation. The value F O1 − SP is presented
d

just for cases where F O1 is minimized. The results presented in Table 2 and
3 were obtained with a program created in a Matlab 7.0, a machine with a
te

1.66 GHz Core Duo Intel Processor, RAM memory of 2 GB, and Windows
Vista Operational System with Service Pack 2. The genetic algorithm (more
p

details about parameters on Appendix J) was executed 5 times to illustrate

the potential of the developed approach.
ce

Tables 2 and 3 present the worst and best solutions found by merely
minimizing the total workload time and by merely minimizing the instability
measure, respectively.
Ac

The results presented in Tables 2 and 3 show that the Genetic Algorithm
combined with the Representation by Rules approach was able to provide
solutions in less than nine minutes for instances with 30 ports and a container
ship with 5 bays, 6 rows, and 50 columns. In this case, each solution had
40, 545, 000 binary variables only in the planning phase.
The difference between F O1 and F O1 − SP is an important issue since,
for all instances, F O1 − SP underestimated the necessary time to unload
and load the ship through all ports. F O1 was at least 18.36% higher than
F O1 − SP for Instance 3, and at most 84.92% higher than F O1 − SP for

17

Page 19 of 60
 Table 2. Results obtained for GA when minimizing only total time.

I N M FO1 FO2 T(s) FO1-SP

t
Worst Best Worst Best Worst Best Worst Best

ip
1 10 1 5468 5359 776.11 621.87 222.95 206.00 3779 3896
2 10 2 3565 3448 650.49 715.86 186.79 199.73 2417 2131
3 10 3 10579 10397 418.42 554.81 218.32 226.01 8938 8555

cr
4 15 1 7841 7460 415.64 436.76 347.34 429.65 5753 5301
5 15 2 4788 4588 658.87 642.27 300.37 298.93 2980 2694

us
6 15 3 16095 15729 640.24 372.60 429.42 336.57 12680 12870
7 20 1 8420 8050 337.20 334.08 494.63 451.82 5513 5787
8 20 2 5437 5066 473.00 561.12 414.09 395.05 2945 2897
9 20 3 20510 20288 574.33 665.96 462.93 427.99 16777 16441

an
10 25 1 9403 9110 256.66 274.98 658.27 718.28 5999 6145
11 25 2 4938 4679 596.59 455.26 607.69 571.34 3069 3055
12 25 3 28060 27806 329.34 349.80 636.99 631.45 22604 22329
M
13 30 1 9410 9126 287.76 279.09 621.03 667.37 6479 6355
14 30 2 4918 4709 248.77 225.97 559.20 656.13 2687 2875
15 30 3 34399 33819 423.69 281.54 732.75 707.35 27911 28137
d

instance 8, with a mean deviation of 45.82%. This means that Stowage

te

planning models without careful and detailed features from the Quay Cranes
Scheduling problem will produce inaccurate unloading and loading times that
could lead the charterer to pay demurrage, i.e., which are charges that the
p

charterer pays to the shipowner for the extra use of the vessel.
The Table 4 shows the possible savings to be obtained by computing
ce

the difference of total time for a model with proper quay crane modeling
(F O1 − Best) and without it (F O1 − SP − Best) from Table 2. Each
container unloaded/loaded after the time estimated by the model without
Ac

quay crane data is multiplied by US$ 110.00. This value is a mean since the
cost varies according to container type and how long the container stays in
port. More details are available at [38].
Table 4 shows the savings could be anywhere from US$ 144, 870.00 (in-
stance 2) to US$ 625, 020.00 (instance 15) per container ship trip.
Worse than the money wasted, is the misleading analysis that is carried
out by using models without integrating 3D SP and SQC. For example, in
Instance 1 in Table 2, the best solution had a value of 5, 359 and the worst

18

Page 20 of 60
Table 3. Results obtained for GA when minimizing only the container ship arrangement

t
instability measure.

ip
I N M FO1 FO2 T(s)
Worst Best Worst Best Worst Best

cr
1 10 1 7384 8258 296.33 296.00 205.92 221.99
2 10 2 4988 6856 297.10 296.67 259.74 191.19

us
3 10 3 12588 12596 295.71 294.37 219.40 257.75
4 15 1 14175 14207 202.61 202.10 476.96 368.63
5 15 2 8636 6633 207.63 206.49 283.00 293.62
6 15 3 16654 17341 210.58 208.92 329.30 346.73

an
7 20 1 13920 13533 155.60 154.75 432.06 552.70
8 20 2 10609 9084 158.82 157.56 434.95 427.61
9 20 3 21224 21891 509.77 487.35 539.97 405.29
M
10 25 1 14601 14666 127.57 126.54 691.01 555.37
11 25 2 8590 9383 209.86 206.82 554.53 471.54
12 25 3 30941 28680 162.92 154.31 576.99 611.04
13 30 1 15967 15153 109.12 107.51 628.86 623.65
d

14 30 2 7992 8627 107.34 107.24 551.69 547.36

te

15 30 3 35190 36805 180.72 163.12 700.12 696.65

p
ce

Table 4. Savings obtained minimizing total time considering quay crane operations (F O1)
instead of (F O1 − SP ).
Ac

Instance Saving(US$) Instance Saving(US$) Instance Saving(US$)

1 160,930 6 314,490 11 178,640
2 144,870 7 248,930 12 602,470
3 202,620 8 238,590 13 304,810
4 237,490 9 423,170 14 201,740
5 208,340 10 326,150 15 625,020

19

Page 21 of 60
had value of 5, 468 according to F O1. But, according to F O1 − SP the best
solution (found observing F O1) was 3, 896 and the worst solution was 3, 779.
So, F O1 − SP does not recommend the same solution that F O1 does. In

t
other words, it leads to a faulty loading and unloading sequence. This is not

ip
an isolated case. It happened in 40% of the instances: instances 6, 7, 10, 14,
and 15 in Table 2. The next section explains in detail how these differences

cr
are related to SQC features.
Results show that optimization methodology must consider the trade-
off between minimization of ship stay time and the stability measure in the

us
determination of a proper arrangement of containers in a ship. In one option,
the decision maker has to decide which combination of weights produces
solutions that constitute a better trade-off in terms of stability and ship stay

an
time. The weights represent the decision makers preferences in each instance.
However, this approach may suffer with the effect of not reaching certain
points on Pareto front [39]. Another option is to simultaneously consider
stowage planning and stability measures, using the Pareto curve as detailed
M
in [40].

5. A Detailed Analysis of Results

d

The purpose of Figure 8 is to illustrate how the container ship arrange-

te

ment is connected to quay crane scheduling in the first four ports for the best
solution found in Instance 1 from Table 2. Figure 9 shows the corresponding
quay crane scheduling for arrangements in Figure 8.
p

Figure 8a shows the container ship arrangement after departing from Port
1 and Figure 9a shows its consequent quay crane scheduling for unloading
ce

service in Port 2. All other subfigures have the same relationship. In the first
ports, the containers are organized so that very few bays require unloading
service. This happens because the objective function minimizes the total
Ac

time spent, but completely ignores any stability measurement.

Figure 9a better presents the reason why F O1−SP is an underestimation
of unloading and loading total time of F O1. The ship has a total time of
112 units of unloading service time, but this could not be shared by the two
quay cranes because the containers were not adequately distributed among
the bays.
Another possible situation that also leads to underestimation of unloading
and loading total time of F O1 is presented in Figure 9b. In this case, the
two cranes could operate in parallel, but it does not imply in the division of

20

Page 22 of 60
 t
ip
cr
(a) Port 1. (b) Port 2.

us
an
M
(c) Port 3. (d) Port 4.

Fig. 8. Detailed Container ship arrangement after loading operations in the first four ports
for the best solution found with the total time minimization objective function.
d
te

total time by two. Quay crane 1 will finish the bay 1 at time 83, but cannot
move to Bay 2 until quay crane 2 finishes its job at time 121.00. So, the ship
unloading operation, instead of taking 357 = 178.5 units of time, will need
p

2
121 + 153 = 274 units of time, meaning F O1 will be 53.93% higher than
ce

F O1 − SP .
Figures 10 and 11 also illustrate the best solution found for instance 1,
but considering a completely different parameter setting, α = 0 and β = 1.
This setting tries to minimize only the instability measure for the container
Ac

ship arrangement. The values of number of units of time, and the instability
measure are found in Table 3.
Figure 10a shows the container ship arrangement after departing from
Port 1. This container ship arrangement will produce a corresponding quay
crane schedule to unload containers at Port 2 as shown in Figure 11a. All
other subfigures of Figures 10 and 11 have the same relationship. Figure
10 also represents a more stable container ship arrangement than the one
presented in Figure 8. The more stable arrangement of Figure 10 was ob-

21

Page 23 of 60
 t
ip
cr
(a) Port 1. (b) Port 2.

us
(c) Port 3.
an (d) Port 4.
M
Fig. 9. Detailed quay crane scheduling before unloading operations in the first four ports
for the best solution found with the total time minimization objective function.
d

tained by ignoring the corresponding increment in time necessary to perform

unloading and loading movements. This resulted in an extremely time con-
te

suming quay crane operation as shown in Figure 11. This is most evident,
for example, in Port 5 where it may be seen by comparing Figures 9d and
11d. Also, the best solution for instance 1 when instability is minimized,
p

as shown in Table 3, has a 54.09% higher value in total time than the best
solution found when total time was minimized as shown in Table 2.
ce

Another set of results reinforces the fact that minimization of instability

does not imply in the minimization of number of movements. For example,
in Instance 1 from Table 3, the worst solution in terms of FO2, had a sig-
Ac

nificantly lower value of FO1 (7, 384) than the best solution (8, 258). This
happened for 66% of the instances: Instances 1, 2, 3, 4, 6, 9, 10, 11, 14, and
16 from Table 3.
The most important conclusion from the results is that by using represen-
tation by rules, it was possible to quickly build the Stowage Plan and study
its impact on port quay crane operation.

22

Page 24 of 60
 t
ip
cr
(a) Port 1. (b) Port 2.

us
an
M
(c) Port 3. (d) Port 4.

Fig. 10. Detailed Container ship arrangement in the first four ports for the best solution
found with the instability minimization objective function.
d
te

6. Explaining an Economical Phenomena

During the last years the container ship experienced an increase in its size
p

[41]. The Figure 12 illustrates the evolution of ships dimensions through years
until 2013. The early container ship's at 1950's had a maximum capacity of
ce

800 TEUs. During 1980's the maximum capacity had a great improvement
from 3, 000 TEUs to 5, 000 TEUs. But, the 2000's had the most impressive
evolution since in thirteen years (from 2000 to 2013) the capacity increased
Ac

from 6, 000 to 18, 000 TEUs.

Nowadays exists ships with a capacity around 22, 000 TEU capacity, and
there are plans now to produce container ships with a 27, 000 to 30, 000 TEU
capacity [43]. The reason why ship's size is still increasing is because it
allows the integration of routes reducing the number of necessary ships to
carry goods [41, 43].
To verify the impact of continuous increasing of container ship's size in
computational effort to find a quality solution and the necessary time to per-
form unloading and loading operations, a completely new set of instances had

23

Page 25 of 60
 t
ip
cr
(a) Port 1. (b) Port 2.

us
(c) Port 3.
an (d) Port 4.
M
Fig. 11. Detailed quay crane scheduling in the first four ports for the best solution found
with the instability minimization objective function.
d
p te
ce
Ac

Fig. 12. Container ship's size evolution from 1956 until 2013. Source: [42].

been created and employed. This set is based on a hypothetical integration

of three existing routes, as presented on Figure 13, which could be served by
only one large container ship with a high carrier capacity [44].
The ship dimensions adopted for the instances presented in this article are

24

Page 26 of 60
 t
ip
cr
us
Fig. 13. Container ship routes through southeast Asia countries. Source: [44].

an
D = 30, 33, 36, 39, 42, 45, 48, 51, 54, 57, 60 (number of bays), R = 6, C = 25,
N = 5, 10, 15 (number of ports). All instances are available at the following
M
site:
https://drive.google.com/open?id=1attXqQZewMWr545-HxFZOAk5gSKwI7M-.
This set enables an analysis of what happens in terms of the necessary
time to unload and load a container ship due to an increment in the container
d

ship capacity and the number of total ports.

te

In this sense, each subplot presented in Figure 14 shows what happens

for a specific container ship size when the operation planning consider 5, 10
and 15 ports. All results considered that the transportation matrix is type 1
p

(mixed).
For example, the first subplot in Figure 14 was generated considering a
ce

container ship with a total capacity of 11, 250 TEUs (“Dimension 11, 250”).
Horizontal axis is about number of ports, and vertical axis points the cor-
responding time to perform unloading and loading operations. The subplot
Ac

shows an almost linear relation between an increase in number of ports and

an increase in necessary time to perform operations. Although, for some
container ships size there is not a linear increasing law. This happens, for
example, with a ship with a total capacity of 13, 500, 16, 875, 18, 000, 19, 125,
20, 250, 21, 375, and 22, 500 TEUs.
Each subplot presented in Figure 15 shows what happens for a specific
container ship size when the operation planning consider 5, 10 and 15 ports.
All results considered that the transportation matrix is type 2 (long).
Some subplots in Figure 15 shows a different behavior when compared

25

Page 27 of 60
 × 104 Dimension11250 × 104 Dimension12375 × 104 Dimension13500 × 104 Dimension14625
6 6 5 8

4 6
4 4
Time

Time

Time

Time
3 4

t
2 2
2 2

ip
0 0 1 0
05 10 15 05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports Number of ports
× 104 Dimension15750 × 104 Dimension16875 × 104 Dimension18000 × 104 Dimension19125
8 8 8 8

6 6

cr
6 6
Time

Time

Time

Time
4 4
4 4
2 2

2 0 2 0
05 10 15 05 10 15 05 10 15 05 10 15

us
Number of ports Number of ports Number of ports Number of ports
× 10 4 Dimension20250 × 10 4 Dimension21375 × 10 4 Dimension22500
10 8 8

8
6 6
Time

Time

Time
6
4 4
4

an
2 2 2
05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports

Fig. 14. Improvement of Container ship size and number of ports and its impact on
M
necessary time to perform operations considering mixed transportation matrix.

× 104 Dimension11250 × 104 Dimension12375 × 104 Dimension13500 × 104 Dimension14625

3 3 2.6 2.6
d
2.5 2.5 2.4 2.4
Time

Time

Time

2 2 2.2 Time 2.2

te

1.5 1.5 2 2
05 10 15 05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports Number of ports
× 10 4 Dimension15750 × 10 4 Dimension16875 × 10 4 Dimension18000 × 10 4 Dimension19125
3.5 5 4.5 4.5

3 4 4
p

4
Time

Time

Time

Time

2.5 3.5 3.5

3
2 3 3
ce

1.5 2 2.5 2.5

05 10 15 05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports Number of ports
× 104 Dimension20250 × 104 Dimension21375 × 104 Dimension22500
5 6 6

4.5
5 5
Time

Time

Time

4
Ac

4 4
3.5

3 3 3
05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports

Fig. 15. Improvement of Container ship size and number of ports and its impact on
necessary time to perform operations considering long transportation matrix.

to the ones presented in Figure 14. For example, in subplots corresponding

to Container ships with dimension of 13, 500, 20, 250, and 22, 500 TEUs, the

26

Page 28 of 60
necessary time to perform unloading and loading operations for 15 ports is
lower than for 10 ports. This is interesting feature since the improvement of
Container ships size does not necessarily results in higher times to perform

t
operations, and could results in more efficient operations. This partially

ip
explains the recent race for Container ships size increase, since the time to
perform unloading and loading operations in a bigger ship that performs a

cr
longer route will not increase more than linear. In some cases a reduction in
this time could even happen.
For the transportation matrix is type 3 (short), each subplot presented

us
in Figure 16 shows what happens for a specific container ship size when the
operation planning consider 5, 10 and 15 ports.

an
× 104 Dimension11250 × 104 Dimension12375 × 104 Dimension13500 × 104 Dimension14625
15 15 15 15

10 10 10 10
Time

Time

Time

Time
5 5 5 5
M
0 0 0 0
05 10 15 05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports Number of ports
× 105 Dimension15750 × 105 Dimension16875 × 105 Dimension18000 × 105 Dimension19125
2 2 2 2

1.5 1.5 1.5

1.5
Time

Time

Time

Time
1 1 1
1
d
0.5 0.5 0.5

0 0 0 0.5
05 10 15 05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports Number of ports
te

× 10 5 Dimension20250 × 10 5 Dimension21375 × 10 5 Dimension22500

2 2 2.5

2
1.5 1.5
Time

Time

Time

1.5
1 1
p

0.5 0.5 0.5

05 10 15 05 10 15 05 10 15
Number of ports Number of ports Number of ports
ce

Fig. 16. Improvement of Container ship size and number of ports and its impact on
necessary time to perform operations considering short transportation matrix.
Ac

The same pattern of linear increment in time to perform operation when

number of ports are increased, that appeared in some subplots from Figure
14, appears in all subplots in Figure 15.
Finally, the complete data employed to create the Figures 14, 15, and 16
could be found at the Appendix K.

27

Page 29 of 60
7. Conclusions
This paper showed a new approach to solving the 3D Stowage Planning

t
Problem for Container ship (3D SP) integrated with the Scheduling of quay

ip
cranes (SQC) problem. It has four modeling advantages:
• It shows how the 3D SP will affect other problems related to port

cr
operation, like the SQC problem.
• It provides better accuracy in estimating total time, which can save the

us
charterer charges for the the extra use of the vessel. In the instances
studied, the solution from the integrated approach provided solutions
with a 45.82% higher total time spent, on average, and prevented an
underestimation of necessary time for ship travel.

an
• In 40% of the instances, the Integrated 3D SP and SQP problem helped
to avoid a misleading analysis, where the adoption of good practices
M
for 3D SP produced a worse total time to unload and load the ship.

• The developed approach showed that a linear increment in container

d

ship size could produce a linear or lesser grow on time to perform un-
loading and loading operations. This partially explains the advantages
te

on use bigger container ships and justifies a long term tendency of

continuous increasing on container ship size.
p

The aforementioned advantages were attained without much computa-

tional effort. The reason for this is that the representation by rules encoding
ce

approach saves considerable computational time. For example, large-scale in-

stances with a 30-port horizon were solved in less than 12 minutes, showing
that this approach may eventually have practical commercial value. Another
Ac

advantage is that the integrated decision process makes it possible to observe

the detailed impact of different options of Stowage Planning in terms of the
total time necessary to perform the total workload by quay cranes.
Different or more complex rules would be needed to ensure other stability
features and to address specific unloading/loading constraints, or even quay
crane special features at a particular port. Actually, this should not be a
problem, since the proposed approach can deal with these practical situations
by merely changing the set of rules.

28

Page 30 of 60
8. Future Works
A promising idea for future work is to codify the previous approaches

t
proposed for the 3D SP or interview crane operators to produce rules from

ip
crew experience and observe their impact on total time necessary to complete
the total workload. Future work could also address this integrated problem
with methods that deal with the Pareto-optimal frontier in order to choose

cr
the solutions.
Since the port yard has the same stack structure that appears on a con-

us
tainer ship, one possible future work is to adapt the developed methodology
to determine the yard cargo arrangement and operation of equipment like
gantry cranes.
Another possibility of future work is to couple container ship loading and

an
unloading operations with retrieving and storing cargo into port yard by
determining a proper control on routing vehicles like AGVs or trucks.
Finally, a promising research branch is to consider different types of con-
M
tainers observing its size (20 feet, 40 feet, 40 feet extended), format (open,
flat, and others) or cargo type (dry, explosive, reefer, and others), and its
corresponding constraints of allocation.
d

9. Compliance with Ethical Standards

te

• Funding:This research was supported by the Fund to Support Teaching,

Research and Extension (FAEPEX) from the University of Campinas
p

(UNICAMP) and by the Foundation for the Support of Research of

the State of São Paulo (FAPESP) under the process 2010/51274-5 and
ce

2015/24295-5.

• Conflict of Interest: The authors declare that they have no conflict of

interest.
Ac

• The authors would like to thank the anonymous reviewers for their
valuable comments and suggestions to improve the quality of the paper.

29

Page 31 of 60
References
References

t
[1] D. Steenken, S. Voss, R. Stahlbock, Container terminal operation and

ip
operations research: a classification and literature review, OR Spectrum
26 (1) (2004) 3–49.

cr
[2] Y. Guan, K.-H. Yang, Z. Zhou, The crane scheduling: models and solu-
tion approaches, Annals of Operations Research 203 (1) (2013) 119–139.

us
[3] C. Bierwirth, F. Meisel, A survey of berth allocation and quay crane
scheduling problems in container terminals, European Journal of Oper-
ational Research 202 (3) (2010) 615 – 627.

an
[4] D. Chang, Z. Jiang, W. Yan, J. He, Integrating berth allocation and
quay crane assignments, Transportation Research Part E 46 (6) (2010)
975–990.
M
[5] A. Imai, H. Chen, E. Nishimura, S. Papadimitriou, The simultaneous
berth and quay crane allocation problem, Transportation Research Part
d
E 44 (5) (2008) 900–920.

[6] F. Meisel, C. Bierwirth, Heuristics for the integration of crane produc-

te

tivity in the berth allocation problem, Transportation Research Part E:

Logistics and Transportation Review 45 (1) (2009) 196–209.
p

[7] C. Yang, X. Wang, Z. Li, An optimization approach for coupling problem

ce

of berth allocation problem and quay crane assignment in container

terminal, Computers & Industrial Engineering 63 (1) (2012) 243–253.

[8] M. P. M. Hendriks, E. Lefeber, J. T. Udding, Simultaneous berth allo-

Ac

cation and yard planning at tactical level, OR Spectrum 35 (2) (2013)

441–456.

[9] K. Braekers, A. Caris, G. Janssens, Integrated planning of loaded and

empty container movements, OR Spectrum 35 (2) (2013) 457–478.

[10] C. Bierwirth, F. Meisel, A follow-up survey of berth allocation and quay

crane scheduling problems in container terminals, European Journal of
Operational Research 244 (3) (2015) 675–689.

30

Page 32 of 60
[11] Q. Zeng, Z. Yang, Integrating simulation and optimization to sched-
ule loading operations in container terminals, Computers & Operations
Research 36 (6) (2009) 1935–1944.

t
ip
[12] M. Avriel, M. Penn, S. Wittenboon, Stowage planning for container
ships to reduce the number of shifts, Annals of Operations Research
76 (1) (1998) 55–71.

cr
[13] I. Wilson, P. Roach, Container stowage planning: a methodology for
generating computerised solutions, Journal of the Operational Research

us
Society 51 (11) (2000) 1248–1255.

[14] A. Imai, K. Sasaki, E. Nishimura, S. Papadimitriou, Multi-objetive si-

an
multaneous stowage and loading planning for a container ship with con-
tainer rehandle in yard stacks, European Journal of Operational Re-
search 171 (2) (2006) 373–389.
M
[15] O. Dubrovsky, G. Levitin, M. Penn, A genetic algorithm with a com-
pact solution encoding for the containership stowage problem, Journal
of Heuristics 6 (2002) 585–599.
d

[16] A. Delgado, R. M. Jensen, K. Janstrup, T. H. Rose, K. H. Andersen,

A constraint programming model for fast optimal stowage of container
te

vessel bays, European Journal of Operational Research 220 (1) (2012)

251–261.
p

[17] A. Azevedo, C. Ribeiro, G. Sena, A. Chaves, L. Leduino, A. Moretti,

ce

Solving the 3d container ship loading planning problem by representa-

tion by rules and meta-heuristics, International Journal of Data Analysis
Techniques and Strategies: Special Issue on Optimisation and Simula-
tion in Realistic Scenarios 6 (3) (2014) 228–260.
Ac

[18] M. Monaco, M. Samarra, G. Sorrentino, The terminal-oriented ship

stowage planning problem, European Journal of Operational Research
239 (1) (2014) 256–265.

[19] D. Ding, M. Chou, Stowage planning for container ships: A heuristic

algorithm to reduce the number of shifts, European Journal of Opera-
tional Research 246 (1) (2015) 242 – 249.

31

Page 33 of 60
[20] S. Amaran, N. Sahinidis, B. Sharda, S. Bury, Simulation optimization:
a review of algorithms and applications, 4OR 12 (4) (2014) 301–333.

t
[21] G. Figueira, B. Almada-Lobo, Hybrid simulationoptimization methods:

ip
A taxonomy and discussion, Simulation Modelling Practice and The-
ory 46 (2014) 118–134, simulation-Optimization of Complex Systems:
Methods and Applications.

cr
[22] J. April, F. Glover, J. Kelly, M. Laguna, Simulation-based optimization:
Practical introduction to simulation optimization, in: Proceedings of the

us
35th Conference on Winter Simulation: Driving Innovation, WSC ’03,
Winter Simulation Conference, 2003, pp. 71–78.

an
[23] M. Avriel, M. Penn, N. Shpirer, Container ship stowage problem: com-
plexity and connection to the coloring of circle graphs, Discrete Applied
Mathematics 103 (1-3) (2000) 271–279.
M
[24] D. Ambrosino, D. Anghinolfi, M. Paolucci, A. Sciomachen, An experi-
mental comparison of different heuristics for the master bay plan prob-
lem, Lecture Notes in Computer Science 6049 (2010) 314–325.
d

[25] D. Pacino, R. Jensen, Constraint-based local search for container

stowage slot planning, in: In: Proceedings of the International Mul-
te

tiConference of Engineers and Computers Scientists II, 2012, pp. 1–6.

[26] L. Fan, M. Low, H. Ying, H. Jing, Z. Min, W. Aye, Stowage planning of

p

large containership with tradeoff between crane workload balance and

ce

ship stability, in: In: Proceedings of the International MultiConference

of Engineers and Computers Scientists III, 2010, pp. 1–7.

[27] D. Ambrosino, A. Sciomachen, E. Tanfani, A decomposition heuristics

Ac

for the container ship stowage problem, Journal of Heuristics 12 (3)

(2006) 211–233.

[28] L. Cruz-Reyes, P. Hernández H., P. Melin, H. J. Fraire H., J. Mar O.,

Constructive algorithm for a benchmark in ship stowage planning, in:
O. Castillo, P. Melin, J. Kacprzyk (Eds.), Recent Advances on Hybrid
Intelligent Systems, Vol. 451 of Studies in Computational Intelligence,
Springer Berlin Heidelberg, 2013, pp. 393–408.

32

Page 34 of 60
[29] Y. Zhu, A. Lim, Crane scheduling with non-crossing constraint, Journal
of the Operational Research Society 57 (12) (2006) 1464–1471.

t
[30] H. Javanshir, S. S. Ganji, Yard crane scheduling in port container ter-

ip
minals using genetic algorithm, Journal of Industrial Engineering Inter-
national 6 (11) (2010) 39–50.

cr
[31] D. Lee, D. Wang, L. Miao, Quay crane scheduling with non-interference
constraints in port container terminals, Transportation Research E
44 (1) (2008) 124–135.

us
[32] P. Legato, R. Mazza, R. Trunfio, Simulation-based optimization for the
quay crane scheduling problem, in: WSC ’08 Proceedings of the 40th

an
Conference on Winter Simulation, 2008, pp. 2717–2725.

[33] K. Mak, D. Sun, Scheduling yard cranes in a container terminal using a

new genetic approach, Engineering Letters 17 (4) (2009) 274–280.
M
[34] K. Kim, Y.-M. Park, A crane scheduling method for port container
terminals, European Journal of Operational Research 156 (3) (2004)
752 – 768.
d

[35] K. Dompere, Fuzziness and Foundations of Exact and Inexact Sciences,

te

Vol. 290 of Series in Fuzziness and Soft Computing, Springer, 2012.

[36] C. Bierwirth, F. Meisel, A fast heuristic for quay crane scheduling with
p

interference constraints, J. of Scheduling 12 (4) (2009) 345–360.

ce

[37] J. Chen, D.-H. Leeb, M. Gohc, An effective mathematical formulation

for the unidirectional cluster-based quay crane scheduling problem, Eu-
ropean Journal of Operational Research 232 (1) (2014) 198–208.
Ac

[38] H. S. L. Services, Import demurrage: Freetime and tariffs (2016).

URL http://www.hamburgsud-line.com/hsdg/en/hsdg/
regionalinformation/southamerica/brazil/surcharges_7/
demurrage_1/demurrage_1.jsp

[39] M. Caramia, P. Dell’Olmo, Multi-objective Management in Freight Lo-

gistics: Increasing Capacity, Service Level and Safety with Optimization
Algorithms, Springer London, London, 2008, Ch. Multi-objective Opti-
mization, pp. 11–36.

33

Page 35 of 60
[40] E. J. Arajo, A. A. Chaves, L. L. de Salles Neto, A. T. de Azevedo, Pareto
clustering search applied for 3d container ship loading plan problem,
Expert Systems with Applications 44 (2016) 50–57.

t
ip
[41] J.-P. Rodrigue, The geography of transport systems - the largest avail-
able containership (2017).
URL https://people.hofstra.edu/geotrans/eng/ch3en/conc3en/

cr
largestcontainerships.html

[42] E. Altena, Optimization of modal shift and container (re-)positioning

us
at maersk line (2013).
URL http://repository.tudelft.nl/assets/uuid:
66811c74-3c70-4b88-ab02-72f2fe86839d/Report_final_altena.

an
pdf

[43] J.-P. Rodrigue, The geography of transport systems - containerships

evolution (2017).
M
URL https://people.hofstra.edu/geotrans/eng/ch3en/conc3en/
containerships.html

[44] C. from China, Shipping routes from china (2017).

d

URL https://cargofromchina.com/shipping-routes/
te

[45] C. Derrett, B. Barrass, Ship Stability for Masters and Mates, 5th Edi-
tion, Butterworth-Heinemann, 1999.
p

URL https://www.123library.org/ebook/isbn/9780080480886/
ce

[46] A. Diabat, E. Theodorou, An integrated quay crane assignment and

scheduling problem, Computers & Industrial Engineering 73 (2014)
115123.
Ac

[47] Y.-M. Fu, A. Diabat, I.-T. Tsai, A multi-vessel quay crane assignment
and scheduling problem: Formulation and heuristic solution approach,
Expert Systems with Applications 41 (15) (2014) 6959–6965.

[48] J. Bisschop, AIMMS - Optimization Modeling, AIMMS B.V., 2017.

URL https://download.aimms.com/aimms/download/manuals/
AIMMS3_OM.pdf

34

Page 36 of 60
Appendices
A. A Detailed Description of Loading and Unloading Rules Description

t
This appendix describes in detail the eight loading and two unloading

ip
rules called in the evaluation of Algorithm 1 in lines 15 and 8, respectively.
The loading rules vary in terms of:

cr
• Filling the ship by forming and positioning vertical stacks among bays
or bay by bay;

us
• Forming flooring layers in the ship basement among bays or bay by
bay.

an
The first six rules are detailed in [17], but Rules 7 and 8 are new and an
original contribution of this work. The detailed description of the rules are:

• LR1: This rule fills the ship row by row, from left to right, starting
M
from the bottom row for each bay in a manner that the containers with
the farthest destination are placed on the lowest rows and each bay is
filled before the next. Figure A-1 shows the application of the loading
d

rule for the container ship at Port 1. The zero numbers indicate empty
space. Positive numbers indicates spaces occupied by containers whose
te

destination port is the corresponding positive number.

• LR2: This rule fills the ship row by row, from left to right, starting
p

from the first bay and filling only one row per bay in a manner that the
containers with the farthest destination are placed on the lowest rows
ce

and distributed among the bays.

• LR3: This rule is the reverse of LR1, which means the ship is filled row
Ac

by row, from right to left, starting from the bottom row for each bay in
a manner that the containers with the farthest destination are placed
on the lowest rows and each bay is filled before the next is begun.

• LR4: This rule is the reverse of LR2 in the sense that it fills the ship
row by row, from right to left, one row per bay, starting from the
first bay, until it reaches the last in a manner that the containers with
the farthest destination are placed on the lowest rows and distributed
among the bays.

35

Page 37 of 60
 t
ip
cr
Fig. A-1. The container arrangement in ship after applying LR1.

us
• LR5: This rule fills the ship, row by row, from left to right, with
containers destined for the nearest port, starting from the first bay
and continuing until the number of elements θp in a column has been

an
reached. The value θp is computed by Eq. (A-1).

p N
M
X X
Tij
i=1 j=p+1
θp = ⌈ ⌉ (A-1)
D×C
d

Then another bay is filled in a manner that the containers with the
te

nearest destination are placed first to form the stacks.

• LR6: This rule is the reverse of LR5 in the sense that it fills the ship,
p

row by row, from right to left, with containers destined for the nearest
port, starting from the first bay and continuing until the number of el-
ce

ements θp in a column has been reached. The value θp is also computed

by Eq. (A-1).

• LR7: All previous rules start from the ship bow. This rule fills the ship
Ac

by alternating between filling from the middle to bow and from middle
to stern. Additionally, it fills the ship from left to the right, first filling
the lower part of each stack of each line with loads whose destination
is farthest away.

• LR8: This rule is similar to LR7, but it begins it fills from middle to
stern and from bow to middle.

There are only two unloading rules:

36

Page 38 of 60
 • UL1: Remove all containers to allow a complete rearrangement in con-
tainers and avoid future re-handles or;

t
• UL2: Remove only the containers that have currently arrived at the

ip
destination port or the containers that are blocking the former ones.

B. An example of the function related to Quay Crane Rule 1

cr
This Appendix presents the m-code function that computes the total time
necessary to perform two quay cranes operations in container ship bays.

us
1 f u n c t i o n [ tmax ] = f q c 1 ( vtbay )
2 % Inputs :
3 % ( 1 ) vtbay − a v e c t o r o f t o t a l workload per bay .

an
4 %
5 % Parameters :
6 % ( 1 ) mdist − minimal d i s t a n c e between quay c r a n e s
7 % ( 2 ) esp − a c t u a l s p a c e between quay c r a n e s
M
8 % ( 3 ) nbay − number o f bays (= l e n g t h ( vtbay ) )
9 % ( 4 ) nquay − number o f quay c r a n e s ( q c s ) −( suppo se 2 )
10 % (5) tf1 − t o t a l time f o r qc1 f i n i s h a l l s e r v i c e s
d
11 % (6) tf2 − t o t a l time f o r qc2 f i n i s h a l l s e r v i c e s
12 % (7) i1 − i ndex f o r bay s e r v e d with qc1
te

13 % (8) i2 − i ndex f o r bay s e r v e d with qc2

14 % ( 9 ) tmove − time t o move a qc t o a n o t h e r bay :
15 % − c o n s i d e r e d 3 x ( time t o l o a d ) o r
p

16 % − ( time t o unload ) one u n i t from s h i p

%
ce

17

18 % Outputs :
19 % ( 1 ) tmax − t o t a l time need
20 %
Ac

21

22 % I n i t i a l i z a t i o n of values .
23 nbay = l e n g t h ( vtbay ) ; % number o f s h i p bays .
24 tf1 = 0; % i n i t i a l t o t a l time o f qc1 o p e r a t i o n .
25 tf2 = 0; % i n i t i a l t o t a l time o f qc2 o p e r a t i o n .
26 i1 = 1; % S t a r t i n g qc 1 i n t he prow .
27 i2 = min ( round ( nbay / 2 ) +1 ,nbay ) ; %qc2 a t middle | s t e r n
.
28 s1 = i2 − 1;

37

Page 39 of 60
29 s2 = nbay ;
30 esp = i2 − i1 − 1; % s p a c e between q c s .
31 mdist = 1 ; % minimum d i s t a n c e between q c s .

t
32 tmove = 3 ; % time t o move qc from one bay t o a n o t h e r .

ip
33 % Ini t i al position for both c r a n e s .
34 t f 1 = t f 1 + vtbay ( i 1 ) ; % To c r e a t e a time l i n e f o r qc1 .

cr
35 t f 2 = t f 2 + vtbay ( i 2 ) ; % To c r e a t e a time l i n e f o r qc2 .
36

37 % Three main c a s e s :

us
38 % ( 1 ) Qcs l a s t bay : no more d i s t . and time−l i n e check .
39 % ( 2 ) With o r wi t ho ut minimal d i s t a n c e .
40 % ( 3 ) P r o c e s s i n g o f qc1 o p e r a t i o n ( t f 1 < t f 2 )

an
41 % o r qc 2 ( t f 2 < t f 1 ) o r both ( t f 1 == t f 2 ) .
42

43 % Loop t o l o a d o r unload a l l t he s h i p .
44 w h i l e ( ( i 1 <= s1 ) | | ( i 2 <= s2 ) )
M
45

46 % Case ( 3 ) : q c s a r e not i n t he l a s t o p e r a t i o n bay .

47 i f ( ( i 1 < s1 )&&( i 2 < s2 ) )
d
48 % Case ( 2 . A) : min d i s t a n c e ok , update qc p o s i t i o n s
.
te

49 i f ( esp > mdist )

50 % Case ( 3 . A) : qc1 o p e r a t i o n ( t f 1 < t f 2 ) .
51 i f ( tf1 < tf2 )
p

52 esp = esp − 1 ;
i1 = i1 + 1;
ce

53

54 t f 1 = t f 1 + vtbay ( i 1 ) + tmove ;
55 % Case ( 3 . B) : qc2 o p e r a t i o n ( t f 2 < t f 1 ) .
56 el s ei f ( tf2 < tf1 )
Ac

57 esp = esp + 1 ;
58 i2 = i2 + 1;
59 t f 2 = t f 2 + vtbay ( i 2 ) + tmove ;
60 % Case ( 3 . C) : qc1 & 2 o p e r a t i o n ( t f 2 == t f 1 ) .
61 else
62 i1 = i1 + 1;
63 t f 1 = t f 1 + vtbay ( i 1 ) + tmove ;
64 i2 = i2 + 1;
65 t f 2 = t f 2 + vtbay ( i 2 ) + tmove ;

38

Page 40 of 60
66 end
67

68 % Case ( 2 . B) : minimal d i s t a n c e p r e s e r v e d o r not .

t
69 % Depends on what qc opr . ( esp == mdist ) .

ip
70 e l s e i f ( esp == mdist )
71 % Case ( 3 . A) : qc2 o p e r a t i o n ( t f 2 < t f 1 ) .

cr
72 i f ( tf2 < tf1 )
73 esp = esp + 1 ;
74 i2 = i2 + 1;

us
75 t f 2 = t f 2 + vtbay ( i 2 ) + tmove ;
76 % Case ( 3 . B) : qc1 c a n d i d a t e f o r opr . ( t f 1 < t f 2 ) ,
77 % but i t r e s u l t s i n d i s t a n c e v i o l a t i o n ! !

an
78 % S y n c r o n i z a t i o n o f qc1 & 2 o p e r a t i o n s .
79 el s ei f ( tf1 < tf2 )
80 t f 1 = t f 2 ; % s y n c h r o n i z a t i o n o f q c s movement ! !
81 i1 = i1 + 1;
M
82 i2 = i2 + 1;
83 t f 1 = t f 1 + vtbay ( i 1 ) + tmove ;
84 t f 2 = t f 2 + vtbay ( i 2 ) + tmove ;
d
85 % Case ( 3 . C) : qc1 & 2 opr . ( t f 2 == t f 1 ) .
86 else
te

87 i1 = i1 + 1;
88 t f 1 = t f 1 + vtbay ( i 1 ) + tmove ;
89 i2 = i2 + 1;
p

90 t f 2 = t f 2 + vtbay ( i 2 ) + tmove ;
end
ce

91

92 % Case ( 2 . C) : min . d i s t . not ok . Move o n l y qc2 .

93 else
94 t f 1 = t f 2 ; % s y n c h r o n i z a t i o n o f q c s movement ! !
Ac

95 end % end o f i f f o r minimal d i s t a n c e c h e c k i n g .

96

97 % Case ( 1 ) : At l e a s t one qc o p e r a t e s i n i t s l a s t bay .

98 % Space between q c s i s no l o n g e r a problem , then i t
99 % i s not updated anymore .
100 else
101 % Case ( 1 . A) : qc1 on t he l a s t bay − move o n l y qc2 .
102 i f ( ( i 1 >= s1 )&&( i 2 < s2 ) )
103 i2 = i2 + 1;

39

Page 41 of 60
104 t f 2 = t f 2 + vtbay ( i 2 ) + tmove ;
105 end
106 % Case ( 1 . B) : qc2 on t he l a s t bay : move o n l y qc1 .

t
107 i f ( ( i 1 < s1 )&&( i 2 >= s2 ) )

ip
108 i1 = i1 + 1;
109 t f 1 = t f 1 + vtbay ( i 1 ) + tmove ;

cr
110 end
111 % Case ( 1 . C) : qc1 & 2 a r e on l a s t bay : both move .
112 i f ( ( i 1 >= s1 )&&( i 2 >= s2 ) )

us
113 i1 = i1 + 1;
114 i2 = i2 + 1;
115 end

an
116 end % end o f i f f o r l a s t bay o p e r a t i o n check .
117

118 end % end o f f o r t he l o o p o f u s i n g a l l qc movement .

119
M
120 % To t a l p r o c e s s i n g time i s t he max . time o f p r o c e s s i n g
121 % a l l t a s k s d e s i g n a t e d f o r qc1 & 2 .
122 tmax = max( t f 1 , t f 2 ) ;
d

The previous code could be generalized for more than two quay canes by
observing the following conditions:
te

1. Suppose there are two quay cranes: qc1 and qc2 . The position of each
quay crane i is tracked by a corresponding variable tfi .
p

2. The lowest value on index i indicates which is the leftmost quay crane,
ce

i.e., qc1 .
3. Suppose the two cranes qc1 and qc2 are moving to the right, because,
according to [37], when quay cranes move in the same direction, the
Ac

optimal solution can still be found.

4. The distance between cranes should be at least equal or higher than a
minimal distance mdist.
5. A quay crane cannot overpass another quay crane.
6. From these suppositions, the leftmost quay crane’s, qc1 , move depends
on the rightmost quay crane’s position, qc2 . This may be expressed in
the following condition: tf1 + mdist ≤ tf2 .

40

Page 42 of 60
 7. This results in the following: qc2 is free to move, but the position of
qc1 is determined by the position of qc2 .

t
The previous conditions lead to the fact that a quay crane rule may

ip
be applied to N quay cranes by applying the following recursive pairwise
computation:

cr
1. Check distance between the rightmost quay crane and its immediate
predecessor, i.e, the positions of qcN and qcN −1 .

us
2. Once the position of qcN −1 has been computed, the same previous check
can be carried out for qcN −i and qcN −(i+1) , i = 1.

an
3. This sequence can be applied recursively, i.e., ∀i = 2, · · · , N − 2.

C. Generating Instances
As suggested by [12], it is possible to create three types of transportation
M
matrix: 1 - Mixed, 2 - Long and 3 - Short. This classification describes how
long the majority of containers must be carried by the container ship for a
specific transportation matrix (instance). The Short transportation matrix
d

indicates that the majority of the containers will remain for a small number
of ports until unloading. The Long transportation matrix indicates that for
te

most of the containers will remain on board the container ship. The Mixed
is created in a manner that mixes Short and Long instance characteristics.
The generation of all instances follows the procedure described in [12], which
p

ensure ship will carry all containers available in ports, but with the modifi-
ce

cation proposed in [17] has been added, however, to consider the available
space of 3D container ship:
Ac

k
X N
X
Tij ≤ R × C × D, ∀i = 1, · · · , N − 1, (C-1)
i=1 j=k+1

where Tij is an element in line i and column j from the transportation

matrix T ; N is the number of ports; R is the number of the rows; C is the
number of columns; and D is the number of bays.
For mixed matrices the distance matrix Tij is random such that Eq. (C-
1) is satisfied and, additionally, each distance matrix element Tij is subject

41

Page 43 of 60
 D1 D2 D3 D4 D1 D2 D3 D4
O1 0 2 1 2 O1 0 0 0 8
0 0 2 0 0 0 0 1

t
O2 O2

ip
O3 0 0 0 0 O3 0 0 0 0
O4 0 0 0 0 O4 0 0 0 0
(a) Mixed distance. (b) Long distance.

cr
D1 D2 D3 D4
O1 0 3 0 0

us
O2 0 0 9 0
O3 0 0 0 0
O4 0 0 0 0

an
(c) Short distance.

Fig. C-1. Small examples of three types of distance matrix.

M
to Tij ≤ 0.2 × R × C × D. This constraint prevents the number of matrix
elements equal to zero from being large.
For long distance matrices, the constraint Tij ≤ 0.2×R×C ×D is replaced
d

by Tij ≤ ((j − i)/(N − 1))2 to prevent near-diagonal elements from having

large values.
te

Finally, the short distance matrices consider only Eq. (C-1) and a special
order to fill the distance matrix in a manner such that near-diagonal matrices
p

become somewhat large.

Examples of the three matrix types are illustrated in Fig. C-1 using the
ce

same parameters of Fig. 3: D = 1, R = 3, C = 3, and N = 4.

D. Stowage Planning Mathematical Model

The mathematical model in terms of linear programming with binary
Ac

variables for the 3D SP is given by (D-1)-(D-8).

min f (xijv (r, c, d), yi(r, c, d)) = αφ1 (xijv (r, c, d)) + βφ2 (yi (r, c, d)) (D-1)
subject to

42

Page 44 of 60
 j R X
C X
D i−1 X
R X
C X
D
X X X
xijv (r, c, d) − xkij (r, c, d) = Tij
v=i+1 r=1 c=1 d=1 k=1 r=1 c=1 d=1

t
i = 1, · · · , N − 1; j = i + 1, · · · , N (D-2)

ip
i N j
X X X
xkjv (r, c, d) = yi (r, c, d)

cr
k=1 j=i+1 v=i+1
i = 1, · · · , N − 1; r = 1, · · · , R; c = 1, · · · , C; d = 1, · · · , D (D-3)
yi (r, c, d) − yi (r + 1, c, d) ≥ 0

us
i = 1, · · · , N − 1; r = 1, · · · , R − 1; c = 1, · · · , C; d = 1, · · · , D (D-4)
j−1 N j−1 N p
X X X X X
xipj (r, c, d) + xipv (r + 1, c, d) ≤ 1

an
i=1 p=j i=1 p=j+1 v=j+1
j = 2, · · · , N; r = 1, · · · , R − 1; c = 1, · · · , C; d = 1, · · · , D (D-5)
xijv (r, c, d) = 0 or 1
M
i, j, v = 1, · · · , N; r = 1, · · · , R; c = 1, · · · , C; d = 1, · · · , D. (D-6)
where: the binary variable xijv (r, c, d) is defined as follows: if, in port i,
the compartment (r, c, d) has a container whose destination is port j and this
d
container was moved in port v, then the variable assumes value 1; otherwise
value 0 is assumed. The term compartment (r, c, d) represents row r, column c
te

for the container ship bay d. Similarly, variable yi (r, c, d) is defined as follows:
if, in port i, the compartment (r, c, d) has a container; then the variable
assumes value 1; otherwise value 0 is assumed. The objective function (D-1)
p

is composed of two terms: the first is the total cost of moving a container
and, the second is the sum of instability measures for the container ship
ce

configuration in each port. It is assumed that, for all ports, the container
movement costs the same and is equal to one. Constraints (D-2) express
the total number of containers that will be shipped from port i to port j.
Ac

Constraints (D-3) require that each compartment (r, c, d) of the container

ship is always occupied by at most one container. Constraints (D-4) are
related to the physical storage of the containers in the ship, and it imposes
that, for each container in row r + 1, there be another container in the
row r for all r = 1, · · · , R − 1. Constraints (D-5) define how a container
can be unloaded from the ship in port j by requiring that, if a container
occupies the position (r, c, d) at port j, and it will be unloaded, then, there
are no containers above or the containers above have already been unloaded
at previous ports. Finally, Constraints (D-6) defines the variables nature.

43

Page 45 of 60
 The two terms which compose the objective function (Eq. (D-1)) de-
fine two optimization criteria: the first term is a function of the number of
containers moved, φ1 (x), and the second depends on how the container ship

t
is organized in each port, φ2 (y). The two criteria are combined by values

ip
given by weights α and β in a manner that forms a bi-objective optimization
framework.

cr
The term φ1 (xijv (r, c, d)) assumes that for all ports, the container move-
ment cost is the same and is equal to one which may be translated as Eq.
(D-7).

us
N −1 N j−1 C X
R X D
X X X X
φ1 (xijv (r, c, d)) = xijv (r, c, d) (D-7)

an
i=1 j=i+1 v=i+1 r=1 c=1 d=1

The term φ2 (yi (r, c, d)) refers to the container ship’s transverse stability
and assumes that every container has the same mass and is equal to one.
M
This term is to control the container ship transverse stability before leave
the port which means after all loading movements had been performed as
described by Eq. (D-8).
d

N
te

X
φ2 (yi (r, c, d)) = (−∆GMi + ∆Li ) (D-8)
i=1
p

where:
ce

R D X
C
! !
X X
∆GMi = yi (r, c, d) · (GYship − r + 0.5) ,
r=1 d=1 c=1
∆Li = hpi + hni ,
Ac

C D X
R
!!
X X
yi (r, c, d) · (GXship − c + 0.5) = hpi − hni ,
c=1 d=1 r=1
hpi , hni ≥ 0

where: the values GYship and GXship represent vertical and horizontal
coordinates of gravity center of the ship, respectively. The variables ∆GMi
represent the variance in metacentric height GM in each port i after loading
all containers. Since GM is the distance between the centre of gravity of

44

Page 46 of 60
a ship and its metacentre, as much metacentric height is increased with a
∆GMi > 0, it turns more difficult the ship to overturn. The variables ∆Li
also helps with the reduction of angle of list after loading all containers.

t
Angle list measures the vessel leaning to either port or starboard. More

ip
discussion about metacentric height increasing and reduction of angle of list
and corresponding objective functions could be seen at [14, 45].

cr
One model feature is that the number of containers that must be loaded
at a certain port is given by a transportation matrix T of dimension (N −
1) × (N − 1), whose element Tij represents the number of containers from

us
port i that must be transported to the destination port j. This matrix is an
upper triangular matrix, since Tij = 0 for every i ≥ j.
Another feature is that the container ship has a rectangular format and

an
can be represented by a matrix with rows (r = 1, 2, · · · , R), columns (c =
1, 2, · · · , C) and bays (d = 1, 2, · · · , D) with maximum capacity of R×C ×D
containers. Irregular formats could be achieved by simply adding constraints
which represents imaginary containers that occupies same spaces during the
M
whole voyage [19].
Finally, this model presents an original contribution to the literature since
it precisely computes the number of movements, instead of doing an estima-
d
tion as done in [14], and the computation of stability issues is done by using
an objective function linear, instead of using non-linear one as done in [17].
te

E. Quay Crane Scheduling Problem

The mathematical model in terms of linear programming with binary
p

variables for the Scheduling Quay Cranes (SQC) had been developed to avoid
indexing in time that could be advantageous as observed by [46, 47] and used
ce

by [36, 37]. The model is given by (E-1)-(E-8).

Ac

min τ (E-1)
subject to

45

Page 47 of 60
 K
X
wj,k = 1, j ∈ ΩB (E-2)
k=1

t
t1,k ≥ a1 w1,k , k ∈ ΩK , j ∈ ΩB (E-3)

ip
tj+1,k ≥ tj,k + aj+1 wj+1,k , k ∈ ΩK , j ∈ ΩB−1 (E-4)
wj+1,k + wj,k+1 = zj,k , j ∈ ΩB−1 , k ∈ ΩK−1 (E-5)

cr
tj+1,k ≥ tj,k+1 + aj+1 (zj,k − 1) , j ∈ ΩB , k ∈ ΩK−1 (E-6)
τ ≥ tj,k , j ∈ ΩB , k ∈ ΩK (E-7)

us
zj,k ≥ 0, wj,k = 0 or 1,
ad = f (xijk (r, c, d), yi (r, c, d)), j ∈ ΩB , k ∈ ΩK . (E-8)
From the given model the following decisions variables had been consid-

an
ered:

• The binary variable wj,k assumes value 1, if the crane k is allocated to

bay or bay group j;
M
• The integer auxiliary variable zj,k helps to prevent quay cranes overpass
each other;
d

• The real variable tj,k represents the total time necessary to perform
operations for all k cranes until bay or group of bays j;
te

The objective function is the minimization of the time necessary to finish

the necessary services on the container ship, and this could be unloading or
p

loading movements. This is represented by the minimization of the finish

time of the last work unit, i.e., the makespan. Constraint (E-2) determines
ce

that each group of bays j should be attend by only one crane k. Constraints
(E-3) and (E-4) helps in the computation of total time necessary to perform
operations till bay j. Constraint (E-5) defines auxiliary variable that helps to
Ac

identify possible points where quay cranes could overpass each other. Con-
straint (E-6) compute total processing time by avoiding quay crane overpass.
Constraints (E-7) gives the makespan. Finally, Constraints (E-8) defines the
variables nature. For instance, once a crane starts to work on a section, it
has to finish the workload in this section before it moves to other sections.

F. Integrated Mathematical Model for 3D SP and SQC

The SQC problem defined by Eq. (E-1)-(E-8) could be coupled with the
3D SP Eq. (D-1)-(D-8), but by solving the SQC problem two times at every

46

Page 48 of 60
port i: once after ship perform unloading and after ship loading containers.
Also, indexation should be adjusted to be compatible with the one used in
3D SP which means index i is for port and s is kind of operation performed

t
(1 for unloading and 2 for loading). It was also assumed, for simplicity, that

ip
every group of bays b corresponds exactly to one bay d. Finally, some ad-
ditional constraints connect the decisions made in 3D SP and correspondent

cr
unloading and loading effort through ports. These assumptions lead to the
model given by Eq. (F-1)-(F-10).

us
N X
X S
min τi,s (F-1)
i=1 s=1

an
subject to
M
d
p te
ce
Ac

47

Page 49 of 60
 K
X
wi,s,j,k = 1,
k=1

t
i = 1, · · · , N, s = 1, · · · , S, j = 1, · · · , D (F-2)

ip
ti,s,1,k ≥ ai,s,1 wi,s,1,k ,
i = 1, · · · , N, s = 1, · · · , S, k = 1, · · · , K, j = 1, · · · , D (F-3)

cr
ti,s,j+1,k ≥ ti,s,j,k + ai,s,j+1wi,s,j+1,k ,
i = 1, · · · , N, s = 1, · · · , S, k = 1, · · · , K, j = 1, · · · , D − 1 (F-4)

us
wi,s,j+1,k + wi,s,j,k+1 = zi,s,j,k ,
i = 1, · · · , N, s = 1, · · · , S, j = 1, · · · , D − 1, k = 1, · · · , K − 1 (F-5)
ti,s,j+1,k ≥ ti,s,j,k+1 + ai,s,j+1 (zi,s,j,k − 1) ,

an
i = 1, · · · , N, s = 1, · · · , S, j = 1, · · · , D, k = 1, · · · , K − 1 (F-6)
τi,s ≥ ti,s,j,k ,
i = 1, · · · , N, s = 1, · · · , 2, j = 1, · · · , D, k = 1, · · · , K (F-7)
M
zi,s,j,k ≥ 0, wj,k = 0 or 1. (F-8)
XN X i−1 X R X C
ψi,1 (r, c, d)xkji (r, c, d) = ai,1,d
d
j=1 k=1 r=1 c=1
i = 1, · · · , N, d = 1, · · · , D (F-9)
te

N
X N X
X R X
C
ψi,2 (r, c, d)xijv (r, c, d) = ai,2,d
j=i+1 v=j r=1 c=1
p

i = 1, · · · , N, d = 1, · · · , D (F-10)
ce

The constraints (F-2)-(F-8) are just the SQC model adapted to give the
minimal time of loading and unloading to 3D SP through feasible quay cranes
scheduling. Constraints (F-9)-(F-10) couple the decision of how to perform
Ac

container ship arrangement (xi,j,v (r, c, d)) through ports with the total work-
load per bay employed in unloading (ai,1,d ) and loading cranes scheduling
service (ai,2,d ). Finally, Eq. (F-1) should replace the term φ1 (x) in objective
function (Eq. (D-7)) since the term is just an estimation of how much time
will be spent by ship through ports.
Although, this integrated mathematical model has constraints (F-3), (F-
4) and (F-6) as nonlinear ones. Constraints (F-3) and (F-4) is a product
of ai,s,d , which is a sum of xkji (r, c, d) variables multiplied by a real num-
ber ψi,1 (r, c, d) and wi,s,j+1,k binary variables. To avoid such problem, it is

48

Page 50 of 60
necessary to apply some integer programming techniques, in particular, the
elimination of products of variables as described in [48]. Let wi,s,j+1,k be a
binary variable, and ai,s,j be a continuous variable for which 0 ≤ ai,s,j ≤ amax
i,s,j

t
holds. Now, a new continuous variable, qi,s,j,k , is introduced to replace the

ip
product qi,s,j+1,k = ai,s,j wi,s,j+1,k . The following constraints must be added
to force qi,s,j+1,k to take value of ai,s,j wi,s,j+1,k .

cr
qi,s,j+1,k ≤ amax
i,s,j wi,s,j+1,k (F-11)

us
qi,s,j+1,k ≤ ai,s,j (F-12)
max
qi,s,j+1,k ≥ ai,s,j − ai,s,j (1 − wi,s,j+1,k ) (F-13)
qi,s,j+1,k ≥ 0 (F-14)

an
Additionally, constraints (F-3) and (F-4) should be replaced by (F-15)
and (F-16).
M
ti,s,1,k ≥ qi,s,1,k ,
i = 1, · · · , N, s = 1, · · · , 2, k = 1, · · · , K, j = 1, · · · , D (F-15)
ti,s,j+1,k ≥ ti,s,j,k + qi,s,j+1,k ,
d

i = 1, · · · , N, s = 1, · · · , 2, k = 1, · · · , K, j = 1, · · · , D − 1 (F-16)
te

The same technique is applied to (F-6) constraint by defining a new con-

tinuous variable, ui,s,j,k to replace the product ui,s,j,k = ai,s,j zi,s,j,k .
p
ce

ui,s,j,k ≤ amax
i,s,j zi,s,j,k (F-17)
ui,s,j,k ≤ ai,s,j (F-18)
max
ui,s,j,k ≥ ai,s,j − ai,s,j (1 − zi,s,j,k ) (F-19)
Ac

ui,s,j,k ≥ 0 (F-20)
Finally, constraint (F-6) is replaced by (F-21).

ti,s,j+1,k ≥ ti,s,j,k+1 + ui,s,j,k − ai,s,j+1,

i = 1, · · · , N, s = 1, · · · , 2, k = 1, · · · , K, j = 1, · · · , D − 1 (F-21)
The mathematical model for 3D SP integrated with SQC is formed by
Eqs. (D-1)-(D-8), (F-1),(F-2), (F-5),(F-7)- (F-10), and (F-11)-(F-21).

49

Page 51 of 60
 Table G-1. Transportation information for a five port example.

P2 P3 P4 P5
P1 2 5 0 0

t
P2 0 2 3 1

ip
P3 0 0 2 2
P4 0 0 0 1

cr
G. Small numerical example for 3D SP Model

us
The mathematical model has been applied in a small example just to
illustrate how container ship arrangement could be affected by stability mea-
sures and the proposed model had been successful to provide more stable

an
ship arrangements through ports.
The numerical example consists on a container ship with dimensions R =
4, C = 4 and D = 1. The number of ports is N = 5, and in each port there
are K = 2 quay cranes available for unloading and loading operations. It was
M
also considered, for simplicity, each column as bay. Each element Ti,j from
transportation matrix gives the number of containers that should be loaded
in port i which destination is port j. The transportation matrix used in this
example is shown in Table G-1.
d

The mathematical model had been solved using GUSEK (the GLPK In-
te

teger Optimizer v4.55 with all cuts enabled) and the corresponding model
has 176 constraints, and 400 binary variables. The integer solution had been
found in 0.4 seconds in a laptop with Intel Core i7-4500U CPU and 1.80 GHz
p

of RAM.
When the model is set to minimize only the total number of container
ce

movements (α = 1 and β = 0), the resulting container ship arrangement is

given in Figure G-1. A different container ship arrangement is produced for
only stability measures maximization (α = 0 and β = 1) as shown in Figure
Ac

G-2. Each square in both figures has a pair of numbers (i, j) representing a
space of the ship occupied by a container, or 0 representing an empty space.
The first number of the pair gives the container loading port information and
the second gives the container unloading port information. The number of
movements per port for both solutions is presented in Table G-2 which leads
to a total cost of US$7200 (since each movement costs US$200). The article
[17] provides a detailed procedure on how to compute stability measure.
Figure G-2 shows a more stable container ship arrangement specially in
ports 2 and 3 when compared with the arrangement in Figure G-1, but

50

Page 52 of 60
 Table G-2. Number of unloading and loading movements per port.

Port 1 2 3 4 5 Total
loading 7 6 4 1 0 18

t
unloading 0 2 7 5 4 18

ip
cr
us
an
M
Fig. G-1. The container ship arrangement for minimization of φ1 (·).
d
p te
ce

Fig. G-2. The container ship arrangement for minimization of φ2 (·).

Ac

without an additional number of container movements.

H. Small numerical example for SQC Model

The mathematical model (as given on Appendix E) has been applied in a
small example just to illustrate how quay crane scheduling could be affected
by the constraint of one crane could not cross another.
The numerical example consists on a container ship with number of bays
equals to B = 9. The number of quay cranes available to perform operations

51

Page 53 of 60
 Table H-1. Time to process for a nine bay container ship.

Bay 1 2 3 4 5 6 7 8 9
Time 22 46 8 70 10 38 40 16 22

t
ip
Table H-2. Scheduling of quay cranes through bays of the container ship.

Bay 1 2 3 70 5 6 7 8 9

cr
QC2 0 0 0 70 80 0 120 0 132
QC1 22 68 76 0 0 118 0 136 0

us
is K = 2. It was also considered, for simplicity, no time is spent on moving
quay cranes and minimal distance between them could be zero. Each element

an
on vector represents a bay of the container ship and gives the total workload
in terms of amount of time necessary to process a task (unloading or loading).
The vector used in this example is shown in Table H-1 and comes from the
article [6].
M
The mathematical model had been solved using GUSEK (the GLPK In-
teger Optimizer v4.55 with all cuts enabled) and the corresponding model
has 62 constraints, and 18 binary variables. The integer solution had been
d

found in not significant time in a laptop with Intel Core i7-4500U CPU and
1.80 GHz of RAM.
te

The optimal scheduling found is presented in Table H-2 which leads to a

makespan of 136 units of time. Each element corresponds to the instant a
task on a bay had been finished for a corresponding quay crane. Figure H-1
p

shows the corresponding quay crane movement and the constraint of quay
crane 1 could overpass quay crane 2 is fulfilled.
ce

Furthermore, Figure H-1 shows two moments in scheduling that the over-
pass constraints leads to an increasing in the quay crane 1 processing time.
The first is when quay crane 1 finished process on bay 3 at the time 76, but
Ac

have to wait quay crane 2 finish its work on bay 5 at time 80. As a conse-
quence only after time 80, quay crane 1 and 2 could move to bay 6 and 7,
respectively. The processing on these two bays leads to second moment and
quay crane 1 had to wait two additional units of time to move to bay 8.
The allocation of quay cranes found with the proposed model is the same
found by [6] with only a difference in makespan caused by considering zero
time to perform a quay crane movement. Although, the proposed model on
Appendix F is more compact.

52

Page 54 of 60
 t
ip
cr
us
an
Fig. H-1. Optimal QC schedule with makespan equals to 136.

Table J-1. GA Parameters and Values.

Parameter Value
M
Population Size 10
Number of generations 1000
Crossover probability 0.8
d

Mutation probability 0.15

te

I. Small numerical example for Integrated 3D SP and SQC Model

Finally, tests had been carried for Integrated Mathematical Model for
p

3D SP and SQC which means the inclusion of two quay cranes to perform
unloading and loading operations in each port. The data is the same given
ce

on D.
The Integrated model (as given on Appendix F) with 796 rows and 1290
columns (960 binary variables) demanded more than 11 hours without re-
Ac

turning an integer optimal solution (the process stopped with a GAP of

23.9%).

J. Genetic Algorithm Parameters and Corresponding Values

The adopted parameters for Genetic Algorithm and its corresponding
values are given in Table J-1.
The crossover employed is the crossover OX. The mutation is simply the
random selection of a vector element and changes it to a random integer
number in the interval [1, N], where N is the number of ports.

53

Page 55 of 60
 Table K-1. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
1 11,250 05 1 9,792 273.23

t
ip
2 2 16,724 353.36
3 3 18,147 266.66
4 10 1 34,090 776.52

cr
5 2 19,703 910.71
6 3 65,577 915.01
7 15 1 46,741 1,471.31

us
8 2 25,413 900.49
9 3 113,948 1,276.74

an
Finally, tests had been carried for Integrated Mathematical Model for
3D SP and SQC which means the inclusion of two quay cranes to perform
unloading and loading operations in each port. The data is the same given
M
on Appendix D.

K. Detailed results from the new data set

In Table K-1, the column index I corresponds to instance number; S
d

corresponds to container ship total capacity; N corresponds to how many

te

ports the container ship has to pass through; the column index M refers to
the type of transportation matrix (1 - Mixed, 2 - Long, 3- Short; there are
more details on Appendix C); F O1 is the total time to perform all crane
p

movements (see subsection 2.2 for constraints and parameter details); T (s)
is the computational time spent in seconds to obtain the solution.
ce

The results presented from Table K-1 to Table K-11 were obtained with
a program created in a Matlab 7.0, a machine with a 1.66 GHz Core Duo
Intel Processor, RAM memory of 2 GB, and Windows Vista Operational
Ac

System with Service Pack 2. The genetic algorithm (more details about
parameters on Appendix J) was executed 5 times to illustrate the potential
of the developed approach.
The crossover employed is the crossover OX. The mutation is simply the
random selection of a vector element and changes it to a random integer
number in the interval [1, N], where N is the number of ports.
Finally, tests had been carried for Integrated Mathematical Model for
3D SP and SQC which means the inclusion of two quay cranes to perform
unloading and loading operations in each port.

54

Page 56 of 60
 Table K-2. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)

t
10 12,375 05 1 14,724 287.88

ip
11 2 16,440 340.59
12 3 38,337 448.42
13 10 1 33,646 783.76

cr
14 2 25,571 842.21
15 3 56,915 848.20

us
16 15 1 53,757 1,151.08
17 2 25,985 983.77
18 3 114,232 1,355.01

an
Table K-3. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
M
19 13500 05 1 16,333 305.02
20 2 21,655 399.52
21 3 45,001 433.35
22 10 1 41,613 795.61
d

23 2 25,588 776.74
24 3 90,586 1,002.22
te

25 15 1 46,318 1,179.21
26 2 23,397 985.71
p

27 3 136,163 1,651.72
ce

Table K-4. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
28 14625 05 1 15,943 334.15
Ac

29 2 23,936 405.13
30 3 41,769 493.24
31 10 1 37,831 865.11
32 2 21,867 767.56
33 3 95,953 1,114.09
34 15 1 63,692 1,309.32
35 2 25,591 1,170.09
36 3 146,765 1,812.58

55

Page 57 of 60
 Table K-5. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)

t
37 15750 05 1 22,214 367.77

ip
38 2 17,995 314.25
39 3 37,206 437.42
40 10 1 46,509 3,591.25

cr
41 2 25,325 819.16
42 3 98,872 1,218.65

us
43 15 1 67,543 1,438.74
44 2 30,361 1,500.57
45 3 167,953 1,881.28

an
Table K-6. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
M
46 16875 05 1 19,389 440.81
47 2 24,629 432.70
48 3 42,310 615.95
49 10 1 58,090 1,108.71
d

50 2 31,206 870.19
51 3 108,407 1,247.28
te

52 15 1 66,632 1,794.58
53 2 41,967 1,254.07
p

54 3 185,187 2,137.30
ce

Table K-7. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
55 18000 05 1 20,417 390.65
Ac

56 2 28,319 501.97
57 3 45,898 2,509.26
58 10 1 52,922 1,153.40
59 2 32,334 907.81
60 3 127,759 1,495.79
61 15 1 64,176 1,495.48
62 2 41,993 2,049.09
63 3 178,191 2,729.01

56

Page 58 of 60
 Table K-8. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)

t
64 19125 05 1 16,130 359.14

ip
65 2 26,812 370.56
66 3 55,575 646.24
67 10 1 61,650 1,118.45

cr
68 2 40,288 402.46
69 3 106,395 1,324.42

us
70 15 1 75,876 1,668.47
71 2 40,621 1,450.72
72 3 168,456 1,958.39

an
Table K-9. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
M
73 20250 05 1 20,101 153.20
74 2 30,161 557.02
75 3 64,474 269.33
76 10 1 67,197 1,302.42
d

77 2 45,722 1,045.01
78 3 129,923 1,586.77
te

79 15 1 83,589 1,742.81
80 2 42,966 1,823.37
p

81 3 173,287 2,353.86
ce

Table K-10. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
82 21375 05 1 27,632 508.25
Ac

83 2 35,050 642.47
84 3 59,906 624.90
85 10 1 63,059 1,573.77
86 2 45,667 1,167.78
87 3 118,192 1,661.03
88 15 1 72,384 2,096.61
89 2 59,110 1,808.01
90 3 192,097 2,774.22

57

Page 59 of 60
 t
ip
cr
us
an
Table K-11. Results obtained for GA when minimizing only total time.

I S N M FO1 T(s)
M
91 22500 05 1 27,564 487.22
92 2 30,278 677.47
93 3 50,194 860.48
94 10 1 68,137 1,456.33
d

95 2 53,442 1,271.27
96 3 155,321 1,787.77
te

97 15 1 66,788 2,002.10
98 2 44,935 1,971.58
p

99 3 212,738 2,701.42
ce
Ac

58

Page 60 of 60