Advanced Engineering Informatics 42 (2019) 100972

Contents lists available at ScienceDirect

Advanced Engineering Informatics

journal homepage: www.elsevier.com/locate/aei

Full length article

An augmented self-adaptive parameter control in evolutionary computation: T

A case study for the berth scheduling problem
Masoud Kavoosia, Maxim A. Dulebenetsb, , Olumide F. Abioyea, Junayed Pashaa, Hui Wangc,
⁎

Hongmei Chid
a
Department of Civil & Environmental Engineering, Florida A&M University-Florida State University, 2525 Pottsdamer Street, Building B, Suite B339, Tallahassee, FL
32310-6046, USA
b
Department of Civil & Environmental Engineering, Florida A&M University-Florida State University, 2525 Pottsdamer Street, Building A, Suite A124, Tallahassee, FL
32310-6046, USA
c
Department of Industrial & Manufacturing Engineering, Florida A&M University-Florida State University, 2525 Pottsdamer Street, Building B, Suite 373D, Tallahassee, FL
32310-6046, USA
d
Department of Computer and Information Sciences, Florida A&M University, 1333 Wahnish Way, 308-A Banneker Technical Bldg. 211-C, Tallahassee, FL 32307, USA

ARTICLE INFO ABSTRACT

Keywords: The demand for international seaborne trade has substantially increased over the last three decades and is
Marine transportation predicted to continue increasing during the upcoming years. A marine container terminal, as an important node
Marine container terminals in supply chains, should be able to successfully cope with increasing demand volumes. Berth scheduling can
Berth scheduling significantly influence the general throughput of marine container terminals. In this study, a mixed-integer linear
Optimization
programming mathematical model is proposed for the berth scheduling problem, aiming to minimize the
Evolutionary algorithms
summation of waiting costs, handling costs, and late departure costs of the vessels that are to be served at a
Parameter control
Parameter tuning marine container terminal. An innovative Evolutionary Algorithm is designed to solve the developed mathe-
matical model. The proposed solution algorithm relies on the augmented self-adaptive parameter control
strategy, which is developed in order to effectively change the algorithmic parameters throughout the search
process. Performance of the designed algorithm is evaluated against nine alternative state-of-the-art meta-
heuristic-based algorithms, which have been frequently used for berth scheduling in the marine container
terminal operations literature. The results demonstrate that all the developed algorithms have a high level of
stability and return competitive solutions at convergence. The computational experiments also prove superiority
of the designed augmented self-adaptive Evolutionary Algorithm over the alternative algorithms in terms of
different performance indicators.

1. Introduction
Waterborne transportation plays an important role for the global
trade. Based on the statistical information, released by United Nations
Conference on Trade and Development (UNCTAD) in 2017, about 80%
of the total volume or 70% of the total value of the global trade was
carried by means of waterborne transportation. Furthermore, UNCTAD
expects that the waterborne trade will increase by 3.2% per year from
2017 to 2022 [1]. Hence, marine container terminal (MCT) operators
should expect a continuously increasing demand for waterborne
transportation and plan their operations accordingly. An efficient
management of container handling processes at MCTs is one of the most
promising alternatives to serve the growing demand in this industry.
The container handling operations that take place at a typical MCT can
be classified as: (1) seaside operations; (2) marshaling yard operations;
and (3) landside operations. This study focuses on the seaside opera-
tions at the MCT, which include mooring of the arriving vessels at the
assigned berthing positions and loading/unloading containers to/from
the vessels using the quay cranes, which are installed alongside the
wharf. More specifically, this study addresses the berth scheduling
problem (BSP), which aims to assign each arriving vessel for service at
one of the available berthing positions along the MCT wharf as well as
determine the order of vessel service at each available berthing posi-
tion.
BSPs are classified as non-deterministic polynomial time hard (NP-
hard) problems, as the large-size instances of BSPs cannot be solved

Corresponding author.
⁎

E-mail addresses: mk17@my.fsu.edu (M. Kavoosi), mdulebenets@eng.famu.fsu.edu (M.A. Dulebenets), olumide1.abioye@famu.edu (O.F. Abioye),
jp17j@my.fsu.edu (J. Pasha), hwang10@eng.famu.fsu.edu (H. Wang), hongmei.chi@famu.edu (H. Chi).

https://doi.org/10.1016/j.aei.2019.100972
Received 10 March 2019; Received in revised form 22 June 2019; Accepted 30 July 2019
Available online 17 August 2019
1474-0346/ © 2019 Elsevier Ltd. All rights reserved.
M. Kavoosi, et al.
using the exact optimization algorithms in a reasonable computational
time. It has been proven in the BSP and operations scheduling literature
that EA-based algorithms (EAs) can return promising solutions for the
large-size problem instances within an acceptable computational time
[2–7]. However, determining an appropriate set of the algorithmic
parameter values has always been a challenging issue in the process of
developing an EA. Generally, parameter selection approaches can be
classified in two major groups [8,9]: (1) parameter tuning; and (2)
parameter control. The parameter tuning is a conventional method,
which is based on a massive set of computational experiments. More
specifically, several values (selected from the known ranges, reported in
the EA literature, and preliminary algorithmic runs) are assigned to all
the algorithmic parameters, and the algorithm is executed for different
combinations of the selected parameter values. Finally, the combination
of the algorithmic parameter values, which returns high-quality solu-
tions within a reasonable computational time, is selected for the de-
veloped algorithm.
Based on the MCT operations and EA literature [9,10], the para-
meter control strategies can be classified as: (1) deterministic; (2)
adaptive; and (3) self-adaptive. The deterministic parameter control
strategy updates the algorithmic parameter values based on a defined
counter (e.g., specific number of generations, specific number of fitness
function evaluations, target computational time), where the feedback
from the search (e.g., the changing pattern of the fitness function va-
lues) is not considered in the process of updating the algorithmic
parameter values. In order to reflect the results of the search in the
process of selecting the parameter values, the adaptive parameter
control was designed, where the algorithmic parameter values are
changed based on the feedback from the search (e.g., changes in the
fitness function values). As for the self-adaptive parameter control
strategy, the algorithmic parameter values are encoded in the solutions
and evolve with the algorithm without any feedback from the search.
Although the parameter control strategies were found to be more
promising for the search process as compared to a typical parameter
tuning strategy [9,11,12], such strategies have some drawbacks too.
More specifically, the deterministic parameter control strategy changes
the algorithmic parameter values without taking into consideration the
changing pattern of the fitness function values, which may not be de-
sirable for the search process. The adaptive parameter control strategy
changes the algorithmic parameter values based on the feedback from
the search, however, it may be difficult to establish an appropriate
adaptive criterion (i.e., set a threshold for the fitness function changes
that will trigger changes in the algorithmic parameter values). More-
over, for the decision problems with a complex search space (e.g.,
“deceptive problems” [13]), the algorithmic parameter values might
have to be changed more/less frequently than based on the established
adaptive criterion. The self-adaptive parameter control strategy allows
updating the algorithmic parameter values more frequently, as they are
encoded in the population chromosomes and evolve along with the
algorithm; however, the self-adaptive parameter control strategy does
not consider any feedback from the search.
In this study, a new approach for parameter control, named as
augmented self-adaptive, is proposed in order to address the drawbacks
of the existing parameter control strategies. Based on the proposed
augmented self-adaptive parameter control strategy, not only the al-
gorithmic parameter values are encoded in the solutions, but they are
also periodically updated based on the feedback from the search. A
mixed-integer linear mathematical model is proposed for the BSP,
where the objective aims to minimize the total vessel service cost. An
augmented self-adaptive EA (ASAEA) is applied to solve the mathe-
matical model. A comprehensive series of numerical experiments are
conducted to assess performance of the proposed ASAEA algorithm.
Specifically, ASAEA is evaluated against the alternative EAs, which rely
on the existing self-adaptive parameter control, adaptive parameter
control, deterministic parameter control, and parameter tuning strate-
gies. Furthermore, the ASAEA algorithm is compared against a group of
2
Advanced Engineering Informatics 42 (2019) 100972
other state-of-the-art metaheuristic-based algorithms, which have been
frequently used for berth scheduling in the MCT operations literature,
including Tabu Search, Simulated Annealing, Variable Neighborhood
Search, Ant Colony Optimization, and Particle Swarm Optimization.
The remainder of this manuscript is organized as follows. In the
second section, the related BSP studies are discussed. The problem of
interest is described in detail in the third section. The model formula-
tion is presented in the fourth section. The main algorithmic steps of the
ASAEA algorithm are explained in the fifth section. The conducted
numerical experiments are presented in the sixth section. The study
conclusions and the potential future research works are described in the
seventh section.
2. Literature review
MCTs, as an infrastructure for intermodal change of cargoes be-
tween different types of transportation systems, have attracted a sig-
nificant attention from the researchers during the last years [14].
Fransoo and Lee [4], Bierwirth and Meisel [5], and Lee and Song [6]
presented a comprehensive review of the studies related to the MCT
operations and waterborne logistics. Generally, BSPs can be classified
from different operational perspectives. Based on configuration of the
MCT layout for berthing vessels, BSPs can be classified as: (1) discrete –
the wharf is divided into several berthing positions and only one vessel
can be served at each berthing position at a time; (2) continuous – the
wharf is not partitioned into separate berthing positions and the ar-
riving vessels can moor at arbitrary positions within the boundaries of
the wharf; and (3) hybrid – a large vessel can occupy more than one
berthing position or a number of small vessels can be served at one
berthing position [14]. The literature review in this study focuses on
discrete berth scheduling problems (DBSPs). Such terminal configura-
tion was found to be the most common in the BSP literature [5,14]. The
collected DBSP studies have been divided into the following categories:
(1) green DBSPs; (2) uncertainty modeling in DBSPs; (3) collaborative
agreements; and (4) general DBSPs – the studies that did not fall under
categories (1)–(3). The following sections present a description of the
collected studies. For a more detailed review of the BSP literature this
study refers to Bierwirth and Meisel [5] and Carlo et al. [14].
2.1. Green DBSPs
Environmental issues are among the aspects of the BSP literature,
which have attracted a lot of attention from the scientists over the past
decade. Lang and Veenstra [15], Golias et al. [16], Du et al. [17], and
Wang et al. [18] directly accounted for the fuel consumption by vessels
in the presented BSP models. Such models can be extended towards
estimation of the emissions, produced by vessels, and quantification of
negative environmental effects. Venturini et al. [19] studied the BSP,
considering a strong cooperation between the ports and liner shipping
companies. The objective of the presented mathematical model mini-
mized the total cost, including the idle time cost, the operational cost,
the delay cost, and the total fuel cost. CPLEX was adopted as a solution
approach. The numerical experiments demonstrated that selection of
the appropriate vessel sailing speed discretization level could sig-
nificantly improve the results from both economic and environmental
perspectives. Dulebenets et al. [20] proposed an innovative BSP model
for minimizing the total service cost of vessels, including the carbon
dioxide emission cost due to container handling. A set of hybrid EAs
were developed to solve the proposed mathematical model. Throughout
the computational experiments, it was found that the unit carbon di-
oxide emission cost could substantially affect the design of berth
schedules.
A number of other studies modeled the emissions, produced by the
handling equipment, and/or the energy consumption by the handling
equipment throughout service of vessels at MCTs. For example, Chang
et al. [21] adopted a rolling-horizon approach for the berth allocation
M. Kavoosi, et al.
and quay crane assignment problem. A total of three objective functions
were optimized, including the following: (1) minimize the total devia-
tion from the desired berthing positions; (2) minimize the total cost
associated with late berthing and departure time of vessels; and (3)
minimize the total energy consumption by quay cranes. A hybrid par-
allel EA, which combined the features of a parallel EA and a heuristic
algorithm, was presented to solve the problem. It was found that an
integration of a parallel EA and a heuristic algorithm improved the
quality of the solutions, which were produced by the developed solu-
tion approach. Li et al. [22] developed a multi-objective optimization
model to minimize the total turnaround time of the arriving vessels and
the total travel distance by the yard trucks due to potential deviations
from the desired berthing positions. An improved Particle Swarm Op-
timization (PSO) was adopted to solve the problem. The conducted
numerical experiments showed that application of the proposed meth-
odology could increase the overall utilization rates of the available MCT
handling equipment, reduce the fuel costs and the associated environ-
mental pollution. Different types of emissions, produced by the internal
transport vehicles and the outbound trucks at MCTs, as well as energy
consumption, were also captured by the mathematical models, pro-
posed by Esmer et al. [23], He et al. [24–26], and Chen et al. [27].
2.2. Uncertainty modeling in DBSPs
Some BSP literature accounted for various sources of uncertainty in
the proposed BSP mathematical formulations. Golias et al. [28] focused
on modeling uncertainty in the vessel arrival times and handling times
throughout berth scheduling. An EA-based heuristic algorithm was
adopted to minimize the average total vessel turnaround time and the
total vessel turnaround time range. Findings demonstrated that the
proposed approach could provide more robust berth schedules, espe-
cially under significant congestion effects. Ursavas and Zhu [29] stu-
died the BSP, taking into consideration the uncertain nature of the
vessel arrival and handling times. A framework, inspired by the sto-
chastic dynamic programming approach, was presented to model the
problem and characterize the optimal policies. It was found that the
proposed methodology was promising and could assist the MCT op-
erators with berth scheduling under uncertain vessel arrival and
handling times. Furthermore, the developed framework could be ex-
tended towards modeling uncertainty for the storage yard allocation
problem.
Umang et al. [30] focused on berth scheduling at the MCT, taking
into account uncertainty in the vessel arrival times and handling times.
The objective of the proposed mathematical model aimed to minimize
the total vessel service cost. An Optimization-based Recovery Algorithm
was adopted to solve the problem. The algorithm was based on set
partitioning and a smart greedy algorithm, which re-assigned the ves-
sels in case of a disruptive event. The computational experiments were
performed using the operational data, collected from the SAQR Port
(United Arab Emirates). It was found that the proposed methodology
could significantly reduce a potential increase of the total vessel service
cost due to uncertainty in the vessel arrival times and handling times as
compared to the current practices, which are adopted at the port.
Xiang et al. [31] proposed a mathematical model for the integrated
berth scheduling and quay crane assignment, considering various
sources of uncertainty (e.g., uncertain vessel arrival times, uncertain
vessel handling times, calling of unscheduled vessels, quay crane
breakdowns). The objective function of the proposed model aimed to
minimize the total baseline cost, associated with deviation from the
desired berthing positions, waiting time, late vessel departures, and
quay crane set up. The reactive strategy was proposed, which con-
sidered the baseline schedule as a reference, in order to minimize the
recovery cost. A rolling-horizon heuristic was used to solve the pro-
blem. The numerical experiments indicated that the proposed solution
approach was able to solve the considered problem instances to the
global optimality. A number of other studies captured uncertainty in
3
Advanced Engineering Informatics 42 (2019) 100972
the vessel arrival times [32,33] and uncertainty in both vessel arrival
times and vessel handling times [34,35] throughout berth scheduling as
well as proposed solution approaches for developing robust berth
schedules.
2.3. Collaborative agreements
Certain studies specifically focused on contractual agreements be-
tween liner shipping companies, including Alix et al. [36], Ding and
Liang [37], Lei et al. [38], Panayides and Wiedmer [39], and Yang et al.
[40]. A number of studies aimed to model the effects of contractual
agreements between the MCT operators. For example, Imai et al. [41]
focused on the BSP at a multi-user MCT, assuming that the vessels could
be diverted for service at an external MCT in case of excessive waiting
times. The presented mathematical model aimed to minimize the total
vessel service time at the external MCT. The authors proposed an EA-
based heuristic to solve the problem. The proposed policy showcased
significant advantages, especially during the peak hours.
Peng et al. [42] studied a BSP at bulk marine terminals, assuming
that the berthing space and capacity of the storage yard could be shared
between different terminals. The proposed mathematical model mini-
mized the total vessel waiting time, the total vessel handling time, and
the total storage yard assignment variations. An EA was developed to
solve the problem. The computational experiments indicated that the
proposed EA was able to return good-quality solutions within a rea-
sonable computational time. Dulebenets et al. [43] proposed a berth
scheduling policy, where the demand could be diverted from a multi-
user MCT to an external MCT at an additional cost. A Memetic Algo-
rithm was developed to minimize the total vessel service cost, including
the cost of serving the diverted demand. The conducted numerical ex-
periments confirmed that the suggested form of a collaborative agree-
ment between the MCT operators could yield remarkable cost savings
during the peak demand periods. A few studies evaluated contractual
agreements between liner shipping companies and MCT operators (see,
for example, Golias and Haralambides [44] and Imai et al. [45]).
2.4. General DBSPs
The collected studies, which did not fall under the aforementioned
study categories, are described in this section of the manuscript. Arango
et al. [46] focused on the BSP, aiming to minimize the total vessel
service time. The problem was solved using a hybrid approach, which
was based on a simulation model and an EA-based algorithm. A case
study, conducted for the Port of Seville (Spain), showcased that the
proposed methodology was superior to the existing berth management
strategy. De Oliveira et al. [47] developed a Clustering Search (CS)
method for the BSP, where the solutions were generated using Simu-
lated Annealing (SA). The objective of the mathematical formulation
minimized the total vessel service cost. The numerical experiments
showed that the developed solution algorithm outperformed the alter-
native algorithms, which have been previously used in the BSP litera-
ture. Cubillos et al. [48] developed an agent-based model in order to
schedule the arriving vessels among the available berthing positions at
the MCT. Different agents were responsible for particular functions,
including creation of the vessel service requests for the berthing posi-
tions, stowage planning for the vessels, and others. The developed de-
cision support system was validated by considering different scenarios,
ranging from maximizing the berth utilization to maximizing the vessel
throughout.
Many ports are subject to the tidal effect, where the depth of the
access channel and the available berthing positions may change by time
of day. In some cases, a given vessel is not able to safely navigate as a
result of the tidal effect and has to wait in the dedicated area of the port
until the depth of the access channel and the assigned berthing position
becomes adequate. The latter important operational feature has been
captured by some of the previously conducted BSP studies [49–52].
 Table 1
Summary of the reviewed DBSP literature.
Author(s) Year Objective Solution approach Notes/important considerations
M. Kavoosi, et al.

Green DBSPs
Lang and Veenstra [15] 2010 Minimize the total cost Simulation + CPLEX Up to 16 vessels were considered due to CPU time issues
Golias et al. [16] 2010 Minimize the total vessel service time and emissions EA Basic EA algorithm
Du et al. [17] 2011 Minimize the total delay and the total fuel consumption ε-constraint method, CPLEX The proposed solution approach was applied for the cases with up to 28 vessels
Wang et al. [18] 2013 Minimize the total delay cost and the total fuel cost CPLEX Proposed an alternative approach for addressing non-linearity of the fuel consumption
function
Venturini et al. [19] 2017 Minimize the total cost CPLEX The proposed solution approach was applied for the cases with up to 20 vessels
Dulebenets et al. [20] 2017 Minimize the total cost Hybrid EAs Local search heuristics improved the solution quality without significantly impacting the
CPU time
Chang et al. [21] 2010 Minimize berthing deviations, the total cost, the total energy Hybrid Parallel EA Integration of a parallel EA and a heuristic algorithm improved the solution quality
consumption by quay cranes
Li et al. [22] 2017 Minimize the total vessel turnaround time and the total yard truck Hybrid PSO Local search heuristics improved the solution quality but increased the CPU time
travel distance
Esmer et al. [23] 2010 Minimize emissions from yard trucks Simulation Computational efficiency of the solution approach was not evaluated
He et al. [24] 2013 Minimize the overflowed workloads and the total transferring cost Simulation + EA The proposed algorithm was more effective as compared to CPLEX for the large-size
problem instances
He et al. [25] 2015 Minimize the total delay and the total energy consumption Simulation + EA + PSO The proposed algorithm was more effective as compared to CPLEX for the large-size
problem instances
He et al. [26] 2015 Minimize the total delay and the total energy consumption Simulation + EA + PSO The proposed algorithm was more effective as compared to CPLEX for the large-size
problem instances
Chen et al. [27] 2013 Minimize the total truck waiting time and arrival pattern changes EA The proposed algorithm outperformed NSGA-II in terms of the solution quality but was
worse in terms of the CPU time

Uncertainty Modeling in DBSPs

Golias et al. [28] 2014 Minimize the average total vessel turnaround time and the total vessel Hybrid EA Local search heuristics improved the CPU time without significantly impacting the

4
Ursavas and Zhu [29]
Umang et al. [30]
Xiang et al. [31]
Moorthy and Teo [32]
Hendriks et al. [33]
Xu et al. [34]
Zhen and Chang [35]
turnaround time range solution quality
2016 Minimize the total cost Stochastic dynamic programming Computational efficiency of the solution approach was not evaluated
2017 Minimize the total cost Heuristic The proposed algorithm was more effective as compared to the alternative algorithms
2018 Minimize the total cost Heuristic The proposed algorithm was more effective as compared to CPLEX for the large-size
problem instances
2007 Minimize the total cost Heuristic Computational efficiency of the solution approach was not evaluated
2010 Minimize the maximum quay crane reservation capacity Exact optimization Computational efficiency of the solution approach was not evaluated
2012 Minimize the total delay and maximize the buffer time Hybrid SA The proposed algorithm was more effective as compared to LINGO for the large-size
problem instances
2012 Minimize the total cost and maximize the robustness Heuristic Computational efficiency of the solution approach was not evaluated

Collaborative Agreements
Imai et al. [41] 2008 Minimize the total vessel service time EA Computational efficiency of the solution approach was not evaluated
Peng et al. [42] 2015 Minimize the total vessel turnaround time and the total storage yard EA The proposed algorithm was more effective as compared to CPLEX for the large-size
assignment variations problem instances
Dulebenets et al. [43] 2018 Minimize the total cost Hybrid EA Local search heuristics improved the solution quality without significantly impacting the
CPU time
Golias and Haralambides [44] 2011 Minimize the total cost EA Computational efficiency of the solution approach was not evaluated
Imai et al. [45] 2014 Minimize the total vessel turnaround time, berth idle time, penalty Heuristic The proposed algorithm was more effective as compared to the alternative algorithms
cost

General DBSPs
Arango et al. [46] 2011 Minimize the total vessel service time Simulation + EA Validation results demonstrated accuracy of the model
De Oliveira et al. [47] 2012 Minimize the total cost CS + SA The proposed algorithm was more effective as compared to the alternative algorithms
Cubillos et al. [48] 2013 Maximize the berth utilization and maximize the vessel throughout Simulation Computational efficiency of the solution approach was not evaluated
Xu et al. [49] 2012 Minimize the total cost Heuristic The proposed algorithm was more effective as compared to CPLEX for the large-size
problem instances
Lalla-Ruiz et al. [50] 2016 Minimize the total weighted vessel turnaround time Heuristic The proposed algorithm was more effective as compared to the alternative algorithms
Dadashi et al [51] 2017 Minimize the total weighted delay CPLEX CPLEX was applied for the cases with up to 39 vessels
Zhen et al. [52] 2017 Minimize the total weighted waiting time and delay Heuristic The proposed algorithm was more effective as compared to the alternative algorithms
Ting et al. [53] 2014 Minimize the total vessel turnaround time PSO The proposed algorithm was more effective as compared to the alternative algorithms
Advanced Engineering Informatics 42 (2019) 100972

(continued on next page)

M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Ting et al. [53] formulated a mixed-integer programming model for the

BSP to minimize the total vessel turnaround time. A PSO-based algo-
The proposed algorithm was more effective as compared to CPLEX for the large-size
rithm was adopted to solve the problem. The proposed PSO algorithm
was found to be superior to CPLEX, CS, Tabu Search (TS), and popu-
lation training algorithm with linear programming in terms of the so-
lution quality and the computational time. Emde and Boysen [54]
proposed a mathematical model for the MCT, handling deep-sea vessels
and service feeder vessels. The objective of the model aimed to mini-
mize the total weighted vessel waiting time and the total number of
delayed containers. An SA-based algorithm was developed to solve the
problem. It was found that the developed solution algorithm out-
performed other real-world priority rule-based approaches.
A number of studies focused on evaluation of various parameter
selection approaches for the EA-based algorithms, which were devel-
oped to solve the BSP. Dulebenets [11] developed a mixed-integer
nonlinear mathematical model for the BSP, aiming to minimize the
Deterministic parameter control
Notes/important considerations

Self-adaptive parameter control

total vessel service cost. A hybrid EA with a deterministic parameter

Adaptive parameter control

control was designed to solve the proposed model, where the mutation
rate was changed using a piecewise function. The computational ex-
periments showed that the developed solution approach returned su-
problem instances

perior objective function values as compared to the Variable Neigh-

borhood Search (VNS), typical EA, and SA algorithms. Dulebenets [12]
developed an EA-based algorithm with an adaptive parameter control
strategy for the BSP at the MCT, where the mutation rate was changed
by the algorithm based on the feedback from the search. The total
weighted vessel service cost was minimized in the developed mathe-
matical model. A series of numerical experiments indicated that the
adaptive parameter control strategy could yield more promising solu-
tions as compared to a typical EA without substantially affecting the
computational time. Dulebenets et al. [9] proposed a self-adaptive EA
Solution approach

for the BSP, where the crossover and mutation probabilities were en-
coded in the EA chromosomes. The model aimed to minimize the total
Hybrid EA
Hybrid EA
Hybrid EA

weighted vessel turnaround time and the total weighted vessel late
departures. The adopted parameter control strategy was found to be
SA

superior to other parameter selection approaches, which were pre-

viously used in the BSP literature.
Minimize the total weighted vessel waiting time and the number of

2.5. Literature summary and contributions

Minimize the total weighted vessel turnaround time and late

The results of a conducted DBSP literature review are summarized

in Table 1, highlighting the following aspects: (1) author(s); (2) year;
(3) objective; (4) solution approach; and (5) notes/important con-
siderations (related to the developed solution approach). Many studies
underlined a high BSP computational complexity and inability of the
exact optimization approaches to provide the optimal solutions in a
reasonable computational time for the large-size problem instances
Minimize the total weighted cost

[9,11,12,15,19,20,24–26,31,34,42,43,49,51,54]. Several studies

showed that introduction of local search heuristics within the proposed
Minimize the total cost

solution algorithms generally improves the solution quality at con-

delayed containers

vergence but incurs additional computational time

[9,11,12,20–22,28,43]. The majority of the solution algorithms, pro-
posed for DBSPs, rely on the parameter tuning strategy, where the al-
departures
Objective

gorithmic parameter values remain unchanged throughout the search

process. Several recent DBSP studies presented the EA-based algorithms
with deterministic parameter control, adaptive parameter control, and
2016

2017
2018
2018

self-adaptive parameter control in order to solve DBSPs [9,11,12]. It

Year

was found that the parameter tuning strategy and the aforementioned
parameter control strategies have a number of drawbacks. A self-
adaptive parameter control strategy, which was found to be superior to
the adaptive parameter control, deterministic parameter control, and
Emde and Boysen [54]

Dulebenets et al. [9]

Table 1 (continued)

parameter tuning strategies, does not consider any feedback from the
Dulebenets [11]
Dulebenets [12]

search [9].
In order to address the drawbacks of the existing EA parameter
Author(s)

selection strategies, this study proposes an augmented self-adaptive

parameter control approach, which not only encodes the major algo-
rithmic parameters in the chromosomes and allows evolving these

5
M. Kavoosi, et al.
parameters throughout the search process, but also periodically
changes the parameter values based on the feedback from the search.
The proposed solution algorithm will be evaluated against the exact
optimization algorithm and other nine alternative state-of-the-art me-
taheuristic-based algorithms, which have been frequently used for berth
scheduling in the MCT operations literature.
3. Detailed problem description
The main focus of this study is to model the seaside operations at the
MCT, where the containers have to be processed for the vessels that are
managed by different liner shipping companies. A set of vessels, ar-
riving for service at the MCT, will be referred to as V = {1, , m} . Based
on a detailed review of the BSP literature, a large variety of different
berthing layouts were reported, such as discrete, continuous, indented,
hybrid, and channel berthing layouts [5]. The MCT, which is modeled
in this study, is assumed to have a discrete berthing layout that has been
widely used in the previous BSP research efforts. A set of MCT berthing
positions available for vessel service will be referred to as B = {1, , n} .
A dynamic vessel arrival pattern is adopted in this study, where the
vessels are not present at the MCT immediately at the beginning of a
planning horizon; however, the MCT operator expects the vessels to
arrive at certain times, which have been previously negotiated with the
corresponding liner shipping companies. The vessel arrival times will
be further referred to as t va, v V (measured in hours). Note that
uncertainty in the arrival time of vessels due to unforeseen events (e.g.,
extreme weather conditions, vessel technical issues, human errors, etc.)
is not explicitly captured in this study.
Once a vessel arrives at the MCT, a set of push boats tow it to the
predetermined berthing position. However, if the predetermined
berthing position cannot accommodate the vessel temporarily (e.g.,
another vessel is being served at that berthing position, certain main-
tenance activities are conducted at that berthing position), the vessel
will be towed to a specific waiting area, which is located near the MCT
(see Fig. 1). The MCT operator is assumed to incur the vessel waiting
cost c vwt , v V (measured in USD/hour). An increasing number of
vessels in the waiting area is not desirable, as it may lead to congestion
in the seaside area and cause difficulties for navigation of vessels. It
should be noted that a given vessel cannot moor at the assigned
berthing position sooner than the berthing position becomes available
in the considered planning horizon (tbber , b B – measured in hours).
The MCT operator typically assigns a berthing position to each one
of the arriving vessels in advance (berthing plans are commonly de-
signed on weekly/bi-weekly basis). In order to assign a berthing posi-
tion to a vessel, there are four key factors that have to be considered by
the MCT operator, including: (1) spatial requirements; (2) negotiated
handling productivity; (3) available handling equipment; and (4) access
Fig. 1. Illustration of the MCT layout and basic MCT seaside operations.
6
Advanced Engineering Informatics 42 (2019) 100972
to the intended storage space in the marshaling yard. As for the spatial
requirements, a berthing position with a length of lbber , b B (mea-
sured in feet) and a depth of hbber , b B (measured in feet) should be
sufficient in order to accommodate a given vessel with a length of
l vves , v V (measured in feet) and a draft of h vves , v V (measured in
feet). The minimum horizontal clearance (d vhor , v V – measured in
feet) and the minimum vertical clearance (d vver , v V – measured in
feet) should be considered by the MCT operator throughout the berth
scheduling to ensure that the vessels will be able to safely moor at the
assigned berthing positions.
The handling productivity, which defines the number of loaded and
unloaded twenty foot equivalent units (TEUs) per hour for each vessel,
is negotiated between the MCT operator and a given liner shipping
company. The handling productivity will be further referred to as
pvb , v V , b B (measured in TEUs/hour). The vessel handling
time (t vb ht
, v V , b B – measured in hours) can be computed based
on the total number of containers (qv , v V – measured in TEUs) to be
handled for a given vessel and the associated handling productivity as
q
follows: t vb ht
= pv v V , b B . Based on the negotiated handling
vb
productivity, the MCT operator has to utilize the available container
handling recourses accordingly and allocate a sufficient amount of
handling equipment to each vessel. The following types of handling
equipment are primarily used for container handling and transfer at the
MCT: (1) quay cranes for loading/unloading the containers on/from the
vessels; (2) internal transport vehicles (e.g., straddle carriers, auto-
mated guided vehicles, yard trucks, automated lifting vehicles) for
transporting containers between the berthing positions and the storage
areas in the marshaling yard; and (3) gantry cranes for handling and
reshuffling containers in the marshaling yard. Based on the handling
rate requested for a given vessel, the MCT operator will endure the
handling cost c vht , v V (measured in USD/hour).
Given the expected arrival time of a vessel at the MCT and the
handling rate, which was negotiated with the corresponding liner
shipping company, the MCT operator has to ensure that the service of a
vessel will be completed in a timely manner (i.e., on or before the re-
quested departure time t vd, v V – measured in hours). Potential de-
lays in vessel service (e.g., due to MCT congestion, handling equipment
breakdowns, inefficient scheduling of the available MCT handling
equipment) can significantly disrupt liner shipping schedules, as the
vessels may not be able to arrive within the agreed time windows to the
next ports of call. The MCT operator is assumed to endure the late
departure cost c vlt , v V (measured in USD/hour) in case of a delayed
vessel service completion. The overall objective of the MCT operator is
to develop such berth schedule, which will allow minimizing the total
cost that is associated with service of the arriving vessels. More speci-
fically, the MCT operator aims to minimize the total vessel service cost,
including the total vessel waiting cost, the total vessel handling cost,
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Table 2
The main components of the proposed mathematical model.
Model Component Description

Type Nomenclature

Sets V = {1, , m} set of vessels arriving for service at the MCT (vessels)
B = {1, , n} set of the MCT berthing positions available for vessel service (berthing positions)

Decision Variables x vb , v V , b B =1 if vessel v is assigned for service at berthing position b (=0 otherwise)
y vv¯ , v, v¯ V , =1 if vessel v̄ is assigned for service immediately after vessel v at the same berthing position (=0 otherwise)
_ _ _

v v̄
_
fv , v V =1 if vessel v is assigned for service as the first vessel at a given berthing position (=0 otherwise)
lv , v V =1 if vessel v is assigned for service as the last vessel at a given berthing position (=0 otherwise)

Auxiliary Variables st
v , v V service start time for vessel v (hours)
ft service finish time for vessel v (hours)
v , v V
wt
v , v V waiting time of vessel v (hours)
lt
v, v V late departure time of vessel v (hours)

Parameters m number of vessels arriving for service at the MCT (vessels)

n number of the MCT berthing positions available for vessel service (berthing positions)
t va, v V arrival time of vessel v at the MCT (hours)
tbber , b B time when berthing position b is available at the first time in the considered planning horizon (hours)
qv , v V number of containers to be handled for vessel v (TEUs)
pvb , v V, b B handling productivity that can be provided by the MCT operator for vessel v at berthing position b (TEUs/hour)
ht ,
t vb v V, b B handling time of vessel v at berthing position b (hours)
t vd, v V requested departure time for vessel v (hours)
h vves , v V draft of vessel v (feet)
hbber , b B depth of berthing position b (feet)
d vver , v V minimum vertical clearance for vessel v (feet)
l vves , v V length of vessel v (feet)
lbber , b B length of berthing position b (feet)
d vhor , v V minimum horizontal clearance for vessel v (feet)
c vwt , v V waiting cost for vessel v (USD/hour)
c vht , v V handling cost for vessel v (USD/hour)
c vlt, v V late departure cost for vessel v (USD/hour)
K large positive number

and the total vessel late departure cost. Constraint set (4) guarantees that each vessel will be moored at the
assigned MCT berthing position without violation of the minimum
4. Model formulation horizontal clearance requirement.

(l vves + d vhor ) x vb lbber v V, b B (4)

This section of the manuscript provides a detailed description of the
nomenclature, adopted throughout this study, and a mixed-integer Constraint set (5) ensures that only one vessel can be assigned for
linear mathematical formulation for the discrete berth scheduling service as the first vessel at each MCT berthing position.
problem with dynamic vessel arrivals and spatial restrictions (DBSP). f v + fv¯ 3 x vb x vb v, v¯ V, v v¯, b B
¯
The DBSP mathematical model components are described in Table 2. _ _ _ _ (5)
The objective function (1) of the DBSP mathematical model, which
Constraint set (6) ensures that only one vessel can be assigned for
will be further denoted as TC , minimizes the total cost of serving the
service as the last vessel at each MCT berthing position.
vessels arriving at the MCT. The total vessel service cost includes a total
of three components: (1) total vessel waiting cost; (2) total vessel l v + l v¯ 3 x vb x vb
¯ v , v¯
_
V, v
_
v¯ , b B
(6)
_ _
handling cost; and (3) total vessel late departure cost.
Constraint set (7) indicates that each vessel can be either assigned
wt wt ht
min TC = [ ( v cv ) + (t vb x vb c vht ) + ( lt lt
v c v )] for service as the first vessel at a given MCT berthing position or served
v V v V b B v V

(1) after another vessel.

Constraint set (2) indicates that each vessel, arriving for service at fv¯ + y vv¯ = 1 v¯ V
the MCT, will be assigned to one of the available MCT berthing posi- _
v V : v v¯
_
_
(7)
tions.
Constraint set (8) indicates that each vessel can be either assigned
x vb = 1 v V for service as the last vessel at a given MCT berthing position or served
b B (2) before another vessel.
Constraint set (3) guarantees that each vessel will be moored at the lv+ y vv¯ = 1 v V
assigned MCT berthing position without violation of the minimum _
v¯ V : v¯ v _ _
vertical clearance requirement. _ (8)
Constraint set (9) ensures that service of each vessel at the assigned
(h vves + d vver ) x vb h vber v V, b B (3)
MCT berthing position can start only once that berthing position

7
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

becomes available for service. Start

st
v tbber x vb v V
b B (9) Input problem and
algorithm data

Constraint set (10) guarantees that service of each vessel at the

assigned MCT berthing position can start only after its arrival at the
Create the first
population and AlgPa
MCT.
Repair the first
st
v t va v V (10) population

Constraint set (11) estimates the handling time of each vessel at the
Evaluate the first
available MCT berthing positions. population

ht qv
t vb = v V, b B
pvb (11) Stopping criterion YES Return the best
is met? solution
Constraint set (12) calculates the service start time for each vessel
based on the service start time and the handling time of the vessel, NO
End
which is assigned for service before that vessel at the same MCT Select parents
berthing position.
st st
v¯ v + (t ht
vb x vb ) K (1 y vv¯ ) v , v¯ V, v v¯ Update AlgPa
_ b B _ _ _ _ _ (12) Criterion of AlgPa YES
and apply to
update is met?
selected parents
Constraint set (13) computes the service finish time for each vessel
at the assigned MCT berthing position.
NO

Produce
ft st ht offspring
v v + (t vb x vb ) v V, b B
b B (13)
Repair offspring
Constraint set (14) estimates the waiting time of each vessel before
its service start at the assigned MCT berthing position.
Evaluate
wt
v
st
v t va v V (14) offspring

Constraint set (15) calculates the late departure time for each vessel Select offspring for
the next generation
assigned for service at one of the MCT berthing positions.
lt
v v
ft
t vd v V (15) Fig. 2. Main steps of ASAEA.
Constraint sets (16)–(18) showcase the ranges of decision variables,
auxiliary variables, and parameters, which are used in the proposed generated, and a separate data structure (denoted as “AlgPa” in Fig. 2)
DBSP mathematical model. is created for storing the most promising crossover and mutation
x vb , y vv¯ , fv , l v {0, 1} v V, v, v¯ V, v v¯, b B probability values. Next, ASAEA repairs the infeasible individuals of the
_ _ _ (16) initial population. After evaluating fitness of the initial population, if
the stopping criterion is met, the algorithm returns the best solution,
m , n , qv N v V (17)
and the search process is terminated; otherwise, the algorithm enters
st ft wt lt a ber ht
, t vd, h vves , hbber , d vver , l vves , lbber , d vhor , c vwt , c vht the main loop. Parent selection is the next step. Then, the defined cri-
v , v , v , v , t v , tb , pvb , t vb
terion for updating the algorithmic parameters is checked. If the cri-
, c vlt , K R+ v V, b B (18) terion is met, the algorithmic parameters will be updated (more details
are provided in Section 5.6). After that, the offspring chromosomes are
5. Solution algorithm description generated using the algorithmic operators (crossover and mutation). If
there are some infeasible solutions in the newly generated offspring, the
Values of the algorithmic parameters have substantial effects on algorithm repairs them in the next step. After that, the offspring eva-
performance of a given EA. The ASAEA algorithm, proposed in this luation is implemented, and then a group of the offspring chromosomes
study, tackles the drawbacks of the existing strategies that are used for is selected via the offspring selection operation. Once the stopping
setting the major algorithmic parameters in the process of evolution of criterion is met, the algorithm returns the chromosome with the best
the algorithm (crossover and mutation probabilities are considered in fitness value in the population, which corresponds to the berth schedule
this study, as they enable the algorithm with explorative and ex- with the lowest total vessel service cost, including the total vessel
ploitative capabilities – [13]). More specifically, ASAEA relies on the waiting cost, the total vessel handling cost, and the total vessel late
augmented self-adaptive parameter control strategy, where not only the departure cost.
crossover and mutation probabilities are encoded in the solutions and
evolve throughout the search, but they are also periodically updated 5.2. Solution encoding
based on the feedback from the search. In the following sections of the
manuscript, each step of the ASAEA algorithm design is described in As it is shown in Fig. 3, a two-dimensional hybrid chromosome was
detail. adopted to represent the solutions to the DBSP mathematical model in
this study. The proposed chromosome representation is classified as
5.1. The main algorithmic steps hybrid, as it includes a combination of real values and integer values.
The integer-coded portion is used to represent the vessel to berthing
The main algorithmic steps of ASAEA are illustrated in Fig. 2. First position assignment. On the other hand, the real-coded portion is used
of all, the required input data for the DBSP mathematical model and the for storing the crossover and mutation probabilities for the ASAEA al-
ASAEA algorithm are assigned. Then, the initial population is gorithm. According to Fig. 3, a total of 7 vessels are assigned to 2

8
M. Kavoosi, et al.
Fig. 3. A chromosome representation model.
berthing positions. Vessels “5”, “1”, “3” and “6” have been assigned to
berthing position “1”; while vessels “4”, “7”, and “2” have been as-
signed to berthing position “2”. Besides, the values of 0.35 and 0.02
have been set for the crossover probability (r c ) and the mutation
probability (r m ), respectively. The individual components of each
ASAEA chromosome will be referred to as “genes” [13]. The term
“locus” (plural: “loci”) will be used to denote a location of a gene along
each ASAEA chromosome, while the term “allele” will be used to denote
a value of a gene [13]. Since the adopted chromosome representation is
two-dimensional, two genes are placed in each locus of a given ASAEA
chromosome. In the considered example, genes with alleles “5” (for a
vessel identifier) and “1” (for a berthing position) are placed in locus
“1”.
5.3. Initialization of the chromosomes and population
The developed ASAEA algorithm applies the First Come First Served
with Spatial Requirements (FCFS-SR) heuristic for generating the in-
teger portion of the initial population chromosomes. The FCFS-SR
heuristic specifically accounts for the geometric characteristics of the
arriving vessels and MCT berthing positions. Application of the FCFS-SR
heuristic avoids generation of the solutions that do not meet constraints
related to the spatial specifications (i.e., the minimum vertical and
horizontal clearance requirements should be satisfied for each vessel to
be served at the MCT). For a detailed description of the FCFS-SR
heuristic this study refers to Dulebenets et al. [9]. Since the FCFS-SR
heuristic is deterministic, it will generate exactly the same vessel to
berthing position assignments. Such population initialization is not
desirable, as the ASAEA algorithm will start the search from just one
point in the search space. In order to avoid the latter drawback, half of
the initial ASAEA population will be initialized with FCFS-SR, while the
other half will be initialized randomly. As for the real-coded portion of
the chromosomes, it will be generated randomly for the whole popu-
lation, assuming the ranges of [0.30, 0.80] and [0.01, 0.06] for the
crossover and mutation probabilities, respectively. The latter ranges
were adopted based on the available literature [8,9]. The number of
chromosomes in the ASAEA population (i.e., population size – PopSize )
will be determined during the process of parameter tuning (more de-
tails are provided in Section 6.2 of the manuscript).
5.4. Fitness function
Once the initial population chromosomes are generated by the
ASAEA algorithm, their fitness values will be estimated. This study
assumes that the fitness function will be equal to the objective function
(1) of the DBSP mathematical model (i.e., the fitness function will be
equal to the total cost of serving the vessels arriving at the MCT).
5.5. Selection of the parent chromosomes
The main loop of the ASAEA algorithm starts with parent selection.
The parent chromosomes will be further used within ASAEA to produce
and mutate the offspring chromosomes via the crossover and mutation
operations. The Boltzmann selection mechanism was chosen for
9
Advanced Engineering Informatics 42 (2019) 100972
identification of the parent chromosomes, which is based on the fol-
lowing equation [13]:
e u (Xj )/ Tg
P (Xj ) = u (Xi)/ Tg
i
e (19)
where Xj is a solution; u (Xj ) is the fitness function value of solution Xj ;
Tg is the temperature in generation g ; P (Xj ) is the assigned selection
probability for a given solution.
The temperature value changes from one generation to another
within the proposed ASAEA algorithm as follows: Tg = T0 T ·g ,
where T0 is the initial temperature; and T is the temperature interval.
For a detailed description of the Boltzmann Selection mechanism this
study refers to Eiben and Smith [13]. Unlike the canonical Boltzmann
probability formula (see Eq. (19)), the Boltzmann probability formula,
adopted within the developed ASAEA algorithm, will divide the fitness
value of each solution by a normalizing coefficient ( ). Hence, u (X )/
will be used instead of u (X ) . Application of the normalizing coefficient
will allow avoiding abnormal temperatures due to fitness function
fluctuations in the ASAEA population.
5.6. ASAEA algorithmic parameters update
The developed ASAEA algorithm stores the most promising values
of the crossover and mutation probabilities (i.e., the crossover and
mutation probabilities that yield the largest relative improvements in
terms of the fitness values) in an external archive in each generation
throughout the algorithmic evolution. The archive keeps the maximum
of pairs of the crossover and mutation probabilities. The effect of each
set of the encoded crossover and mutation probabilities on each chro-
mosome is tracked by calculating the relative fitness value improve-
ment for a given chromosome as compared to the previous generation.
In fact, the augmentation concept is conducted by implementing two
updating processes: (1) updating the stored crossover and mutation
probabilities in the archive in each generation; and (2) updating the
encoded crossover and mutation probabilities in the population every
generations. In order to keep the archive updated, a total of most
promising crossover and mutation probability sets in the population
replace least promising crossover and mutation probability sets in the
archive in each generation. Then, every generations, one set of al-
gorithmic parameters out of stored sets in the archive is selected
randomly. The selected set, which belongs to one of the previous gen-
erations, is used to replace the algorithmic parameters encoded in
randomly selected individuals of the current population. It should be
noted that the values of , , , and will be determined during the
process of parameter tuning (more details are provided in Section 6.2 of
the manuscript). The sets of the algorithmic parameters, which led to
the best solutions, have a higher probability to be encoded within the
chromosomes throughout the process of evolution.
5.7. ASAEA operations
The selected parents undergo the algorithmic operations in order to
produce and mutate the offspring chromosomes. Based on Fig. 3, each one
of the chromosomes in the population has an integer-coded portion and a
real-coded portion. Thus, the customized crossover and mutation opera-
tors should be designed; so, they can be applied to the integer and real-
coded portions, separately. Moreover, according to the defined problem, a
solution is feasible when each vessel is assigned to only one berthing
position. It means that there should not be any repetitive number in the
second row of the integer portion of the chromosomes (see Fig. 3). Hence,
the designed algorithmic operators should consider the latter aspect for
the integer portion of the chromosome. The following sections describe
the developed crossover and mutation operators in detail.
M. Kavoosi, et al.
Fig. 4. Cycle and Laplace crossover operations.
5.7.1. Crossover operator
EAs are based on exploration and exploitation of the search space,
aiming to find the best possible solutions, where the crossover operator
gives the capability to the algorithm of exploring different domains of
the search space [13]. In this study, in order to apply the crossover
operation to the hybrid chromosomes, two different crossover operators
were combined together. The cycle crossover was applied for the in-
teger-coded portion, and the Laplace crossover was adopted for the
real-coded portion of the chromosomes within the developed ASAEA
algorithm. The cycle crossover mates two integer-coded chromosomes
in a way, which prevents generating infeasible solutions, while the
canonical crossover types (e.g., one-point crossover) would generate
infeasible solutions. Besides, the Laplace crossover, introduced in 2007,
was found to be efficient for the real-coded chromosomes [55]. Fig. 4
depicts how the cycle crossover operates.
Based on the example provided in Fig. 4, the following steps should
be taken to identify a cycle: (1) select the first allele of the first parent
(the vessel-related alleles are considered); (2) consider the same locus
in the second parent; (3) look for the same allele (as the allele in the
second parent) in the first parent and select that allele; and (4) do steps
“2” and “3” until the first allele, identified in step “1”, is found in the
second parent. Then, the genes from parent “1”, belonging to the cycle,
are copied into child “1” [13]. In the considered example, the gene
arrays (including vessels “5” and “6”), belonging to the cycle, are
copied from parent “1” in order to create child “1”. The missing genes
(i.e., the vessels that have not been assigned for service at the MCT) are
copied from parent “2” into child “1”. In the considered example, the
missing gene arrays (including vessels “1”, “3”, “4”, “7”, and “2”) are
copied from parent “2” into child “1”. Child “2” is produced in a similar
fashion.
The Laplace crossover, applied to the real-coded portion of the
chromosomes, is based on the Laplace distribution. Firstly, a random
coefficient, denoted as µ , is calculated using the following equation:
·[ln( )], 0.5
µ=
+ ·[ln( )], > 0.5 (20)
where is a normally distributed random number in the range [0, 1];
R+ and R+ are the coefficients of the Laplace density function
(in this study, it was assumed that = 0 , and = 0.75). The smaller the
value of , the closer the genotype of children to the genotype of par-
ents.
Suppose that the following two parents are used to generate chil-
dren by applying the Laplace crossover:
Y1 = (y11 , y12 , , y1n )
Y2 = (y21 , y22 , , y2n )
(21)
Then, each one of the alleles of the children can be calculated using
10
Advanced Engineering Informatics 42 (2019) 100972
the following equation:
1i = y1i + µ|y1i y2i |
2i = y2i + µ|y1i y2i | (22)
where y1i and y2i are the alleles of the parents; 1i and 2i are the alleles
of the generated children.
An example of the Laplace crossover operation is illustrated in
Fig. 4. Two children were produced from two parent chromosomes with
the crossover probabilities of 0.50 and 0.30 and the mutation prob-
abilities of 0.01 and 0.03. The crossover and mutation probabilities for
child “1” were set by the Laplace crossover to 0.47 and 0.02, respec-
tively. On the other hand, the crossover and mutation probabilities for
child “2” were set by the Laplace crossover to 0.27 and 0.04, respec-
tively.
5.7.2. Mutation operator
Each one of the offspring chromosomes, generated using the de-
signed crossover operator, represents a domain of the search space. The
mutation operator can exploit the domains in an effective way to en-
hance the quality of solutions. More precisely, a mutation operator
should change a solution gently to exploit a domain; otherwise, it may
guide the search process to another domain instead of exploiting a
specific domain. In this study, a customized mutation operator was
designed for the integer-coded portion of the chromosomes, which is a
combination of the swap and insert mutation operators [13]. An ex-
ample of a mutation operation is presented in Fig. 5. In case of the swap
mutation, two randomly selected genes switch their alleles. In the
considered example, berthing position “1” of locus “2” is swapped with
berthing position “2” of locus “7”, while vessel “5” of locus “1” is
swapped with vessel “4” of locus “5” (see Fig. 5). As for the insert
mutation, the second gene (i.e., the outmost right gene, selected for
mutation) is moved to the right locus of the first gene (i.e., the outmost
left gene, selected for mutation), and the rest of the genes are shifted to
the right. In the considered example, berthing position “2” of locus “7”
is inserted in locus “3”, while vessel “4” of locus “5” is inserted in locus
“2” (see Fig. 5).
On the other hand, the floating point mutation was applied for the
real-coded portion of the chromosomes. The floating point mutation
generates a random value from a given range and replaces the old allele
with the newly generated random value (Fig. 5). The crossover and
mutation probabilities were updated from 0.40 and 0.020 to 0.50 and
0.035 for the swap mutation example (Fig. 5). Besides, the crossover
and mutation probabilities were updated from 0.40 and 0.020 to 0.48
and 0.018 for the insert mutation example (Fig. 5). Algorithm 1 shows
how the aforementioned three mutation operators are applied to the
chromosomes within the developed ASAEA algorithm.
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Fig. 5. Swap, insert, and floating point mutation operations.

5.8. Chromosome repairing

Algorithm 1: Customized Mutation

CustMutation (Offspring ) The processes of generating the initial population and applying the
in: Offspring – the offspring chromosomes generated by the crossover operator algorithmic operators (i.e., crossover and mutation) may create in-
out: MutOffspring – the mutated offspring chromosomes feasible chromosomes. In this study, a repairing operator is applied to
0: MutOffspring Offspring ◁ Initialization infeasible chromosomes. Two factors can independently make a chro-
1: i 1 mosome to be infeasible: (1) disruption of the vessel service order in a
2: while i |Offspring| do
given chromosome; and (2) assignment of a vessel to the berthing po-
3: RandNum1 Rand (0, 1) ◁ Generate a random number
sition, which does not meet the spatial requirements. As it is illustrated
4: if rim > RandNum1 then
c c in Fig. 6, vessels “6” and “4”, which are assigned to berthing position
5: MutOffspringir Float (Offspringir ) ◁ Mutate the crossover probability value
m m
“1”, are placed in the loci between vessels “1”, “3” and “7”, which are
6: MutOffspringir Float (Offspringir ) ◁ Mutate the mutation probability value assigned to berthing position “2”. The latter causes disruption in the
7: end if
vessel service order at berthing positions “1” and “2”. The developed
8: RandNum2 Rand (0, 1) ◁ Generate a random number
9: if 0 RandNum2 0.25 or 0.51 RandNum2 0.75 then
ASAEA algorithm deploys the repairing operator to adjust the service
10: MutOffspringiber Swap (Offspringiber , r m ) ◁ Apply swap to berthing positions
order of vessels (see Fig. 6).
11: MutOffspringives Swap (Offspringives , r m ) ◁ Apply swap to vessel identifiers
12: else
Algorithm 2: Chromosome Repairing
13: MutOffspringiber Insert (Offspringiber , r m ) ◁ Apply insert to berthing positions
14: MutOffspringives Insert (Offspringives, r m ) ◁ Apply insert to vessel identifiers ChromRepair (MutChrom, V , B, hves , hber , d ver , l ves , lber , dhor )
15: end if in: MutChrom – the mutated offspring chromosome in the current generation;
16: i i+1 V = {1, , m} – set of vessels; B = {1, , n} – set of berths; hves – draft of vessels;
17: end while
hber – depth of berths; d ver – vertical clearance requirements; l ves – length of ve-
18: return MutOffspring
ssels; lber – length of berths; dhor – horizontal clearance requirements
In step “0”, a new data structure (MutOffspring ) is generated for storing out: ChromRep – the repaired chromosome
0: ChromRep MutChrom ◁ Initialization
the mutated offspring chromosomes. Then, the customized mutation
1: cntr 1; v 1; b 1 ◁ Initialization
operator enters the main loop (steps “2”–“17”). The random value 2: while cntr m do
(RandNum1) between “0” and “1” is generated in step “3”. The muta- 3: v ves ; b
ChromRepcntr ber
ChromRepcntr ◁ Determine the vessel and berth identifiers
tion probability is checked in step “4”. If the mutation probability is 4: if (hbber < hvves + dvver ) or (lbber < lvves + dvhor ) do ◁ Check the spatial requirements
greater than RandNum1, the mutation operation will be executed for
5: FitBerths Find (hber h vves + d vver and lber l vves + d vhor ) ◁ Find the
the crossover and mutation probability values. Note that rim is the appropriate berths
mutation probability encoded in each one of the ASAEA chromosomes, 6: ber
ChromRepcntr Rand (FitBerths ) ◁ Select one of the appropriate berths
as the developed algorithm relies on the self-adaptive parameter control 7: end if
strategy. The floating point mutation is applied to the real-coded por- 8: cntr cntr + 1
tion of the given chromosome (i.e., the crossover and mutation prob- 9: end while
ability values) in steps “5” and “6”. Another random number 10: ChromRep Sort (ChromRep , B ) ◁ Sort the gene arrays by berthing positions
11: return ChromRep
(RandNum2 ), varying between “0” and “1”, is generated in step “8” to
determine whether the swap or insert mutation should be executed. If The main steps of the repairing operator are presented in Algorithm 2.
RandNum2 falls into the range [0, 0.25] or the range [0.51, 0.75], the In step “0”, a data structure (ChromRep ) is initiated to store the repaired
swap mutation operator will be executed for the integer-coded portion chromosome. Then, the repairing operator enters the loop (steps
of the chromosomes (steps “10” and “11”); otherwise, the insert mu- “2”–“9”), where it checks satisfaction of the spatial requirements for
tation operator will be executed for the integer-coded portion of the each vessel, encoded in the given chromosome. In case if the spatial
chromosomes (steps “13” and “14”). Algorithm 1 returns the mutated requirements are not met for a given vessel, all of the berthing posi-
offspring chromosomes (MutOffspring ) in step “18”. tions, which have adequate depth and length to serve the vessel, are
selected and stored in the FitBerths data structure (step “5”). In order to
avoid any bias in the repairing process, one of the suitable berthing

11
M. Kavoosi, et al.
Fig. 6. Repairing an infeasible chromosome.
positions, which satisfies the minimum vertical and horizontal clear-
ance requirements for a given vessel, is selected randomly (step “6”).
After checking the spatial requirements for each vessel, the repairing
operator sorts the gene arrays of the mutated offspring chromosome
based on the berthing positions to prevent any disruptions in the vessel
service order (step “10”). The algorithm returns the repaired offspring
chromosome in step “11”.
5.9. Selection of the offspring chromosomes
A Stochastic Universal Sampling (SUS) selection operator was adopted
within the developed ASAEA algorithm, in order to determine the off-
spring chromosomes that will be transferred to the next generation. Unlike
certain selection mechanisms (e.g., ranking selection), SUS is not sub-
stantially biased by fitness of the chromosomes. Fig. 7 illustrates how the
SUS mechanism operates schematically. According to Fig. 7, firstly, 8 in-
dividuals are assigned a portion of the range [0, 1] based on their fitness.
Suppose that Nchrom = 6 chromosomes should be selected. Hence, a
random number is generated in the range of [0, 1/6], which will be
considered as the start location of the pointers. The distance between the
pointers is equal to 1/ Nchrom . The offspring chromosomes are selected
based on the pointer locations. According to the example illustrated in
Fig. 7, individuals “1”, “2”, “3”, “4”, “5”, and “8” have been selected for
the next generation (based on the pointer locations) from the generated
offspring. This operation should be repeated until the necessary number of
the offspring chromosomes have been selected [13].
5.10. Elitism strategy
Deployment of the algorithmic operators generally causes a lot of
changes in the solutions, and there is no guarantee that better solutions
would be generated as compared to the previous generation. Elitism is a
strategy that is utilized to keep the best solution in each generation and
transfer it to the next generation without any change. More specifically,
the best solution is recorded before implementing the algorithmic op-
erators (i.e., parent selection, crossover, mutation, and offspring se-
lection), and, then, it will be added to the next generation offspring
chromosomes after the offspring selection. The elitism strategy will be
applied within the developed ASAEA algorithm before deployment of
the Boltzmann selection mechanism at the parent selection stage.
Fig. 7. Stochastic universal sampling selection.
12
Advanced Engineering Informatics 42 (2019) 100972
5.11. Convergence criterion
The evolution process will never stop in nature, but a convergence
or stopping criterion should be defined to have a balanced tradeoff
between the final solution quality and the computational time [13].
Based on changes in the quality of discovered solutions from one gen-
eration to another, the proposed ASAEA algorithm will be terminated
after a certain number of generations ( gens ).
6. Computational experiments
A set of comprehensive computational experiments were under-
taken in order to assess performance of the developed ASAEA algo-
rithm, which relies on the augmented self-adaptive parameter control
strategy. A detailed description of the computational experiments is
provided in this section of the manuscript. The computational experi-
ments include the following four major steps: (1) input data generation;
(2) algorithmic parameter setting; (3) performance evaluation against
the benchmark objective function values; and (4) comprehensive
comparative analysis against the alternative EA-based and state-of-the-
art metaheuristics. In the first step, the input data were generated for
the DBSP mathematical model. In the second step, a parameter tuning
analysis was conducted to select the appropriate values of parameters
for the ASAEA algorithm and other alternative algorithms. The alter-
native solution algorithms, which have been considered in this study,
can be divided into two groups.
The first group of the alternative algorithms includes the EA-based
algorithms with different strategies for setting the crossover and mu-
tation probabilities, including the following: (1) a typical EA with
constant values for the crossover and mutation probabilities that re-
main unchanged throughout evolution of the algorithm; (2) an EA with
a deterministic parameter control strategy (DPCEA), where the cross-
over and mutation probabilities are changed based on a defined counter
(e.g., a specific number of generations) throughout evolution of the
algorithm; (3) an EA with an adaptive parameter control strategy
(AEA), where the crossover and mutation probabilities are updated
based on the feedback from the search throughout evolution of the
algorithm; and (4) an EA with a self-adaptive parameter control
strategy (SAEA), where the crossover and mutation probabilities are
encoded in the population chromosomes and evolve along with the
algorithm itself. A detailed description of the EA, DPCEA, AEA, and
SAEA algorithms can be found in Dulebenets et al. [9]. Apart from the
EA-based algorithms, there are a lot of different metaheuristic algo-
rithms in the literature, which have been inspired by different phe-
nomena [53,54,56–58]. The second group of the alternative algorithms
includes the state-of-the-art metaheuristic algorithms, which have been
widely used in the MCT and freight terminal operations literature, in-
cluding the following: (1) Tabu Search (TS); (2) Simulated Annealing
(SA); (3) Variable Neighborhood Search (VNS); (4) Ant Colony Opti-
mization (ACO); and (5) Particle Swarm Optimization (PSO). A detailed
description of the TS, SA, VNS, ACO, and PSO algorithms can be found
in Cordeau et al. [56], Emde and Boysen [54], Hansen et al. [57],
Arabani et al. [58], and Ting et al. [53], respectively.
In the third step of the computational experiments, the objective
function values, returned by the considered solution algorithms for the
small-size problem instances, will be evaluated against the benchmark
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

objective function values. The benchmark objective function values will
be determined using CPLEX, which is able to return the global optimal
solutions for mixed-integer linear programming models. General
Algebraic Modeling System (GAMS, [59]) will be used to encode the
DBSP mathematical model and solve it with CPLEX. Note that CPLEX
will not be applied to solve the large-size problem instances, as it was
not able to solve the large-size problem instances of the DBSP mathe-
matical model within 2 h based on preliminary computational experi-
ments. In the fourth step, the candidate solution algorithms will be
comprehensively analyzed for the large-size problem instances in terms
of the objective function and computational time values. The scope of
the numerical experiments also included a detailed evaluation of the
convergence patterns and the objective function stochastic variations
for the considered solution algorithms (more information can be found
in Appendices A and B, respectively). Furthermore, Appendix C ana-
lyzes evolution of the algorithmic parameters, encoded within the SAEA
and ASAEA chromosomes.
Note that all the solution algorithms, considered in this study, were
encoded using MATLAB 2016a [60]. Alienware CPU with an Intel®
Core™ i7-7700 K processor, 32.0 GB of RAM, and Windows 10 Oper-
ating System was utilized in order to conduct the computational ex-
periments in this study. A description of each step of the computational
experiments is provided in Sections 6.1–6.4 of the manuscript.
6.1. Input data generation
The first step of the computational experiments aimed to generate
the input data for the DBSP mathematical model. The input data, re-
quired to conduct the computational experiments, were adopted from
the available BSP literature, freight terminal literature, and appropriate
online resources [9,11,41,44,45,61,63–65]. Table 3 elaborates on how
the values of the DBSP mathematical model parameters were generated.
According to the common assumption used in the BSP literature, it was
assumed that vessels arrive at the MCT according to an exponential
distribution [9,11,41]. An average inter-arrival time of vessels was set
equal to 2 h as follows: t va+ 1 = t va + t v V , t = exp [2] (hours),
where term “exp [num1]” denotes the exponential distribution with a
mean of num1. It was assumed that each MCT berthing position becomes
available at the beginning of planning period (i.e., tbber = 0 b B ).
The number of containers that should be handled for each vessel was
generated randomly between 1000 TEUs and 3000 TEUs based on the
uniform distribution as follows: qv = U [1000; 3000] v V (TEUs),
where term “U [num1; num2]” denotes the uniform distribution with
lower bound num1 and upper bound num2 . The handling productivity
for each vessel at a given berthing position was assigned as follows:
pvb = U [130; 160] v V , b B (TEUs/hour). The requested de-
parture time for a given vessel was estimated based on the arrival time
of that vessel at the MCT, vessel handling time at the berthing position,
which has the highest handling productivity, and a tolerance factor,
varying between 20% and 50% and capturing a potential increase in the
vessel handling time (e.g., due to diversion of the vessel from the
berthing position with the highest handling productivity to the alter-
native berthing position), as follows: t vd =
t va + minb (t vb
ht
)·U [1.20; 1.50] v V (hours).
Some of the common vessel types in the maritime transportation
industry were modeled in this study, including the following [65]: (1)
Panamax (length: 820.2 feet, draft: 41.0 feet); (2) Panamax Max
(length: 951.4 feet, draft: 41.0 feet); (3) Post Panamax (length: 935.0
feet, draft: 42.7 feet); (4) Post Panamax Plus (length: 984.3 feet, draft:
47.6 feet); (5) New Panamax (length: 1200.8 feet, draft: 49.9 feet); and
(6) Triple E (length: 1312.3 feet, draft: 50.9 feet). The dimensional
specifications of the MCT berthing positions were generated based on
the dimensional specifications of the aforementioned vessel types.
Specifically, the depth of a given berthing position was assigned as
follows:
hbber = 1.20· min v (h vves ) + U [0; 1.20·max v (h vves ) min v (h vves )] b B
(feet). On the other hand, the length of a given berthing position was set
as follows: lbber = 1.20·min v (l vves ) + U [0; 1.20· max v (l vves ) min v (l vves )]
b B (feet). The minimum vertical clearance for each vessel was
assigned as follows: d vver = U [4; 8] v V (feet). The minimum
horizontal clearance for each vessel was assigned as follows:
d vhor = U [50; 100] v V (feet).
The waiting cost for each vessel was generated randomly between
1000 USD/hour and 5000 USD/hour based on the uniform distribution
as follows: c vwt = U [1000; 5000] v V (USD/hour). On the other
hand, the handling cost for each vessel was assumed to vary between
50,000 USD/hour and 70,000 USD/hour and was estimated as follows:
c vht = U [50, 000; 70, 000] v V (USD/hour). The late departure cost
for each vessel was assumed to vary between 5000 USD/hour and
10,000 USD/hour and was set as follows:
c vlt = U [5000; 10, 000] v V (USD/hour). The value of a large po-
sitive number was fixed to K = 10, 000 . The number of arriving vessels
(m ) and the number of available berthing positions (n ) were defined
independently for each one of the problem instances. A total of 60
problem instances were developed by using the generated input data

Table 3
Data generation for the DBSP mathematical model.
Parameter Selected Value

Number of arriving vessels: m (vessels) The values differ based on the problem instance
Number of available berthing positions: n (berthing positions) The values differ based on the problem instance
Arrival time of vessel v at the MCT: t va, v V (hours) t va+ 1 = t va + t v V
Average inter-arrival time of vessels: t (hours) t = exp[2]
First time availability of berthing position b in the planning horizon: tbber , b B (hours) tbber = 0 b B
Number of containers to be handled for vessel v : qv , v V (TEUs) qv = U [1000; 3000] v V
Handling productivity for vessel v at berthing position b : pvb , v V, b B (TEUs/hour) pvb = U [130; 160] v V, b B
Requested departure time for vessel v : t vd, v V (hours) t vd = t va + minb (t vb
ht )·U [1.20; 1.50] v V
Draft of vessel v : h vves , v V (feet) [41.0; 41.0; 42.7; 47.6; 49.9; 50.9]
Depth of berthing position b : hbber , b B (feet) hbber = 1.20·minv (h vves ) + U [0; 1.20·max v (h vves ) minv (h vves )] b B
Minimum vertical clearance for vessel v (feet): d vver , v V (feet) d vver
= U [4; 8] v V
Length of vessel v : l vves , v V (feet) [820.2; 951.4; 935.0; 984.3; 1200.8; 1312.3]
Length of berthing position b : lbber , b B (feet) lbber = 1.20·minv (l vves ) + U [0; 1.20·max v (l vves ) minv (l vves )] b B
Minimum horizontal clearance for vessel v : d vhor , v V (feet) d vhor = U [50; 100] v V
Waiting cost for vessel v : c vwt , v V (USD/hour) c vwt = U [1000; 5000] v V
Handling cost for vessel v : c vht , v V (USD/hour) c vht = U [50, 000; 70, 000] v V
Late departure cost for vessel v : c vlt, v V USD/hour) c vlt = U [5000; 10, 000] v V
Large positive number: K 10, 000

13
 M. Kavoosi, et al.

Table 4
The parameter selection analysis results for the considered EA-based metaheuristic algorithms.
Algorithm Parameter Description Candidate Values Best Value Algorithm Parameter Description Candidate Values Best Value

EA Population size (PopSize ) [40; 50; 60] 50 AEA Crossover probability values (r cv )2 [0.80–0.40; 0.80–0.30; [0.80–0.30]
0.30–0.80]
EA Initial temperature (T0 ) [1200; 1500; 1800] 1500 AEA Mutation probability values (r mv )3 [0.06–0.01; 0.06–0.02; [0.06–0.02]
0.01–0.06]
EA Temperature interval ( T ) [0.10; 0.15; 0.20] 0.20 AEA Adaptive criterion ( gensmax ) [100; 200; 300] 300
EA Normalizing coefficient ( )1 [0.10; 0.15; 0.20] 0.10 AEA Maximum number of generations ( gens ) [4000; 5000; 6000] 6000
EA Crossover probability (r c ) [0.30; 0.50; 0.80] 0.80 SAEA Population size (PopSize ) [40; 50; 60] 50
EA Mutation probability (r m ) [0.01; 0.03; 0.06] 0.03 SAEA Initial temperature (T0 ) [1200; 1500; 1800] 1200
EA Maximum number of generations ( gens ) [4000; 5000; 6000] 6000 SAEA Temperature interval ( T ) [0.10; 0.15; 0.20] 0.20
DPCEA Population size (PopSize ) [40; 50; 60] 50 SAEA Normalizing coefficient ( )1 [0.10; 0.15; 0.20] 0.10
DPCEA Initial temperature (T0 ) [1200; 1500; 1800] 1500 SAEA Maximum number of generations ( gens ) [4000; 5000; 6000] 6000
DPCEA Temperature interval ( T ) [0.10; 0.15; 0.20] 0.20 ASAEA Population size (PopSize ) [40; 50; 60] 50

14
DPCEA Normalizing coefficient ( )1 [0.10; 0.15; 0.20] 0.10 ASAEA Initial temperature (T0 ) [1200; 1500; 1800] 1200
DPCEA Crossover probability values (r cv ) [0.80, 0.40; 0.80, 0.30; 0.30, [0.80, 0.30] ASAEA Temperature interval ( T ) [0.10; 0.15; 0.20] 0.20
0.80]
DPCEA Mutation probability values (r mv ) [0.06, 0.01; 0.06, 0.02; 0.01, [0.06, 0.02] ASAEA Normalizing coefficient ( )1 [0.10; 0.15; 0.20] 0.10
0.06]
DPCEA Number of segments in the piecewise function [15; 20; 25] 20 ASAEA Generational interval between the parameter updates( ) [1; 10; 20] 1
(nseg )
DPCEA Maximum number of generations ( gens ) [4000; 5000; 6000] 6000 ASAEA The external archive size ( )4 [0.20; 0.25; 0.30] 0.30
AEA Population size (PopSize ) [40; 50; 60] 50 ASAEA Number of individuals for the parameter update ( )4 [0.20; 0.25; 0.30] 0.20
AEA Initial temperature (T0 ) [1200; 1500; 1800] 1200 ASAEA Percentage of parameter values to be updated in the external [0.80; 0.90; 1.00] 1.00
archive ( )
AEA Temperature interval ( T ) [0.10; 0.15; 0.20] 0.20 ASAEA Maximum number of generations ( gens ) [4000; 5000; 6000] 6000
AEA Normalizing coefficient ( )1 [0.10; 0.15; 0.20] 0.10

Notes: 1 – the normalizing coefficient for the Boltzmann selection mechanism was set as a percentage of the average population fitness; 2 – the crossover probability values were altered with an increment of 0.10 for the
AEA algorithm; 3 – the mutation probability values were altered with an increment of 0.01 for the AEA algorithm; 4 – the external archive size and the number of individuals for the parameter update were set as a
percentage of the ASAEA population.
Advanced Engineering Informatics 42 (2019) 100972
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Notes: 1 – the exchange rate parameter corresponds to the number of vessel to berth assignments to be changed in a given solution in order to create a new solution during the local search; 2 – the normalizing coefficient
and changing the number of arriving vessels and the number of avail-
Best Value
able berthing positions. The developed problem instances were divided

6000

6000
into the small-size problem instances ([PI-1 through PI-30], where m
0.20
0.30

1.00
1.50

2.00
2.00
0.20
50

50
was changed from 8 to 17 with an increment of 1 vessel, while n was
changed from 2 to 4 with an increment of 1 berthing position), and the
large-size problem instances ([PI-31 through PI-60], where m was

[4000; 5000; 6000]

[4000; 5000; 6000]

[0.20; 0.30; 0.40]
[0.30; 0.40; 0.50]

[1.00; 1.50; 2.00]

[1.00; 1.50; 2.00]

[1.50; 2.00; 3.00]

[1.50; 2.00; 3.00]
[0.10; 0.20; 0.30]
Candidate Values

changed from 75 to 120 with an increment of 5 vessels, while n was

changed from 6 to 10 with an increment of 2 berthing positions).
[40; 50; 60]

[40; 50; 60]

Note that the input data, which were generated for evaluation of the
candidate solution algorithms (see Table 3), can be downloaded in a
MATLAB format using the following link: Data_ASAEA. Alternatively,

for the Boltzmann selection mechanism was set as a percentage of the average solution fitness; 3 – the inertia weight for PSO was set as follows: W PSO = 0.50 + U [0.00; 0.50] [53].
the input data can be also requested by contacting the corresponding
author.
v¯ )
v
_

6.2. Algorithmic parameter setting

V,

Maximum number of iterations (iters )

Maximum number of iterations (iters )

As indicated at the beginning of Section 6, the developed ASAEA

vv¯ , v, v¯

PSO 3
)
( ACO )

algorithm will be compared against nine alternative state-of-the-art

Maximum velocity increase (Vmax
_
0

metaheuristic algorithms, which have been frequently used in the MCT

_
Initial pheromone amount (

Heuristic parameter ( ACO )

Pheromone trail parameter

and freight terminal operations literature. Each one of the considered

Cognition component (c1)
Population size (PopSize )

Population size (PopSize )

algorithms has a set of parameters, which remain constant throughout

Parameter Description

Social component (c2 )

Evaporation rate ( )

the search process (while some of the considered algorithms have cer-
tain parameters that change throughout the search process – i.e., the
DPCEA, AEA, SAEA, and ASAEA algorithms). The second step of the
computational experiments aimed to select the appropriate values for
parameters of the considered algorithms, which is critical in order to
ensure adequate algorithmic performance [61–64]. Since each algo-
rithm has a significant number of parameters, the Taguchi’s scheme was
Algorithm

adopted in order to conduct the parameter selection analysis in a timely

manner. Based on the Taguchi’s scheme, only “the most favorable”
ACO
ACO
ACO

ACO
ACO
ACO
PSO
PSO
PSO
PSO
PSO

parameter values are considered [58]. Specifically, one of the algo-

rithmic parameters is selected at a time, and a series of computational
experiments are conducted, where the value of the selected parameter
Best Value

is changed, while the other parameters are assigned certain constant

6000
1200

6000

6000

values. The most promising values of a given parameter are determined

0.20
0.10
10

10

10
2

based on a tradeoff between the solution quality at convergence and the

The parameter selection analysis results for the considered state-of-the-art metaheuristic algorithms.

computational time required. Once the most promising values are de-
termined for a given parameter, the same procedure is applied for the
[4000; 5000; 6000]
[1200; 1500; 1800]

[4000; 5000; 6000]

[4000; 5000; 6000]

[0.10; 0.15; 0.20]
[0.10; 0.15; 0.20]
Candidate Values

rest of the parameters.

A total of 3 problem instances were selected at random from the
[10; 15; 20]
[5; 7; 10]

[5; 7; 10]

generated large-size problem instances (which are described in Section

[2; 4; 6]

[2; 4; 6]

[2; 4; 6]

6.1 of the manuscript) in order to conduct the parameter selection

analysis. A total of 10 replications were performed for each algorithm
and each most favorable parameter combination in order to obtain the
average objective function and computational time values. Tables 4 and
TS )

5 contain the results from the conducted parameter selection analysis,

Number of solutions evaluated during the local search (

including the algorithm, the parameter description, the candidate va-

lues for a given parameter, and the best value for a given parameter.

6.3. Performance evaluation against the benchmark objective function

VNS )

values
Maximum number of iterations (iters )

Maximum number of iterations (iters )

Maximum number of iterations (iters )

Number of neighborhood structures (

The third step of the computational experiments aimed to evaluate

performance of the considered algorithms against the benchmark ob-
Normalizing coefficient ( )2
Temperature interval ( T )

jective function values. CPLEX is an exact optimization algorithm,

Initial temperature (T0 )
Parameter Description

Exchange rate (r VNS )1

which has been widely used in the MCT and operations research lit-
Exchange rate (r SA )1
Exchange rate (r TS )1

erature to obtain the global optimal solutions for mixed-integer linear

Tabu List size ( )

programming mathematical models [57]. CPLEX was used in this study

in order to obtain the benchmark objective function values (i.e., the
optimal objective function values) and compare them against the ob-
jective function values, returned by the considered algorithms (i.e., the
EA, DPCEA, AEA, SAEA, ASAEA, TS, SA, VNS, ACO, and PSO algo-
rithms), for the small-size instances of the DBSP mathematical model
Algorithm

[PI-1 through PI-30]. The maximum CPLEX computational time was

Table 5

limited to 7200 s, while the relative optimality gap was limited to

VNS
VNS
VNS
SA
SA
SA
SA
SA
TS
TS
TS

TS

0.01%.

15
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Table 6
Comparison of the considered EA-based metaheuristic algorithms against the exact optimization algorithm (CPLEX) for problem instances [PI-1 through PI-30].
Instance No. of No. of CPLEX EA DPCEA AEA SAEA ASAEA
Vessels Berths
Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec.
106 USD 106 USD 106 USD 106 USD 106 USD 106 USD

PI-1 8 2 7.9382 0.31 7.9382 6.63 7.9382 6.86 7.9382 7.90 7.9382 9.58 7.9382 10.59
PI-2 8 3 7.1958 0.81 7.1958 6.65 7.1958 6.85 7.1958 7.93 7.1958 9.61 7.1958 10.64
PI-3 8 4 6.8256 0.22 6.8256 6.62 6.8256 6.81 6.8256 7.88 6.8256 9.55 6.8256 10.57
PI-4 9 2 9.0248 0.59 9.0248 6.79 9.0248 6.96 9.0248 8.04 9.0248 9.73 9.0248 10.84
PI-5 9 3 8.0258 2.89 8.0258 6.78 8.0258 6.92 8.0258 8.02 8.0258 9.72 8.0258 10.82
PI-6 9 4 7.5863 0.72 7.5863 6.78 7.5863 6.94 7.5863 8.03 7.5863 9.77 7.5863 10.85
PI-7 10 2 10.2635 0.78 10.2635 6.97 10.2635 7.09 10.2635 8.20 10.2635 10.02 10.2635 11.16
PI-8 10 3 9.0151 12.09 9.0151 7.00 9.0151 7.11 9.0151 8.23 9.0151 10.07 9.0151 11.21
PI-9 10 4 8.5098 1.69 8.5098 6.97 8.5098 7.09 8.5098 8.21 8.5098 10.02 8.5098 11.16
PI-10 11 2 11.4298 1.15 11.4298 7.10 11.4298 7.19 11.4298 8.32 11.4298 10.21 11.4298 11.43
PI-11 11 3 9.9455 59.43 9.9455 7.13 9.9455 7.21 9.9455 8.35 9.9455 10.26 9.9455 11.45
PI-12 11 4 9.3916 18.44 9.3916 7.15 9.3916 7.22 9.3916 8.36 9.3916 10.31 9.3916 11.49
PI-13 12 2 12.9442 2.03 12.9442 7.35 12.9442 7.39 12.9442 8.64 12.9442 10.63 12.9442 11.92
PI-14 12 3 11.2608 209.34 11.2608 7.28 11.2608 7.34 11.2608 8.52 11.2608 10.52 11.2608 11.76
PI-15 12 4 10.4824 47.50 10.4824 7.30 10.4824 7.37 10.4824 8.55 10.4824 10.53 10.4824 11.80
PI-16 13 2 14.4643 5.19 14.5251 7.52 14.4643 7.53 14.4643 8.71 14.4643 10.81 14.4643 12.11
PI-17 13 3 12.4245 975.45 12.5388 7.49 12.4941 7.52 12.4941 8.72 12.4245 10.82 12.4245 12.13
PI-18 13 4 11.5453 655.94 11.7222 7.51 11.6954 7.54 11.6954 8.74 11.5453 10.87 11.5453 12.17
PI-19 14 2 15.8371 8.70 16.2517 7.67 16.1942 7.68 16.1570 8.90 15.8971 11.11 15.8371 12.45
PI-20 14 3 13.4706 5453.10 13.8648 7.69 13.8244 7.71 13.8168 8.90 13.5906 11.16 13.4706 12.50
PI-21 14 4 12.4624 4674.83 12.8537 7.68 12.8128 7.70 12.7864 8.91 12.6624 11.12 12.5624 12.48
PI-22 15 2 17.1866 14.59 17.7417 7.86 17.7264 7.89 17.6901 9.14 17.4966 11.41 17.3966 12.84
PI-23 15 3 14.4428 7203.65 14.0908 7.87 14.0395 7.88 13.7550 9.13 13.7275 11.39 13.5847 12.84
PI-24 15 4 13.2278 7203.71 13.1679 7.91 13.0275 7.91 12.7758 9.19 12.7540 11.43 12.5721 12.88
PI-25 16 2 18.7701 28.98 19.4133 7.99 19.3787 8.05 19.3294 9.28 19.1830 11.69 19.0101 13.12
PI-26 16 3 15.5692 7202.86 15.0605 8.00 14.9772 8.04 14.6616 9.28 14.6100 11.68 14.5859 13.17
PI-27 16 4 14.1527 7202.95 14.0320 8.01 13.9020 8.08 13.6510 9.33 13.6272 11.71 13.4999 13.20
PI-28 17 2 20.3710 72.00 21.0907 8.13 21.0606 8.23 20.9897 9.45 20.8502 11.94 20.6358 13.45
PI-29 17 3 16.7674 7202.66 16.1970 8.16 16.1116 8.27 15.7584 9.47 15.7084 12.00 15.7014 13.50
PI-30 17 4 15.1667 7202.42 15.0326 8.17 14.8568 8.27 14.6034 9.50 14.5534 11.99 14.4534 13.52

Average: 12.1899 1848.83 12.2474 7.41 12.2135 7.49 12.1489 8.66 12.0978 10.72 12.0528 12.00

Note that all the considered algorithms can be classified as sto-
chastic search algorithms; therefore, a total of 10 replications were
performed for each algorithm in order to obtain the average objective
function and computational time values for the small-size problem in-
stances. Tables 6 and 7 contain the results from the conducted analysis,
including the instance number, the number of vessels arriving for ser-
vice at the MCT, the number of the MCT berthing positions available for
vessel service, the obtained objective function values (i.e., the bench-
mark objective function values, obtained by CPLEX, and the average
objective function values, returned by the considered algorithms), and
the required computational time values.
Based on the analysis results, it was found that the CPLEX compu-
tational time was substantially influenced with the problem size. The
latter finding underlines a high computational complexity of the DBSP
mathematical model and justifies the need for the development of
heuristic and/or metaheuristic algorithms to tackle the large-size pro-
blem instances. Furthermore, problem instances PI-23, PI-24, PI-26, PI-
27, PI-29, and PI-30 could not be solved to the global optimality within
the computational time limit imposed (i.e., 7200 sec). On the other
hand, the computational time, required by the considered metaheuristic
algorithms to solve the small-size problem instances, did not exceed 14
sec. As for the objective function values, the considered algorithms
were able to return the global optimal solutions for many of the gen-
erated small-size problem instances (i.e., the optimality gaps were equal
to zero). However, for certain small-size problem instances, the con-
sidered algorithms were not able to find the global optimal solutions.
For example, the average objective function value, returned by the EA
algorithm for problem instance PI-16, was 0.42% worse than the op-
timal one.
An additional analysis was performed in order to estimate the
maximum optimality gaps for the considered algorithms over the
generated small-size problem instances. A total of 50 scenarios were
developed for each small-size problem instance by changing the vessel
arrival times (t va+ 1 = t va + t v V , t = exp [2]) and the handling
productivities ( pvb = U [130; 160] v V , b B ). A total of 10 re-
plications were performed for each one of the considered algorithms in
order to obtain the average objective function values for each scenario
and each small-size problem instance. The optimality gaps were esti-
mated based on the following relationship:
TCAlg TCCPLEX
OGAlg =
TCCPLEX (23)
where OGAlg is the optimality gap of algorithm Alg ; TCCPLEX is the op-
timal objective function value (i.e., the total vessel service cost), ob-
tained by CPLEX; TCalg is the average over 10 replications objective
function value, returned by algorithm Alg .
The optimality gaps over the generated scenarios for the considered
algorithms and all the generated small-size problem instances are pre-
sented in Fig. 8 using a boxplot format. Each boxplot is represented by a
rectangle. The boxplot rectangle covers the range, varying between
25th and 75th optimality gap percentiles (the bottom of a rectangle
corresponds to the 25th optimality gap percentile, while the top of a
rectangle corresponds to the 75th optimality gap percentile). An addi-
tional line within each rectangle corresponds to the median optimality
gap. The dashed lines (or “whiskers”) extend from the bottom and the
top of each rectangle, covering 99.3% of the optimality gap values. The
optimality gap values, falling outside the range that is covered by
whiskers, are considered as outliers and are denoted using symbol “+”.
Based on the conducted optimality gap analysis, it can be observed that
all the considered algorithms demonstrated a high accuracy. The
maximum optimality gaps did not exceed ≈4.5% over the generated
scenarios and considered small-size problem instances. However,
smaller optimality gaps were generally recorded for the SAEA and

16
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Table 7
Comparison of the considered state-of-the-art metaheuristic algorithms against the exact optimization algorithm (CPLEX) for problem instances [PI-1 through PI-30].
Instance No. of No. of CPLEX TS SA VNS ACO PSO
Vessels Berths
Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec. Objective, CPU, sec.
106 USD 106 USD 106 USD 106 USD 106 USD 106 USD

PI-1 8 2 7.9382 0.31 7.9382 0.17 7.9382 0.13 7.9382 0.11 7.9382 10.44 7.9382 6.79
PI-2 8 3 7.1958 0.81 7.1958 0.17 7.1958 0.13 7.1958 0.11 7.1958 10.44 7.1958 6.78
PI-3 8 4 6.8256 0.22 6.8256 0.17 6.8256 0.13 6.8256 0.11 6.8256 10.43 6.8256 6.91
PI-4 9 2 9.0248 0.59 9.0248 0.18 9.0248 0.14 9.0248 0.11 9.0248 10.87 9.0248 7.20
PI-5 9 3 8.0258 2.89 8.0258 0.18 8.0258 0.14 8.0258 0.11 8.0258 10.89 8.0258 7.22
PI-6 9 4 7.5863 0.72 7.5863 0.18 7.5863 0.14 7.5863 0.11 7.5863 10.91 7.5863 7.33
PI-7 10 2 10.2635 0.78 10.2635 0.18 10.2635 0.14 10.2635 0.12 10.2635 11.30 10.2635 7.63
PI-8 10 3 9.0151 12.09 9.0151 0.18 9.0151 0.14 9.0151 0.12 9.0151 11.07 9.0151 7.46
PI-9 10 4 8.5098 1.69 8.5098 0.18 8.5098 0.14 8.5098 0.12 8.5098 11.06 8.5098 7.59
PI-10 11 2 11.4298 1.15 11.4298 0.18 11.4298 0.14 11.4298 0.12 11.4298 11.49 11.4298 7.84
PI-11 11 3 9.9455 59.43 9.9455 0.18 9.9455 0.14 9.9455 0.12 9.9455 11.49 9.9455 7.85
PI-12 11 4 9.3916 18.44 9.3916 0.19 9.3916 0.15 9.3916 0.12 9.3916 11.50 9.3916 8.00
PI-13 12 2 12.9442 2.03 12.9442 0.19 12.9442 0.15 12.9442 0.13 12.9442 11.92 12.9442 8.24
PI-14 12 3 11.2608 209.34 11.2608 0.19 11.2608 0.15 11.2608 0.13 11.2608 11.95 11.2608 8.27
PI-15 12 4 10.4824 47.50 10.4824 0.19 10.4824 0.15 10.4824 0.13 10.4824 11.93 10.4824 8.36
PI-16 13 2 14.4643 5.19 14.5251 0.20 14.5251 0.16 14.5251 0.13 14.5251 12.37 14.4643 8.71
PI-17 13 3 12.4245 975.45 12.5488 0.20 12.5488 0.16 12.5488 0.13 12.5388 12.41 12.5388 8.70
PI-18 13 4 11.5453 655.94 11.7422 0.20 11.7322 0.16 11.7322 0.13 11.7222 12.41 11.7054 8.83
PI-19 14 2 15.8371 8.70 16.2817 0.20 16.2747 0.16 16.2617 0.14 16.2517 12.95 16.2042 9.09
PI-20 14 3 13.4706 5453.10 13.8948 0.20 13.8878 0.16 13.8748 0.14 13.8648 12.99 13.8344 9.09
PI-21 14 4 12.4624 4674.83 12.8737 0.21 12.8657 0.16 12.8637 0.14 12.8437 13.01 12.8228 9.23
PI-22 15 2 17.1866 14.59 17.7664 0.21 17.7584 0.17 17.7464 0.15 17.7364 13.53 17.7344 9.51
PI-23 15 3 14.4428 7203.65 14.1108 0.21 14.1028 0.17 14.0908 0.15 14.0808 13.57 14.0495 9.54
PI-24 15 4 13.2278 7203.71 13.1879 0.21 13.1799 0.17 13.1779 0.15 13.1579 13.60 13.0375 9.70
PI-25 16 2 18.7701 28.98 19.4333 0.21 19.4273 0.17 19.4233 0.15 19.4033 13.98 19.3887 9.92
PI-26 16 3 15.5692 7202.86 15.0805 0.21 15.0725 0.17 15.0705 0.15 15.0505 13.98 14.9872 9.92
PI-27 16 4 14.1527 7202.95 14.0520 0.22 14.0440 0.17 14.0420 0.15 14.0220 13.98 13.9120 10.04
PI-28 17 2 20.3710 72.00 21.1207 0.22 21.1077 0.18 21.1007 0.16 21.0807 14.46 21.0706 10.36
PI-29 17 3 16.7674 7202.66 16.2270 0.22 16.2070 0.18 16.2070 0.16 16.1870 14.46 16.1216 10.33
PI-30 17 4 15.1667 7202.42 15.0626 0.23 15.0426 0.18 15.0426 0.16 15.0226 14.48 14.8668 10.46

Average: 12.1899 1848.83 12.2582 0.20 12.2539 0.15 12.2516 0.13 12.2442 12.33 12.2193 8.56

ASAEA algorithms. The maximum ASAEA optimality gap did not ex-
ceed 1.59%. Note that the optimality gaps are not shown in boxplots for
problem instances PI-23, PI-24, PI-26, PI-27, PI-29, and PI-30 (see
Fig. 8), as CPLEX was not able to find the global optimal solutions for
the latter instances within the computational time limit imposed.
6.4. Comprehensive comparative analysis against the alternative
metaheuristics
The fourth step of the computational experiments aimed to conduct
a comprehensive comparative analysis of the developed ASAEA algo-
rithm against the alternative EA-based and state-of-the-art metaheur-
istics for the large-size problem instances [PI-31 through PI-60] in
terms of the objective function and computational time values. A total
of 10 replications were performed for each algorithm in order to obtain
the average objective function and computational time values for the
large-size problem instances. Tables 8 and 9 present the instance
number, the number of vessels arriving for service at the MCT, the
number of the MCT berthing positions available for vessel service, the
average objective function values that were returned by the considered
algorithms, and the required computational time values. Based on the
information, provided in Tables 8 and 9, it can be observed that the
ASAEA algorithm clearly outperforms the alternative EA-based and
state-of-the-art metaheuristics in terms of the objective function values
at convergence. Specifically, the ASAEA algorithm outperformed the
EA, DPCEA, AEA, and SAEA algorithms on average by 16.41%, 15.03%,
12.50%, and 8.29%, respectively, in terms of the objective function
values for the generated large-size problem instances. Furthermore, the
ASAEA algorithm outperformed the TS, SA, VNS, ACO, and PSO algo-
rithms on average by 19.25%, 19.20%, 18.88%, 16.30%, and 16.08%,
respectively, in terms of the objective function values for the generated
large-size problem instances.
Superiority of the ASAEA algorithm over the alternative EA-based
and state-of-the-art metaheuristics confirms efficiency of the proposed
augmented self-adaptive strategy for parameter control, where not only
the algorithmic parameter values are encoded in the solutions, but they
are also updated based on the feedback from the search. Other meta-
heuristic algorithms that rely on parameter control (i.e., DPCEA, AEA,
and SAEA) returned lower objective function values as compared to the
metaheuristic algorithms that solely rely on parameter tuning (i.e., EA,
TS, SA, VNS, ACO, and PSO). The latter finding can be supported by the
fact that the parameter tuning strategy does not allow the parameter
values evolving throughout the algorithmic run, which may not be fa-
vorable for the search process (as it was demonstrated throughout the
conducted analysis). Also, single-solution-based algorithms (i.e., TS, SA,
and VNS) were generally outperformed by population-based algorithms
(i.e., EA, DPCEA, AEA, SAEA, ASAEA, ACO, and PSO), as they work
with one solution at a time, which imposes substantial limitations on
explorative capabilities of such algorithms. On the other hand, popu-
lation-based algorithms work with a population of solutions and can
explore different domains of the search space in each generation/
iteration.
The scope of the numerical experiments also included an assessment
whether the difference between the average objective function values,
returned by the ASAEA algorithm for the large-size problem instances,
and the average objective function values, returned by the EA, DPCEA,
AEA, SAEA, TS, SA, VNS, ACO, and PSO algorithms, was statistically
significant. A set of paired z-tests were conducted in order to accom-
plish the latter task. Based on the null hypothesis (H0 ), the average
objective function values, returned by the ASAEA algorithm, were as-
sumed to be equal to the average objective function values, returned by
alternative algorithm Alg H0: TCASAEA = TCAlg . On the other hand,

17
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Fig. 8. The optimality gaps over the generated scenarios for the considered algorithms and problem instances [PI-1 through PI-30].

Table 8
Objective function and computational time values obtained by the considered EA-based metaheuristic algorithms for problem instances [PI-31 through PI-60].
Instance No. of No. of EA DPCEA AEA SAEA ASAEA
Vessels Berths
6 6 6 6
Objective, 10 CPU, sec. Objective, 10 CPU, sec. Objective, 10 CPU, sec. Objective, 10 CPU, sec. Objective, 106 CPU, sec.
USD USD USD USD USD

PI-31 75 6 60.877 55.29 60.354 55.66 59.336 98.25 57.077 98.29 52.086 142.88
PI-32 75 8 57.821 55.23 56.585 55.58 55.410 98.97 53.444 99.31 49.843 146.88
PI-33 75 10 56.294 55.46 55.693 56.07 54.596 99.86 53.018 99.91 48.380 145.21
PI-34 80 6 65.827 58.72 64.967 59.11 63.632 105.19 62.017 105.30 57.017 152.98
PI-35 80 8 61.976 57.83 61.664 58.28 59.583 104.91 57.682 104.96 54.168 153.04
PI-36 80 10 61.920 58.23 60.648 58.60 59.384 105.12 57.058 105.18 53.270 153.28
PI-37 85 6 71.701 60.96 70.843 61.63 69.835 110.95 66.792 111.65 61.688 164.79
PI-38 85 8 67.176 61.58 66.822 61.98 65.221 111.28 61.948 111.69 57.048 165.27
PI-39 85 10 65.764 61.00 65.712 61.76 64.430 110.50 61.978 110.68 56.478 163.23
PI-40 90 6 76.910 64.37 76.537 65.11 73.823 116.66 71.594 116.95 66.843 172.19
PI-41 90 8 70.656 64.90 70.080 65.06 69.390 117.65 66.484 117.74 60.813 175.49
PI-42 90 10 71.143 65.08 70.207 65.24 67.971 117.69 65.637 119.11 60.319 173.14
PI-43 95 6 82.389 68.87 81.070 69.38 79.836 122.88 76.118 123.34 70.788 182.74
PI-44 95 8 76.146 68.97 75.318 69.58 72.512 125.38 70.659 125.42 64.761 184.10
PI-45 95 10 75.491 67.73 73.968 68.57 72.788 123.20 69.799 123.31 65.230 182.12
PI-46 100 6 86.876 71.55 86.681 72.12 85.159 131.35 81.433 131.64 75.549 192.83
PI-47 100 8 79.747 72.69 78.781 72.79 77.777 131.93 74.292 131.97 68.251 193.00
PI-48 100 10 79.136 71.39 78.070 72.98 76.524 129.53 73.625 130.38 67.892 192.10
PI-49 105 6 93.265 74.40 91.115 75.07 89.702 135.36 85.991 135.61 79.010 200.57
PI-50 105 8 83.025 74.65 82.853 75.41 81.266 136.47 78.121 136.57 71.926 201.19
PI-51 105 10 82.580 74.39 82.513 76.23 79.537 135.62 76.909 135.79 72.214 200.00
PI-52 110 6 97.995 77.39 97.218 77.96 94.429 141.03 90.845 141.22 83.603 208.26
PI-53 110 8 87.435 78.94 85.777 79.40 84.606 143.78 81.786 144.20 74.692 212.45
PI-54 110 10 86.655 77.58 86.199 78.61 82.986 140.72 80.906 141.99 74.317 208.53
PI-55 115 6 104.241 79.90 101.522 80.49 100.522 146.28 96.251 146.31 88.620 213.36
PI-56 115 8 92.329 80.74 90.786 81.50 88.430 147.56 85.099 148.29 78.791 214.20
PI-57 115 10 91.314 79.60 88.696 80.15 87.251 145.71 84.549 145.81 78.404 212.61
PI-58 120 6 109.817 82.73 108.911 83.28 105.920 151.30 101.529 151.50 93.805 222.81
PI-59 120 8 96.543 83.86 95.123 84.65 92.848 153.40 89.727 154.14 83.543 223.54
PI-60 120 10 94.593 82.71 94.216 83.51 92.290 151.70 87.718 151.91 81.097 220.97

Average: 79.588 69.56 78.631 70.19 76.900 126.34 74.003 126.67 68.348 185.79

18
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Table 9
Objective function and computational time values obtained by the considered state-of-the-art metaheuristic algorithms for problem instances [PI-31 through PI-60].
Instance No. of No. of TS SA VNS ACO PSO
Vessels Berths
6 6 6 6
Objective, 10 CPU, sec. Objective, 10 CPU, sec. Objective, 10 CPU, sec. Objective, 10 CPU, sec. Objective, 106 CPU, sec.
USD USD USD USD USD

PI-31 75 6 62.703 1.17 62.452 1.06 62.485 0.99 61.524 58.67 61.208 49.53
PI-32 75 8 58.549 1.18 58.734 1.07 58.464 1.01 56.740 59.36 56.930 49.26
PI-33 75 10 58.392 1.18 57.945 1.07 58.135 1.00 56.091 59.44 56.992 50.29
PI-34 80 6 67.715 1.23 67.864 1.11 67.185 1.05 66.283 61.50 66.010 52.07
PI-35 80 8 63.631 1.23 63.517 1.11 63.854 1.05 62.160 61.81 62.560 52.50
PI-36 80 10 63.006 1.23 62.746 1.12 63.030 1.05 61.593 62.40 61.553 53.05
PI-37 85 6 73.374 1.30 73.558 1.17 73.552 1.13 70.947 64.93 71.361 54.72
PI-38 85 8 68.782 1.29 68.457 1.17 68.629 1.14 67.022 65.59 66.444 55.37
PI-39 85 10 67.952 1.28 68.134 1.16 67.665 1.12 66.724 65.04 66.766 55.13
PI-40 90 6 78.446 1.34 79.200 1.22 78.721 1.18 76.346 68.06 76.105 58.10
PI-41 90 8 73.112 1.36 72.964 1.24 72.137 1.21 71.437 69.44 70.858 59.19
PI-42 90 10 72.514 1.35 72.530 1.23 72.262 1.18 71.054 69.09 70.690 58.79
PI-43 95 6 84.371 1.41 84.157 1.28 84.252 1.25 82.148 71.58 82.092 61.12
PI-44 95 8 77.579 1.43 77.249 1.30 76.529 1.27 76.224 72.91 75.666 62.29
PI-45 95 10 76.791 1.40 76.802 1.27 75.738 1.23 75.498 71.63 75.223 61.80
PI-46 100 6 89.494 1.49 89.551 1.36 89.311 1.33 86.547 76.25 87.344 65.39
PI-47 100 8 82.071 1.50 81.681 1.37 82.216 1.35 79.114 76.78 79.863 65.61
PI-48 100 10 81.415 1.48 81.050 1.36 81.291 1.32 79.094 76.21 78.690 64.86
PI-49 105 6 94.534 1.53 94.934 1.39 94.700 1.36 93.431 78.58 91.998 67.00
PI-50 105 8 86.255 1.53 86.127 1.40 85.240 1.38 84.114 78.63 83.392 66.93
PI-51 105 10 84.567 1.56 85.129 1.42 84.881 1.39 82.981 79.90 82.937 68.09
PI-52 110 6 99.835 1.60 99.862 1.46 99.534 1.43 97.028 81.96 97.048 70.72
PI-53 110 8 89.488 1.61 90.194 1.47 88.911 1.44 88.351 82.93 87.399 71.02
PI-54 110 10 89.117 1.59 88.761 1.45 88.803 1.42 87.308 81.72 86.748 69.85
PI-55 115 6 106.296 1.65 106.399 1.52 106.326 1.46 104.038 83.71 102.364 72.98
PI-56 115 8 94.402 1.65 93.780 1.52 94.153 1.46 91.538 83.93 91.100 72.88
PI-57 115 10 92.808 1.65 92.777 1.51 92.848 1.44 90.153 83.51 90.500 72.51
PI-58 120 6 112.379 1.72 111.961 1.58 111.562 1.52 108.751 87.22 108.809 76.18
PI-59 120 8 98.321 1.71 98.574 1.59 98.023 1.51 96.117 87.37 95.842 76.04
PI-60 120 10 97.304 1.71 97.351 1.58 97.240 1.51 94.335 86.89 95.009 75.90

Average: 81.507 1.45 81.481 1.32 81.256 1.27 79.490 73.57 79.317 62.97

Table 10 tests, including the test number, the algorithms compared, the number
The results obtained from the conducted paired z-tests. of observations (which is equal to 30, as all the generated large-size
Test Algorithms Number of TC , 106 STD , 106 Test
problem instances were considered), the average objective function
Compared Observations USD USD Statistic (z) values (TC ), the objective function value standard deviations (STD ),
and the test statistic values (i.e., z-values).
1 ASAEA 30 68.348 12.023 3.312 Considering a 5.00% significance level (which corresponds to a
EA 30 79.588 14.174
critical z-value of 1.96), the null hypothesis was rejected for all the
2 ASAEA 30 68.348 12.023 3.063 conducted paired z-tests, except test “4”. Therefore, the average ob-
DPCEA 30 78.631 13.910
jective function values, returned by the ASAEA algorithm for the large-
3 ASAEA 30 68.348 12.023 2.579 size problem instances, were statistically different (i.e., statistically
AEA 30 76.900 13.616
lower) than the average objective function values, returned by the EA,
4 ASAEA 30 68.348 12.023 1.750 DPCEA, AEA, TS, SA, VNS, ACO, and PSO algorithms. Although, the
SAEA 30 74.003 12.989 null hypothesis was not rejected for test “4” at a 5.00% significance
5 ASAEA 30 68.348 12.023 3.844 level, the ASAEA algorithm outperformed the SAEA algorithm on
TS 30 81.507 14.387 average by 8.29% in terms of the objective function values for the
6 ASAEA 30 68.348 12.023 3.832 generated large-size problem instances, which can be considered as
SA 30 81.481 14.418 substantial from the practical standpoint.
7 ASAEA 30 68.348 12.023 3.774 As for the computational time, the highest computational time va-
VNS 30 81.256 14.366 lues were generally recorded for ASAEA, which can be explained by two
8 ASAEA 30 68.348 12.023 3.305 major factors. First, the crossover and mutation probabilities are en-
ACO 30 79.490 14.012 coded within the ASAEA chromosomes. Therefore, the additional op-
9 ASAEA 30 68.348 12.023 3.277 erations are required in order to alter the crossover and mutation
PSO 30 79.317 13.842 probabilities throughout evolution of the algorithm. Second, storage of
the promising crossover and mutation probability values in an external
archive in each generation as well as periodic update of the crossover
the alternative hypothesis (Ha ) assumed that the average objective and mutation probability values in the current population chromo-
function values, returned by the ASAEA algorithm, were different from somes incur additional computational time. However, the maximum
the average objective function values, returned by alternative algorithm ASAEA computational time did not exceed 3.73 min. Therefore, the
Alg Ha: TCASAEA TCAlg . A total of 9 paired z-tests were performed, MCT operators will be able to develop and revise berthing plans in a
where the ASAEA objective function values, were compared against the timely manner. Furthermore, other algorithms that rely on parameter
objective function values, returned by 9 alternative metaheuristic al- control (i.e., DPCEA, AEA, and SAEA) typically required more time to
gorithms. Table 10 contains the results from the conducted paired z-

19
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

converge as compared to the algorithms that solely rely on parameter
tuning (i.e., EA, TS, SA, VNS, ACO, and PSO). Such pattern can be
justified by the fact that the additional operations have to be conducted
in order to update the algorithmic parameter values, which further
increases the computational time. Also, single-solution-based algo-
rithms (i.e., TS, SA, and VNS) required less computational time to
converge as compared to population-based algorithms (i.e., EA, DPCEA,
AEA, SAEA, ASAEA, ACO, and PSO), as they work with one solution at a
time and, therefore, require less fitness function evaluations for the
same number of generations/iterations (which was set to 6000 gen-
erations/iterations for single-solution-based algorithms and population-
based algorithms – see Tables 4 and 5).
Based on the conducted computational experiments, the developed
ASAEA algorithm outperformed the alternative EA-based algorithms
and the alternative state-of-the-art metaheuristic algorithms, which
have been widely used in the MCT and freight terminal operations lit-
erature in terms of the objective function values at termination for the
small-size and large-size problem instances. Therefore, the proposed
solution algorithm will be able to assist the MCT operators with the
development of cost-effective berth schedules and minimize the total
vessel waiting time, the total vessel handling time, and the total vessel
late departures (as the costs associated with the total vessel waiting
time, the total vessel handling time, and the total vessel late departures
are the components of the objective function of the DBSP mathematical
model). Minimization of the total vessel late departures would be also
critical for liner shipping companies, since the vessels would be served
in a timely manner at a given port of call and would be able to deliver
the cargo to consecutive ports of call in a timely manner.
7. Concluding remarks and future research directions
An increasing demand for maritime transportation has been ob-
served during the recent decades. Proper planning of the marine con-
tainer terminal operations is required in order to cope with the growing
demand. Effective berth schedules are considered as critical compo-
nents throughout planning of the marine container terminal operations,
as berth schedules directly influence the terminal throughput as well as
service completion of the arriving vessels. This study presented a
mixed-integer linear programming mathematical model for the berth
scheduling problem, where the summation of waiting costs, handling
costs, and late departure costs of the vessels was minimized. An in-
novative Evolutionary Algorithm, which relied on the augmented self-
adaptive parameter control strategy, was designed to solve the
developed mathematical model. Based on the proposed augmented self-
adaptive parameter control strategy, not only the algorithmic para-
meter values were encoded in the solutions, but they were also peri-
odically updated based on the feedback from the search.
According to the conducted computational experiments, the objec-
tive function values, returned by the developed augmented self-adap-
tive Evolutionary Algorithm, did not differ from the benchmark ob-
jective function values, obtained by CPLEX, by more than 1.59% over
the considered problem instances. The proposed augmented self-adap-
tive parameter control strategy was found to be more efficient as
compared to the deterministic, adaptive, and self-adaptive parameter
control strategies, which were previously used in the marine container
terminal operations literature. Furthermore, the augmented self-adap-
tive Evolutionary Algorithm outperformed the alternative state-of-the-
art metaheuristic algorithms, which have been widely used in the
marine container terminal operations literature in terms of the objec-
tive function values at convergence. The computational time of the
developed algorithm did not exceed 3.73 min, which will allow the
marine container terminal operators developing and revising berthing
plans in a timely manner. It was also found that the proposed aug-
mented self-adaptive parameter control strategy could improve the al-
gorithmic stability and decrease the stochastic variations in the objec-
tive function values at convergence.
There are several future research extensions for this study, including
the following: (1) development of new heuristics for generating the
initial population considering both population diversity and quality of
the produced solutions; (2) consideration of the alternative constraint
handling techniques within the developed algorithm; (3) design of the
alternative chromosome representations (e.g., the encoding may in-
clude additional algorithmic parameters, such as population size,
parent selection parameters, offspring selection parameters); (4) de-
velopment of various strategies for updating the algorithmic parameters
throughout the search process; (5) evaluation of the proposed algorithm
for multi-objective mathematical formulations; (6) consideration of the
alternative berthing layouts (e.g., continuous, hybrid, channel); and (7)
modeling uncertainties in the marine container terminal operations
(e.g., uncertainty in the vessel arrivals, uncertainty in the vessel
handling times, potential handling equipment breakdowns).
Declaration of Competing Interest
The authors declared that there is no conflict of interest.

Appendix A

The analysis of convergence patterns allows keeping track of the objective function value improvements from one generation to another (in case
of the considered EA-based metaheuristics – EA, DPCEA, AEA, SAEA, and ASAEA) or from one iteration to another (in case of the considered state-of-
the-art metaheuristics – TS, SA, VNS, ACO, and PSO). The convergence patterns were recorded for each one of the considered metaheuristic
algorithms for each replication and each one of the generated large-size problem instances. The results are provided in Fig. 9 for the EA, DPCEA,
AEA, SAEA, and ASAEA algorithms, while the convergence patterns of the TS, SA, VNS, ACO, and PSO algorithms are illustrated in Fig. 10. The
convergence patterns are shown only for the last replication and problem instances [PI-52 through PI-60], while similar tendencies were observed for
other replications and large-size problem instances. Also, the ASAEA convergence patterns are presented in Fig. 10 as well for comparison purposes.
The convergence patterns demonstrate that all the considered metaheuristic algorithms start the search process with the same solution, which
can be explained by the fact that the initial solutions of all the algorithms were generated using the same heuristic (i.e., the FCFS-SR heuristic – see
Section 5.3 of the manuscript). Based on the convergence pattern analysis, it can be noticed that ASAEA was able to identify the promising solutions
of the search space much faster as compared to the alternative EA-based and state-of-the-art metaheuristics. Therefore, the proposed augmented self-
adaptive parameter control strategy, which is deployed within the developed ASAEA algorithm, allows effective exploration of the search space and
identification of the domains with high-quality solutions. Moreover, the proposed augmented self-adaptive parameter control strategy improved the
ASAEA exploitative capabilities as well, since the ASAEA algorithm was able to find the solutions with superior objective function values within the
identified domains of the search space before convergence (i.e., quite significant objective function improvements were recorded even after gen-
eration ≈5000).
The SAEA and AEA algorithms, which rely on the self-adaptive and adaptive parameter control strategies, respectively, were also able to discover
the promising solutions fairly quickly. It can be observed that the explorative capabilities declined for DPCEA, when comparing to AEA, SAEA, and
ASAEA. The latter finding underlines that the augmented self-adaptive, self-adaptive, and adaptive parameter control strategies are more favorable
for the search process than the deterministic parameter control strategy, which is adopted within DPCEA. Lack of the explorative and exploitative

20
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Fig. 9. The EA, DPCEA, AEA, SAEA, and ASAEA convergence patterns for problem instances [PI-52 through PI-60].

Fig. 10. The TS, SA, VNS, ACO, PSO, and ASAEA convergence patterns for problem instances [PI-52 through PI-60].
21
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

capabilities was noticed for the algorithms that solely rely on parameter tuning (i.e., EA, TS, SA, VNS, ACO, and PSO). Such tendency can be
supported by the fact that the parameter tuning strategy does not alter the parameter values throughout the algorithmic run, which substantially
slows down the search process. Furthermore, a premature convergence was generally observed for single-solution-based algorithms (i.e., TS, SA, and
VNS). Since single-solution-based algorithms perform operations with one solution at a time, they are not able to effectively explore various domains
of the search space, which further results in a premature convergence.

Appendix B

The EA-based and state-of-the-art metaheuristic algorithms were found to be efficient for solving complex decision problems [13]; however, they
deploy stochastic operators (e.g., crossover, mutation, selection). The latter leads to the stochastic variations in the objective function values,
returned by such algorithms at convergence, from one algorithmic run to another. Significant stochastic variations in the objective function values,
obtained by a given algorithm, are not favorable, as such algorithm cannot be implemented in practice (e.g., the MCT operator will not be able to
develop a berth schedule using a given algorithm, if the total vessel service cost drastically fluctuates from one replication to another). The scope of
the computational experiments included the estimation of the objective function stochastic variations for the candidate metaheuristic algorithms.
The coefficient of variation in the objective function values, obtained by each algorithm, was estimated over 10 replications for each large-size
problem instance. The results from the conducted analysis are presented in Fig. 11 for each algorithm over all the generated large-size problem
instances using a boxplot format. A detailed description of the boxplot components is provided in Section 6.3 of this manuscript.

Fig. 11. The objective function coefficient of variation values for the developed algorithms and problem instances [PI-31 through PI-60].

It was found that the coefficient of variation in the objective function values did not exceed 1.4% over the considered metaheuristic algorithms,
which can be considered as acceptable from the practical standpoint. Lower coefficient of variation values were typically recorded for the EA-based
algorithms that rely on parameter control (i.e., DPCEA, AEA, SAEA, and ASAEA). Moreover, the least objective function coefficient of variation
values were determined for the ASAEA algorithm. Specifically, the ASAEA objective function coefficient of variation values did not exceed 0.5% for
the generated large-size problem instances. The latter finding indicates that the proposed augmented self-adaptive parameter control strategy can
reduce the stochastic variations in the objective function values, returned by the algorithm at convergence, and improve the algorithmic stability.

Appendix C

As it was mentioned earlier, the crossover and mutation probabilities are encoded in the chromosomes of the SAEA and ASAEA algorithms, and
evolve in each generation throughout the search process. The scope of the computational experiments included the analysis of changes in the
crossover and mutation probability values, encoded within the SAEA and ASAEA chromosomes, from the search start towards the convergence. The
average crossover and mutation probability values (over all the individuals within the SAEA population and the ASAEA population, respectively)
were recorded in each generation and for each one of the generated large-size problem instances. Evolution of the crossover probability values for the
SAEA and ASAEA algorithms is provided in Fig. 12, while Fig. 13 illustrates evolution of the mutation probability values for the SAEA and ASAEA
algorithms. Evolution of the crossover and mutation probability values is shown only for the last replication and problem instances [PI-52 through
PI-60], while similar tendencies were observed for other replications and large-size problem instances.
Based on the results from the conducted analysis, it was found that both SAEA and ASAEA similarly adopted fairly high values for the crossover
probabilities (about 0.70–0.80) and for the mutation probabilities (about 0.05–0.06) at the beginning of the search process. High crossover and
mutation probability values allowed the SAEA and ASAEA algorithms fast identification of the search space domains, which contain good-quality
solutions. However, throughout the search process, both algorithms started decreasing the crossover and mutation probability values in order to
effectively exploit the discovered search space domains for more promising solutions. Lower crossover and mutation probability values were

22
M. Kavoosi, et al. Advanced Engineering Informatics 42 (2019) 100972

Fig. 12. Evolution of the crossover probability values for the SAEA and ASAEA algorithms and problem instances [PI-52 through PI-60].

Fig. 13. Evolution of the mutation probability values for the SAEA and ASAEA algorithms and problem instances [PI-52 through PI-60].

23
M. Kavoosi, et al.
generally recorded for the ASAEA chromosomes as compared to the SAEA chromosomes. Therefore, lower crossover and mutation probability values
were more favorable for the search process towards convergence, since ASAEA outperformed SAEA in terms of the objective function values (more
details are provided in Section 6.4 of the manuscript). The proposed augmented self-adaptive parameter control strategy allows ASAEA sampling the
most promising crossover and mutation probability values periodically from the external archive. On the other hand, SAEA does not use the external
archive and was not able to adjust the crossover and mutation probability values as effectively as ASAEA throughout the search process.
References
[1] UNCTAD, Review of Maritime Transport 2017, United Nations Conference on Trade
and Development, New York, NY, Geneva, 2017http://unctad.org/en/
PublicationsLibrary/rmt2017_en.pdf.
[2] E. Elbeltagi, T. Hegazy, D. Grierson, Comparison among five evolutionary-based
optimization algorithms, Adv. Eng. Inf. 19 (1) (2005) 43–53.
[3] J. Garcıa, A. Berlanga, J.M. Molina, Evolutionary algorithms in multiply-specified
engineering. The MOEAs and WCES strategies, Adv. Eng. Inf. 21 (2007) 3–21.
[4] J.C. Fransoo, C.Y. Lee, The critical role of ocean container transport in global supply
chain performance, Prod. Oper. Manage. 22 (2) (2013) 253–268.
[5] C. Bierwirth, F. Meisel, A follow-up survey of berth allocation and quay crane
scheduling problems in container terminals, Eur. J. Oper. Res. 244 (3) (2015)
675–689.
[6] C.Y. Lee, D.P. Song, Ocean container transport in global supply chains: Overview
and research opportunities, Transport. Res. Part B: Methodol. 95 (2017) 442–474.
[7] M. Guerrero, F.G. Montoya, R. Baños, A. Alcayde, C. Gil, Community detection in
national-scale high voltage transmission networks using genetic algorithms, Adv.
Eng. Inf. 38 (2018) 232–241.
[8] M. Srinivas, L.M. Patnaik, Adaptive probabilities of crossover and mutation in ge-
netic algorithms, IEEE Trans. Syst. Man Cybern. 24 (4) (1994) 656–667.
[9] M.A. Dulebenets, M. Kavoosi, O. Abioye, J. Pasha, A self-adaptive evolutionary
algorithm for the berth scheduling problem: towards efficient parameter control,
Algorithms 11 (7) (2018) 1–35.
[10] X.S. Yang, Nature-inspired Optimization Algorithms, Elsevier, 2014.
[11] M.A. Dulebenets, A novel Memetic Algorithm with a deterministic parameter con-
trol for efficient berth scheduling at marine container terminals, Maritime Bus. Rev.
2 (4) (2017) 302–330.
[12] M.A. Dulebenets, Application of evolutionary computation for berth scheduling at
marine container terminals: parameter tuning versus parameter control, IEEE Trans.
Intell. Transp. Syst. 19 (1) (2018) 25–37.
[13] A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, Springer, 2015.
[14] H.J. Carlo, I.F. Vis, K.J. Roodbergen, Seaside operations in container terminals:
literature overview, trends, and research directions, Flexible Serv. Manuf. J. 27
(2–3) (2015) 224–262.
[15] N. Lang, A. Veenstra, A quantitative analysis of container vessel arrival planning
strategies, OR Spectrum 32 (3) (2010) 477–499.
[16] M. Golias, M. Boile, S. Theofanis, C. Efstathiou, The berth-scheduling problem:
Maximizing berth productivity and minimizing fuel consumption and emissions
production, Transport. Res. Rec.: J. Transport. Res. Board 2166 (2010) 20–27.
[17] Y. Du, Q. Chen, X. Quan, L. Long, R.Y. Fung, Berth allocation considering fuel
consumption and vessel emissions, Transport. Res. Part E: Logist. Transport. Rev. 47
(6) (2011) 1021–1037.
[18] S. Wang, Q. Meng, Z. Liu, A note on “Berth allocation considering fuel consumption
and vessel emissions”, Transport. Res. Part E: Logist. Transport. Rev. 49 (1) (2013)
48–54.
[19] G. Venturini, Ç. Iris, C.A. Kontovas, A. Larsen, The multi-port berth allocation
problem with speed optimization and emission considerations, Transport. Res. Part
D: Transport Environ. 54 (2017) 142–159.
[20] M.A. Dulebenets, R. Moses, E.E. Ozguven, A. Vanli, Minimizing carbon dioxide
emissions due to container handling at marine container terminals via hybrid
evolutionary algorithms, IEEE Access 5 (2017) 8131–8147.
[21] D. Chang, Z. Jiang, W. Yan, J. He, Integrating berth allocation and quay crane
assignments, Transport. Res. Part E: Logist. Transport. Rev. 46 (6) (2010) 975–990.
[22] M.W. Li, W.C. Hong, J. Geng, J. Wang, Berth and quay crane coordinated sche-
duling using multi-objective chaos cloud particle swarm optimization algorithm,
Neural Comput. Appl. 28 (11) (2017) 3163–3182.
[23] S. Esmer, I.B. Cetin, O. Tuna, A simulation for optimum terminal truck number in a
Turkish port based on lean and green concept, Asian J. Ship. Logist. 26 (2) (2010)
277–296.
[24] J. He, W. Zhang, Y. Huang, W. Yan, A simulation optimization method for internal
trucks sharing assignment among multiple container terminals, Adv. Eng. Inf. 27 (4)
(2013) 598–614.
[25] J. He, Y. Huang, W. Yan, Yard crane scheduling in a container terminal for the
trade-off between efficiency and energy consumption, Adv. Eng. Inf. 29 (1) (2015)
59–75.
[26] J. He, Y. Huang, W. Yan, S. Wang, Integrated internal truck, yard crane and quay
crane scheduling in a container terminal considering energy consumption, Expert
Syst. Appl. 42 (5) (2015) 2464–2487.
[27] G. Chen, K. Govindan, M.M. Golias, Reducing truck emissions at container terminals
in a low carbon economy: proposal of a queueing-based bi-objective model for
optimizing truck arrival pattern, Transport. Res. Part E: Logist. Transport. Rev. 55
(2013) 3–22.
24
Advanced Engineering Informatics 42 (2019) 100972
[28] M. Golias, I. Portal, D. Konur, E. Kaisar, G. Kolomvos, Robust berth scheduling at
marine container terminals via hierarchical optimization, Comput. Oper. Res. 41
(2014) 412–422.
[29] E. Ursavas, S.X. Zhu, Optimal policies for the berth allocation problem under sto-
chastic nature, Eur. J. Oper. Res. 255 (2) (2016) 380–387.
[30] N. Umang, M. Bierlaire, A.L. Erera, Real-time management of berth allocation with
stochastic arrival and handling times, J. Sched. 20 (1) (2017) 67–83.
[31] X. Xiang, C. Liu, L. Miao, Reactive strategy for discrete berth allocation and quay
crane assignment problems under uncertainty, Comput. Ind. Eng. 126 (2018)
196–216.
[32] R. Moorthy, C.P. Teo, Berth management in container terminal: the template design
problem, Container Terminals and Cargo Systems, Springer, 2007, pp. 63–86.
[33] M. Hendriks, M. Laumanns, E. Lefeber, J.T. Udding, Robust cyclic berth planning of
container vessels, OR Spectrum 32 (3) (2010) 501–517.
[34] Y. Xu, Q. Chen, X. Quan, Robust berth scheduling with uncertain vessel delay and
handling time, Ann. Oper. Res. 192 (1) (2012) 123–140.
[35] L. Zhen, D.F. Chang, A bi-objective model for robust berth allocation scheduling,
Comput. Ind. Eng. 63 (1) (2012) 262–273.
[36] Y. Alix, B. Slack, C. Comtois, Alliance or acquisition? Strategies for growth in the
container shipping industry, the case of CP ships, J. Transp. Geogr. 7 (3) (1999)
203–208.
[37] J.F. Ding, G.S. Liang, Using fuzzy MCDM to select partners of strategic alliances for
liner shipping, Inf. Sci. 173 (1–3) (2005) 197–225.
[38] L. Lei, C. Fan, M. Boile, S. Theofanis, Collaborative vs. non-collaborative container-
vessel scheduling, Transport. Res. Part E: Logist. Transport. Rev. 44 (3) (2008)
504–520.
[39] P.M. Panayides, R. Wiedmer, Strategic alliances in container liner shipping, Res.
Transport. Econ. 32 (1) (2011) 25–38.
[40] D. Yang, M. Liu, X. Shi, Verifying liner shipping alliance’s stability by applying core
theory, Res. Transport. Econ. 32 (1) (2011) 15–24.
[41] A. Imai, E. Nishimura, S. Papadimitriou, Berthing ships at a multi-user container
terminal with a limited quay capacity, Transport. Res. Part E: Logist. Transport.
Rev. 44 (1) (2008) 136–151.
[42] J. Peng, Z. Zhou, R. Li, A collaborative berth allocation problem with multiple ports
based on genetic algorithm, J. Coastal Res. 73 (1) (2015) 290–297.
[43] M.A. Dulebenets, M.M. Golias, S. Mishra, A collaborative agreement for berth al-
location under excessive demand, Eng. Appl. Artif. Intell. 69 (2018) 76–92.
[44] M.M. Golias, H.E. Haralambides, Berth scheduling with variable cost functions,
Maritime Econ. Logist. 13 (2) (2011) 174–189.
[45] A. Imai, Y. Yamakawa, K. Huang, The strategic berth template problem, Transport.
Res. Part E: Logist. Transport. Rev. 72 (2014) 77–100.
[46] C. Arango, P. Cortés, J. Muñuzuri, L. Onieva, Berth allocation planning in Seville
inland port by simulation and optimisation, Adv. Eng. Inf. 25 (3) (2011) 452–461.
[47] R.M. De Oliveira, G.R. Mauri, L.A.N. Lorena, Clustering search for the berth allo-
cation problem, Expert Syst. Appl. 39 (5) (2012) 5499–5505.
[48] C. Cubillos, R. Díaz, E. Urra, D. Cabrera-Paniagua, G. Cabrera, G. Lefranc, An agent-
based solution for the berth allocation problem, Int. J. Comput. Commun. Control 8
(3) (2013) 384–394.
[49] D. Xu, C.L. Li, J.Y.T. Leung, Berth allocation with time-dependent physical limita-
tions on vessels, Eur. J. Oper. Res. 216 (1) (2012) 47–56.
[50] E. Lalla-Ruiz, C. Expósito-Izquierdo, B. Melián-Batista, J.M. Moreno-Vega, A set-
partitioning-based model for the berth allocation problem under time-dependent
limitations, Eur. J. Oper. Res. 250 (3) (2016) 1001–1012.
[51] A. Dadashi, M.A. Dulebenets, M.M. Golias, A. Sheikholeslami, A novel continuous
berth scheduling model at multiple marine container terminals with tidal con-
siderations, Maritime Bus. Rev. 2 (2) (2017) 142–157.
[52] L. Zhen, Z. Liang, D. Zhuge, L.H. Lee, E.P. Chew, Daily berth planning in a tidal port
with channel flow control, Transport. Res. Part B: Methodol. 106 (2017) 193–217.
[53] C.J. Ting, K.C. Wu, H. Chou, Particle swarm optimization algorithm for the berth
allocation problem, Expert Syst. Appl. 41 (4) (2014) 1543–1550.
[54] S. Emde, N. Boysen, Berth allocation in container terminals that service feeder ships
and deep-sea vessels, J. Oper. Res. Soc. 67 (4) (2016) 551–563.
[55] K. Deep, M. Thakur, A new crossover operator for real coded genetic algorithms,
Appl. Math. Comput. 188 (1) (2007) 895–911.
[56] J.F. Cordeau, G. Laporte, P. Legato, L. Moccia, Models and tabu search heuristics for
the berth-allocation problem, Transport. Sci. 39 (4) (2005) 526–538.
[57] P. Hansen, C. Oğuz, N. Mladenović, Variable neighborhood search for minimum
cost berth allocation, Eur. J. Oper. Res. 191 (3) (2008) 636–649.
[58] A.B. Arabani, S.F. Ghomi, M. Zandieh, Meta-heuristics implementation for sche-
duling of trucks in a cross-docking system with temporary storage, Expert Syst.
Appl. 38 (3) (2011) 1964–1979.
[59] GAMS. Cutting Edge Modeling, 2019. [Online]. Available: https://www.gams.
com/.
[60] Mathworks, Release 2016a, 2019. [Online]. Available: https://www.mathworks.
M. Kavoosi, et al.
com/products/new_products/release2016a.html.
[61] M.A. Dulebenets, A delayed start parallel evolutionary algorithm for just-in-time
truck scheduling at a cross-docking facility, Int. J. Prod. Econ. 212 (2019) 236–258.
[62] M.A. Dulebenets, Evaluation of non-parametric selection mechanisms in evolu-
tionary computation: a case study for the machine scheduling problem, IntechOpen,
United Kingdom, 2018, pp. 23–45 (ISBN: 978-1-78923-329-2).
[63] M.A. Dulebenets, A Comprehensive evaluation of weak and strong mutation
25
Advanced Engineering Informatics 42 (2019) 100972
mechanisms in evolutionary algorithms for truck scheduling at cross-docking
terminals, IEEE Access 6 (2018) 65635–65650.
[64] M.A. Dulebenets, A diploid evolutionary algorithm for sustainable truck scheduling
at a cross-docking facility, Sustainability 10 (5) (2018) 1–23.
[65] Eniram, Evolution of shipping, 2019. [Online]. Available: https://www.eniram.fi/
evolution-of-shipping/.