Omega 103 (2021) 102444

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

An expanded robust optimisation approach for the berth allocation

problem considering uncertain operation time ✩
Xi Xiang a,b, Changchun Liu c,∗
a
Center of Excellence in Modeling and Simulation for Next Generation Logistics, National University of Singapore, 117602, Singapore
b
Department of Industrial Systems Engineering and Management, National University of Singapore, Singapore 117602, Singapore
c
Institute of Operations Research & Analytics, National University of Singapore, Singapore 117602,Singapore

a r t i c l e i n f o a b s t r a c t

Article history: Container terminals play a vital role as representative logistic facilities for contemporary trade by han-
Received 26 February 2020 dling outbound, inbound, and transshipment containers to and from the sea and hinterland. The increas-
Accepted 2 March 2021
ing number of containers and vessels poses new challenges to port management and resource schedul-
Available online 3 March 2021
ing, because of scarce land, high labour cost, and limited technical equipment. This study investigates the
Keywords: berth allocation planning problem at a tactical level considering uncertain operation time. Based on the
Tactical berth allocation problem historical data, we formulate a data-driven expanded robust optimisation model to minimise the total
Expanded robust optimisation cost of deviations between the planned and expected berthing time of the vessel. To solve the model,
Column-and-constraint generation we firstly use K-means clustering to construct the uncertainty set. Secondly, we present a column-and-
Data-driven constraint generation algorithm to solve the model. Extensive computational experiments are conducted
to verify the effectiveness of the proposed model and algorithm. Experiment results show that the pro-
posed model can not only guarantee the out-of-sample performance, which overcomes the vulnerability
of the sample average approximation approach but also avoid the over-conservatism of the traditional
robust optimisation model.
© 2021 Elsevier Ltd. All rights reserved.

1. Introduction [34], BAP can be divided into three variants in accordance with the
planning level: strategic, tactical, and operational.
The throughput of container terminals worldwide is increasing In previous decades, most scholars have focused on BAP at the
rapidly, and the importance of maritime transport for trade and operational level [20]. In this study, we focus on tactical berth al-
development cannot be overemphasised [43]. The increasing num- location planning (TBAP). In the TBAP, port planners must negoti-
ber of containers poses new challenges to port management, and ate with shipping liners to decide when and where the vessels are
efficient management of container terminals has become a criti- berthed. The shipping liners firstly inform the ports about their
cal component in port decision support systems [2]. A fundamen- expected berthing time, and the ports make the plan to satisfy
tal aspect of improving service quality is a reasonable allocation their requirements as much as possible [52]. Besides, the allocated
of limited resources. Scholars have paid attention to berth alloca- berthing position influences vessel operation time. To reduce the
tion problem (BAP) [6], yard allocation problem [29], empty con- distance between the allocated berth and storage locations, the
tainer repositioning problem [23], container pre-marshalling prob- vessel tends to dock at its favourite berth (home berth), resulting
lem [48], integrated problem [26], and so on. Among these sur- in the shortest operation time [34].
veys, the BAP plays an important role in container terminal man- Fig. 1 illustrates an example of TBAP. For TBAP, vessels ar-
agement. Amongst terminal resources, berths are closely related to rive periodically. As mentioned in [34,52,54], most vessels have a
and have a great influence on technical equipment and yards. Tra- weekly arrival pattern. For this kind of vessel, shipping liners tend
ditionally, the goal of BAP is to determine the berthing position to berth as close as possible to their expected berthing times to re-
and time for each arriving vessel at the terminal. As described in duce energy consumption. Therefore, the port planner must make
a tactical berth allocation plan to minimise the deviations between
the expected and planned berthing time of vessels. In this study,
✩
we aim to make a tactical plan for vessels with a weekly arrival
Area: Production Management, Scheduling and Logistics. This manuscript was
processed by Associate Editor Yagiura. pattern. Based on our plan, the port planners can adjust this tacti-
∗
Corresponding author. cal plan at the operational level to achieve an operational schedule
E-mail addresses: [email protected] (X. Xiang), [email protected] (C. Liu). each week when the information is accurate.

https://doi.org/10.1016/j.omega.2021.102444
0305-0483/© 2021 Elsevier Ltd. All rights reserved.
X. Xiang and C. Liu Omega 103 (2021) 102444

Fig. 1. An illustration of the tactical berth allocation problem.

In container operation management problems, prior data are be addressed to achieve our goal: how to obtain an uncertainty
widely used. Most studies have assumed that prior information, set depending on the collected historical data; and how to solve
i.e. the number of containers, is accurate and reliable. The oper- the ERO problem effectively given uncertainty set(s). To overcome
ation time of a vessel depends on the number of containers that these two challenges, we firstly present a data-driven approach for
will be handled. A vessel visits a port periodically, and its total uncertainty set construction via K-means clustering. A column and
number of containers fluctuates in each period since of some un- constraint generation (C&CG) algorithm is then proposed to solve
predicted factors. Furthermore, the unproductive reshuffles caused the ERO formulation.
by the stack position of containers also exert an influence on the The contributions of this study are listed as follows:
operation time. As information technology evolves, port planners
• First, we consider a TBAP for vessels with a weekly arrival
can collect historical data and make decisions responsive to certain
pattern and stochastic fluctuation in the number of contain-
changes that rely on objective data. Thus, we study a TBAP given
ers, which can obtain a tactical plan that can be given as
uncertain operation time based on a data-driven approach.
a reference for the port planners to make an operational
Zhen [52] proposed a stochastic programming (SP) formulation
schedule for one week when the information is accurate.
and a robust optimisation (RO) formulation for TBAP. SP has been a
• Second, we propose an ERO-TBAP model by employing the
popular mathematical method for finding optimal decisions under
K-means clustering technique to divide the uncertainty set
uncertainty over the past few decades. A key assumption in the
into K non-overlapping clusters, which could overcome the
SP is that the decision-maker must have complete information on
vulnerability of SP and avoid the over-conservatism of RO.
the distribution of the uncertainty, either through empirical data
• Finally, we propose an exact algorithm to solve the proposed
or subjective judgment. However, this requirement might turn out
model. To our best knowledge, few scholars have studied the
to be difficult or impossible to satisfy when a strategic decision
exact algorithm to solve BAP under uncertainty.
must be made in advance of the realisation of the uncertainty. For
the RO framework, no assumption is made on the probability dis- The remainder of this paper is organised as follows:
tribution of the uncertainty. The traditional proposed measure of Section 2 reviews the related works. Section 3 introduces the
robustness is the min-max cost approach in which the cost associ- models for the TBAP. Section 4 presents a solution methodology
ated with the worst-case scenario is minimised. However, the solu- for solving the ERO problem. Section 5 discusses the numerical ex-
tions obtained by RO always tend to be over-conservative. Different periments and analyses the results. Section 6 presents conclusions
from SP and RO, this study proposes an expanded robust optimi- and future research directions.
sation (ERO) model, which can effectively utilise data information
and avoid over-conservative solutions. In the ERO, multiple uncer- 2. Related works
tainty sets are used to characterise uncertainty instead of only one
set in the RO formulation. BAP has been greatly addressed in the literature. Vis and Koster
In recent years, research on data-driven methods has been well [46], Steenken et al. [42], Stahlbock and Voß [41], Bierwirth and
studied [14,40]. The data-driven approach used in this study is mo- Meisel [5], Bierwirth and Meisel [6] provided overviews on con-
tivated by Bertsimas et al. [3] which provided a data-driven RO tainer terminal operations. As mentioned before, BAP can be di-
method. Its main idea is to use limited data to design uncertainty vided into strategic, tactical, and operational levels. We will first
sets for RO. In this framework, it is assumed that planners only review BAP from these three aspects.
possess the information about history data of the number of con- For the strategic level, it considers a time frame ranging from
tainers. In previous literature, Bertsimas et al. [3] assumed that one year to several years, including decisions such as establish-
the uncertainty sets are defined by certain structures and sizes ing shared and dedicated berths, the arrangement of contracts, and
based on the data points available. While the structure of uncer- cooperation between terminals and shipping companies. Hendriks
tainty set in our study is not predefined, we consider the uncer- et al. [15] considered a terminal operator who provides container
tainty set which is constructed by historical data. Then we develop handling services at multiple terminals within the same port and
a data-driven ERO framework for TBAP with a stochastic num- address the problem of spreading a set of cyclically calling ves-
ber of containers. Two following major research challenges must sels over the various terminals and allocating a berthing and de-

2
X. Xiang and C. Liu Omega 103 (2021) 102444

parture time to each of them. Imai et al. [19] proposed a strategic TBAP and developed an SP and an RO model to decide berth posi-
berth template problem that selects the ships among the request- tions for vessels. The former can handle deviation in vessel opera-
ing ones to be served and arranges their berth-windows within a tion time with an arbitrary probability distribution, whereas the
limited planning horizon. latter can deal with situations with limited probability informa-
The decisions arising at the tactical level contains tactical berth tion. Iris and Lam [21] considered the concept of recoverable ro-
and yard templates, quay crane (QC) assignments, which covers op- bustness and integrated proactive and reactive approaches to han-
erations in a time horizon that ranges from one week up to sev- dle the uncertainty, which is on vessel arrival time and QC han-
eral months. Moorthy and Teo [34] first mentioned the concept dling rate. The present study considers a TBAP given the uncer-
of TBAP[9]. proposed a novel concept called “QC-profile” to study tainty in operation times and proposes an ERO approach. On one
the tactical discrete BAP and QC assignment problem. Vacca et al. hand, ERO could effectively leverage data information and opti-
[45] proposed an exact branch and price algorithm for the problem mise the objective value without any presumption about the prob-
introduced in [9]. A similar problem was subsequently studied by ability distribution of uncertainties, which SP typically relies on.
Xie et al. [51], who presented a branch-and-price algorithm based On the other hand, ERO can avoid over-conservative solutions by
on the framework of Dantzig–Wolfe decomposition to produce op- incorporating partial stochastic information, which is disregarded
timal solutions. Zhen et al. [54] integrated TBAP with yard tem- by RO.
plate planning. Meisel and Bierwirth [33] proposed a framework RO has been widely studied in the academic field. Bertsimas
for aligning all decisions of BAP, QC assignment, and QC scheduling and Sim [4] proposed the gamma uncertainty set to control the
problems in an integrative manner. Lalla-Ruiz et al. [24] proposed degree of conservatism and introduced the concept of the price
a biased random key genetic algorithm for TBAP. Zhen [52] pro- of robustness. With the increasing access to uncertainty data and
posed SP and RO models to solve TBAP under uncertainty. Huang recent advances in machine learning techniques, data-driven opti-
et al. [16] studied the continuous berth template problem. In our misation paves a new way to decision-making under uncertainty.
study, we will consider a TBAP for vessels with a weekly arrival Campbell and How [8] presented a Bayesian nonparametric, data-
pattern, which can obtain a tactical plan that can be given as a driven, nonconvex uncertainty set construction for RO. Shang et al.
reference for the port planners to make an operational schedule [39] constructed a data-driven uncertainty set for static RO by ex-
for one week. ploiting the kernel-based support vector clustering technique. Bert-
The operational level considers the shortest time horizon simas et al. [3] proposed a data-driven static RO framework based
among the three planning levels, which covers decisions ranging on statistical hypothesis tests. Ning and You [35] proposed a novel
from one up to several days. Imai et al. [17] considered a static data-driven RO framework that leverages the power of machine
BAP to minimise the sum of port staying times of vessels and learning and big data analytics for decision-making under uncer-
the dissatisfaction of the vessels in terms of berthing order. Imai tainty. In the current study, we use historical data as the uncer-
et al. [18] studied the problem of determining a dynamic berth tainty set and employ the K-means clustering technique to divide
assignment to vessels in a public berth system. Park and Kim the uncertainty set into K non-overlapping clusters. Based on the
[36] proposed a sub-gradient optimisation method and used a sim- obtained sets, we propose an ERO-TBAP model.
ulated annealing method. Guan and Cheung [12] considered the
problem of allocating space at berth for vessels to minimise total
weighted flow time. Lee and Chen [25] presented an optimisation- 3. Tactical berth allocation models
based approach for BAP. Buhrkal et al. [7] reviewed and described
three main models of the discrete dynamic BAP and compared In the paper of Zhen [52], they first introduced the determin-
them from a computational perspective. Mauri et al. [31] and istic TBAP (DTBAP), then extended it to the stochastic programming
Ribeiro et al. [38] used an adaptive, large neighbourhood search for TBAP (SPTBAP) model (see Section 3.3) and robust optimisation TBAP
BAP. (ROTBAP) models (see Section 3.4) by considering uncertainty. In
Most of the above-mentioned articles studied BAP under a de- our study, we form the basis for developing the expanded robust
terministic case. However, container terminal operations face a optimisation TBAP (EROTBAP) model. Notations are displayed as fol-
complex environment, i.e., typhoons, seaquakes, and rainstorms. lows:2
Therefore, several scholars have begun studying BAP under uncer- Indices
tainty recently. As mentioned in [27], mixed integer programming i Index of vessels, i = 1, . . . , V
(MIP) [11], SP [27,44,53,55], and RO [28,49] are the main meth- j Index of berths, j = 1, . . . , B
ods for addressing uncertainty problems. From the perspective of Parameters
the operational level, [11] introduced a risk measure to characterise V Total number of vessels
robustness and formulated a bi-objective MIP model. Given uncer- B Total number of berths
tain arrival and operation times, [55] formulated an SP model to T Time planning horizon
solve operational BAP with the assumption that the probability dis- ai Expected berthing time of vessel i, which is proposed by its
tribution of uncertain parameters is known in advance. Zhen and shipping liner
Chang [53] developed a bi-objective MIP model to find a trade-off ti j Operation time of vessel i if it berths at berth j
between system cost and robustness with the consideration of un- ci1 The waiting cost per unit time for vessel i
certain arrival and operation times. Liu et al. [27] proposed two SP ci2 The penalty cost per unit time of berth earlier than ai for
models by introducing a concept of service level for BAP under un- vessel i
certainty. Subsequently, [28] developed three two-stage RO mod- Decision variables
els for the BAP, including RO, expanded RO, and risk-constrained xi j =1, if vessel i berths at berth j; 0 otherwise
RO models, wherein the baseline schedule is made in the first yi j Continuous variable, berthing time of vessel i at berth j
stage and the recovery decision is made in the second stage. Xi- zii j =1, if vessel i is subsequently after vessel i at berth j; 0
ang et al. [49] made a trade-off between economic performance otherwise
and customer satisfaction by formulating a bi-objective RO model τi ∈ {−1, 0, 1}, auxiliary variables
for BAP. For the tactical level, Moorthy and Teo [34] used a novel θi j ∈ {0, 1}, auxiliary variables
sequence pair-based approach for a robust berth template to min- u+i
The waiting time for vessel i
imise total connectivity cost and expected delays[52]. considered u−i
The length of time that vessel i berths earlier than ai

3
X. Xiang and C. Liu Omega 103 (2021) 102444

Fig. 2. TBAP for vessels with a weekly arrival pattern and stochastic fluctuation in the number of containers.

3.1. Problem description tuation. Thus, operation time is assumed to be stochastic. Such an
assumption was also used in [13,28,44,49,52,55]. A stochastic pa-
This paper studies a tactical problem and we use Fig. 2 to il- rameter ξi is defined to denote the extra operation time needed.
lustrate the planning flow with regards to the creation of a tactical Instead of having a deterministic operation time vector t , we use
berthing plan. In the first step, we do not know the accurate infor- the notation t + ξ for the stochastic operation time.
mation about several future weeks. Therefore, we use the collected Remark. It is known that the deviation of vessels’ arrival time
historical data to make a tactical plan. In this tactical plan, every is also one of the major uncertainties in berth scheduling. This
week has the same schedule. Based on this plan, the terminal can paper focuses on the uncertain operation time which is caused
do some resource preparation. As time varies, we can obtain al- by the fluctuation of the demand for freight transportation. How-
most accurate information. Using this information, we adjust the ever, our model still works if we also consider the deviation of
original tactical plan to make an operational schedule. In this pa- a vessels’ arrival time. Suppose that ξi1 and ξi2 denote the devi-
per, we only focus on how to generate a tactical plan based on ation of vessels’ arrival time with respect to its expected arrival
historical data. time and extra operation time needed, respectively. On one hand,
The problem considered in this study has two important fea- if ξi1 ≤ 0, which means vessel i arrives before their expected ar-
tures: vessels with a weekly arrival pattern and stochastic fluctu- rival time, it does not influence the baseline schedule, since the
ation in the number of containers. Next, we will introduce these terminal can let vessel i wait at sea. In fact, vessels typically do
two features. not arrive early because they rather save fuel by reducing speed.
We first introduce the feature of vessels’ weekly arrival pattern, On the other hand, if ξi1 > 0, we can define the stochastic param-
which has been addressed in [47,52,54]. As shown in Fig. 2, ves- eter ξi = ξi1 + ξi2 to denote the sum of the deviations of a vessel’s
sels arrive at the terminal once a week, which will result in the arrival time and operation time, our model still works. In conclu-
periodicity of TBAP as shown in Fig. 3. In this paper, we aim to sion, our model can consider the deviation of vessels’ arrival time
make a tactical plan for BAP for several future weeks, while we by setting ξi = (ξi1 )+ + ξi2 .
do not consider the problem of making an operation schedule for
each week. In the notations, two auxiliary variables, i.e. θi j , τi , are 3.2. DTBAP
introduced to consider periodicity. For example, in Fig. 3, the cycle
time of a period is 24. Vessels 1, 2 and 3 expect to berth at time BAP aims to assign the berthing position and service time for
a1 = 23, a2 = 9 and a3 = 21. However, in the schedule, the planned each arriving vessel. This study aims to minimise the total cost, in-

berthing times ( Bj=1 yi j ) are 1, 10 and 22, respectively. Their op- cluding the waiting and earliness costs. In the study of Imai et al.
eration times are 9, 9, and 3, respectively. Based on the constraint [18], Park and Kim [37], Liu et al. [30], the waiting and earliness
(2), the deviations of three vessels are 2, 1, 1. More specifically, costs were considered. If the berthing time is later than the ves-
2 = |1 + 1 × 24 − 23|, 1 = |10 − 0 × 24 − 9|, 1 = |22 − 0 × 24 − 21|. sel’s arrival time, it may incur complaints from vessel owners, re-
The other feature is the stochastic fluctuation in the number of sulting in waiting cost. On the other hand, if the vessel’s scheduled
containers. TBAP involves uncertainties that stem from the unpre- berthing time is earlier than its expected arrival time, the vessel
dictability of the number of containers. As shown in Fig. 2, the ves- needs to speed up to meet the port’s berthing time, resulting in
sels inform the accurate number of containers, which can be trans- extra fuel consumption cost. Generally, the tardiness cost caused
formed into ti j , to the port before making a tactical plan. However, by delay in departure time is an important objective, which can

its total number of containers fluctuates in each period, since a be represented by ci3 ( Bj=1 (yi j + ti j xi j ) − di ), where ci3 and di de-
vessel visits a port periodically. Although information technology note penalty cost per unit time caused by delay in departure time
is very advanced nowadays, it is difficult to predict accurate fluc- and the departure time of vessel i. Since the departure time is re-

4
X. Xiang and C. Liu Omega 103 (2021) 102444

Fig. 3. TBAP with considering the periodicity.

lated to the arrival time and operation time, thus it can be rep-
resented by the sum of expected arrival time and operation time,
 
V
i.e, ai + Bj=1 ti j xi j . Furthermore, we do not consider the quay crane zii j = xi j , ∀i = 1, ., V ; j = 1, . . . , B (7)
scheduling problem and the operation time is supposed to be con- i =1,i =i
stant once we know the number of containers. Therefore, we use
the similar modeling method in [22,55], that is, we use the min-

V
imisation of the waiting cost instead of the tardiness cost. In other zi i j = xi j , ∀i = 1, ., V ; j = 1, . . . , B (8)
words, the cost coefficient ci2 can be seen as the sum of waiting i =1,i =i
and delay costs per unit time for vessel i if we consider the de-
lay cost in departure. The deterministic model is formulated as fol-
lows:
yi j ≤ T xi j , ∀i = 1, ., V ; j = 1, . . . , B (9)


V
[DTBAP] min (ci1 u+i + ci2 u−i ) (1) yi j , u+ , u− ≥ 0, ∀i = 1, . . . , V, j = 1, . . . , B (10)
i i
i=1


B xi j , zii j , θi j ∈ {0, 1}, ∀i, i = 1, . . . , V, i = i ; j = 1, . . . , B (11)
s.t. y i j − T τi − a i = u +
i
− u−
i
, ∀i = 1 , . . . , V (2)
j=1

τi ∈ {−1, 0, 1}, ∀i, = 1, . . . , V (12)


B
The objective function (1) aims to minimise the total cost of
xi j = 1, ∀i = 1 , . . . , V (3)
the deviations. Constraint (2) is the linearisation of the deviations
j=1
between the planned and expected berthing times of the vessel.
Constraint (3) ensures that each vessel is assigned to one berth.

V
Constraint (4) indicates that the total operation tasks of vessels al-
ti j xi j ≤ T , ∀ j = 1, . . . , B (4) located to each berth must be completed within the planning hori-
i=1
zon. Constraint (5) ensures that the precedence among vessels in
each berth is satisfied. Constraint (6) ensures that a ring exists for
yi j + ti j − T θi j ≤ yi j + T (1 − zii j ), ∀i, i = 1, . . . , V, i = i ; j = 1, . . . , B (5) each berth, which is caused by considering periodicity. For exam-
ple in Fig. 3, there exists a ring among vessels 1, 2, and 3. Con-
straints (7)-(8) indicate that each vessel must have one vessel sub-

V sequently before and behind. Especially, this rules out that some
θi j = 1 , ∀ j = 1 , . . . , B (6) berth is only visited by a single vessel. However, it is impossible
i=1 to have only one vessel in one berth in reality, since we consider a

5
X. Xiang and C. Liu Omega 103 (2021) 102444


planning horizon for one week. Thus, these two constraints can be For the non-linear form maxs∈ Vi=1 bi v+ is
, we linearise it by
used to solve TBAP in this paper. Constraint (9) defines the vari- introducing additional decision variable, i.e., C max . The linearised
able upper bounds of yi j . Constraints (10)-(12) specify the domains model can be presented as follows:
of the decision variables.

N
[ROTBAP] min (ci1 u+i + ci2 u−i ) + C max (22)
3.3. SPTBAP
i=1

In this study, we use a time buffer to characterise the robust-

ness of the schedule. The time buffer of vessel i is denoted by ηi , s.t. Constraints (2 ) − (4 ), (6 ) − (12 ), (15 ) − (20 )
which is defined as the subsequent vessel’s start time minus the 
V

ending time of vessel i. If (ξi − ηi ) > 0, then a conflict will occur C max ≥ bi v+
is
, ∀s = 1 , . . . , S (23)
between vessel i and its subsequent vessel. To resolve this conflict, i=1
the ports must increase productivity, resulting in a penalty cost
(the cost rate is denoted by bi ). Therefore, the objective function 3.5. EROTBAP
of SPTBAP is given as follows:
  The RO model has two key features, namely, characterising un-

V 
V certainties with an uncertainty set and searching for solutions with
[SPTBAP] min ( ci1 u+
i
+ ci2 u−
i
)+E b i ( ξi − η i ) +
(13) the minimum recovery cost associated with the worst-case sce-
i=1 i=1 nario(s) in the uncertainty set(s). Capturing randomness with only
We then use the sample average approximation (SAA) scheme to re- one uncertainty set may not achieve the result expected. A small
formulate SPTBAP. Assume that there are S scenarios, and ξi is de- uncertainty set (if used) hardly captures randomness and maybe
noted by ξis , i = 1, . . . , V, s = 1, . . . , S. Based on the above analyses, unreliable towards uncertainties. However, a large uncertainty set,
an SP model formulation is proposed for the TBAP under uncer- e.g., with any possible realisation within it, tends to be overpro-
tainty: tective and costly. In this study, we classify uncertain scenarios
into multiple uncertainty sets, assign a weight to the performance

V
1  +
S V
of the worst-case scenario(s) in each uncertainty set(s), and cal-
[SAATBAP] min (ci1 u+i + ci2 u−i ) + bi vis (14)
S culate the weighted summation of the worst-case performance of
i=1 s=1 i=1
these uncertainty sets. Using multiple uncertainty sets can effec-
tively avoid over-conservative solutions.
s.t. Constraints (2 ) − (4 ), (6 ) − (12 ) Before formulating the model, we describe how to construct an
ξis − ηi = v − v−is , ∀i = 1, . . . , V, s = 1, . . . , S
+
is
(15) uncertainty set to achieve the first goal introduced in Section 1:
how to obtain an uncertainty set depending on the collected his-
torical data.
0 ≤ y i j + t i j + ηi − T θ i j ≤ T , ∀i = 1, ., V ; j = 1, . . . , B (16) We introduce the procedure to estimate the number of con-
tainers from historical data and characterise it by constructing
the uncertainty set via a data-driven approach, i.e, K-means clus-
yi j + ti j + ηi − T θi j ≤ yi j + M (1 − zii j ), ∀i, i = 1, . . . , V, tering. Since we study the problem with the vessels’ weekly ar-
i = i ; j = 1 , . . . , B (17) rival pattern, we can collect the number of containers inflated
for each week. We assume that we collect historical data con-
sisting of S sample paths. Each sample path s comprises realised

V 
B 
V
demands ξis . To implement K-means clustering, we partition the
ti j xi j + ηi = T B (18) 1 S
set of {ξ , . . . , ξ } into K non-overlapping, full-dimensional clusters.
i=1 j=1 i=1
The K-means clustering algorithm outputs the centroids, which we
denote as μk , μk = {μk1 , μk2 , . . . , μkV }, for k ∈ K = {1, . . . , K }. The
ηi ≥ 0, ∀i = 1, ., V (19) uncertainty sets are described formally as,
 
k = ξ ∈ [ξ , ξ ]|(ξ − μk ) (ξ − μk ) ≤ (ξ − μl ) (ξ − μl ), ∀l ∈ K ,
v+is , v−is ≥ 0, ∀i = 1, . . . , V, s = 1, . . . , S (20)
(24)
Objective (14) aims to minimise the baseline and penalty costs
which can be simplified into:
for resolving the conflict. Two non-negative variables v+ is
and v−
is  
are introduced to linearise the non-linear form (ξi − ηi )+ in Con- k = ξ ∈ [ξ , ξ ]|2ξ  (μl − μk ) + μk μk − μl μl ≤ 0, ∀l ∈ K , (25)
straint (15). Constraint (5) in the DTABP is substituted by Con-
straints (16) and (17). Constraint (18) ensures that the sum of all
where ξ and ξ are (respectively) the upper and lower bounds of ξ .
the operation and slack times equals T B.
We can now determine the weight, w = P[ξˆ ∈  ] as, k k
3.4. ROTBAP S
I ( ξ ∈ k )
s
s=1
wk = , (26)
S
The RO model assumes an uncertainty set to capture an uncer-
tainty factor, and the solutions derived by this method are robust where I(· ) is the indicator function.
to any disruptions within the uncertainty set. Here, we assume This uncertainty set allows us to model a rich variety of struc-
1 2 S tural information about ξ in resolutions that vary depending on
that an uncertainty set  = {ξ , ξ , . . . , ξ } is known. We can then
the number of clusters. Based on the constructed uncertainty sets,
obtain the ROTBAP model as follows:
the corresponding EROTBAP can be formulated as follows:

V 
V
[ROTBAP] min (ci1 u+i + ci2 u−i ) + max bi v+
is
(21) 
V 
K
s∈
i=1 i=1 [EROTBAP] min (ci1 u+i + ci2 u−i ) + wkCkmax (27)
s.t. Constraints (2 ) − (4 ), (6 ) − (12 ), (15 ) − (20 ) i=1 k=1

6
X. Xiang and C. Liu Omega 103 (2021) 102444

Fig. 4. Results of the example.

−LB
problem (MP). When Gap, Gap = UBUB is smaller than the optimal-
ity tolerance, the procedure is terminated. The C&CG algorithm is
s.t. Constraints (2 ) − (4 ), (6 ) − (12 ), (15 ) − (20 )
presented as follows:

V
Ckmax ≥ bi v+ , ∀ ξ s ∈ k , k = 1 , . . . , K (28) • Input: Uncertainty sets k and parameters of TBAP.
is
i=1 • Output: Robust tactical berth allocation plan.
• Step 1: Initialisation. Set LB = 0, UB = +∞, k = ∅, k =
Next, we display a small instance to demonstrate the benefits
1, . . . , K, and iteration number r = 1.
of EROTBAP against DTABP and ROTABP. • Step 2: Obtain an initial solution of the first stage. We
Example. Suppose there is only one berth and the time hori- 
firstly select the set with the largest Vi=1 μki among K un-
zon is 20 time units. Consider three vessels with expected arrival
certainty sets. Suppose that the selected set is denoted by
time 1, 7, 12, and expected operation time 6, 4, 5. For the sake of
k . We then solve the following problem and obtain an ini-
simplicity, we set the cost rate ci1 = ci2 = 1, bi = 1 for i = 1, 2, 3. As-
tial solution (xˆ, yˆ, zˆ, θˆ, τˆ , ηˆ , uˆ+ , uˆ− ).
sume that there is no fluctuation for vessels 1 and 3, i.e., ξ1 = ξ3 =
 
0. And the extra operation time for vessel 2 is stochastic, which 
V 
V
is ranging from 0 to 3. Assume we have sufficient historical data, min (ci1 u+i + ci2 u−i ) + max bi v+
is
(31)
s ∈ k 
which means S is very large. By using three models, i.e., DTBAP, i=1 i=1
ROTBAP, and EROTBAP, we can obtain the schedules as shown in s.t. Constraints (2 ) − (4 ), (6 ) − (12 ), (15 ) − (20 )
Fig. 4. In addition, we set K = 2 in EROTBAP. As shown in the fig-
ure, the only difference among schedules obtained by three mod- • Step 3: Identify the worst-case scenarios. Based on the so-
els is the berthing time of vessel 3, which results in a different lution of the first stage (xˆ, yˆ, zˆ, θˆ, τˆ , ηˆ , uˆ+ , uˆ− ), we can ob-
time buffer of vessel 2, i.e., η2 . Obviously, the schedule obtained tain the solution of second-stage. We identify a worst-case
r
by EROTBAP can not only conduct uncertainty to some extend but scenario ξ k for each uncertainty set which has the largest
also achieve a solution that is not too conservative. objective value as follows:


V

4. Solution approach ξ rk = arg s max bi (ξis − ηˆi )+ (32)

ξ ∈k \k i=1

We propose a C&CG algorithm to solve the EROTBAP, which can We use Zkr to denote the largest objective value for uncer-
achieve the second goal presented in Section 1: how to solve the tainty set k at iteration r. Then,
ERO problem effectively given uncertainty sets.
The EROTBAP model can be transformed into a two-stage 
V 
V
Zkr = ci (uˆ+
i
+ uˆ−
i
)+ max r bi (ξis − ηˆi )+
model:
i=1
ξ s ∈k ∪{ξ k } i=1

V 
K 
V
We can update the upper bound as follows:
min (ci1 u+i + ci2 u−i ) + wk max bi v+
is
(29) 
ξ ∈k i=1
s
i=1 k=1 
K

s.t. Constraints (2 ) − (4 ), (6 ) − (12 ), (16 ) − (20 ) UB = min UB, wk Zkr (33)

k=1
where v+
is
satisfies
• Step 4: Solve the MP. Based on the selected worst-case sce-
v+is − v−is = ξis − ηi , ∀i = 1, . . . , V, s = 1, 2, . . . , S (30) r
nario ξ k , we create new second-stage variables (v+ , v− ) and
Constraint (35), which are added to the subsequent MP.
where continuous variables v+is
and v−
is
denote the second-stage de-
cisions to be determined after ξ is realised. 
V 
K

The algorithm solves the problem iteratively, and each iteration min ci ( u +
i
+ u−
i
)+ wk κk (34)
r
r will identify a worst-case scenario ξ for each uncertainty set. i=1 k=1
r
Based on the selected scenario ξ , second-stage variables (v+ , v− )
and corresponding constraints are created and added to the master s.t. Constriants (2 ) − (4 ), (6 ) − (12 ), (15 ) − (20 )

7
X. Xiang and C. Liu Omega 103 (2021) 102444


V • The maximum number of iterations is that all the scenarios
bi v+
is
≤ κk , ∀ ξ s ∈ k , k = 1 , . . . , K (35) are added into the MP. In that case, the number of iterations
i=1 is maxk=1,.,K |k |. Actually, if all the worst-case scenarios in
We solve the above model and set the lower bound as the set k have been added into k , the algorithm converges to
optimal objective value of the MP. The output of MP is de- optimality.
noted by (u+ , u− , κk ).
Therefore, we obtain the desired result. 

V 
K
LB = (ci1 u+i + ci2 u−i ) + wk κk (36) Up to now, we have arrived at a holistic framework based on
i=1 k=1 ERO for solving TBAP, which integrates information from historical
data into final decisions. The ERO-framework is described as fol-
Update r = r + 1.
lows:
• Step 5: Iterate until the algorithm terminates. Check if
−LB
Gap = UBUB ≤  , then an  -optimal value is obtained; oth- • Step 1. Collect historical data about the number of contain-
erwise, return to Step 3. ers.
• Step 2. Construct the uncertainty set using K-means cluster-
For the proposed approach, it must generate an optimal solu-
ing as discussed in Section 3.5.
tion in a finite number of iterations. Theorem 1 is given as follows:
• Step 3. Formulate the EROTBAP model and linearise it.
Theorem 1. The C&CG approach converges in a finite number of it- • Step 4. Solve the resulting model using the C&CG algorithm
erations, i.e., maxk=1,.,K |k |. introduced in Section 4 to generate robust optimal sched-
∗ ules.
Proof. Suppose that Z ∗ and (x∗ , y∗ , z∗ , θ , τ ∗ , η∗ , u+∗ , u−∗ ) repre-
sent the optimal objective value and optimal solution of EROTBAP. 5. Computational experiments
We prove this theorem in four steps:
• First, we can observe that no matter what the value of ηˆ Extensive numerical experiments are conducted to validate the
is obtained in the first-stage, the second-stage always has performance of the proposed model and algorithm. This section
feasible solutions. Therefore, the solutions obtained in Steps includes four parts. Section 5.1 describes how to generate the in-
2 and 4 are both feasible for the original problem. stances. Section 5.2 verifies the effectiveness of the proposed C&CG
• We then prove that Step 3 must obtain an upper bound. As algorithm by comparing it with a commercial solver, i.e., CPLEX,
mentioned above, the second-stage is always feasible with a some practical policies, i.e., first come first served (FCFS), rolling
given solution (xˆ, yˆ, zˆ, θˆ, τˆ , ηˆ , uˆ+ , uˆ− ). Thus, we have, horizon heuristic (RHH), and Benders Decomposition. Section 5.3 in-
vestigates the performance of EROTBAP compared with other

V 
K 
V
mathematical models, i.e., SPTBAP and ROTBAP. Section 5.4 investi-
Z∗ ≤ (ci1 uˆ+i + ci2 uˆ−i ) + wk max bi (ξis − ηˆi )+ (37)
s
ξ ∈k i=1 gates how the number of clusters and uncertainty degree influence
i=1 k=1
the system performance.

V 
K 
V
= (ci1 uˆ+i + ci2 uˆ−i ) + wk max r bi (ξis − ηˆi )+ 5.1. Instance generation
i=1 k=1
ξ ∈k ∪{ξ k } i=1
s

(38) Because of lacking of real data, we randomly generate instances

for the experiments with the rules of instance generation in [52],

K
in which the generation method is convincible. The planning hori-
= wk Zkr
zon is set to 168, and one unit time is 1 h, i.e. T = 168. Six groups
k=1
of instances are considered, as shown in Table 1.
Formula (37) uses the fact that the objective value of feasi-
The operation time of vessel i at berth j, i.e. ti j , is generated as
ble solution is always larger than that of the optimal one
follows: Firstly, similar to [30,37,49,52], the average operation time
for min problem. Formula (38) employs the definition in is generated by a uniform distribution. Here, a mean value ti is
Equation (32).
generated in accordance with ti ∼ U (18, 36 ). Secondly, ti j is gener-
• Subsequently, we prove that Step 4 must obtain an B
ti j
lower bound. Since k ⊂ k , the optimal solution ated, satisfying j=1B = ti . In the previous literature, the stochas-
∗
(x∗ , y∗ , z∗ , θ , τ ∗ , η∗ , u+∗ , u−∗ ) of EROTBAP must be a feasible tic operation time is often generated by a uniform distribution
solution of MP. Then we have, [10,21,50,52]. Therefore, we assume that the random parameter ξis
follows uniform distributions [ζil , ζiu ] and are generated indepen-

V 
K 
V
Z∗ = (ci1 ui+∗ + ci2 u−∗
i
)+ wk max bi (ξis − ηi∗ )+ dently of each other. ζil and ζiu are randomly generated with mean
i=1
s
ξ ∈k i=1 values 5 and 19, respectively. Table 1 reports the average berth
k=1


V 
K 
V utilisation of the six groups of instances. ai follows a uniform dis-
≤ (ci1 ui+∗ + ci2 u−∗
i
)+ wk max bi (ξis − ηi∗ )+ (39) tribution of U[1, 168]. For the cost rates ci1 , ci2 and bi , we refer to
i=1 k=1
s
ξ ∈k i=1 [32] and generate them randomly with mean values, 20 0 0, 20 0 0,
and 60 0 0 USD/hour, respectively. For the EROTBAP, 500 historical

V 
K 
V samples are collected to compute the parameters of the uncer-
= (ci1 ui+∗ + ci2 u−∗
i
)+ wk max bi vis+∗ (40) tainty set, and K is set to 5 unless explicitly mentioned. Besides,
i=1 k=1
s
ξ ∈k i=1
algorithms are coded with Java and run on a PC with the following
configuration: 64 bit, 2.80 GHz, Intel(R) Core(TM) i7-7600U CPU.

V 
K
= (ci1 ui+∗ + ci2 u−∗
i
)+ wk κk (41) 5.2. Algorithm performance
i=1 k=1

Formula (39) holds since k ⊂ k . Formula (40) is just This section verifies the effectiveness of the proposed algorithm
an equivalent transformation. Because of the strong duality by comparing it with CPLEX, some practical policies, and Benders
property, Formula (41) holds. Decomposition.

8
X. Xiang and C. Liu Omega 103 (2021) 102444

Table 1
Berth utilisation rates.

Para. Berth utilisation

∗
Group V B T Average operation time V/B Average berth utilisation (%)

G1 10 3 168 39 3.33 77.38

G2 20 5 168 39 4.00 92.86
G3 30 8 168 39 3.75 87.05
G4 40 10 168 39 4.00 92.86
G5 50 13 168 39 3.85 89.29
G6 60 16 168 39 3.75 87.05
Average: 87.75

Note: ∗ : The average operation time is equal to 18+36

2
+ 5+19
2
= 39.

Table 2
Computational time of CPLEX and C&CG for solving EROTBAP model.

Group CPU time (s)

CPLEX C&CG Ratio

Average Largest Smallest Average Largest Smallest (%)

G1 1188 3855 299 170 447 70 14.3

G2 10577 20979 3777 2459 3806 1511 23.2
G3 19048 22251 15312 4979 6191 3437 26.1
G4 31051 32961 28868 7851 9660 5519 25.3
G5 41372 52572 35265 11578 15427 7883 28.0
G6 64231 68311 54939 16633 22339 11520 25.9

Table 3
Comparisons between FCFS, RHH, and C&CG (Instances in G1-G3).

Para. FCFS RHH EROTBAP Comparison

S EC1 Time(s) EC2 Time(s) EC3 Time(s) gap1 (%) gap2 (%)

Results of G1
2 107067 1 86756 3 85285 4 25.54 1.72
4 107673 1 88549 3 85829 5 25.45 3.17
8 114423 1 92984 3 90472 5 26.47 2.78
16 117139 1 94764 5 92010 7 27.31 2.99
32 121188 2 95169 8 92558 13 30.93 2.82
64 121787 3 95082 19 92938 29 31.04 2.31
128 123050 7 94816 44 93337 74 31.83 1.58
256 122911 11 96006 129 93415 216 31.58 2.77
512 124560 18 95813 221 93418 354 33.34 2.56
Results of G2
2 797555 1 643134 5 628580 6 26.88 2.32
4 800521 1 647121 8 632306 8 26.60 2.34
8 847538 1 673029 18 657350 11 28.93 2.39
16 877068 1 691476 37 675178 19 29.90 2.41
32 901087 2 699729 67 683094 67 31.91 2.44
64 906396 3 702940 90 686339 98 32.06 2.42
128 908784 7 706591 199 688915 671 31.92 2.57
256 912426 14 711182 681 689152 2078 32.40 3.20
512 927819 39 711629 1398 691067 3213 34.26 2.98
Results of G3
2 1747585 1 1398620 6 1358902 11 28.60 2.92
4 1746480 1 1387606 6 1366345 12 27.82 1.56
8 1845968 2 1428576 10 1400043 17 31.85 2.04
16 1929604 2 1478001 24 1450777 49 33.00 1.88
32 1967058 5 1507131 75 1476486 127 33.23 2.08
64 1984538 9 1536412 146 1482637 302 33.85 3.63
128 1960541 17 1525250 395 1488209 745 31.74 2.49
256 1985957 29 1545148 1416 1490496 3054 33.24 3.67
512 2029144 40 1549878 2614 1498700 4937 35.39 3.41
EC1 −EC3 EC2 −EC3
Note. gap1 = EC3
× 100%, gap2 = EC3
× 100%.

5.2.1. Comparison with CPLEX Table 2 shows that the computational time of the C&CG algo-
This section evaluates the performance of the proposed algo- rithm is much shorter than that of CPLEX for all groups of in-
rithm compared with CPLEX. It is noted that both CPLEX and our stances. The computational times consumed by C&CG are 14.3%,
algorithm will terminate when  = 10−4 . Table 2 reports the com- 23.2%, 26.1%, 25.3%, 28.0% and 25.9% of those consumed by CPLEX
putational results, i.e., the average, largest, and smallest CPU times for six groups, respectively. CPLEX may spend over 19 hours to
of 10 instances in each instance group. The detail results are re- solve the large-scale instances in instance group G6. By contrast,
ported in Table A.9 in appendix. C&CG can solve most instances in G6 to optimality in less than 6

9
X. Xiang and C. Liu Omega 103 (2021) 102444

Table 4
Comparisons between FCFS, RHH, and C&CG (Instances in G4-G6).

Para. FCFS RHH EROTBAP Comparison

S EC1 Time(s) EC2 Time(s) EC3 Time(s) gap1 (%) gap2 (%)

Results of G4
2 2635086 2 2080871 111 2033651 182 29.57 2.32
4 2648426 2 2100065 126 2045273 192 29.49 2.68
8 2781951 3 2105002 129 2065491 205 34.69 1.91
16 2934041 3 2191085 195 2162144 246 35.70 1.34
32 2942493 7 2251995 210 2214228 543 32.89 1.71
64 2973194 13 2293123 278 2221827 1065 33.82 3.21
128 2988804 26 2316954 621 2227722 1918 34.16 4.01
256 3003563 42 2323884 1867 2232220 4398 34.55 4.11
512 3072236 61 2337902 3672 2251920 7677 36.43 3.82
Results of G5
2 3631560 2 2831798 311 2787251 622 30.29 1.60
4 3670483 2 2833627 316 2787936 700 31.66 1.64
8 3827976 4 2849079 403 2788676 736 37.27 2.17
16 4013173 4 2967803 443 2944325 1401 36.30 0.80
32 4090316 10 3069587 500 3029243 2437 35.03 1.33
64 4093117 18 3153402 649 3042309 3417 34.54 3.65
128 4109788 37 3191671 962 3050530 4231 34.72 4.63
256 4150371 57 3195254 2723 3056528 5849 35.79 4.54
512 4246710 84 3220100 5545 3090819 10989 37.40 4.18
Results of G6
2 5104100 2 3876147 372 3840326 1073 32.91 0.93
4 5179558 2 3881364 613 3854271 1109 34.38 0.70
8 5362609 5 4072111 652 3913445 1236 37.03 4.05
16 5581728 5 4098259 769 4087869 2192 36.54 0.25
32 5734186 12 4268958 816 4228830 3522 35.60 0.95
64 5752680 22 4417358 947 4244930 4751 35.52 4.06
128 5800626 48 4478883 1587 4255437 6151 36.31 5.25
256 5841654 75 4480538 4231 4269983 8131 36.81 4.93
512 5979737 116 4525401 8162 4329168 16131 38.13 4.53
EC1 −EC3 EC2 −EC3
Note. gap1 = EC3
× 100%, gap2 = EC3
× 100%.

hours for EROTBAP. Therefore, the proposed C&CG algorithm is ac- practice. In conclusion, from the perspective of practice, EROTBAP
ceptable in terms of smaller computational time. can achieve a significant improvement to the expected cost with
an acceptable computational time compared with FCFS and RHH
5.2.2. Comparison with other practical policies approaches.
This section quantifies the performance of different solution ap-
proaches via simulation and evaluates the average expected cost 5.2.3. Comparison with benders decomposition
and computational time. FCFS and RHH are applied as benchmarks This section compares our method with Benders Decomposi-
to solve this problem. tion. In this experiment, each group contains 10 instances, and the
For the detail of RHH, we refer to Agra and Oliveira [1], and procedures of Benders Decomposition and our algorithm will be
the algorithm is displayed in Appendix B. In the RHH, we use nv terminated within a CPU time limit of 180 0 0s. Table 5 reports the
to denote the number of new vessels to consider in each iteration. average values of 10 instances. In Table 5, the column “Itr” denotes
nt and ut represent the numbers of time periods to consider for the number of iterations, the column “gap” reports the gap be-
each sub-problem after the arrival time of the last vessel, and be- tween upper and lower bounds, and the column “UB” reports the
fore the arrival time of the first new vessel, respectively. nv = 5 best feasible solutions obtained when the procedure is terminated.
and nt = ut = 10. In this experiment, we use instance groups G1- As shown in Table 5, as the instance size increases, the objec-
G6, and S is set to 2,4,8,16,32,64,128,256,512. Tables 3 and 4 report tive values and the CPU time all tend to increase for both two
the results, where EC and ’time’ denote the average expected cost methods. Obviously, the objective values, CPU time, and optimality
and computational times, respectively. Fig. 5 shows the results for gaps provided by C&CG dominate those solved with Benders De-
different methods. It is noted that Fig. 5 only displays the results composition for all six groups. Moreover, the number of iterations
for G1 and G3 since the results for the six groups are similar. needed by the C&CG method is much lower than that of Benders
Fig. 5 shows that EROTBAP performs better in the indicator Decomposition. For example, in the instance group G3, C&CG can
of the expected cost compared with FCFS and RHH. On average, obtain the optimal solution with an average CPU time of 4979s,
EROTBAP can save 29.28%, 30.54%, 32.08%, 33.48%, 34.28%, and and the corresponding gap is much better than 2.78% obtained by
35.91% expected cost compared with FCFS for six groups. Although Benders Decomposition with a large average CPU time of 180 0 0s.
RHH performs better than FCFS, it still invests 2.52%, 2.56%, 2.63%, This phenomenon can be explained as follows. ERO model focuses
2.79%, 2.73%, and 2.85% more cost than EROTBAP for six groups. on the worst-case scenarios and our C&CG algorithm selects the
Consistent with the conclusion in Section 5.3, we can find that the worst-case scenarios in each iteration. For most instances, our al-
value of the expected cost obtained by the ERO model tends to be gorithm will terminate after very few iterations as shown in the
stable when S = 32. The computational times of all the three meth- table. Besides, at the beginning of our algorithm, we just need to
ods generally increase as S increases. We observe that the EROTBAP solve a problem with very few scenarios in each uncertainty set.
has the largest computational times among the three methods be- Thus, it is easy to solve the master problem and subproblems in
cause it is solved by an exact algorithm. However, it remains com- each iteration. For Benders Decomposition, it needs to consider all
putationally tractable for all tested instances, which is applicable in the scenarios in each iteration and the gap of Benders Decompo-

10
X. Xiang and C. Liu Omega 103 (2021) 102444

(a) Expected cost for G1 (b) CPU times for G1

FCFS FCFS

CPU time (second)

300
RHH RHH
Expected cost

ERO ERO
110000

200
100
90000

0
2 4 8 16 32 64 128 256 512 2 4 8 16 32 64 128 256 512

Number of scenarios (S) Number of scenarios (S)

(c) Expected cost for G3 (d) CPU times for G3

FCFS FCFS

CPU time (second)

4000
RHH RHH
1800000
Expected cost

ERO ERO

2000
1400000

2 4 8 16 32 64 128 256 512 0 2 4 8 16 32 64 128 256 512

Number of scenarios (S) Number of scenarios (S)

Fig. 5. Expected cost and computational times with varying S, S = 2, 4, 8, 16, 32, 64, 128, 256, 512 for instance groups G1 and G3.
Table 5
Comparisons between Benders Decomposition and C&CG.

Para. Benders Decomposition C&CG

Itr CPU Time (s) UB gap(%) Itr CPU Time (s) UB gap(%)

G1 212.2 9013 92869 0.00 3.9 170 92869 0.00

G2 308.2 13019 662125 0.00 5.2 2459 662125 0.00
G3 419.7 18000 1486778 2.78 7.1 4979 1451047 0.00
G4 389.6 18000 2312683 8.49 8.6 7851 2160347 0.00
G5 309.5 18000 3427316 15.21 9.9 11578 3020436 0.00
G6 218.4 18000 4990835 21.67 11.2 15982 4201673 0.12

sition decreases at a slow rate. Therefore, the C&CG method per- lows:
forms better than Benders Decomposition in both CPU time and Sˆ V
m 
V
bi (ξis − ηˆi )+
solution quality. ZS = ci (uˆ+
i
+ uˆ−
i
)+ s=1 i=1

i=1
Sˆ
• Step 4: Report the following three performance metrics:
5.3. Model performance – Tractability: We use the computational time averaged
over 50 training sets to characterise the tractability of
In this section, we will verify the effectiveness of the proposed each method.
model by comparing it with other mathematical models and prac- – Optimality gap: The relative difference between the ex-
m
tical policies. 50 training datasets are generated with varying sizes. pected cost Z S and the optimal expected cost Z ∗ , which
For each model and policy, we perform the following procedure in is computed by SAA approach with S = 10 0 0 data points.
each case. This indicator is employed to evaluate the performance
of out-of-sample, which is computed as follows:
• Step 1: For training datasets m, solve the models based on m
1  ZS − Z∗
50
the collected S random samples and obtain the first-stage ZS =
50 Z∗
decisions, (xˆ, yˆ, zˆ, θˆ, τˆ , ηˆ , uˆ+ , uˆ− ), and corresponding objec- m=1
tive value which is denoted by ZˆSm . – Prediction error: The relative gap between method’s op-
timal objective value (ZˆSm ) and its corresponding expected
S V

V
bi (ξis − ηˆi )+ m
cost (Z S ):
ZˆSm = ci (uˆ+
i
+ uˆ−
i
)+ s=1 i=1
S m
1  ZˆSm − Z S
50
i=1
ZˆS = m
50 ZS
• Step 2: Collect another Sˆ = 10 0 0 random samples. m=1
• Step 3: Based on the first-stage decisions determined, sim- This section compares three models, i.e., SAATBAP, ROTBAP, and
ulate each sample and calculate the expected cost as fol- EROTBAP, introduced in Section 3. In this experiment, we use in-

11
X. Xiang and C. Liu Omega 103 (2021) 102444

Table 6
Computational results of different models.

Para CPU times Optimality gap (%) Prediction error (%)

S SAA RO ERO SAA RO ERO SAA RO ERO

Results of G1
2 6 3 4 -0.220 2.147 0.520 0.354 0.426 0.181
4 7 4 5 -0.273 5.645 1.161 1.818 -0.096 -0.147
8 8 4 5 -0.098 28.548 6.633 -1.552 -0.148 -0.132
16 9 6 7 -0.232 34.960 8.447 -0.174 -0.192 -0.183
32 16 11 13 0.230 40.903 9.092 -0.780 -0.121 -0.090
64 32 25 29 -0.187 43.875 9.540 0.034 0.000 0.000
128 70 60 74 -0.206 43.875 10.010 0.296 0.000 0.000
256 240 202 216 -0.009 43.875 10.102 -0.130 0.000 0.000
512 371 322 354 0.016 43.875 10.105 0.027 0.000 0.000
Results of G3
2 14 6 11 -0.209 2.191 0.569 -0.368 0.211 0.104
4 16 9 12 -0.300 16.071 1.120 -0.860 -0.176 -0.146
8 23 11 17 -0.100 28.039 3.614 1.517 -0.129 -0.156
16 72 25 49 -0.211 38.737 7.369 -0.176 -0.180 -0.075
32 186 74 127 0.212 40.684 9.272 -0.841 -0.131 -0.082
64 453 229 302 -0.192 44.468 9.727 0.036 0.000 0.000
128 1077 597 745 -0.212 46.472 10.139 0.296 0.000 0.000
256 3892 2153 3054 -0.017 47.120 10.308 -0.136 0.000 0.000
512 6093 3290 4937 0.024 47.485 10.916 0.028 0.000 0.000

Table 7
Computational results of EROTBAP with different sizes of uncertainty set.

Group K=1 K=3 K=5 K=7 K=9

Obj Time(s) Obj Time(s) Obj Time(s) Obj Time(s) Obj Time(s)

G1 95867 186 94420 197 92869 170 90889 245 86196 326
G2 700706 2603 694014 2651 681953 2459 662125 3567 624397 4413
G3 1519417 4982 1516559 5401 1481345 4979 1451047 7994 1364564 9340
G4 2293812 8091 2292851 8662 2251196 7851 2160347 11913 2037985 14842
G5 3131431 13151 3122512 13337 3090252 11578 3020436 17626 2849785 19863
G6 4415449 19051 4368636 17507 4329592 16633 4196596 25077 3951880 28342

(a) CPU times for G1 (b) Optimality gap for G1 (c) Prediction error for G1
400

SAA SAA SAA

RO RO RO
40

ERO ERO ERO

300

1
Prediction error (%)
CPU time (second)

Optimality gap (%)

30
200

0
20
100

−1
10
0

−2
0

2 4 8 16 32 64 128 256 512 2 4 8 16 32 64 128 256 512 2 4 8 16 32 64 128 256 512

Number of scenarios (S) Number of scenarios (S) Number of scenarios (S)

(d) CPU times for G3 (e) Optimality gap for G3 (f) Prediction error for G3
50

SAA SAA SAA

RO RO RO
ERO ERO ERO
5000

40

1
Prediction error (%)
CPU time (second)

Optimality gap (%)

30
3000

0
20

−1
10
1000

−2
0

2 4 8 16 32 64 128 256 512 2 4 8 16 32 64 128 256 512 2 4 8 16 32 64 128 256 512

Number of scenarios (S) Number of scenarios (S) Number of scenarios (S)

Fig. 6. Performance analysis of different models.

stance groups G1 and G3, and S is set to 2,4,8,16,32,64,128,256,512. by SAA. The corresponding results are shown in Table 6 and Fig. 6,
The reasons why we just use instances in G1 and G3 are explained where computational time, optimality gap, and prediction error are
as follows. First, the results for the six groups are similar, which reported. SAATBAP model is solved using Benders decomposition,
can be observed from the results of G1 and G3. It is unneces- and ROTBAP is solved using C&CG.
sary to report similar results. Second, for large-scale instances in Tractability. Table 6 and Fig. 6(a,d) display the computational
G5 and G6, it is difficult or even can not be solved to optimality times over 50 training datasets. It is obvious that the computa-

12
X. Xiang and C. Liu Omega 103 (2021) 102444

(a) Objective Value (G1) (b) CPU time (G1) (c) Objective Value (G2) (d) CPU time (G2)

700000
300
92000

Time (s)

Time (s)

3500
Cost

Cost

640000
200
86000

2500
1 3 5 7 9 1 3 5 7 9 1 3 5 7 9 1 3 5 7 9

K K K K

(e) Objective Value (G3) (f) CPU time (G3) (g) Objective Value (G4) (h) CPU time (G4)

2250000
5000 7000 9000

12000
Time (s)

Time (s)
Cost

Cost
1400000

2050000

8000
1 3 5 7 9 1 3 5 7 9 1 3 5 7 9 1 3 5 7 9

K K K K

(i) Objective Value (G5) (j) CPU time (G5) (k) Objective Value (G6) (l) CPU time (G6)

4400000

26000
2850000 3050000

18000
Time (s)

Time (s)
Cost

Cost

4000000

18000
12000

1 3 5 7 9 1 3 5 7 9 1 3 5 7 9 1 3 5 7 9

K K K K

Fig. 7. Optimisation performances of EROTBAP with different K.

Table 8
Computational results of EROTBAP under different levels of uncertainties (G1).

α ROTBAP SPTBAP EROTBAP

EC C1 C2 T Cw EC C1 C2 T Cw EC C1 C2 T Cw

0.25 19300 19300 0 19300 17917 16000 1917 27500 18397 17000 1397 23000
0.50 32000 32000 0 32000 30007 25800 4207 49300 30520 29133 1387 34633
0.75 44956 43998 958 44956 39357 35300 4057 60067 39677 38300 1377 45800
1.00 120798 55255 65543 120798 88224 36767 51458 163533 93415 39731 53684 143495
1.25 275959 134308 141650 275959 164851 38121 126730 306600 207009 101433 105575 281633

tional times of three models generally scale linearly in S. We can Prediction error. Fig. 6(c,f) displays the prediction error of the
observe that the computational time consumed by EROTBAP is various methods. Both RO and ERO can consistently provide mean-
slightly longer than that of ROTBAP and slightly smaller than that ingful upper bounds on the expected cost. Furthermore, ERO per-
of SAATBAP, which is acceptable in practice because it is a deci- forms better than RO. In contrast, SAA may underestimate or over-
sion at the tactical level, thus not requiring a quick decision. For estimate the true performance, depending on the number of data
example, when the size of data is set as S =256, the ERO model points.
with G3 can also be solved with an average computational time We can conclude that EROTBAP can not only guarantee the
less than 50 0 0 seconds, which is applicable in practice. out-of-sample performance which overcomes the vulnerability of
Optimality gap. Fig. 6(b,e) displays the optimality gaps for the SAA but also avoids the over-conservatism of traditional RO model,
various methods. As the figure is shown, the optimality gap of ERO with an acceptable computational time in practice. Therefore, if
increases with more data obtained. The gap is about 10% when planners focus on the out-of-sample performance and avoid over-
S = 512. Indeed, when S = 32, the optimality gap of ERO is already conservative solutions, then the ERO model is recommended.
very stable, which means it is enough to collect 32 data points to
guarantee the good performance of out-of-sample for ERO. These 5.4. Sensitivity analysis
results can provide useful guidelines for planners that using lim-
ited data can also achieve a good tactical plan. In practice, there is There are several parameters influencing the decisions, i.e., the
no need to collect a lot of data to generate a tactical plan, which number of clusters (K) and the level and uncertainty. This section
helps planners not only reduce the burden of data collection but will investigate how these parameters influence the decisions.
also generate a robust schedule quickly with a small dataset. For
the RO model, as the size of data increases, the optimality gap
5.4.1. Influence of the number of clusters
tends to perform worse and worse. As shown in the figure, when
This section investigates the influence of the number of clusters
N ≥ 16, the optimality gap of RO is even larger than 35%, which
(K) on optimisation performance. When the number of clusters is
means it becomes more conservative when more data is obtained.
smaller, more conservative solutions are obtained. We vary K from
For the SAA model, the obtained mean value in each case is near
1 to 9 with a step of 2. Table 6 and Fig. 7 report the optimisation
the corresponding optimal expected cost. However, it fluctuations
results of computational time and objective values.
up and down near zero. In other words, SAA can not guarantee the
Fig. 7 shows that as K increases, a smaller objective value is
performance of out-of-sample performance.
obtained, and the schedules tend to become less conservative ex-

13
X. Xiang and C. Liu Omega 103 (2021) 102444

Fig. 8. Computational results of EROTBAP under different levels of uncertainties (G1).

pectedly. We find some interesting observations for computational cost (see Fig. 8d), which is equal to the first-stage cost plus the
time. The computational time appears to decrease when K varies second-stage cost of the worst-scenario, EROTBAP performs better
from 1 to 5, although the model becomes more complex. This re- than SPTBAP, indicating that EROTBAP can protect against worst-
sult may be caused by the setting of Step 2 in the algorithm. More- case realisation. Therefore, EROTBAP is suggested to the planners,
over, the computational time decreases or increases slightly when which aim to control the worst-case performance. Due to the un-
K increases from 1 to 5. However, when K increases from 5 to 9, certain factors in the future, the planners are suggested to make a
the computational time tends to increase rapidly. Results can pro- robust plan which can well resist the possible risks in the future.
vide useful guidelines for planners to achieve a desirable trade-off Our ERO model can achieve this goal well at the expense of a slight
between the computational time and the conservatism of a so- increase in the cost. Furthermore, our model can also prevent ex-
lution. For example, if the planners tend to reduce the objective treme bad situations. In addition, if the planners face an environ-
value as much as possible, they are suggested to select a large K ment with a high degree, our ERO model is more suggested.
with a huge expense of computational time. While if the planners
tend to reduce the computational time as much as possible while 6. Conclusions
ensuring the solution quality, they are recommended to set K = 5.
This study investigates TBAP to assign the berthing positions
5.4.2. Impact of uncertainty degree and service times for vessels that periodically arrive at terminals.
For the BAP, considering uncertainties can directly influence The operation time of a vessel depends on the number of contain-
the first- and second-stage costs. The following experiments are ers that will be handled. A vessel visits a port periodically, and
conducted to evaluate how system costs, i.e. first-stage cost (C1 ), its total number of containers fluctuates in each period. This de-
second-stage cost (C2 ), expected cost (EC), and worse-case total mand fluctuation causes many uncertainties. Thus, we study TBAP
system cost (T Cw ), are influenced by the uncertainty degree. given an uncertain operation time. Due to the uncertain factors
Five cases in instance group G1 with different levels of in the future, the planners are suggested to make a robust plan
uncertainty, α = 0.15, 0.5, 0.75, 1, 1.25, ξ = αξ , are considered. which can well resist the possible risks in the future. In order to
Table 8 reports C1 , C2 , EC, and T Cw for each case, showing that four achieve this goal, we propose a holistic ERO-framework for solv-
indicators all tend to increase as α increases for all three models. ing TBAP, which provides useful guidelines for planners that using
To analyse the results further, Fig. 8 is provided. limited data can also achieve a good tactical plan. In practice, there
Fig. 8(a) shows that ROTBAP has the worst performance, and is no necessary to collect a lot of data to generate a tactical plan,
SPTBAP has the best performance on the indicator of expected cost which helps planners not only reduce the burden of data collection
amongst the three models. For the first-stage cost (see Fig. 8b), but also generate a robust schedule quickly with a small dataset.
SPTBAP tends to have a small value, whereas EROTBAP has a signif- From the perspective of practice, the ERO-framework can achieve
icant growth trend when α increases from 1 to 1.25 because EROT- a significant improvement to the expected cost with an acceptable
BAP tends to invest more cost at the first stage to improve the ro- computational time compared with FCFS and RHH approaches. Fur-
bustness of the schedule. Therefore, EROTBAP has a better perfor- thermore, our model can also prevent extreme bad situations. In
mance on the indicator of second-stage cost (see Fig. 8c), indicat- addition, our ERO model is more suggested to solve the problem
ing that EROTBAP can improve the schedule’s robustness at the ex- in the environment with a high uncertain degree. Through exper-
pense of a slight increase in the first-stage cost. For the worst-case iments, we also provide useful guidelines for planners to achieve

14
X. Xiang and C. Liu Omega 103 (2021) 102444

a desirable trade-off between the computational time and the con- Appendix B. Procedure of Rolling horizon optimisation
servatism of a solution by setting an appropriate value of K. For ex- algorithm
ample, if the planners tend to reduce the objective value as much
as possible, they are suggested to select a large K with a huge ex- The procedure of RHH is given in Algorithm 1 .
pense of computational time. While if the planners tend to reduce
the computational time as much as possible while ensuring the so-
Algorithm 1: Rolling horizon optimisation algorithm.
lution quality, they are recommended to set K = 5.
This study only considers the TBAP under uncertain operation 1 Initialisation. Set nv = 5, nt = ut = 10
time. In the future, we will study the TBAP and QC assignment 2 itrmax =  nVv  ← the total number of iterations
problems given uncertainty. Furthermore, we will consider other 3 Solve the first subproblem with nv vessels within the time
types of uncertainties, i.e. breakdown of equipment, calling of un- horizon [0, ans + nt]
scheduled vessels, and other unforeseen events. There are many 4 for it r = 2; it r ≤ it rmax ; it r + + do
clustering methods to partition the uncertainty set into several 5 Fix xi , yi j ∀i, j for the vessels scheduled before
clusters. In future studies, other methods will be explored. ans×(itr−1 ) − ut
6 Solve the itrth subproblem with vessels from
CRediT authorship contribution statement (itr − 1 ) ∗ ns + 1 to itr ∗ ns within the time horizon
[ans×(itr−1 ) − ut, anv×itr + nt].
Xi Xiang: Writing - review & editing, Writing - original draft, 7 end
Software, Investigation. Changchun Liu: Software, Conceptualiza-
tion.

Acknowledgment

This research has been made possible by the funding support

from the Singapore Maritime Institute.

Appendix A. The detail results of experiments

Table A.9 displays the detail results of Section 5.2.

Table A9
The detail results of CPLEX and C&CG for solving EROTBAP model.

Results of G1 Results of G2 Results of G3

No. Obj T1 T2 No. Obj T1 T2 No. Obj T1 T2

1 110147 187 1194 1 730432 2160 9311 1 1469791 5050 20990

2 66062 70 299 2 515484 1511 3777 2 1456468 6191 22251
3 64140 72 512 3 787238 3232 13658 3 1459249 3437 15312
4 154765 447 3855 4 696225 1652 7661 4 1504243 4010 16425
5 78937 76 515 5 713191 3806 20979 5 1554179 6006 21337
6 96637 184 1405 6 672731 2910 11146 6 1465688 3556 15861
7 117834 211 1423 7 589192 3015 11844 7 1478423 4892 18876
8 57527 177 1373 8 710529 2271 7827 8 1464259 5438 18647
9 93488 145 980 9 646782 2098 10965 9 1502998 5249 18689
10 89152 133 322 10 757724 1931 8599 10 1458156 5964 22089
Avg. 92869 170 1188 Avg. 681953 2459 10577 Avg. 1481345 4979 19048
Results of G4 Results of G5 Results of G6

No. Obj T1 T2 No. Obj T1 T2 No. Obj T1 T2

1 2185201 7544 29123 1 2979383 10580 38558 1 4378667 14926 59250
2 2341405 9660 32961 2 3319710 15427 52572 2 4635102 22339 62896
3 2215986 7842 31943 3 2961627 11008 41228 3 4077225 16993 62347
4 2208230 5519 28868 4 3057205 7883 35265 4 4209868 11606 64939
5 2313646 8882 31948 5 3109772 12586 42858 5 4480512 17544 67052
6 2156347 7902 31721 6 2982093 9690 38789 6 4218427 19777 59264
7 2229876 9253 31909 7 3147819 14168 37897 7 4032961 14741 64052
8 2266668 6984 29552 8 3030400 11359 36923 8 4492993 15639 68213
9 2309518 9289 31301 9 3039171 9954 50261 9 4362880 11520 65985
10 2285086 5639 31182 10 3275343 13125 39367 10 4407285 21249 68311
Avg. 2251196 7851 31051 Avg. 3090252 11578 41372 Avg. 4329592 16633 64231

Note. T1 and T2 denote the CPU time of C&CG and CPLEX, respectively.

15
X. Xiang and C. Liu Omega 103 (2021) 102444

References [28] Liu C, Xiang X, Zheng L. A two-stage robust optimization approach for the
berth allocation problem under uncertainty. Flexible Services and Manufactur-
[1] Agra A, Oliveira M. Mip approaches for the integrated berth allocation ing Journal 2019.
and quay crane assignment and scheduling problem. Eur J Oper Res [29] Liu C, Zhang C, Zheng L. A bi-objective model for robust yard allocation
2018;264(1):138–48. scheduling for outbound containers. Eng Optim 2017;49(1):113–35.
[2] Alfandari L, Davidoviç T, Furini F, Ljubiç I, Maraš V, Martin S. Tighter mip mod- [30] Liu C, Zheng L, Zhang C. Behavior perception-based disruption models for
els for barge container ship routing. Omega (Westport) 2019;82:38–54. berth allocation and quay crane assignment problems. Computers and Indus-
[3] Bertsimas D, Gupta V, Kallus N. Data-driven robust optimization. Math Pro- trial Engineering 2016;97:258–75.
gram 2018;167(2):235–92. [31] Mauri GR, Ribeiro GM, Lorena LAN, Laporte G. An adaptive large neighborhood
[4] Bertsimas D, Sim M. The price of robustness. Oper Res 2004;52(1):35–53. search for the discrete and continuous berth allocation problem. Computers
[5] Bierwirth C, Meisel F. A survey of berth allocation and quay crane scheduling and Operations Research 2016;70(C):140–54.
problems in container terminals. Eur J Oper Res 2010;202(3):615–27. [32] Meisel F, Bierwirth C. Heuristics for the integration of crane productivity in the
[6] Bierwirth C, Meisel F. A follow-up survey of berth allocation and berth allocation problem. Transportation Research Part E: Logistics and Trans-
quay crane scheduling problems in container terminals. Eur J Oper Res portation Review 2009;45(1):196–209.
2015;244(3):675–89. [33] Meisel F, Bierwirth C. A framework for integrated berth allocation and crane
[7] Buhrkal K, Zuglian S, Ropke S, Larsen J, Lusby R. Models for the discrete berth operations planning in seaport container terminals. Transportation Science
allocation problem: acomputational comparison. Transportation Research Part 2013;47(2):131–47.
E Logistics and Transportation Review 2011;47(4):461–73. [34] Moorthy R, Teo C-P. Berth management in container terminal: the template
[8] Campbell T, How JP. Bayesian nonparametric set construction for robust op- design problem. OR Spectrum 2006;28(4):495–518.
timization. In: 2015 American Control Conference (ACC); 2015. p. 4216–21. [35] Ning C, You F. Data-driven decision making under uncertainty integrating ro-
doi:10.1109/ACC.2015.7171991. bust optimization with principal component analysis and kernel smoothing
[9] Giallombardo G, Moccia L, Salani M, Vacca I. Modeling and solving the tacti- methods. Computers & Chemical Engineering 2018;112:190–210.
cal berth allocation problem. Transportation Research Part B: Methodological [36] Park KT, Kim KH. Berth scheduling for container terminals by using a sub–
2010;44(2):232–45. gradient optimization technique. Journal of the Operational Research Society
[10] Golias M, Portal I, Konur D, Kaisar E, Kolomvos G. Robust berth scheduling at 2002;53(9):1054–62.
marine container terminals via hierarchical optimization. Computers and Op- [37] Park YM, Kim KH. A scheduling method for berth and quay cranes. Or Spec-
erations Research 2014;41(1):412–22. trum 2003;25(1):1–23.
[11] Golias MM. A bi-objective berth allocation formulation to account for [38] Ribeiro GM, Mauri GR, Beluco SDC, Lorena LAN, Laporte G. Berth allocation in
vessel handling time uncertainty. Maritime Economics and Logistics an ore terminal with demurrage, despatch and maintenance. Computers and
2011;13(4):419–41. Industrial Engineering 2016;96(C):8–15.
[12] Guan Y, Cheung RK. The berth allocation problem: models and solution meth- [39] Shang C, Huang X, You F. Data-driven robust optimization based on kernel
ods. Or Spectrum 2004;26(1):75–92. learning. Computers & Chemical Engineering 2017;106:464–79. ESCAPE-26
[13] Han X, Lu Z-q, Xi L-f. A proactive approach for simultaneous berth and quay [40] Soyster A, Murphy F. Data driven matrix uncertainty for robust linear program-
crane scheduling problem with stochastic arrival and handling time. Eur J Oper ming. Omega (Westport) 2017;70:43–57.
Res 2010;207(3):1327–40. [41] Stahlbock R, Voß S. Operations research at container terminals: a literature
[14] Hanks RW, Lunday BJ, Weir JD. Robust goal programming for multi-ob- update. Or Spectrum 2008;30(1):1–52.
jective optimization of data-driven problems: a use case for the united [42] Steenken D, Voß S, Stahlbock R. Container terminal operation and operations
states transportation command’s liner rate setting problem. Omega (Westport) research - a classification and literature review. Or Spectrum 2004;26(1):3–49.
2020;90:101983. [43] UNCTAD. Review of maritime transport 2017.. Review of Maritime Transport
[15] Hendriks MPM, Armbruster D, Laumanns M, Lefeber E, Udding JT. Strategic 2017;volume 14(4):251–7.
allocation of cyclically calling vessels for multi-terminal container operators. [44] Ursavas E, Zhu SX. Optimal policies for the berth allocation problem under
Flexible Services and Manufacturing Journal 2012;24(3):248–73. stochastic nature. Eur J Oper Res 2016;255(2):380–7.
[16] Huang K, Suprayogi, Ariantini. A continuous berth template design model with [45] Vacca I, Salani M, Bierlaire M. An exact algorithm for the integrated plan-
multiple wharfs. Maritime Policy & Management 2016;43(6):763–75. ning of berth allocation and quay crane assignment. Transportation Science
[17] Imai A, Nagaiwa K, Tat CW. Efficient planning of berth allocation for container 2013;47(2):148–61.
terminals in asia. Journal of Advanced Transportation 1997;31(1):75–94. [46] Vis IFA, Koster RD. Transshipment of containers at a container terminal: an
[18] Imai A, Nishimura E, Papadimitriou S. The dynamic berth allocation prob- overview. Eur J Oper Res 2003;147(1):1–16.
lem for a container port. Transportation Research Part B: Methodological [47] Wang K, Zhen L, Wang S, Laporte G. Column generation for the integrated
2001;35(4):401–17. berth allocation, quay crane assignment, and yard assignment problem. Trans-
[19] Imai A, Yamakawa Y, Huang K. The strategic berth template prob- portation Science 2018;52(4):812–34.
lem. Transportation Research Part E: Logistics and Transportation Review [48] Wang N, Jin B, Lim A. Target-guided algorithms for the container pre-mar-
2014;72:77–100. shalling problem. Omega (Westport) 2015;53:67–77.
[20] Iris C, Lalla-Ruiz E, Lam JSL, Voß S. Mathematical programming formulations [49] Xiang X, Liu C, Miao L. A bi-objective robust model for berth allocation
for the strategic berth template problem. Computers and Industrial Engineer- scheduling under uncertainty. Transportation Research Part E Logistics and
ing 2018;124:167–79. Transportation Review 2017;106:294–319.
[21] Iris C, Lam JSL. Recoverable robustness in weekly berth and quay crane plan- [50] Xiang X, Liu C, Miao L. Reactive strategy for discrete berth allocation and quay
ning. Transportation Research Part B: Methodological 2019;122:365–89. crane assignment problems under uncertainty. Computers & Industrial Engi-
[22] Kim KH, Moon KC. Berth scheduling by simulated annealing. Transportation neering 2018;126:196–216.
Research Part B: Methodological 2003;37(6):541–60. [51] Xie F, Wu T, Zhang C. A branch-and-price algorithm for the integrated
[23] Kuzmicz KA, Pesch E. Approaches to empty container repositioning prob- berth allocation and quay crane assignment problem. Transportation Science
lems in the context of eurasian intermodal transportation. Omega (Westport) 2019;53(5):1427–54.
2019;85:194–213. [52] Zhen L. Tactical berth allocation under uncertainty. Eur J Oper Res
[24] Lalla-Ruiz E, González-Velarde JL, Melián-Batista B, Moreno-Vega JM. Biased 2015;247(3):928–44.
random key genetic algorithm for the tactical berth allocation problem. Appl [53] Zhen L, Chang DF. A bi-objective model for robust berth allocation scheduling.
Soft Comput 2014;22:60–76. Computers and Industrial Engineering 2012;63(1):262–73.
[25] Lee Y, Chen CY. An optimization heuristic for the berth scheduling problem. [54] Zhen L, Chew EP, Lee LH. An integrated model for berth template and
Eur J Oper Res 20 09;196(2):50 0–8. yard template planning in transshipment hubs. Transportation Science
[26] Liu C. Iterative heuristic for simultaneous allocations of berths, quay cranes, 2011;45(4):483–504.
and yards under practical situations. Transportation Research Part E: Logistics [55] Zhen L, Lee LH, Chew EP. A decision model for berth allocation under uncer-
and Transportation Review 2020;133:101814. tainty. Eur J Oper Res 2011;212(1):54–68.
[27] Liu C, Xiang X, Li Z. Two decision models for berth allocation problem un-
der uncertainty considering service level. Flexible Services and Manufacturing
Journal 2017(3):1–33.

16