Transportation Research Part B 139 (2020) 102–131

Contents lists available at ScienceDirect

Transportation Research Part B

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

Machine learning-driven algorithms for the container

relocation problem
Canrong Zhang a,b, Hao Guan a,b, Yifei Yuan a,b, Weiwei Chen c, Tao Wu d,∗
a
Research Center for Modern Logistics, Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China
b
Department of Industrial Engineering, Tsinghua University, Beijing 100084, China
c
Department of Supply Chain Management, Rutgers Business School-Newark and New Brunswick, Rutgers University, New Jersey 08854,
United States
d
School of Economics & Management, Tongji University, Shanghai 200092, China

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

Article history: The container relocation problem is one of important issues in seaport terminals which
Received 27 August 2019 could bring a significant saving on the operating cost even with a slight improvement due
Revised 1 March 2020
to the huge number of containers processed across the world each year. Given a specific
Accepted 24 May 2020
layout and container retrieval priorities, the container relocation problem aims to find the
optimal movement sequence to minimize the total number of container relocation oper-
Keywords: ations. In this paper, we propose novel machine learning-driven algorithms, which inte-
Container relocation grate optimization methods and machine learning techniques, to solve the problem. More
branch-and-bound algorithm specifically, we propose a new upper bound method called MLUB that incorporates branch
beam search pruners. These pruners are derived from some machine learning techniques through using
machine learning-driven technique the optimal solution values of many small-scale instances. The tightened upper bounds
generated by MLUB are used subsequently both in the exact branch-and-bound algorithm
called IB&B and the hybrid beam search heuristic called MLBS. Moreover, we also provide
a tighter lower bound for the problem by additionally considering the interaction between
consecutive target containers. Based on the benchmark data published recently in the lit-
erature, extensive experiments are conducted to test the performance of the proposed al-
gorithms. The experimental results demonstrate that the proposed algorithms outperform
the state-of-the-art algorithms reported in the literature, and some managerial insights
regarding the load intensity of the bay and some algorithm parameters such as the look-
ahead depth and the beam width are drawn from the results.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction

Globalization facilitates a steady rise in the volume of international trade, and the global logistics network is undertaking
more and more pressure from the demanding requirement. As an important part of the network, seaborne transportation
accounts for more than 60% of the trade amount and maintains a high growth rate in recent years. As a consequence, the
seaport container terminal, a critical hub in the seaborne transportation, is facing increasing pressure from the carriers,
shippers, and consignees in terms of providing higher-efficiency and more-flexibility services.

∗
Corresponding author.
E-mail address: [email protected] (T. Wu).

https://doi.org/10.1016/j.trb.2020.05.017
0191-2615/© 2020 Elsevier Ltd. All rights reserved.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 103

Fig. 1. An illustration of a bay.

Roughly speaking, the operations on the container terminals include four parts: receiving containers for temporary stor-
age before the arrival of destined vessels, the discharging operation, the loading operation, and the pickup operation by con-
signees to take their containers away from the terminal. This paper focuses on the loading operation, and, more specifically,
restricts our focus on the relocation operation during loading. For a more comprehensive understanding of the container
terminal operation, the interested readers are referred to Steenken et al. (2004) and the papers surveyed therein.
Before starting the loading operation, containers are transported to the terminal by water or land for temporary storage.
To increase the utilization of the yard space, a stack is formed by storing containers vertically, and a horizontal row of stacks
forms a bay, while a number of adjacent bays make up a block as shown in Fig. 1. Due to some loading rules required by
the stowage plan, for example, heavier containers required to be stacked under lighter ones when loaded onto the vessel,
each container will be given a retrieval priority during the loading operation. As the arrival of containers is uncontrollable,
containers with higher retrieval priorities are often buried beneath those with lower priorities, leading to the relocation
operation inevitable. This paper focuses exactly on this kind of relocation operations occurring during the loading operation.
To be more specific, we assign a smaller number to the container with a higher retrieval priority, and vise versa. For
example, as shown in Fig. 1, container 1 should be retrieved first, container 2 should be retrieved next, and so on. When
we try to retrieve one container, there may be two cases: if the target container happens to be on the top of its stack, then
we can just perform the retrieval operation; otherwise, if some containers are stored on the top of the target one, then the
relocation operation is required. For example, as container 7 is stored on the top of container 1, it should be relocated to
other stacks. Therefore, the container relocation problem (CRP) studied in this paper aims to give an optimized operation
sequence for retrieving containers from a bay with the goal of minimizing the relocation operations as much as possible.
As such relocation operations are time-wasting and unproductive, a method which can generate an operation sequence
with fewer relocation operations is appealing due to a huge number of containers handled across the world each year.
In addition, the relocation problem distinguishes itself as a difficult one in the combinatorial field because the relocated
container causes a chain effect on the subsequent retrieval operations. Due to these two considerations, this paper attempts
to propose a novel method, which integrates optimization methods and machine learning techniques, to solve the problem,
and validates the method using the benchmark test instances reported on the recently published literature. Accordingly, this
paper makes the following contributions:
1. Valuable information regarding the promising nodes selection (called the branch pruner in the later section) is extracted
first and then incorporated into a heuristic method called MLUB to generate tight upper bounds. To the best of our
knowledge, it is the first attempt to integrate optimization methods and machine learning techniques in the algorithm
design for the CRP;
2. An exact branch-and-bound algorithm which relies on the MLUB to determine the branching precedence is proposed.
Contrary to using the best lower bound to select the next node to explore, this exact algorithm adopts the best upper
bound to select the next node to explore;
3. To handle large-scale instances, we also propose a beam search algorithm that makes use of the MLUB in its beam node
selection. In addition, some screening techniques are also proposed and used in the searching process to further enhance
its convergence;
4. Based on the benchmark data, the numerical experiments reveal that our proposed algorithms outperform the state-of-
the-art algorithms reported in the literature, and some managerial insights regarding the load intensity of the bay and
some algorithm parameters such as the look-ahead depth and the beam width are drawn from the results.
The remainder of this paper is organized as follows. Section 2 reviews the related works. Section 3 defines the container
relocation problem. Sections 4 and 5 describe the proposed exact branch-and-bound algorithm and the beam search heuris-
104 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

tic, respectively, with their machine learning-driven branching pruners presented in Section 6. Results of the numerical
experiments are presented in Section 7. Finally, conclusions and suggestions for future research are given in Section 8.

2. Related work

As an essential intermodal interface in the international shipping network, container terminals are crucial to the efficient
circulation of cargoes and play an important role in the global competition. In recent years, with more and more research
studies dedicated to the quay side of container terminals (Zhen et al., 2017; Wang et al., 2018; Iris and Lam, 2019; Iris et al.,
2018), increased pressure has transferred from quay side to yard side. As such, the yard optimization urgently needs to be
addressed.
Yard optimization primarily consists of two parts: the yard crane planning and container handling, at the operational
level. As summarized in Zehendner et al. (2016), the former delves into the yard crane assignment/scheduling problem to
assign yard cranes to storage areas and to define a schedule of storage and retrieval operations for each crane; the latter
focuses on the optimization for the handling process of containers, including the storage allocation problem to determine
storage locations for incoming containers, the remarshalling/premarshalling problem to reorganize parts of the storage area
(a block or a bay) in less busy periods, and the container relocation problem to retrieve all containers from a bay in a given
sequence with a minimum number of relocations. For the specific introduction and review of the three problems, we refer
the interested readers to the works given by Bortfeldt and Wäscher (2013), Carlo et al. (2014) and Lehnfeld and Knust (2014).
Plenty of studies have been performed on the aforementioned two parts of yard optimization. For the yard crane plan-
ning, Zhang et al. (2002) is among the first few research studies that addressed the crane deployment problem, which
was formulated as an MIP model and solved by an improved Lagrangean relaxation method. On this basis, the single-crane
scheduling (Gharehgozli et al., 2014b; Zheng et al., 2018) and the scheduling with crane interference (Gharehgozli et al.,
2014a; Virgile et al., 2018) are widely studied. For the container handling, most of the research studies focus on the storage
allocation problem to determine storage locations for incoming containers. Under two cases with different objectives, the
container allocation problem was studied in Kim and Kim (2002). In order to assign containers to routes with minimum cost,
Bell et al. (2011), Bell et al. (2013) and Wang et al. (2015) studied frequency-based, cost-based and profit-based container as-
signment problems under maritime liner shipping networks, respectively. Based on Kim et al. (20 0 0) and Zhang et al. (2010),
Zhang et al. (2014a) considered the stack height and state-changing magnitude information when interpreting the punish-
ment parameter, and proposed two new “conservative” allocation models for outbound containers in container terminals.
Zhang et al. (2014b) presented two new dynamic programming models for the container allocation problem with additional
consideration of adjusted probability on the types of arriving containers. By contrast, there are relatively few research studies
dedicated to the remarshalling/premarshalling problem (Lee and Chao, 2009; Hottung et al., 2020). The review on the CRP
will be presented separately a moment later. In addition, abundant studies were conducted for the integration of the yard
crane planning and container handling, for example, Zhang et al. (2003), Zhen (2016) and Jiang and Jin (2017). And there
are also some research studies that studied from the strategic and tactical points of view, for example, Liu et al. (2016) and
Zhen (2016). Since strategic and tactical decisions are not the focus of this paper, we refer the interested readers to the
review by Lee and Song (2017) for more information.
As mentioned previously, the focus of this paper is the container relocation problem. In the following, we will focus on
the review on this topic. As an NP-hard problem (Caserta et al., 2012), the CRP is firstly studied by Kim and Hong (2006), in
which they proposed a branch-and-bound method to solve the CRP to optimality. As branch-and-bound algorithms can be
time-consuming for large-instances, they established a heuristic using “expected number of additional relocations” (ENAR) to
determine the stack for relocated containers with the goal to minimize the number of relocations during container loading.
Caserta et al. (2011) proposed a “two-dimensional” corridor method, in which a dynamic programming algorithm is used
as a subroutine, to tackle the exponential growth of the number of states for large-scale problems, and the validity of this
method was proved by comparing with Kim and Hong (2006) in numerical experiments. Forster and Bortfeldt (2012) pro-
posed a heuristic tree search procedure for the CRP that consists of four elements. Zhu et al. (2012) developed iterative
deepening A∗ algorithms to solve the CRP. They presented three different lower bounds (LB) calculation methods and four
probe heuristics embedded in the algorithms. The computational results show that their method can solve most cases to the
optimum in a time suitable for practical application. On this basis, Tanaka and Takii (2016) proposed a new LB calculation
method called LB4 and proved that LB4 is greater than or equal to LB3 established by Zhu et al. (2012). In this paper, we
will provide a tighter lower bound, based on LB4, by additionally considering the interaction between consecutive target
containers.
Jin et al. (2015) constructed an improved greedy look-ahead heuristic for the unrestricted CRP, which can be split into
three levels. The top level keeps executing the relocation and retrieving operation until the layout is empty. The stack to
take in the relocated container is given by the middle level using limited tree search and the bottom level applies sev-
eral heuristic rules to evaluate the leaf nodes generated by the middle level. Jovanovic and Voß (2014) proposed a chain
heuristic, based on the classical Min-Max method proposed by Caserta et al. (2011), considering both the relocated con-
tainer and the container to be relocated next. Ting and Wu (2017) improved the Min-Max heuristic and proposed a vir-
tual relocation heuristic (VRH), which took all the blocking containers above the target container into account at the same
time. Galle et al. (2018) improved upon and enhanced an existing binary encoding and using it, formulated the restricted
CRP as a binary integer programming problem in which they exploited structural properties of the optimal solution. Later,
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 105

Jin (2020) identified two deficiencies contained in the model of Galle et al. (2018) in formulating the “last in, first out” (LIFO)
policy, and provided the corrected integer programming formulation. Jovanovic et al. (2019) presented an ant colony opti-
mization algorithm for the CRP, which can be applied to both the restricted and unrestricted CRP and outperform existing
methods for the CRP. Bacci et al. (2019) proposed a new beam search approach, called the Bounded Beam Search algorithm,
to solve the CRP and compared it with other methods in the literature, showing that it outperforms most of the existing
approaches.
As machine learning techniques are incorporated in the design of our algorithm, we also give a brief review on this field.
Machine learning is usually used in learning (Siripirote et al., 2015; Wei and Liu, 2013; Allahviranloo and Recker, 2013),
estimation (Yin et al., 2015; Zheng and Zuylen, 2013) and prediction (Andres and Nair, 2017) of parameters, as well as to
enhance the modeling process (see, e.g., Hofleitner et al. (2012), Sopasakis and Katsoulakis (2016), Kouvelas et al. (2017),
Arentze and Timmermans (2004), and Dong et al. (2018)).
The study which adopts machine learning techniques for the container location-related problems is very scant in the
literature. Hottung et al. (2020), to the best of our knowledge, is the only published report which proposed a new method
called Deep Learning Heuristic Tree Search (DLTS) for the container pre-marshalling problem (CPMP). In the paper, the
deep neural networks are used to learn solution strategies through analyzing existing (near-) optimal solutions to CPMP
instances, and then the networks are integrated into a tree search procedure to decide which branch to choose next and
to prune the search tree. Although Hottung et al. (2020) and our paper utilizes the similar idea, their differences are evi-
dent and can be summarized as follows: (i) The CRP studied in this paper considers the retrieval of containers, while the
CPMP studied in Hottung et al. (2020) does not; (ii) In terms of the method for generating the feasible solutions (that is,
upper bounds in both papers), our method combines a learning mechanism and a heuristic method to yield a feasible so-
lution, while there are no heuristics involved in Hottung et al. (2020) which relies heavily on the deep neural networks;
(iii) Hottung et al. (2020) proposes a heuristic method without considering the global optimality, while our paper further
incorporates the heuristic into the exact branch-and-bound framework to generate optimal solutions; (iv) The deep neural
networks proposed in Hottung et al. (2020) encounter the scalability issue, while this paper handles the scalability issue
well.
In summary, different types of optimization algorithms including the exact and heuristic ones were proposed in the lit-
erature to address the CRP, and several types of machine learning techniques such as the decision tree, the neural network
and so on were adopted to parameter estimation and system optimization in a wide range of application. In the CRP, how-
ever, the algorithms that combines the optimization methods and the machine learning techniques are very scant. Inspired
by the fact that more valuable information regarding the structure of the optimal solution can be extracted and facilitates
the solution for large-scale instances, we are motivated to make such an attempt in this regard.

3. Problem description

Generally, the container relocation problem can be divided into two categories by two classification criteria: (i) the re-
stricted and unrestricted problems; (ii) the problem with distinct priorities and the one with duplicate priorities. The former
classification depends on the range of containers that can be chosen to be relocated. If only the blocking containers right
above the target container can be relocated, it is a restricted CRP; otherwise, it is called unrestricted CRP. The latter classifi-
cation is determined by the uniqueness of retrieval priorities. The problem with distinct priorities implies a unique retrieval
order of containers, and the other specifies precedence relations between groups of containers. In this paper, we consider
the restricted CRP in which each container is given a distinct retrieval priority. More specifically, the problem is to deter-
mine, arrange and adjust the horizontal and vertical positions of the containers in a bay, in order to obtain an optimal
sequence of operations for retrieving containers according to a given retrieval priority, with the objective of minimizing the
number of relocations.
For the problem, we make the following assumptions:

1. The container layout and retrieval priority are given beforehand;

2. All containers are uniform in length and height, and they can be described in unit size;
3. There exists a stack height limit for the consideration of security;
4. The relocation operation only occurs within the same bay due to the consideration that the horizontal movement of yard
cranes carrying containers is time-consuming;
5. The retrieval priority of each container is unique, and there is no priority shared by multiple containers, i.e., this paper
studies the problem with distinct priorities;
6. To better fit the worker’s operational convenience, only the containers right above the target container can be relocated,
i.e., the restricted CRP is studied in this paper.

4. The exact algorithm (IB&B)

In this section, we propose an exact algorithm, which is essentially an iterative branch-and-bound algorithm (IB&B), for
the CRP, and present the lower bound (LB) generating method called LB4 and the upper bound (UB) generating method
called MLUB, respectively.
106 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

4.1. The framework of the algorithm

Before presenting the framework of the IB&B, we first give a brief discussion on the classical branch-and-bound algorithm
(CB&B). The basic idea of the CB&B is dividing and conquering which exploits the relationship between the local lower bound
(with respect to the node currently under consideration) and the global upper bound (with respect to the original problem
for a minimization objective) to trim some unpromising nodes and facilitate the convergence. To abate confusion and ease
exposition, the bounds without any modifiers such as “global” or “local” refer to the global ones throughout the paper, while
the bounds with “global” or “local” refer to those with respect to the original problem and the local node, respectively. As
mentioned in Morrison et al. (2016), three components can be fine-tuned to improve the performance of the algorithm: the
branching strategy, the search strategy, and the pruning rules. In the following, we will present them one by one.
The branching strategy represents how the present node is partitioned to produce new nodes to expand the tree. For
the container relocation problem we study, the branching process is straightforward. Given a specific layout of the bay
under consideration, if the relocation operation is inevitable, which means the target container is buried under those with
lower retrieval priority (called blocking containers), then the blocking containers need to be relocated sequentially from the
top down. For the first blocking container, we should initially identify the set of all the stacks that can accommodate the
blocking container, and then make the decision about reposition. To be more specific, let c denote the first blocking container
that is on the top of the stack where the target container is buried. Then, the set of all stacks that can accommodate the
blocking container c can be denoted by CS(c ) = {s|s ∈ S, s = s(c ), h(s ) < H }, where S denotes the set of stacks indexed by s,
s(c) denotes the stack where container c is originally located, h(s) represents the height of stack s, and H is the stack height
limit. As a result, we can relocate container c to each stack in the set of CS(c), and update the bay state accordingly, thereby
generating |CS(c)| states (or called nodes) in total. In the following, the nodes and states will be used interchangeably. From
each newly generated node, the branching process can be restarted for the second blocking container (if any) until the target
container is exposed to the top.
The search strategy refers to how to expand the searching tree consisting of the nodes generated by the branching
strategy mentioned previously. Under the breadth-first search strategy (BFS), the searching tree is expanded uniformly with
the frontier moving forward evenly. Under this strategy, there is no notable difference regarding which node will be treated
first, thereby making its implementation straightforward and easy, but it has a significant drawback that is a feasible solution
will not be revealed until the frontier cannot move forward any further. To circumvent this drawback, this paper intends to
adopt the depth-first search strategy (DFS) which explores from the selected node and goes forward as deep as possible, and
backtracks to the most recently visited node if it cannot proceed any further. For this search strategy, which node will be
investigated first makes a significant difference. The previous studies usually used the LB as the criterion, that is, the node
with the best lower bound will be chosen to be investigated first. It is called the LB-priority strategy in this paper. This
strategy may misguide the search process especially when the quality of the lower bound is low. Although the upper bound
also suffers the same drawback, a tighter upper bound is more likely to be a good indicator for a better neighborhood.
Therefore, in this paper, rather than using the lower bound, we attempt to utilize the UB as the criterion to select the next
node. This strategy is called the UB-priority strategy. The comparison between the two strategies will be conducted through
numerical experiments in Section 7.3.3.
With respect to the pruning strategy, a conventional way is adopted by exploiting the relationship between the local LB
and the global UB. More specifically, there are primarily two situations: (i) if the local LB and UB are equal at one node,
then with no need to explore further, this node has arrived at the local optimality and can be used to update the global UB
of the original problem whenever possible; (ii) if the local LB of one node plus the number of relocation operation executed
so far is greater than or equal to the incumbent global UB of the original problem, then the node certainly cannot produce
a better solution and will be pruned.
Similar to Tanaka and Takii (2016), the framework of the IB&B is shown in Algorithm 1 , with the embedded branch-and-
bound algorithm (EB&B) showed in Algorithm 2 . Given the layout of a yard bay and the retrieval priority for each container,

Algorithm 1: IB&B.
Input: L ← initial layout
LB ← lower bound of L by LB4
UB ← upper bound of L by MLUB
while LB < UB do
Call EB&B with designated LB and UB
if there exists feasible solution = LB then
An optimal solution is found
break
end
Update UB with the better UB obtained from the execution of EB&B (if any)LB = LB + 1
end
return LB
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 107

Algorithm 2: EB&B.
Input: L ← initial layout; LB ← lower bound of L by LB4; UB ← upper bound of L by MLUB
LayoutList ← null
F indingF easibl eSol utionF l ag ← 0
Identify all relocation choices for L, perform these relocation operations to generate child states, and sort these child
states by their upper (lower) bounds for the UB-priority (LB-priority) version of the EB&B
Add L to LayoutList
while Layout List .size! = 0 do
Get the next operation to perform of the most recently visited layout in LayoutList
if no more operations to perform then
Remove the most recently visited layout in LayoutList
end
else
Generate a new layout L1 by performing the pre-specified relocation operation
Identify all relocation choices for L1, perform these relocation operations to generate child states, and sort these
child states by their upper (lower) bounds for the UB-priority (LB-priority) version of the EB&B
lb ← lower bound of L1 by LB4
ub ← upper bound of L1 by MLUB
if ub = lb and lb + Layout List .size = LB then
F indingF easibl eSol utionF l ag = 1
break
end
else if ub! = l b and l b + Layout List .size ≤ LB then
if ub + Layout List .size < UB then
Update UB by ub + Layout List .size
end
Add L1 to LayoutList
end
end
end
if F indingF easibl eSol utionF l ag = 1 then
return True
end
else
return False
end

it is able to obtain an LB and an UB (whose generation algorithms will be presented in Sections 4.2 and 4.3, respectively).
Different from the CB&B, the EB&B is partially implemented in the framework to search for a specific target in each calling.
Initially, the global LB is set as the search target, and all the nodes that satisfy either one of the following conditions will be
removed: (i) arriving at the local optimality with an optimal value not equal to the search target; (ii) becoming unpromising
when compared to the LB. That is to say, only the nodes with the local LB plus the number of relocation operation executed
so far smaller than or equal to the current target value are retained, with the aim to reduce the searching effort. If no
feasible solution equal to the target value is found after one search, the global UB will be replaced with the newly found
best UB (if any), and the global LB will be increased by 1, then the implementation of the EB&B with the updated global LB
and UB will be performed again until the feasible solution is found or the global LB is equal to the global UB. Since IB&B
terminates when the global LB equals the global UB, the method generates optimal solutions and is an exact algorithm.

4.2. The method for calculating lower bounds (LB4)

To calculate lower bounds, our proposed algorithms hybridize the LB4 method proposed by Tanaka and Takii (2016),
which takes into consideration the combinations of destination stacks for the blocking containers above the target one. We
refer the interested readers to Tanaka and Takii (2016) for its details. We notice that, in LB4, the blocking containers on a
target container are independent of those of the next (remaining) target container. Actually, the blocking containers of the
first target container affect the relocation of the blocking containers of the next (remaining) target container. That is to say,
they are interdependent. We, therefore, are able to propose a tighter lower bound called the enhanced LB4 (or ELB4 for
short) which is based on LB4 but additionally considers the interaction between two or more consecutive retrieval of target
containers. The underlying idea of ELB4 can be illustrated through Fig. 2.
108 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Fig. 2. Comparison between LB4 and ELB4.

As shown in Fig. 2(a), according to LB4, the initial bay status (a1) branches into (a2) and (a3), and then the calculation
for LB4 corresponding to the first target container (that is, container 1) is terminated. To calculate the LB4 for the second
target container (that is, container 2), container 1 and all the blocking containers above it (that is, container 3) are removed,
thereby generating the initial bay status for container 2, which corresponds to (a4), and then the LB4 for container 2 can
be calculated. Herein, the lower bounds of relocation numbers required for the first two target containers are calculated
independently, and the sum of them are equal to 0. Under ELB4, however, we reserve the container whose priority changes
the priority status of the accommodating stack, corresponding to Fig. 2(b5). Starting from (b4) and (b5), we further conduct
the LB4 process for container 2, and obtain a tighter LB for the first two containers, which is ELB4 = 1. Herein, ELB4 obtains
a better lower bound than LB4 because it considers the interaction between two consecutive target containers. We omit the
proof process of this claim due to its straightforwardness.
Fig. 2 only demonstrates the interaction between two consecutive target containers. Actually, the interaction can be ex-
tended to 3 or more consecutive target containers. For simplicity, in this paper, we only consider the interaction between
two consecutive target containers. Although ELB4 is tighter than LB4, our preliminary experiments show that the ELB4
method takes much longer time than LB4 while the improvement on the lower bounds is not that significant relative to
the extra time consumed, as discussed in Section 7.1. Considering that the LB needs to be calculated at each node in the
branch-and-bound tree, we still use the LB4 method for our proposed algorithms. It is noted that ELB4 is a good choice when
high-quality lower bounds are required and long runtime is allowed, and, for the particular case, the number of consecutive
target containers whose interactions are considered simultaneously needs to be set to a larger value.

4.3. The method for calculating the machine learning-driven upper bounds (MLUB)

In this section, we propose a machine learning-driven method for the upper bound generation. The method is based
on the virtual relocation heuristic (VRH) proposed by Ting and Wu (2017) which in turn is based on the Min-Max method
proposed by Caserta et al. (2011). For completeness, the explanation on VRH is given in Appendix based on a given example,
while the Min-Max method is proposed below as some definitions are needed in the later sections. Moreover, a more unified
formula is proposed in this paper to replace the original one to simplify the implementation. After that, the MLUB for
generating the upper bound will be presented.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 109

Fig. 3. Selection principle for Type-1 stacks.

Let us first define the priority number of a non-empty stack s as p(s ) = min{ p(c )|s(c ) = s}, where p(c) is the retrieval
priority of container c and the priority number of the stack is expressed as the highest priority (i.e., the smallest priority
number) among all the containers belonging to that stack. If stack s is empty, p(s) is set to N + 1. Assume c denotes the
container to be relocated, and then the available stacks to accommodate the relocated container can be divided into two
categories according to whether p(s) is larger than p(c) or not. For the stack whose priority number is larger than p(c),
relocating the container onto it will not cause future relocation, then we call such a relocation a well-relocated one and call
the corresponding stack a Type-1 stack. On the contrary, if the future relocation of the relocated container is inevitable, we
call the accommodating stack a Type-2 stack.
The Min-Max method first checks the existence of the well-relocated stacks when choosing a stack to accept the relo-
cated container. On the one hand, if there is more than one well-relocated stack, the stack with the highest priority will be
chosen, that is, s∗ = arg min{ p(s )|s ∈ S, p(s ) > p(c ), h(s ) < H }, where S and H denote the set of stacks and the stack height
limit, respectively. This selection policy is aimed at reserving more well-relocated stacks for the future blocking containers.
As shown in Fig. 3(a), the current target container is 1 and the container to be relocated is container 2. Under the assump-
tion that the stack height is limited to 3, the feasible receiving stacks are stacks 2 and 3, both of which are well-relocated
ones. If container 2 is relocated to stack 2, stack 3 remains Type-1 stack for the next blocking container 6 as shown in
Fig. 3 (b). If container 2 is relocated to stack 3, however, neither the two stacks meet the well-relocated condition for next
relocation and container 6 is bound to be relocated at least twice as shown in Fig. 3(c).
On the other hand, if none of the available stacks meets the well-relocated condition, the stack with the lowest prior-
ity is selected, that is, s∗ = arg max{ p(s )|s ∈ S, p(s ) < p(c ), h(s ) < H }. This selection policy tries to put aside the relocated
container so that it has no or little impact on the retrieval for the immediately subsequent several target containers. In
summary, under the Min-Max rule, if there exist stacks with lower priority than that of the container to be relocated, the
one with the highest priority will be chosen; otherwise, the stack with the lowest priority will be chosen. We can unify the
policy for the two scenarios by introducing a new stack priority of stack s denoted by p∗ (s) as follows.

p( s ) − p ( c ) i f p( s ) > p( c )
p∗ (s ) = (1)
2 N + 1 − p( s ) otherwise

where N is the total number of containers at the beginning, p(s) is the highest container priority of stack s and set to N + 1
when the stack is empty. The use of 2N + 1 ensures that the p∗ (s) value of stacks satisfying the well-relocated condition
is always smaller than that of stacks not satisfying the condition. With the introduction of p∗ (s), the aforementioned two
policies applied to two different scenarios have been unified by adopting the smallest number of p∗ (s) no matter whether
they are Type-1 or Type-2 stacks.
The basic idea of the MLUB is shown in Fig. 4(b) which is essentially an enhanced version of the look-ahead method. As
shown in Fig. 4(a), we expand from node 1 to get a complete set of searching branches for the first D layers. Assume that
at the Dth layer, there are a total number of K states. And then by using the VRH method proposed by Ting and Wu (2017),
we can get at most K upper bounds (denoted by ub), among which, min{ub} + D will be set as the UB of the initial node.
It is evident that, with the increase of D, the calculation time of this method increases exponentially. Accordingly, we are
motivated in this paper to eliminate some branches which are unlikely to be selected in the D layers in terms of generating
good-quality UBs in order to reduce the searching efforts. To the end, a branch pruner (or called filter) extracted from the
machine learning-driven methods will be proposed to prune unpromising branches and expedite the convergence. As shown
in Fig. 4(b), if branch 1 → 4 is pruned by the filter in the first layer, and branches 2 → 7 and 3 → 9 are pruned in the
second layer, the number of nodes to calculate UBs will decrease dramatically. To the best of our knowledge, it is the first
paper attempting to increase the searching efficiency of the algorithm by adopting the learning mechanism for the CRP. How
to generate the filter by the learning mechanism will be introduced in Section 6.
110 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Fig. 4. (a): Illustration of the look-ahead method without branch pruners; (b) Illustration of the look-ahead method with branch pruners.

Fig. 5. Illustration of the MLBS.

5. The beam search algorithm (MLBS)

Even though the exact algorithm has been enhanced as presented in the previous section, it still encounters difficulty
when solving large-scale instances. We, therefore, have to resort to heuristic algorithms to handle such large-scale instances.
By well utilizing the branching process provided by the branch-and-bound algorithm, we propose the MLBS in the following
section.

5.1. The framework of the algorithm

Beam search is a classical heuristic method, which is similar to the breadth-first tree search, with the difference lying
in the number of nodes to explore. More specifically, the beam search strictly restricts the number of reserved nodes equal
to or smaller than the beam width at each level to reduce the searching space. The reserved nodes at each level are called
beam nodes and only the beam nodes can create their successors while others will be discarded. There are two types of
beam search algorithms in terms of how to generate succeeding beam nodes: the interdependent one and the independent
one. The former one generates only one descendant from each beam node while the latter one selects next-generation beam
nodes from all “legal” descendant nodes. As shown in Fig. 5, suppose the beam width is 2, node 1 can branch into nodes
2 and 3, node 2 has 3 successors, and node 3 has 2 successors. Fig. 5(a) and (b) schematically illustrate the node selection
process by the interdependent and independent beam searches, respectively.
In this paper, we use the independent beam search method, that is, all the descendent nodes in each layer are put to-
gether and compared with each other to screen out the beam nodes. The pseudocode of the MLBS is shown in Algorithm 3 .
In the algorithm, the evaluation function is particularly important as it directly determines which node will be left for the
subsequent exploration and which will be discarded immediately. We adopt the UB obtained by the MLUB presented in
Section 4.3 as the evaluation function. The smaller the UB, the better the quality of the node, and the more likely it is to get
a better solution. The LB of each node as calculated by the LB4 presented in Section 4.2 is also calculated at each level to
discard unpromising nodes with the help of the UB. In addition, a set of screening techniques presented in Section 5.2 will
also be used to further reduce the candidate nodes for beam node selection. The global UB will be updated whenever a
better UB is found. Once the global UB equals LB, the optimal solution has been found and the search terminates. Another
termination condition is when the beam node set becomes empty in one certain level, UB will be the final solution provided
by the algorithm.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 111

Algorithm 3: MLBS.
Input:LB ← lower bound of L by LB4
UB ← upper bound of L by MLUB
if LB = UB then
return UB
end
else
Retrieve the target container until the target container is inaccessible
L ← the layout after retrieving until the target container is inaccessible
i ← 0; Ci ← ∅; f lag ← 1
Add L to Ci
while f lag = 1 do
i=i+1
for each beam node in beam node set Ci do
Get the blocking container B on the top of the target container
for each optional stack do
Relocate B to the stack and get a new layout L1
lb ← lower bound of L1 by LB4
ub ← upper bound of L1 by MLUB
if lb + i < UB then
Add the new layout to beam node set C(i+1 )
end
if ub + i < UB then
UB = ub + i
end
if UB = LB then
f lag = 0
end
end
end
Delete duplicate nodes, sort the remaining nodes in C(i+1 ) by their upper bounds in an ascending order, and
only retain the first W nodes
for each beam node in beam node set C(i+1 ) do
Retrieve the target container until the target container is inaccessible
end
if C(i+1 ) is empty then
f lag = 0
end
end
return UB
end

5.2. Screening techniques

It is evident that the quality of the beam nodes has a significant impact on the quality of the algorithm, which makes it
necessary for us to provide high-quality node pool for selection. To this end, a set of screening techniques is presented as
follows.

1. Calculate the UB and LB of the node simultaneously. If the UB equals the LB, then this node needs not to be added into
the candidate set for beam nodes selection. In this case, the node has arrived at the local optimality, with no need to
explore further, and can be used to update the global UB if possible;
2. If the LB of one node plus the number of relocation performed so far is greater than or equal to the current global UB,
this node cannot produce a better solution, and therefore will not enter the candidate set for beam nodes selection;
3. During the execution, some beam nodes might be exactly the same in a certain layer as shown in Fig. 6, in which the
stack limit is set to 3, and states (f) and (g) in the figure are actually the same. Therefore, before sorting the nodes, we
will perform a de-duplication screening to ensure that all the nodes in the candidate set are distinctive;
4. In order to deal with the situation when there are multiple nodes with the same UB in the candidate set, we propose to
select the node whose relocation leads to the smallest p∗ (s) value defined by Eq. (1), contrary to the random selection
adopted in Ting and Wu (2017).
112 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Fig. 6. Illustration of the same beam nodes.

6. The machine learning-driven branch pruners

In this section, we will present the machine learning-driven branch pruners, called filters for short, to facilitate the
convergence of the upper bound.

6.1. Adopted datasets

The datasets chosen to generate exact solutions for filter generation generally need to meet the following two conditions:
(i) the instances must be small-scale, so that the exact solution can be relatively easy to obtain; (ii) the instances must be
representative and diversified as well, so that the extracted knowledge, to a large extent, will be meaningful and useful.
Based on the aforementioned conditions, we choose, from Zhu et al. (2012), the datasets of the form H − S − N, including
3-7-18, 3-7-19, 3-7-20, 3-8-21, 3-8-22, 3-8-23, 3-9-24, 3-9-25, 3-9-26, 3-10-27, 3-10-28, 3-10-29, 4-6-20, 4-6-21, 4-6-22, 4-
6-23, 4-7-24, 4-7-25, 4-7-26 and 4-7-27, where H is the stack height limit, S is the total number of stacks, and N is the
initial number of containers, which varies between H (S − 1 ) to HS − 1. Besides, we choose, from Caserta et al. (2011), the
datasets of the form T − S, including 3-3, 3-4, 3-5, 3-6, 3-7, 3-8, 4-4, 4-5, 4-6 and 4-7, where T is the stack height, S is
the number of stacks, and the number of containers is by default TS, meaning that each stack contains T containers at the
initial state. There is no stack height limit in the original dataset in Caserta et al. (2011), while in order to make it more
consistent with the reality and comparable with the recent research, we assume the stack height limit is the initial stack
height plus 2, i.e., H = T + 2. The dataset from Zhu et al. (2012) represents the situation where the bay is relatively full and
randomly placed, while the dataset from Caserta et al. (2011) considers the situation where the bay is relatively empty and
evenly placed. Each dataset in Zhu et al. (2012) contains 100 instances and each dataset in Caserta et al. (2011) consists of
40 instances, leading to a total number of 20 × 100 + 10 × 40 = 2400 instances. The exact solutions of the aforementioned
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 113

small-scale instances will be generated by enumeration, represented as the optimal operation sequences, from which the
filter will be constructed.

6.2. Random forest based branch pruner

The random forest is a classifier that utilizes multiple decision trees to train and predict samples. As the basis of the
decision tree, information entropy is an index to quantitatively describe uncertainty from the perspective of probability
theory. The more chaotic the system, the greater the uncertainty of the system, the greater the information entropy, and
vice versa. Classification by reasonable properties, however, can effectively reduce the chaos of the whole system, thereby
reducing the information entropy of the system. In view of this observation, given a dataset that contains data objects with
properties describing the data objects, we can extract the classification effectiveness under each property, primarily due to
the identification of the change ratio of information entropy relative to the initial status without classification. There are
multiple algorithms such as ID3 and C4.5 supporting the implementation of random forest. Since the relative change in
the entropy rather than the absolute change is considered in this paper, the C4.5 algorithm which focuses more on the
information gain ratio better fits our case, and, therefore, is adopted in this paper. Accordingly, the property leading to
the most effective classification is firstly selected as the current dividing point of the decision tree. Then we exclude this
property and repeat the process at the frontier node in the current decision tree until the establishment of the whole tree
is completed. Different decision trees are established for the samples drawn by the replacement sampling method, and then
aggregated to a random forest, where the modal number of the results under all decision trees in the random forest will be
adopted as the final prediction output of the random forest.
Based on the random forest, the proposed branch pruner is actually a classifier which classifies stacks into several cate-
gories, and only keeps the target category of stacks and discards the rest to achieve the goal of pruning. Accordingly, each
relocation in the optimal operation sequence of the instances introduced in Section 6.1 corresponds to a data object, and
we need to find reasonable properties to describe the data object. To this end, we first need to define the representation of
the bay for each configuration of containers piled up in the stacks. With the definition, we are able to identify the optional
stacks for each container to be relocated, among which, the best stack will be chosen and the corresponding index will be
recorded. It should be noted that when defining the representation of the bay, the information on the size of the bay should
not be incorporated, as the information on size may prevent it from being extended to larger-scale instances. To represent
the bay, we introduce the following two indicators, both of which are unitless so that the size-independent requirement can
be met.

1. R1(s): the ratio to represent the relationship between the selection preference of stack s and the total number of optional
stacks that can accommodate the container to be relocated;
2. R2(s): the ratio of the number of the blocking containers to the total number of containers in the same stack s.

Taking Fig. 7 as an example, the current target container is container 1 and container 9 is to be relocated. With the
assumption that the stack height limit is 3, the set of optional stacks to accommodate container 9 is S3 ={2, 4, 5, 6}, and
p∗ (2 ) = 18 (as p∗ (2 ) = 2N + 1 − p(2 ) = 2 ∗ 12 + 1 − 7 = 18 according to the definition of p∗ (s) as shown in Eq. (1)), p∗ (4 ) =
17 (whose calculation is the same as p∗ (2)), p∗ (5 ) = 1 (as p∗ (5 ) = p(5 ) − p(c ) = 10 − 9 = 1 according to Eq. (1)) and p∗ (6 ) =
19 (whose calculation is the same as p∗ (2)), then we can get a stack selection sequence by sorting the p∗ (s) values in an
ascending order, i.e. {5, 4, 2, 6}. Given the stack selection preference ranking, we can calculate R1(s) for each stack as:
R1 ( 2 ) = 2/3, R1 ( 4 ) = 1/3, R1 ( 5 ) = 0/3 = 0, R1 ( 6 ) = 3/3 = 1.
The calculation of R2 is straightforward. As shown in Fig. 7, considering container 12 is a blocking container in stack
2, and the total number of containers in stack 2 is 2, thereby R2(2 ) = 1/2. Following the same fashion, we can calculate
R2(4 ) = 0, R2(5 ) = 1/2, and R2(6 ) = 0.
To validate the significance of the two indicators, we calculate the values of R1(s) and R2(s) for different stacks in the 30
datasets introduced in Section 6.1, and then we observe the average values, with respect to the two indicators, of the stacks
that are chosen in the exact solution (labeled 1 as shown in Figs. 8 and 9), and those that are not in the exact solution
(labeled 0). It is evident that the average value of stacks labeled 1 are always below those of stacks labeled 0, which implies
that the two indicators are effective to differentiate the selected stacks from the unselected ones.
With the previous description, we are able to describe a stack and further a bay. As a bay is composed of many stacks,
we can divide stacks according to the two ratios R1(s) and R2(s) and count the number of stacks belonging to different
stack types to describe the bay. To this end, our problem turns out to determining two dividing points, or thresholds, for
the two ratios, respectively. The average of the two ratios for the 30 datasets in terms of different stack types (as before, 0
represents the stack not chosen in the exact solution, and 1 represents the chosen stack), are shown in Fig. 10.
According to the distribution of Fig. 10, we choose 0.5 as the dividing point for R1(s), so that the pruner can cut off at
least half of the candidate branches, and 1/3 as the dividing point for R2(s). As a consequence, the stacks can be divided
into 2 × 2 = 4 categories according to their ratio values against their respective dividing points. For each bay, we assign
the stacks to different stack categories, and add up the number of stacks belonging to different categories to generate the
representation of the bay. More specifically, an array of 1 × 4 named TEMP is introduced to represent the bay, with the first
element counting the number of stacks with R1(s) within [0,0.5] and R2(s) within [0,1/3], the second element corresponding
to R1(s) within [0,0.5] and R2(s) within (1/3,1], the third to R1(s) within (0.5,1] and R2(s) within [0,1/3], and the last one to
114 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Fig. 7. Illustration of the tie breaking principle.

Fig. 8. The average score of R1(s) for different datasets.

Fig. 9. The average score of R2(s) for different datasets.

R1(s) within (0.5,1] and R2(s) within (1/3,1]. For the example as shown in Fig. 7, its ratio values are R1(2 ) = 2/3, R1(4 ) =
1/3, R1(5 ) = 0/3 = 0 and R1(6 ) = 3/3 = 1 for the first indicator, and R2(2 ) = 1/2, R2(4 ) = 0, R2(5 ) = 1/2 and R2(6 ) = 0
for the second indicator. Therefore, TEMP=[1,1,1,1] is obtained for the particular case as shown in Fig. 11. Assume stack 5 is
chosen in the exact solution, the corresponding outcome will be labeled 2, as stack 5 falls into the second element in array
TEMP. In order to eliminate the impact of scalability, we will normalize the data again to adopt TEMP = [0.25, 0.25, 0.25,
0.25] rather than the original form TEMP = [1,1,1,1]. That is to say, the scalability issue is handled through the normalized
TEMP.
For each relocation, we will have a normalized TEMP and an element index corresponding to the optimal solution. We
record all these data from the running on the 30 datasets and generate 6693 records. In order to determine the proper
number of decision trees in each random forest, we varied the number of decision trees in the range {1,...,10 0 0} with incre-
mental interval set to 25. We observed that the estimation errors keep the same when the number of decision trees is larger
than 50, and therefore, we set this number to 100 in the subsequent experiments. By randomly dividing the data into two
sets according to the ratio of 7 to 3, we form the training set and testing set, respectively. The classifier is constructed by
Random Forest in RStudio software, and, as a consequence, the prediction accuracy is 80.5% for the testing set. The runtime
for generating the classifier by invoking the RStudio software is 6.5 min, which can be ignored when compared to the total
runtime of the test dataset. After the random forest based branch pruner is constructed, by injecting the TEMP of the current
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 115

Fig. 10. The average values of two ratios for the 30 datasets in terms of different stack types.

Fig. 11. The example of TEMP for the case as shown in Fig. 7.

bay, for example, TEMP = [0.25, 0.25, 0.25, 0.25] for the case as shown in Fig. 7, into the random forest, it will identify the
element number of the array to indicate which category of stacks will proceed to the branching process and which will be
pruned at this stage. For example, if the returned value is 3, then it means that only the third category of branches (in this
case, corresponding to stack 6) will be retained and all other categories of branches (in this case, corresponding to stacks 2,
4 and 5) will be discarded. Through this process, the random forest based branch pruner fulfills the pruning function.

6.3. Association rules based branch pruner

The branch pruner based on random forest has primarily encountered the following two shortcomings which will be
further illustrated in the numerical experiments: (i) the time for recording all the required data and the time to generate
the classifying results from the RStudio file is relatively long, which, to a certain extent, increases the running time of the
algorithm; (ii) the numerical experiments also show that the classification accuracy is not high when it applies to large-
scale problems. The possible reason is that the statistical indicators, i.e. R1(s) and R2(s), may not be representative enough,
and the normalization operation with the aim to eliminate the impact of scalability may remove some important features.
Due to these considerations, we attempt to find some more intuitive and more efficient rules to guide the pruning, that is,
the association rules based pruner in this subsection. The basic idea is that we divide the stacks into two categories as we
did in the Min-Max method presented in Section 4.3, then we retain partial stacks to represent each type with the aim to
reduce the search space, and finally the selection recommendations will be generated by the association rules based on the
remaining data.
We count, separately, the number of choosing Type-1 stacks (not necessarily with the highest stack priority) and choos-
ing Type-1 stacks with the highest stacks priority for relocation in the exact solution for each instance covered in the 30
datasets. The result is shown in Fig. 12, in which the right-side vertical coordinate “Ratio” refers to the ratio of the number
116 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Fig. 12. Type-1 stack selection statistics.

Table 1
The selected features.

Id Definition

N1 a binary variable, 1, if there are Type-1 stacks; 0, otherwise

N2 the priority-based ranking number of the stack chosen in the exact solution among
all the Type-2 stacks. For example, in Fig. 7, the priorities of Type-2 stacks 2, 4
and 6 are 7, 8 and 6, respectively, and, therefore, their priority-based ranking
numbers are 2, 3 and 1, respectively. If stack 2 is chosen in the exact solution,
then N2=2.
N3 the number of Type-2 stacks
N4 the difference between the number of Type-2 stacks and the priority ranking
number of the stack chosen in the exact solution, i.e. N3-N2
N5 the priority ranking number of the container to be relocated among all the
containers above the target one. For example, in Fig. 7, the two blocking
containers 9 and 2 have ranking number 2 and 1, respectively.
N6 a binary variable, 1, if the container to be relocated has the largest priority number
among all the containers above the target one; 0, otherwise
N7 the number of containers above the target one
N8 the difference between the stack height limit and the height of the stack
recommended by the Min-Max rules

of Type-1 stacks with highest priority selected in the exact solution to the number of Type-1 stacks selected in the cor-
responding exact solution for each dataset. From the result, we find that the ratio of each dataset is more than 90%, with
the average equal to 95.73%. Based on this observation, we, to a large extent, only need to keep the Type-1 stack with the
highest stack priority for further consideration and discard the rest to reduce the search space.
As for the selection of Type-2 stacks for further considerations, because statistical results are not that intuitive as in
the selection of Type-1 stacks, we will resort to the association learning method to generate the recommendation. The
features used in the learning method are shown in Table 1, most of which are extracted directly from those widely used
in the upper bound calculation methods reported in the published literature, including the stack priority ranking, stack
height, container ranking and other aspects. Roughly speaking, the 8 features can be categorized into 4 types, with the
first type (including N1) related to Type-1 stacks, the second type (including N2 to N4) associated with Type-2 stacks, the
third type (including N5-N7) pertaining to the stack where the target container is located, and the fourth type (including
N8) suggesting a preferred stack. Therefore, the 8 features are both representative and diversified, providing a relatively
complete representation for the bay.
We collect the data from all the operations related to the Type-2 stacks before the number of blocking containers equals
0 for the 30 datasets introduced in Section 6.1 as after the number equals 0 there will be no more relocation operations.
For each operation in the exact solution, we recorded all the values of the features as shown in Table 1 and accumulated
3498 records in total. In order to make the data records processable for association rules, we define multiple 0-1 variables
for possible values of each feature, and the number of variables of each feature depends on the chosen data. We need to
conduct statistical analysis for the chosen data first before determining the number of variables of each feature. For the
3498 data records used in this paper, the possible values and the number of variables of each feature obtained through
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 117

Table 2
The variables of each feature for the accumulated 3498 records.

Id Possible values Variables Number of variables (N)

N1 {0,1} hastypeone/notypeone 2
N2 {1,2,...,8} typetworankN 8
N3 {1,2,...,8} typetwonumN 8
N4 {0,1,...,5} rankdiffN 6
N5 {1,2,3} blockconrankN 3
N6 {0,1} ismaxupid/notmaxupid 2
N7 {1,2,3,4} blockconnumN 4
N8 {1,2,...,6} heightdiffN 6

Table 3
Rules list in the first execution of the Apriori algorithm.

LHS RHS support confidence

{} {blockconnum1} 0.6134934 0.6134934

{} {rankdiff0} 0.6409377 0.6409377
{} {blockconrank1} 0.7238422 0.7238422
{} {ismaxupid} 0.8624929 0.8624929
{blockconnum1} {blockconrank1} 0.6134934 1.0000000
{blockconrank1} {blockconnum1} 0.6134934 0.8475513
{blockconnum1} {ismaxupid} 0.6134934 1.0000000
{ismaxupid} {blockconnum1} 0.6134934 0.7113026
{rankdiff0} {ismaxupid} 0.5457404 0.8514719
{ismaxupid} {rankdiff0} 0.5457404 0.6327478
{blockconrank1} {ismaxupid} 0.6134934 0.8475513
{ismaxupid} {blockconrank1} 0.6134934 0.7113026
{blockconnum1, blockconrank1} {ismaxupid} 0.6134934 1.0000000
{blockconnum1, ismaxupid} {blockconrank1} 0.6134934 1.0000000
{blockconrank1, ismaxupid} {blockconnum1} 0.6134934 1.0000000

blockconnum corresponding to N7 in Table 1; blockconrank to N5; ismaxupid to N6;

rankdiff to N4.

statistics are listed in Table 2. Using RStudio software and applying Apriori association learning algorithm to the generated
data, we first extract the rules as shown in Table 3 with both support and confidence greater than 0.5.
As we want to use rules to guide branch pruning, what we are concerned about is which features can deduce the “rankd-
iff” feature, which means the difference between the number of Type-2 stacks and the ranking sequence of the selected
stack in the exact solution. The “rankdiff” features serve as the indicator for the selected stack, which, to a certain extent,
can eliminate the influence of the stack size. For example, “rankdiff0” means the difference is 0, meaning the selected stack
is the one with the lowest stack priority in Type-2 stacks. Each row in Table 3 represents a rule, in which the left-hand side
(LHS) and the right-hand side (RHS) both indicate the set of features. For example, the fifth row represents the rule that the
feature “blockconnum1” leads to the feature “blockconrank1”, namely, the chosen stack with the feature “blockconnum1”
has the feature “blockconrank1” at the same time, for which the support is 0.6134934 and the confidence is 1.0 0 0 0 0 0 0.
From the results, we observe two rules leading to the “rankdiff” feature. As the first one dominates the second one, we can
only consider the first one which means that the stack with the lowest stack priority needs to be kept.
Then we delete the “rankdiff0” data and run a further analysis on the remaining 1256 records. We get 80 rules by Apriori
association learning algorithm, whose support and confidence are both greater than or equal to 0.5. After checking the rules
leading to “rankdiff”, we obtain the results as shown in Table 4. As other rules as shown in Table 4 can be dominated by
the first rule, it convinces us to keep the stack with the second lowest priority and discard the rest. Considering that only
64 records are left after deleting “rankdiff1” data, we will stop there and select the two stacks with first two lowest stack
priorities to represent the Type-2 stacks. The total runtime for applying the Apriori association learning algorithm in the
RStudio software is less than 1 min. In conclusion, at most three stacks are reserved each time, which include the Type-1
stack with the highest stack priority and the Type-2 with first two lowest stack priority.

7. Numerical experiments

All experiments are coded in Java, and run on a personal computer with Intel(R) Core i7-7500U CPU @2.70GHz and
RAM-8GB.

7.1. Test on the calculation methods for the lower bounds

In this subsection, we compare the performance of the LB4 and the newly proposed ELB4. In order to compare the
performance of different calculation methods for the LB on large-scale problems, we choose, from Zhu et al. (2012), the
118 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Table 4
The association rules leading to rankdiff1.

LHS RHS support confidence

{} {rankdiff1} 0.949045 0.949045

{blockconnum1} {rankdiff1} 0.589968 0.969895
{blockconrank1} {rankdiff1} 0.675159 0.966933
{ismaxupid} {rankdiff1} 0.838376 0.950361
{heightdiff1} {rankdiff1} 0.880573 0.975309
{blockconnum1, blockconrank1} {rankdiff1} 0.589968 0.969895
{blockconnum1, ismaxupid} {rankdiff1} 0.589968 0.969895
{blockconnum1, heightdiff1} {rankdiff1} 0.566083 0.982044
{blockconrank1, ismaxupid} {rankdiff1} 0.589968 0.969895
{blockconrank1, heightdiff1} {rankdiff1} 0.643312 0.984166
{heightdiff1, ismaxupid} {rankdiff1} 0.784236 0.973320
{blockconnum1, blockconrank1, ismaxupid} {rankdiff1} 0.589968 0.969895
{blockconnum1, blockconrank1, heightdiff1} {rankdiff1} 0.566083 0.982044
{blockconnum1, heightdiff1, ismaxupid} {rankdiff1} 0.566083 0.982044
{blockconrank1, heightdiff1, ismaxupid} {rankdiff1} 0.566083 0.982044
{blockconnum1, blockconrank1, heightdiff1, ismaxupid} {rankdiff1} 0.566083 0.982044

blockconnum corresponding to N7 in Table 1; blockconrank to N5; ismaxupid to N6; heightdiff to N8;

rankdiff to N4.

Table 5
The comparison of the calculation methods for LBs.

Dataset LB4 ELB4 GAPR GAPT

Relocation1 Time1 (s) Relocation2 Time2 (s)

7-10-63 4605 1.111 4796 1.833 4.15% 64.99%

7-10-64 4706 0.766 4920 1.329 4.55% 73.50%
7-10-65 4892 1.128 5070 1.451 3.64% 28.63%
7-10-66 4914 0.891 5074 1.164 3.26% 30.64%
7-10-67 5086 0.984 5277 1.354 3.76% 37.60%
7-10-68 5073 0.797 5237 1.191 3.23% 49.44%
7-10-69 5186 0.906 5363 1.349 3.41% 48.90%
Average 4923 0.940 5105 1.382 3.71% 47.67%
∗ ∗
GAPR = (Relocation2-Relocation1)/Relocation1 100 GAPT = (Time2-Time1)/Time1 100

datasets of the form H − S − N (whose specific meaning is illustrated in Section 6.1), including 7-10-63, 7-10-64, 7-10-65,
7-10-66, 7-10-67, 7-10-68 and 7-10-69, and there are 100 instances for each dataset. As shown in Table 5, the quality of LBs
calculated by ELB4 has improved by 3.71% when compared to LB4, but the time has increased by 47.67%. It is observed that
the improvement is costly in terms of the increase in runtime. Considering that the LB needs to be calculated at each node
in the branch-and-bound tree, we decide to use LB4 instead for the subsequent experiments to save the total runtime.

7.2. Test on the calculation methods for the upper bounds

In this section, we will test the calculation methods for the upper bounds, that is, MLUB. As several variants of MLUB
are involved, to abate the confusion, we make the following definitions:

• Complete-VRH: the MLUB without adopting pruners in its look-ahead process, that is, the look-ahead method in con-
junction of the heuristic method proposed in Ting and Wu (2017);
• RF-VRH: the MLUB adopting the random forest pruner in its look-ahead process;
• AR-VRH: the MLUB adopting the association rules pruner in its look-ahead process.

7.2.1. The look-ahead method with and without the branch pruner
In order to compare the performance of different calculation methods for the LB on large-scale problems, we still resort
to dataset 7-10-63, 7-10-64, 7-10-65, 7-10-66, 7-10-67, 7-10-68 and 7-10-69 from Zhu et al. (2012). For the convenience
of comparison, we take a small look-ahead depth for each of the three methods, i.e., D = 1, and the results are shown in
Table 6.
For comparison, a calculation method based on the PR4 heuristic proposed in Zhu et al. (2012) (denoted by Zhu) and
the VRH proposed in Ting and Wu (2017) (denoted by VRH), are also included in this experiment. In Table 6, “Relocation”
represents the sum of the UB for the 100 instances obtained by the corresponding method for the considered dataset, and
“Time” represents the total running time.
From the perspective of the quality of the UB (evaluated by the value of “Relocation”), the VRH gets better UBs than the
method proposed in Zhu et al. (2012), and most of the UBs obtained by Complete-VRH, RF-VRH and AR-VRH are better than
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 119

Table 6
The comparison of the calculation methods for UBs (D = 1).

Dataset Zhu (Zhu et al., 2012) VRH (Ting and Wu, 2017) Complete-VRH RF-VRH AR-VRH

Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s)

7-10-63 5912 0.016 5748 0.031 5684 0.062 5739 0.546 5712 0.048
7-10-64 6111 0.016 5963 0.036 5910 0.093 5958 0.484 5933 0.094
7-10-65 6386 0.016 6277 0.032 6216 0.062 6259 0.500 6235 0.116
7-10-66 6307 0.016 6206 0.063 6161 0.046 6200 0.446 6171 0.080
7-10-67 6668 0.015 6516 0.031 6488 0.078 6517 0.406 6496 0.070
7-10-68 6706 0.031 6589 0.047 6576 0.047 6595 0.282 6579 0.046
7-10-69 6964 0.020 6811 0.030 6811 0.051 6811 0.071 6811 0.040

Fig. 13. The gap of UB under different datasets.

that of the VRH, which implies that the application of the look-ahead method can significantly improve the quality of the
UB. Among the methods, the UBs of the Complete-VRH are the tightest, which should be taken for granted, as it explores all
the possibilities of relocation. With pruning strategies adopted to improve the solution efficiency, the UB quality of RF-VRH
and AR-VRH is slightly worse than that of Complete-VRH. Paradoxically, RF-VRH has slightly worse results than VRH for the
datasets 7-10-67 and 7-10-68, and it is probably because the initial state of bay is relatively full and the look-ahead method
is conducted for only 1 layer.
In terms of the running time, the method in Zhu et al. (2012), which produces the worst UB, has the shortest running
time. Among the other four methods, using VRH alone has the shortest running time, followed by AR-VRH. The running time
of RF-VRH is the longest, since there are few available relocation stacks and branches under these datasets, and the random
forest based filter needs some operating time to solve and output prediction results, restricting the advantage of pruning. As
the number of alternative stacks increases and the look-ahead depth increases, the advantage of pruning gradually emerges,
which is further verified in Section 7.2.2.

7.2.2. Sensitivity analysis

As can be seen from the above analysis, how full the bay is (that is, the load intensity) and how deep the look-ahead
depth is have certain influences on the performance of the look-ahead method and different pruning mechanisms. In this
subsection, sensitivity analysis is further carried out for both parameters.
As the situations with D = 0 (i.e. using VRH without look-ahead) and D = 1 are considered in 7.2.1, D = 2, 3, 4, 5, 6, 7, 8, 9
are selected for numerical experiments in this subsection. Partial results for D = 2, 4, 6, 8 are shown in Table 7 as other
results display the same trend.
Under different load intensities, the quality of the UB obtained by the three look-ahead methods is firstly analysed. As
shown in Fig. 13 (where only D=1 and 6 are shown as other depths reveal the same trend), the y-coordinate represents
the UB obtained by VRH minus the UB obtained by the corresponding method, e.g., the y-coordinates of the points on line
“Complete-VRH” represent the UB obtained by VRH minus the UB obtained by Complete-VRH. It can be found that, with
the increase of the container number (i.e. the decrease of available stacks for each relocation), the gap between the UB
obtained by the three look-ahead methods and that by VRH decreases, which implies that, if there are more available stacks
for relocation (i.e., low load intensity of the bay), it is more likely to get tighter UBs by the look-ahead methods.
Under different look-ahead depths, the UBs obtained by Complete-VRH and AR-VRH are all tighter than that obtained by
VRH, and the quality of the UBs increases with the look-ahead depth. The RF-VRH also displays the same trend. However,
when the number of containers is large (i.e., the bay is relatively full), the UB quality of RF-VRH will be partially worse
than that of VRH and the solution performance will be unreliable. In terms of the running time, the efficiency of the three
look-ahead methods, i.e. Complete-VRH, RF-VRH and AR-VRH, decreases with the look-ahead depth, and the decreasing rate
of the Complete-VRH is obviously faster than the other two methods with pruning strategies.
 120
C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131
Table 7
The comparison of the calculation methods for UBs (D = 2, 4, 6, 8).

Dataset D=2 D=4

Complete-VRH RF-VRH AR-VRH Complete-VRH RF-VRH AR-VRH

Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s)

7-10-63 5647 0.210 5739 1.068 5690 0.080 5590 0.797 5737 1.499 5642 0.225
7-10-64 5854 0.187 5959 0.710 5901 0.640 5807 0.468 5951 1.234 5854 0.162
7-10-65 6165 0.110 6241 0.926 6187 0.107 6110 0.313 6229 0.985 6154 0.181
7-10-66 6130 0.078 6192 0.734 6140 0.111 6098 0.242 6182 0.989 6113 0.202
7-10-67 6461 0.078 6511 0.719 6471 0.106 6437 0.141 6524 0.734 6456 0.084
7-10-68 6565 0.083 6595 0.394 6572 0.060 6529 0.078 6583 0.594 6539 0.096
7-10-69 6797 0.047 6810 0.313 6797 0.047 6781 0.047 6806 0.375 6781 0.062
Dataset D=6 D=8
Complete-VRH RF-VRH AR-VRH Complete-VRH RF-VRH AR-VRH
Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s) Relocation Time(s)
7-10-63 5537 2.109 5719 1.828 5583 0.416 5493 10.357 5692 2.985 5546 1.090
7-10-64 5766 1.203 5947 1.484 5823 0.323 5736 4.628 5955 2.545 5799 0.682
7-10-65 6062 0.875 6234 1.561 6124 0.314 6021 3.346 6215 2.504 6081 0.624
7-10-66 6074 0.406 6177 1.250 6089 0.178 6035 1.421 6182 1.874 6062 0.414
7-10-67 6403 0.234 6530 0.812 6432 0.188 6374 0.530 6544 1.263 6411 0.302
7-10-68 6480 0.187 6580 0.750 6500 0.135 6447 0.360 6574 1.187 6469 0.266
7-10-69 6758 0.093 6808 0.578 6762 0.084 6741 0.156 6814 0.958 6745 0.127
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 121

Fig. 14. The comparison of running time under different depths.

As shown in Fig. 14, the y-coordinate represents the ratio of the difference of total solution time between RF-VRH/AR-
VRH and Complete-VRH under different depths to the solution time of Complete-VRH. More specifically, when the running
time of RF-VRH/AR-VRH is less than Complete-VRH, the y-coordinate is positive, otherwise it is negative, which means, when
this value is positive, the branch pruner of the corresponding method has an improved effect on the solution efficiency, and
the larger the value, the more significant the improvement. It shows that the improvements of solution efficiency of random
forest based and association rules based filters for the look-ahead VRH both increase with D. When D reaches a certain value,
the y-coordinate reaches positive values, and pruning through the two filters can improve the solution efficiency, reflective
of the advantage of pruning.
In summary of this section, using the look-ahead method to improve VRH can effectively enhance the quality of the UB,
and further pruning by random forest and association rules would slightly decrease the quality of the UB compared with the
complete look-ahead UB, but still ensure that the UB quality is higher than that of VRH for most cases. Considering the time
consumed by the construction of pruning filters, the solution efficiency is lower than Complete-VRH when the look-ahead
depth is small or there are only a few available stacks for relocation. However, with the increase of D, and the decrease of
the load intensity of the bay, the improvement of pruning filters on the solution efficiency emerges. Among the two filters,
the experimental results show that the association rules based filter is more reliable than the random forest based filter in
both UB quality and solution efficiency. In addition, it should be noted that, though the quality of UB improves with the
increase of D, running time also increases, and the increase of running time will be further amplified when the obtained UB
is applied to the calculation of each node by the IB&B. Therefore, there is a trade-off between the UB quality and solution
efficiency in the searching framework of the IB&B, which will be further verified in Sections 7.3 and 7.4.

7.3. Test on the exact method IB&B

7.3.1. Test on performance

We compared the IB&B with the CB&B based on the datasets adopted in Zhu et al. (2012) in terms of the number of
nodes explored to get optimal solutions as well as the running time, and, to avoid the impact of the quality of UB, the
method based on the PR4 heuristic (without look-ahead method and branch pruner) in Zhu et al. (2012) is applied to the
calculation of upper bounds. As each dataset contains 100 instances, we report the average results in Table 8, from which
we see that the IB&B performs better than the CB&B in terms of both performance indicators.

7.3.2. Sensitivity analysis

As the look-ahead depth can affect the number of nodes explored as well as the search time, we conduct the sensitivity
analysis to observe the impact of the look-ahead depth D on the two performance indicators under the UB-priority branching
strategy. The experimental results are shown in Table 9, from which we see that the running time increases when the look-
ahead depth D increases, and the number of nodes explored has the trend to decrease when D increases, but the reduction
is not evident when compared with the total number. We, therefore, choose D = 1 in the subsequent experiment regarding
the comparison between the UB-priority and LB-priority branching strategies.

7.3.3. Branching strategies

We compare the UB-priority with the LB-priority (using LB4) branching strategies based on the datasets adopted in
Zhu et al. (2012) in terms of the number of nodes explored as well as the running time for generating the optimal solution.
The results are shown in Table 10, from which we see that the UB-priority branching strategy can effectively reduce the
total number of nodes explored, and the runtime of the UB-priority branching strategy is relatively shorter when compared
with the LB-priority branching strategy, except for RF-VRH, which is because of the long time consumed to solve and output
prediction results by the random forest based filter. It is noted that although the UB-priority branching strategy outperforms
122 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Table 8
Test on the IB&B (adopting PR4 heuristic (Zhu et al., 2012) to calculate UBs).

Dataset Optimal solution CB&B IB&B GAPN GAPT

Node number1 Time1(s) Node number2 Time2(s)

5-6-25 1674 108042 25.935 90648 10.814 −16.10% −58.30%

5-6-26 1859 219609 18.481 187820 15.544 −14.48% −15.89%
5-6-27 1888 108099 11.124 83777 7.052 −22.50% −36.61%
5-6-28 1969 120795 12.806 56709 5.823 −53.05% −54.53%
5-6-29 2113 498046 48.592 249971 22.238 −49.81% −54.24%
5-7-30 2011 756815 74.103 618241 60.160 −18.31% −18.82%
5-7-31 2186 1908334 185.981 1016409 111.438 −46.74% −40.08%
5-7-32 2268 1067219 108.531 640883 62.371 −39.95% −42.53%
5-7-33 2340 1557448 122.869 1021655 83.952 −34.40% −31.67%
5-7-34 2454 768616 83.128 656812 72.836 −14.55% −12.38%
6-6-30 2295 5295414 617.146 5241436 601.926 −1.02% −2.47%
6-6-31 2476 7505345 982.455 6251112 919.500 −16.71% −6.41%
6-6-32 2682 20614627 2924.011 12945238 1821.959 −37.20% −37.69%
6-6-33 2873 39792068 5909.837 12472842 1799.120 −68.65% −69.56%
6-6-34 2920 33474780 3495.543 23438601 2422.007 −29.98% −30.71%
∗ ∗
GAPN = (Node number2-Node number1)/Node number1 100 GAPT = (Time2-Time1)/Time1 100

Table 9
Sensitivity analysis of the look-ahead depth.

Dataset Optimal solution Complete-VRH

D=1 D=2 D=3

Node number Time(s) Node number Time(s) Node number Times(s)

5-6-25 1674 79381 10.427 80666 18.053 80471 35.486

5-6-26 1859 188418 20.686 192124 38.321 187551 78.233
5-6-27 1888 82674 3.781 81488 17.092 81489 33.772
5-6-28 1969 42290 5.373 42271 7.839 42170 13.701
5-6-29 2113 249221 28.771 249021 57.281 249227 126.971
Dataset Optimal solution RF-VRH
D=1 D=2 D=3
Node number Time(s) Node number Time(s) Node number Times(s)
5-6-25 1674 93602 98.444 95672 162.354 101240 273.520
5-6-26 1859 189354 198.148 190087 290.504 187843 450.690
5-6-27 1888 82791 81.043 82759 201.705 83269 345.579
5-6-28 1969 46889 41.883 46929 65.770 47025 101.363
5-6-29 2113 249405 222.192 249999 562.925 250036 711.764
Dataset Optimal solution AR-VRH
D=1 D=2 D=3
Node number Time(s) Node number Time(s) Node number Times(s)
5-6-25 1674 78469 9.725 79754 12.266 79784 18.034
5-6-26 1859 188864 15.741 192595 24.025 188022 36.596
5-6-27 1888 82498 8.328 81486 10.269 81486 16.410
5-6-28 1969 42267 5.191 42266 6.834 42276 7.648
5-6-29 2113 249222 21.947 248928 30.760 249134 50.099

the LB-priority branching strategy and fits our test data very well, there is no guarantee that the former strategy outperforms
the latter one for general cases. We believe that the relationship between the two strategies deserves deep investigation.

7.4. Test on the heuristic method MLBS

7.4.1. Solution quality of the heuristic method

We test our method on the datasets adopted in Caserta et al. (2011) with the look-ahead depth D = 3 and the beam
width W = 5. The results are shown in Table 11 where the CM represents the method proposed by Caserta et al. (2011).
It can be seen from the table that Complete-VRH can generate better results than VRH, and using the branch pruners can
improve the efficiency of the algorithm but at the slight cost of solution quality compared with Complete-VRH. In fact, when
the look-ahead depth is increased to D = 4, all the results generated by AR-VRH become better than those by VRH as shown
in Table 11.
The runtime of the MLBS using the three aforementioned methods for calculating UBs is shown in Table 12, from which
we see that, with the increase of problem sizes, the computing time of the method using Complete-VRH increases rapidly,
and the advantage of using the branch pruner becomes more evident. From Tables 11 and 12, we can see that using
Complete-VRH in the MLBS can generate better results than using VRH at the cost of runtime, and incorporating branch
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131
Table 10
Comparsion between LB- and UB-priority branching strategies.

Dataset Optimal solution LB-priority UB-priority

VRH (Ting and Wu, 2017) Complete-VRH RF-VRH AR-VRH

Node number Time(s) Node number Time(s) Node number Time(s) Node number Time(s) Node number Time(s)

5-6-25 1674 90648 23.623 81501 4.559 79381 10.427 93602 98.444 78469 9.725
5-6-26 1859 187820 44.158 189863 7.535 188418 20.686 189354 198.148 188864 15.741
5-6-27 1888 83777 22.591 82583 9.506 82674 3.781 82791 81.043 82498 8.328
5-6-28 1969 56709 16.461 46532 3.399 42290 5.373 46889 41.883 42267 5.191
5-6-29 2113 249971 62.333 249297 10.003 249221 28.771 249405 222.192 249222 21.947
5-7-30 2011 618241 169.187 605804 31.269 599098 92.116 608706 543.962 599506 62.769
5-7-31 2186 1016409 270.020 967282 47.085 966409 222.438 988183 1115.593 966331 123.099
5-7-32 2268 640883 154.024 629694 25.843 635582 182.848 628137 869.474 632391 104.292
5-7-33 2340 1021655 239.517 1014436 35.957 1014732 168.512 1014860 1593.988 1014738 80.521
5-7-34 2454 656812 172.892 612972 29.474 619442 102.265 615714 539.604 620843 93.604

123
 124
C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131
Table 11
Average relocation number of different UBs on CM dataset.

Bay info Average relocation number

Stack height Stack number Container number CM VRH D=3 D=4

(Caserta et al., 2011) (Ting and Wu, 2017) Complete-VRH RF-VRH AR-VRH Complete-VRH RF-VRH AR-VRH

5 4 20 16.6 15.45 15.425 15.425 15.425 15.4 15.425 15.425

5 5 25 18.8 19 18.95 18.925 18.95 18.725 18.925 18.9
5 6 30 22.1 22.22 22.25 22.275 22.3 22.2 22.25 22.2
5 7 35 25.8 24.37 24.35 24.425 24.37 24.35 24.425 24.37
5 8 40 30.1 27.8 27.75 27.775 27.775 27.725 27.775 27.775
5 9 45 33.1 30.52 30.52 30.55 30.55 30.52 30.55 30.52
5 10 50 36.4 33.4 33.375 33.45 33.375 33.25 33.4 33.325
6 6 36 32.4 31.2 31.2 31.375 31.25 31.15 31.375 31.175
6 10 60 49.5 46.15 46.025 46 46.075 45.925 46 45.975
10 6 60 102 79.97 78.5 81.475 78.35 78.375 81.475 78.65
10 10 100 128.3 113.25 112.325 115.575 113.175 111.9 114.85 112.425

The results in column “CM (Caserta et al., 2011)” are copied directly from the referred paper.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131
Table 12
Average running time of different UBs when D = 3 on CM dataset.

Bay info Average runing time(s)

Stack height Stack number Container number CM (Caserta et al., 2011) VRH (Ting and Wu, 2017) Complete-VRH RF-VRH AR-VRH

5 4 20 0.5 0.011 0.016 0.084 0.012

5 5 25 0.8 0.018 0.049 0.215 0.027
5 6 30 0.8 0.031 0.137 0.305 0.056
5 7 35 1.43 0.041 0.414 0.431 0.077
5 8 40 1.46 0.045 0.793 0.631 0.076
5 9 45 1.41 0.056 1.947 0.839 0.101
5 10 50 1.87 0.098 3.806 1.272 0.156
6 6 36 1.74 0.081 0.469 0.702 0.130
6 10 60 1.95 0.230 16.047 5.552 0.572
10 6 60 4.73 0.914 4.838 4.773 1.597
10 10 100 6.34 4.946 290.335 66.570 9.920

The results in column “CM (Caserta et al., 2011)” are copied directly from the referred paper.

125
 126
C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131
Table 13
Test results of different UBs when D = 3 on the dataset in Lee and Lee (2010).

Case LL (Lee and Lee, 2010) VRH (Ting and Wu, 2017) Complete-VRH RF-VRH AR-VRH

Number Time(s) Number Time(s) Number Time(s) Number Time(s) Number Time(s)

R011606_0070_001 48 6304.280 37 0.012 37 0.108 37 0.347 37 0.010

R011606_0070_002 47 11081.030 38 0.405 38 5.750 38 2.395 38 0.523
R011606_0070_003 40 5501.920 38 0.067 38 7.552 38 2.395 38 0.226
R011606_0070_004 88 9026.420 45 0.515 45 21.741 45 4.153 45 1.116
R011606_0070_005 54 9107.970 40 0.219 40 1.428 40 1.050 40 0.303
R011608_0090_001 100 13268.670 61 0.895 61 357.731 61 19.244 61 1.985
R011608_0090_002 101 11134.630 61 0.088 61 0.569 61 0.508 61 0.168
R011608_0090_003 126 21583.130 65 2.316 65 323.351 65 10.451 65 1.836
R011608_0090_004 88 7042.380 61 0.416 61 5.945 61 2.508 61 0.478
R011608_0090_005 92 13738.000 60 1.160 59 69.677 59 4.370 59 2.298
U011606_0070_001 55 17326.310 55 0.015 55 1.601 55 1.571 55 0.066
U011606_0070_002 60 11243.400 58 0.012 58 1.962 58 1.015 58 0.068
U011608_0090_001 85 21586.810 76 0.054 76 0.594 76 0.362 76 0.038
U011608_0090_002 90 8021.310 78 0.023 78 3.098 78 1.207 78 0.045

The results in column “LL (Lee and Lee, 2010)” are copied directly from the referred paper.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 127

Table 14
Sensitivity anlysis of the look-ahead depth.

Dataset Complete-VRH

D=1 D=2 D=3

Average relocation Time(s) Average relocation Time(s) Average relocation Time(s)

5-4 15.475 0.013 15.45 0.022 15.425 0.016

5-5 18.95 0.031 18.925 0.049 18.95 0.049
5-6 22.225 0.071 22.275 0.092 22.25 0.137
5-7 24.425 0.081 24.35 0.173 24.35 0.414
5-8 27.775 0.127 27.75 0.312 27.75 0.793
5-9 30.575 0.176 30.6 0.522 30.525 1.947
5-10 33.425 0.246 33.35 0.878 33.375 3.806
6-6 31.225 0.132 31.075 0.225 31.2 0.469
6-10 46.125 1.092 45.975 4.016 46.025 16.047
10-6 79.15 1.565 78.875 3.128 78.5 4.838
10-10 113.525 16.983 112.575 64.088 112.325 290.335
Dataset RF-VRH
D=1 D=2 D=3
Average relocation Time(s) Average relocation Time(s) Average relocation Time(s)
5-4 15.475 0.052 15.475 0.065 15.45 0.084
5-5 18.95 0.078 18.975 0.098 19.025 0.215
5-6 22.325 0.114 22.375 0.156 22.325 0.305
5-7 24.5 0.155 24.35 0.199 24.35 0.431
5-8 27.8 0.163 27.775 0.259 27.775 0.631
5-9 30.52 0.199 30.52 0.330 30.55 0.839
5-10 33.425 0.252 33.4 0.438 33.425 1.272
6-6 31.3 0.247 31.25 0.342 31.425 0.702
6-10 46.175 0.944 45.975 1.524 46 5.552
10-6 79.925 1.742 80.1 2.169 79.875 4.773
10-10 113.875 9.164 113.925 18.035 114.125 66.570
Dataset AR-VRH
D=1 D=2 D=3
Average relocation Time(s) Average relocation Time(s) Average relocation Time(s)
5-4 15.45 0.010 15.425 0.012 15.425 0.012
5-5 18.95 0.022 18.95 0.022 18.9 0.027
5-6 22.275 0.034 22.3 0.040 22.2 0.056
5-7 24.35 0.044 24.37 0.077 24.37 0.077
5-8 27.775 0.053 27.775 0.062 27.775 0.076
5-9 30.575 0.080 30.55 0.085 30.52 0.101
5-10 33.375 0.105 33.375 0.111 33.325 0.156
6-6 31.225 0.093 31.25 0.113 31.175 0.130
6-10 46.1 0.285 46.075 0.418 45.975 0.572
10-6 79.25 1.174 78.35 1.314 78.65 1.597
10-10 113.1 6.350 113.175 7.368 112.425 9.920

Table 15
Sensitivity analysis of beam width when D = 2.

Beam width Complete-VRH RF-VRH AR-VRH

Average relocation Time(s) Average relocation Time(s) Average relocation Time(s)

5 33.35 0.878 33.4 0.438 33.375 0.111

10 33.325 1.300 33.4 0.888 33.35 0.186
20 33.325 2.110 33.35 1.180 33.3 0.397

pruners into the UB calculation methods can save the runtime significantly. To make the conclusion more convincing, we
also test the algorithm on the datasets adopted in Lee and Lee (2010) with the look-ahead depth D = 3 and the beam width
W = 5. The results are as shown in Table 13.

7.4.2. Sensitivity analysis

As the look-ahead depth D and beam width W may greatly affect the UB quality as well as running time, we will conduct
the sensitivity analysis on the two factors. To analyse the influence of the look-ahead depth, keeping the beam width W = 5
unchanged and varying the values of D for Complete-VRH, RF-VRH and AR-VRH, we obtain the result as shown in Table 14.
From these sensitivity analyses, we can see that, as the look-ahead depth increases, the solution quality is not notably
improved. Therefore, in view of the requirement of solution efficiency, the optimal depth D = 1 is adoptable.
As for the influence of beam width, setting the look-ahead depth D to 2 and varying the beam width W for the MLBS,
we obtain the result as shown in Table 15 for the dataset 5–10 of Caserta et al. (2011). It can be seen from the result that,
128 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Table 16
Effect of deleting duplicate nodes.

Bay info Average relocation number

Stack height Stack number Container number No deletion Deletion

5 4 20 15.45 15.45
5 5 25 18.925 18.925
5 6 30 22.275 22.275
5 7 35 24.375 24.35
5 8 40 27.75 27.75
5 9 45 30.6 30.6
5 10 50 33.35 33.35
6 6 36 31.075 31.075
6 10 60 46.075 45.975
10 6 60 79 78.875
10 10 100 112.65 112.575

Table 17
Effect of using the Min-Max rules to break ties.

Bay info Random selection Min-Max selection

Stack height Stack number Container number Range Number Number

5 4 20 2 15.463 15.45
5 5 25 3 18.95 18.925
5 6 30 7 22.25 22.275
5 7 35 6 24.43 24.35
5 8 40 4 27.775 27.75
5 9 45 3 30.598 30.6
5 10 50 6 33.423 33.35
6 6 36 8 31.155 31.075
6 10 60 5 46.033 45.975
10 6 60 18 78.975 78.875
10 10 100 18 112.82 112.575

with the increase of beam width, the quality of the solution is improved, but the improvement is not significant. To save
the running time, we will set the beam width to a relatively small number.

7.4.3. Screening techniques analysis

We compare the average relocation number of algorithms with and without deleting duplicate nodes, leading to the
result shown in Table 16, from which we can see that solution quality can be improved by deleting duplicate nodes.
In Ting and Wu (2017), when the upper bounds for some nodes are the same, a random selection method is used to
break ties. We conduct numerical experiments to invoke the random selection on the CM datasets 10 times to get the
average relocation number, as well as the results under the same conditions except using the Min-Max rules to break ties,
and the comparison result is shown in Table 17. As can be seen from the table, the average number of relocation using the
Min-Max method to break ties is significantly lower than that of using random selection method for most cases. The range
in the table represents the difference between the best solution and the worst solution in the 10 random experiments and
we find that the performance of the random selection method varies greatly for different problem sizes. With the increase
of problem scale, the range increases significantly, implying that the random selection to break tie is not a reliable method.

8. Conclusions

For an important restricted container relocation problem in container terminals, this paper proposes a new calcula-
tion method for upper bounds, which incorporates the look-ahead method and machine learning to facilitate the conver-
gence. Furthermore, the machine learning-driven upper bounds are injected into an exact branch-and-bound algorithm and
a beam search method to enhance their respective performance. The numerical experiments show that the proposed ma-
chine learning-driven UB, the exact algorithm and the heuristic display the better performance than the state-of-the-art
methods reported in the literature based on the benchmark test instances.
Further research can be done in the following directions: (i) considering the relocation across bays and combining the
operation time of yard cranes; (ii) extending the framework of the algorithms to the unrestricted container relocation prob-
lem; and (iii) considering the container relocation problem with duplicate priorities.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 129

Acknowledgments

This work was supported by the National Key R&D Program of China under grant No. 2018AAA0101705, and the National
Natural Science Foundation of China under grant No. 71872092.

Appendix A. The virtual relocation heuristic (VRH)

One of the drawback of the Min-Max method is that it only considers one container at a time. To overcome this draw-
back, Ting and Wu (2017) proposed a virtual relocation heuristic (VRH) to calculate the upper bound, which takes into con-
sideration all blocking containers above the target one simultaneously. Here we will give an example to specifically present
the two stages of this algorithm, as shown in Fig. 15.
In the first stage of the VRH, we initially rank the blocking containers above the target container in a descending order
by the priority number and obtain the container sequence to be relocated Q = {8, 7, 6, 5}, then determine the well-relocated
stacks for the blocking containers sequentially according to the Min-Max condition. As shown in Fig. 15(a), stack 3 satisfies
the well-relocated condition for container 8, then the container will be virtually relocated to stack 3. It should be noted that
when relocating containers by the Min-Max condition, the first-in-last-out (FILO) principle of containers should be taken
into account, i.e., containers originally placed in the lower layer should be above the upper-layer containers after relocation.
This can be realized by recording the original minimum layer number of all the accommodated containers in a particular
stack s, denoted by tmin (s). As shown in Fig. 15(a), with container 8 relocated to stack 3, the tmin (s) of the stack s is updated
from the initial stack height limit 5 to the original layer number of container 8, which is 4. Assume the next container c
needs to be relocated to this stack, and then t(c) < tmin (s) must be satisfied, i.e., the original layer number of container c
is smaller than the minimum layer number of all the accommodated containers in this stack. As in Fig. 15(b), for the next
container 7, as stack 3 satisfies the well-relocated principle, container 7 can be relocated to stack 3, and its corresponding
tmin (s) value is updated to 2 accordingly. For the following container 6 and container 5, though stack 3 meets the well-
relocated condition, the condition t(c) < tmin (s) is not satisfied, which violates the FILO principle, implying that these two
containers will not be processed in this stage. To summarize the first stage of VRH, all blocking containers above the target
container is ranked by the priority number in a descending order, and then according to the well-located condition of the
Min-Max method and the FILO principle, the well-located stack for each blocking container is searched sequentially. If no
stack satisfies the both conditions for one container, this container will be put aside and relocated in the next stage.
In the second stage of the VRH, the remaining containers in the first stage will be ranked according to the priority
number in an ascending order and the sequence Q={5, 6} is obtained, which means container 5 should be considered first
and container 6 follows. In this case, if container 5 is relocated to a stack that has accommodated other containers before, it
is bound to be inserted into the middle of those containers that have been virtually relocated in the first stage. As shown in
Fig. 16(a), if container 5 is relocated to stack 3, it needs to be inserted between containers 8 and 9. Here we can determine
the target stack based on virtual relocation index (VRI) of the stack s as follows.

min{lsc − p(c ), p(c ) − usc } if U ≤ 1
V RIs = (2)
−N + ( p(c ) − usc ) otherwise

where U denotes the number of blocking containers above the inserted container c. For container 5 in Fig. 16(a), as there
are containers 7 and 8 above it, both of which are blocking containers relative to container 5, U=2. N is the initial total
number of containers. In terms of stack s, lsc is the highest priority (i.e. the minimum priority number) of all containers
below the inserted container c, and usc is the highest priority (i.e. the minimum priority number) of all containers above the
inserted container c. For Fig. 16(a), l35 = min{9} = 9, and u35 = min{7, 8} = 7. Therefore, V RI3 = −9 + 5 − 7 = −11 is obtained
for stack 3, V RI2 = min{2 − 55 − 0} = −3 is obtained for stack 2, and then the stack with the highest VRI will be selected for

Fig. 15. The first stage of the VRH (Ting and Wu, 2017)
130 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131

Fig. 16. The second stage of the VRH (Ting and Wu, 2017)

relocation, which means container 5 will be relocated to stack 2. In the same way, we can get the recommended relocation
stack for container 6 is stack 3. Repeat the above operation until all containers are extracted, and then a relatively tight UB
is obtained.

References

Allahviranloo, M., Recker, W., 2013. Daily activity pattern recognition by using support vector machines with multiple classes. Transp. Res. Part B-Methodol.
58, 16–43.
Andres, M., Nair, R., 2017. A predictive-control framework to address bus bunching. Transp. Res. Part B-Methodol. 104, 123–148.
Arentze, T.A., Timmermans, H.J.P., 2004. A learning-based transportation oriented simulation system. Transp. Res. Part B-Methodol. 38 (7), 613–633.
Bacci, T., Mattia, S., Ventura, P., 2019. The bounded beam search algorithm for the block relocation problem. Comput. Oper. Res. 103, 252–264.
Bell, M.G.H., Liu, X., Angeloudis, P., Fonzone, A., Hosseinloo, S.H., 2011. A frequency-based maritime container assignment model. Transp. Res. Part
B-Methodol. 45 (8), 1152–1161.
Bell, M.G.H., Liu, X., Rioult, J., Angeloudis, P., 2013. A cost-based maritime container assignment model. Transp. Res. Part B-Methodol. 58, 58–70.
Bortfeldt, A., Wäscher, G., 2013. Constraints in container loading c a state-of-the-art review. Eur. J. Oper. Res. 229 (1), 1–20.
Carlo, H.J., Vis, I.F.A., Roodbergen, K.J., 2014. Transport operations in container terminals: literature overview, trends, research directions and classification
scheme. Eur. J. Oper. Res. 236 (1), 1–13.
Caserta, M., Schwarze, S., Voß, S., 2012. A mathematical formulation and complexity considerations for the blocks relocation problem. Eur. J. Oper. Res. 219
(1), 96–104.
Caserta, M., Voß, S., Sniedovich, M., 2011. Applying the corridor method to a blocks relocation problem. OR Spectrum 33 (4), 915–929.
Dong, C.J., Shao, C.F., Clarke, D.B., Nambisan, S.S., 2018. An innovative approach for traffic crash estimation and prediction on accommodating unobserved
heterogeneities. Transp. Res. Part B-Methodol. 118, 407–428.
Forster, F., Bortfeldt, A., 2012. A tree search procedure for the container relocation problem. Comput. Oper. Res. 39 (2), 299–309.
Galle, V., Barnhart, C., Jaillet, P., 2018. A new binary formulation of the restricted container relocation problem based on a binary encoding of configurations.
Eur. J. Oper. Res. 267 (2), 467–477.
deGharehgozli, A.H., Laporte, G., Yu, Y., Koster, R., 2014. Scheduling twin yard cranes in a container block. Trans. Sci 49 (3), 686–705.
deGharehgozli, A.H., Yu, Y., Koster, R., Udding, J.T., 2014. An exact method for scheduling a yard crane. Eur. J. Oper. Res 235 (2), 431–447.
Hofleitner, A., Herring, R., Bayen, A., 2012. Arterial travel time forecast with streaming data: a hybrid approach of flow modeling and machine learning.
Transp. Res. Part B-Methodol. 46 (9), 1097–1122.
Hottung, A., Tanaka, S., Tierney, K., 2020. Deep learning assisted heuristic tree search for the container pre-marshalling problem. Comput. Oper. Res. 113,
1–13. 104781
Iris, C., Christensen, J., Pacino, D., Ropke, S., 2018. Flexible ship loading problem with transfer vehicle assignment and scheduling. Transp. Res. Part
B-Methodol.. 111, 113–134.
Iris, C., Lam, J.S.L., 2019. Recoverable robustness in weekly berth and quay crane planning. Transp. Res. Part B-Methodol. 122, 365–389.
Jiang, X.J., Jin, J.G., 2017. A branch-and-price method for integrated yard crane deployment and container allocation in transshipment yards. Transp. Res.
Part B-Methodol. 98, 62–75.
Jin, B., 2020. On the integer programming formulation for the relaxed restricted container relocation problem. Eur. J. Oper. Res. 281 (2). 475–C482
Jin, B., Zhu, W., Lim, A., 2015. Solving the container relocation problem by an improved greedy look-ahead heuristic. Eur. J. Oper. Res. 240 (3), 837–847.
Jovanovic, R., Tuba, M., Voß, S., 2019. An efficient ant colony optimization algorithm for the blocks relocation problem. Eur. J. Oper. Res 274 (1), 78–90.
Jovanovic, R., Voß, S., 2014. A chain heuristic for the blocks relocation problem. Comput. Ind. Eng 75, 79–86.
Kim, K.H., Hong, G.P., 2006. A heuristic rule for relocating blocks. Comput. Oper. Res. 33 (4), 940–954.
Kim, K.H., Kim, H.B., 2002. The optimal sizing of the storage space and handling facilities for import containers. Transp. Res. Part B-Methodol. 36 (9),
821–835.
Kim, K.H., Park, Y.M., Ryu, K.R., 20 0 0. Deriving decision rules to locate export containers in container yards. Eur. J. Oper. 124 (1), 89–101.
Kouvelas, A., Saeedmanesh, M., Geroliminis, N., 2017. Enhancing model-based feedback perimeter control with data-driven online adaptive optimization.
Transp. Res. Part B-Methodol. 96, 26–45.
Lee, C.Y., Song, D.P., 2017. Ocean container transport in global supply chains: overview and research opportunities. Transp. Res. Part B-Methodol. 95,
442–474.
Lee, Y., Chao, S.L., 2009. A neighborhood search heuristic for pre-marshalling export containers. Eur. J. Oper. Res. 196 (2), 468–475.
Lee, Y., Lee, Y.J., 2010. A heuristic for retrieving containers from a yard. Comput. Oper. Res. 37 (6), 1139–1147.
Lehnfeld, J., Knust, S., 2014. Loading, unloading and premarshalling of stacks in storage areas: Survey and classification. Eur. J. Oper. Res. 239 (2), 297–312.
Liu, M., Lee, C.Y., Zhang, Z., Chu, C., 2016. Bi-objective optimization for the container terminal integrated planning. Transp. Res. Part B-Methodol. 93,
720–749.
Morrison, D.R., Jacobson, S.H., Sauppe, J.J., Sewell, E.C., 2016. Branch-and-bound algorithms: a survey of recent advances in searching, branching, and prun-
ing. Discret. Optim. 19, 79–102.
Siripirote, T., Sumalee, A., Ho, H.W., Lam, W.H.K., 2015. Statistical approach for activity-based model calibration based on plate scanning and traffic counts
data. Transp. Res. Part B-Methodol. 78, 280–300.
 C. Zhang, H. Guan and Y. Yuan et al. / Transportation Research Part B 139 (2020) 102–131 131

Sopasakis, A., Katsoulakis, M.A., 2016. Information metrics for improved traffic model fidelity through sensitivity analysis and data assimilation. Transp. Res.
Part B-Methodol. 86, 1–18.
Steenken, D., Voß, S., Stahlbock, R., 2004. Container terminal operation and operations research c a classification and literature review. OR Spectrum 26 (1),
3–49.
Tanaka, S., Takii, K., 2016. A faster branch-and-bound algorithm for the block relocation problem. IEEE Trans. Autom. Sci. Eng. 13 (1), 181–190.
Ting, C.J., Wu, K.C., 2017. Optimizing container relocation operations at container yards with beam search. Transp. Res. Part E-Logist. Transp. Rev. 103, 17–31.
Virgile, G., Cynthia, B., Patrick, J., 2018. Yard crane scheduling for container storage, retrieval, and relocation. Eur. J. Oper. Res 271 (1), 288–316.
Wang, S., Liu, Z., Bell, M.G.H., 2015. Profit-based maritime container assignment models for liner shipping networks. Transp. Res. Part B-Methodol. 72,
59–76.
Wang, T.S., Wang, X.C., Meng, Q., 2018. Joint berth allocation and quay crane assignment under different carbon taxation policies. Transp. Res. Part
B-Methodol. 117, 18–36.
Wei, D., Liu, H., 2013. Analysis of asymmetric driving behavior using a self-learning approach. Transp. Res. Part B-Methodol. 47, 1–14.
Yin, K., Wang, W., Bruce Wang, X.B., Adams, T.M., 2015. Link travel time inference using entry/exit information of trips on a network. Transp. Res. Part
B-Methodol. 80, 303–321.
Zehendner, E., Feillet, D., Jaillet, P., 2016. An algorithm with performance guarantee for the online container relocation problem. Eur. J. Oper. Res. 259 (1),
48–62.
Zhang, C., Chen, W., Shi, L., Zheng, L., 2010. A note on deriving decision rules to locate export containers in container yards. Eur. J. Oper. 205 (2), 483–485.
Zhang, C., Liu, J., Wan, Y., Murty, K.G., Linn, R.J., 2003. Storage space allocation in container terminals. Transp. Res. Part B-Methodol. 37 (10), 883–903.
Zhang, C., Wan, Y.W., Liu, J., Linn, R.J., 2002. Dynamic crane deployment in container storage yards. Transp. Res. Part B-Methodol. 36 (6), 537–555.
Zhang, C., Wu, T., Kim, K.H., Miao, L., 2014. Conservative allocation models for outbound containers in container terminals. Eur. J. Oper. Res. 238 (1), 155–165.
Zhang, C., Wu, T., Zhong, M., Zheng, L., Miao, L., 2014. Location assignment for outbound containers with adjusted weight proportion. Comput. Oper. Res
52, 84–93.
Zhen, L., 2016. Modeling of yard congestion and optimization of yard template in container ports. Transp. Transp. Res. Part B-Methodol. 90, 83–104.
Zhen, L., Liang, Z., Zhuge, D., Lee, L.H., Chew, E.P., 2017. Daily berth planning in a tidal port with channel flow control. Transp. Res. Part B-Methodol. 106,
193–217.
VanZheng, F., Zuylen, H., 2013. Urban link travel time estimation based on sparse probe vehicle data. Transp. Res. Part C-Emerg. Technol. 31, 145–157.
Zheng, F.F., Man, X.Y., Chu, F., Liu, M., Chu, C.B., 2018. A two-stage stochastic programming for single yard crane scheduling with uncertain release times of
retrieval tasks. Int. J. Prod. Res 1–16.
Zhu, W., Qin, H., Lim, A., Zhang, H., 2012. Iterative deepening a∗ algorithms for the container relocation problem. IEEE Trans. Autom. Sci. Eng. 9 (4), 710–722.