You are on page 1of 14

European Journal of Operational Research 207 (2010) 13271340

Contents lists available at ScienceDirect

European Journal of Operational Research


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

Production, Manufacturing and Logistics

A proactive approach for simultaneous berth and quay crane scheduling


problem with stochastic arrival and handling time
Xiao-le Han *, Zhi-qiang Lu, Li-feng Xi
Department of Industrial Engineering and Management, Shanghai Jiao Tong University, 800 Dong Chuan Rd, Shanghai 200240, China

a r t i c l e

i n f o

Article history:
Received 18 June 2009
Accepted 17 July 2010
Available online 23 July 2010
Keywords:
Genetic Algorithm
Stochastic simulation
Berth and quay crane scheduling
Uncertainty

a b s t r a c t
For a container terminal system, efcient berth and quay crane (QC) schedules have great impact on the
improvement of both operation efciency and customer satisfaction. In this paper we address berth and
quay crane scheduling problems in a simultaneous way, with uncertainties of vessel arrival time and container handling time. The berths are of discrete type and vessels arrive dynamically with different service
priorities. QCs are allowed to move to other berths before nishing processing on currently assigned vessels, adding more exibility to the terminal system. A mixed integer programming model is proposed,
and a simulation based Genetic Algorithm (GA) search procedure is applied to generate robust berth
and QC schedule proactively. Computational experiment shows the satised performance of our developed algorithm under uncertainty.
 2010 Elsevier B.V. All rights reserved.

1. Introduction
Container transportation has been the mainstream of modern
logistics, and seaport container terminals are considered as key
nodes of international transportation. High performance is vital
for a container terminal due to its high capitalization and erce
competition with others. A container terminal system utilizes
space resources such as quay berths and yards, as well as transshipment equipment such as cranes and trucks to perform designated container loading/unloading and transit tasks (Vis and de
Koster, 2003). A well functional container terminal system requires
effective and collaborative decision making processes for resource
allocation and equipment scheduling. Typical decision problems
include ship planning process related to task arrangement and
coordination of arriving vessels, storage and stacking logistics in
yard, and transport optimization of container transshipment
equipment (Steenken et al., 2004).
When container vessels are sailing to the destination port, they
periodically update their ETA (estimated time of arrival) to the
port, and the terminal acquires information about container loading/unloading tasks from vessel company and cargo agents
through EDI (electronic data interchange). Based on this information the container terminal generates predictive ship operation
plan. After the container vessel arrives at the terminal, it rst waits
at anchor ground until its planned berth space becomes available.
Then this vessel is berthed at the designated quay position, and
* Corresponding author. Tel.: +86 13585930507; fax: +86 21 34206782.
E-mail addresses: hanxl@sjtu.edu.cn (X.-l. Han), zhiqianglu@sjtu.edu.cn (Z.-q.
Lu), lfxi@sjtu.edu.cn (L.-f. Xi).
0377-2217/$ - see front matter  2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2010.07.018

some pre-scheduled quay cranes are assigned to serve this vessel


and start processing its container transshipment tasks. During its
processing, the vessels position is usually xed, because its repositioning is time consuming, which involves unfasten of mooring
cables and the help from a tugboat. But quay cranes can move
around along the guide rail as long as they do not interfere with
each other, to perform handling tasks of containers located at different holds of the vessel, or at different vessels. When all container loading/unloading tasks of a vessel are nished, the service
terminates and this vessel leaves the quayside, its occupied berth
segment is then released and becomes available again. Fig. 1 is a
sketch of 3 berthed vessels and 6 serving quay cranes at a container
terminal with 3 berths.
The above process is a brief description of quayside operation,
and it follows berth and QC plans, the arrangement of which are
among the rst scheduling problems encountered by terminal
operation planners. Generally, berth scheduling is to determine
the berthing locations for arriving vessels in planning time horizon,
along with sequences and time windows for their loading/unloading tasks; and QC scheduling is to determine each available QCs
processing vessels and related processing time windows. The berth
and QC scheduling can be inuenced by each other, hence these
two resources should be simultaneously considered while making
operational plans. Because of the physical limitation of berth space
(i.e. wharf length, water depth) and capacity limitation of quay
crane handling, well organized berth allocation and QC scheduling
plans have great effect on terminals container turnover increase,
hence the operational cost reduction for both terminal and vessels,
and customer satisfaction improvement. As the service interface
between the terminal and vessels, high performance of quayside

1328

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

Fig. 1. Quay cranes serving berthed vessels.

operation also has direct or indirect impacts on the efciency of


other operational sections such as yard, yard cranes and transshipment trucks, etc.
The real-time execution of operational schedules at container
terminal is affected by different kinds of uncertainties lying in vessel arrival time, task handling time, equipment reliability, container information inaccuracy, weather variability, etc. Like the
machine scheduling in production, these uncertainties will cause
extra cost and degrade the performance of the original schedule.
One way to deal with these uncertainties is to generate a perturbation-insensitive robust schedule by taking into account uncertainties while making plans. Other ways include revision or
rescheduling when uncertainty is realized, called the predictivereactive approach (Aytug et al., 2005).
Based on practical container terminal operation, we look into
the simultaneous berth and quay crane scheduling problem, in
which each vessels arrival time and handling time are stochastic
distributed. This paper is organized as follows. The next section
provides a literature review. The mathematic formulation of our
problem is given in Section 3. Section 4 proposes a solution procedure, and numerical experiments are presented in Section 5. Section 6 concludes this paper.

2. Literature review
Many research works have been reported in the literature on
either berth or QC scheduling problem. The berth scheduling problem can be divided into two categories: discrete and continuous,
according to different indexing methods to locate the berthed vessels. In discrete situation, the wharf is divided into several separated segments, and no vessels berthing location can stride over
adjacent segments (Imai et al., 2001; Wang and Lim, 2007). In
the continuous situation, there is no partition of the wharf and vessels can be berthed anywhere along the wharf, as long as enough
space exists (Kim and Moon, 2003; Imai et al., 2005). For the QC
scheduling, the most important constraint is that quay cranes all
move along a shared guide rail, which forbids them from getting
across each other in position (Lee et al., 2008; Tavakkoli-Moghaddam et al., 2009). Bierwirth and Meisel (2010) provided a comprehensiveness survey on both berth allocation and quay crane
scheduling problems.
The amount of researches on simultaneous berth and QC
scheduling problem is relatively small. Bierwirth and Meisel
(2010) also provided a profound classication scheme of these
researches based on two kinds of integration concepts: deep integration and functional integration. Most researches of integrated
berth and QC scheduling problem focus on continuous berth situation. Park and Kim (2003) rst proposed a scheduling method
for berth and quay cranes under continuous berth situation. Both
time and space coordinates were discretized to formulate the MIP
model, and a two phased solution procedure was adopted. In rst
phase berth allocation and rough quay crane allocation was
determined, then in second phase detailed crane scheduling
was generated considering minimal setups times. Meisel and
Bierwirth (2009) investigated a similar problem with the rst
phase problem in Park and Kim (2003), assuming that when mul-

ti cranes are serving the same vessel, each cranes productivity


was decreased by interference with each other. They applied
two meta-heuristics, Squeaky Wheel Optimization and Tabu
Search, respectively to alter the vessel priority list, and proposed
a heuristic for searching better solutions under a given priority
list. In these two studies under continuous berth situation, quay
cranes can be moved to other vessel before its current vessel nishes processing, however such movement related time was not
considered when evaluating vessel processing time. Zhang et al.
(2010) considered the coverage ranges for quay cranes when
addressing the simultaneous berth and quay crane scheduling
problem under continuous berth situation, and applied a sub-gradient optimization algorithm based on Lagrangian relaxation to
search for near-optimal solutions. By limiting the quay crane
adjustments, each decomposed sub-problem was solved by a
polynomial-time enumeration procedure. Imai et al. (2008) considered berth and QC simultaneously under discrete berth situation. The authors used GA to generate the berth allocation of
vessels, and then proposed a heuristic to schedule crane transferring. The feasibility of berth schedule was checked by converting
the crane transferring problem into a network ow problem, and
crane tasks were rescheduled if necessary. Different from the two
studies under continuous berth situation, the authors assumed
that quay crane cannot move from one berth to another via other
berths if the other berths were engaged in vessel task operation,
and the assigned QC amount for each vessel was pre-determined,
hence the handling time of a vessel remained unchanged if more
QCs than needed were assigned to it. Liang et al. (2009) investigated a similar problem with Imai et al. (2008), and a combined
Genetic Algorithm with heuristic was proposed, in which a chromosome included both berth and quay crane assignment information. However the quay crane moving time was not
considered and detailed algorithm for quay crane movement
scheduling was not provided.
No impact from uncertainty on berth and quay crane scheduling
is considered in these studies, which is quite important for scheduling of complex and variable systems such as container terminal.
However there exist many studies on uncertainty based scheduling, with applications in other different areas. Singer (2000) studied a job shop scheduling problem with random processing time
and applied it in production planning. Acar et al. (2009) proposed
a generalized MIP formulation with iterative optimization-simulation procedure and applied it in facility location problem. More
studies were reviewed in Sahinidis (2004) and Aytug et al.
(2005). In this paper, we address the simultaneous berth and quay
crane scheduling problem in which QC operation on one vessel can
be interrupted and resumed by other QCs later. Uncertainty factors
such as stochastic vessel arrival and container handling time are
considered.

3. Problem description and formulation


In our problem the berth is of discrete type, i.e. the wharf is
divided into discrete segment, and each vessel can berth into any
segment as long as the space constraints are satised, but cannot
be berthed across two successive segments. The vessels are arrived

1329

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

Aj

arrival time of vessel j, which is a stochastic parameter


with probability density function uAj
task handling time of vessel j by single QC, which is also a
stochastic parameter with probability density function uCj
due time of vessel j
max QC amount that can be assigned to vessel j
weight value for vessels with priority h
setup time for QC reassignment
number large enough

Cj
Ej
Bj
bh
P
M

Fig. 2. Example of Gantt chart for vessel processing.

dynamically, i.e. their arrival time can be after the planning moment. Arriving vessels have different priority levels, representing
relative customer importance. All berths and QCs are available at
the beginning of scheduling time horizon. In our problem we make
planning decisions on each vessels berth position, service sequence and assigned QC amount. We assume when a new vessel
berths, QCs can be reassigned among berths if its necessary to provide enough QCs as planned for all berthed vessels. This means
each QC can interrupt its service operation to one vessel and move
along the guide rail to start or resume processing another vessel,
under the condition that after the movement, the assigned amount
of QCs for each berthed vessel remains unchanged. In this way the
vessels are not assigned with specic QCs but the amount of QCs.
This assumption may cause more setup time because more frequent changes of QC positions, but brings into the system more
exibility. Instead of waiting until another vessel nishes operation (which may be quite a long time), a newly arrived vessel can
start loading/unloading operation shortly after its arrived, as long
as there are enough QCs to serve all berthed vessels.
Fig. 2 is a Gantt chart of 7 vessels processed at 4 berths. Each
rectangle represents a corresponding vessels processing starting
time and ending time, and is labeled with the vessels ID and assigned quay crane amount. For instance the rectangle in berth-2 labeled with 6(4) indicates vessel-6 will be moored at this berth,
and 4 quay cranes will be assigned to handle its containers. Suppose there are 6 available QCs along the quayside. When vessel-3
berths and starts processing at time t1, because there are already
enough QCs at berth-3 (2 QCs have just nished their tasks on vessel-1), thus no QC reassignment is required. However when vessel6 berths at time t2, there willnot be enough QCs at berth-2 because
all 6 QCs are busy or idle at other berths, if no QC moves among
berths from time t1 to t2. Hence QC reassignment among berths
and additional setup time for vessel-7 and vessel-2 are required
at time t2. In Gantt chart the two shadowed segments represents
the setup time for vessel-7 and vessel-2 due to QC reassignment.
According to the classication scheme in Bierwirth and Meisel
(2010), our problem can be viewed as BAP; QCAPnumber 
QCAPspecific. We use the following notation to represent
parameters and variables that are involved in this problem:

Decision variables
xijk
= 1, if vessel j is processed at berth i as the kth vessel; = 0,
otherwise
tijk
processing starting time of vessel j at berth i as the kth vessel
cijk
processing time of vessel j at berth i as the kth vessel
assigned amount of QCs of vessel j at berth i as the kth vesbijk
sel
v jj0
= 1, if vessel j starts operation when vessel j0 is under processing; = 0, otherwise
r jj0
= 1, if vessel j starts operation after vessel j0 starts processing; = 0, otherwise
sjj0
= 1, if vessel j starts operation before vessel j0 ends processing; = 0, otherwise
ej
= 1, if vessel j is delayed by its due date; = 0, otherwise
nji
quay crane amount at berth i when vessel j enters
ujj0
= 1, if vessel j is the rst one to start operation after vessel
j0 starts operation; = 0, otherwise
wjj0
= 1, if vessel j and j0 start operation at the same time; = 0,
otherwise
zj
= 1, if reassignment of quay cranes among berths is needed
when vessel j enters; = 0, otherwise
Given the two parameters Aj and Cj taking stochastic values, our
objective is to proactively plan a robust berth and QC schedule
which has statistically good performance under these uncertainties
without rescheduling. Survey and analysis of actual terminal
operation data show that under most ordinarily circumstances,
these parameters are normal distributed, i.e. Aj  NlAj ; rAj C j 
NlCj ; rCj . Under some extreme circumstances like bad weather
condition or major failure of equipment, Aj or Cj uctuates so much
that rescheduling is usually required, which induces a different
problem from ours. All decision variables except xijk and bijk will
be stochastic variables because they are related to these stochastic
parameters. Our problem is formulated as the following model:

min Z Ef rf
XXX
s:t: f
t ijk cijk  Aj xijk
i

Wi
Qi
Dj
Tj

water depth of berth i


wharf length of berth i
tonnage of vessel j, including vertical safety distance
length of vessel j, including horizontal safety distance

XXXX

XX
X

Parameters
i 2 I f1; . . . ; eIg set of berths
j 2 J f1; . . . ; eJg set of vessels
k 2 K f1; . . . ; eJg set of processing sequences of vessels at any
berth
l 2 L f1; . . . ; e
Lg set of QCs
e set of service priorities of vessels
h 2 H f1;
.
.
.
; Hg
n
o
subset of V, vessels with priority level h
j 2 J 1; . . . ; Je

j2V h

bh xijk t ijk cijk  Ej ej ;

xijk 1 8j;

xijk 6 1 8i; k;

xijk1 P

xijk

8i; k P 2;

8i; j; k;
8i; j; k;
8i; j; k;

Aj Mxijk  1 6 t ijk 6 Mxijk

Mxijk  1 < cijk 6 Mxijk

Mxijk  1 < bijk 6 xijk Bj


X
t ij0 k1cij0 k1 P Mxijk  1 8i; j; k;

t ijk 

8
9

j0

XX
W i  Dj xijk P 0 8j;
i

10

1330

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

XX
Q i  T j xijk P 0 8j;
i

r jj0  1M 6

XX

t ijk 

11

XX

8j j0 ;

t ij0 k < r jj0 M

12

k
k
XX
XX
0
sjj0  1M <
tij0 k cij0 k 
tijk 6 sjj0 M 8j j ;
i

13
0

0:5rjj sjj  1 6 v jj 6 0:5rjj sjj 8j j ;


0

14
0

0:5v jj0 v j0 j  1 6 wjj0 6 0:5v jj0 v j0 j 8j j ;


X
nji 6 e
L 8j;

15
16

XX
Mej  1 <
tijk cijk  Ej 6 Mej
nji P

XX
0

v jj bij k
0

j j

bijk

8j;

8j; i;

18

8j j0 ;

r jj0 P ujj0

17

19
0

00

ujj0 P r jj0  r jj00 8j j j ;


X
nj0 i ujj0 8j; i;
Mzj P nji 

20
21

j0 j

cijk  C j

,
XX
i0

bi0 jk0  P

k0

v j j zj
0

j0 j

wj0 j00 P xijk  1M

8i; j; k;

22

j00 1;...;j0 1

4. Simulation based GA search

xijk 2 f0; 1g 8i; j; k;

23

t ijk ; cijk ; bijk ; nji P 0 8i; j; k;

24

v jj ; rjj ; sjj ; ej ; ujj ; wjj ; zj 2 f0; 1g 8j j;


0

xij0 0;

ensures if xijk = 0, cijk = 0; else the processing time of vessel j should


be positive. Constraint (8) ensures if xijk = 0, bijk = 0; else the assigned QC amount for vessel j should be positive and below the
up-limit. Constraint (9) ensures each vessel starts processing after
the leaving of its predecessor at the same berth. Constraints (10)
and (11) are space constraints. The vessel size cannot exceed its
berth size in vertical or horizontal dimension. Constraints (12)
(15) are denitions of r jj0 ; sjj0 ; v jj0 and wjj0 . Constraint (16) ensures
each time a new vessel is berthed, the total amount of needed
QCs will not exceed the up-limit of all available QCs at container
terminal. Constraint (17) is the denition of ej. Constraint(18) is
the denition of nji. Constraints (19) and (20) are denition for
ujj0 . Constraint (21) is the denition of zj. Constraint (22) ensures
each vessels processing time consists of task handling time and
needed setup time when QC reassignment happens. Here we assume the handling task can be equally assigned to each QC with
same handling ability, so the task handling time of a vessel can be
averaged by its assigned QC amount, as in Liang et al. (2009). The
setup time is set to a xed amount P, which is enough for crane
movement between any two berths according to the quay length
and QC moving speed. In this way the detailed quay crane position
and movement are not modeled for simplication, and results of
our problem can be viewed as an upper bound for the actual
problem.

t ij0 0;

cij0 0;

bij0 0 8i; j:

25
26

The objective function (1) is to minimize the expected value plus


standard deviation of total service time and weighted tardiness
time for all vessels in a planning time horizon. The service time is
dened in rst part of (2) as the difference between vessel processing nishing time and arrival time. It consists of waiting time and
processing time, and represents the operational efciency of the
container terminal. The tardiness time is dened in second part of
(2) as the difference between vessel processing nishing time and
due time, representing the customer satisfaction. To evaluate the
performance and robustness of berth and QC plans, both expected
value and standard deviation objectives are considered. Constraint
(3) ensures each vessel has to be processed by exact one berth. Constraint (4) ensures that at most one vessel can be processed by one
berth at the same time. Constraint (5) is the physical interpretation
of processing sequence. Constraint (6) ensures if xijk = 0, tijk = 0; else
the processing of vessel j should start after its arrival. Constraint (7)

Our problem is formulated as a non-liner mixed integer programming problem, with stochastic parameters and statistical valued objective. We adopt a simulation based Genetic Algorithm
procedure to search for robust solutions. Our procedure applies a
GA framework, and incorporates a simulation module using Monte
Carlo sampling to evaluate the performance of each single chromosome. The search procedure terminates when the max number of
generations is achieved. The owchart for this procedure is shown
in Fig. 3.
4.1. Representation
The solution the problem is represented as a chromosome
which is a string consists of 2eI eJ  1 integer items. In such a
chromosome, the rst eI eJ  1 items represent berth assignment
and service sequence for vessels, and the last eI eJ  1 items represent each vessels assigned QC amount. The items in the berth
segment are vessel IDs ordered by their processing sequences at
each berth, with 0 items representing the switch from one berth
to another. The items in the QC segment represent assigned QC

Fig. 3. Procedure of simulation based GA.

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

1331

4.3. Mutation
Mutation operation is performed on chromosomes chosen out
of the population by mutation probability. The mutation procedure
for each chosen chromosome is:

Fig. 4. Example of a chromosome.

Step1.

Step2.

Fig. 5. Gantt chart for the example chromosome.

amounts of the vessels in the corresponding positions in berth segment, also with 0 items representing the switch between berths.
Note that items in berth segment and QC segment are one-toone corresponded. All items of berth segment in a chromosome
compose the integer set G f1; . . . eJ; 0 . . . 0g, and jGj eI eJ  1.
Fig. 4 is a chromosome example with 5 vessels and 3 berths, and
its related Gantt chart is in Fig. 5. In this example vessel-3 is berthed as the 1st one at berth-1 with 4 QCs assigned, vessel-4 is berthed as the 2nd one at berth-2 with 3 QCs assigned, etc. Hence this
chromosome represents the solution with x131, x211, x242, x351,
x322 = 1, b131 = 4, b211 = 5, b242 = 3, b351 = 4, b322 = 5, and xijk, bijk = 0
for other i, j, k. Note that information about possible QC reassignment is not included in the chromosome; instead the necessity
of QC reassignment is further determined during the evaluation
phase of GA in Section 4.5. Other auxiliary decision variables for
the solution should be decided in the evaluation phase as well.

Step3.

The mutation operation does not generate illegal solutions. Two


examples of mutation operation are given in Fig. 6. Gantt charts related to the chromosomes before and after mutation in these two
examples are provided in Fig. 7.
4.4. Crossover
Crossover operation is performed on pairs of chromosomes chosen out of the population by crossover probability and randomly
matched. The crossover procedure for each pair of chromosomes
(ch1, ch2) is:
Step1.
Step2.

4.2. Initialization
In initialization, chromosomes are randomly generated to construct the initial population. To generate a chromosome, rst we
derive its berth segment by generating a random sequence of elements in G. Then each item in the QC segment is generated as a
random integer between 1 and the max QC amount of its corresponding vessel in the berth segment, unless for those 0 items in
berth segment, their corresponding QC amounts in QC segment
are set to 0. If such a chromosome satises the space constraints
(10) and (11), its added into the initial population. The generating
process is repeated until the population size reaches a given
number.

Randomly choose an item from the chromosome. Let g


denote its index. If this gth item is in QC segment, go to
Step2, else go to Step3.
Randomly choose another item from QC segment. Let g0
denote its index. For each item indexed from g to g0 ,
change its value to a random integer between 1 and
the max QC amount of its corresponding vessel in the
berth segment, if its original value is not 0. Terminate
procedure.
Randomly choose another item from berth segment. Let
g0 denote its index. Randomly reorder the items indexed
from g to g0 , along with their corresponding items in QC
segment. If the generated chromosome satises space
constraints (10) and (11), terminate procedure; else go
to Step1.

Step3.

Step4.

Randomly choose two items from the berth segment of


chromosome ch1. Let f, g(f < g) denote their indexes.
Interchange the gene segments from the fth item to the
gth item between ch1 and ch2. Interchange corresponding items in QC segments between ch1 and ch2.
List redundant items of ch1s berth segment by comparing it with G. Note that for a pair of nonzero redundant
items, the one outside the gene segment from the fth
item to the gth item is chosen and added to the list. For
a group of redundant zero items, those outside the gene
segment from the fth item to the gth item are chosen and
added to the list.
List redundant items of ch2s berth segment, using the
same method with Step3.

Fig. 6. Two examples of mutation operation.

1332

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

Fig. 7. Gantt chart representation of chromosomes before and after mutation in the two examples.

mean value E[f(ch)] and standard deviation r[f(ch)] are estimated


to calculate the objective function value in (1).
Since only vessel-to-berth allocation and QC-to-vessel assignment information are explicitly provided by the given chromosome
ch, a heuristic is applied in each simulation round to derive detailed berth and QC schedule from ch. In the heuristic, as time
evolves along the planning horizon, berths are scanned for QC
assignment feasibility when new vessels start processing. Then
possible vessel determent and QC reassignment are imposed to
make necessary adjustment to the feasibility of solution implied
by ch. The detailed simulation procedure including the heuristic
is as follows:
Step1.
Step2.
Fig. 8. Example of crossover operation.

Step3.

Step4.
Step5.

Randomly exchange redundant items between ch1 and


ch2. Interchange corresponding items in QC segments
between ch1 and ch2. If the generated chromosome satises space constraints (10) and (11), terminate procedure; else go to Step1.

The crossover operation may generate illegal solutions after


Step2, because there could be redundant and missing items in
two chromosomes due to interchange operation. Note that the
redundant items from one chromosome is exactly the missing
items in the other chromosome, hence Step35 are used to repair
the illegal solutions. An example of crossover operation is given
in Fig. 8. Gantt Charts related to ch1 and ch2 before and after crossover are provided in Fig. 9.

4.5. Evaluation by simulation


To evaluate the performance of a chromosome ch under stochastic arrival and handling time, Monte Carlo sampling and simulation is adopted. Denote sample size as S, and denote f(ch) as
chromosome chs f value which is dened in Eq. (2). For f(ch), its

Step5.

Step6.
Step7.
Step8.

Set {fr} = {0, . . . , 0} (r = 1, . . . , S); get xijk and bijk values from
ch.
"j 2 J, generate random samples fnAj;r g; fnCj;r g r
1; . . . ; S for Aj and Cj according to uAj and uCj . Set r = 1.
Set Jo = J. "i, j, k, if xijk = 1, set cijk nCj;r =bijk , else cijk = 0. "i,
P
j, k, if xijk = 1, set tijk maxfnAj;r ; j0 tij0 k1 cij0 k1 g, else
tijk = 0.
Calculate r jj0 ; sjj0 , v jj0 and nji("j0 j 2 J, i 2 I), based on current x , c , t
and bijk values. Let j1 arg minj2Jo
P P ijk  ijk ijk
j , go
i
k t ijk . If constraint (17) is satised by vessel
P P1
to Step5; otherwise let j2 arg minj2J f i k t ijk
PP
PP
PP
cijk g; s:t:
cij2 k .
i
k t ij2 k 6
i
k t ij1 k <
i
k t ij2 k P
P
Let i1
i; k1
k; s:t: xij1 k 1, and ti1 j1 k1 i k
P
0
t ij2 k cij2 k . For each k
k; s:t: k 2 k1 1; eJ;
j xi1 jk
0
1, let j
j; s:t: xi1 jk0 1, and t i1 j0 k0 maxfti1 j0 k0 ;
P
0
0
j t i1 jk 1 ci1 jk 1 g. Go back to the beginning of this
step.
Calculate ujj0 ; wjj0 , zj values based on values of other decision variables. If zj1 0, go to Step6; otherwise, for each
j3
j; s:t: v jj1 1 let i3
i; k3
k; s:t: xi3 j3 k3 1, and
00
ci3 j3 k3 ci3 j3 k3 P;
for
each
k
k; s:t: k 2 k3
P
00
1; eJ; j xi3 jk3 1, let j
j; s:t: xi3 jk00 1, and ti3 j00 k00
P
00
00
maxfti3 j00 k00 ;
j t i3 jk 1 ci3 jk 1 g.
PP
PP
J o J o n fjj i k tijk i k t ij1 k g. If jJoj > 0, go to Step4,
else calculate fr value by (2).
If r < S, r = r + 1, go to Step3.
e f ch,
Calculate
Zf ch e
Ef ch r
in
which
q
P
P
P
2
e
e f ch
Ef ch r fr =S, and r
r fr 
r 0 fr 0 =S =S;
r; r0 1; 2; . . . ; S.

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

1333

Fig. 9. Gantt chart representation of ch1 and ch2 before and after crossover.

4.6. Reproduction
Chromosomes are randomly selected and copied to generate the
next generation, using a Roulette Wheel selection procedure. The
probability of a chromosome being copied is proportional to the
tness value related to the evaluated objective function value
acquired by simulation in Section 4.5. For a given chromosome
ch, its tness value is dened as in (27), where a is a problem scale
related parameter.

Fch f1 expfZf ch=agg1 :

27

4.7. Local search


When an improvement occurs during GA process, local renement is taken to identify potential better solutions along some
promising searching directions. Such local search is also under stochastic circumstances, and uncertainties are taken into account
through simulation as in Section 4.5. The procedure of local search
based on chromosome ch is:
Step1.

Apply evaluation procedure in Section 4.5 on ch, and calculate expected value of each vessels waiting time and
P

P

handling time, i.e. E
i;k t ijk  Aj and E
i;k cijk .

Step2.

Get J0  J, which is the top 20% elements of J with biggest


P 
E
t
Aj
i;k ijk
P
 ratio values.
E

Step3.

Step4.

Step5.

c
i;k ijk

For each vessel j in J0 , change the position of its representative gene item in berth segment of ch, without changing other items relative locations. And reposition js
corresponding item in QC segment.
For each vessel j in J0 , regenerate the corresponding QC
amounts of vessels that are assigned to the same berth
with j and processed before j starts operation.
Evaluate newly generated chromosomes in Step3 and
Step4 by simulation. If better solution than the bestso-far solution of GA process is found, update the bestso-far solution.

5. Computational experiment
The performance test and analysis of our algorithm for the
simultaneous berth and QC scheduling problem are conducted in
this Section. First in Section 5.1 the parameters for both GA and

simulation procedures are determined by preliminary experiments


on test instances. Then in Section 5.2, our algorithm is applied under deterministic environment to compared with two existing literature on discrete berth and QC scheduling problem. In Section
5.3, our algorithm is applied under stochastic environment and
the solutions are compared with those acquired under deterministic environment. All computations are conducted on an Intel T2400
processor with 1.83 GigaHertz clock.
5.1. Parameter selection
5.1.1. Determining the sample size for simulation
Since sampling and simulation are incorporated in the evaluation phase of GA procedure, computational time will increase signicantly. To acquire high quality solution by GA in reasonable
time, the sample sizeS is a key parameter which needs to be properly chosen. If not enough samples are generated, the simulation
may provide deviated results when estimating chromosomes statistical performance; however if too many samples are generated,
the evaluation process may require more computational time than
acceptable.
To determine the value of sample size, we generate 5 test instances; and for each instance, 10 random chromosomes are generated and evaluated under different sample sizes, i.e. S takes the
value of 50, 100, 250, 500, 750, 1000, 5000, 10000, 50,000 and
100,000, respectively. Let chi,j denote the generated jth chromosome for the ith instance (i = 1, . . . , 5 and j = 1, . . . , 10). For chi,j, let
e
e S chi;j  denote respectively the estimated E[f(chi,j)]
E S f chi;j  and r
and r[f(chi,j)] values through simulation with the sample size of S.
e Si denote respecE Si and D r
Besides, for the ith test instance, let D e
tively the mean error of estimated mean value and standard deviation on the 10 generated chromosomes, through simulation with
the sample size of S. Furthermore we consider e
E 100000 f chi;j  and
e 100000 f chi;j  as the sufciently well estimated E[f(chi,j)] and
r
e Si can be expressed as:
r[f(chi,j)] values for chi,j, thus D eE Si and D r




1 X  e
E S f chi;j 

S
e
DEi
 1;


10 j1;...;10  e
E 100000 f chi;j 



e S f chi;j 
1 X  r
:
e Si

1
Dr


e 100000 f chi;j 
10 j1;...;10 r
e S and D r
e Si valFor the ve instances, the comparison among D E
i
ues under different sample sizes are shown in Fig. 10. Generally
e Si values tend to decrease as sample size increases
E Si and D r
both D e
for each instance i. The mean errors of estimated mean value are

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340
S

S
1

E3

E4

S
4
S

E5

100

250

500 750 1000 5000 10000 50000


Sample size S

objective function values

155
135
115
95
75
55
35
15
101

201

301

401

501
601
Generation

701

801

901

1001

Fig. 11. Performance comparison of different mutation rates for an instance under
population size of 50 and crossover rate of 0.8.

crossover rate=0.1
crossover rate=0.2
crossover rate=0.4
crossover rate=0.6
crossover rate=0.8

175
155
135
objective function value

115
95
75
55

15
1

101

201

301

401

501
601
generation

701

801

901

Fig. 13. Performance comparison of different population sizes for an instance under
crossover rate of 0.8 and mutation rate of 0.8.

mutation rate=0.1
mutation rate=0.2
mutation rate=0.4
mutation rate=0.6
mutation rate=0.8

175

135

35

Fig. 10. Mean errors of estimated mean value and standard deviation for ve
instances under different sample sizes.

195

population=30
population=50
population=70
population=100

155

E2

50

175

E1

0.11
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0

objectvie function value

Mean errors

1334

115
95

5.1.2. Determining the GA parameters


GA searches with different parameter settings are applied on
the ve instances to select an effective parameter combination
for our problem. The population size takes the value of 30, 50, 70
and 90, respectively, and both the crossover rate and mutation
rates take the value of 0.1, 0.2, 0.4, 0.6 and 0.8, respectively. Each
GA terminates until 1000 generations are reproduced, and the
sample size S takes the value of 750, as determined in Section 5.1.1.
Figs. 1113 illustrate some comparisons among these different
parameter settings. Fig. 11 shows the convergence curves on different mutation rates for a typical instance with the crossover rate of
0.8 and population size of 50. Fig. 12 shows the convergence curves
on different crossover rates for the instance with the mutation rate
of 0.8 and population size of 50. Fig. 13 shows the convergence
curves on different population sizes for the instance with the crossover rate of 0.8 and mutation rate of 0.8. The mutation rate of 0.8
outperforms other values in Fig. 11, and the crossover rate of 0.8
outperforms other values in Fig. 12. Hence we set mutation rate
and crossover rate to 0.8 in GA search. In these two gures, the
objective function values are not improved after 500 generations
when crossover and mutation rate are set to 0.8, so the maximum
generation of 500 will be sufcient to acquire a near-optimal solution. Besides, in Fig. 13, the objective function values of the best
found solutions after 500 generations are close to each other under
population size of 50, 70 and 90, so the population size is set to 50
for GA search.
5.2. Performance under deterministic circumstance

75

Fig. 12. Performance comparison of different crossover rates for an instance under
population size of 50 and mutation rate of 0.8.

In literature, Imai et al. (2008) and Liang et al. (2009) studied


the simultaneous berth and QC scheduling problem with no allowing of quay crane reassignment during any vessels processing,
which is a main difference from our assumption. The consideration
behind our proposal is that by allowing quay crane reassignment,
the quay side handling system is given higher exibility, although
at the risk of more frequent setup. To examine the inuence of adding such exibility into the problem, computational results of our
algorithm under deterministic environment are compared with
these two literatures, respectively.

within 2% for all instances and sample sizes, while the mean errors
of estimated standard deviation are within 2% for all instances
when sample size exceeds 750, which implies that 750 samples
will be sufcient for the estimation of stochastic environment. As
a result we set S = 750 for the simulation process.

5.2.1. Computational experiment on a real case


We rst apply our algorithm to the problem demonstrated in
Liang et al. (2009), which is a real case. The calculation is performed under deterministic circumstance, i.e. the simulation is
performed only once when evaluating each chromosome, and each

55
35
15
1

101

201

301

401

501
601
generation

701

801

901

1001

1335

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

38
36
34
32
Bj=4
Bj=5
Bj=6

30
28

0.15
0.25
0.35
1
Setup time for QC reassignment, P (hr)
Fig. 14. Solutions for problem in Liang et al. (2009).

2000

lue increases as the setup time increases, which accords with intuition. The objective function value of the approximate solution
found in Liang et al. (2009) is 2165 minutes for all vessels, i.e.
36.08 hours. For P = 0.15, 0.25 or 0.35, the approximation solution
found by our algorithm is better than 36.08, but for P = 1, the
approximation solution exceeds 36.08. It demonstrates that when
setup time is small, its worth adjusting and moving quay cranes
before the end of its service at currently assigned vessel, but when
setup time is too large, doing this will cost more time than it saves.
Considering quay crane moving speed and quay length of berth,
such setup time will not exceed 0.5 hours, which makes our
assumption of QC service preemption preferable.
5.2.2. Computational experiment on randomly generated cases
We apply our algorithm to problems with same data congurations as in Imai et al. (2008): Vessels arriving interval follows
exponential distribution with expected value Int, and handling
time follows 3-Erlang distribution with expected value Se. Calculation is also conducted under deterministic circumstance, and as in
Imai et al. (2008), the objective function of total service time is
adopted, i.e. f is dened as only the rst half of (2)s right side.
Problems under different scales are randomly generated using
one seed according to the given parameter distributions, i.e.
Int = 2, 3, 4, and 5 hours, Se = 6 and 8 hours, total QC amount
e
L 8; 10; 12; 14; 16; 18, berth amount eI 4 and 5, respectively.
The vessel amount eJ is roughly inverse proportional to Int, and
takes the value of 88, 64, 45 and 34, respectively.
The results are illustrated following the patterns in Imai et al.
(2008): Fig. 15 shows the relationship between total service time
and total quay crane amount under different average arriving

2000

Int=2

Int=2
Int=3

Total service time (hr)

Int=4
Int=5

1200

800

Total service time (hr)

Int=3
1600

1600

Int=4
Int=5

1200

800

400

400

0
8

10

12
14
Total QC

16

18

4 berths and Se=6 hr


6000

12

14

16

18

5 berths and Se=6 hr


5000

Int=2

Int=2

Int=3

5000
Total service time (hr)

10

Total QC

Int=4
4000

Int=5

3000
2000

Int=3
Total service time (hr)

Total service and delay time


(hr)

stochastic parameter is xed to its nominal value given in the paper. For GA evaluation, the objective function of total waiting, handling and delay time used in Liang et al. (2009) is adopted, which
equals E(f) in (1). In Liang et al. (2009), the maximum allowed
amount of QCs assigned is the same for each vessel, i.e. Bj takes
the same value for each j. Since such Bj value is not indicated in
the paper, we set Bj value from 4 to 6, respectively. The QC reassignment setup time P takes the value of 0.15, 0.25, 0.35 and 1,
respectively.
The computational results are shown in Fig. 14. The objective
function value decreases as the max amount of quay cranes increases in Fig. 14. This is because in this case, arriving vessels are
relatively sparse, thus allowing assigning more quay cranes to
one vessel will shorten their processing time without causing other
vessels waiting for longer time. Besides, the objective function va-

4000

Int=4
Int=5

3000

2000

1000

1000
0

0
8

10

12
14
Total QC

16

4 berths and Se=8 hr

18

10

12
14
Total QC

16

5 berths and Se=8 hr

Fig. 15. Total service time under different problem scales.

18

1336

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

25

25

Total QC=8

Total QC=8

20

Total QC=12
Total QC=14
Total QC=16

15

Total QC=18
10

Total QC=10
Average service time (hr)

Average service time (hr)

Total QC=10
20

Total QC=12
Total QC=14
Total QC=16

15

Total QC=18
10

2
3
4
5
Average arriving interval, Int (hr)

2
3
4
5
Average arriving interval, Int (hr)

4 berths and Se=6 hr

5 berths and Se=6 hr

70

Average service time (hr)

60
50
40
30
20

Total QC=8
Total QC=10
Total QC=12
Total QC=14
Total QC=16
Total QC=18

50
40
30
20
10

10
0

Average service time (hr)

60
Total QC=8
Total QC=10
Total QC=12
Total QC=14
Total QC=16
Total QC=18

2
3
4
5
Average arriving interval, Int (hr)

2
3
4
5
Average arriving interval, Int (hr)

4 berths and Se=8 hr

5 berths and Se=8 hr

Fig. 16. Average service time under different problem scales.

The average service time in Fig. 16 is calculated by dividing all


vessels total service time by the vessel amount, and can be viewed
on average as the sum of each vessels task handling time, waiting
time and setup time for QC reassignment. For most of the time, the
average service time decreases as arriving interval increases. Because as fewer vessels are arriving, both the probability and intensity of vessels competing for berth and QC resources will reduce,
which leads to the reduction of both waiting time and setup time

1.6
Ratio of S2 and S1s f values

intervals; and Fig. 16 shows the relationship between average service time and average arriving interval under different total quay
crane amounts. As in Imai et al. (2008), total service time in
Fig. 15 decreases as fewer vessels are arriving during the planning
horizon, or less average task handling time Se is required. The total
service time monotonously decreases as total QC increases, which
is different from Imai et al. (2008). This is because by allowing QC
reassignment during vessel processing, its easier to utilize idle QCs
at other berths, and more service time can be saved than the time
spent on QC movement and adjustment. Besides, as total QC increases, total service time under different arriving intervals get closer, and when total QC is more than 14, the reduction range of total
service time under each arriving interval gets smaller. This is because when enough QCs are available, most vessels are assigned
with their maximum allowed QC amounts; hence no obvious
improvement can be further achieved. Moreover, the total service
time decreases with the increasing berth amount, especially for
the situations which require higher productivity, i.e. more frequently arriving vessels (Int = 2), longer average task handling time
(Se = 8), or less available quay cranes (Total QC = 8 or 10). This is
because in such situations, the QC reassignment operations are
more benecial: they help improving QC utilization in concert with
the augmented berth capacity by reducing vessels possible waiting time caused by the lack of QC exibility.

Int=2
Int=3
Int=4
Int=5

1.5
1.4
1.3
1.2
1.1
1

10

12
14
Total QC

16

18

Fig. 17. Ratio of S2 and S1s f values under deterministic circumstance (5 berths and
Se = 8 hours).

1337

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

1
0.95
0.9

Int=2

0.85

Int=3
Int=4

0.8

Int=5

0.75
0.7
8

10

12
14
Total QC

16

18

Fig. 18. Ratio of S2 and S1s expected f values under stochastic circumstance (5
berths and Se = 8 hours).

for QC reassignment. However a partial increasing trend can be observed when the arriving interval changes from 3 to 4 hours, which
implies the increasing of task handling time offsets the reduction
of waiting time and setup time. Similar to the total service time,
average service time tend to be steady when total QC is more than
14, indicating the redundancy of available quay cranes.
5.3. Performance under stochastic circumstance
In previous Section 5.2, solutions are acquired under deterministic circumstance, and we denote them by S1. In this section, our
algorithm is applied under stochastic environment, i.e. the simulation during evaluation phase is performed for S iterations based on

5.3.1. Comparing the statistical performance of S1 and S2s


The generated cases in Section 5.2.2 with 8 hours of average
handling time and 5 berths are used in this section. SetP = 0.25
and bh = 1 for each h in (2). Vessels arrival and task handling time
are assumed to follow normal distributions with rCj lCj =3 and
rAj Int=3. The detailed procedure of fASj;r ; C Sj;r g sample generation
is in Appendix.
Fig. 17 shows the fdet(S2)/fdet(S1) ratios for all cases, and Fig. 18
shows the E[fsto(S2)]/E[fsto(S1)] ratios for the same cases. Fig. 19
shows the [l  r, l + r] ranges of both fsto(S1) and fsto(S2). Fig. 20
illustrates the sampled probability distributions of fsto(S1) and
fsto(S2) for several cases. In Fig. 17, all fdet(S2)/fdet(S1) ratios exceed

[-, +] range of S1 and S2 f values

8000

4500

7000
6000
5000
S2

4000

S1
3000
2000
1000
0

[-, +] range of S1 and S2s f values

0.65

the sampled vessel arriving time and task handling time, and these
solutions found under such stochastic circumstance are denoted as
S2. To restrict the computational time, the sample size S is set to
750 during GA search as previously determined.
After both S1 and S2 have been acquired, the statistic performance of S1 and S2 are compared by applying simulation procedure on them and setting S to 100,000. To differentiate between
the f values of a given solution ch(ch = S1 or S2) under two kinds
of circumstances, we denote fdet(ch) and fsto(ch) as the f(ch) values
under deterministic and stochastic circumstances, respectively. For
fsto(ch), its expected value and standard deviation, i.e. E[fsto(ch)] and
r[fsto(ch)] can be estimated by the simulation procedure in Section
4.5, and the sample size S is set to 100,000 in order to provide more
accurate estimations for statistic performance comparison. Correspondingly, fdet(ch), the f(ch) value under deterministic circumstance can be calculated after a single-iteration simulation with
nominal parameter values.

4000
3500
3000
2500
S2

2000

S1

1500
1000
500
0

10

12
14
16
Total QC

18

10

Arrving interval=2

[-, +] range of S1 and S2s f values

2000
1800
1600
1400
1200

S2

1000

S1

800
600
400
200
0
8

10

12
14
16
Total QC

Arrving interval=4

12
14
16
Total QC

18

Arrving interval=3

18

[-, +] range of S1 and S2s f values

Ratio of S2 and S1s expected f


values

1.1
1.05

700
600
500
S2
400

S1

300
200
100
0
8

10

12
14
16
Total QC

18

Arrving interval=5

Fig. 19. Statistical range of S1 and S2s f values under stochastic circumstance (5 berths and Se = 8 hours).

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

0.16

0.14

0.14

0.12

0.12

S1

0.1

S2

0.08

S1
S2

0.1
Probability

Probability

1338

0.08
0.06

0.06
0.04

0.04

0.02

0.02
0
1150 1650 2150 2650 3150 3650
f value

0
590

Arrving interval=2, Total QC=14

915 1240 1565 1890 2215 2540


f value

Arrving interval=3, Total QC=12


0.18

0.25

0.16
0.14

0.15

S1
S2

0.1

Probability

Probability

0.2

0.12
S1
S2

0.1
0.08
0.06
0.04

0.05

0.02
0

0
450

725 1000 1275 1550 1825 2100


f value

Arrving interval=4, Total QC=10

330 450 570 690 810 930 1050 1170


f value

Arrving interval=5, Total QC=8

Fig. 20. Probability distributions of S1 and S2s f values under stochastic circumstance (5 berths and Se=8 hours).

1, which means under deterministic circumstance, S1 always outperforms S2. However in Fig. 18 under stochastic circumstance, the
expected f value of S2 is within 105% range of S1s expected f value,
and in some occasions E[fsto(S2)] is smaller than E[fsto(S1)]. In addition to that, r[fsto(S2)] is smaller than r[fsto(S2)] for all cases in
Fig. 19, leading to more stability of the f value for S2 under variable
parameters. Such character becomes more obvious when fewer
QCs are available or more vessels are arriving. The probability distributions of fsto(S1) and fsto(S2) in Fig. 20 also suggest the robustness of S2 when facing uncertainties: although with a price of
wasting opportunity for efciency improvement under some benecial situations, its efciency will not degenerate as much as S1
does under those pernicious situations, especially when high productivity is required at the terminal. Hence, its necessary to incorporate uncertainties into the decision process in order to acquire
proactive schedule plans with strong stability and better statistical
performance.
5.3.2. Impact of the degree of uncertainty
We generate 10 random cases under the problem scale of
8 hours of average handling time and 5 berths. In each case there
are 34 arriving vessels with 5 hours of average arrival interval.
Our algorithm is then applied to each case under different stochastic circumstances by setting rCj =lCj and rAj =Int values to be 1/4, 1/3
and 1/2, respectively. The probability of fsto(S2) being better than
fsto(S1) can also calculated during simulation.
Results are shown in Table 1. On average, fdet(S2) is 5.67% worse
than fdet(S1), but E[fsto(S2)] is 6.27% better than E[fsto(S1)], and
r[fsto(S2)] is 30.05% better than r[fsto(S1)]. As a result, while
E[fsto(S2)]  r[fsto(S2)] decreases by 1.47% from E[fsto(S1)]  r[fsto-

(S1)], E[fsto(S2)) + r(fsto(S2)) decreases by 9.56% from E[fsto(S1))


+ r(fsto(S1)), which leads to the shrink of [l + r, l + r] range by
8.09% from fsto(S1) to fsto(S2). Besides, there is an average probability
of 73.21% that S2 has smaller total service and weighted delay time
than S1 under stochastic parameters. These values indicate that
though S1 outperforms S2 when uncertainty does not exist, S2 outperforms S1 under stochastic circumstances in both average and
deviation aspects. As the degree of uncertainty, i.e. rCj =lCj and
rAj =Int increases, these differences between S2 and S1 get more obvious, which makes S2 more preferable.
6. Conclusion
This paper addresses the problem of combining discrete berth
and QC scheduling under uncertainty, which is quite important
for container terminal operation for two reasons. Firstly, when performing quay side operation, there exists inherent interrelationship between the allocation and utilization of berth and QCs,
which are the major valuable resources of the terminal and quite
often the bottlenecks for serving vessel customers. So the operational berth and QC schedules should be simultaneously considered in decision making process. Secondly, the uncertainty is
naturally inevitable when planned schedules are executed in terminal system and has its inuence on the actual performance of
terminal operation. Since the entire container terminal system is
capital intensive and complicated with lots of operational regulations, terminal operators tend to have risk aversion, and wish to
maintain a stable running of the whole system despite these uncertainties. In this paper we propose a method to cover these two
aspects.

215.93
268.32
24.71
243.61
293.03
199.47
292.58
50.15
242.43
342.73
0.76
228.18
299.45
34.58
264.87
334.03
214.47
311.97
50.37
261.61
362.34
0.61
239.63
333.02
42.33
290.69
375.34
231.82
346.00
57.37
288.63
403.36
0.62
249.20
342.54
54.95
287.59
397.48
236.31
375.22
69.73
305.49
444.95
0.87
341.26
519.79
107.10
412.70
626.89
563.94
826.54
132.27
694.27
958.81
fdet
E(fsto)
r(fsto)
E(fsto)  r(fsto)
E(fsto) + r(fsto)
P (fsto(S2) < fsto(S1))
1/2

458.60
1150.85
359.51
791.34
1510.37
0.89

310.41
564.94
136.13
428.81
701.06
0.85

205.65
238.57
16.60
221.97
255.18
199.97
250.12
26.44
223.68
276.56
0.76
214.83
266.18
25.83
240.35
292.01
214.47
267.06
27.47
239.59
294.53
0.52
232.01
287.87
28.56
259.31
316.44
231.32
296.25
33.19
263.06
329.44
0.76
242.95
301.58
31.15
270.43
332.72
235.77
311.75
37.46
274.30
349.21
0.78
347.41
443.23
54.01
389.22
497.24
488.68
627.56
72.13
555.42
699.69
fdet
E(fsto)
r(fsto)
E(fsto)  r(fsto)
E(fsto) + r(fsto)
P (fsto(S2) < fsto(S1))
1/3

450.45
721.20
125.21
595.98
846.41
0.89

308.99
462.28
86.05
376.23
548.34
0.63

202.43
227.22
12.23
214.99
239.45
199.47
230.97
16.93
214.04
247.91
0.62
214.61
247.20
16.74
230.46
263.93
214.47
248.39
17.83
230.55
266.22
0.64
236.73
272.07
19.89
252.18
291.96
232.07
274.01
21.95
252.06
295.97
0.65
244.09
278.24
18.73
259.51
296.98
234.85
291.10
28.70
262.40
319.80
0.81
322.42
397.74
39.80
357.94
437.53
311.41
412.80
57.60
355.20
470.40
0.70
486.18
569.27
45.55
523.72
614.81
452.60
654.06
127.84
526.23
781.90
0.84
fdet
E(fsto)
r(fsto)
E(fsto)  r(fsto)
E(fsto) + r(fsto)
P (fsto(S2) < fsto(S1))
1/4

18
16
14
12
10

S1
Solution

S2
8

S1

Total QC

rcj =lcj & rAj /Int

Table 1
Average performance of solutions to 10 cases (5 berths, 34 vessels, Se = 8 hours, Int = 5 hours).

S2

S1

S2

S1

S2

S1

S2

S1

S2

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

1339

On one hand, to make the GA simultaneously search for better


berth and QC schedules, chromosome structure with both berth
and QC allocation information are designed, along with corresponding GA operators. The interrelationship between berth and
QC are handled by a heuristic, in which we allow vessel processing
interruption for QC reassignment, in order to obtain higher capacity utilization. Such treatment is actually a kind of tradeoff between increased setup time and saved waiting time among
different vessels. Although the setup time P takes a xed value as
an upper bound, our proposed treatment and algorithm outperforms Liang et al. (2009), in which vessel operation is uninterrupted. Besides, a monotonic changing trend of total service time
by total QC amount is observed comparing with Imai et al.
(2008), and indicates the stable performance of our algorithm:
available QCs can be effectively utilized, instead of remaining idle
and waiting for the end of some uninterrupted vessel operation.
However some non-monotonic changing trend of average service
time by arriving interval implies the positive effect of QC reassignment may be offset by its negative effect, and indicates that the
heuristic can be further improved with capability of identifying
such situations. With the development of automated container terminal handling system and the progress of control technology, QC
reassignment setup time, including moving and readjustment time
will be further reduced, and container terminals will attain better
performance by exible QC reassignment when facing increasing
demand and erce competition.
On the other hand, to use the information of stochastic parameters proactively in order to generate robust solutions, Monte Carlo
sampling and simulation procedure are applied. In order to save
computational time, sample size is selected as the minimum one
with an acceptable relative error. Such sample size value can be revised in practical use, based on the consideration of balancing the
estimation accuracy and computational time. We use the sum of
mean value and standard deviation as the objective function to
be optimized, since we are looking for solutions with both efciency and stability, which are illustrated by statistical performance comparisons between S2 and S1. These results also
suggest that the proactive method is more effective when the terminals processing capacity is less sufcient relative to handling
task demand, or when the uncertain degree is higher, and thus
its worth the time to run the simulation based GA search to acquire a schedule plan with both high efciency and stability. When
processing capacity is redundant, since the redundant capacity can
be used as buffers when uncertainty occurs, the proactive method
is less effective considering its time consumption, but it still generates solutions better than S1.
Our proposed model and solution procedure can serve terminal
operators in two ways. Firstly the algorithm can be applied in daily
routine of operational management by providing efcient and robust schedule plans for quay side operation. Secondly, when
upgrading or introducing new hardware such as automated container terminal handling system, the return on investment can be
assessed by applying the algorithm under the anticipative terminal
conguration and evaluating the saved cost related to processing
time, customer satisfaction, etc.
In problem modeling process, vessels processing time is
roughly estimated to be inverse proportional to its assigned QC
amount, because to get vessels exact processing time, the detailed
container stowage plan and QC working plan on each bay should
be incorporated into the model. For the same reason, the setup
operation due to QC reassignment is simplied by interrupting
and inuencing all vessels being processed at QC reassignment
time, and unnecessary setup time may be added onto these inuenced vessels. So the acquired solution can be viewed as an upper
bound. Besides, uncertainty lies in the arrival and task handling
time, and is viewed as regular disturbance. However in practice

1340

X.-l. Han et al. / European Journal of Operational Research 207 (2010) 13271340

there is another kind of irregular uncertainty of accidental events,


like equipment break down or major delay caused by bad weather,
etc. To incorporate such uncertainty into the problem may require
different approaches like reactive or rescheduling. These two aspects are our future research directions.
Acknowledgements
This research is supported by National Natural Science Foundation of China (Nos. 70771065, 70802040) and National High Technology Research and Development Program of China (863
Program) (Nos. 2009AA043000, 2009AA043001). Thanks are due
to two anonymous reviewers for comments that improved the
quality of this paper.
Appendix. Sample generation procedure for stochastic
parameters following normal distribution
The sample generation procedure for fnAj;r g; fnCj;r g r 1; . . . ; S;
j 2 eJ is as follows. uAj and uCj are specied to be normal distribution in computation experiment.
Step1. Generate 4 sets of [0, 1] Uniform distributed numbers:
fe11 ; . . . ; e1 g, fe21 ; . . . ; e2 g, fe31 ; . . . ; e3 g and fe41 ; . . . ; e4 g,
eJ s
eJ s
eJ s
eJ s
using 4 different seeds.
Step2. Let

nAj;r lAj rAj


nCj;r lCj rCj

q
2 ln e1rsj1 sin2pe2rsj1 ;

q
2ln e3rsj1 sin2pe4rsj1 r 1;. .. S; j 2 J:

References
Acar, Y., Kadipasaoglu, S.N., Day, J.M., 2009. Incorporating uncertainty in optimal
decision making: Integrating mixed integer programming and simulation to

solve combinatorial problems. Computers & Industrial Engineering 56 (1), 106


112.
Aytug, H., Lawley, M.A., McKay, K., Mohan, S., Uzsoy, R., 2005. Executing production
schedules in the face of uncertainties: A review and some future directions.
European Journal of Operational Research 161 (1), 86110.
Bierwirth, C., Meisel, F., 2010. A survey of berth allocation and quay crane
scheduling problems in container terminals. European Journal of Operational
Research 202 (3), 615627.
Imai, A., Nishimura, E., Papadimitriou, S., 2001. The dynamic berth allocation
problem for a container port. Transportation Research Part B: Methodological
35 (4), 401417.
Imai, A., Sun, X., Nishimura, E., Papadimitriou, S., 2005. Berth allocation in a
container port: Using a continuous location space approach. Transportation
Research Part B: Methodological 39 (3), 199221.
Imai, A., Chen, H.C., Nishimura, E., Papadimitriou, S., 2008. The simultaneous berth
and quay crane allocation problem. Transportation Research Part E: Logistics
and Transportation Review 44 (5), 900920.
Kim, K.H., Moon, K.C., 2003. Berth scheduling by simulated annealing.
Transportation Research Part B: Methodological 37 (6), 541560.
Lee, D.H., Wang, H.Q., Miao, L., 2008. Quay crane scheduling with non-interference
constraints in port container terminals. Transportation Research Part E:
Logistics and Transportation Review 44 (1), 124135.
Liang, C., Huang, Y., Yang, Y., 2009. A quay crane dynamic scheduling problem by
hybrid evolutionary algorithm for berth allocation planning. Computers &
Industrial Engineering 56 (3), 10211028.
Meisel, F., Bierwirth, C., 2009. Heuristics for the integration of crane productivity in
the berth allocation problem. Transportation Research Part E: Logistics and
Transportation Review 45 (1), 196209.
Park, Y.M., Kim, K.H., 2003. A scheduling method for berth and quay cranes. OR
Spectrum 25 (1), 123.
Sahinidis, N.V., 2004. Optimization under uncertainty: State-of-the-art and
opportunities. Computers & Chemical Engineering 28 (67), 971983.
Singer, M., 2000. Forecasting policies for scheduling a stochastic due date job shop.
International Journal of Production Research 38 (15), 36233637.
Steenken, D., Vo, S., Stahlbock, R., 2004. Container terminal operation and
operations research A classication and literature review. OR Spectrum 26
(1), 349.
Tavakkoli-Moghaddam, R., Makui, A., Salahi, S., Bazzazi, M., Taheri, F., 2009. An
efcient algorithm for solving a new mathematical model for a quay crane
scheduling problem in container ports. Computers & Industrial Engineering 56
(1), 241248.
Vis, I.F.A., de Koster, R., 2003. Transshipment of containers at a container terminal:
An overview. European Journal of Operational Research 147 (1), 116.
Wang, F., Lim, A., 2007. A stochastic beam search for the berth allocation problem.
Decision Support Systems 42 (4), 21862196.
Zhang, C., Zheng, L., Zhang, Z., Shi, L., Armstrong, A.J., 2010. The allocation of berths
and quay cranes by using a sub-gradient optimization technique. Computers &
Industrial Engineering 58 (1), 4050.

You might also like