Original Article

Yard Crane Management by Simulation

and Optimisation

Pa s q u a l e L e g a t o * , P i e t ro C a n o n a c o a n d R i n a M . M a z z a

DEIS – Dipartimento di Elettronica, Informatica e Sistemistica, Università

della Calabria, I-87030 Arcavacata di Rende (CS), Cosenza, Italy.
E-mail: [email protected]
*Corresponding author.

A b s t r a c t In a terminal container the yard management is a critical and dynamic

component that requires governance and flexibility to adapt to and address even the
most complex issues arising out of yard planning and operations. With particular reference
to the assignment and deployment of Rubber Tired Gantry Cranes among yard blocks, an
optimisation model is proposed. Its purpose is to determine the block pairs between which
yard cranes will be transferred during the period under examination in order to satisfy the
crane capacity requirements and minimise the total cost for block matching and crane
activation. The use of this stand-alone optimisation approach is then extended by in-
troducing the architecture of an integrated framework, which includes both optimisation
and simulation techniques and is based on an all-inclusive queuing network representing
the main subsystems of a terminal. During scenario analysis, the framework is designed to
evaluate which policy for assigning yard cranes to yard blocks is best for meeting the
dynamic and constantly updating requirements of the yard subsystem. To this purpose,
some Ranking and Selection (R&S) techniques are considered. The ongoing numerical
experiments mean to demonstrate how a newly proposed R&S procedure is sufficiently
robust for use in practice.
Maritime Economics & Logistics (2009) 11, 36–57. doi:10.1057/mel.2008.23

Keywords: maritime container terminal; logistics; yard operations; discrete-

event simulation; combinatorial optimisation; ranking and selection

36 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
www.palgrave-journals.com/mel/
 Yard Crane Management

Introduction

According to the figures provided by UNCTAD (2004), containerised trade,

measured in twenty-foot equivalent units, is forecasted to grow by an average
annual rate of 5.32 per cent between 2003 and 2025.
As a result of this trend, the number of container terminals worldwide keep
increasing and competition has become both price driven and service driven. In
this market struggle among container terminals, the success of individual
companies depends upon their ability to fulfil customer demand with high
standard quality service and keep their operations lean; otherwise, they are
bound to lose clients to competition.
Operational efficiency is generally measured in terms of container
throughput, vessel/vehicle turn-a-round time and/or unproductive times. In
light of the literature investigation on container terminals, various activities can
turn out to be the choke point of efficiency in this very complex logistic system.
Among these difficulties, it appears that the management of yard cranes has
been quite experience-based and did not receive a great deal of attention until
the last decade. In particular, despite the popularity of rubber tired gantry
cranes (RTGCs) and the importance of their role in the yard operation, there still
has been a limited number of systematic studies on the RTGC deployment
problem. Zhang et al (2002) address the RTGC deployment problem; given the
forecasted workload of each block in each period of a day, they formulate a
mixed-integer programming model with the objective of finding the times and
routes of crane movements among blocks that minimise the total delayed
workload in the yard. Linn et al (2003) present an algorithm and a mathematical
model for the optimal yard crane deployment that minimises the total amount
of work overflow to the next period. Cheung et al (2002) propose an alternative
mixed-integer linear problem to Zhang et al (2002) for the inter-block crane
deployment problem by removing a restriction on crane movements and con-
sidering time periods whose length can be very small. Lim et al (2002) give a
bipartite graph representation model of the crane scheduling problem with
three different spatial constraints (non-crossing constraints, neighbourhood
constraints and job-separation constraints) and they provide algorithms based
on squeaky wheel optimisation with local search technique to solve it. Kim et al
(2006) use simulation to address dispatching and task sequencing rules for
container yards with multiple yard cranes in a block. However, they consider
dual rail-mounted gantry cranes (DRMGCs) where one DRMGC is larger than
the other and, thus, one can pass the other without interfering with the
movement of the other.
In this paper we continue focusing on the deployment of RTGCs among
yard blocks, but within a greater dynamic and constantly updating system
r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 37
 Canonaco et al

represented by an integrated queuing network for the terminal under study.

In the Problem Statement and Mathematical Formulation section, RTGC de-
ployment is described and followed by the formulation of an optimisation
model for the problem solution. The closing part of this section illustrates
the limits of adopting a similar stand-alone optimisation approach. In the
Simulation and Optimisation Model section attention is first directed for pre-
senting some (classic and new) Ranking and Selection (R&S) techniques
that are used to select the best system design when comparing different alter-
natives. An integrated simulation framework based on a queuing network
is then introduced to deal with stochastic and dynamic data and optimisa-
tion issues as well. Experiments and numerical results for selecting the best
policy when assigning RTGCs to yard blocks are reported in the Numerical
Results section. Conclusions and future work are both discussed in the
last section.

Problem Statement and Mathematical Formulation

The role of a yard in a maritime container terminal is to provide storage space

for containers from their import by truck to their export by vessel and vice versa,
and during their (pure) transhipment from vessel to vessel.
Our modelling and simulation efforts are referred to a pure transhipment
terminal, where a unique storage area is shared among a certain number of
shipping companies to which properly sized portions of the yard must be as-
signed to stack/retrieve container batches (that is, a set of containers sharing
some common properties). Such a situation really occurs at the port of Gioia
Tauro, in Southern Italy, where oceanic (mother) vessels from the Eastern Asia
maritime route discharge containers mainly addressed to other ports in
Northern Europe. Owing to the lack of adequate rail services and road infra-
structures, containers stored at Gioia Tauro are almost completely retrieved by
secondary vessels (both dedicated and common feeders).
Generally speaking, from an organisational point of view, a yard may be
divided into large areas called zones. In each zone, containers are stacked into
blocks. As shown in Figure 1, a block has: a number of lanes or rows ranging
from 6 to 13 (plus one for trucks) placed side by side; usually five containers in
height called tiers for each lane; 20 or more containers in length. A vertical
section of a block (for example, 5 tiers
6 lanes) is normally referred to as bay.
The number and which of the RTGCs are to be assigned to a block usually
depends on the expected daily workload and, therefore, on the total crane
capacity (measured in minutes) required to complete container stacking/
retrieval operations. To do so, RTGCs can be transferred from one block to
38 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

Figure 1: Organisation of a yard operated by cranes.

Figure 2: Crane transfer between yard blocks.

another: Figure 2 illustrates how RTGCs can travel between adjacent yard
blocks without any turning motion (for example, from block 1 to block 3) or by
changing lanes (for example, from block 1 to block 4). In the former case, crane
transfer can take about 10 min; in the latter, about five additional minutes are
required to perform 901 turns. These movements are exclusively referred to
inter-block (and not inter-zone) crane transfer.
The point to remark is that, in many terminals the current RTGC deploy-
ment is done manually and no formal procedures are used. When congestion
arises, very often an idle RTGC is randomly pulled out of a block and moved to
another block to perform operations and reduce work backlog.
To tackle these issues and support crane transfer during a planning period
(for example, of 6 hours), in the following we formulate and discuss a general
model belonging to the class of location and assignment problems illustrated in
Nemhauser and Wolsey (1988).
r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 39
 Canonaco et al

Let:

M number of blocks with additional crane capacity;

N number of blocks with insufficient crane capacity;
K maximum number of cranes allowed simultaneously in the same block;
cij cost for matching block i with block j for crane transfer;
fi cost to activate an idle crane in block i;
ai amount of crane capacity available in block i (expressed in time units);
bj amount of crane capacity requested by block j (expressed in time units).
For i ¼ 1, y, M and j ¼ 1, y, N, let the decision variables be:


1; if crane transfer occurs from block i to block j
xij ¼
0; otherwise
yi ¼number of cranes activated in block i

The purpose of the optimisation model is to determine the block pairs xij,
i ¼ 1, y, M, j ¼ 1, y, N between which cranes will be transferred during the
period under examination in order to satisfy the crane capacity requirements
and to minimise the total cost for block matching and crane activation. Under
the basic, practical assumption that a crane can be transferred out of a block at
the most once, the model is formulated as

X
M X
N X
M
Min cij xij þ fi yi ð1Þ
i¼1 j¼1 i¼1

subject to
X
N
xij pyi ; i ¼ 1; . . . ; M ð2Þ
j¼1

X
M
ai xij Xbj ; j ¼ 1; . . . ; N ð3Þ
i¼1

X
M
xij pK; j ¼ 1; . . . ; N ð4Þ
i¼1

xij 2 f0; 1g; yi 2 f0; 1; . . . ; Kg; i ¼ 1; . . . ; M; j ¼ 1; . . . ; N ð5Þ

Constraints (2) ensure that crane transfer can occur from block i to block
j only if cranes are initially assigned and activated in block i. Constraints (3)
40 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

specify that the crane capacity transferred from block i to block j must
satisfy the crane capacity requested in block j. Constraints (4) guarantee that
the number of cranes working simultaneously in block j cannot exceed K.
Constraints (5) are the constraints on the decision variables.
Much alike the other optimisation models already presented in the In-
troduction, in the above formulation some of the underlying modelling
assumptions are necessary to simplify a complex analysis, yet questionable.
Practically speaking, they are seldom met in the real world and can be
misleading when investigating the performance of different RTGC deployment
rules.
To begin with, in the (1)–(5) model all data are assumed to be deterministic
and static. However, on-the-field experience bears out that daily workloads vary
according to contingent requirements and/or circumstances, among which the
calling vessels’ ETA–expected time of arrival. Therefore, workloads are often
known and revised in just a matter of hours. In such a case, the crane de-
ployment problem should be solved again and a periodic update of the crane
deployment schedule should follow immediately.
In the second place, converting the daily workload in terms of crane time is
not as straightforward as assumed. As a matter of fact, the service time of a yard
crane is not deterministic and the amount of work done in a container block per
time period is not typically proportional to the number of cranes operating in
the block during that period. As the number of cranes in a block increases,
crane productivity may decrease because the space for cranes to manoeuvre
and work becomes more limited. In addition to this, other ‘irregular phenom-
ena’, such as crane starvation, blocking and/or failure, could and should
be taken into account. Any of the previous can occur during the ordinary
work cycle for container retrieval/stacking in the yard and cause delay upon
yard operations.
In the end, the solutions obtained from the above models are likely to
provide us only with partial guidance when modelling time-evolving systems
such as a storage yard in a container terminal. To this end, one may recognise
the usefulness of reproducing the system dynamics over multiple periods, under
some conditions of uncertainty because of randomly occurring events and
random duration of logistic activities.
Besides this, one may further observe that in the model the maximum
number of crane transfers in or out of a block is restricted to one per period
(with the purpose of reducing traffic blockages). On the other hand, if this
assumption is replaced with the different hypothesis that an RTGC can be
transferred more than once in a single period, then new constraints must
be added to the formulation, along with a new decision variable (that is,
number of crane transfers per period).
r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 41
 Canonaco et al

A Simulation and Optimisation Model

In an attempt to address the problems described in the previous section, there is

an emerging awareness that stand-alone optimisation techniques, especially in
industry-oriented research and development activities, need to interact with
other methodologies to deliver maximum results. In particular, we believe that
the integer programming model (1)–(5) may give a satisfactory, practical con-
tribution to the complex problem of both yard organisation and crane deploy-
ment for managing the work in progress at the yard. This can be accomplished
only by resorting to a combined approach in which discrete-event simulation
plays its role by representing and evaluating the different policies upon which
the work in progress may be carried out by cranes.
To evaluate whatever system (managed by complex decisions and policies)
via simulation experiments and statistical analysis of simulated output data,
one needs to define a suitable performance measure for mutual comparison. To
deal with this issue in our context, we choose to simulate the storage/retrieval
operations at several yard-blocks, as performed by deployed cranes, and
then measure the time to finish the work in progress at the yard block where
operations are last to be completed. The related sample mean statistic is
the performance measure adopted for ranking purposes and selection of the
best policy.
In the following, we first introduce the R&S procedure used for confidence-
based statistical comparisons among sample means, as discussed by Kim and
Nelson (2003), and later propose an integrated framework based on a joint
optimisation-simulation approach that provides a wider, and more nuanced,
lens to look through when planning a multi-period crane deployment schedule.

Selecting the best system

In Goldsman et al (2002), R&S is defined as a natural statistical technique used

to identify the best among a set of k competing designs, policies or system
configurations. This method is applicable when system parameters (for ex-
ample, allocated resources, scheduling policies and so on) are discrete and the
number of competing designs is both discrete and small (for example, k ¼ 2,
y, 20). The method is applied once a sample mean for a measure of system
performance has been constructed from simulated or real data. At the basis of
the method there is the evaluation of the sample variance associated to each
sample mean to be compared. The smaller the sample variance is, the more one
is confident that a sample mean is better (smaller or higher) than the other
and, therefore, that the related system is to be preferred. As usual in classical
42 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

statistics, the normality assumption of the sample means is of great help in

determining the confidence level of the selection process.
Nearly all of the research work in R&S can be classified into the following
general approaches:

K indifference zone procedures, where either the best or whatever solution

evaluated within a fixed distance from the best can be selected, with a user-
specified probability;
K subset selection procedures, which aim at producing a subset of (small)
random size that contains the best system, with a user-specified probability;
K other approaches.
We will focus on the indifference-zone formulation.
A certain number of observations of any given performance index are taken
from each of the k systems and used to generate sample mean statistics for
selecting the best system. However, because the above observations are random
variates, there are no guarantees of selecting the best design, despite it being
truly representative of the best system configuration. On the other hand, if the
best design is selected, then a correct selection (CS) is said to have been made.
Let:

k the number of alternative designs (i ¼ 1, y, k);

n the number of observations (j ¼ 1, y, n);
m[1]p?pm[k] the unknown, ordered, k expected values of a performance
measure;
m1, y, mk the k estimated values of a performance measure;
d the indifference-zone, a positive real number fixed by the user;
P* the confidence level fixed by the user;
P{CS} the probability of correct selection,
in a maximisation problem the probability of performing a CS with at least
probability P* is

PfCSg¼Pf^ mi 8i 6¼ kj^
^ mk 4^ ^i XdgXP
mk  m

The probability of CS was first computed in Rinott (1978) by resorting to

numerical integration under the hypothesis of normality of the statistics in-
volved. The numerical solution requires the use of a parameter (Rinott’s con-
stant h), whose value depends on the sample size, the number of alternative
system designs and the confidence level. Numerical values for h are tabled in
Wilcox (1984).
r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 43
 Canonaco et al

A general scheme of a two-stage R&S procedure is addressed in the fol-

lowing and a simple numerical example is given in Appendix.
Stage 1

Step 0 Select the confidence level, indifference-zone parameter and first-stage

sample size, respectively 1a, d>0 and n0X2.
Step 1 Obtain Rinott’s constant h ¼ h(n0, k, 1a).
Step 2 Obtain n0 observations Xij, j ¼ 1, 2, y, n0 from each system i ¼ 1, 2,
y, k.
Step 3 Compute sample mean and variance for each system i ¼ 1, 2, y, k,
respectively X̄i(n0), Si2(n0).
Stage 2

Step 4 Calculate the total no. of observations for each system i ¼ 1, 2, y, k

Ni ¼ max{n0, Jh2Si2/d2n}
Step 5 If n0Xmaxi Ni then stop the procedure and select the system with the
‘best’ X̄i(n0); otherwise, take Nin0 additional observations Xi;n0 þ1 ;
Xi;n0 þ2 ; . . . ; Xi;Ni from each system i for which Ni>n0.
Step 6 Select the configuration with the ‘best’ overall sample mean X̄i(Ni).
The number of observations required to select the best system design with
P{CS}XP* is a major impact factor on procedure performance. As shown in
step 4 of the general scheme presented above, this amount mostly depends on
sample variance and, thus, on how the sample mean is computed earlier. To this
end, different methods use different approaches. In pioneer two-stage R&S
procedures, Rinott (1978) uses a classic sample mean, while Dudewicz and
Dalal (1975) use a weighted sample mean during the second stage. In more
recent and advanced multi-step procedures, Kim and Nelson (2001) take a
single observation from the systems still in play and then choose whether or not
to cease sampling from the systems that no longer appear to be competitive;
Chen and Kelton (2000) take into account both the sample variances and the
differences between sample means when determining the sample sizes.
A moving-average sample mean is investigated in our procedure. Our ap-
proach was inspired by Law and Kelton’s (2000) description of the graphical
technique used by Welch to deal with the problem of the initial transient or the
start-up problem. Our procedure builds an ‘artificial’ process with the same
mean as the output observations of interest, but with a smaller variance. It then
smoothes out the high-frequency random deviations by introducing a moving-
average with a moving window of w values. In brief, the underlying purpose of
the method is to give an unbiased estimator that has lower variance than other
unbiased estimators for all possible values of the system performance measure
44 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

under examination. When found, a similar estimator allows us to choose the

best system at a lower computational cost.
Practically, n output observations of a performance index are organised
into b groups (one per simulation run), each of size m (mb1), and then used
to compute an average value of the ith observation across these groups (see
Figure 3)

1 X
b
Y
i ¼ Yji
b j¼1

The above values, Y
1, Y
2, y, Y
m, are then used to define the moving-
average Y
i(w) with a window length of w as follows:
Pw


s¼w Yiþs
Y
i ðwÞ ¼ ; i ¼ w þ 1; . . . ; m  w
2w þ 1

For example, to compute the moving average Ȳ2(1), one must first compute
Ȳ2 (in our context, the average completion time for batch no. 2 across the b
simulation runs), and then average this result out with the w values to its left
and the w values to its right. In this specific case, if w ¼ 1, then
Ȳ2(1) ¼ (Ȳ1 þ Ȳ2 þ Ȳ3)/3 and, in general, Ȳi(1) ¼ (Ȳi1 þ Ȳi þ Ȳi þ 1)/3 for i ¼ 2,
y, m1.
To conclude, the Ȳi(w)’s become the observations that will be used when
computing the sample mean and variance for each system, as required by steps
3 and 6 of the R&S procedure.

Figure 3: The moving-average-based R&S approach with w ¼ 1.

r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 45
 Canonaco et al

An integrated framework

All of the simulation and optimisation techniques earlier described are in-
corporated in the integrated framework illustrated in Figure 4, which consists of
four interacting modules, each with a specific focus on one of the key com-
ponents of the container terminal.
The (three) discrete-event simulation modules provide all the data that are
not assumed to be deterministic. In more detail:

K The Vessel Arrival Simulation Module generates, according to specific

patterns or proper distribution functions, the periodic vessel calls to the
port, along with vessel characteristics – these include physical characteristics
(for example, length and draft), operational settings (for example, GCP –
gross crane productivity and crane intensity) and requirements (for example,
tugs and/or pilots), among which the number of container moves and cover
moves that represent the vessel workload.
K The Channel Manoeuvring & Berthing Simulation Module decides vessel
access in and out of the port by implementing the admission policy currently
operated by the berth management – company procedures for vessel
entrance/exit account for berth availability and assignment prevalently
based on resource guarantee and priority mechanisms.

Figure 4: The architecture of the integrated crane deployment framework.

46 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

K The Operations Simulation Module represents the heart of the pure

transhipment terminal system by reproducing container discharge/loading
activities on the quayside performed by quay cranes, while shuttle vehicles
transport containers to and from the yard area where their stacking/retrieval
is carried out – the dynamics of this work cycle illustrated in Figure 5 are
strongly dependent on the crane service times in both the quay and yard
subsystems.
Owing to its crucial role in the network and for the purposes of the present
work, from here on, attention is drawn to the so-called yard side. Transfer
cranes are first planned or ‘activated’ for the initiation of yard operations, then
assigned to the yard blocks in which these operations are to be performed.
However, not every block receives a crane immediately. Therefore, periodically
and not necessarily only at the end of a work shift (for example, every 6 hours),
after stacking/retrieval activities are completed in some blocks, available cranes
are transferred to meet the requests of other blocks with insufficient crane
capacity. The queuing network model in Figure 6 represents this transfer, but in
order to define the block pairs between which cranes are to be transferred
during each period, we use the alternative optimisation model devised for this
aim in the Problem Statement and Mathematical Formulation section and
embody it in the framework’s Optimisation Module.

Figure 5: The integrated terminal network model.

r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 47
 Canonaco et al

To solve this yard queuing network model, a possibility could lie in as-
suming that the arrival process of container stacking/retrieval requests follows
a Poisson process with the expected arrival rate of l and then by representing
container service operations with an M/G/1 queuing model as proposed by
Kim and Kim (1999) under the assumption that the queuing discipline is
work-conserving, as defined in Heyman and Sobel (1982). As suggested by the
formulas in Table 1, the computational effort required to determine the per-
formance measures of interest would be quite modest. However, the modelling

Figure 6: The yard queuing network model.

Table 1: Some formulas for performance measures in the M/G/1

Parameters and performance measures Formulas

Average arrival rate l
Average service rate m=1/{E(Td)+E(Tr)+E(Tt)}a
Traffic intensity r=l/m
Variance of the service time V=V(Tc)(Tc=Td+Tr+Tt)
Expected number of vehicles in the system E(Ls)=r+(l2V+r2/2(1r))
Expected waiting time in the system E(Ws)=E(Ls)/l
a
Td is the handling time for a target container, Tt is the travel time between two bays in the same yard zone
and Tr is the re-handling time to pick up the target container.

48 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

capability provided by this basic queuing system is too simple to fully account
for a correct representation of the complex nature of the real system.
Furthermore, the above assumptions are not realistic. The arrival process of
container requests in the terminal yard is likely not to follow a Poisson process,
nor must stacking/retrieval operations in a block be performed by only one
fixed crane-server. As a matter of fact, in our model we assume that as many as
K cranes can be fixed to work simultaneously in a block and this number can
vary dynamically during yard operations. These limitations necessarily lead to
solving the model in Figure 6 by discrete-event simulation techniques.
Simulation allows to investigate many views of the same system: in our
early research on scenario analysis at a container terminal in Legato and Mazza
(2001), simulation was used to focus on the berth subsystem and design a
specialised quantitative model for bottleneck analysis, operations management
and resource optimisation. Today, the potential insight provided by the in-
tegrated simulation framework is even greater. Its open-module architecture
allows the possibility to incorporate additional optimisation-guided rules of
work and low-level operational policies developed for the other terminal sub-
systems – for example, the dynamic allocation of straddle carriers to a specific
quay crane and/or scheduling different holds of the same vessel to quay cranes
as reported in Canonaco et al (2008).
With the current framework configuration, we will focus on the new op-
erational scenarios generated by five alternative management policies for as-
signing yard cranes to yard blocks and accounting for order, times and routes of
the crane transfer. The objective is to select, by way of our R&S procedure, the
policy which allows us to minimise the maximum average time to complete
stacking/retrieval operations of suitable batches of containers in the yard. In a
pure transhipment container terminal, each batch of containers may correspond
to a specific group of containers that arrive with the same vessel and need to be
stacked in the same yard block or, vice versa, depart with the same vessel and,
therefore, need to be retrieved from the same yard block.

Numerical Results

The simulation model and experiments, as well as the robustness of the R&S
procedure proposed have been verified and validated in compliance with the
classes of techniques conventionally used for this purpose in Banks et al (2001).
They have also been carefully designed and conducted within Rockwell’s Arena
(2000), one of the major general-purpose discrete-event simulation packages
available on the market. The two Arena flow diagrams used to model the core
logistic processes at the heart of all yard crane activity (see Figures 7 and 8) are
r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 49
 Canonaco et al

Figure 7: A flow diagram for the crane monitoring process in Arena.

Figure 8: A flow diagram for container generation and retrieval/stacking in Arena.

composed of standard and user-defined blocks, which allow us to account for

non-customary policies and procedures. The former diagram periodically
monitors the status of the available cranes and schedules their future use ac-
cording to the idle, busy or in-transfer condition of each resource. In the specific
case in hand, crane transfer can occur in agreement with one of the following
five alternative policies or system configurations:

K crane transfer to a random yard block;

K crane transfer to the closest yard block;
K crane transfer to the yard block with greatest priority;
K crane transfer to the yard block with shortest workload;
K crane transfer to the yard block with greatest workload.
Observe that results obtained when applying the first option are conceived
as a lower bound reference in comparison with the results of the other policies.
The latter Arena diagram generates multiple job requests consisting of
batches of containers that require retrieval if scheduled for departure from the
yard and/or stacking if planned for storage on the yard. Although a pure
transhipment terminal is under examination, the statistical model currently
used to represent the above stacking/retrieval requests may be extended to
include import and export flows as well. In particular, this is possible by in-
troducing different random arrival/departure processes whose purpose is to
represent import, export and transhipment flows.
50 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

During numerical experimentation performed on a personal computer

equipped with a 1.73 GHz Intel Pentium M processor and 1 Gb of RAM, in order
to favour an easier and less time-consuming data input and output for every
scenario, separate VBA (Visual Basics for Applications) windows have been
used to obtain an integrated interaction with Arena and Microsofts Excel 2002.
For instance, a great number of experiments have been rapidly performed
(approximately 2 seconds per run) by limiting data specification to the fields
portrayed in Figures 9 and 10 and are listed below:

K average workload per yard block;

K number of available RTGCs;
K policy for RTGC assignment to yard block and, if applicable, priority
specification;
K yard layout.
For the time being, four different classes of problems have been defined by
providing combinations of smoothly varying container traffic intensity (low,

Figure 9: The VB window for general data input.

r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 51
 Canonaco et al

Figure 10: The VB window for priority specification among yard blocks.

medium and high) on the four yard blocks considered and the time (low and
high) required, and therefore the distance to cover depending on the yard size,
when transferring an idle crane from one yard block to another bearing in-
sufficient crane capacity (see Table 2).
For these problems no transient suppression schemes have been used for
the following reasons. First of all, given a certain yard definition (that is,
number and position of yard blocks), terminating simulations have been per-
formed in order to minimise the average maximum time to complete a container
batch processing under different crane transfer policies. Therefore, there has
been no need to evaluate steady-state parameters. Secondly, the ‘requests’ to
stack/retrieve a batch of containers have been assumed to occur according to a
renewal process, as defined in Heyman and Sobel (1982).
The nonexistence of the warm-up period for the performance index of in-
terest is illustrated in Figure 11.
This stated, our procedure (Our) has been applied and compared with
Rinott’s (R) (1978) and Dudewicz and Dalal’s (D&D) (1975) on each class of
problems to verify which of the five assignment policies is likely to be the best.
This comparison has also provided a measure of how our procedure responds in
terms of the total number of observations (nobs) to be taken from each alter-
native configuration in order to obtain the statistics used in selecting the best
52 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

Table 2: Classes of problems tested in the yard simulation

Problems Description
Class 1 Low traffic, high transfer time
Class 2 Medium traffic, high transfer time
Class 3 Medium traffic, low transfer time
Class 4 High traffic, low transfer time

Figure 11: Example of the nonexistence of the warm-up period.

system design with a least probability of P*. As demonstrated by the results

reported in Table 3, in the worst case (that is, for Class 1 and 3 problems) the
other procedures have reached the same correct result but with an average of 5
per cent reduction in the number of total observations drawn from the two-
phase sampling procedure; whereas in the best case (that is, for Class 2 and 4
problems), these procedures have been outperformed by ours by an average of
40 per cent.
More precisely, let us consider a Class 2 scenario and the following
specific settings required by the R&S algorithms: d ¼ 5 min, P* ¼ 0.90, h ¼ 3.317,
n0 ¼10. All three procedures select the ‘closest’ policy as likely being the best.
Observe that, when dealing with cases like these in which the workload and
crane transfer time are approximately of the same order, because of the
yard size and layout, this particular result is expected to emerge and, therefore,
it is also used as a validation sample. However, our procedure arrives to this
conclusion with only 80 observations, whereas Rinott and Dudewicz and
r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 53
 Canonaco et al

Table 3: Classes of problems tested in the yard simulation

Problems R D&D Our

Class 1 5% 5% nobs
Class 2 nobs nobs 40%
Class 3 5% 5% nobs
Class 4 nobs nobs 40%

Table 4: Results of the R&S procedures for a Class 2 problem

Procedure Nobs Max average container-batch completion time (min)

Shortest Greatest Closest Priority Random

R 151 97.21 93.67 79.76 101.57 95.52
D&D 151 97.19 93.41 82.32 101.01 90.95
Our 80 95.54 91.73 79.51 103.00 92.72

Dalal need more than 150 observations for the same accomplishment as shown
in Table 4.

C o n c l u s i o n s a n d F u t u r e Wo r k

We have focused on the efficient planning and use of the terminal yard to
increase the competitiveness of a container terminal by improving the internal
organisation and operations management. With particular reference to the
assignment and deployment of yard cranes, an optimisation model supporting
crane transfer has been proposed and examined. To enhance the use of
stand-alone approaches, an integrated framework based on both optimisation
and simulation techniques has been presented. Herein, attention has been
drawn to the queuing model representing the yard subsystem and to
selecting the best policy for assigning cranes to the yard blocks. Discrete-event
simulation and a new moving-average based R&S technique have been
jointly proposed as the basis for problem solution. Validation and numerical
experiments conducted in Rockwell’s Arena are encouraging, especially when
considering the limited computation effort required by the R&S procedure and
the related improvement in efficiency. Our future direction of research is to
perform sensitivity analysis on our R&S procedure for different lengths of the
moving-average-window and then investigate procedure extension from 2 to n
54 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

stages in comparison with the latest multi-stage procedures appearing in the

literature.

References

Arena – Version 5.00.02 Copyright r 2000 Rockwell Software Inc.

Banks, J., Carson, J.S., Nelson, B.L. and Nicol, D.M. (2001) Discrete-Event System Simulation, 3rd
edn., Upper Saddle River, NJ: Prentice-Hall.
Canonaco, P., Legato, P., Musmanno, R. and Mazza, R.M. (2008) A queuing network model
for the management of berth crane operations. Computers and Operations Research 35:
2432–2446.
Chen, E.J. and Kelton, W.D. (2000) An Enhanced Two-Stage Selection Procedure. In: J.A. Joines,
R.R. Barton, K. Kang and P.A. Fishwick (eds.) Proceedings of the 2000 Winter Simulation
Conference. pp. 727–735.
Cheung, R.K., Li, C.-L. and Lin, W. (2002) Interblock crane deployment in container terminals.
Transportation Science 36(1): 79–93.
Dudewicz, E.J. and Dalal, S.R. (1975) Allocation of observations in ranking and selection with
unequal variances. Sankhya B7: 28–78.
Goldsman, D., Kim, S.-H., Marshall, W.S. and Nelson, B.L. (2002) Ranking and selection for steady-
state simulation: Procedures and perspectives. INFORMS Journal on Computing 14(1): 2–19.
Heyman, D.P. and Sobel, M.S. (1982) Stochastic Models in Operations Research, Volume I: Stochastic
Processes and Operating Characteristics. New York: McGraw-Hill.
Kim, K.H. and Kim, H.B. (1999) Segregating space allocation models for container inventories in
port container terminals. International Journal of Production Economics 59: 415–423.
Kim, S.-H. and Nelson, B.L. (2001) A fully sequential procedure for indifference-zone selection in
simulation. ACM TOMACS 11: 251–273.
Kim, S.-H. and Nelson, B.L. (2003) Selecting the Best System: Theory and Methods. In: S. Chick,
P.J. Sánchez, D. Ferrin, and D.J. Morrice (eds.) Proceedings of the 2003 Winter Simulation
Conference. pp. 101–112.
Kim, K.H., Lee, S.-J., Park, Y.-M., Yang, C.H. and Bae, J.-W. (2006) Dispatching yard cranes in port
container terminals. TRB – Transportation Research Board Annual Meeting.
Law, A.M. and Kelton, W.D. (2000) Simulation Modeling and Analysis, 3rd edn., New York:
McGraw-Hill.
Legato, P. and Mazza, R.M. (2001) Berth planning and resources optimisation at a container terminal
via discrete event simulation. European Journal of Operational Research 133(3): 537–547.
Lim, A., Rodrigues, B. and Zhu, Y. (2002) Crane Scheduling Using Squeaky Wheel Optimization with
Local Search. Proceedings of the 4th Asia-Pacific Conference on Simulated Evolution and Learning,
Singapore (SEAL 2002).
Linn, R., Liu, J., Wan, Y., Zhang, C. and Murty, K.G. (2003) Rubber tired gantry crane deployment for
container yard operation. Computers & Industrial Engineering 45: 429–442.
Nemhauser, G.L. and Wolsey, L.A. (1988) Integer and Combinatorial Optimization. New York: John
Wiley.
Rinott, Y. (1978) On two-stage selection procedures and related probability-inequalities. Commu-
nications in Statistics – Theory and Methods 7(8): 799–811.
UNCTAD Transport Newsletter No. 24. (2004). UNCTAD/WEB/SDTE/TLB/2004/2. United Nations
Conference on Trade and Development (UNCTAD) Transport Section, Trade Logistics Branch,
Geneva, www.UNCTAD.org.
Wilcox, R.R. (1984) A table for Rinott’s selection procedure. Journal of Quality Technology 16(2):
97–100.

r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 55
 Canonaco et al

Zhang, C., Wan, Y., Liu, J. and Linn, R.J. (2002) Dynamic crane deployment in container storage
yards. Transportation Research Part B 36: 537–555.

Appendix

In a simple service system, an operator can perform customer service according

to three (k ¼ 3) policies. A two-stage R&S procedure is used to select with
probability 0.9, the policy corresponding to the best (minimum) estimate of
the system performance index m or whatever of the three within 1 unit (d ¼ 1)
of the best.
Stage 1

Step 0 1a ¼0.90, d ¼ 1 minute and n0 ¼10.

Step 1 h ¼ h(n0, k, 1a) ¼ h(10, 3, 0.90) ¼ 2.587.
Step 2 n0 ¼10 observations Xij, j ¼ 1, 2, y, 10 from the system with policy
i ¼ 1, 2, 3

System/observation j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8 j=9 j=10
Policy i=1 16.6 17.6 19.6 17.7 18.0 17.7 17.7 17.3 17.6 15.8
Policy i=2 18.2 18.7 17.3 18.1 16.8 18.6 17.0 19.3 16.4 18.9
Policy i=3 19.8 17.5 18.7 19.1 18.4 19.9 19.1 17.5 17.3 16.0

Step 3 Sample mean and variance for the system with policy i ¼ 1, 2, 3

System/estimate X¯i(10) Si2 (10)

Policy i=1 17.6 0.96
Policy i=2 17.9 0.99
Policy i=3 18.3 1.57

Stage 2

Step 4 Total no. of observations Ni ¼ max{n0, Jh2Si2/d2n} for the system with
policy i ¼ 1, 2, 3
System/estimate Ni
Policy i=1 Ni=max(10, 7)=10
Policy i=2 Ni=max(10, 7)=10
Policy i=3 Ni=max(10, 17)=17

56 r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57
 Yard Crane Management

Step 5 10Xmaxi Ni ¼ max(10, 10, 17) ¼ 17 is not verified -Nin0 additional

observations from the system with policy i for which Ni>10

System/observation j=11 j=12 j=13 j=14 j=15 j=16 j=17

Policy i=1 — — — — — — —
Policy i=2 — — — — — — —
Policy i=3 21.1 18.7 17.6 19.9 17.9 18.6 20.75

Step 6 Policy i ¼ 1 has the ‘best’ overall sample mean estimate X
2(10) ¼ 17.6
for performance index m. Policy i ¼ 2 can also be chosen because it is
within 1 unit of the best (17.9); either choice is said to be ‘statistically
indifferent’. This is not the case for policy i ¼ 3 (18.7).

System/estimate X¯i(Ni)
Policy i=1 17.6
Policy i=2 17.9
Policy i=3 18.7

r 2009 Palgrave Macmillan 1479-2931 Maritime Economics & Logistics Vol. 11, 1, 36–57 57