Original Article

Prediction of arrival times and human

resources allocation for container terminal

G i a n f r a n c o Fa n c e l l o a , C l a u d i a P a n i a , M a r c o P i s a n o a ,
P a t r i z i a S e r r a a , P a o l a Z u d d a s b a n d P a o l o Fa d d a a

a
CIREM: Centro Interuniversitario di Ricerche Economiche e Mobilitá,
University of Cagliari, Italy.
E-mails: [email protected], [email protected],
[email protected], [email protected], [email protected]
b
Dipartimento di Ingegneria del Territorio (DIT), University of Cagliari, Italy.
E-mail: [email protected]

A b s t r a c t Increasing competition in the container shipping sector has meant that

terminals are having to equip themselves with increasingly accurate analytical and
governance tools. A transhipment terminal is an extremely complex system in terms of both
organisation and management. Added to the uncertainty surrounding ships’ arrival time in
port and the costs resulting from over-underestimation of resources is the large number of
constraints and variables involved in port activities. Predicting ships delays in advance
means that the relative demand for each shift can be determined with greater accuracy, and
the basic resources then allocated to satisfy that demand. To this end, in this article we
propose two algorithms: a dynamic learning predictive algorithm based on neural networks
and an optimisation algorithm for resource allocation. The use of these two algorithms
permits on the one hand to reduce the uncertainty interval surrounding ships’ arrival in
port, ensuring that human resources can be planned around just two shifts. On the other
hand, operators can be optimally allocated for the entire workday, taking into account
actual demand and operations of the terminal. Moreover, as these algorithms are based on
general variables they can be applied to any transhipment terminal. Future integration of
the two models within a broader decision support system will provide an important support
tool for planners for fast, flexible planning of the terminal’s operations management.

Maritime Economics & Logistics (2011) 13, 142–173. doi:10.1057/mel.2011.3

Keywords: decision support system; containership arrivals prediction; container

handling optimisation; neural network; resource allocation; container hub
planning

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
www.palgrave-journals.com/mel/
 Prediction of arrival times and human resources allocation

Introduction

A seaport terminal is a complex system within which a broad variety of opera-

tions is carried out involving a wide array of resources that need to interact in a
24-hour operating cycle. The terminal operator aims at maximising efficiency,
that is achieve the largest number of movements at the least possible cost.
The operations carried out in a terminal are characterised by a large number of
variables and constraints that increase their complexity. Several factors, not always
easy to control, can affect the quality of services provided and overall efficiency.
Demand uncertainty further complicates the task of planners and as a
result, the effectiveness of planning itself. To be able to predict the delay of
ships, operators need to know the actual time of arrival in port so as to be able
to determine more accurately the demand for each work shift. In this way, the
resources required for meeting that demand can be allocated with greater cer-
tainty, avoiding under- or overmanning at the planning stage.
This is a major issue in a port system where the cost of manpower is
relatively high. The efficiency of a seaport terminal cannot therefore be divorced
from optimising human resources management, an essential factor especially in
systems having a low level of automation. Human resources management is a
very complex issue, terminals have to be manned 24 hours a day, considering
that numerous workers interact with each other and that the operations are
intrinsically random.
Additionally, because the different operations carried out in terminals are
strongly interrelated, there is a need not only to maximise efficiency, but also to
ensure proper coordination, hence to solve integrated decision-making pro-
blems. However, the decision-making processes involved in terminal operations
can sometimes be so complex that they become unmanageable without the
support of suitable methodological tools.
The work presented here forms part of a broader research project aimed at
developing a decision support system (DSS), comprising different but strongly
interconnected modules, designed to assist terminal planning.
The DSS incorporates four modules:

K forecasting module;
K optimal human resources allocation module;
K optimal equipment allocation module;
K maintenance module.
This article describes the studies conducted to date for the first two modules.
Using forecasting algorithms, the first module provides an answer to the problem
of demand uncertainty. The second module, on the other hand, uses output from
the first module as input data for determining optimal manpower allocation.
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 143
 Fancello et al

Though traditional statistical tools are currently used as planning support,

day-to-day management of terminal operations still relies heavily on planner
experience. Therefore, tools need to be devised that provide planners with
objective and analytical answers. Of the most effective tools available for ac-
curate forecasting we opted for a dynamic learning model based on neural
networks, able to solve information uncertainty on inbound flows. The main
reasons for choosing this type of model lie in the irregular time series for ship
delays and the complexity of the phenomenon being studied.
On the other hand, the complexity of daily resources planning requires
tools that take variables and constraints into account and deliver the best so-
lution without upsetting the planning arrangements at a higher level.
As can be seen, a tool is required that is capable of preventive activity
planning and rapid rescheduling in the event of disruptions caused by unfore-
seen events. In the light of these considerations, an optimising approach of the
resources allocation type appears to be the most suitable as, among other
things, it also matches the requirements for efficiency and non-intrusiveness.
The use of an integrated tool such as DSS, ensures fast and flexible plan-
ning of terminal operations, while achieving a substantial reduction in costs,
greater efficiency and as a result enhanced terminal competitiveness.

State of the Art

State of the art: Forecasting module

Demand uncertainty plays a decisive role in terminal organisation. Attempts

have been made to develop an innovative model for short-term forecasting, the
most crucial time period in that it concerns updated information for resources
optimisation.
On a daily basis, there remains the uncertainty of ships’ arrival time
in port. Knowing the extent of this delay is essential for planning at least over a
24-hour period. Several approaches can be adopted to address this problem.
Comparison of traditional statistical methods with neural networks used for
predictive purposes conducted in 1998 (Zhang et al, 1998) showed that neural
networks perform better when major irregularities are present, for dis-
continuous time series and for short time series. Furthermore, their flexibility,
nonlinearity and arbitrary function mapping, makes neural networks useful
predictive tools.
Celik uses a neural network for modelling freight traffic distribution over
short time horizons. The results were an improvement over those obtained
using a gravity model developed by the same author (Celik, 2004).
144 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

The work that most closely matches the application concerned with here
was published by Bilegan and colleagues and aimed at predicting the number
and type of railway wagons required daily at the land interface of an intermodal
container terminal in Canada (Bilegan et al, 2006). Again promising results
were obtained with a neural network model, specified using an iterative trial-
and-error procedure with the aid of javaNNS software.
Carbonneau and colleagues conducted a comparative analysis of dynamic
learning (neural networks, recurrent neural networks and support vector ma-
chines) and traditional methods (naı̈ve forecasting, trend, moving average,
multiple linear regression and time series models) applied to logistics. The
findings showed that neural networks outperformed the other techniques
(Carbonneau et al, 2007).
On the other hand, two main limitations for neural networks are reported in
the literature. First of all, no definite methodology exists for correctly specifying
the model, which explains why the iterative trial-and-error procedure is more
frequently used, and second, the result obtained is strongly data dependent and
dependent on model specification.
Closer examination of the literature on NN model specification, revealed
that the major focus is on choice of input variables as this influences network
topology and, as a result, computational properties, generalisation ability and
above all, prediction efficiency. For this reason, in the present study attention
has also been focused on data pre-processing.
The methods used for choosing the input variables are classified in the
literature into two broad groups:

K Model-free methods: whereby using chiefly statistical techniques, it is

possible to select/reduce input variables on the basis of their significance
and/or importance. This process takes place regardless of the neural network
model used (La Rocca and Perna, 2005; Papadonkontantakis et al, 2005).
These models are used in the present study.
K Model-based methods: whereby input variables are chosen on the basis of the
predictive efficiency of the neural network. Using an iterative procedure, a
different set of input variables is selected for each iteration and at the end of
the procedure the best performing set is chosen (Bowden et al, 2004; Li and
Peng, 2006; Saxèn and Petterson, 2006).

Furthermore, Zhang et al (1998) point out that apart from the number of input
and output nodes, the factors that most influence network performance, and
hence the results obtained are
K variable normalisation;
K choice of learning algorithm and related parameters;
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 145
 Fancello et al

K choice of training and validation sets;

K choice of number of hidden layers and nodes.
Determining the number of hidden layers and nodes requires special attention
in order to avoid overfitting, which results in the loss of network generalisation
ability. Thus network complexity needs to be checked.

State of the art: Optimal human resources allocation module

Optimum allocation of human resources is also a major challenge in seaport

terminals. Current research focuses on planning at the tactical and operational
levels, developing optimisation models. These models overcome the limi-
tations of simulation models, which are impracticable when dealing with a large
number of alternatives, require long development and validation times and are
not generalisable to other settings.
A plethora of literature exists on shift scheduling problems, proposing
planning models that have however been developed for sector-specific appli-
cations and therefore differ from one to another. Some universally applicable
models have also been developed (Beasley and Cao, 1998) but they perform
poorly in applications requiring flexible and rapid responses as is the case of
port terminals, where the number of ships to be handled in a single shift is far
from certain and not known well in advance.
Essentially three approaches are proposed in the literature for addressing
daily scheduling of human resources in a container terminal:

K Scheduling: An algorithm widely used in the literature creates a work queue

system for each terminal staff member in order to reduce overall delays to a
minimum (Hartmann, 2004). Meersmans and Wagelmans (2001) address the
problem of integrated equipment scheduling for an automated terminal
adopting an optimising ‘beam search’ approach; Park and Kim (2003) on the
other hand propose a specific scheduling algorithm for the berthing plan and
quayside cranes.
K Shared platforms (Dell’Olmo and Lulli, 2004): The terminal is modelled using
a network of complex platforms, each having an engineered and an operating
capacity. The problem consists of a kind of generalised scheduling among
platforms.
K Resource allocation (Gaudioso et al, 1999; Gambardella et al, 2001; Imai et al,
2001; Legato and Monaco, 2004; Cordeau et al, 2005): Gambardella for
instance, formulates a network design problem for determining, for a 24-hour
time horizon, the number and type of resources required per shift to satisfy
container handling demand. Gaudioso has provided a major contribution to
human resource allocation. He defines two planning levels in transhipment
146 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

terminals, monthly and daily. For monthly scheduling, a sequence of work

and rest shifts is prepared for each worker that is then refined at the daily
level. Legato and Monaco contributed to furthering progress in this direction.
Drawing on the work of Gaudioso, they developed a model for daily human
resources allocation in a transhipment terminal, which forms the basis of the
optimisation model proposed here.
The resource allocation approach appears to be the most suitable for those
cases where it is required to make the best use of a limited amount of resources
in order to maximise benefits. This approach has been adopted by a number of
authors and appears to best match the characteristics of efficiency and non-
intrusiveness on the system.

G e n e r a l L a y o u t o f a C o n t a i n e r Te r m i n a l

Container terminals usually work four 6-hour shifts around the clock, every day
as follows:
K first shift: 01:00–07:00;
K second shift: 07:00–13:00;
K third shift: 13:00–19:00;
K fourth shift: 19:00–01:00.
Two-level personnel scheduling is used, monthly and daily (Monaco et al,
2008). Monthly scheduling ensures personnel is available for each working day,
in conformity with shift arrangements as well as contractual obligations and
labour regulations. Because of the uncertainty of demand for that period, the
schedule for each day assigns ‘fixed’ workers to one specific shift and ‘flexible’
workers to a shift to be decided during daily scheduling, once demand has been
determined with greater certainty. Contractual terms for flexible shifts specify
that a flexible worker will be assigned to a shift with only 24 hours advance
notice, though the monthly schedule establishes the days on which he/she will
work. Once the daily schedule is prepared, only the distribution of workers
across shifts can be altered, the total number of workers per shift remaining
unchanged. In the event, demand requires additional manpower then external
workers are hired, usually used by stevedoring companies for ground opera-
tions and truck trailer driving. Clearly the cost of hiring external manpower is
higher and should be avoided as much as possible.
Clearly both scheduling levels are characterised not only by process com-
plexity but also by temporal fragmentation of the information and the longer
the scheduling horizon, the greater the information uncertainty. Under these
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 147
 Fancello et al

circumstances, the problem of predicting ship delays has negative effects on

planning efficiency. Poor forecasting, combined with the possibility of unfore-
seen events, translates into higher costs because of the additional workload
involved. The ability to predict ship delays means knowing their exact time of
arrival in port so that the demand for each shift can be determined more ac-
curately. Once this information is known, human resources can be allocated
accordingly, avoiding under- or overmanning.

Methodology

Methodology: Forecasting module

For predicting ships arrival at the terminal, the study examined the calibration
of a neural network-based simulation model.
After a review of the pertinent literature, it was decided to opt for neural
networks basically for two aspects associated with the phenomenon under
study (Haykin, 1994; Zhang et al, 1998; Potvin and Smith, 2001), namely its
complexity and the irregular time series of arrivals.
The main steps involved in model structuring are listed below:

(a) choice of predictive approach;

(b) choice of paradigm;
(c) choice of input variables;
(d) variable normalisation;
(e) choice of network architecture;
(f) choice of number of hidden layers and nodes;
(g) choice of learning algorithm and related parameters;
(h) interpretation of results.

The study was initially based on a model, developed at Cagliari University

(Pisano, 2008), with the following characteristics:

(a) A predictive causal approach was used: the network was trained to
recognise relationships between a given number of appropriately chosen,
independent input variables and the output variable, the ship’s delay.
(b) An MLP paradigm was used (fully connected, feedforward), the most widely
used in the literature (Liao and Fildes, 2005). Information flow across the
network is feedforward. Information propagates unilaterally moving from the
units of one level to the next: output cannot be used as input.
(c) Input variables were chosen using a priori knowledge without the aid of
special descriptive or statistical techniques. Eight input variables were
148 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

chosen: ‘ship name’, ‘ship length’, ‘transit time’, ‘number of dockers

required for unloading’, ‘number of dockers required for loading’, ‘ETA
month’, ‘ETA day of the week’, ‘ETA hour’.
(d) It was necessary to normalise the variables, appropriately scaling them to
the transfer function used. In this case, an along-channel normalisation was
used: as a logistic transfer function was used the range [0, 1] was chosen.
(e) The 194 patterns contained in the database were divided up among the
training (80 per cent) and validation sets (20 per cent). Once the network
had been trained, predictive ability was evaluated on a different set of data,
the validation set.
(f) In determining the number of hidden layers and nodes, using a trial-and-error
procedure. Additionally, based on the considerations reported by Hornik et al
(1989), it was decided to use networks having one or two hidden layers.
(g) The learning algorithm and related parameters were also determined
using a trial-and-error procedure: an iterative procedure that consists in
training several networks with different architecture and choosing the one
that minimises error. For equal error rates, the network with the simplest
architecture was chosen.
With the aid of JavaNNS software, also used by Bilegan et al (2006), the
model that performed best was obtained with just one hidden layer and 30
nodes. The batch-back propagation algorithm yields the best results defined
as follows:
K learning rate (coefficient that determines the extent of weight
change) ¼ 0.2;
K number of learning cycles (that achieves the best compromise between
accuracy and generalisation) ¼ 20 000.
(h) The smallest validation error expressed in per cent was 8.2 per cent; equated
to the range [24 hour, þ 24 hours], this corresponds to a mean error of
roughly 4 hours in delay prediction. The absolute value of this result should
be considered as it takes into account a prediction of 4 hours late/early,
giving an uncertainty range for ship arrival in port of around 8 hours.

Though satisfactory in scientific terms, the result obtained is not yet accep-
table for the specific operating context. In fact, it is easy to deduce that
K the risk exists that the ship’s actual time of arrival covers three shifts;
K the possibility of univocally determining the demand for each shift can be
completely ruled out.
These considerations form the starting point for the implementation of the new
model discussed in the ‘Discussion: Forecasting module’ section.
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 149
 Fancello et al

Methodology: Optimal human resources allocation module

The optimisation model was also based on an earlier integer linear program-
ming model for optimal allocation of drivers in a transhipment terminal, de-
veloped at the University of Cagliari (Pisano, 2008).
The optimisation model formulation is shown below, along with a brief
description of the architecture.

OBJECTIVE FUNCTION:
X X X X
Min
i¼1; nt j¼1; 4 k¼qc; rt; ra z¼1; l

 ðci;k  ððaz ðprodi; qc þ prodi; rt þ prodi; ra Þ=3Þgz Þxi; j; k; z Þ

X X X X X X
þ ðbvj; k; z Þ þ ðduj; k; z Þ
j¼1; 4 k¼qc; rt; ra z¼1; l j¼1; 4 k¼qc; rt; ra z¼1; l

Subject to

P
(1) xi; j; k; z þ uj; k; z þ vj; k; z ¼ nmj; k; rj; z 8z 2 Na; 8j 2 J; 8k 2 K
i¼1;nt

P
(2) xi; j; k; z þ uj; k; z þ vj; k; z ¼ nhj; k; hj; z 8z 2 H; 8j 2 J; 8k 2 K
i¼1; nt

(3) yi; j ¼ yti; j 8i 2 Tj

P P
(4) xi; j; k; z ¼ yi;j 8i 2 N; 8j 2 J
k¼qc; rt; ra z¼1; l

P P P
(5) xi; j; k; z ¼ 1 8i 2 N
j¼1; 4 k¼qc; rt; ra z¼1; l

(6) vj; k; z X0 integer 8j 2 J; 8k 2 K; 8z 2 Z

(7) uj; k; z X0 integer 8j 2 J; 8k 2 K; 8z 2 Z

150 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

The optimisation problem is treated as a least cost problem. The cost

function is associated to the Boolean variable xi, j, k, z (equal to 1 if the worker i is
assigned to shift j, task k, activity z; 0 otherwise) and is defined as
ðci; k  ððaz ðprodi; qc þ prodi; rt þ prodi; ra Þ=3Þgz Þ

In addition to the cost of permanent staff, the objective function is required to

minimise the cost of any external workers (v on the right hand side) and
indicates personnel shortfall (u in third term). The problem variables are briefly
described in Tables 1 and 2.
The two criteria adopted for human resources allocation are specified
within the objective function. These are labour costs and productivity, which
have been treated taking into account two important aspects:
K Multi-skilled workers, and hence the need to minimise additional costs (ci, k)
incurred by assigning workers to lower level tasks than their main task.
K The need to assign more experienced crews to high priority activities.
More precisely, the deterministic quantity (prodi, qc þ prodi, rt þ prodi, ra )/3)
is introduced to provide a rough estimate of crew productivity (prodi, k
represents productivity time series achieved by worker i performing task k).
A standard crew comprises: one quayside crane operator (qc), two yard crane
operators (rtg) and three truck trailer drivers (ra). This combination ensures
the best mix of productivity and minimum idle time. Housekeeping, which
concerns handling operations inside the yard, does not require quayside
cranes. In order to account for actual operating conditions, crew productivity
is corrected using a coefficient az, a function of the operating conditions for
activity z.
On the other hand, the coefficient gz ensures that the more experienced crews
are assigned to high priority activities.
Lastly the constraints for defining the set of admissible solutions are
summarised in Table 3. The critical properties of the model described together
with the new model formulation are discussed in the ‘Discussion: Optimal
human resources allocation module’ section.

Table 1: Variables of the integer linear programming model for personnel allocation

Variable Type Description

Xi, j, k, z Permanent staff Binary variable: x=1 if worker i is assigned shift j, task k, activity z; 0
otherwise
Vj, k, z External workers Number of external works required for shift j, activity z for task z=ra
(truck trailer) alone
Uj, k, z Shortfall Indicates personnel shortfall, number of workers, shift j, task k and
activity z

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 151
 Fancello et al

Table 2: Symbols of the integer linear programming model for personnel allocation

Symbol Description
Notation j Shift
i ith operator
k Task (qc = quay crane operator, rt = yard crane operator, ra = tractor trailer driver)
z Activity

Set J Shift (1, 2, 3, 4)

K Tasks (qc, rt, ra)
Na Number of ships to be handled (1, y nav)
H Housekeeping operations
Tj Dockers on duty on shift j assigned in monthly schedule (1, y na)
F Dockers on flexible shifts (na þ 1, y nt)
N Dockers on duty (Tj U F )
Z Activities to be performed (Na U H)

Parameters nt Number of dockers on flexible shifts

na Number of dockers already assigned in monthly schedule
l Total activities to be performed (ship þ housekeeping)
nav Number of ships to be handled
yt Docker assigned to shift j in monthly schedule (i in Tj, j in J )
p Priority of task k for docker i (i in N, k in K )
prod Average productivity of docker i for task k (i in N, k in K )
c Cost of assigned task (i in N, k in K )
g Handling gain on ship z; allows to monetise productivity and prioritise ships (z in Z )
a Constant X0 for a decrease/increase in average productivity depending on
operating conditions for activity z (z in Z )
d Unit cost of dummy worker performing task k on shift j (integer)
b Cost of external worker performing the task (integer)

Table 3: Constraints of the integer linear programming model for personnel allocation

Constraint Description
1 Ensure manpower demand is satisfied for the working vessel maintaining correct crew
composition
2 Ensure manpower demand is satisfied for housekeeping operations maintaining correct crew
composition
3 Adhere to monthly scheduling for workers with shifts already assigned
4 Account for logic connection for monthly scheduling adherence
5 Ensure that each worker is assigned only one shift, one task and one activity
6 Concerns integrity and non-negativity of the variable v
7 Concerns integrity and non-negativity of the variable u

Discussion

Discussion: Forecasting module

The main shortcoming of the initial prediction model, as mentioned in the

‘Methodology: Forecasting module’ section, concerned the choice of input
152 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

variables, which was made without the aid of specific variable selection and
reduction techniques. In an attempt to improve predictive quality, the attention
was focused on data pre-processing.
Starting from the initial set of eight variables, first of all a ninth was added
using a priori knowledge, ‘ship’s port of departure’.
Analysis of the correlations made it possible to eliminate, using the w2 tests,
the redundant variables, variables strongly correlated with others that do not
provide any additional information on the phenomenon being studied. Using
multivariate statistical techniques (multiple correspondence analysis and clus-
ter analysis), it was possible to analyse simultaneously the remaining variables
and reorganise them, using combined simultaneous analysis1 of the w2 and
Valor-test,2 depending on their significance.3
The results of the significance test were then analysed and four sets of
alternative inputs determined (Table 4).
Considering the four sets of input data selected, a new neural network was
specified, and its predictive ability assessed. The study comprised two separate
test phases.
The first phase consisted of comparing the results obtained implementing
numerous networks, varying the following:

K the number of input variables using alternately the first, second, third and
fourth set;
K the learning algorithm: using backpropagation, backpropagation with
momentum, batch backpropagation or resilient propagation, the algorithms
most widely used in the literature;
K the learning parameters: number of learning cycles, learning rate and so on;
K the number of hidden nodes: 5, 10, 20 or 30 nodes per hidden layer.

All the networks implemented in this phase had just one hidden layer.
Table 5 shows the minimum prediction errors, expressed in per cent, for the
screening.4
An important result already emerged in the first screening for
the chosen set of variables. The best networks had three variables in
the input level. Table 6 shows the substantial reduction in minimum
prediction error, from 9.3 per cent (seven variables) to 6 per cent (three
variables).
It is important to note that in choosing the best model, we considered not
only the error values but also the training error and validation error curves
obtained during learning.
Considering the sets with four, five and seven input variables, we
found that in practically all the tests, these curves exhibited two different
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 153
154
Fancello et al

Table 4: w2 and Valor-test values for the variables in the four sets chosen

Variables First set (seven variables) Second set (three variables) Third set (four variables) Fourth set (five variables)
2 2 2
w v-test w v-test w v-test w2 v-test
Name of ship 739.9 21.1 895.1 25.4 1000.4 26.0 841.2 23.1
Departure port 449.2 18.6 352.9 16.6 463.7 18.9 406.3 17.4
No, crew loading 289.9 10.5 187.4 8.1 314.8 11.4 407.4 14.2
Number of crew unloading 218.3 6.5 — — 199.1 5.6 188.2 5.1
ETA month 219.0 7.8 — — — — 226.4 8.1
ETA day of week 139.0 6.9 — — — — — —
ETA hour 261.54 4.77 — — — — — —

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Table 5: Minimum errors, expressed in per cent for the first test phase

Learning cycles: 40 000

Algorithm Backpropagation Backpropagation with momentum Batch backpropagation Resilient propagation

Hidden nodes dmax=0.1 dmax=0.1 dmax=0.1 dmax=0.1

h=0.1 h=0.2 h=0.1 h=0.2 h=0.1 h=0.2 d=0, 1 d=0

d=50 d=30
a=4, 0 a=2, 0
First set: seven variables 5 12.12 14.83 13.42 13.78 9.33 9.36 14.14 10.25
10 21.45 113.84 16.12 17.89 9.64 9.75 18.17 10.25
20 16.73 18.57 17.89 21.91 9.38 9.80 18.14 10.25
30 19.75 19.24 22.14 25.69 8.89 9.38 24.23 10.25
Second set: three variables 5 7.75 8.37 8.48 9.06 5.95 5.99 10.05 6.16
10 8.43 8.83 8.53 8.94 5.55 5.85 10.49 6.16
20 8.66 8.49 8.57 8.72 5.66 6.32 10.49 6.16
30 8.19 8.77 8.79 9.02 5.83 6.32 10.49 6.16
Third set: four variables 5 11.83 12.25 13.15 18.11 9.21 9.21 — —
10 13.67 14.87 13.08 17.32 9.03 9.44 13.49 10.49
20 13.86 15.10 13.42 16.40 9.46 9.38 12.96 10.44
30 — — — — — — — —
Forth set: five variables 5 12.00 — 13.78 — 9.00 — 12.25 —
10 13.71 — 14.14 — 9.27 — 13.42 —
20 14.83 — 18.97 — 8.83 — 12.45 —
30 — — — — — — — —
Prediction of arrival times and human resources allocation

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
155
 Fancello et al

Table 6: Minimum prediction error during learning for the four sets of variables chosen

First set Second set Third set Fourth set

(seven variables) (three variables) (four variables) (five variables)
Error (%) 9.3 6 9 8.8

Figure 1: Trend of training error (continuous) and validation error (dashed) curves during training.

trends during training:

K Practically constant for both: this can be explained by the fact that the
addition of less significant variables to the set with three variables, might
constitute an impediment in the learning process: the network ‘does not
learn’ (Figure 1a).
K The training error curve slopes downwards, whereas the validation error
curve slopes practically upwards (irregularly) and in some cases even
intersects and surpasses the training error curve. In the latter case, overfitting
has occurred, a problem associated largely with network architecture design:
the network loses generalisation ability (Figure 1b).

On the basis of these results, in the second test phase, a new specification was
developed keeping the number of input variables (3) and learning algorithm
(batch-backpropagation) unchanged and attempting to improve the network
architecture by varying:

K the number of hidden layers: 1, 2;

K the number of hidden nodes: 5, 10, 20 or 30 for just one hidden layer 4 þ 4 (8),
5 þ 5 (10), 6 þ 6 (12) and 7 þ 7 (14) for two hidden layers;
K learning parameters.

Table 7 shows the results for the second test phase (minimum error
expressed in per cent).
156 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

Table 7: Minimum error expressed in per cent for the second test phase

Learning parameters

Hidden h=0.25 h=0.20 h=0.15 h=0.1

nodes
dmax=0.05 dmax=0.1 dmax=0.05 dmax=0.1 dmax=0.05 dmax=0.1 dmax=0.05 dmax=0.1

8 5.91 5.65 5.65 5.38 6.08 5.83 5.65 5.65

10 6 5.65 5.47 5.91 5.83 5.38 5.91 5.47
15 5.47 5.83 5.65 5.83 6.16 5.83 5.74 5.47
4þ4 6.16 5.47 5.74 5.47 5.83 5.91 9.7 9.64
5þ5 5.83 6 6.16 5.74 6.16 5.74 6.24 5.65
6þ6 — — 5.91 6 — 5.91 5.47 5.74
7þ7 — — — — — — 6.16 5.83

Over 120 models were implemented in the first and second phases with the
aid of JavaNNS software. The best results were obtained for the following
model:

K size of learning set: 158 patterns;

K size of validation set: 39 patterns;
K paradigm: feedforward multi-layer perceptron fully connected;
K input nodes: 3;
K hidden layers: 2;
K hidden nodes: 4 þ 4;
K output node: 1;
K learning algorithm: batch-backpropagation;
K learning rate: 0.25;
K learning cycles: 40 000.
The analysis conducted with the neural network yielded positive results. It was
possible to reduce validation error by almost 34 per cent, from 8.2 per cent to
5.47 per cent, as shown in Table 7. Mean delay prediction error decreased from
4 hours to around 2 hours 40 min (absolute value), obtaining an uncertainty
range of o6 hours (5.20 hours).
Two basic considerations emerge from this analysis:

K the possibility that the ship’s predicted arrival time spans three shifts (most
critical condition) is eliminated;
K the possibility exists, first ruled out, that the predicted arrival time falls within
a single shift, enabling to determine univocally the demand for that shift.

So in practice, the certainty exists that personnel can be scheduled for two shifts
instead of three.
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 157
 Fancello et al

Discussion: Optimal human resources allocation module

A number of personnel allocation tests were carried with the optimisation

model described in the ‘Methodology: Optimal human resources allocation
module’ section, assuming demand falls short of, is equal to or exceeds man-
power resources. For demand shortfall constraint 5 had to be modified to in-
clude the possibility of worker ‘i’ not being assigned a shift.
Analysis of the structure identified two aspects that leave room for im-
proving the formulation and obtaining a more realistic level:
(a) Interperiodicity
(b) Productivity.
(a) Interperiodicity: The model described is static. The demand refers to a single
shift so one needs to determine the number and composition of the crews for
satisfying that demand, assuming a given manpower availability for meeting the
needs of the different shifts. Should insufficient personnel be allocated to handle
the workload then the model indicates the shortfall per shift and per task, re-
sorting to ‘dummy’ workers: the uj, k, z. This type of solution does not however
realistically reflect the operational situation in seaport terminals where, by con-
trast, any tasks left undone at the end of a shift, are passed on to the next one.
Thus workers starting the next shift not only have to meet the demand envisaged
for that shift but also complete any work left undone in the previous one.
Thus a dynamic model is required, in other words that envisages work
undone in one shift due to insufficient manning being passed on to the next,
considering the interperiodicity over the shifts and over the 24 hours.
In the static model, the constraint for satisfying demand is posed in terms of
demand balance: Xj þ Vj þ Uj ¼ period demand j.
When Xj þ Vjoperiod demand j, part of the demand predicted for that shift
cannot be satisfied because of undermanning. This shortfall is indicated by the
non-null value taken by the variables ‘uj, k, z’. In the cost budget, these ‘dummy’
workers, which actually represent undermanning for the work period, will
necessarily result in much higher costs compared with actual workers, be they
permanent staff or external workers.
The variable of interperiodic nature will therefore represent undermanning
of a shift, which equates to demand not being satisfied in that shift.
Bearing in mind the above considerations, the dynamicity constraint of
demand on shifts can be formulated as follows:
For cargo handling:
X
xi; j; k; z þ uj; k; z þ vj; k; z  uj1; k; z ¼ nmj; k; rj; z
(a) i¼1; nt

8z 2 Na; 8j 2 J; 8k 2 K; with j 6¼ 1

158 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

X
(b) xi; j; k; z þ uj; k; z þ vj; k; z ¼ nmj; k; rj; z
i¼1; nt

8z 2 Na; 8j 2 J; 8k 2 K; with j ¼ 1

For housekeeping:
X
(c) xi; j; k; z þ uj; k; z þ vj; k; z  uj1; k; z ¼ nhj; k; hj; z
i¼1; nt

8z 2 H; 8j 2 J; 8k 2 K; with j 6¼ 1

X
(d) xi; j; k; z þ uj; k; z þ vj; k; z ¼ nhj; k; hj; z
i¼1; nt

8z 2 H; 8j 2 J; 8k 2 K; with j ¼ 1

With the new system of constraints work left undone in shift j is auto-
matically added to the workload for shift j þ 1 overcoming the staticity of the
previous formulation.
The undermanning detected in the fourth shift will on the other hand in-
crease the workload predicted for the first shift ( j ¼ 1) on the next day.
The uncompleted work predicted by the model for the last shift of the
previous day is a known value (though it does suffer from some degree of
uncertainty) and the planner will input this directly into the database, thereby
increasing the demand predicted for the first shift of the day for which per-
sonnel is yet to be scheduled. The constraint for the shift j ¼ 1 therefore remains
unchanged with respect to the original formulation.

(b) Productivity: The deterministic description included in the crew productivity

model is unnatural. Studies reported in the literature show the productivity
parameter to be influenced by numerous factors, only some of which can be
controlled by the planner, and is ill-suited to being interpreted deterministically.
It is important to identify those aspects influencing operational efficiency in one
way or another to enable the planner to introduce the most suitable boundary
conditions directly into the model.
The shift in particular has a major influence on operator fatigue and per-
formance (Fancello et al, 2008). This effect can be incorporated into the ob-
jective function using a new parameter ‘e’, a function of the shift j. The value it
takes, calibrated using statistical procedures or determined by the planner, will
be input into the data files. In this way, the planner is able to optimise pro-
ductivity using more detailed data, ensuring a greater degree of adherence. The
model does not provide a definitive invariant indicator, but a general and
generalisable tool that planners can utilise according to experience.
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 159
 Fancello et al

On the basis of the above considerations, the new model can be defined as
follows:

X X X X
Min ðci; k  ðej ðaz ðprodi; qc þ prodi; rt þ prodi; ra Þ=3Þgz Þxi; j; k; z Þ
i¼1; nt j¼1; 4 k¼qc; rt; ra z¼1; l

X X X X X X
þ ðbvj; k; z Þ þ ðduj; k; z Þ
j¼1; 4 k¼qc; rt; ra z¼1; l j¼1; 4 k¼qc; rt; ra z¼1; l

Subject to
X
(a) xi; j; k; z þ uj; k; z þ vj; k; z  uj1; k; z ¼ nmj; k; rj; z
i¼1; nt

8z 2 Na; 8j 2 J; 8k 2 K; with j 6¼ 1

X
(b) xi; j; k; z þ uj; k; z þ vj; k; z ¼ nmj; k; rj; z
i¼1; nt

8z 2 Na; 8j 2 J; 8k 2 K; with j ¼ 1

X
(c) xi; j; k; z þ uj; k; z þ vj; k; z  uj1; k; z ¼ nhj; k; hj; z
i¼1; nt

8z 2 H; 8j 2 J; 8k 2 K; with j 6¼ 1

X
(d) xi; j; k; z þ uj; k; z þ vj; k; z ¼ nhj; k; hj; z
i¼1; nt

8z 2 H; 8j 2 J; 8k 2 K; with j ¼ 1

X X
(e) xi; j; k; z ¼ yi; j 8i 2 N; 8j 2 J
k¼qc; rt; ra z¼1; l

X X X
(f) xi; j; k; z p1 8i 2 N
j¼1; 4 k¼qc; rt; ra z¼1; l

(g) vj; k; z X0 integer 8j 2 J; 8k 2 K; 8z 2 Z

(h) uj; k; z X0 integer 8j 2 J; 8k 2 K; 8z 2 Z

160 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

As can be observed, the constraint yi, j ¼ yti, j, has been relaxed as it was
superfluous. There is no sense in getting the model to do something already
determined in the data files.
The integer linear programming problem was formulated and solved with
open source GLPK software using the GLPSOL solver, included in the GLPK
package.
The model was tested for three typical situations likely to arise in terminal
management:

1. Demand ¼ Supply (Figure 2).

2. DemandoSupply (Figure 3).
3. Demand4Supply (Figure 4).

Using the above software, it was possible to determine, in a short time

(o100 ), the optimal personnel allocation while observing the constraints

Figure 2: Result of allocation for situation 1: Demand ¼ Supply (Assumption: Demand ¼ 4 loading/
unloading gangs þ 2 housekeeping gangs; Supply ¼ 34 dockers). (a) New model: FO ¼ 18 733 and
(b) initial model: FO ¼ 18 733.

Figure 3: Result of allocation for situation 2: Demand oSupply (Assumption: Demand ¼ 11 loading/
unloading gangs þ 2 housekeeping gangs; Supply ¼ 96 dockers). (a) New model: FO ¼ 76 033 and
(b) initial model: FO ¼ 76 033.

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 161
 Fancello et al

Figure 4: Result of allocation for situation 3: Demand4Supply (Assumption: Demand ¼ 18 loading/

unloading gangs þ 7 housekeeping gangs; Supply ¼ 96 dockers). (a) New model: FO ¼ 360 333 and
(b) initial model: FO ¼ 360 333.

and the ability of the model to realistically represent terminal operating

characteristics.
Tests conducted assuming demand equal to or less than manpower supply
for the period (situations 1 and 2) correctly reproduce the same result obtained
with the original formulation. In fact, as there is no undermanning, the dynamic
constraints do not come into play and the dynamic formulation coincides with
the original static one (Figures 2a–b and 3a–b).
However, the same cannot be said for the third situation. In this case,
the original and new models yield different results (Figures 4a and b). The
assumption that demand exceeds manpower supply in the period necessitates
assigning the ‘u’, thereby activating the dynamic constraints for moving the
unsatisfied demand from one shift to another (Figure 4a). The new model
yields a more realistic representation of actual operating characteristics in a
seaport terminal. This is confirmed by the higher value taken by the objective
function, indicating the additional costs incurred by the terminal due to
undermanning.
Further tests conducted introducing the shift parameter into the dynamic
formulation but using the same data as for situation 3, yielded variations in the
value of the objective function (5581.46) and in the assignment of ‘flexible’
permanent staff. This was to be expected insomuch as the parameter ‘e’, by
changing the historical productivity data, makes it more convenient to assign to
the worker one shift rather than another.

Conclusion

The difficulties inherent in managing port operations, due to the uncertainty

of demand and to the complexity of planning processes, means that planners
162 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

need to be assisted at each stage by tools that support the decision-making

process.
This article describes two models that support the planner in the different
planning activities: a forecasting model and an optimisation model.
With the neural network-based forecasting model proposed here, it was
possible to reduce the uncertainty interval of ships’ arrival time in port, thereby
increasing the accuracy of demand forecasting, with the certainty, in practice,
of being able to plan resources around just two work shifts instead of three.
The demand thus calculated was used as input for the subsequent optimi-
sation model. The optimisation formulation is of an interperiodic nature
and represents the operations actually carried out in a terminal more realisti-
cally. With this model, it was possible to optimise worker allocation over the
24-hour period, taking into account actual operational requirements.
Considering the promising results achieved, the two models, appro-
priately revisited and integrated, can provide a useful planning support tool.
This part of the research falls within a broader project aimed at developing a
complete DSS capable of providing 3601 support to planners. This will reduce
terminal operating costs while maximising efficiency, thereby enhancing
competitiveness.

Notes

1 w2: Test of independence of two variables. The null hypothesis is tested that two classification
criteria when applied to the same data set are independent. If the distribution with respect to
one criterion is not influenced by the classification with respect to the other, then the two
classification criteria are said to be independent.
2 Valor test, this represents the deviation of the significant variable with respect to a normal
distribution. The principle is as follows: for a sample size of n individuals, q nominal variables
have been observed. A particular group of nk individuals is identified. How do we classify in
order of importance the variables that best characterise that group? A variable will not
characterise the group if the nk values found appear to have been drawn at random from the n
values observed. The more doubtful the hypothesis of a random draw, the more significant that
variable will be for characterising the group. The Valor test can thus be written as

n
njk  nk  nj
V:T: ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
nj  nj 
nk  nn n1  n  1  n
k

If v-test 42, the mean of the group will differ significantly from the sample. Thus the higher the
v-test, the more significant and characterising the variable will be for that group.
3 Statistical analyses were performed with the aid of the French SPAD.N software (Système
Portable Pour l’Analyse des Données).
4 The percentage error is calculated as (OMSE)  100, where MSE is the mean square error.

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 163
 Fancello et al

References

Beasley, J.E. and Cao, B. (1998) A dynamic programming based algorithm for the crew scheduling
problem. Computers and Operations Research 25(7/8): 567–582.
Bilegan, I.C., Crainic, T.G. and Gendreau, M. (2006) Fleet Management for Advanced Intermodal
Services. Final Report, Centre de Recherche sur les Transports (CRT) CRT-2006-13.
Bowden, G.J., Dandy, G.C. and Maier, H.R. (2004) Input determination for neural network models
in water resources applications. Part 1 – background and methodology. Journal of Hydrology
301(1–4): 75–92.
Carbonneau, R., Laframboise, K. and Vahidov, R. (2007) Application of machine learning techniques
for supply chain demand forecasting, O.R. applications. European Journal of Operational
Research 184(3): 1140–1154.
Celik, H.M. (2004) Modeling freight distribution using artificial neural networks. Journal of
Transport Geography 12(2): 141–148.
Cordeau, L., Laporte, G., Legato, P. and Moccia, L. (2005) Models and Tabu search heuristics for the
Berth allocation problem. Transportation Science 39(4): 526–538.
Dell’Olmo, P. and Lulli, G. (2004) Planning activities in a network of logistic platforms with shared
resources. Annuals of Operations Research 129(1–4): 155–169.
Fancello, G., D’Errico, G. and Fadda, P. (2008) Processing and analysis of ship-to-shore gantry
crane operator performance curves in container terminals. Journal of Maritime Research
5(2): 39–58.
Gambardella, L.M., Mastrolilli, M., Rizzoli, E. and Zaffalon, M. (2001) An optimization metho-
dology for intermodal terminal management. Journal of Intelligent Manufacturing 12(5–6):
521–534.
Gaudioso, M., Giallombardo, G. and Legato, P. (1999) Ottimizzazione e simulazione nella
gestione di un terminal container portuale: il caso di Gioia Tauro. Meeting’s documents
“Metodi e Tecnologie dell0 Ingegneria dei Trasporti – Seminario 1999”, Reggio Calabria,
11–13 December 1999, Cantarella G., Russo F., Milano: Franco Angeli, 2001, Vol. 1798/10,
pp. 273–290.
GNU Linear Programming Kit: Modeling Language GNU MathProg. Draft edition, January 2004.
Version 4.4.
Hartmann, S. (2004) A general framework for scheduling equipment and manpower at container
terminals. OR Spectrum 26: 51–74.
Haykin, S. (1994) Neural Networks. New York: Macmillan College Publishing Company.
Hornick, K., Stinchcombe , M. and White, H. (1989) Multilayer feedforward networks are universal
approximators. Neural Networks 2(5): 359–366.
Imai, A., Nishimura, E. and Papadimitriou, S. (2001) The dynamic berth allocation problem for a
container port. Transportation Research 35B: 401–417.
La Rocca, M. and Perna, C. (2005) Variable selection in neural network regression models
with dependent data: A subsampling approach. Computational Statistics & Data Analysis
48(2): 415–429.
Legato, P. and Monaco, M.F. (2004) Human resources management at a marine container terminal.
European Journal of Operation Research 156(3): 769–781.
Li, K. and Peng, J.-X. (2006) Neural input selection – a fast model-based approach. Neurocomputing
70(4–6): 762–769.
Liao, K.P. and Fildes, R. (2005) The accuracy of a procedural approach to specifying feedforward
neural networks for forecasting. Computers & Operations Research 32(8): 2151–2169.
Meersmans, P.J.M. and Wagelmans, A.P.M. (2001) Dynamic Scheduling of Handling Equipment at
Automated Container Terminals. Erasmus University, Rotterdam, The Netherlands. Econometric
Insitute Report, EI 2001-33.

164 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

Monaco, M.F., Moccia, L. and Sammarra, M. (2008) Operations research for the management
of a transshipment container terminal. The Gioia Tauro case. Maritime Economics & Logistics
11(1): 7–35.
Papadonkontantakis, S., Machefer, S., Schnitzlain, K. and Lygeros, A.I. (2005) Variable selection and
data pre-processing in NN modelling of complex chemical processes. Computers & Chemical
Engineering 29(7): 1647–1659.
Park, Y.M. and Kim, K.H. (2003) A scheduling method for berth and quay cranes. OR Spectrum
25(1): 1–23.
Pisano, M. (2008) A decision support system for planning operations at a transshipment terminal
container. PhD thesis, University of Cagliari.
Potvin, J.-Y and Smith, K.A. (2001) Artificial neural networks for combinatorial optimization.
Rapport CRT-2001-02, Centre de recherche sur les transports.
Saxen, H. and Petterson, F. (2006) Method for the selection of inputs and structure of feedforward
neural networks. Computers & Chemical Engineering 30(6–7): 1038–1045.
Zhang, G., Patuwo, B.E. and Hu, M.Y. (1998) Forecasting with artificial neural networks: The state of
the art. International Journal of Forecasting 14(1): 35–62.

Appendix

Instance 1: Demand ¼ Supply

The workday, divided into four shifts, is characterised by a demand that can be
fully satisfied by available workers without under- or overmanning.
The allocation hypothesis is as follows:

K Supply: 34 dockers available, 10 have been assigned the shift in monthly

schedule, the remaining 24 are on flexible shifts. Productivity time series
associated with each worker, differentiated by task, is shown in the following
(Table A1).
K Demand (D): the workday envisages two ship handling and one house-
keeping operation. For the terminal operator, these three activities are
identical in that they have the same priority and produce the same economic
benefit. The following table shows the personnel required for each activity
for each shift (Table A2).

The allocation results obtained using the static and dynamic models are shown
below (Table A3).
Figures 2a and b summarise the allocation results for each shift. Quite
rightly, the model does not assign any u or v, the demand for each shift being
fully satisfied by the terminal’s personnel. As no priority rules have been de-
fined for the different activities, the dockers are assigned purely on the basis of
cost and productivity time series data, thereby ensuring the objective function is
minimised.
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 165
 Fancello et al

Table A1: Instance 1: Dockers’ productivity

Dockers
Task 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Qc 25 24 23 22 0 0 0 0 0 0 0 0 0 0 0 0 0
Rt 15 14 13 12 15 14 13 12 15 14 13 12 0 0 0 0 0
Ra 13 12 11 10 13 12 11 10 13 12 11 10 13 12 11 10 13

Task 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
Qc 0 0 0 0 0 0 0 25 24 23 22 0 0 0 0 0 0
Rt 0 0 0 0 0 0 0 15 14 13 12 15 14 13 12 15 14
Ra 12 11 10 13 12 11 10 13 12 11 10 13 12 11 10 13 12

Table A2: Instance 1: Demand

ID activities Type of activity

Ship Housekeeping

Number of activities: 2 Number of activities: 1

1 2 3
D1 1 0 0
Demand/shift D2 1 0 0
D3 0 1 1
D4 0 1 1

As no dummy workers come into play here, the dynamic constraints of the
new formulation coincide with the two original constraints. Thus the objective
function will have the same value in both models.

166 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

Table A3: Instance 1: Allocation results

Dockers Static model: OF=187 333 Dynamic model: OF=187 333

j k z j k z
1 1 qc 1 1 qc 1
2 2 qc 1 2 qc 1
3 3 qc 2 3 qc 2
4 4 qc 2 4 qc 2
5 1 ra 1 1 rt 1
6 1 rt 1 1 rt 1
7 2 rt 1 2 ra 1
8 2 rt 1 2 rt 1
9 3 ra 3 3 ra 2
10 3 rt 2 3 ra 2
11 1 rt 1 3 ra 2
12 3 rt 2 3 rt 2
13 2 ra 1 3 ra 3
14 2 ra 1 3 ra 3
15 2 ra 1 2 ra 1
16 1 ra 1 4 ra 3
17 4 ra 2 3 ra 3
18 4 ra 2 1 ra 1
19 3 ra 2 1 ra 1
20 3 ra 2 4 ra 2
21 3 ra 2 4 ra 2
22 3 ra 3 2 ra 1
23 1 ra 1 1 ra 1
24 3 ra 3 4 ra 3
25 4 rt 2 4 rt 3
26 4 rt 2 4 rt 3
27 4 rt 3 4 rt 2
28 4 rt 3 3 rt 2
29 3 rt 3 4 ra 2
30 4 ra 2 4 rt 2
31 3 rt 3 4 ra 3
32 4 ra 3 3 rt 3
33 4 ra 3 3 rt 3
34 4 ra 3 2 rt 1
v, u: not assigned v, u: not assigned

Instance 2: DemandoSupply

The number of activities envisaged for scheduling the workday requires a

smaller number of dockers than those available, resulting in overmanning with
respect to actual requirements.
Allocation hypothesis:
K Supply: 96 dockers available, 48 have been assigned the shift in monthly
schedule, the remainder are on flexible shifts. Productivity time series data
r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 167
 Fancello et al

Table A4: Instance 2: Dockers’ productivity

Dockers
Task 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Qc 25 24 23 22 0 0 0 0 0 0 0 0 0 0 0 0 0
Rt 15 14 13 12 15 14 13 12 15 14 13 12 0 0 0 0 0
Ra 13 12 11 10 13 12 11 10 13 12 11 10 13 12 11 10 13

Task 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
Qc 0 0 0 0 0 0 0 25 24 23 22 0 0 0 0 0 0
Rt 0 0 0 0 0 0 0 15 14 13 12 15 14 13 12 15 14
Ra 12 11 10 13 12 11 10 13 12 11 10 13 12 11 10 13 12

Task 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
Qc 0 0 0 0 0 0 0 0 0 0 0 0 0 0 25 24 23
Rt 13 12 0 0 0 0 0 0 0 0 0 0 0 0 15 14 13
Ra 11 10 13 12 11 10 13 12 11 10 13 12 11 10 13 12 11

Task 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
Qc 22 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Rt 12 15 14 13 12 15 14 13 12 0 0 0 0 0 0 0 0
Ra 10 13 12 11 10 13 12 11 10 13 12 11 10 13 12 11 10

Task 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
Qc 0 0 0 0 25 24 23 22 0 0 0 0 0 0 0 0 0
Rt 0 0 0 0 15 14 13 12 15 14 13 12 15 14 13 12 0
Ra 13 12 11 10 13 12 11 10 13 12 11 10 13 12 11 10 13

Task 86 87 88 89 90 91 92 93 94 95 96
Qc 0 0 0 0 0 0 0 0 0 0 0
Rt 0 0 0 0 0 0 0 0 0 0 0
Ra 12 11 10 13 12 11 10 13 12 11 10

associated with each worker, differentiated by task, are shown in the

following (Table A4).
K Demand (D): activities envisaged for the workday include handling five ships
plus one housekeeping operation. Ship handling 1 and 2 have higher priority
and produce a greater economic benefit (from a modelling standpoint, this
means that the parameter gz will be greater). The following table shows the
gangs required for each activity per shift (Table A5).
The allocation results obtained using both the original static model and the
dynamic formulation are shown below (Tables A6 and A7).
Again quite rightly the model does not assign any ‘v’ or ‘u’.
168 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

Table A5: Instance 2: Demand

ID activities Type of activity

Ship Housekeeping

Number of activities: 5 Number of activities: 1

1 2 3 4 5 6
Demand/shift D1 1 0 1 0 1 0
D2 1 0 1 0 1 0
D3 0 1 0 1 0 1
D4 0 2 0 1 0 1

Table A6: Instance 2: Allocation results using the original model

i j k z i j k z i j k z
1 1 qc 1 34 3 rt 2 67 not assigned
2 2 qc 1 35 4 rt 4 68 not assigned
3 3 qc 2 36 4 rt 4 69 4 ra 2
4 4 qc 2 37 1 ra 5 70 not assigned
5 1 rt 1 38 1 ra 5 71 not assigned
6 1 ra 1 39 1 ra 5 72 not assigned
7 2 rt 5 40 2 ra 5 73 4 qc 2
8 2 rt 5 41 2 ra 5 74 4 rt 2
9 3 ra 2 42 2 ra 5 75 1 qc 3
10 3 ra 2 43 3 ra 6 76 1 qc 5
11 4 rt 6 44 3 ra 6 77 2 rt 1
12 4 rt 6 45 3 ra 6 78 4 ra 2
13 1 ra 3 46 4 ra 6 79 3 rt 4
14 1 ra 3 47 4 ra 6 80 1 rt 3
15 1 ra 3 48 4 ra 6 81 2 rt 1
16 2 ra 3 49 4 ra 2 82 4 ra 2
17 2 ra 1 50 4 rt 2 83 3 rt 4
18 2 ra 3 51 2 qc 5 84 1 rt 3
19 3 ra 4 52 2 qc 3 85 3 ra 4
20 3 ra 4 53 4 ra 2 86 not assigned
21 3 ra 2 54 4 rt 2 87 not assigned
22 4 ra 4 55 1 rt 5 88 not assigned
23 4 ra 4 56 3 rt 6 89 2 ra 3
24 4 ra 4 57 2 ra 1 90 not assigned
25 1 rt 1 58 4 rt 2 91 not assigned
26 2 ra 1 59 1 rt 5 92 not assigned
27 3 qc 4 60 3 rt 6 93 4 ra 2
28 4 qc 4 61 not assigned 94 not assigned
29 1 ra 1 62 not assigned 95 not assigned
30 1 ra 1 63 not assigned 96 not assigned
31 2 rt 3 64 not assigned
32 2 rt 3 65 not assigned v not assigned
33 3 rt 2 66 not assigned u not assigned

Original model: OF=76 033.

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 169
 Fancello et al

Table A7: Instance 2: Allocation results using the new formulation

i j k z i j k z i j k z
1 1 qc 1 34 3 rt 2 67 not assigned
2 2 qc 1 35 4 rt 6 68 not assigned
3 3 qc 4 36 4 rt 4 69 4 ra 2
4 4 qc 4 37 1 ra 5 70 not assigned
5 1 ra 1 38 1 ra 5 71 not assigned
6 1 ra 1 39 1 ra 5 72 not assigned
7 2 rt 5 40 2 ra 5 73 4 ra 2
8 2 rt 5 41 2 ra 5 74 4 rt 2
9 3 ra 2 42 2 ra 5 75 1 qc 5
10 3 ra 2 43 3 ra 6 76 2 qc 5
11 4 rt 6 44 3 ra 6 77 2 ra 1
12 4 rt 4 45 3 ra 6 78 2 rt 1
13 1 ra 3 46 4 ra 6 79 1 rt 5
14 1 ra 3 47 4 ra 6 80 3 rt 4
15 1 ra 3 48 4 ra 6 81 4 ra 2
16 2 ra 3 49 4 qc 2 82 4 ra 2
17 2 ra 1 50 4 rt 2 83 3 rt 4
18 2 ra 3 51 1 qc 3 84 3 rt 6
19 3 ra 4 52 2 qc 3 85 2 ra 3
20 3 ra 4 53 3 rt 2 86 not assigned
21 3 ra 4 54 4 rt 2 87 not assigned
22 4 ra 4 55 1 rt 3 88 not assigned
23 4 ra 4 56 1 rt 3 89 4 ra 2
24 4 ra 4 57 4 rt 2 90 not assigned
25 1 rt 1 58 2 rt 1 91 not assigned
26 2 ra 1 59 3 rt 6 92 not assigned
27 3 qc 2 60 1 rt 5 93 4 ra 2
28 4 qc 2 61 not assigned 94
29 1 rt 1 62 not assigned 95
30 1 ra 1 63 not assigned 96
31 2 rt 3 64 not assigned
32 2 rt 3 65 not assigned ‘v’ not assigned
33 3 ra 2 66 not assigned ‘u’ not assigned

New formulation: OF=76 033.

As no ‘u’ is assigned, the dynamic constraint does not come into play and
the formulation coincides with the original one (Figures 3a and b). The model
recognises the surplus of personnel and only allocates the number of dockers
needed to satisfy the demand forecast. Twenty of the 96 dockers available are
not assigned to any shift. To ensure that the monthly schedule is observed, the
remaining dockers are allocated in such a way as to minimise the objective
function. Ship handling operations 1 and 2, which have higher priority, are
allocated to dockers with higher productivity.
Note that there are some minor differences in allocation. It should be re-
called that for determining the best solution, the model performs an overall
170 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

balance. Hence, it may happen that several allocation options exist that are
identical in terms of objective function value.

Instance 3: Demand4Supply

Available personnel is insufficient to ensure completion of operations predicted

for the workday.
K Supply: same as instance 2, 96 dockers available, half of whom have been
assigned a shift in the monthly schedule (Table A4).
K Demand (D): activities envisaged for the workday include handling six ships
plus three housekeeping operations. For the terminal operator these nine
operations differ in that ship handing 1, 2 and 3 have priority twice as high as
the others.
Table A8 shows the number of gangs required for each operation for each shift.
The allocation results obtained using the dynamic formulation are shown
below (Table A9).
Once all available dockers have been allocated, to satisfy the demand the
least cost model draws first from the less expensive external workers ‘v’, who
however can only be assigned the task of truck trailer drivers, all the other
workers required to fill the shortfall being drawn from the more expensive ‘u’.
This explains the high value taken by the objective function as a result of the
personnel shortfall (Table A10).
Allocation of the u activates the dynamic constraints to ensure that any
unfulfilled demand passes from one shift to the next, thereby reproducing
what actually happens in the terminal. This is illustrated in Figures 4a and b,
which clearly show the difference between the new interperiod formulation
(Figure 4a), whereby any unfulfilled demand passes from one shift to the next,
and the static formulation where this does not happen (Figure 4b).

Table A8: Instance 3: Demand

ID activities Type of activity

Ship Housekeeping

Number of activities: 6 Number of activities: 3

1 2 3 4 5 6 7 8 9
Demand/shift D1 2 0 2 1 2 1 0 1 1
D2 1 0 2 0 1 0 0 1 0
D3 0 1 0 1 0 1 1 0 1
D4 1 1 0 0 1 0 1 1 0

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 171
 Fancello et al

Table A9: Instance 3: Allocation results using the new formulation; internal workers

i j k z i j k z i j k z
1 1 qc 1 33 3 rt 3 65 2 ra 1
2 2 qc 1 34 3 rt 1 66 1 ra 3
3 3 rt 3 35 4 rt 2 67 1 ra 3
4 4 qc 1 36 4 rt 2 68 2 ra 8
5 1 rt 3 37 1 ra 3 69 3 ra 2
6 1 rt 5 38 1 ra 1 70 1 ra 1
7 2 rt 3 39 1 ra 9 71 2 ra 3
8 2 rt 5 40 2 ra 8 72 3 ra 6
9 3 rt 2 41 2 ra 3 73 1 rt 1
10 3 rt 2 42 2 ra 3 74 1 qc 3
11 4 rt 1 43 3 ra 9 75 1 qc 5
12 4 rt 1 44 3 ra 7 76 1 qc 6
13 1 ra 1 45 3 ra 2 77 2 rt 1
14 1 ra 1 46 4 ra 1 78 1 rt 9
15 1 ra 1 47 4 ra 8 79 1 rt 8
16 2 ra 8 48 4 ra 7 80 1 rt 9
17 2 ra 3 49 1 qc 1 81 1 rt 1
18 2 ra 3 50 2 qc 3 82 1 rt 6
19 3 ra 4 51 1 qc 5 83 1 rt 6
20 3 ra 6 52 1 qc 4 84 1 rt 5
21 3 ra 2 53 1 rt 3 85 4 ra 2
22 4 ra 1 54 1 rt 3 86 4 ra 2
23 4 ra 1 55 1 rt 4 87 4 ra 2
24 4 ra 5 56 1 rt 5 88 4 ra 5
25 1 qc 3 57 1 rt 3 89 2 ra 1
26 2 qc 3 58 1 rt 4 90 1 ra 1
27 3 qc 2 59 1 rt 8 91 4 ra 7
28 4 qc 2 60 1 rt 5 92 4 ra 7
29 1 rt 1 61 1 ra 3 93 2 ra 1
30 1 rt 1 62 1 ra 3 94 2 ra 3
31 2 rt 3 63 1 ra 3 95 4 ra 8
32 2 rt 8 64 3 ra 6 96 4 ra 5

Table A10: Instance 3: Allocation results using the new formulation; external workers and shortfall

Allocation external workers Identifying shortfall

j k z Number of v j k z Number of u
1 ra 5 6 2 rt 3 2
1 ra 4 3 2 qc 5 1
1 ra 6 3 2 rt 1 1
1 ra 8 3 2 Rt 5 1
1 ra 9 2 2 rt 8 1
2 ra 5 3 3 rt 4 2
3 ra 4 2 3 rt 6 2
3 ra 7 2 3 rt 7 2
3 ra 9 2 3 rt 9 2

172 r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173
 Prediction of arrival times and human resources allocation

Table A10 Continued

Allocation external workers Identifying shortfall

j k z Number of v j k z Number of u
4 ra 8 1 3 qc 4 1
3 qc 5 1
3 qc 6 1
3 rt 5 1
3 rt 8 1
4 rt 7 4
4 rt 5 3
4 rt 8 3
4 qc 5 2
4 rt 4 2
4 rt 6 2
4 rt 9 2
4 qc 4 1
4 qc 6 1

Note: In all the tests performed the total labour cost, referred to the main
task of a docker, was taken as 25h/h for quayside crane operators and external
workers, 23h/h for yard crane operators, 21h/h for truck trailer drivers and
300h/h for the ‘dummy’ workers representing personnel shortfall.
Composition of the ship handling and housekeeping gangs has been kept
fixed throughout as
K Ship handling gang ¼ 1Q þ 2 RT þ 3 RA.
K Housekeeping gang ¼ 2 RT þ 3 RA.

r 2011 Macmillan Publishers Ltd. 1479-2931 Maritime Economics & Logistics Vol. 13, 2, 142–173 173