Multi-step Optimization of Logistics Networks:

Strategic, Tactical and Operational Decisions

Maarten Hendriks
Contents

1 Introduction 3
1.1 Logistics Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Decision levels in logistics networks . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Decision levels in a distribution network . . . . . . . . . . . . . . . 5
1.2.2 Decision levels in a container port . . . . . . . . . . . . . . . . . . . 6
1.3 Optimization of logistics networks . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 Illustrative example . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Multi-step optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4.1 Multi-step optimization of a distribution network . . . . . . . . . . 11
1.4.2 Multi-step optimization of a multi-terminal container operation . . 13
1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2 Design of a Distribution Network Topology 20

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 Bi-level Network Design Problem . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.1 Operational Decision Making . . . . . . . . . . . . . . . . . . . . . 26
2.2.2 Tactical Decision Making . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Computational Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.1 Comparison Branch and Bound versus Heuristic . . . . . . . . . . . 32
2.3.2 Performance of the Heuristic for Large Networks . . . . . . . . . . . 33
2.3.3 Statistical Robustness Analysis . . . . . . . . . . . . . . . . . . . . 33
2.4 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . 38

3 Strategic Allocation in a Container Operation 40

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2.2 Straightforward MILP . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.3 Alternative MILP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3.1 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3.2 Feasibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4 Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4.1 Two-step approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

1
CONTENTS 2

3.4.2 Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.3 Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.5 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . 67

4 Generation of a Robust Timetable 69

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.2 MILP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3 Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.1 Benefit of plan modification . . . . . . . . . . . . . . . . . . . . . . 77
4.3.2 WB-plan vs. nominal plan . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . 82

5 Joint Berth Allocation and Yard Design Problem 83

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2.2 MIQP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.3 Alternating Optimization . . . . . . . . . . . . . . . . . . . . . . . 90
5.2.4 Initial Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.3.1 Random Initial Conditions . . . . . . . . . . . . . . . . . . . . . . . 93
5.3.2 Sophisticated Initial Conditions . . . . . . . . . . . . . . . . . . . . 94
5.4 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . 96

6 Disruption Management in a Container Terminal 98

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.2 Berth time allocation . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.2.3 Berth position allocation . . . . . . . . . . . . . . . . . . . . . . . . 108
6.2.4 Crane scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.3.1 Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.3.2 Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.4 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . 114

7 Conclusions and Recommendations 116

7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2 Recommendations and Ongoing Work . . . . . . . . . . . . . . . . . . . . . 121
Chapter 1

Introduction

In the last decades a remarkable growth is noticeable in the number of people and goods
that have to be transported from one place to another by different kinds of resources, e.g.
buses, trains, airplanes and ships, but also transport belts, cranes, elevators and robots.
A set of these resources linked together, is called a logistics network. Such a logistics
network is usually run by a logistics provider, who faces the problem of delivering the
right amount in the right place at the right time. Due to the ever-growing complexity of
these networks, a logistics provider needs efficient tools to support its decisions leading to
a maximal service level at minimal costs. The decision making can typically be classified
into three levels: strategic, tactical and operational. The combined decisions on these
inter-related levels, or even on one level specifically, are often too complex to be solved
at once. A possible approach is to cut the total decision making problem into several
subproblems, which then are solved step by step and where necessary alternatingly.
This dissertation proposes, develops and tests multi-step optimization methodologies
to support the decisions of logistics providers in two particular logistics networks. Case
studies suggest the suitability of the developed methodologies for practical applications.
For each of the two logistics networks considered, the following generic conceptual ap-
proach is applied: we first reason about how to cut the overall problem into appropriate
steps and compare our choice to existing methods. Next, for each chosen step, a math-
ematical model is constructed, which is optimized for a real-life data set. Finally, the
results of the individual steps are discussed separately and the performance of the multi-
step procedure as a whole is evaluated.

1.1 Logistics Networks

Logistics networks are present in every-day life, in all sorts and sizes. A chip assembly
facility, the bus network in a town, and the luggage handling system at an airport are
just a few examples among many. Figure 1.1 shows some pictures of (parts of) typical
logistics networks.
A logistics network is commonly run by a logistics provider, who aims to deliver
people and/or goods in the right place at the right time. In order to meet these goals
at minimal costs, intelligent decisions at different levels of the logistics networks are
required. In the past, a logistics provider could easily keep track of all decisions to be
made. However, size and complexity of present logistics networks have grown that large,
that intelligent decision support systems are required to help the provider in making the
right decisions. This dissertation develops methodologies, insights and advices to support

3
CHAPTER 1. INTRODUCTION 4

(a) Luggage handling. (b) Railway switch region.

(c) Parcel service.

Figure 1.1: Examples of logistics networks.

service providers in their decision making in the following two particular logistics networks
(see also Figure 1.2):

1. A supply and demand distribution network,

2. A multi-terminal container port.

In the next section, we first discuss the common categorization for the decision making
in logistics networks. Then we discuss these levels in detail for the two logistics networks
addressed in this dissertation.

1.2 Decision levels in logistics networks

The three levels of decision making are categorized based on their time scale and the
extent of influence they have on the network’s performance:

• The strategic level is concerned with the number, location and size of particular
network nodes (e.g. facilities and resources). Since the establishment of such a
node (e.g. an airport or a train station) usually requires a lot of money and effort,
these decisions have a long lasting effect on the network performance. Although
CHAPTER 1. INTRODUCTION 5

(a) Distribution truck. (b) Container terminal.

Figure 1.2: Parts of the two logistics networks addressed.

strategic moves are continuously considered, actual changes are only executed after
a number of years. Hence, these decisions are defined to be long term decisions.

• On the tactical level, a logistics provider has to decide on the design of i) the
network topology, i.e. how the established nodes should be connected, and ii) the
detailed layout of the particular nodes, i.e. the location of resources and products at
one facility. To preserve continuity for employees and customers, and to restrict the
network complexity this network design should not change too frequently. Decisions
regarding the tactical level are therefore reconsidered after a number of months and
defined as medium term decisions.

• The operational level is concerned with short term decisions, i.e. every day, every
minute or every second (re)routing and (re)scheduling of people or goods through
the established network.

The above described categorization in the decision making is now applied to the two
particular logistics networks addressed in this dissertation.

1.2.1 Decision levels in a distribution network

A distribution network consists of a production company, which supplies different prod-
ucts at various facilities, and several retailers, which have a demand for these products.
Dependent on the size of the network, production facilities and retailers are spread over a
region, country or continent. Mismatch in supplies and demands might lead to a surplus
or shortage of a particular product at a particular time. Since neither the supplier nor the
retailers want to deal with buffering this variability in supply and demand, they hand the
job to a third-party logistics service provider (LSP). His task is therefore to ship the right
amounts from production facilities to retailers, possibly using intermediate warehouses
for i) temporary storage of products to compensate the variability in supply and demand,
and ii) consolidation of products so as to leverage on the economy of scales principle. The
CHAPTER 1. INTRODUCTION 6

above described three level decision making can be applied to such a distribution network
as follows:
For an LSP, the number, locations and capacities of suppliers and retailers in the
distribution network are given. On a strategic level, the LSP has to decide on the number,
location and capacity of intermediate warehouses. While making these decisions, not
only has the current suppliers and retailers, but also potentially new markets have to
be considered. These decisions have a long lasting effect on the performance of the
distribution network and are therefore reconsidered on the long term time scale.
Once the location and sizes of all warehouses are known and fixed, the LSP has to
make tactical decisions on the links/connections between these facilities, i.e. the network
topology. In the distribution network considered, these links are the so-called line hauls
being the roads, rails and/or waterways involved in the movement of freight between two
facilities. According to the economy of scales principle, a line haul’s efficiency increases if
more products flow through it. The average line haul utilization can simply be increased
by decreasing the total number of line hauls in the topology. Hence, an LSP strives to
construct a topology with a small number of line hauls, which still performs well on an
operational level. A line haul however cannot be established and removed on a short term
basis, since it has a large impact on the operations of both involved facilities. Namely, after
establishing a line haul, routings and schedules have to be modified and employees have
to change their tasks according to these modified plans. Hence, to provide consistency,
a network topology should not change too frequently. The establishment of line hauls is
therefore reconsidered on the medium term time scale.
The operational tasks involve short term decisions: day-to-day routing and scheduling
of the shipments along the chosen line hauls between the different facilities. Given the
network topology and given supply and demand for a limited time horizon, the LSP has
to decide on how much to send through which link facing costs for transportation, storage
and penalties for early and late delivery.

1.2.2 Decision levels in a container port

Since 1960, containerization has grown rapidly and nowadays annually up to 114 million
TEU’s (1 TEU, Twenty feet Equivalent Unit, is a container of length 20 feet, width 8 feet
and height 8 feet) are transported all over the world. In this world-wide network, a con-
tainer port not only serves as a connection between land and sea container transportation,
but also as a transshipment hub for forwarding containers between vessels.
A terminal operator coordinates and performs the logistics processes of discharging,
loading, transporting and storing containers of various vessel lines in a particular terminal.
Each vessel line owns a vessel fleet to maintain several repetitive loops along ports all over
the world. Commonly, the number and phasing of the vessels of one loop are such that
one vessel calls on each of its ports exactly once a week. Hence, the terminal operator has
to service each customer (line) according to a cyclic timetable, which is repeated week
after week. One can compare such a timetable with a bus or train schedule. The general
levels of decision making in logistics networks can be applied to a multi-terminal container
port as follows:
Dependent on its assets, a terminal operator provides its services in a certain number
of terminals along the world. Once a new terminal is to be built or an existing one is
to be overtaken somewhere, a terminal operator can take place in a competitive bidding
procedure for operating this terminal in the future. These are long term decisions and
CHAPTER 1. INTRODUCTION 7

have a large impact on the overall turnover of the operator. The expansion of the number
of terminals is very expensive, however necessary to cope with the exponential growth
of container transport. Hence, an operator has to continuously anticipate on the future
market while considering the services in another terminal. Next, strategic decisions have
to be made on the way to operate a terminal, e.g. which kind of resources (straddle
carriers vs. trucks and stacking cranes) are used to transport the containers between
quay and yard and how many quay cranes are required at the quay.
In many multi-terminal container ports, various operators take care of the logistics
processes for container handling. Commonly, the tasks are divided such that one terminal
operator is responsible for one terminal (or at least for the major share of one terminal).
In an increasing number of ports (e.g. Singapore, Rotterdam and Antwerp) however,
one terminal operator is responsible for multiple terminals. Given the quay lengths and
storage capacities of the terminals, and given the load compositions of the calling loops,
the first tactical problem is then to i) allocate a terminal and ii) a berthing time interval
to each of the loops. Secondly, a berth position has to be allocated to each loop and an
appropriate yard layout (which containers to stack where) has to be constructed. Together
this results in a tactical timetable, which is reconsidered on the medium term time scale.
The tactical timetable depicts the allocation if all vessels arrive perfectly in time.
However in practice, vessels are sometimes early or late (e.g. due to breakdown or bad
weather conditions), and may have different call sizes and/or compositions each week.
Moreover, quay cranes and other resources may brake down for an unknown period of
time. The daily operational tasks of a port operator involve the management of the
disrupted system to serve the vessel lines as good as possible at minimal costs. First,
each vessel has to be allocated to a specific berth position within its terminal. The
reference berth position of a vessel is usually taken closest to the position of its export
containers in the stack. In this way, the travel distance of container carriers between
vessels and stacks is reduced. Second, a schedule for quay cranes along with resources
and its drivers has to be constructed to process a vessel within the agreed service time.
These decisions are usually made every eight hours (one shift), and sometimes even a
replanning takes place after four hours (half a shift). The detailed sequence of actual
discharging, loading, transporting and stacking containers is updated at every container
pick-up and drop-down.

1.3 Optimization of logistics networks

From the previous section it becomes clear that a logistics provider faces many decisions so
as to run a logistics network efficiently. To satisfy his customers and to maximize his own
profit, he has to strive for decisions that lead to a maximum customer satisfaction (service
level) at minimum costs. These two objectives however are conflicting and the provider
continuously considers a higher service, at the expense of additional resources (equipment
and personnel). Customer satisfaction can be quantified easily, since logistics providers
are usually charged by their customers when services are not provided in time. This
enables the logistics provider to explicitly trade off his service level against his investment
costs.
An appropriate way to optimize the decisions is by constructing and optimizing a
mathematical model of (a part of) the logistics network. First, one has to select a vector
of decision variables x, which have to be decided on by the, in this case, logistics provider.
CHAPTER 1. INTRODUCTION 8

Next, a suitable objective function f (x) has to be constructed, which describes how the
system performance (total costs) depends on the decision variables x. Finally, physical
limitations or theoretical bounds on (parts of) the system have to be incorporated. To
model this, one distinguishes between inequality constraints g(x) ≤ 0 and equality con-
straints h(x) = 0 represented by functions g(x) and h(x) of x that are bounded by or
fixed to zero, respectively.
An optimization tool can then be used to optimize the model with respect to the
objective function f (x) (e.g. consisting of investment costs and costs for not delivering in
time). The output of the optimization is a set of optimal decisions that leads to minimal
objective costs. The generic mathematical form of an optimization problem looks as
follows:

min f (x)
x
subject to g(x) ≤ 0
h(x) = 0.

1.3.1 Illustrative example

To illustrate the translation of a logistics problem into a mathematical problem, we apply
the above described approach to a simple, every day example. Imagine, it is Monday
afternoon 17:55 and suddenly you realize today is the last day to return a number of
borrowed books to the library without paying a fine of 0.20 euro per book. The library
closes at 18:00 and is located 4 km outside your home town, so going by bike today is not
an option. You can take the bike tomorrow, but then you have to pay the 0.20 euro fine
per book. Furthermore, there is the restriction that you are able to carry only 10 books
on your bike and are not willing to bike the route more than once. If you take the car
right away, you will deliver the books just in time, but spend 0.19 euro on gasoline per
kilometer. What do you do? It is clear that the books cannot be returned without having
some costs, no matter what decision you make: if you go by bike, you pay the fine, if you
go by car, you pay the gasoline.
Assume you take the bike and you have to return a number of N books. The fine for
one book is 0.20 euro, which makes the fine for N books N · 0.20 euro. Moreover, there
is the condition that you can only carry up to 10 books on your bike and you do not
want to have to ride the bike to the library more than once. This means that you can
only choose between the bike and the car if no more than 10 books are to be returned, in
mathematical terms if N ≤ 10.
Now assume you go by car and pay 0.19 euro per kilometer on gasoline. To drive to
the library and back, a distance of 8 kilometers has to be covered, which will cost you
8 · 0.19 = 1.52 euro. Now dependent on your decision on the transport and the number
N of books, you pay either N · 0.20 euro or 1.52 euro.
The question here is, given these conditions how to incur minimal costs. Mathemati-
cally, the choice of your transport can be represented by a variable, which we denote by x.
Dependent on the choice, this variable x is given a different value. We define the variable
x to be 1 if you take the car, and zero if you take the bike. This can be represented as
follows:
CHAPTER 1. INTRODUCTION 9


1 if you drive your car today,
x = (1.1)
0 if you drive your bike tomorrow.

The costs can now be written as a function f (x) of this decision variable x:

f (x) = 0.20 · N · (1 − x) + 1.52 · x. (1.2)

If you would decide to take the car (x = 1), then the term (1 − x) in (1.2) becomes 0
and costs for fines are excluded. In this case, the only cost contribution remaining is the
gasoline costs equal to 1.52 · 1. If you would decide to take the bike (x = 0), then you
only pay the fine of 0.20 · N and 0 for gasoline (1.52 · 0).
Finally, the restriction on the number of books N you can carry on your bike has to
be formulated dependent on the decision variable x. We already saw that if N ≤ 10,
you can still decide on taking the bike or the car, and hence x can be either zero or one.
If N > 10, you have to take the car, and hence for this case we should enforce x to be
one. Mathematically, this can be achieved as follows. We propose the next inequality
constraint:
g(x) = −Sx + (N − 10) ≤ 0, (1.3)
where S is a number still to be determined. The inequality constraint g(x) in (1.3) implies
that the term −Sx + (N − 10) has to be smaller than or equal to zero. For the case where
x = 0 this is valid only if N ≤ 10, for the case where x = 1 this is valid if S ≥ N − 10.
Hence, we define S ≥ N − 10, since then independent on N , you can always choose the
car (x = 1) and still satisfy (1.3). However, to satisfy (1.3) for x = 0, it has to be that
N ≤ 10, and this is exactly what we strived for.
Now, a decision on x has to be made such that the costs represented by the objective
function f (x) are minimal, and the inequality constraint g(x) is satisfied. This can be
depicted in the general form of an optimization model as follows:

min 1.52 · x + 0.20 · N (1 − x)

x (1.4)
subject to −Sx + (N − 10) ≤ 0

For the given parameter values, the dots in Figure 1.3a depict the minimal costs as
a function of the number of books N you have to return. The thin black dotted lines
represent the individual contributions of costs for going either by car or bike. Furthermore,
the vertical grey dotted line depicts the maximum number of books you can carry on your
bike. From this figure, the following is noticed: for a small number of books, the minimum
costs equal the fine costs (you should go by bike), but from the break-even point on, the
minimum costs follow the gasoline costs (you should go by car). Apparently, for a small
number of books (up to seven), it is less costly to take the bike the next day (x = 0).
If you need to return between 8 and 10 books it is cheaper to drive the car right away
(x = 1). Finally, if you have to carry more than 10 books you have to use your car anyway
(x = 1).
The parameter values have a crucial effect on the best choice and the corresponding
minimal costs. Figure 1.3b for instance depicts the minimal costs for exactly the same
problem, except that now the fine per book is 0.10 euro. From this figure it can be
 CHAPTER 1. INTRODUCTION 10

2 2
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
Costs

Costs
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
N N
(a) Fine per book is 0.20 euro. (b) Fine per book is 0.10 euro.

Figure 1.3: Minimal costs as a function of the number of books to be returned N .

concluded that it is cheaper to drive the bike if you have to return up to fifteen books.
However, if you have to deliver more than 10 books you have to take the car, resulting in
a large cost increase at N = 11.
In the above described example, only one decision has to be made, which can eas-
ily be optimized in one’s head. One can imagine however that in real-life logistics net-
works, many decisions have to be made facing a multi-objective function, subject to many
constraints. Hence, finding the optimal decisions is far from trivial. For these kinds of
systems, optimization by means of a mathematical model can be very helpful. This disser-
tation considers the decision making in two complex logistics networks, defines practically
relevant and interesting subproblems, formulates appropriate mathematical optimization
models and applies them to real-life case studies.

1.4 Multi-step optimization

As has been mentioned before, the considered logistics networks are of such a size, that the
combined decisions on the strategic, tactical and operational levels or, even at one level,
cannot be solved at once within the time allowed. Operational decisions for instance have
to be made every second or minute of the day. A model that runs for an hour to propose
these operational decisions is useless. The approach in this dissertation is to cut the
overall problem into a number of subproblems each of which i) is practically interesting in
its own right, and ii) can be solved within the time allowed at the level(s) concerned. The
proposed models for instance enable to construct decisions on the long and medium term
i i i
time scale within hours, while short term decisions can be constructed within minutes.
i The price we pay by solving thei subproblems sequentially is that the solutions found i
are no longer guaranteed to be optimal. In this dissertation, limited attention is paid to
quantify the (expected) deviations between the found solution and an optimal solution.
The focus is on the gains that can be achieved with respect to the solutions as currently
applied in practice. Although some currently applied solutions might result from man-
agerial decisions or negotiations that are not covered by our models, a quantification of
the induced additional costs is made explicitly. This provides insights in the costs of
CHAPTER 1. INTRODUCTION 11

modification and can support a logistics provider in his future decision making.
Solving the multiple subproblems can be done either i) sequentially, i.e. the output
of one subproblem is fixed and used as an initial condition for the next subproblem, or
ii) alternatingly, i.e. one alternates between different subproblems where the output of
one subproblem is the input for the other and vice versa. In this dissertation, dependent
on the subproblems discussed either a sequential or a alternating solution approach is
applied.
Several studies on multi-step optimization for large systems can be found in literature,
where either sequential or alternating solution approaches are applied. The study in [12]
for instance discusses several approaches to solve problems in fluid and solid mechanics,
in which coarse scale phenomena influence local phenomena and vice versa. These studies
cut the global problem into different time and/or space scales to retrieve tractable sub-
problems, which can be solved sequentially or alternatingly. In this dissertation as well,
most subproblems are formulated based on the different scales/decision levels involved.
However, sometimes a problem is cut into two or more subproblems, which are all at the
same scale/decision level. Therefore, the terminology ”multi-step optimization” is used
rather than ”multi-scale optimization”.
Some of the subproblems we consider are, to some extent, similar to problems ad-
dressed individually in literature, others, however, are not and result from the cuts in our
overall multi-step approach. In the next two subsections, we first highlight the overall
problem that is considered for the distribution network (Section 1.4.1) and the multi-
terminal container operation (Section 1.4.2). Moreover, the multiple steps and solution
methods for the corresponding network are highlighted and main results are summarized.
Relevant references are cited, however a detailed literature review for each subproblem is
only given at the beginning of the chapter that discusses the concerning subproblem.

1.4.1 Multi-step optimization of a distribution network

The three-level decision making in a distribution network has been discussed before and
can be summarized as follows: i) Strategic decisions involve the number, location and
capacity of warehouses, ii) on a tactical level one has to decide on the network topology,
i.e. which links (highways, railways, waterways) to use between suppliers, warehouses
and retailers and iii) the daily routing of vehicles (trucks, trains, vessels) through the
chosen topology are considered to be operational decisions. In Figure 1.4, an illustration
of the distribution networks considered is depicted. Production facilities, warehouses and
retailers of different sizes are, in this example, spread among the Netherlands. The red
lines represent the main highways connecting all facilities.
In this study, the strategic decisions are assumed to have been made and hence the
number, location and capacity of warehouses are fixed. Furthermore, the number, location
and average supply rate of production facilities, and the number, location and average
demand of retailers are assumed to be given. Additionally, the actual daily supplies and
demands per product type for a limited future horizon ahead (couple of days) are assumed
to become available each day. Having this information, a third-party logistics provider
deals with the joint problem of i) the generation of a tactical network topology on the
medium term time scale and ii) the daily scheduling and storing of different types of
products in the generated network.
Looking at Figure 1.4 again, this means that the provider has to select i) which of the
red links (highways) to use for a longer time period and ii) how much to send through the
 “FMcartoon1˙temp” — 2008/10/21 — 15:04 — page 1 — #1
i i

CHAPTER 1. INTRODUCTION 12

Legend
Production
facility

Warehouse

Consumer

Highway

Figure 1.4: Illustration of a distribution network.

selected links each day. The LSP strives for a topology that has i) a small number of line
hauls to have a high link utilization so as to leverage on the economy of scales principle
and ii) a fixed number of links to provide consistent routings and schedules and reduce the
organizational complexity, while still the operational performance is satisfactory. For real-
life distribution networks, however, it is very time consuming to evaluate the operational
performance of each possible topology to find the one that fits best these two objectives
best. Hence, the complex problem of tactical and operational decisions is cut into two
problems, i) the topology design problem and ii) the daily routing problem, which are
solved in an alternating fashion: a coupled bi-level optimization method is proposed that
evaluates only a very limited number of topologies to find one with a small number of
i i
links that still has a satisfactory operational performance.
Aimathematical model is constructed to describe the product flows as a function
i of the
daily operational decisions given a topology and given supply and demand on a restricted
future horizon. This model is used to determine the minimal operational costs for that
particular topology and these particular supply and demand time series. The bi-level
method alternates between the tactical level of decision making and the operational level
of decision making. The method is based on iteratively dropping links from a topology
on the tactical level dependent on its performance at the operational level.
Results of a case study suggest that this method is very fast and still yields accurate
solutions. Moreover, results suggest that the constructed network topologies are insensi-
tive to changes in second and higher order moments of supply and demand distributions.
The topology however appears to be sensitive to changes in the means of supplies and
demands.
Many theories and solution approaches on the design of logistics networks and supply
chains can be found ([11], [18], and [44]). The problem considered in this dissertation
however investigates a specific and practically interesting problem, which to our knowledge
has not yet been explored.
CHAPTER 1. INTRODUCTION 13

1.4.2 Multi-step optimization of a multi-terminal container op-

eration
The three level decision making in a container operation can be summarized as follows:
On the strategic level, the operator considers the enlargement of the number of terminals,
the kind of operations within each particular terminal and the number of quay cranes
in each particular terminal. Given the terminals and their capacities, the operator can
service a certain number of periodically calling container vessels, and has to construct
an appropriate tactical time table and berth allocation plan. Since in practice, container
operations are heavily disturbed due to bad weather conditions and breakdowns, the
terminal operator continuously needs to reschedule on an operational level in order to
return to the tactical plan as soon as possible.
This study addresses decision problems faced by a terminal operator on the strategic,
tactical as well as on the operational level. Similar to the distribution network problem,
the ioverall problem cannot be solved to optimality and is for this case cut into four i
subproblems, which are “PSAintro7˙temp”
then solved — step by step
2008/12/3 (either
— 10:04 — page sequentially
1 — #1 or alternatingly).
i i
Figure 1.5 shows a graphical illustration of the decisions at the various levels. The colored
dotted rectangles indicate the selected subproblems, which are shortly discussed in the
next subsections.

Strategic decisions
- Number of terminals
- Type of operation
- Quay crane capacity per terminal

Tactical decisions 1

- Terminal, arrival & departure, and crane

capacity for each call

- Robust timetable per terminal 2

- Berth positions for vessels at the quay, and

3
stack positions for containers in the yard

Operational decisions
- Arrivals and departures for each vessel
4
- Berth positions for each vessel
- Crane scheduling

Figure 1.5: Illustration of the chosen subproblems for the container operation.

i i
CHAPTER 1. INTRODUCTION 14

Since the decisions on the strategic level induce the highest costs and have the largest
impact on the system’s performance, it makes sense to start solving the subproblem(s)
at the strategic level, and use the results as inputs for solving the subproblem(s) at
the tactical level. These solutions in turn are input for the solution methods for the
subproblem(s) at the operational level.
In literature many studies can be found on container operations, some of which discuss
problems that to some extent are similar to one of the four main steps addressed here. The
authors in [45], [46], [49] and [43] give extensive overviews of the main issues in container
ports and discuss existing solution approaches. Comparisons between our studies and
existing ones are made in the next subsections.

Step 1: Terminal, arrival and departure, and crane capacity allocation to

vessels
Nowadays, mega-container ports consist of multiple terminals, which are relatively close
to each other. In most of these ports, the logistics terminal operations are performed by
multiple container operators. Typically, one terminal operator takes care of the logistics
processes in one terminal. However, in an increasing number of ports (e.g. Singapore,
Rotterdam and Antwerp), one terminal operator is responsible for a number of terminals.
The logistics problems in such ports can no longer be considered per terminal for two
main reasons. Firstly, peaks and troughs in quay crane utilization should be avoided and
vessel calls should be spread evenly over the various terminals. Secondly, transshipment
containers will very likely generate inter-terminal traffic resulting in costs that should be
i i

taken into consideration. All possible flows of containers through a number of terminals
i
“portn7nc˙temp” — 2007/8/16 — 16:28 — page 1 — #1
i

managed by one operator are depicted in Figure 1.6.

Seaside
1
inal

Import Import
Termina

Export Export
Term

l3

Termin
al 2

Inter-terminal
Transport
Import
Land side
Export

Figure 1.6: Container flows in a cluster of multiple terminals.

i i

This research is supported by the company of PSA HNN, that operates a number of
i i

container terminals in Antwerp, Belgium. The terminals’ length, and the set of current
customers (shipping lines) calling at this port are. The actual remaining question on
CHAPTER 1. INTRODUCTION 15

the strategic level is hence about the number of quay cranes minimally required, or in
particular: ”Can we do more business with the same number of quay cranes and where is
this spare capacity?”.
The current policy of the terminal operator is to strive for satisfying the preferred
tactical arrival and departure times and preferred berthing terminal of shipping lines to
provide a high service level. It is interesting to determine the potential quay crane spare
capacity and the potential reduction in inter-terminal transport if small modifications
to the shipping lines’ preferences are allowed. An optimization model that allocates a
terminal, arrival and departure times, and crane capacity to a set of cyclically arriving
vessels with known call sizes can already yield nice insights in potential spare capacity
and cost reductions. This actually is the first in the sequence of optimization steps with
respect to container operations. The allocation of exact positions of the vessels within a
terminal and the actual quay cranes assigned to them is not relevant yet and is left to
be determined in a subsequent subproblem. In this way, we are left with an practically
interesting, well-defined problem that is still tractable from a computational point of view.
Note that in this first step, we strive for improvements at the strategic level (spread work
evenly, identify spare capacity) as well as improvements on the tactical level (reduction
in carriers for inter-terminal transport).
An appropriate mathematical optimization model is proposed, which includes decisions
regarding i) which terminal to call, ii) arrivals and departures and iii) crane capacities and
incorporates flow equations for the container transport between the different terminals.
The multi-objectives are to minimize i) deviations from preferred arrival and departure
times, ii) the minimally required crane capacity, and iii) the amount of inter-terminal
container transport. For randomly generated problem instances, the proposed model can
be solved substantially faster than a more common approach. Subsequently, represen-
tative data, provided by PSA HNN, is used to set the parameter values in the model.
Results of a case study suggest that significant reductions in both minimally required
quay crane capacity and inter-terminal transport can be gained if only small modifica-
tions to the current allocations of only a part of the shipping lines were allowed. Many
studies exist on the berth allocation problem within one terminal. Surprisingly, to our
knowledge, there is no study on the allocation to a cluster of inter-related terminals so
far. A recommendation in [39] however indicates the growing need for the optimization
of a multi-terminal container port.

Step 2: Increasing the robustness of a tactical timetable per terminal

The first optimization step results in a cyclic timetable per terminal. If all vessels arrive
accordingly, this timetable can be executed over and over again. However, the arrival
times of container vessels in a terminal are heavily disturbed by delays in other terminals
in their loop and by bad weather conditions during travel. Therefore, a terminal operator
cannot expect the vessels to arrive exactly in time. On the other hand, a shipping line
cannot demand immediate service if it is delayed for instance one day. As a compromise,
the terminal operator and each shipping line agree on two kinds of arrivals: i) within and
ii) out of a so-called arrival window, which is placed around the scheduled arrival time and
typically has a width of eight hours. If a vessel arrives within its window, the operator
guarantees to operate the vessel within an agreed process time. If a vessel arrives out of
its window, the terminal operator is not bound to any process time at all.
The goal of the second step optimization is to slightly modify a terminal’s tactical berth
CHAPTER 1. INTRODUCTION 16

plan, determined in the first step, into a tactical berth plan that is robust to all scenarios
where all vessels arrive anywhere within their arrival window. In our definition, a berth
plan is robust with respect to a given set of arrival windows if a feasible solution exists for
each arrival scenario where all vessels arrive within their windows. The price for achieving
this robustness is then the additional crane capacity reservation that is required in the
worst case arrival scenario where all vessels arrive within their windows. The problem
is hence to construct a window-based berth plan that minimizes the maximally required
crane capacity for all scenarios where vessels arrive within their arrival windows. Arrival
windows, and quay and crane capacity reservations have to be allocated to each of the
vessels in the set. Note that although sufficient quay and crane capacities are reserved,
the actual vessels’ berth positions and quay cranes operating are left to be allocated in
a subsequent subproblem. However, still we end up with a tractable problem that is
practically interesting on its own right.
A mathematical optimization model is proposed that constructs a window-based plan
and minimizes the maximally required crane capacity. Results on representative data
provided by PSA HNN suggest that slightly modifying a berth plan, generated in the first
optimization step, yields a significant reduction in the maximally required crane capacity.
As a particular case, the optimization model finds a nominal berth plan, i.e. a berth
plan that neglects disturbances and hence ignores the arrival window agreements. Results
suggest that the window based plan requires slightly more crane capacity if vessels’ arrivals
would only slightly deviate from the scheduled arrivals for zero and narrowly bounded
arrivals¿ However, the window based plan is much more robust than the nominal plan in
case of relatively large deviations (which are still within the arrival window bounds).
The concept of building in pro-active robustness in tactical timetables has already
been introduced in airline applications [7], [1] and [29], and railway applications [5], [50]
and [5]. To the best of our knowledge, only one study [37] addresses the stochastic berth
allocation problem by building in some kind of pro-active robustness. Given a timetable
and delay distributions, the study in [37] allocates berth positions to vessels to minimize
overlaps of vessels (i.e. two vessels at the same place at the same time) and deviations from
preferred berth positions. In contrast to this study, we explicitly guarantee no overlaps
and do use the flexibility of modifying i) the timetable and ii) crane capacity allocation
to increase the robustness.

Step 3: Allocation of vessels’ berth and containers’ stack positions

Once the (window-based) berth plan has been constructed for each terminal and hence
the tactical arrival and departure times are known, the next step is to allocate berth
positions to the vessels. Containers to be loaded onto a vessel arrive at the terminal
between a few weeks up to a few hours prior to the vessel’s departure time. These
containers are temporarily stored in designated areas in the terminals’ yard. Once the
vessel has been positioned along the terminal quay, carriers in between the quay and the
yard start moving containers from the vessel to designated areas (unloading process) and
from designated areas to the vessel (loading process). The berth position of a vessel and
the position of its designated container area(s) hence determine the distance that has
to be covered by the carriers. An illustration of a container terminal and the relevant
logistics activities are depicted in Figure 1.7.
Given the tactical timetable and average call sizes, the goal of the third optimization
step is to allocate i) vessels’ berth positions at the quay and ii) container area positions
CHAPTER 1. INTRODUCTION 17

Group of Containers
Straddle Carrier
Quay Crane
Vessel

Figure 1.7: Logistics activities in a container terminal.

in the yard such that the total carrier travel distance is minimized. A reduction in this
travel distance would not only lead to a smaller fuel consumption of the carriers, but also
to a possible decrease in the idle time of the quay cranes and consequently to a certain
amount of spare crane capacity.
For computational complexity reasons, this problem in itself is split into two subprob-
lems, being the vessels’ berth position problem and the container area position problem.
For both these problems, we formulate independent mathematical optimization problems,
which are coupled in the objective function that minimizes the total carrier travel distance.
Due to this structure, the problem can be efficiently solved in an alternating fashion: we
start solving the berth position problem for chosen initial values of the container areas
positions. Next, the generated values of the berth positions are passed to the container
area position problem and fixed as parameters. Subsequently, the container area posi-
tion problem is solved and generated values are passed back as parameters to the berth
position problem. This procedure is repeated until the objective value does no longer
decrease. Since this alternating method finds a local minimum and heavily depends on
the initial condition, an additional optimization model is proposed that finds a proper ini-
tial condition. Starting from this initial condition, the alternating procedure finds a local
minimum that outperforms the best of all solutions found by starting from an extensive
number of random initial conditions.
Results of a case study on a representative timetable of one of Antwerp’s terminals
depict vessels’ positions and a yard lay out that induce a substantially smaller travel
distance than the one as currently applied. Although a feasible solution in the second
step optimization might turn out to be infeasible at this third level, not one of many
experimental instances for typical quay utilizations in the world’s busiest terminals reveals
this problem. Although many studies on the individual berth allocation problem and the
individual yard lay-out design problem can be found, ([15] and [33], respectively), a study
that addresses the joint problem has, to the best of our knowledge, not yet been conducted.
CHAPTER 1. INTRODUCTION 18

Step 4: Online reallocation of a terminal under disturbances

The results of the former three steps together generate i) a robust timetable, ii) vessels’
berth positions and iii) container positions in the yard. Such a tactical cyclic allocation
is commonly reconsidered on a medium term time scale. However, as mentioned before,
a container operation is exposed to all kinds of disturbances, which requires an online
management system that observes the disturbances and reacts to them by reallocating
the affected vessels and resources.
In the fourth optimization step, an approach is proposed to act upon disturbances
on vessels’ arrivals, changes in load compositions and break-downs of quay cranes. The
approach is similar to the one constructed for the operational routing of product flows
through the distribution network: In each iteration (time) step of the approach, (expected)
parameter realizations over a limited future horizon are taken into consideration while
determining the operational decisions for the current time step.
Each iteration step in itself consists of three sequential steps. In the first step, the
expected arrivals and the expected call sizes of all vessels within the horizon are considered.
Crane capacities are allocated such that vessels expected to arrive within their windows
depart in time and vessels expected to arrive outside their windows are processed as fast
as possible (without spending too much additional resources). Once the start and end
berth times of each vessel within the horizon have been determined, berth positions at
the quay are allocated in the second step, considering the call sizes and call compositions
and the position of the containers in the yard. In the third step, the actual quay crane
scheduling is performed for all vessels in a more limited future time horizon. After solving
these three steps sequentially, the operational decisions of only the current time step are
actually executed. Then, the sequence of the three optimization steps is performed for
the next time step.
Since each iteration step can be solved within a couple of minutes, this approach is
very suitable for practical setting. Namely, in practice, an operational plan is typically
updated each hour. Simulation experiments are performed to trade off the carrier travel
distance against the deviations from preferred berthing positions. Namely, dependent on
the actual load compositions of the vessels, the optimal berth position (optimal in the
sense that the carrier travel distance is minimum) of a vessel might deviate from the one
derived in the tactical timetable.
Several studies consider the dynamic berth allocation problem in which vessels arrive
in a terminal while work is in progress (see [45] and [15] for overviews). A rolling horizon
approach like in this study, that observes and reacts on stochastic arrivals, stochastic load
compositions as well as crane break downs however cannot be found so far.

1.5 Outline
The outline of the dissertation is as follows: In the chapters 2 through 6, the distribution
network problem and the four main steps for the multi-terminal container operations are
dealt with. For all of these chapters, the structure is the same. Firstly, a detailed literature
review is given and the considered problem is positioned within existing studies. Secondly,
the problem is formally phrased, and assumptions and model parameters are properly
arranged. Third, a solution approach is proposed and mathematically formulated. Fourth,
experiments on random instances or representative data are performed and results are
discussed. Finally, conclusions and recommendations are given. In Chapter 7, the main
CHAPTER 1. INTRODUCTION 19

findings from the various chapters are summarized and conclusions and drawbacks on the
overall performance of the multi-step approaches in the two considered logistics networks
are discussed.
Chapter 2

Design of a Distribution Network

Topology

2.1 Introduction
The logistics networks considered in this chapter consist of production facilities, ware-
houses and consumers, which are geographically connected by links, e.g. roads, railways,
waterways. In long-distance transportation networks, the distribution of products is often
performed by a third-party logistics service provider (LSP). As opposed to the common
supply chain studies, the problem addressed here is not to control the amount of products
in the supply chain. Instead, this study addresses one of the typical services an LSP pro-
vides: supply and demand cannot be influenced by the LSP, but are simply a (stochastic)
reality, that is revealed only a few days in advance. A supplier pushes its products from
several production facilities to the logistics provider. By the same token, consumers pull
products from the network. The logistics provider then has to decide whether to store
the products in a warehouse or to immediately match them with a consumer demand.
The decisions regarding design and operation of distribution networks can typically be
classified into three levels: The strategic level, the tactical level and the operational level.
The strategic level deals with decisions regarding the number, location and capacities
of warehouses. These decisions have a long-lasting effect on the system’s performance.
The tactical level includes decisions on which transportation links to actually use, i.e.,
the design of the network topology. To preserve continuity for employees and to restrict
the complexity of organizational tasks such a network topology should not change too
frequently. These tactical decisions are therefore reconsidered somewhere between once
a month and once a year. The operational level refers to day-to-day decisions such as
scheduling and routing of the shipments given the tactical topology.
In this chapter, we focus on the interplay of the operational level and the tactical level,
i.e., we are interested to determine the topology of the network for a given operational
strategy and given operational parameters. We specifically strive for a robust topology,
which can be established for a relatively long period of time (months or years) and is still
cost-effective when operational parameters (supply and demand distributions) change.
This research is supported by the LSP ”Koninklijke Frans Maas Groep”, one of the
leading logistics service providers in Europe. One of the typical services this LSP pro-
vides can be illustrated by the following scenario: a big company produces different
types of products at various production facilities throughout Europe. These products
are consumed by all kinds of industries at various plants throughout Europe. Average

20
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 21

consumption of each type of product by each consumer is known, but the daily demands
fluctuate. On the other hand, due to production in batches, machine break down etc.,
also the supply of the different types of products fluctuates on a daily basis. Regardless
of this, the consumers still expect, dependent on the type of product, a certain level of
in-time delivery on a daily basis, where it does not matter from which production facility a
consumer receives its products. Neither the consumers nor the supplier want to deal with
the logistics of buffering the variability of supplies and demands and decide to hand this
task to an LSP. Its task therefore is to supply the desired quantities at the desired day.
The LSP can compensate for the stochasticity in supplies and demands by temporarily
storing products in warehouses rather than shipping them directly from supplier to con-
sumer. Furthermore, shipping via a warehouse enables product mixing so as to leverage
on the economy of scales. On the other hand shipping through a warehouse introduces
additional delay due to the handling activities, e.g. (un)loading and consolidation. The
task of the LSP is therefore to decide which transportation links to use on the long run
and how much of which product(s) to ship through them each day, such that total costs,
including transportation and storage costs and penalties for early and late deliveries, are
minimized.
In this chapter, we address the above described problem and determine the structure
of the tactical network topology assuming the underlying operational control problem is
approximately solved via a rolling horizon approach. We specifically strive for a network
with a very small number of links that still has close to minimal operational cost. Reducing
the number of links (or, equivalently, assuming unit cost per link) is a surrogate for a
more detailed optimization at the tactical level for which the associated costs are hard to
quantify:

• A reduction in the number of links reduces the complexity of the network and with
that the complexity of the organizational tasks of an LSP,

• Each link involves a fixed cost due to contracts and overhead,

• Reducing the number of links leads to thicker flows per link.

The overall problem can thus be characterized as a bi-level bi-objective joint network
design and operation problem. The upper (tactical) level chooses the links and the lower
(operational) level, for given choice of links, chooses the material flow. Our goal is to
provide an efficient method to approximate the trade-off front of link cost versus opera-
tional cost for a given time series of supply and demand, and in particular to determine
a network topology that is (i) approximately optimal in both objectives (small number of
links and minimal operational costs) and (ii) robust with respect to stochasticity (second
and higher order moments) of supply and demand. An extensive amount of experiments
is performed to confirm these properties.

2.1.1 Related Work

In [34], [9], [11], [25], [30], [52], [53], [2] networks for different applications are designed
based on demand and supply. For instance in [34], properties of logistics networks struc-
tures are determined by optimizing the structural design on a strategic level: given the
location of a plant producing one type of product and given the number and location
of the consumers, who have to be satisfied, the number and locations of warehouses is
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 22

determined. In [34] robustness is the extent to which the system is able to carry out its
functions despite some damage done to it, such as the removal of some of the nodes and/or
edges in the network. Results show that if the maximal robustness level is increased, more
warehouses are present in the optimal network, decreasing the network efficiency and in-
creasing its complexity. These studies all assume the deterministic problems (i.e. supply
and demand are constant). In this chapter, however, we consider the distribution of mul-
tiple products and take stochastic supply and demand into consideration while designing
a network for an LSP. Our definition of robustness is then to which extent the network
topology is able to minimize operational costs despite some changes in supply and demand
distributions.
A common setting for stochastic supply chain problems can be found in [48], [28], [6]
[18] and [44]. These studies investigate different policies for supply chain management:
orders are placed to manufacturing facilities to keep inventory positions in warehouses
at a desired level and satisfy consumer demand. Only a restricted number ([18], [44])
mimic transportation costs by a stepwise cost-function dependent on the capacity of
a transportation device (economy of scales). Such an approach reflects practice more
accurately than a linear model at the expense of higher computational time. In this
chapter, a heuristics is proposed that generates approximately similar close to optimal
network topologies for both the stepwise and the linear model for medium size problems.
For large network sizes, which become intractable with the stepwise model, we have to
rely on the approach with linear transportation costs. Again, we have to stress that the
decision space in a supply chain is crucially different from that of an LSP. Namely, in a
supply chain, each member can place orders to the one upstream. For an LSP however,
both supply and demand are given and cannot be influenced. Despite this difference,
some of the supply chain studies mentioned above are related to ours as they use similar
approaches on a tactical and/or strategic level.
In [48] for instance, the strategic design of a multi-echelon, multi-product supply chain
network under demand uncertainty is studied. Given demand, decisions have to be made
on the production amount at each supply facility. The objective is to minimize total costs
taking infrastructure as well as operational costs into consideration. The authors consider
a steady-state form of this problem and thus all flows between nodes are considered to
be time-averaged quantities. A set of (only) three demand scenarios is generated and the
objective function is expanded by adding weighted costs for meeting demand in each of
these three scenarios. In this chapter, however, the amount of supply at the production
facilities is given as well as the consumers’ demands and cannot be controlled. Moreover,
we do consider time-variant supply and demand of multiple products and use a rolling
horizon approach to decide on the product flows on an operational level. Dependent on
the link usage during a large number of time steps in this rolling horizon approach, a
close to minimal number of links is established to construct a close to optimal topology.
This topology is robust to changes in second and higher order moments of the supply and
demand distributions.
In [28] the number and locations of transshipment hubs (strategic level) in a supply
network is determined dependent on the product flows. The authors start from a network
in which all potential hubs are present and minimize the costs for the operational activities
in the network. Then the decrease in costs is evaluated for each of the cases where one
hub is deleted from the network (establishing a hub introduces fixed costs). Next, the
hub that causes the largest decrease is deleted from the network. This process is repeated
and terminates whenever deleting a hub does not significantly affect the costs. In this
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 23

chapter, a similar approach on a tactical level is applied: we start from a fully connected
network, run the operational decision making policy for a certain time period (100 time
steps) and in the end delete a least used link over this period. This procedure is repeated
until the network is minimally connected. Additionally, experiments suggest that this
heuristic generates a topology with a small number of links that is robust to changes in
supply and demand distributions.
In [6], models and algorithms for a one product, multi-stage stochastic distribution
problem with recourse are developed. In the first stage, the amount of products to be
supplied by the plants to warehouses has to be decided on, without knowing the demand.
Then in the second stage, the demand becomes available and products have to be shipped
from the warehouses to the consumers. In this chapter, we consider the distribution
of multiple products by an LSP and investigate the robustness of generated network
topologies. Since supply and demand over a restricted time horizon in the future are
known, we can apply a rolling horizon approach rather than a multi-stage approach. As
we noticed from practice, such a rolling horizon approach is often used by an LSP.
To our knowledge, only a restricted number of studies consider the viewpoint of an LSP
where supply and demand are both stochastic and uncontrollable. Of particular interest
is the study in [47]: A company owns several production plants and has to distribute
only one type of product to different regional markets. In each time period, a random
(uncontrollable) amount of products becomes available at each of these plants. Before
the random demand becomes available, the company has to decide which proportion of
the products should be shipped directly and which proportion should be held at the
production plants. Linear costs are assigned to transportation, holding and backlog. A
look ahead mechanism is introduced by using approximations of the value function and
improving these approximations using samples of the random quantities. It is numerically
shown that this dynamic programming method yields high quality solutions. One of the
results of [47] is that most improvements on the operational costs are made when a part
of the consumers is served by two plants, rather than one. The study in [47] investigates
the distribution of only one type of product for only one network configuration with linear
transportation costs. In this chapter, however, we consider the distribution of multiple
products and start from considering a stepwise cost function to represent the economy
of scales principle. Hence, our methodology takes the possible consolidation of high and
low value products to ship full trucks into consideration. Another difference is that in the
setting we consider, the exact supplies and demands over a restricted future horizon are
known and we thus can apply a rolling horizon approach. This is another major difference
with the study in [47], where decisions on how much to send or store have to be made
before the actual demand becomes available (recourse model). Furthermore, we expand
our study by investigating the robustness of the constructed network topologies, i.e., the
dependency of the operational performance of the constructed topologies on supply and
demand distributions.
We apply the proposed bi-level optimization method to several real-life networks. The
results suggest that in each considered case a distribution network with only a few links
provides a large portion of the operational efficiency of the fully connected network. These
results confirm the findings of other studies on different kinds of two-echelon networks.
Results in [26] for instance suggest that, for a small theoretical problem, limited flexibility
in manufacturing processes (i.e., each plant builds only a few products) yields most of
the benefits of total flexibility (i.e., each plant builds all products). However, the authors
mention that for more realistic cases they have no guidelines or general approach to add
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 24

flexibility in an arbitrary network. In this chapter, this issue is addressed by developing

a heuristic that for more realistic cases and for more than a two echelon system generates
a network with a small number of links that still performs well. Moreover, the study
in [26], does not investigate the robustness of the constructed networks to the system’s
parameters. In this chapter, besides proposing a generic design methodology, we also
investigate robustness properties of the designed topologies.

2.1.2 Contributions
We proceed in Section 2.2 by first addressing the lower-level operational problem (Sec-
tion 2.2.1) of controlling the material flow for a given network topology. A stepwise cost
function is introduced in the flow model to represent the costs for transportation by trucks
(economies of scale). The supplies and demands of each product type are scaled to the
trucks’ capacity. Storage costs per unit per day are (therefore) the same for each product
type. Costs for early and late delivery do depend on the product type. A model pre-
dictive control with rolling horizon (MPC) [13] is used, which determines a sub-optimal
operational routing schedule for a particular topology and particular supply and demand
realizations over a certain time period (in this case 100 time steps). This amounts to
solving a sequence of dependent non-linear minimum-cost network flow problems.
In Section 2.2.2, the upper level is addressed, which in turn refers to the tactical
decision making. We present a branch-and-bound method that computes the exact trade-
off front of the bi-objective problem for given multiple-product supply and demand time
series and the given control policy. As the method is impractical for larger instances, we
propose a heuristic that works by iteratively dropping least used links (after determining
a routing schedule for a certain time period) to generate an approximation of the trade-off
front. The heuristic is accurate and much faster than the branch and bound, however still
large real-life networks instances are intractable due to the model’s non-linearity. Hence,
we replace the stepwise cost model by a linear cost model. Since in the heuristic the link
usage depends on the product type(s) through it (the higher the demand for just-in-time
delivery, the higher the weight factor), links with products that do not require just-in-
time delivery will be dropped first. In the end, these products are then shipped through
another link together with products that do require just in time delivery. In this way,
different product types are consolidated even in the linear approach. We compare the
results to those obtained with the stepwise cost model and find similar results in both
the costs curve and the topology structure. For the heuristic approximation, which is
very efficient, we can usually provide a network with a small number of links and close to
minimal operational costs. From that we can conclude that the topology is also close to
optimal.
Results of the branch and bound method and the heuristics approach for small prob-
lems are compared in Section 2.3. Next, the results for the heuristics applied to the
stepwise and the linear model for medium size problems are compared. Moreover, results
of the heuristics approach applied to the linear model for several large network config-
urations are shown in Section 2.3. These results suggest the validity of the proposed
approximative bi-level optimization. Finally, we perform a large amount of experiments
to formulate generic insights in the robustness of the heuristic network topology. Since this
heuristic network topology results from a particular supply and demand time series, we
are interested in the performance of this topology for different time series. Interestingly,
the experimental results suggest that the heuristic network topology is only sensitive to
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 25

the first moment of supply and demand distributions. Namely, as long as the heuristic
network topology is run operationally for a different supply and demand time series with
the same means, the operational costs are still close to minimal (relative to the operational
costs in the fully connected network). However, if the means are changed, the operational
costs grow large (relative to the operational costs in the fully connected network). Having
reasonable forecasts about the individual means of supply and demand we would a priori
i i

be able to construct a cost-effective network topology, robust to any time series with these
i
“calKfin˙temp” — 2008/6/11 — 9:38 — page 1 — #1
i

means. We end with conclusions and recommendations for future work in Section 2.4.
Suppliers Warehouses Consumers
1

...
1 K 1 1

. . .

. . .

. . .

S W D

Figure 2.1: Illustration of the considered logistics networks.

i i

i i

2.2 Bi-level Network Design Problem

We consider the distribution of K types of products by trucks through a network with S
production facilities, W warehouses, and D consumers. Unless stated differently, we use
the indices s ∈ {1, ..., S} for the production facilities, w ∈ {1, ..., W } for the warehouses,
d ∈ {1, ..., D} for the consumers, and k ∈ {1, ..., K}, the product types. Products can
be sent either directly from supplier(s) to consumer(s) or indirectly via a warehouse,
assuming that each consumer can receive its products from each (and more than one) of
the production facilities. Shipments from one warehouse to another are not taken into
account, but could easily be incorporated into this framework as well. An example network
for this type of three-echelon multi-item distribution system is depicted in Figure 2.1.
Even though each link in the graph has a certain length representing the geographical
distance between the corresponding nodes, Figure 2.1 does not show the spatial positioning
of the nodes for ease of presentation.
Each production facility supplies one or multiple product types, and each consumer
demands one or multiple product types, at each time step (day) t ∈ {1, ..., T }. The supply
and demand time series Sik (t) (supply of product type k by supplier i at day t) and Djk (t)
(demand of product type k by consumer j at day t) are randomly distributed according
to a joint distribution that is defined by the following sampling procedure: for each
combination of product type, supplier and consumer, we assume a uniform distribution
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 26

on an interval [σ̄ik (t) − σ̂ik (t), σ̄ik (t) + σ̂ik (t)] with given mean σ̄ik (t) and half-length σ̂ik (t).
Next, T preliminary samples out of each of the K · (S + D) distributions are taken. The
final samples are then determined by adding a constant either to each supply or each
demand sample for each product type to enforce that for each product type, the sum of
T supplies equals the sum of T demands over the horizon.
On a given network, the operational task of the LSP is to decide the amount of products
to be transported over each link, which incurs transportation cost on the links, storage
cost in the warehouses, and penalty costs for early or late delivery at the consumers. On
the tactical level, the task of the LSP is to choose a subset of links such that it can carry
out its operational task well, e.g., with minimal expected costs. Of course, there is an
immediate trade-off between the number of links and the operational costs, as additional
links can only improve the operational costs due to the added flexibility. In the following,
we present an approach to approximate the trade-off front of this bi-objective bi-level
problem numerically, and in particular to determine a network structure with a close to
minimal number of links and close to minimal operational costs.

2.2.1 Operational Decision Making

To formalize the operational task of the LSP, we define the operational decision variables
xkij (t) ≥ 0 as the amount of products of type k transported from node i to node j on time
step t, where (i, j) ∈ U, the set links in the given network. On the suppliers side, we
require that all available products Sik (t) have to picked up in the same period such that
X
Sik (t) = xkij (t) ∀k, i, t (2.1)
j∈ Pi

where the set Pi contains the indices of all the nodes to which production facility i is
connected. On the consumers side, both early and late deliveries might occur and have
to be accounted for. For this we introduce state variables bkj (t) as the backlog of product
type k for consumer j on day t, whose dynamics is
X
bkj (t) = bkj (t − 1) + Djk (t) − xkij (t) ∀k, j (2.2)
i∈ Qj

where the set Qj contains the indices of all the nodes connected to consumer j. In
the warehouses, several processes (e.g. inbound, consolidation) between arriving at and
storing in a warehouse cause delay. The total delay is captured in a constant τw ∈ N so
that products arriving in warehouse w at time t can be shipped out τw time steps later
at the earliest. To model this time delay, we introduce additional state variables ywk (t) as
the inventory of product k in warehouse w on day t, whose dynamics is
X X
ywk (t) = ywk (t − 1) + xkiw (t − τw ) − xkwj (t) ∀k, w, t (2.3)
i∈ Iw j∈ Ow

where the set Iw contains the indices of all nodes that are sources of warehouse w and the
set Ow contains the indices of all nodes that are destinations of warehouse w.
The operational cost for time step t is the sum of transportation cost, storage cost,
and backlog cost. It can be expressed as a function of the decision variables xkij (t) and
state variables bkj (t) and ywk (t) at time t as
K K X
D K
k k2
X X X X X
α(xkij (t), bkj (t), ywk (t)) ≡ cij φ( k
xij (t)) + β bj (t) + hw ywk (t) (2.4)
(i,j)∈ U k=1 k=1 j=1 w∈ W k=1
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 27

where cij represents the geographical distance between nodes i and j, hw represents the
inventory holding costs in warehouse w, and β k represents the “value” of product k for
consumer j. The reason for not using the absolute value |bkj (t)| as a penalty for early or
late delivery is that in case of shortages, the quadratic function favors equal distribution
of shipments among consumers rather than shipping everything to one consumer. The
function φ(·) expresses quantity-dependent transportation costs. We consider two cases,
(i) φ(x) ≡ d Vx e to express stepwise-constant transportation costs due to the use of trucks
with unit capacity V , and (ii) the identity φ(x) ≡ x to express the simplest case of linear
transportation costs.
We further assume that suppliers and consumers commit to their real supplies and
demands for a constant number of time steps Ω in advance. Thus, the exact supplies
and demands for the current and the next Ω − 1 time steps are known and can be taken
into account when deciding the transportation quantities. By using state augmentation
we can include this information in the current state of the distribution system at time t.
Consequently, the operational problem can be posed as a discrete-time stochastic optimal
control problem, with the objective to minimize the expected operational cost over the
considered fixed time horizon T . Under the assumptions given above it is reasonable to
assume that there exists an optimal state-feedback control policy {µ∗ (t)}Tt=1 that maps
the augmented state to its corresponding optimal transportation quantities at each time
step t.
The standard method to solve this type of stochastic optimal control problem is
stochastic dynamic programming. Unfortunately, the state space dimension in our case
is very large, so that a numerical solution, even with a very coarse state discretization,
seems intractable. We therefore have to resort to solving the operational control problem
only approximately.
One principle method for approximate dynamic programming is to use a limited look-
ahead scheme in a rolling horizon fashion, also called model predictive control (MPC).
The basic idea is that optimization is only performed over a limited horizon where more
information about uncertain data is available, or even well representable by deterministic
or nominal values, and to choose some reasonable approximation of the value function for
the time steps outside the horizon. Here, we use a horizon of Ω steps as the exact supply
and demand data is known over this time frame, and approximate the value function of
the state at the horizon with the constant zero. Thus, we obtain an approximate control
policy µ̂ by solving the following time-expanded network flow problem:
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 28

 
t+Ω−1
X X K X
X D X K
X
minimize  cij uij (q)) + β k [bkj (q)]2 + hw ywk (q) (2.5)
q=t (i,j)∈ U k=1 j=1 w∈ W k=1

subject to
Sik (q) xkij (q)
P
= ∀k, i, q
j∈ Pi
ywk (q) ywk (q xkiw (q − τw ) − xkwj (q) ∀k, w, q
P P
= − 1) +
i∈ Iw j∈ Ow

bkj (q) = bkj (q − 1) + Djk (q) − xkij (q)

P
i∈ Qj ∀k, i, j, q
K
xkij (q)
P
V · uij (q) ≥ ∀k, i, j, q
k=1

xkij (q) ≥ 0 ∀k, i, j, q

ywk (q) ≥ 0 ∀k, w, q

Minimization is performed over the variables xkij (q), uij (q), bkj (q), and ywk (q), where
q ∈ {t, ..., t + Ω − 1}. The additional variables uij (q) are introduced to linearize the trans-
portation cost function φ(·) and represent the number of trucks of capacity V necessary to
transport the total amount of products over link (i, j) at time step q. To model the step-
wise constant transportation cost, these variables have to be restricted to integer values;
otherwise a continuous relaxation can be used to model linear transportation costs.
For all time steps prior to t, the values of the state variables and decision variables are
known so that they serve as deterministic data in the above limited look-ahead problem.
The same holds for the values of supplies Sik (q) and demands Djk (q) within the considered
horizon {t, ..., t + Ω − 1}. Consequently, the problem is a mixed-integer quadratic program
(in case of stepwise-constant transportation costs) or a quadratic program (in case of
linear transportation costs). The approximate policy µ̂ is now given by the mapping of
the current (augmented) state at time t to the values xkij (t) in an optimal solution of the
above optimization problem.

2.2.2 Tactical Decision Making

In a given collection of suppliers, warehouses, and consumers, let L be the union of all
potential links (directed edges) from suppliers to warehouses, warehouses to consumers,
and suppliers to consumers. The tactical problem can then be defined as choosing a subset
U ⊆ L of a given cardinality L with smallest expected operational cost. In doing so, we
tacitly assume that there exists some mapping γµ that assigns the expected operational
costs γµ (U) to the network of chosen links U when using a particular control policy µ.
As it is unclear how this expectation can be computed exactly in the present set-up, we
resort to a variant of sample average approximation by Monte Carlo simulation, where
we simulate the process for a fixed number of time steps while applying the given policy
µ. Here, we use the the approximate limited look-ahead policy µ̂ defined above, and refer
to the resulting estimator of the expected operational cost as γ̂µ̂ (U).
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 29

In a bi-objective formulation, this tactical problem is to be solved for any value of

L ∈ {Lmin , . . . , |L|}, where Lmin is the smallest number of links such that no node is
isolated and each warehouse is connected to at least one supplier and one consumer. Our
conjecture about this trade-off between the number of chosen links and the operational
costs is that the operational costs are only marginally sensitive to a reduction of the
number of links from the fully connected network until a critical value, after which the
costs increase sharply. In this case, the trade-off frontier would form a pronounced knee.
Hence, our particular goal is to determine a representative network from this region, as this
would constitute in a sense a best-possible bi-objective approximation. Such a network
could be considered “cost-effective” as it would yield close-to-minimal operational costs
with a close-to-minimal number of links.
To determine the trade-off frontier, we first develop a bi-objective branch and bound
method that computes the minimal cost topology for any given number of links simulta-
neously. As this method still searches a large part of the solution tree, we additionally
propose a heuristic, which is able to generate a satisfactory solution for larger instances.

Bi-objective Branch and Bound Method

We develop a branch and bound algorithm that defines a tree search over a set of binary
decision variables zij associated with each link (i, j) ∈ L, where

1 if the link from facility i to facility j is present,
zij = (2.6)
0 otherwise.

A pseudo-code description of the algorithm is given below (Algorithm 1). The al-
gorithm maintains and updates a current approximation to the trade-off frontier whose
initial values are given in Table 2.1.

L Lmin Lmin + 1 . . . . L
opt(L) ∞ ∞ ∞ ∞ ∞ ∞ ∞
Zopt (L) - - - - - - (1,1,1,...,1)

Table 2.1: Initial state of the table for the minimal costs opt(L) and the corresponding
solution (network) Zopt (L) .

During the search, the L-entry is updated whenever a network z with L links is
identified to have lower operational cost than the current best network with L links (lines
14-16). At each branching operation (lines 19-27), a natural lower bound is given by
setting all undecided variables to one as all solutions in this subtree can only be composed
of a subset of those links. The subtree can be pruned if this lower bound is not lower than
the current best value for any number of links L for which there are potential solutions
in the subtree; otherwise, two children are created by setting the first undecided variable
to zero and one, respectively, and added to the list on unexplored nodes A.

Heuristic
Since the branch and bound method evaluates a large number of possible topologies, only
small problems can be solved. Hence, we propose a heuristic that evaluates less than
|L| topologies and still yields a very good solution. The heuristic starts from the fully
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 30

connected network and approximates the operational costs γ̂µ̂ (L) when using the control
policy µ̂ over the fixed number of time steps. Then the link(s) that are least used over
this time period are deleted. In determining the usage of a link, the value of the product
type(s) through this link is taken into account by a weight factor. We define the weighted
average daily usage Pij of the link between facilities i and j as
K T
1 XX bk
Pij = · xk (t). (2.7)
T k=1 t=1 max{bk } ij
bk

In this way, links that provide just-in-time delivery of high value products are not (easily)
deleted. This procedure of deleting links is repeated until a minimally connected network
remains. A pseudo-code description of the algorithm is given below (Algorithm 2).

Algorithm 1 Branch and Bound

1: let z := (?, ?, ..., ?)
2: let LB(z) := −1
3: initialize the list of unexplored nodes A with z
4: while A is not empty do
5: pick z as the last element from A
6: delete z from A
7: let p be the number of ones in z and q be the number of zeros in z
8: if p + q = |L| then
9: if LB(z) = −1 then
10: let c := γ̂µ̂ (z)
11: else
12: let c := LB(z)
13: end if
14: if c < opt(p) then
15: let opt(p) := c
16: let Zopt (p) := z
17: end if
18: else
19: if LB(z) = −1 then
20: let z 0 be a copy of z where all ? are replaced by 1
21: let LB(z) := γ̂µ̂ (z 0 )
22: end if
23: if LB(z) < Zopt (L) for some L with p ≤ L ≤ |L| then
24: let z 00 be a copy of z where the first ? is replaced by 0 and append z 00 to A
25: let z 000 be a copy of z where the first ? is replaced by 1 and append z 000 to A
26: let LB(z 00 ) := −1 and LB(z 000 ) := LB(z)
27: end if
28: end if
29: end while
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 31

Algorithm 2 Heuristic
1: let U be the fully connected network L
2: while the network is not minimally connected do
3: apply the policy µ̂ during T time steps and determine γ̂µ̂ (U)
4: determine usage Pij for each link in U
5: delete the link(s) from U with minimal usage
6: end while

2.3 Computational Study

In the previous section, we proposed two algorithms to find a network topology with
a close to minimal number of links that is still cost-effective on the operational level.
In this section, the performance of both these algorithms is evaluated and compared.
Experiments suggest that the heuristic is very effective in constructing a network topology
with a small number of links that yields close to minimal operational costs. However, large
real-life problems are still intractable. Hence, we replace the stepwise cost function by a
linear cost function and find similar results much faster. Some large real-life instances are
solved for the linear cost structure and results are presented.
To evaluate the performance of the heuristic, a bi-objective relative approximation
factor (c , L ) is used as a performance measure. The factor denotes the relative deviation
from the ideal point, the point composed of the single-objective optima. In this case,
those values are known, so we can define the two components as

γ̂µ̂ (U) |U|

c (U) := and L (U) := , (2.8)
γ̂µ̂ (L) Lmin

denoting the operational costs of the considered topology U relative to the operational
costs in the fully connected network, and the number of links |U| relative to the number
of links in a minimally connected network. Even though we are not able to give some
a priori performance guarantee of the heuristic, this factor can be used to bound the
approximation quality a posteriori.
In addition to evaluating the performance of the heuristic approach, a computational
study is performed to investigate the robustness of the so found network topology with re-
spect to changes in the supply and demand distributions. Results suggest that a “heuristic
network“ with a satisfactory (c (U), L (U)) value is cost-effective as long as the means of
supply and demand distributions remain the same. Finally, we formulate the hypothesis
that increasing the initial inventory enables to compensate for the backlog costs em-
bedded in a particular supply and demand time series. The hypothesis is confirmed by
experimental results.
For the logistics networks in this chapter, we consider M suppliers, W warehouses, D
consumers and the distribution of K product types. The networks considered typically
consist of few suppliers and warehouses and a lot of consumers, such that W < M  D.
Furthermore, we generate stochastic time series of K · S supplies and K · D demands from
uniform distributions with random means and variances for 100 time steps. We assume
that supplies and demands are only known 2 days in advance, so the window size Ω is equal
to 3. Moreover, we focus on networks of relatively small geographical extent. Therefore,
it is assumed that products can be transported from any arbitrary node to another within
one day. In addition, we assume that the delay for going through warehouse n is equal
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 32

to one day, which implies that τw = 1. The costs are chosen such that the proportion of
transportation and storage costs ci,j : hn =1 : 0.6 on average and the early/late delivery
costs bk are randomly chosen between the bounds of 0.001 and 100.

2.3.1 Comparison Branch and Bound versus Heuristic

First, we compare the results of the branch and bound method and the results of the
heuristic applied to the stepwise cost model. Hence, we generate different problem in-
stances (supply and demand distributions as well as spatial location of network nodes) for
many small network configurations and run the branch and bound method as well as the
heuristic. Comparison of the resulting trade-off fronts of operational cost versus number
of links show that the heuristic generates a topology that is close to optimal.
As an illustration (for a network with S = 2, W = 1, D = 3 and K = 1), we show
results for ten different parameter sets being 10 supply and demand time series, each
accompanied by a specific spatial location of nodes. Results for these sets for both the
branch and bound method and the heuristic are shown in Figure 2.2. As expected, both
methods generate equal operational costs for the fully connected network (number of
deleted links equals zero). From Figure 2.2, it also appears that the critical value in the
efficient frontier found by the heuristic occurs at 5 or 6 deleted links, whereas it always
occurs at 6 deleted links for the networks generated by the branch and bound approach.
i i
The difference in the number of links in the resulting network for these instances is thus
“BBH˙temp” — 2008/12/1 — 11:20 — page 1 — #1
at most i1 link. This suggests that the heuristic finds a satisfactory solution for small i

network instances. Since the CPU time of the branch and bound grows extremely fast
with increasing network size, we have to rely on the heuristic for larger instances.
7
10
Heuristic
6
B&B
10

5
10
Costs

4
10

3
10

2
10
0 1 2 3 4 5 6 7
Number of deleted links

Figure 2.2: Comparison of branch and bound with heuristic for 10 supply and demand
instances, S = 2, W = 1, D = 3, K = 1, Branch and bound: c ≤ 1.01, L = 1.5,
Heuristics: c ≤ 1.01, L = 1.63.

The heuristic, applied to the stepwise cost model, enables to find a close to optimal
topology for a network with up to 4 suppliers, 3 warehouses, 25 consumers and 6 product
types (M = 4, W = 3, C = 25 and K = 6, respectively). Due to the non-linearity,
larger real-life problems are still intractable. We therefore remove the integer variables
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 33

uij from the model and introduce linear transportation costs (φ(x) ≡ x). We apply the
heuristic to this linear model, which is now able to construct cost-effective topologies for
large real-life network sizes. In the next section, results of the heuristic for the stepwise
and linear model are compared. Additionally, we present results of the heuristic for the
linear model for large real-life network sizes.

2.3.2 Performance of the Heuristic for Large Networks

Results for both the linear and the stepwise model are presented in Figure 2.3 for two
different network configurations, respectively. For larger network sizes, we have to rely
on the heuristic model with linear costs. Two cost curves for large real-life networks are
presented in Figure 2.4. Costs are plotted on a logarithmic scale versus the number of
deleted links. In these plots, zero deleted links corresponds to the fully connected net-
work, whereas the maximal number of deleted links corresponds to a minimally connected
network.
All these curves (in Figures 2.3 and 2.4) exhibit the same typical characteristics from
the left to the right: at first, a large number of links can be deleted without affecting
the costs significantly: Up to 70% of the links is not used at all in the first heuristic
iteration and is thus deleted at once. Still, about 15% of all the links can subsequently be
deleted without substantially affecting costs. Then, after about 85% of the total number
of links that can be deleted, costs start increasing slightly with the number of deleted
links. Finally, when about 90-95% of the total number of links that can be deleted, costs
start increasing dramatically. Below the figures the L -value are given for an upper bound
on the c -value of 1.01.
The comparison between the results of the linear and stepwise model show that in
the fully connected network, as expected, the costs for the stepwise model are slightly
higher than the costs for the linear model (due to the logarithmic scale this can hardly be
noticed in Figure 2.3). Furthermore, Figure 2.3 shows that for both models the dramatic
cost increase starts at approximately the same number of links. Further analysis (not
depicted in the figures here) shows that for the performed experiments, the set of deleted
links at the critical value is approximately the same for both models. A large number of
experiments, not presented in this chapter, exhibit the same characteristics, suggesting
that the heuristic for the linear model is a quite good approximation of the heuristic with
the stepwise model.
Apparently, a topology with only a small number of links suffices to gain close to
minimal operational costs. Since only a small number of links remain to be deleted we
can also conclude that the resulting topology is also close to optimal. In [26], similar
characteristics are presented for a manufacturing system with deterministic supplies and
demands. The results in [26] suggest that a small flexibility (each plant produces only a
couple of product types) can almost achieve the benefits of total flexibility (each plant
produces all product types). Furthermore, it should be noticed that the curves in Fig-
ures 2.3 and 2.4 exhibit approximately horizontal plateaus. The reason for this is that
not one, but a group of links has to be deleted to significantly affect the costs.

2.3.3 Statistical Robustness Analysis

We now formulate the hypothesis that networks generated for a particular supply and
demand time series, are insensitive to changes of second or higher order moments of
 i i i i
“MP1˙temp” — 2008/12/1 — 16:25 — page 1 — #1 “MP2˙temp” —
i i i 2008/12/1 — 11:48 — page 1 — #1

CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 34

8 7
10 10
linear linear
7
stepwise stepwise
10

6
6
10
10
Costs

Costs
5
10
5
10
4
10

3 4
10 10
0 10 20 30 40 50 60 70 0 20 40 60 80 100 120 140
Number of deleted links Number of deleted links
(a) S = 3, W = 2, D = 15, K = 5, c ≤ 1.01, (b) S = 4, W = 3, D = 20, K = 5, c ≤ 1.01,
L = 1.65. L = 1.57.
i i i i

i Figure 2.3: Comparison heuristics results for the linear and stepwise
“MP3˙temp” — 2008/12/1 — 11:44 — page 1 — #1
i
“MP4˙temp” model,
— 2008/12/1
i — 11:44 —T =100,
page 1 — #1 τw =1,

Ω=3.

9 9
10 10

8
8 10
10

7
10
Costs

Costs

7
10
6
10

6
10 5
i i i
10
i i i i

5 4
10 10
0 50 100 150 200 250 300 0 200 400 600 800 1000
Number of deleted links Number of deleted links
(a) S = 5, W = 3, D = 40, K = 6, c ≤ 1.01, (b) S = 8, W = 4, D = 75, K = 10, c ≤ 1.01,
L = 2.35. L = 1.97.

Figure 2.4: Heuristics results for the linear model, T =100, τw =1, Ω=3.

supply and demand distributions. Thus, if information about the individual means of
supplies and demands is available a priori, we only need to generate a stochastic time
series with these means to find a cost-effective topology for other scenarios with these
means. Next, as we optimize over a finite period of time T , we have to address the
influence of the initial inventory. We demonstrate that the steady state value of the costs
converges to the minimal transportation plus storage costs, as initial inventory increases.

Sensitivity
i i i

i
To allow for an automatic statistical sensitivity
i analysis we define
i
a representative “heuris- i

tic network“ to be the smallest network U discovered by the heuristic, whose operational
costs are within a factor F of the operational costs of the fully connected network, i.e.,
 CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 35

where c (U) ≤ F . Deleting one more link will increase the costs above that level. We then
perform the following statistical analysis: We choose fixed mean supplies and demands
for a network with S = 4, W = 3, D = 20 and K = 5 and generate stochastic supply
and demand samples (for 100 time steps). Based on this, we determine the “heuristic
network“ and fix its topology, i.e., its link structure U. For this network U and the same
product mix, we than compute the operational costs for supply and demand time series
i from 20 different distributions (with different i second and higher i order moments but same i

i means) and record the resulting approximation

“scatter1˙temp” — 2008/12/1 — 11:19 — page 1 — #1
i factor c (U) in each case. This experi-
“scatter2˙temp” — 2008/12/1 — 11:19 — page 1 — #1
i

ment is repeated for 99 other heuristic networks, each generated from particular supply
and demand time series.
3 3

2.5 2.5

2 2
εc

εc
1.5 1.5

i i i i
1 “scatter3˙temp” — 2008/12/1 — 11:20 — page 1 — #1
1 “scatter4˙temp” — 2008/12/1 — 11:20 — page 1 — #1
i i i
1 1.5 2 2.5 3 3.5 4 1 1.5 2 2.5 3 3.5 4
εL εL

(a) F = 1.001. (b) F = 1.01.

3 3

2.5 2.5

2 2
εc

εc

1.5 1.5

1 1
1 1.5 2 2.5 3 3.5 4 1 1.5 2 2.5 3 3.5 4
εL εL

(c) F = 1.05. (d) F = 1.10.

i i i

i Figure 2.5: Sensitivity ”heuristic network”

i to 2nd and higher
i order moments. ’o’ product i

mix with 1 high value and 4 low value products, ’+’ product mix with more than 1 high
value product.

Figure 2.5 shows four scatter plots (each for a particular value of F ) of the 100 mean
values over the 20 c samples versus L . As expected, the robustness decreases as F
increases (since then the number of links in the heuristic network decreases which makes
the network more sensitive). We define the network to be robust if the mean value
of c (over 20 experiments) is at most 1.2 with a frequency of 95%. The scatter plot
in Figure 2.5a shows that c is strongly clustered near one indicating that, with high

i i i

i i i i
 i i

“ECDF˙temp” — 2008/12/1 — 11:38 — page 1 — #1

i i

CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 36

0.8

0.6

F(εc)
0.4

0.2
same means
different means
0
0 2 4 6 8 10
εc

Figure 2.6: Empirical cumulative distributions showing the sensitivity of the heuristic
networks to changes in i) variances and ii) means of supplies and demands.

likelihood, our process of choosing a heuristic network (with F = 1.001) will find a
network that is a ”very good” network even when the supply and demand time series
vary randomly with respect to higher order moments. The histogram for the case where
F = 1.01 (Figure 2.5b) is still satisfying since more than 95% of the considered cases
results in an c value equal to or smaller than 1.2. The results for F = 1.05 (Figure 2.5c)
and F = 1.1 (Figure 2.5c) do not satisfy our definition of a robust network anymore. For
these values of F , several means and standard deviations of c are even outside the figure
range (larger than 3). We choose as “heuristic network“ the one with F = 1.01, since (i)
the accompanying results fulfill our definition of a robust network, and (ii) the generated
heuristic networks
i have less links than the case where F = 1.001. These results suggest
i

that the “heuristic

i network“ (F = 1.01) indeed is robust to changes in second andi higher
order moments of supply and demand distributions.
The corresponding values of L suggest that the generated heuristic networks on av-
erage consist of 2.5 times the number of links in a minimally connected network. This
means that each network facility is on average connected to two or three other network
facilities. Apparently, this restricted amount of connectivity suffices to provide a large
portion of the operational efficiency and robustness of the fully connected network. An-
other interesting result is that the number of links in the “heuristic network“ depends
on the product mix. A circle in Figure 2.5 represents an instance with a product mix
with only one high value product and in this case 4 low value products, whereas a plus
represents an instance with a product mix with at least two high value products. The
results suggest that more links are required when more high value products are present.
We give the following explanation for this phenomenon: the structure of the “heuristic
network“ is mostly determined by the number and locations of suppliers and consumers of
high value products. When more high value products are present, in general the number
of suppliers and consumers of high value products increases, and thus more links have to
be established to provide just-in-time delivery.
To check the sensitivity of the heuristic network to the first moments of supply and
demand distributions, we perform the following experiment: we take the 100 heuristic
networks, where F =1.01, and 100 accompanying results of c from the previous experi-
ments where the means of supplies and demands stay the same, but the variances change
 CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 37

(Figure 2.5b). In addition, for each of the 100 heuristics networks we generate supply
and demand time series from completely different distributions and determine the cor-
responding c . Figure 2.6 depicts two empirical cumulative distribution functions of 100
c values generated from i) time series with different variances, but the same means as
the heuristic network originally was generated from (solid line), and ii) time series with
different means (dotted line). The solid line again confirms that the heuristic network is
robust to the variances since the cumulative distribution grows to 100% very fast starting
at c = 1. The cumulative distribution function for the instances with different means
grows much slower. Furthermore, only 80% of the c values are smaller than or equal to
10. These results suggest that indeed the heuristic network (F = 1.01) is sensitive to
changes in the means of supply and demand distributions.

Inventory Flexibility
In all previously performed experiments the parameter of initial inventory ywk (0) has been
set to zero, which is typical in practice for an LSP. It turns out that this initial inven-
tory position is the only controllable parameter that can influence the operational costs
for running the fully connected network. Consider the black dotted lines in Figure 2.7a
and 2.7b, which show results of the heuristic for S = 4, W = 3, D = 20 and K = 5,
for respectively two different supply and demand time series with equal first moments
and ywk (0) = 0. The fact that the means of supply and demand are equal implies that
approximately the same amount of products is shipped along time period T , and thus ap-
proximately the same transportation costs are present for both these cases. Nevertheless,
a large difference in costs for the fully connected network can be noticed. This difference
can therefore only be caused by backlog costs, which are induced by the specific time
series. For instance, a time series, which exhibits a peak in demand at the beginning of
period T , always leads to backlog costs, even in a fully connected network. Therefore,
i we pose the hypothesis that the only icontrollable parameter i that influences the costs for i

i
the fully connected
“IF1˙temp” —network is —the
2008/12/1 — 14:19 page initial
i
1 — #1 inventory. Increasing the initial
“IF2˙temp” — 2008/12/1
i
— 14:18 — inventory
page 1 — #1 will
compensate for early backlog, which eventually will lead to convergence to the inevitable
transportation costs.

zero initial inventory zero initial inventory

8 8
10 1.0% initial inventory 10 1.0% initial inventory
2.5% initial inventory 2.5% initial inventory
Transportation+storage costs Transportation+storage costs

6 6
10 10
Costs

Costs

4 4
10 10

2 2
10 10
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Number of deleted links Number of deleted links
(a) Results time series 1. (b) Results time series 2.

Figure 2.7: Initial inventory flexibility.

To confirm this hypothesis, we perform the following additional experiments for the
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 38

two supply and demand time series as mentioned above: we generate a supply and demand
time series, set the inventory position in each of the warehouses to zero and determine
the operational costs dependent on the number of deleted links (black dotted lines in
Figure 2.7a and 2.7b). Next, we perform two additional optimizations for the same
supply and demand time series where 1.0% and 2.5% of the total shipped amount of
products is present in the warehouses as an initial condition, respectively. The results of
these experiments are depicted in Figure 2.7a and 2.7b. As one can see, for both cases,
the total costs of the fully connected network indeed converge to the transportation and
storage costs as initial inventory increases.

2.4 Conclusions and Recommendations

In this chapter we considered transportation networks from the point of view of a third
party logistics provider, who deals with the problem of distributing different types of
products from suppliers to consumers via transportation links. Warehouses between the
suppliers and consumers may be used to compensate for the stochastic behavior of sup-
plies and demands and to consolidate different products. The amounts of these supplies
and demands are assumed to be uncontrollable for the logistics provider. With only in-
formation about the supplies and demands a few days in advance, the logistics provider
has to decide on which transportation links to use for a long period of time (tactical level)
and how much of which products to ship through them each day (operational level).
We formulated a bi-level joint network design and network operation problem: at
the upper (tactical) level, the network topology has to be constructed, and at the lower
(operational) level routings and schedules for daily shipments have to be decided on. A
model predictive control with a rolling horizon (MPC) was used as decision model on this
operational level. A heuristic was proposed to construct the topology dependent on the
operational decisions and compared to a bi-objective branch and bound method. The
consolidation of product types so as to leverage on the scales economies is taken into
consideration in this network design heuristic.
Experimental results reveal that the cost of such a near-optimally operated network
as a function of the number of links in the network stays almost constant as the vast
majority of network links is removed and explodes once the link number has decreased
below a critical value. The experiments show that our heuristic determines a network
that has close to optimal costs with a very low number of links (typically about 10% of
all links that can be deleted before a minimally connected network remains).
Furthermore, experimental results suggested that the resulting topology is insensitive
to second and higher order moments of the individual supply and demand distributions.
Hence, information about the means of supplies and demands over a certain time period
will suffice to generate a close to optimal network topology robust to any supply and
demand scenarios with the same means.
The distribution network design problem, faced in this chapter, involves decision mak-
ing at different levels: a tactical network topology with a minimal number of links and
optimal operational performance is to be found. For real-life instances, this problem is
cannot be solved to optimality within reasonable time. However, the problem is very
suitable to be decoupled into two related problems, one at the tactical and one at the
operational level. The proposed bi-level approach is solved very fast for real-life instances
and still finds a satisfactory solution. In the next four chapters, we proceed with the ap-
CHAPTER 2. DESIGN OF A DISTRIBUTION NETWORK TOPOLOGY 39

proach of multi-step optimization by cutting the combined decision making problem in a

container operation into four main subproblems, which then are solved step by step. Like
in this chapter, the cuts are chosen such that the remaining subproblems are i) interesting
from a practical point of view and ii) solvable within the computation time allowed by the
specific decision level. A subproblem for instance that considers only strategic decisions
is allowed to run for a couple of days or even a week. An operational subproblem that has
to be solved each hour of the day however, should take no more than a couple of minutes
such that results can still be analyzed and interpreted.

A recommendation is the extension of the current approach to include multi-echelon net-

works, allowing shipments between regional and local warehouses in a whole distribution
network. In addition, the inclusion of more accurate travel times, including some that are
smaller or longer than a day can be considered.
The study on distribution networks as presented in this dissertation only involves tac-
tical and operational decisions. As a recommendation, our bi-level optimization scheme
could be extended to allow for different strategic decisions. In particular, while the cur-
rent research studied the influence of the operational level on the tactical level, we have
ignored the strategic level by fixing the number of warehouses in advance. In reality, ware-
houses can be opened or shut depending on the amount of traffic going through them.
Determining the right number and location of potential warehouses for a given source and
sink distribution in a logistic network therefore is an interesting extension.
Chapter 3

Strategic Allocation in a
Multi-Terminal Container Operation

3.1 Introduction
In the previous chapter, we considered the joint decision making problem at the tactical
and operational level in a distribution network. Since the overall problem is too complex
to be solved within satisfactory time, it was cut into two problems, which were solved
alternatingly. Although this procedure does not find a global optimum, it appears to
be very fast and still finds a satisfactory solution. The approach of cutting the overall
decision making problem into multiple subproblems is also applied for a multi-terminal
container operation. In this chapter and chapters 4, 5 and 6, we subsequently address one
of the four chosen subproblems (see Figure 1.5). In this chapter, we start from combined
problems at the strategic and tactical decision making levels (step 1 in Figure 1.5). The
question at the strategic level is whether the same amount of loops can be processed with
less crane capacity. We think this can be achieved by modifying the terminal and time
allocation at the tactical level.
In many container ports, a number of terminal operators takes care of the logistics
processes for container handling. Commonly, the tasks are divided such that one terminal
operator is responsible for one terminal (or at least for the major share of one terminal),
at which various vessel lines have one of their loops calling. The berth allocation problem
(BAP) then involves the allocation of these loops in time and space in order to minimize
a certain objective function ([15], [35], [24], [8], [27], [16], [51], and [36]). All these studies
have in common that they consider the berth allocation problem for a single terminal.
In an increasing number of ports (e.g. Singapore, Antwerp and Rotterdam) however,
one terminal operator is responsible for multiple terminals. In this dissertation, a multi-
terminal container operation in the port of Antwerp is considered, which is run by only
one terminal operator being PSA HNN. The overall problem then becomes to allocate i)
the loop to which a number of vessels belong, to a terminal, ii) a time interval to a vessel
for berthing, iii) a suitable berth position for a vessel within its terminal, and iv) a number
of quay cranes to a vessel, taking the cyclic nature of the system into account. The BAP
in such ports can no longer be considered per terminal for two main reasons. One is that
it makes sense to avoid peaks and troughs in quay crane utilization and to spread vessel
calls evenly over the various terminals. The other is that transshipment containers will
unavoidably generate inter-terminal traffic, whose costs should be taken into account. All
possible flows of containers through such a port are depicted in Figure 3.1.

40
 i i

“portn7nc˙temp” — 2007/8/16 — 16:28 — page 1 — #1

i i
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 41

Seaside

1
inal
Import Import

Termina
Export Export

Term

l3
Termin
al 2

Inter-terminal
Transport
Import
Land side
Export

Figure 3.1: Container flows in a cluster of multiple terminals.

i i

i i

The multi-terminal allocation problem falls naturally apart in four subproblems, fol-
lowing the general classification of decision making: The first is a strategic problem, i.e.
which terminal and which time interval to allocate to each (vessel in the) loop. This
strategic problem is reconsidered occasionally when contracts for new loops are negoti-
ated or existing contracts are renegotiated. The algorithm to solve this problem optimally
may run for several hours, if not days.
Second, it is important to build in some pro-active robustness (by means of quay
and crane reservations) in these timetables such that the risk of delay propagation is
minimized.
The next problem is of a more tactical nature. Given the terminal and berthing interval
for each call, the problem is where to berth the vessels along the quay. This is different
from the studies mentioned above for two reasons. One is that the time window at this
level is no longer a variable of the problem, it is given. Secondly, the objective is not to
minimize the makespan of each vessel, but to use the time allowed to optimally serve the
vessel. The service to be optimized at the tactical level concerns the carriers in between
quay and yard. The objective is to find berth positions at the quay and container areas
in the yard such that the total carrier travel distance is minimized given the strategic
timetable and expected call sizes and call compositions.
The timetables and yard layout per terminal are kept for a year or two and considered
to be the cyclic reference plan. In practice however, all kinds of disturbances take place.
The operational level is concerned with the disruption management of the disturbed
container operation. This involves the allocation of actual start and end process times
to a vessel, the actual berth position to a vessel and the actual quay cranes processing a
vessel.
In this chapter, the strategic problem of allocating a terminal, and a time interval of
berthing to each vessel is addressed. In Chapter 4, the timetable per terminal, constructed
in Chapter 3, is slightly modified to make it more robust to disturbances on vessels’
arrivals. The tactical problem of finding berth positions and container areas is considered
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 42

in Chapter 5. An on-the-fly approach for allocating positions and quay cranes in the
disturbed system is proposed and discussed in Chapter 6.

3.1.1 Related Work

Descriptions and classifications as well as solution methods for the main logistics processes
in container ports are given in [46], [49] and [43]. These studies determine the so-called
berth allocation problem (BAP) as one of the key issues in a container port. Existing
studies as discussed below all consider different versions of the single terminal BAP, which
allocates a berth position and a berth interval to a set of vessels within one terminal. To
our knowledge and as stated in [39], the BAP for multiple interrelated terminals has not
yet been considered.
We however address the multi-terminal BAP, which is concerned with the allocation
of i) a terminal, ii) a berth interval, iii) a berth position within the terminal, and iv) a
number of quay cranes to a set of vessels in a multi-terminal container operation. This
problem is solved in multiple steps: in this chapter, a terminal and a berth interval
is allocated to vessels. Next in Chapter 4, the constructed timetable per terminal is
slightly modified to increase its robustness. Then, appropriate berth positions within
the allocated terminals are generated for the given (robust) timetable (Chapter 5). The
operational quay crane scheduling is addressed in an online procedure in Chapter 6. Each
of the individual subproblems are not only relevant from a practical point of view, but
are also solvable within the time allowed at the corresponding decision level. Although
the subproblems considered are all typically different from the single terminal BAP, still
some principles from the single terminal BAP are relevant and therefore worth discussing
below.
In the last two decades intensive research has been conducted on the single-terminal
BAP. The single-terminal BAP consists of two interrelated assignment problems: allocate
i) a berthing position and ii) a time interval of berthing to each vessel within one terminal
such that vessels are not overlapping. The objective in these studies is often to minimize
the vessels’ turnaround times.
In this dissertation however, the multi-terminal BAP is addressed and solved in mul-
tiple steps. Accordingly, in this chapter we only allocate i) a terminal and ii) a berth
interval (start and end time of processing) to a set of vessels. Multi-objectives are to min-
imize i) the container transport between different terminals (by minimizing the allocation
of connecting vessels to different terminals), ii) the maximally required crane capacity per
terminal (by evenly distributing the workload among the terminals), and deviations from
preferred berth times of vessel lines.
In existing studies, the single-terminal BAP is modeled either as a discrete or a contin-
uous problem. In the discrete case, the quay is divided into segments with specific lengths
or even points when the quay and vessels’ lengths are ignored. The problem can then be
modeled as a parallel machine scheduling problem [42] and [32] where each vessel is a job
and each berth a machine. However, large segments result in a poor space utilization,
whereas small segments might lead to an infeasible solution. The continuous approach,
where vessels can berth anywhere along the quay, circumvents these difficulties, however
is more complex from a computational point of view.
Although, our problem can be considered as a multiple-job-on-one-processor schedul-
ing problem, where each processor represents a terminal and each vessel a job, we only
guarantee non-exceeding of quay and crane capacities, rather than positioning the ves-
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 43

sels along the quay in this chapter. Subsequently in Chapter 5, for a given timetable, a
continuous position allocation problem is solved and hence vessels are assumed to berth
anywhere along the quay. This cut enables to solve both subproblems very efficiently.
Besides the distinction between a continuous and discrete case, existing studies on the
single-terminal BAP consider either a static or dynamic case. The static case assumes all
vessels to be in the port before the berth allocation is determined. This implies that each
vessel can be allocated anywhere in time. The discrete, static berth allocation problem
[21] is an assignment problem and is solvable in polynomial time with the Hungarian
method [40]. This method assigns jobs to machines by sequentially computing shortest
paths until each job is assigned to a machine. In the dynamic case, vessels arrive while
work is in progress, in which case there may be idle times between successive vessels. The
dynamic BAP is NP-hard for both the continuous and the discrete case [14].
In this dissertation, we address a strategic problem and therefore we first assume
that each vessel can be positioned anywhere in time (although deviations from preferred
berthing intervals are minimized). In an additional case study, lower and upper bounds
are placed around start and end berth times, which restricts the vessel order and makes
the problem dynamic. Although the dynamic BAP is NP-hard for both the continuous
and the discrete case [14], we can still solve real-life instances of our problem within
satisfactory time since we abstract from the position allocation in this chapter.
Studies that present heuristics solution approaches for several versions of the discrete
and/or continuous dynamic single-terminal BAP with fixed process times can be found
in [10], [37], [51], [15] and [27]. Other studies however assume that the process time of a
vessel depends on its berth position [23], [24], [22], [20], [8], [16], and [35]. This assumption
results from the fact that carriers have to cover a certain distance between a vessel’s berth
and the designated stacks for the vessel’s containers in the yard.
A limited number of studies addresses a two phase problem of berth allocation and
crane scheduling [20] and [41]. The authors in [20] assume that the process time of a
vessel depends on its berth position, construct a mixed integer linear program for the joint
problem of berth allocation and crane scheduling, and solve it using a genetic algorithm.
The authors in [41] take into account that a vessel’s process time depends is inversely
proportional to the number of cranes assigned to it. The problem is solved in two phases.
The first phase determines the vessels’ positions and berth times and an integer number of
quay cranes in each time segment for each vessel. A sub-gradient optimization technique is
applied to obtain a near-optimal solution of the first phase. The second phase constructs
a schedule for each individual quay crane guaranteeing non-crossing of cranes.
In this chapter, a vessel’s process time is assumed to be inversely proportional to
the crane capacity allocated to it. The actual crane scheduling problem is addressed in
Chapter 6 of this dissertation. The main difference with [41] is that we consider the
multi-terminal BAP (rather than the single-terminal BAP) and cut the problem into four
subproblems, rather than two. The first three subproblems are typically different from
existing studies. Only the subproblem of crane scheduling is quite similar to the second
phase in [41]. We propose a different solution technique for this problem in Chapter 6.
To solve the multi-terminal BAP, it is necessary to incorporate the complex of interact-
ing terminals in one model. This model should take into account the amount of inbound
and outbound containers and their corresponding destinations. Inbound containers of
an arriving vessel for instance could be partly destined for the hinterland and partly for
another vessel. Hence, allocation of the two involved vessels to different terminals implies
inter-terminal traffic and thus additional costs. However, due to other objectives and
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 44

constraints this may still be the best or only solution.

3.1.2 Contributions
In this chapter we consider one terminal operator, who is responsible for a number of
terminals in one container port. A vessel (or actually the loop to which a number of
vessels belong) is allocated to a terminal for a certain amount of time to be processed by
a number of quay cranes. Although we guarantee that terminal quay lengths as well as
quay crane capacities are never exceeded, i) the berth position and ii) the actual quay
crane scheduling within a terminal are still to be determined at a tactical and operational
level, respectively. The tactical level of allocating berth positions to vessels and positions
for their containers in the yard is addressed in Chapter 5, while the operational on-the
fly vessel allocation and quay crane scheduling are treated in Chapter 6. These cuts into
specific subproblems are depicted in Figure 1.5. The strategic subproblem, considered in
this chapter, involves a number of dependent, one-dimensional packing problems, which
allow capacitated parallel processing. This is a very interesting problem from a practical
point of view, which still can be solved within satisfactory time.
Most of the existing studies after the BAP consider a set of vessels within a certain
time horizon. The corresponding objective in these researches often reduces to fitting
all vessels within a time horizon and minimizing the total weighted handling time for
all vessels. However, in practice most vessels run a regular service on their ports, for
instance once a week, which makes the system cyclic. Vessels can arrive at the end of the
considered time period (cycle) and leave at the beginning of this time period (next cycle).
Relating this to the packing problem implies that rectangles (vessels) can be cut into two
pieces, where one piece is placed at the end of the time horizon and the other piece at the
beginning. The authors in [37] take this into consideration for a single terminal BAP.
The contributions of this chapter are the following:
• We address the strategic problem of allocating vessels to a certain terminal for a
certain time interval of processing.
• The model in this chapter takes the cyclic nature of the system into consideration.
• An alternative approach, introduced here, is much faster solvable than the straight-
forward approach. Using the alternative approach, we are able to construct accurate
allocations for real-life problems within a couple of hours. Since such a strategic
problem is only reconsidered once each year or once each two years, a run time of a
couple of hours or even a day is still satisfactory.
• Applying the alternative approach to a representative data set in a case study
suggests that significant reductions in required crane capacities and inter-terminal
transport can be achieved.
The chapter is structured as follows: in Section 3.2, the problem considered is formally
phrased. Then, we introduce a straightforward mixed integer linear program to solve
this problem. In addition, we propose an alternative mixed integer linear program. We
compare the performance and discuss feasibility aspects in Section 3.3. In Section 3.4,
the alternative approach is applied to a representative data set, provided by the terminal
operator PSA HNN in Antwerp. Results suggest that significant reductions in required
crane capacities and inter-terminal transport of a representative can be achieved. Finally,
in Section 3.5, we draw conclusions and make recommendations for future research.
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 45

3.2 Mathematical Models

In this section, we first describe the problem considered in detail. Next, we propose
and discuss two mathematical formulations to solve the problem. The first approach is
straightforward in the sense that the way of modeling the berthing of vessels is common.
The second model significantly reduces the number of variables. The way of modeling the
cyclic property of the system is similar for both approaches.

3.2.1 Problem Description

For all of this chapter the following holds, unless stated differently: t ∈ {1, 2, . . . , T },
the cluster of terminals, v ∈ {1, 2, . . . , V }, the set of vessels, z ∈ {0, 1, 2, . . . , V }, the set
of container destinations. Furthermore, we assume vessels to call cyclically, where each
vessel in the set arrives exactly once each cycle. In general, the cycle length is a week
for such a container operation. We consider discrete time k and unless stated differently,
k ∈ {1, 2, . . . , K}, is the set of discrete time slots within the cycle.
In the cluster of terminals, the set of container vessels has to be unloaded and loaded.
Vessel v imports a pre-determined number of inbound containers Ivz ∈ N with desti-
nation(s) z, where v 6= z. In this context, z = 0 means that containers are destined
for the hinterland, whereas z = 1, 2, . . . , V means that containers are destined for vessel
v = 1, 2, . . . , V respectively. Besides import containers brought in by vessels, a certain
amount of containers Hv with destination v is imported from the hinterland by trucks
and trains during the cycle. These containers are distributed among the different termi-
nals dependent on their destination vessels. Furthermore, each vessel v exports a number
of outbound containers Ov ∈ N. Container transport between among the terminals is
established by trucks.
Terminal t has a restricted quay length Lt ∈ R+ and a number of quay cranes Nt ∈ N.
Once berthing, vessel v occupies a certain amount of quay meters Mv . In addition, this
length Mv determines the maximum number of quay cranes Sv ∈ N processing vessel v
and the efficiency ηv ∈ [0, 1] of the quay cranes on vessel v. In practice, quay cranes
with different processing rates are present in the terminals. We do not take the specific
allocation of quay cranes to vessels into account yet, but assume the average processing
rate λ̄t ∈ N to be the processing rate of each quay crane in terminal t. So the handling
time of vessel v in terminal t depends on i) the mean processing rate λ̄t in terminal t,
ii) the efficiency ηv of quay cranes operating vessel v, iii) the number of quay cranes
processing vessel v and iv) the total number of inbound and outbound containers Ivz and
Ov of vessel v. We assume the processing time of vessel v to be inversely proportional
to the first three of these items and proportional to the latter. Furthermore, the number
of quay cranes processing vessel v may change from one time slot to another. After the
unloading and before the loading, containers can temporarily be stored in the yard of
terminal t up to the yard’s capacity Wt . The time it takes to transport containers from
terminal p to terminal r is defined as ∆pr ∈ N, p, r ∈ {1, ..., T }. Furthermore, we assume
that the total number of time slots vessel v is berthing, is less then the total number of
time slots K in the cycle. In addition, we assume that vessels arrive at the beginning of
a time slot and depart at the end of a time slot.
Our goal is to minimize the total costs of the system, which consist of three conflicting
elements: first of all, costs are associated with the number of quay cranes that have to
be installed in the terminals in order to satisfy the proposed schedule. We define ct to be
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 46

the average costs of a quay crane in terminal t. Second of all, a fixed amount of money
cpr has to be paid for each container that is transported from terminal p to terminal r.
Finally, the terminal operator has some contractual agreements with each of the vessel
lines with respect to the berthing interval of the corresponding vessels. We define Av ∈ N
to be the preferred arrival time of vessel v in the port and Dv ∈ N to be the preferred
departure time of vessel v. We have to remark that, in our definition, this means that
vessel v prefers to depart at the end of time slot hDv − 2, Dv − 1]. We assign factors of
penalty costs ccv and cdv if in the constructed allocation vessel v departs earlier and later
than its preferred departure time Dv , respectively.

Parameter Definition
T Number of terminals in the cluster
V Number of vessels in the set
K Number of discrete time slots within the cycle
Lt Quay length [m]
Mv Quay length required for vessel [m]
Ivz # inbound containers to be unloaded from vessel v with destination z and v 6= z
Ov # outbound containers to be loaded onto vessel v
Hv # containers with destination v arriving from the hinterland during the cycle
Av Preferred arrival time of vessel v
Dv Preferred departure time of vessel v
Ev Parameter to distinguish between the cases Av < Dv and Dv ≥ Av
Nt # quay cranes available in terminal t
Sv Maximum # quay cranes, which can process vessel v
λ̄t Mean processing rate of quay cranes in terminal t [containers/time slot]
ηv Vessel efficiency with respect to quay crane rate [-]
∆pr # time slots needed to transport containers from terminal p to r
Wt # containers that can be stored in terminal t
cav Factor of penalty costs for vessel v for arriving too late [ euro/container
time slot
]
euro/container
ccv Factor of penalty costs for vessel v for departing too early [ time slot ]
cdv Factor of penalty costs for vessel v for departing too late [ euro/container
time slot
]
cpr Factor of transportation costs from terminal p to r [euro/ container]
ct Factor of costs for required equipment in terminal t [euro/ quay crane]

Table 3.1: Model parameters

With respect to the cyclic property of the considered system, we have two additional
remarks. First, we require conservation with respect to the arrival and departure of
containers:
XV
Iiv + Hv = Ov ∀v (3.1)
i=1

Second, we notice that both Av ≥ Dv and Av < Dv are possible, since we model
cyclically arriving container vessels (see also Figure (3.2)). Therefore, we introduce an
auxiliary parameter Ev , which explicitly distinguishes between both cases:

1 if Av ≥ Dv ,
Ev =
0 if Av < Dv . ∀v
 i i i i

“avdv˙temp” — 2007/8/31 — 10:07 — page 1 — #1 “dvav˙temp” — 2007/8/31 — 10:07 — page 1 — #1

i i i i

CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 47

Day 1 Day 1
7 2 7 2

6 3 6 3

5 4 5 4
(a) Av = 2 and Dv = (b) Av = 7 and Dv =
5: Ev = 0. 4: Ev = 1.

Figure 3.2: Possible arrivals and departures of vessel v in terminal t in a cyclic system
with K = 7.

In the next section it becomes clear why we need this parameter. The sets and parameters
discussed above are conveniently arranged in Table 3.1.

3.2.2 Straightforward MILP

i i i i
Binary variables i i i i


1 if in terminal t vessel v berths during time slot hk − 1, k],
atv (k) =
0 otherwise.

1 if in terminal t vessel v departs during time slot hk − 2, k − 1],
dtv (k) =
0 otherwise.

Auxiliary binary variables


1 if in terminal t vessel v is berthing during time slot hk − 1, k],
btv (k) =
0 otherwise.

 1 if atv (ka ) = 1 and dtv (kd ) = 1 and ka > kd ,
etv = 1 if in terminal t vessel v is continuously berthing,
0 otherwise.


1 if atv (ka ) = 1 and ka < Av ,
eav =
0 if atv (ka ) = 1 and ka ≥ Av .

Continuous variables
Integer variables
Constraints
Vessel v can arrives once each cycle at exactly one terminal:
T X
X K
atv (k) = 1 ∀v (3.2)
t=1 k=1
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 48

mtv (k) = Amount of quay meters consumed in terminal t by vessel v during

time slot hk − 1, k] [m]
qtv (k) = Amount of quay processing vessel v in terminal t during time slot hk − 1, k]
htv (k) = Amount of containers from hinterland transported into terminal t
with destination v during time slot hk − 1, k] [containers/ time slot]
fprv (k) = Amount of containers transported from terminal p to terminal r with
destination v during time slot hk − 1, k] [containers/ time slot], p 6= r
wtv (k) = WIP in terminal t with destination v at time k
nt = Number of quay cranes required in terminal t

∆av = Number of time slots vessel v berths too late

∆cv = Number of time slots vessel v departs too early
∆dv = Number of time slots vessel v departs too late

Furthermore, vessel v departs once each cycle from the same terminal it arrives at:
K
X K
X
atv (k) − dtv (k) = 0 ∀t, v (3.3)
k=1 k=1

If vessel v arrives at and departs from terminal t, terminal t is occupied by vessel v only
between its arrival and departure time.
Now assume that vessel v berths at terminal t = 1, arrives at time slot ka and departs
and time slot kd , where the cycle length K = 10. Since the system is cyclic, we can
distinguish three different cases with respect to the berthing of a vessel:
1. ka < kd
This means that the arrival and departure of vessel v take place in the same cycle.
Assume ka = 3 and kd = 7. The corresponding values for a1v (k), d1v (k) and b1v (k)
are depicted in the table below:

k 1 2 3 4 5 6 7 8 9 10
d1v (k) 0 0 0 0 0 0 1 0 0 0
a1v (k) 0 0 1 0 0 0 0 0 0 0
K
d1v (k 0 )
P
1 1 1 1 1 1 0 0 0 0
k0 =k+1
K
a1v (k 0 )
P
1 1 0 0 0 0 0 0 0 0
k0 =k+1
b1v (k) 0 0 1 1 1 1 0 0 0 0

To model the dependency between atv (k) and dtv (k) on one hand, and btv (k) on the
other for the case where ka < kd , we propose the following constraint:
K
X
dtv (k 0 ) − atv (k 0 )


btv (k) = ∀t, v, k (3.4)
k0 =k+1

2. ka > kd
This means that vessel v arrives in one cycle and departs in the next one. Assume
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 49

ka = 8 and kd = 4. The corresponding values for a1v (k), d1v (k) and b1v (k) are
depicted in the table below:

k 1 2 3 4 5 6 7 8 9 10
d1v (k) 0 0 0 1 0 0 0 0 0 0
a1v (k) 0 0 0 0 0 0 0 1 0 0
K
d1v (k 0 )
P
1 1 1 0 0 0 0 0 0 0
k0 =k+1
K
a1v (k 0 )
P
1 1 1 1 1 1 1 0 0 0
k0 =k+1
b1v (k) 0 0 0 -1 -1 -1 -1 0 0 0

To model the dependency between atv (k) and dtv (k) on one hand, and btv (k) on the
other for the case where ka > kd , we propose the following constraint:
K
X
dtv (k 0 ) − atv (k 0 )


btv (k) − 1 = ∀t, v, k (3.5)
k0 =k+1

3. av = dv
This means that vessel v i) arrives and departs during the same time slot in the
same cycle (no berthing at all) or ii) arrives in a certain cycle and departs in the
next time cycle (continuously berthing).
For the case where no berthing takes place, the accompanying values for btv (k) (see
table below) follow from (3.4).

k 1 2 3 4 5 6 7 8 9 10
btv (k) 0 0 0 0 0 0 0 0 0 0

For the case where vessel v is continuously berthing, the accompanying values for
btv (k) (see table below) follow from (3.5).

k 1 2 3 4 5 6 7 8 9 10
btv (k) 1 1 1 1 1 1 1 1 1 1

We introduce the auxiliary binary variable etv to distinguish between the cases as given
in the definition of etv , resulting from the cyclic nature of the system. The principle used
here is similar to the one for the auxiliary parameter Ev (see also Figures 3.2a and 3.2b).
K
X
dtv (k 0 ) − atv (k 0 )


btv (k) − etv = ∀t, v, k (3.6)
k0 =k+1

Furthermore, vessel v can only berth at one terminal:

T
X
btv (k) ≤ 1 ∀v, k (3.7)
t=1
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 50

The sum of lengths of all vessels berthing at terminal t during time slot hk − 1, k] should
be less than or equal to the total quay length in terminal t:
V
X
Mv · btv (k) ≤ Lt ∀t, k (3.8)
v=1

Vessel v can be operated by up to Sv quay cranes in terminal t during time slot hk − 1, k],
only if this vessel is berthing in terminal t during time slot hk − 1, k]:

qtv (k) ≤ Sv · btv (k) ∀t, v, k (3.9)

Vessel v has to be fully processed during the cycle:

T X
X K Z
X
ηv λ̄t · qtv (k) = Ivz + Ov ∀v (3.10)
t=1 k=1 z=0

We want to minimize the maximum number of quay cranes in terminal t ever required
during the cycle. Therefore, we introduce an auxiliary variable nt , which is a soft upper
bound on the number of quay cranes in terminal t. This variable nt is present in the
objective function:
XV
qtv (k) ≤ nt ∀t, k (3.11)
v=1

The maximum number of quay cranes ever required in terminal t during the cycle cannot
be larger than the number of quay cranes actually available in terminal t:

n t ≤ Nt ∀t (3.12)

The sum over the cycle’s time slots of the number of containers with destination v, trans-
ported from the hinterland into the different terminals, should be equal to the total
number of containers with destination v arriving from the hinterland during the cycle.
K X
X T
htv (k) = Hv ∀v (3.13)
k=1 t=1

Since the system is cyclic, the storage level in the terminals and the inter-terminal trans-
port during time slot hk − 1, k] should equal the storage level in the terminals and the
inter-terminal transport during time slot hk − 1 + αK, k + αK], where α ∈ N:

wtv (k) = wtv (k + αK) ∀t, v, k (3.14)

and
fprv (k) = fprv (k + αK) ∀p, r, v, k (3.15)
We assume that inbound containers with destination 0 (”hinterland”) are transported
into the hinterland directly after they arrive in the terminal and are not counted as stack.
The amount of containers in terminal t with destination v during time slot hk − 1, k]
is equal to the amount of containers in terminal t with destination v during time slot
hk − 2, k − 1] plus all incoming flows (inbound containers from vessels, containers from
other terminals and containers from the hinterland) minus all outgoing flows (outbound
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 51

containers to vessels and containers to other terminals). We assume that loading and
unloading of containers from vessel v with different destinations is divided proportionally
among the time slots vessel v is actually berthing. To this, we first define the constants
βvz = PZ
Ivz
and γv = P Z
Ov
, and derive appropriate constraints:
Ivz +Ov Ivz +Ov
z=0 z=0

V
X
wtv (k) = wtv (k − 1) + βiv ηi λ̄t · qti (k) − γv ηv λ̄t · qtv (k) + htv (k) + (3.16)
i=1
T
X T
X
frtv (k − ∆pr ) − ftrv (k) ∀t, v, k
r=1 r=1

If we start with bringing containers into the yard during time slot hk − 1, k], the following
constraint has to be satisfied:
XV V
X T
X

wtv (k − 1) + βiv ηi λ̄t · qti (k) + htv (k) + frtv (k − ∆pr ) ≤ Wt ∀t,(3.17)
k
v=1 i=1 r=1

If we start with taking away containers from the yard during time slot hk − 1, k], the
following constraint has to be satisfied:
T
X
wtv (k − 1) − γv ηv λ̄t · qtv (k) − ftrv (k) ≥ 0 ∀t, v, k (3.18)
r=1

Whatever order is applied during the cycle, (6.9) and (3.18) guarantee that never too
much and never too less (negative amount of) containers are in the yard.
We use the additional integer variables ∆av , ∆cv and ∆dv to model the number of time
slots vessel v i
berths too late, departs too early and departs too late, respectively. We
assume vessel v arrives in the port at time Av ,“deltaa˙temp”
which means that the actual—berth
— 2007/8/31 10:51time
— page 1 — #1
can onlyi take place exactly at or later than A . We assume that the costs ∆a for late
v v
arrival of vessel v depend linearly on the difference between the preferred arrival time Av
and the actual arrival time. Due to the cyclic nature of the system, both atv (ka ) = 1,
where ka < Av and ka ≥ Av are possible. Hence, a jump in the cost function occurs at
Av as depicted in Figure (3.3).

∆a
v

A K
v
Actual berth time

Figure 3.3: Costs for berthing too late.

To model this jump, we introduce the auxiliary binary variable eav in an additional
constraint:
T X
X K  
a
− Av − k · atv (k) + K · eav


∆v = ∀v (3.19)
t=1 k=1
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 52

where
T X
X K
eav


K· ≥ Av − k · atv (k) ∀v (3.20)
t=1 k=1

With respect to departing too early or too late one can distinguish 4! = 24 permutations
(of av , Av , dv and Dv . It turns out that with help of the introduced auxiliary binary
variables eav and etv , and the auxiliary parameter Ev as defined in Section 3.1, we are able
to construct appropriate constraints for ∆cv and ∆dv to satisfy each of the 24 cases:
T X
X K   T
X
∆cv a


≥ Dv − k · dtv (k) − K · ev + K · Ev − K · etv ∀v (3.21)
t=1 k=1 t=1

where
∆cv ≥ 0 ∀v (3.22)
and
T X
X K   T
X
∆dv a


≥ − Dv − k · dtv (k) + K · ev − K · Ev + K · etv ∀v (3.23)
t=1 k=1 t=1

where
∆dv ≥ 0 ∀v (3.24)
Finally, some of the continuous variables have to be lower-bounded:

qtv (k) ≥ 0 (3.25)

htv (k) ≥ 0 (3.26)
fprz (k) ≥ 0 (3.27)
wtz (k) ≥ 0 (3.28)

Objective function
Linear penalty costs are assigned when vessel v berths later than its arrival time and/
or when vessel v departs too early or too late (cav , ccv , cdv respectively). Furthermore,
a linear unit penalty cost is assigned when containers are transported from one termi-
nal to another (cpr ). Finally, linear costs are assigned to the number of required quay
cranes in terminal t (ct ). The decision variables are represented in a vector ~u(k) =
[atv (k), dtv (k), htv (k), qtv (k), fprv (k)]T and the objective function is formulated:
V
X K X
X T X
T X Z T
X
cav ∆av ccv ∆cv cdv ∆dv


min + + + cpr fprz (k) + ct nt (3.29)
~
u(1),...,~
u(K)
v=1 k=1 p=1 r=1 z=1 t=1

Remark: In the solution of this MILP it could be that an arbitrary amount of containers is
stored in a certain terminal during the entire cycle. This could be prevented by assigning
a small cost for each stored container.
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 53

3.2.3 Alternative MILP

In the previous section, we introduced a straightforward approach of modeling the prob-
lem. The cyclic nature of the system is taken into account by (4.9), (4.6) and (4.13),
and (3.19) through (3.24). In this straightforward formulation, we used binary variables
atv (k) and dtv (k), which indicate whether or not a vessel berths at or departs from ter-
minal t during time slot hk − 1, k]. In the alternative approach we split these variables
into the integer variables av and dv and the binary variable xtv . Here, av and dv denote
the time slots vessel v berths and departs, respectively. In our definition dv means that
the processing of vessel v ends at the end of time slot hk − 2, k − 1]. Additionally, xtv
denotes the terminal in which vessel v berths. Consequently, some of the constraints of
the straightforward approach have to be adapted and even some new constraints have
to be introduced to describe the same problem. In the end, however, the alternative
way of modeling uses only a fraction 3TTK+T
+K+2
+1
of the number of binary variables in the
straightforward way of modeling.

Continuous variables

av = Actual berth time slot of vessel v (start of processing vessel v)

dv = Actual departure time slot of vessel v (processing of vessel v ends at the end of
time slot hk − 2, k − 1]

The rest of the continuous variables are equivalent to the continuous variables as
described in Section 3.2.

Binary variables

1 if in terminal t vessel v berths,
xtv =
0 otherwise.

Auxiliary binary variables


1 if vessel v is berthing during time slot hk − 1, k],
bv (k) =
0 otherwise.


 1 if av > dv ,
0 if av < dv ,

ev =

 1 if av = dv and vessel v is continuously berthing,
0 if av = dv and vessel v does not berth at all.


1 if av < Av ,
eav =
0 if av ≥ Av .

Constraints
Vessel v berths at only one terminal t:
T
X
xtv = 1 ∀v (3.30)
t=1
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 54

The arrival and departure times (av and dv respectively) of vessel v are within the cycle:
1 ≤ av ≤ K ∀v (3.31)
and
1 ≤ dv ≤ K ∀v (3.32)
Since the system is cyclic, we can distinguish three different cases with respect to the
berthing of a vessel:
1. av < dv
This means that vessel v arrives and departs in the same cycle. To better explain
this, we present the next example: assume K = 10, av = 3 and dv = 7. This means
that, according to its definition, the variable bv (k) should have the following values:

k 1 2 3 4 5 6 7 8 9 10
bv (k) 0 0 1 1 1 1 0 0 0 0

To model the dependency between av and dv on one hand, and bv (k) on the other for
the case where av < dv , we propose the following constraints: first of all, the number
of berthing time slots should be equal to the difference between the departure and
arrival time slot:
XK
bv (k) = dv − av (3.33)
k=1
Second of all, after the departure of vessel v, vessel v is not berthing anymore:
k · bv (k) ≤ dv − 1 (3.34)
Finally, before the arrival of vessel v, vessel v is not berthing yet:
(K − k) · bv (k) ≤ K − av (3.35)

2. av > dv
This means that vessel v arrives in a certain cycle and departs in the next one.
To better explain this, we present the next example. Assume K = 10, av =8 and
dv = 4. In the table below, we not only depict corresponding values of bv (k), but
also of bv (k) − 1.

k 1 2 3 4 5 6 7 8 9 10
bv (k) 1 1 1 0 0 0 0 1 1 1
bv (k) − 1 0 0 0 -1 -1 -1 -1 0 0 0

The last row in this table enables to derive constraints similar to the ones we derived
in the first case. First of all, the number of berthing time slots should be equal to
the difference between the departure and arrival time slot plus K:
K
X

bv (k) − 1 = dv − av (3.36)
k=1
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 55

Second of all, after the arrival of vessel v, vessel v is berthing:

−k · bv (k) − 1 ≤ av − 1 (3.37)

Finally, before the departure of vessel v, vessel v is berthing:

−(K − k) · bv (k) − 1 ≤ K − dv (3.38)

3. av = dv
This means that vessel v i) arrives and departs during the same time slot in the
same cycle (no berthing at all) or ii) arrives in a certain cycle and departs in the
next time cycle (continuously berthing).
For the case where no berthing takes place, the accompanying values for bv (k) (see
table below) follow from (4.10).

k 1 2 3 4 5 6 7 8 9 10
bv (k) 0 0 0 0 0 0 0 0 0 0

For the case where vessel v is continuously berthing, the accompanying values for
bv (k) (see table below) follow from (4.14).

k 1 2 3 4 5 6 7 8 9 10
bv (k) 1 1 1 1 1 1 1 1 1 1

We want to derive generic constraints, which relate av and dv to bv (k) as well as bv (k)
to av and dv for each of the upper three cases. To this we introduce the binary variable
ev and combine (4.10) and (4.14) into
K
X

bv (k) − ev = dv − av ∀v (3.39)
k=1

and

1 − av ≤ k · bv (k) − ev ≤ dv − 1 ∀v, k (3.40)
and


dv − K ≤ K − k · bv (k) − ev ≤ K − av ∀v, k (3.41)
Vessel v requires an amount of quay meters Mv at a terminal t during time slot hk − 1, k],
iff the vessel is actually berthing during time slot hk − 1, k] at terminal t.

mtv (k) ≤ Mv · xtv ∀t, v, k (3.42)

and
T
X
mtv (k) = Mv · bv (k) ∀v, k (3.43)
t=1
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 56

Furthermore, the sum of lengths of all vessels berthing at terminal t during time slot
hk − 1, k] should be less than or equal to the total quay length of terminal t:
V
X
mtv (k) ≤ Lt ∀t, k (3.44)
v=1

Vessel v can only be operated in terminal t iff vessel v is berthing in terminal t. Further-
more, a maximum number of quay cranes Sv can be assigned to vessel v:
qtv (k) ≤ Sv · xtv ∀t, v, k (3.45)
and
qtv (k) ≤ Sv · bv (k) ∀t, v, k (3.46)
Next, constraints (4.14) through (3.18) from the straightforward MILP are also valid in
this formulation. Furthermore, we have to slightly adapt constraints (3.19) through (3.24):
∆av = − Av − av + K · eav


∀v (3.47)
where
K · eav ≥ Av − av ∀v (3.48)
and
∆cv ≥ (Dv − dv ) − K · eav + K · Ev − K · ev ∀v (3.49)
where
∆cv ≥ 0 ∀v (3.50)
and
∆dv ≥ − (Dv − dv ) − K · eav + K · Ev − K · ev


∀v (3.51)
where
∆dv ≥ 0 ∀v (3.52)
All 24 permutations for the sequence of av , dv , Av and Dv in the cycle and the corre-
sponding value of ∆dv are depicted in Figures 3.4 and 3.5.

Objective function
The decision variables are represented in a vector ~u(k) = [xtv , av , dv , htv (k), qtv (k), fprz (k)]T
and the objective function is equivalent to the one in (3.29).

3.3 Results
In the previous chapter, two approaches have been formulated, which model the de-
scribed system. As a next step, both approaches are coded in Matlab and solved using
CPLEX. Results for a large set of randomly generated instances suggest that the models
find the same solution, but significantly differ in CPU time. In this section, we statis-
tically compare the CPU times of both models dependent on the number of vessels V
in the set. Results suggest that the alternative approach convincingly outperforms the
straight-forward approach. Next, we investigate the CPU time of the alternative approach
dependent on the number of time slots K in the considered cycle. Finally, we suggest that
from generated terminal and time window allocations of realistic problems, we are able
to construct feasible and satisfactory i) two-dimensional packing solutions and ii) quay
crane allocations.
 i i i i

“aAdD1˙temp” — 2008/10/28 — 10:19 — page 1 — #1 “aAdD2˙temp” — 2008/10/28 — 10:19 — page 1 — #1

i i i i

CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 57

∆dv ∆dv

av Av dv Dv 2K av Av Dv dv 2K

K K

i i i i

“aAdD3˙temp” — 2008/10/28 — 10:19 — page 1 — #1 “aAdD4˙temp” — 2008/10/28 — 10:19 — page 1 — #1

i i
dv i dv i
Dv Dv

(a) (b)
d
∆v ∆dv

av dv Av Dv 2K av Dv Av dv 2K

K K

i i i i

“aAdD5˙temp” — 2008/10/28 — 10:19 — page 1 — #1 “aAdD6˙temp” — 2008/10/28 — 10:19 — page 1 — #1

i i
dv i dv i
Dv Dv

(c) (d)
d
∆v ∆dv

av dv Dv Av 2K av Dv dv Av 2K

K K

i i i i

“aAdD7˙temp” — 2008/10/28 — 10:25 — page 1 — #1 “aAdD8˙temp” — 2008/10/28 — 10:25 — page 1 — #1

i i
dv i dv i
Dv Dv

(e) (f)
i d i i i
∆v ∆dv
i i i i
dv av Av Dv 2K Dv av Av dv 2K

K K

i i i i

“aAdD9˙temp” — 2008/10/28 — 10:25 — page 1 — #1 “aAdD10˙temp” — 2008/10/28 — 10:25 — page 1 — #1

i i
dv i dv i
Dv Dv

(g) (h)
i d i i i
∆v ∆dv
i i i i
dv av Dv Av 2K Dv av dv Av 2K

K K

i i i i

“aAdD11˙temp” — 2008/10/28 — 10:25 — page 1 — #1 “aAdD12˙temp” — 2008/10/28 — 10:25 — page 1 — #1

i i
dv i dv i
Dv Dv

(i) (j)
i d i i i
∆v ∆dv
i i i i
dv Dv av Av 2K Dv dv av Av 2K

K K

Dv dv Dv dv

(k) (l)
i i i i

i
Figure 3.4: Twelve permutations for the sequence of av , dv , Av and Dv in the cycle and
i i i

the corresponding delay in departure ∆dv .

i i i i

i i i i
 i i i i

“aAdD13˙temp” — 2008/10/28 — 10:25 — page 1 — #1 “aAdD14˙temp” — 2008/10/28 — 10:25 — page 1 — #1

i i i i

CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 58

∆dv ∆dv

Av av Dv dv 2K Av av dv Dv 2K

K K

i i i i

“aAdD15˙temp” — 2008/10/28 — 10:25 — page 1 — #1 “aAdD16˙temp” — 2008/10/28 — 10:25 — page 1 — #1

i i
dv i dv i
Dv Dv

(a) (b)
d
∆v ∆dv

Av Dv av dv 2K Av dv av Dv 2K

K K

i i i i

“aAdD17˙temp” — 2008/10/28 — 10:25 — page 1 — #1 “aAdD18˙temp” — 2008/10/28 — 10:25 — page 1 — #1

i i
dv i dv i
Dv Dv

(c) (d)
d
∆v ∆dv

Av Dv dv av 2K Av dv Dv av 2K

K K

i i i i

“aAdD19˙temp” — 2008/10/28 — 10:32 — page 1 — #1 “aAdD20˙temp” — 2008/10/28 — 10:32 — page 1 — #1

i i
dv i dv i
Dv Dv

(e) (f)
i d i i i
∆v ∆dv
i i i i
dv Av av Dv 2K Dv Av av dv 2K

K K

i i i i

“aAdD21˙temp” — 2008/10/28 — 10:32 — page 1 — #1 “aAdD22˙temp” — 2008/10/28 — 10:32 — page 1 — #1

i i
dv i dv i
Dv Dv

(g) (h)
i d i i i
∆v ∆dv
i i i i
dv Av Dv av 2K Dv Av dv av 2K

K K

i i i i

“aAdD23˙temp” — 2008/10/28 — 10:32 — page 1 — #1 “aAdD24˙temp” — 2008/10/28 — 10:32 — page 1 — #1

i i
dv i dv i
Dv Dv

(i) (j)
i d i i i
∆v ∆dv
i i i i
dv Dv Av av 2K Dv dv Av av 2K

K K

Dv dv Dv dv

(k) (l)
i i i i

i
Figure 3.5: Twelve permutations for the sequence of av , dv , Av and Dv in the cycle and
i i i

the corresponding delay in departure ∆dv .

i i i i

i i i i
 CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 59

3.3.1 Performance Analysis

In this section, we first compare the CPU time of both approaches dependent on the
number of vessels in the set. We consider a problem with a cluster of three terminals
T = 3 and a one-week-cycle, where each time slot has a width of one day, hence K = 7.
The rest of the parameters, as described in Section 3.1, are randomly generated. The
parameter set is used as input data and, together with the model, fed into CPLEX. The
mixed integer optimization is terminated as soon as it has found a feasible integer solution
proven to be within 5% of optimal. For each value of V ∈ {6, 7, 8, ..., 40}, we randomly
(within reasonable bounds following from practice) generate 60 parameter sets and solve
both approaches for these instances. Furthermore, for each of these optimizations, we
monitor the time CPLEX’s CPU time. Figure 3.6a shows the corresponding geometric
means and 95% confidence intervals of the CPU time dependent on V for both approaches.
i
Due to relatively large CPU times when i
using the straightforward
i
approach, V is varied i
between six and twenty only in this case.
“CPUn1˙temp” — 2008/12/1 — 15:05 — page 1 — #1
The relative difference between the optima of
“CPUn2˙temp” — 2008/12/1 — 15:17 — page 1 — #1
i i i
both approaches for each parameter set are found to be within the 5% optimality gap,
which suggests that indeed both models describe the same problem.
3 4
10 10

3
2 10
10

2
CPU time[s]

CPU time[s]

10
1
10
1
10

0
10 0
10
Alternative
K
Straightforward
−1 −1
10 10
5 10 15 20 25 30 35 40 6 8 10 12 14 16 18 20
V V
(a) Straightforward versus Alternative approach. (b) Alternative approach for K ∈ {7, 14, 21, 42}.

Figure 3.6: CPU time analysis.

From Figure 3.6a it is obvious that the alternative approach significantly outperforms
the straightforward approach: First of all, the mean CPU time of the alternative approach
is significantly shorter than the mean runtime of the straightforward approach for each
point in the considered interval. Moreover, the fraction between the CPU time of the
straightforward approach and the CPU time of the alternative approach increases expo-
nentially with the number of vessels in the set. Furthermore, the confidence interval of
the straightforward approach starts diverging at a smaller V value than the confidence
interval of the alternative approach. We now assume that the order of diverging remains
the same for larger V values. A real-life instance, where the number of vessels in the set
i is forty, then could CPLEX take weeksi or even months if the
i straightforward approach is

i applied and only a couple of minutes if ithe alternative approach

i is applied. Therefore, in i

the remainder of this chapter, only the alternative approach is used.

The final goal is to consider a real-life case and to improve its current allocation
by applying the alternative approach. In order to generate a more and more detailed
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 60

schedule, the width of a time slot should be decreased, at the expense of larger CPU times.
Nowadays, planners of such container operations construct a schedule composed of time
slots with 2 or 4 hours width. Therefore, we are interested in the CPU time dependent on
the number of time slots K in the cycle. Again, we consider a port with three terminals
T = 3 and a one-week-cycle. Additionally, for each value of V ∈ {6, 7, 8, ..., 20}, we
generate 10 parameter sets and solve the model for K ∈ {7, 14, 21, 42}, sequentially. The
geometric means and the 95% confidence intervals are shown in Figure 3.6b, where K
increases along the direction of the arrow. As expected, the CPU time increases as K
increases. Furthermore, the four graphs for the geometric means exhibit approximately
the same slope, which suggests that, independent of K the CPU time grows with the
same exponential rate as V increases. Additional experiments suggest that a problem of
real-life size (V = 40) can be solved within about 6000 seconds for the case with 42 time
slots (for the one-week-cycle, the width of a time slot is then 4 hours). This suggests
that with the alternative approach a rather accurate allocation can be constructed for a
real-life problem within a couple of hours.

3.3.2 Feasibility
Although the proposed method allocates a terminal, a number of quay cranes and a time
interval of berthing to a vessel, the actual position within that terminal as well as the ac-
tual quay cranes processing that vessel are not specifically generated (see also Figure 1.5).
This cut allows to solve the model relatively fast, at the expense of possible infeasibility of
the found solution on an operational level. If the terminals would continuously be utilized
against their quay lengths capacities, this could lead to a situation where (4.11) is fulfilled,
however a feasible two-dimensional packing solution does not exist. In practice however,
ports require a significant utilization margin to compensate for disturbances (e.g. late
arrivals and departures) on an operational level. Results in Chapters 4 and 6 suggest
that strategic allocations, generated in the first optimization step, always yield feasible
position and quay crane allocation in the following optimization steps.

3.4 Case study

We consider three interacting terminals (T = 3) in the port of Antwerp, where thirty-
seven vessel lines have one of their vessels processed exactly once a week (V = 37).
Furthermore, we assume that each vessel line has a preferred terminal and a preferred
arriving and departure time, which fit best to their schedule. The current policy of PSA
HNN is commercially driven and aims to satisfy these preferred allocation as good as
possible. The induced costs for the required number of quay cranes and inter-terminal
traffic are of lower priority in this policy.
We are interested in potential reductions in required crane capacity and inter-terminal
traffic if we allow small modifications to a representative allocation, constructed while ap-
plying the policy of satisfying shipping lines’ preferences. A simple visual analysis (without
any computation at all) of the port learns the following: each terminal shows heavy fluc-
tuations in the workload distribution over the shifts of the week. At busy shifts, the entire
quay crane capacity is totally occupied, whereas at quiet shifts not one crane is working at
all. A second observation is the relatively large amount of transshipment containers that
have to be transported from one terminal to another due to the allocation of connecting
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 61

vessels to different terminals. Both these phenomena result from the discussed policy of
the terminal operator.
As an illustration, Figure 3.7 depicts the scaled number of quay crane activities in one
of the terminals for each hour in a a week cycle (K = 168) according to an allocation,
which is representative for the situation in Antwerp. The black line represents the scaled
mean crane capacity
i usage per hour. From the high fluctuations in crane usage,i the
following can
i
be concluded: i) during a couple
“cur1˙temp” of—time
— 2008/12/1 16:10 —slots,
page 1 —a#1
high amount of the crane
i
capacity is needed and ii) during a lot of time slots a large percentage of this amount is
simply not used.
1

0.8
# Quay cranes

0.6

0.4

0.2

0
0 50 100 150
k

Figure 3.7: Current quay crane usage in terminal 1 during the hours of a one-week cycle.

We are interested in the benefit of flexibility, i.e. the potential reduction of the required
number of quay cranes and the costs for inter-terminal transport when both the current
terminal and time allocation are allowed to be adapted within certain bounds. We slightly
expand the alternative MILP, such that for an arbitrarily chosen level of flexibility, the
required quay crane capacity and inter-terminal transport is minimized.
Nowadays, the allocations in ports are constructed on a one or two hour(s) time grid.
Since these sizes of time slots lead to large computation times when solving the MILP,
the optimization is performed in two steps: First, the MILP is solved for time slots of
eight hours. For a weekly cycle this means K = 21. The model can then be solved within
minutes while generated allocations are still quite accurately, since port employees work
in shifts and vessels commonly berth during multiples of a shift. In a second step, a
similar MILP is built to refine the constructed allocation per terminal to a one-hour time
grid. Solving
i this MILP takes less than a tenth of a second. i

The resulting
i
two-step optimization approach enables us to efficiently investigate
i
the
dependency of the cost savings on the level of flexibility: a number of vessels and the
maximal modification in their berthing interval can arbitrarily be selected. Next, these
settings can easily be implemented in the model to determine possible savings in quay
crane and inter-terminal transport costs. In Section 3.4.1, we discuss the two-step op-
timization in detail and apply it in two experiments presented in sections 3.4.2.2 and
3.4.3.
In the first experiment, our hypothesis is that a slight adaptation of the allocation
of only a couple of vessels already leads to large cost reductions. Hence, the terminal
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 62

allocation of a small part of the vessels is chosen to be flexible and the time position
of the berthing interval can be shifted between some pre-specified bounds. For different
values of these bounds, we construct Pareto frontiers of the number of required cranes
versus the inter-terminal transport costs, which confirm the hypothesis. One of the points
from one of the Pareto-frontiers is highlighted to further illuminate the hypothesis. This
point represents the allocation after selecting less than one third of the vessels and allowing
i) a change in their terminal allocation and ii) a shift in the time position of the berthing
interval of these vessels of maximal 24 hours (Gv = 24). The results of the two-step
optimization suggest that about 25% of the number of quay cranes can be saved while at
the same time the inter-terminal transportation costs are reduced by 3%.
In the second experiment, we assume that all vessel lines prefer to stay with their
current time position of the berthing interval (Gv = 0), while they allow a different
terminal allocation. Consequently, the terminal allocation of all vessels is chosen to be
flexible while the time positioning of the berthing interval is fixed to the existing one.
Results suggest that with the same number of quay cranes as in the current allocation,
40% of the costs for inter-terminal transport can be saved by adapting the terminal
allocation.

3.4.1 Two-step approach

Step 1
We want to minimize the number of quay cranes required for the current throughput, and
at the same time reduce the costs for inter-terminal transport by adapting the current
terminal and time allocation. We expect that changing the current terminal allocation
of a couple of vessels and slightly shifting their current berthing interval in time already
leads to a significant reduction in both these objectives. Hence, we select 11 vessels (out
of the total set) from the busy peaks arbitrarily and define two sets: the set S to be the
set of 11 vessels, which current terminal and berthing allocation can be adapted, and the
set Fij to be the set of index pairs of all the other vessels together with their current
terminal. We have to specify additional constraints for the vessels in the different sets.
First of all, for the vessels in Fij , the time-position of the berthing interval should be
fixed:

ai = A i , i ∈ Fij (3.53)
di = Di , i ∈ Fij (3.54)
Furthermore, the current terminal allocation for the vessels in Fij should be fixed:

xij = 1, ∀i, j ∈ Fij , (3.55)

The vessels in set S on the other hand are free to be allocated to any of the terminals
according to (3.30). Additionally, we allow some freedom in the time allocation of their
berthing interval, while the length of the berthing interval remains equal to Pv according
to:
XK
bv (k) = Pv , ∀v, (3.56)
k=1

where bv (k) is 1 iff vessel v berths during time interval [k, k + 1i and 0 otherwise. Fur-
thermore, the time position of the berthing interval of the vessels in S can be placed
 i i

“gapsp˙temp”
CHAPTER 3. STRATEGIC ALLOCATION — 2008/4/8
IN—A15:24 — page 1 — #1
CONTAINER OPERATION 63
i i

maximally Gv time slots earlier or Gv time slots later with respect to its current time
position as shown in Figure 3.8.
Avl Av Dv Dvu
Gv Pv Gv
time

av dv

Figure 3.8: Flexibility in the time-position of the berthing interval Pv of vessel v ∈ S.

The corresponding lower and upper bounds for berthing of vessel v ∈ S, Alv and Dvu
respectively, are then given by:

Av − G v if Av − Gv ≥ 1,
Alv = (3.57)
Av − Gv + K otherwise ∀v ∈ S,


Dv + Gv if Dv + Gv ≤ K,
Dvu = (3.58)
Dv + Gv − K otherwise ∀v ∈ S.

Due to the cyclic property of the system, we distinguish two cases and derive appropriate
constraints for the vessels in S accordingly:
1. Alv < Dvu :

ev = 0, (3.59)
av ≥ Alv , (3.60)
dv ≤ Dvu . (3.61)
2. Dvu < Alv (additional binary variables eav and edv are introduced):

K · eav ≥ Alv − av (3.62)

−K · (1 − eav ) ≤ Alv − av (3.63)

K · edv ≥ Dul − dv (3.64)

−K · (1 − edv ) ≤ Dul − dv (3.65)
i i

i dv ≥ av + K · eav − K · edv . (3.66) i

Experiments show that, with the resulting MILP, allocations on a grid of eight-hours
time slots are generated within minutes. Further decreasing the width of the time slots
turns out to lead to large computation times. Hence, as a first step, the model is built of
eight hours time slots. The actual berthing time is therefore rounded up to a multiple of
a shift to express the parameter Pv of vessel v in the model.
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 64

Step 2
Since in today’s ports most allocations are constructed on a two-hours or even on a one-
hours grid, we introduce a second step to refine the constructed allocation per terminal
from eight-hours time slots into one-hour time slots. We define pnv to be the berthing time
of vessel v on the refined time grid and thus pnv ≤ Pv . Additionally, we require that the
refined berthing time interval is positioned between the allocated arrival and departure
time (optimal values a?v and d?v of the first step optimization) on the coarse time grid.
Hence, we introduce the variables anv and dnv to be the arrival and departure time of vessel
v on the refined time grid, respectively. In the second step, we build an MILP with
similar constraints as given in (3.59) through (3.66), where Alv is substituted by a?v , Dvu
by d?v , av by anv , dv by dnv , and eav and edv by ena nd
v and ev , respectively. Since the vessels
are already distributed among the terminals, the remaining objective is to minimize the
required crane capacity. The computation time turns out to be less then a second per
terminal.

3.4.2 Experiment 1
For each value of Gv ∈ {0, 8, 16, 24, 48} (in hours) for the vessels in S, a Pareto frontier
of the total number of required quay cranes versus the costs for inter-terminal transport
is constructed. Each point in a frontier results from a single two-step optimization with a
specific ratio between costs for quay cranes and costs for inter-terminal transport. The re-
sults are depicted in Figure 3.9a. The cross represents the state of the allocation currently
applied in Antwerp. From Figure 3.9a the following can be concluded:

• For Gv = 0 for v ∈ S yet the costs for inter-terminal transport or the number of
required quay cranes can be reduced. Apparently, an adaptation in the terminal
allocation of the vessels in S suffices to achieve this.

• For Gv = 0 for v ∈ S yet a reduction in the number of quay cranes is possible at

the expense of higher inter-terminal costs.

• The improvements going from Gv = 0 to Gv = 8 are relatively large, whereas the

improvements going from Gv = 24 to Gv = 48 are approximately zero.

• All fronts intersect (the upper left point) where the crane costs are zero. Appar-
ently, the inter-terminal costs cannot be further reduced even if Gv grows and the
maximum number of cranes is used.

• The grey bullet suggests that if Gv = 24 for v ∈ S the number of required quay
cranes can be reduced by almost 25% and the costs for inter-terminal by about
3%. This means that besides a possible change in terminal allocation, the time
allocation of only 11 vessels has to be shifted one day maximally to gain significant
improvements.

The quay crane usage of the allocation, represented by the grey bullet in Figure 3.9a,
is depicted in Figures 3.10b, 3.10d and 3.10f. The results are scaled to the quay crane
usage in the current allocation as shown in Figures 3.10a, 3.10c and 3.10e. The black lines
represent the mean quay crane usage per hour in the different terminals. If we compare
Figures 3.10a, 3.10c, 3.10e with Figures 3.10b, 3.10d, 3.10f, the following can be noticed:
 i i i

“case1˙temp” — 2008/12/1 — 18:52 — page 1 — #1 “case2˙temp” — 2008/12/1 — 18:52 — page 1 — #1

i i i

CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 65

1 Current Allocation 1

0.95
Total number of quay cranes

0.95

Scaled total costs

0.9

0.85 0.9

0.8
0.85
0.75

0.7 0.8
0.5 0.6 0.7 0.8 0.9 1 0 10 20 30 40
Inter−terminal Costs Gv[hrs]

(a) Pareto frontiers. (b) Costs vs flexibility level Gv for different cost
ratios.

Figure 3.9: Improvements of introducing a flexibility level.

• The workload in the generated allocation is better balanced than in the current
allocation. This results in the previously mentioned reduction of almost 25% of the
required quay cranes.

• At some points in time still some quay cranes are not working in the generated
allocation. Introducing either a higher level of flexibility (by increasing Gv , v ∈ S)
or including more vessels into the set S would probably fill up these gaps and lead
to an even better workload balance and a smaller number of required cranes.

• The mean quay crane usage inia specific terminal can differ for the current allocation
i i
and the generated allocation. This can be explained by a difference in terminal
i
i i
allocation of the vessels in S. The total quay crane usage however is equal for both
allocations.

Additionally, we depict the benefit of modification in a different way. For a con-

stant ratio of quay crane costs and inter-terminal costs, the scaled total costs are plot-
ted versus the level of flexibility Gv . Figure 3.9b presents the results for 10 ratios
( ccprt ∈ {0, 20, 40, 60, 80, 100, 120, 140, 160, ∞}). From this figure, the following is noticed:

• For the ratio equal to 0 (no costs for quay cranes), the total costs (for this ratio only
inter-terminal costs) are not affected as the level of flexibility increases. Apparently,
the inter-terminal costs are not affected by Gv , because the terminal length and
crane capacity are not binding at Gv = 0, so that the ’best’ (lowest inter-terminal
traffic) solution is already obtained.

• For all ratios larger than zero the total costs decrease as the level of flexibility
increases.

• As the ratio grows between zero and forty, the total costs-decrease becomes relatively
large. Apparently, these are the sensitive ratios, where a slide increase of the crane
costs already leads to a significant reduction in quay cranes. This suggests that the
current number of quay cranes is far from minimal.
 i i i

“cur1˙temp” — 2008/12/1 — 16:10 — page 1 — #1 “gen1˙temp” — 2008/12/1 — 16:10 — page 1 — #1

i i i

CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 66

1 1

0.8 0.8
# Quay cranes

# Quay cranes
0.6 0.6

0.4 0.4

0.2 0.2
i i i

“cur2˙temp” — 2008/12/1 — 16:10 — page 1 — #1 “gen2˙temp” — 2008/12/1 — 16:10 — page 1 — #1

i 0 i 0 i
0 50 100 150 0 50 100 150
k k
(a) Current allocation for terminal 1. (b) Generated allocation for terminal 1.
1 1

0.8 0.8
# Quay cranes

# Quay cranes
0.6 0.6

0.4 0.4

0.2 0.2
i i i

“cur3˙temp” — 2008/12/1 — 16:10 — page 1 — #1 “gen3˙temp” — 2008/12/1 — 16:10 — page 1 — #1

i 0 i 0 i
0 50 100 150 0 50 100 150
k k
(c) Current allocation terminal 2. (d) Generated allocation terminal 2.
1 1
i i i

i i i
0.8 0.8
# Quay cranes

# Quay cranes

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 50 100 150 0 50 100 150
k k
(e) Current allocation for terminal 3. (f) Generated allocation for terminal 3.
i i i
Figure 3.10: Current and generated quay crane usage during the hours of a one-week
i i i
cycle.

• After the cost ratio of forty, the scaled cost curve approximately stays in steady-
state. For these ratios, the relative quay crane costs are that large, that further
increasing them does not significantly affect the costs.

i i i

i i i
 CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 67

3.4.3 Experiment 2
In this experiment we assume that none of the vessel lines is prepared to change their
berthing times in Antwerp, however a change in terminal allocation is allowed by all lines.
We are interested in decreasing the current costs for inter-terminal transport, while the
current berthing times are remained. Additionally, we require that the number of quay
cranes needed is at most equal to the number of quay cranes required for the current
allocation. Hence, we allow a terminal adaptation for each vessel and fix its berthing
interval in time to the current allocation (Gv = 0). Figure 3.11 shows the cumulative
costs for inter-terminal transport for the current allocation and the generated allocation
for each hour in the weekly cycle. The costs are scaled to the total costs in the current
i i i i

i
allocation. These results suggest that, with the same number of quay cranes, about 40%
“IT1˙temp” — 2008/12/1 — 18:52 — page 1 — #1
i i
“IT2˙temp” — 2008/12/1 — 18:52 — page 1 — #1
i

of the costs for inter-terminal transport can be saved.

1 1

0.8 0.8
Cumulative costs

Cumulative costs
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 50 100 150 0 50 100 150
k k

(a) Inter-terminal costs current al- (b) Inter-terminal costs generated

location. allocation.

Figure 3.11: Current and generated cumulative inter-terminal costs for a one-week cycle.

3.5 Conclusions and Recommendations

In this chapter, we considered a port consisting of a number of inter-acting terminals,
i operated by one terminal operator. The strategic problem faced by the operator is about
i i i

allocating vessels to a terminal for a certain time interval. We abstracted from position and
i i i i

quay crane allocation (problems faced at tactical and operational levels, respectively) and
constructed an efficient mixed integer linear program (MILP) to strategically allocate a
terminal and a time window to each of the vessels in the cycle. Moreover, results suggested
that an alternative MILP formulation decreases the CPU time of a straightforward MILP
formulation several orders of magnitude for real-life instances. In fact, real-life problems
can be solved within a couple of hours when the alternative approach is used.
Additionally, a two-step optimization approach was introduced by means of adapting
the MILP. This two-step approach was applied to perform a case study, based on repre-
sentative data provided by a multi-terminal container operator in the port of Antwerp.
The approach enabled us to efficiently investigate the benefit of modifying an existing al-
location, i.e. the potential crane and inter-terminal transport cost savings if the existing
terminal and time allocations were to be adapted. Pareto frontiers were presented to give
insights in the possible reduction of quay cranes at the expense of higher inter-terminal
transportation costs, and vice versa. Results suggest that a small adaptation of an ex-
CHAPTER 3. STRATEGIC ALLOCATION IN A CONTAINER OPERATION 68

isting allocation suffices to gain significant improvements: a reduction of almost 25% of

the number of cranes and at the same time a reduction of more than 3% of the inter-
terminal costs. Furthermore, if the current terminal allocation of all vessels is allowed to
be adapted while the current time allocation is fixed, costs for inter-terminal transport
can be reduced by 40%. Whether such a reallocation is practically and commercially
feasible is beyond the scope of this chapter.
The strategic problem considered in this chapter is a very interesting one from a
practical point of view. A terminal operator that handles a number of terminals in one
port faces the problem of distributing its customers (shipping lines) among these terminals,
such that i) the workload is proportionally balanced over the terminals and over time,
and ii) the amount of inter-terminal transport induced by transshipment containers is
minimized. This chapter proposed a method to solve this strategic problem by abstracting
from actual berth positions and crane scheduling and focusing on terminal and time
allocations. Once this strategic timetable is generated, several subsequent problems are
to be solved. We make the following outline for these problems in the next chapters:
• Usually, a contract between a terminal operator and a shipping line stipulates that
the terminal operator has to process a vessel within a certain time, provided that
the vessel arrives within a certain agreed time window around the scheduled arrival
time. In Chapter 4 of this dissertation, a model is developed that slightly modifies
the timetables, constructed in the current chapter, to increase their robustness to
the within window arrivals.
• In Chapter 5, the robust terminal timetables generated in Chapter 4 are considered
to be a given. The remaining tactical problem is to find appropriate positions for
the vessels at the quay and positions for the containers in the yard such that the
carrier travel distance between quay and yard is minimized.
• The subproblems solved in Chapters 3 through 5 result in a reference timetable
and yard lay out. At an operational level however, the considered processes, e.g.
arrival of vessels and quay crane productivity, are stochastic. Due to these stochastic
properties, the reference allocation has to be continuously adapted. In Chapter 6, a
model predictive control (MPC) is proposed to adapt the allocation under observed
disturbances.
With respect to the research performed in this chapter, the following recommendations
are given:
• In the current chapter, we assumed all vessels to arrive exactly once during the
cycle, implying that all vessels have the same period. However, in practice it can
happen that some vessels have different cycle lengths. An extension of the model,
which incorporates this phenomenon, would be an interesting study.
• In today’s ports, inter-terminal transport is not only established by trucks, but
also by barges. In the current approach we model the resource utilization of the
barges by simply reducing the quay lengths by 200 meters and dedicating one quay
crane in each terminal to barge operations. It is worth investigating the trade-off
between the amount of inter-terminal transport by trucks and barges and therefore
worthwhile extending the current model with the actual loading and unloading of
barges.
Chapter 4

Generation of a Robust Timetable

for a Container Terminal

4.1 Introduction
In Chapter 3 of this dissertation, a timetable is constructed per terminal while assum-
ing deterministic vessel arrivals and hence ignoring disturbances on arrivals. In practice
however, container vessels might arrive respectively earlier or later than their scheduled
arrival time, due to all kinds of events during travel (e.g. tailwind, storms, technical
problems). To cope with these disturbances, the terminal operator and each of the vessel
lines agree on an arrival window, which is placed around the scheduled arrival time. The
arrival window concept distinguishes between two kinds of arrivals: arrivals within and
out of the predetermined window. If a vessel arrives within its window, the terminal
operator has to process this vessel within an agreed process time. If a vessel arrives out
of its window, the terminal operator is not bound to any process time. Nevertheless, he
aims to serve the vessel as soon as possible, but without sacrificing the agreements for
other vessels.
A plan constructed in Chapter 3 might, for particular arrival scenarios within the
windows, i) not yield a feasible operational plan due to lack of quay meters or ii) require
a large amount of crane capacity to fulfill the window agreements. We are interested
in constructing a berth plan, which is robust to all arrival scenarios where vessels arrive
within their windows. In our definition, a berth plan is robust with respect to a given set of
arrival windows if a feasible solution exists for each arrival scenario where all vessels arrive
within their windows. The price for achieving this robustness [3] is then the additional
crane capacity reservation that is required in the worst case arrival scenario where all
vessels arrive within their windows. The problem is hence to construct a window-based
berth plan (denoted as WB-plan for the remainder of this chapter) that minimizes the
maximally required crane capacity for all scenarios where vessels arrive within their arrival
windows.

4.1.1 Related Work

To deal with stochastic disturbances in dense transportation schedules, two (complemen-
tary) approaches are gaining more and more attention as stated in [7]: i) disruption
management, which is concerned with operational recovery after a disruption, and ii)
pro-active robustness, which builds in buffer times and other characteristics into strate-

69
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 70

gic or tactical timetables to absorb disturbances and thus to prevent delay propagation
through a schedule.
So far, most studies on disruption management are conducted for airline operations.
However, over the last years, this approach is gaining more and more attention in railway
applications as well. An overview of disruption management approaches for airline op-
erations and the way these approaches now enter railway applications is given in [7]. In
Chapter 6, a model predictive control approach is applied to online manage the logistics
activities in a container terminal.
A few studies on pro-active robustness in airline scheduling can be found. The authors
in [7] and [1] address a number of robustness ideas. Of particular interest is the approach
of adding slack between connected flights in [29]. Flight schedules are often that tight
that in case of a small plane delay, passengers might miss their connecting flight. Adding
more slack between the flights is beneficial for the passengers but reduces the productivity
of the airline fleet. The authors propose an MILP in which both a flight’s arrival time
and the departure time of its connecting flight(s) can be scheduled somewhere within a
window. Each possible arc between a time slot in the arrival window and a time slot
in the departure window is called a copy. Each copy implies a connecting travel time
and, as determined from historical data, induces a probability of passengers missing their
connected flight (if the travel time exceeds the connecting time). Given a set of flights
within a restricted amount of time, the objective is to select exactly one copy for each
pair of connected flights such that the expected total number of delayed passengers is
minimized.
With respect to pro-active robustness in railway applications, a few approaches can
be found [5], [50]. The authors in [50] consider a stochastic optimization model for the
macro-level for building in time buffers between connecting trips based on arrival and
departure distributions for each train. They propose a model, which allocates a restricted
amount of time supplement to a number of trips to minimize the expected total amount
of delay. Experimental results suggest that applying slight modifications to an existing
timetable can reduce the average passenger delay substantially. The study in [5] presents
embeds robustness into the train timetable by allocating time windows for arrival and
departures rather than single arrival and departure times. The model is stated as a
flexible periodic event scheduling problem, which guarantees that any particular choice
of event times within the computed intervals is feasible. The resulting window-based
timetable is therefore robust to disturbances within these bounds. The total sum of
interval lengths can be traded off against the total sum of travel times in a bi-objective
optimization formulation.
In this chapter on container vessel planning, we also build in robustness by reserving
quay meters and crane capacities to satisfy the agreements in each arrival scenario within
the windows. Our results as well suggest that slight modifications to a representative plan
yield significant reductions in the maximal amount of crane capacity reservation. A major
difference with airline and rail operations however is the following: passengers can enter a
plane or train by themselves, but cranes are required for discharging and loading vessels.
Besides satisfying the window agreements, an additional goal for container operators is
therefore to reduce the maximal amount of crane capacity ever required. Our model
constructs a berth plan that minimizes the maximal crane capacity reservation and still
satisfies the agreements for all arrival scenarios within the windows.
The common berth allocation problem is concerned with allocating container vessels
in space and time. Different solution methods and different objectives are addressed in
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 71

previous studies [8], [23], [24], [22], [17], [19], [27], [32], [38], [41], [51]. However, these
studies do not take arrival uncertainty into consideration while constructing the berth
plan. Although such a berth plan might be optimal with respect to a certain objective,
it can be very sensitive to disturbances on arrivals. A small delay of a single vessel for
instance might be propagated and even amplified through the entire schedule, making it
hard or even impossible to recover on an operational level.
To the best of our knowledge, only one study [36] addresses the stochastic berth
allocation problem by building in pro-active robustness. Arrival times of an existing plan
are given and cannot be changed. The authors derive an expression for the expected delay
for each vessel based on arrival distributions. Given these expected values and a desired
berthing position for each vessel, suitable time buffers and suitable berth positions are
allocated to each vessel. Conflicting objectives are to minimize total expected delay, the
number of overlaps of vessels and the deviations from preferred berth locations. Once a
periodic berth allocation is determined, simulations with stochastic arrivals are performed.
Simulations compare the performance to a model that neglects disturbances. Results
suggest that taking disturbances into consideration yields a reduction in total delay on
the operational level. One of the recommendations of the authors is to incorporate crane
allocations while constructing a robust berth plan.

4.1.2 Contributions
We propose a mixed integer linear program, which explicitly incorporates the within-
window-arrival agreements and minimizes the maximally required crane capacity. Besides
the arrival times of vessels being decision variables, the model also considers time-variant
crane capacity reservations per vessel to be decision variables. The model thus incor-
porates two flexibilities: i) shifting the berth plan of vessels in time and ii) reserving a
time-variant crane capacity for each vessel. These two flexibilities enable to better balance
the workload over time and hence to minimize the maximal crane capacity reservation
ever required.
We arbitrarily select one of the terminals and consider its timetable as provided by PSA
HNN (based on the current policy of satisfying shipping lines’ preferences) as a starting
point for this chapter. Computational results for applying the MILP demonstrate that
with only small modifications to the berth plan of one of the terminals, a significant
reduction in the maximal crane capacity reservation can be obtained. As a comparison
we use the nominal plan, which ignores the arrival window agreements and thus constructs
a berth plan with minimally required crane capacity assuming zero uncertainty on the
arrivals. Note that the same nominal plan is also generated by step two in Section 3.4.1.
Results suggest that both approaches yield significant reductions in the required crane
capacity with respect to the allocation based on policy as currently applied by PSA
HNN. Furthermore, results suggest that although the WB-plan requires slightly larger
crane capacity reservation than the nominal plan for narrow arrival uncertainty bounds,
the WB-plan requires a significantly smaller crane capacity reservation for medium and
wide arrival uncertainty bounds, which are still within the arrival window bounds.
Although it is guaranteed that the terminal length is never exceeded no matter what
arrival scenario occurs, the exact vessel berth position within the terminal is still to be
determined. Similarly, the operational quay crane allocation is still to be determined. In
this chapter, we solve a one-dimensional capacitated packing problem under arrival dis-
turbances within the arrival windows. In Chapter 5, a joint vessel position and container
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 72

stacking problem is addressed. The objective is to minimize the carrier travel distance
between vessels and yard. Experiments suggest that i) the generated berth plans always
yield feasible berth position allocations and ii) significant reductions in the current carrier
travel distance can be achieved.
In Chapter 6, we propose a rolling horizon planning approach to recover from i) all
stochastic arrivals, i.e. arrivals within but also out of the arrival windows, ii) crane break-
downs, and iii) disturbances on vessels’ load compositions. In each iteration step of the
rolling horizon planning process, first berthing times and time-variant crane capacities
are allocated to the vessels within the horizon. Subsequently, based on these allocations,
appropriate berth positions and integer-valued crane allocations are constructed.
In this chapter, we as well aim for embedding robustness into an existing berth plan.
In contrast to the study in [36], the model in this chapter strives for a more sophisticated
crane capacity reservation. Additionally, our model does have the flexibility to modify
the scheduled arrival times and explicitly takes the agreements for arrivals within the
window into consideration. With this WB-plan tool, the arrival times are chosen such
that the maximal crane capacity reservation is minimized while the within-window-arrival
agreements are still satisfied. As a particular case, the model constructs a nominal plan
by simply reducing the arrival window size to zero, which results in a deterministic berth
planning problem (equal to step two in Section 3.4.1). It is interesting to compare i) the
crane capacity required in the WB-plan and in the nominal plan, and ii) the sensitivity
of both plans to different arrival uncertainty bounds, which are still within the window
bounds.
The outline of this chapter is as follows: In Section 2, the problem is formally phrased.
Then, an MILP is proposed to construct a WB-plan with minimally required crane ca-
pacity in the worst case scenario. In Section 3, results of a case study show that with only
small modifications to an existing plan already significant improvements can be achieved.
In a second experiment, the performances of the WB-plan and nominal plan are compared.
We end with conclusions and future work in Section 4.

4.2 Mathematical Model

In this section, an MILP is proposed to construct a WB-plan that minimizes the maximal
crane capacity reservation in the worst of arrival scenarios, where vessels arrive anywhere
within their windows.

4.2.1 Problem description

For all of this chapter the following holds, unless stated differently: v ∈ {1, 2, . . . , V }, the
set of vessels, and k ∈ {1, 2, . . . , K}, the set of discrete time slots. We consider a terminal
with quay length L and a set of V container vessels, where vessel v has length Mv . Each
vessel is assumed to be discharged and loaded at this terminal exactly once a week. We
define Cv to be the total amount of containers that has to be discharged from and loaded
onto vessel v and assume this amount to be the same each week.
Dependent on the length Mv of vessel v, a maximum number Sv of quay cranes can
process vessel v simultaneously. In practice, quay cranes with different processing rates
are present in the terminal. We do not take the specific allocation of quay cranes to vessels
into account yet, but consider an average processing rate λ̄ ∈ N for all quay cranes. The
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 73

Parameter Definition
V Number of vessels in the set
K Number of discrete time slots in the periodic plan
L Terminal quay length [m]
Mv Length of vessel v [m]
Cv Nr. containers to be discharged from and loaded onto vessel v
Sv Maximal nr. quay cranes that can process vessel v simultaneously
λ̄ Mean crane rate [containers/time slot]
ηv Crane efficiency on vessel v
Wv Width of arrival window for vessel v
Pvmin Minimal process time of vessel v
Pvmax Maximal process time of vessel v
αv Fraction between Pvmax and Pvmin

Table 4.1: Model parameters

efficiency ηv ∈ [0, 1] of cranes on vessel v depends on the length Mv of vessel v; the

smaller the length, the lower the efficiency. Then the minimal handling time of vessel v
in terminal t depends on i) the mean processing rate λ̄ in terminal t, ii) the efficiency
ηv of quay cranes operating vessel v, iii) the maximal number of quay cranes that can
proces vessel v simultaneously, and iv) the total number of containers Cv to be discharged
from and loaded onto vessel v. The processing time of vessel v is assumed to be inversely
proportional to the first three of these items and proportional to the latter. The minimal
integer number of time slots Pvmin required to process vessel v can thus be determined as
follows:
 
min Cv
Pv = . (4.1)
ηv Sv λ̄
According to the agreements, the terminal operator has to process a vessel within the
agreed process time only if that vessel arrives within its arrival window. We assume the
width Wv of the arrival window for vessel v to be equal to an integer number of time slots.
In practice, the process time agreed upon by the vessel line of vessel v and the terminal
operator is a factor αv larger than the minimal proces time. This we approximate by

Pvmax = αv Pvmin ,
 
(4.2)
where Pvmax is the maximal number of time slots in which vessel v has to be processed
only if it arrives within its arrival window, and αv ≥ 1. Commonly, the value of αv is
significantly larger than 1, which implies that vessel v not has to be processed with the
maximal number of cranes Sv permanently while berthing. Of course a vessel may also
be processed faster than the Pvmax . Furthermore, we assume Wv + Pvmax ≤ K. The model
parameters are conveniently arranged in Table 5.1. The principle of the within-window-
arrival agreements for vessel v is illustrated in Figure 4.1, where lv and rv represent
respectively the left and right end of an arrival window of width Wv . If vessel v arrives
somewhere within its windows, the terminal operator has to process vessel v no more than
Pvmax time slots later. In Figure 4.1, the upper bound on the departure time is depicted
for nine arrival scenarios.
The problem is to construct a periodic berth plan that is robust to the arrival scenarios
where vessels arrive anywhere within their window. While robustness of this berth plan
 CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 74

Pvmax
9
8
7
6
5
4
3
2
1

1' 2' 3' 4' 5' 6' 7' 8' 9' time
Arrival Window
corresponding upper
bounds on departure time

lv Wv rv

Figure 4.1: Within-arrival-window agreements.

in our definition is the property that the arrival agreements are satisfied for each of the
arrival scenarios within the windows, we aim to achieve this robustness with minimum
cost, that is, the crane capacity reservation required to achieve this in the worst case
arrival scenario should be minimized. In the next subsection, we propose an MILP, which
incorporates both these conditions.

4.2.2 MILP
Decision variables

lv = Left end of arrival window of vessel v

qv (k) = Amount of crane capacity reserved for vessel v during time slot [k, k + 1i

i i
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 75

Auxiliary variables

av = First time slot that quay meters and crane capacity are reserved for vessel v
dv = Last time slot that quay meters and crane capacity are reserved for vessel v

1 if vessel v can possibly berth during time slot [k, k + 1i,
bv (k) =
0 otherwise.

1 if av > dv ,
ev =
0 if av ≤ dv .
rv = Right end of arrival window of vessel v.

1 if time slot [k, k + 1i lies within the arrival window of vessel v,
wv (k) =
0 otherwise.

w 1 if lv > rv ,
ev =
0 if lv ≤ rv .
mv (k) = Amount of quay meters reserved for vessel v during time slot [k, k + 1i.
Q = At least the amount of crane capacity required in the worst case scenario.

Constraints and objective

For vessel v the arrival window has to be positioned in time by determining the left end
time position of the window lv . From lv the values for rv and wv (k) follow accordingly.
First of all, we have to enforce that the values of the binary variable wv (k) are equal to
one if and only if time slot [k, k + 1i lies between the left end right end of the window
(see also [19]):
1 − lv ≤ k · wv (k) − ew


v ≤ rv − 1 ∀v, k, (4.3)
rv − K ≤ K − k · wv (k) − ew



v ≤ K − lv ∀v, k, (4.4)
K
X
wv (k) − ew


v = rv − lv ∀v. (4.5)
k=1

Additionally, the width of the arrival window for vessel v is fixed to Wv . Considering the
discrete time model, this implies that there have to be Wv + 1 time slots within the arrival
window of vessel v:
XK
wv (k) = Wv + 1 ∀v. (4.6)
k=1

For each vessel v, the earliest possible arrival and latest possible departure time (av and
dv , respectively) have to be decided on. Only in between its earliest possible arrival time
and its latest possible departure time, a vessel might be berthing. Before its earliest
possible arrival time and after its latest possible departure time, a vessel cannot berth at
all. This can be formulated in a similar way as has been done for positioning the arrival
window in (4.3), (4.4) and (4.5):


1 − av ≤ k · bv (k) − ev ≤ dv − 1 ∀v, k, (4.7)



dv − K ≤ K − k · bv (k) − ev ≤ K − av ∀v, k, (4.8)
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 76

K
X

bv (k) − ev = dv − av ∀v. (4.9)
k=1

If time slot [k, k + 1i is a possible berthing time slot of vessel v, Mv quay meters have to
be reserved during that time slot:

mv (k) = Mv · bv (k) ∀v, k. (4.10)

The sum of lengths of all vessels possibly berthing during time slot [k, k + 1i should never
exceed the terminal length L:
V
X
mv (k) ≤ L ∀k. (4.11)
v=1

If time slot [k, k +1i is reserved for vessel v, the crane capacity reserved for vessel v during
that time slot is limited by the number of cranes that can process vessel v simultaneously:

qv (k) ≤ Sv · bv (k) ∀v, k. (4.12)

We now have to enforce that the window agreements for vessel v are satisfied. The
agreements state that if vessel v arrives within its window, its process time has to be
within the agreed process time Pvmax . Hence, for each range of sequential time slots that
starts from a time slot within the arrival window of vessel v and ends Pvmax − 1 time
slots later, the sum of reserved crane capacities should be sufficient to process at least Cv
containers. Since the position of the window of vessel v is a decision variable on itself,
we have to explicitly consider the sum of crane reservations for each possible range of
sequential time slots of length Pvmax within the considered cycle. Only if the first time slot
of such a range lies within the arrival window of vessel v, sufficient crane reservations for
vessel v during these time slots are required to process at least Cv containers. To model
this we make use of the value of the binary value wv (k):
k+Pvmax −1
X
ηv λ̄ · qv (i) ≥ Cv · wv (k) ∀v, k. (4.13)
i=k

The sum of reserved crane capacities of all vessels during time slot [k, k + 1i should never
exceed the maximally required crane capacity reservation:
V
X
qv (k) ≤ Q ∀k. (4.14)
v=1

The objective is to minimize the maximal crane capacity reservation:

min Q (4.15)
lv ,qv (k)

4.3 Case study

The MILP proposed in the previous section determines a WB-plan with minimal crane
capacity reservations in the worst case arrival scenario while within-arrival-window agree-
ments are still met. The model thus incorporates the arrival window agreements and
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 77

hence finds a plan robust to these agreements. In this section, we perform two experi-
ments on the representative berth plan provided by PSA HNN. We consider one terminal
and 15 vessels (V = 15), which call exactly once each week. A time slot width of one
hour is chosen, so K = 168. We set the arrival window width of vessel v to eight hours,
so Wv = 8, ∀v, which is a typical value in real-life container operations. This implies
that if a vessel arrives up to four hours earlier or four hours later than its planned arrival
time, the terminal operator still has to process the vessel within the agreed process time.
Typical values for the other parameters in Table 5.1 are provided by PSA HNN as well.
Furthermore, we assume a typical value of αv = 1.4, ∀v.
For this data set, two experiments are performed. In the first experiment, the WB-plan
MILP is applied to the provided berth plan for different values of maximal modification.
Results show that with small modifications, significant reductions in the maximal crane
capacity reservation can already be achieved. In the second experiment, the performances
of the WB-plan and the nominal plan are compared. The nominal plan can be constructed
as a particular case by simply setting the arrival windows width to zero (Wv = 0, ∀v) in the
MILP (or by running step two in Section3.4.1). Optimizing the MILP then results in an
optimal periodic plan with a minimal crane capacity reservation. The performance of both
the WB-plan and the nominal plan are compared for different arrival uncertainty bounds
within the arrival windows. Results suggest that although the WB-plan requires slightly
more crane capacity reservations for narrow arrival uncertainty bounds, it significantly
outperforms the nominal plan for medium and wide arrival uncertainty bounds.

4.3.1 Benefit of plan modification

As mentioned before, vessels have fixed routes and a preferred arrival time in each port
they call on. Negotiations have to point out whether vessel lines are willing to slightly
modify their scheduled arrival times. We thus aim for a reduction in crane capacity
reservation if we assume relatively small modifications to the existing plan can be made.
To obtain some more insight into the improvements that can be made dependent on
the extent of modification, a sequence of 4 experiments is performed. In experiment i,
i ∈ {1, 2, 3, 4}, a vessel subset Vi is selected from the representative data set of PSA HNN,
where |V1 | = 2, |V2 | = 4, |V3 | = 7, |V4 | = V = 15, and Vi ⊂ Vi+1 , i ∈ {1, 2, 3}. For each
of the vessels in the subset Vi of experiment i, we allow a maximal modification of Gv
time slots with respect to the existing plan, by introducing appropriate upper and lower
bounds on the arrival window position lv . In each experiment i ∈ {1, 2, 3, 4}, the WB-plan
MILP with Wv = 8, ∀v is determined consecutively for Gv ∈ {0, 1, 2, ..., 8}, v ∈ Vi .
Results are depicted in Figure 4.2. The (scaled) maximal crane capacity reservation
is plotted versus the maximally allowed plan modification Gv for each v in the selected
vessel subset. Each curve in this plot depicts the outcome of one experiment i. In this
figure, we notice the following:

• Each of the four curves is monotonically decreasing. This makes sense, since in-
creasing the extent of maximal modification will never yield a higher amount of
maximally required crane capacity.

• Along the same line, we can explain that the curve of experiment i + 1 never exceeds
the curve of experiment i. Namely, if the plan of more vessels is allowed to be
modified, the maximal crane capacity reservation will never increase.
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 78

• One interesting result is that with allowing the plan of seven vessels to be modified,
the same improvements can be achieved as by allowing the plan of all fifteen vessels
to be modified.

• iAnother interesting result is that by allowing the plan of only four vessels to be
modified, at least 95% of the improvements can —
“tempimage˙temp” already be obtained
2008/9/3 — 10:25 as
— by allowing
page 1 — #1
i
the plan of all fifteen vessels to be modified.
replacemen
• By modifying the plan of four out of fifteen vessels maximally 2 hours, a reduction
of about 5% in the maximal crane capacity reservation can be achieved.

1
Maximal crane capacity reservation

0.95

0.9

0.85 2 vessels modifiable

4 vessels modifiable
7 vessels modifiable
15 vessels modifiable
0.8
0 2 4 6 8
Maximal plan modification Gv [hrs]

Figure 4.2: Benefit of modification.

4.3.2 WB-plan vs. nominal plan

The MILP as proposed in the previous section explicitly includes the within-arrival-
window agreements for Wv = 8. Another approach to construct a tactical berth plan
is to simply ignore the arrival window agreements Wv = 0 and to determine the optimal
deterministic berth plan. Such a nominal plan is constructed by simply reducing the
window width Wv to zero ∀v and running the MILP.
We are interested in the performance of the WB-plan and the nominal plan for bounded
arrivals, i.e. we assume all actual operational arrivals to be within certain bounds, which
are still within the arrival window bounds. We define the time distance between the
arrival uncertainty bounds to be Uv , where 0 ≤ Uv ≤ Wv , ∀v.
Subset V2 of the previous experiment is chosen since for this instance |V2 | = 4 is a
reasonable guess for the number of vessels to be modifiable. The following procedures are
proposed to evaluate the performance of the WB-plan and nominal plan for a particular
value of Uv :

Procedure to evaluate WB-plan performance:

1. Determine a WB-plan by optimizing the MILP with a window width Wv = 8,
 CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 79

2. Record the optimal objective value QW 8

B
,
3. Substitute the optimal values of the left end of the arrival window lv∗ into the lv vari-
ables in the MILP and fix them,
4. Run the MILP for a window width Wv = Uv , ∀v,
5. Record the optimal objective value QW B
Uv .

i
Procedure to evaluate nominal plan performance:
1. Determine a nominal plan by optimizing the MILP with a window width Wv = 0,
“tempimage˙temp” — 2008/9/4 — 10:28 — page 1 — #1
i2. Record the optimal objective value QW I
0 ,
3. Substitute the optimal values of the left end of the arrival window lv∗ into the lv vari-
ables in the MILP and fix them,
4. Run the MILP for a window width Wv = 8, ∀v,
5. Record the optimal objective value QW B
Uv .

1
Maximal crane capacity reservation

Area WI-plan
Area WB-plan
0.9 QW I
8

0.8
QW
8
B

0.7

QW
0
B

0.6

0 5 10 QW
0
I
15
Maximal plan modification Gv [hrs]

Figure 4.3: Window-Based plan (Wv = 8, ∀v) vs. Window-Ignoring plan (Wv = 0, ∀v).

For each of the values of Gv ∈ {0, 1, 2, ..., 16}, and v ∈ V2 , these procedures are applied
to determine the performance of the WB-plan and nominal plan for Uv = 0, ∀v and Uv =
Wv , ∀v. Figure 4.3 depicts QW 0
B
, QW
8
B
, QW I
0 , Q0
WI
as a function of Gv ∈ {0, 1, 2, ..., 16}
v ∈ V2 . In this figure we can notice the following:
• The grey and shaded area represent the limits of maximal crane capacity reservations
for respectively the nominal plan and the WB-plan for values of 0 ≤ Uv ≤ 8. It can
be noticed that the shaded area in total lies within the grey area. Apparently, the
nominal plan outperforms the WB-plan if zero disturbances on arrivals are present
(Uv = 0), but is much more sensitive for wide arrival uncertainty bounds (Uv = 8).
• The curves for QW 0
I
and QW0
B
are monotonically decreasing. This is to be expected,
since these lines result from the first step optimizations (step 1 in both procedures).
Hence, if the maximal plan modification increases, the maximal crane capacity reser-
vation will never increase.
• The curves for QW0
I
and QW0
B
however, are not monotonically decreasing. From
the procedures we notice that these curves are determined in a second optimization
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 80

(step 4) in which the values from the first optimization are already fixed. Regarding
this the non-monotonic decrease makes sense, since the decisions made in the first
optimization (step 1) do not necessarily have to be optimal in the second optimiza-
tion (step 4), and might result in a higher crane capacity in the second optimization
even when the maximal plan modification increases.
• The curve for QW
0
I
never exceeds the curve for QW I
8 . This makes sense, since a berth
plan incorporating bounded arrivals will always require at least the same amount
of crane capacity as the same berth plan with deterministic arrivals. The same
reasoning can be applied for the observation that QW0
B
never exceeds QW8
B
.
• The curve for QW 0
I
never exceeds the curve for QW0
B
. This is to be expected since
Q0 is determined in a first optimization (step 1) and QW
WI
0
B
in the second optimiza-
WI
tion (step 4). Q0 ignores the arrival window agreements and hence is the optimal
plan if no arrival disturbances are present (Uv = 0). QW 0
B
is constructed from a
WB-plan, which incorporates the window agreements. This may lead to a berth
plan, which is not optimal for deterministic arrivals.
• The curve for QW 8
B
never exceeds the curve for QW I
8 . This is to be expected since
QW8
B
is determined in a first optimization (step 1) and QW 8
I
in the second opti-
WB
mization (step 4). Q8 incorporates the arrival window agreements and hence is
a robust plan if arrival disturbances are present on an operational level. QW 8
I
is
constructed from a nominal plan, which ignores the window agreements during op-
timization. Ignoring these agreements may lead to a plan which is not robust to
arrival disturbances.
• The initial values of QW0
I
and QW0
B
are equal. This is to be expected since in this
case Gv = 0 ∀v and hence no modifications to the berth plan can be made (lv
is fixed). An intelligent time-variant crane capacity allocation for the given plan
results in the same amount of maximal crane capacity reservation when arrivals are
assumed to be deterministic. The same reasoning can be applied for the observation
that the initial values of QW8
I
and QW8
B
are equal.
• The initial values of QW0
I
and QW0
B
on one hand, and QW 8
I
and QW 8
B
on the other
are not (necessarily) equal, although they are all based on the same berth plan (lv is
fixed to the same value in each of these cases). Still, this observation makes sense,
since in the former two cases (QW 0
I
and QW0
B
) deterministic arrivals are assumed
WI WB
and in the latter two cases (Q8 and Q8 ) arrival uncertainty bounds (equal to
the window bounds) are assumed. Arrival bounds will result in at least the same
amount of crane capacity reservation than in the case with deterministic arrivals.
As mentioned before these curves depict only the limits on the performance for all
values where 0 ≤ Uv ≤ 8. We are interested in the dependency of the nominal plan and
the WB-plan performance on specific values of Uv within this range. Hence, we applied
additional experiments to evaluate the performance of the nominal plan and the WB-
plan for Uv ∈ {1, 2, 3, 4, 5, 6, 7}, ∀v. Results for Uv ∈ {0, 1, 2, ..., 8} ∀v are depicted in
Figure 4.4. In this figure, we notice the following:
• Figures 4.4a and 4.4i again present respectively the lower and upper limits on the
crane capacity reservations for stochastic arrival uncertainty bounds within the win-
dow bounds (see also Figure 4.3). As Uv increases from zero to 8, both the curves
 “tempimage˙temp” — 2008/9/2 — 12:56 — page 1“tempimage˙temp”
— #1 — 2008/9/2 — 12:48 — page 1“tempimage˙temp”
— #1 — 2008/9/2 — 12:48 — page 1 — #1
i i i i i

CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 81

1 1 1
WI
WB
0.9 0.9 0.9

0.8 0.8 0.8

0.7 0.7 0.7

i i i i i

“tempimage˙temp” — 2008/9/2 — 12:49 — page 1“tempimage˙temp”

— #1 — 2008/9/2 — 12:49 — page 1“tempimage˙temp”
— #1 — 2008/9/2 — 12:49 — page 1 — #1
i 0.6 i 0.6 i 0.6 i i

0 5 10 15 0 5 10 15 0 5 10 15

(a) Uv = 0. (b) Uv = 1. (c) Uv = 2.

1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

0.7 0.7 0.7

i i i i i

“tempimage˙temp” — 2008/9/2 — 12:49 — page 1“tempimage˙temp”

— #1 — 2008/9/2 — 12:50 — page 1“tempimage˙temp”
— #1 — 2008/9/2 — 12:50 — page 1 — #1
i 0.6 i 0.6 i 0.6 i i

0 5 10 15 0 5 10 15 0 5 10 15

(d) Uv = 3. (e) Uv = 4. (f) Uv = 5.

1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

0.7 0.7 0.7

0.6 0.6 0.6

0 5 10 15 0 5 10 15 0 5 10 15

(g) Uv = 6. (h) Uv = 7. (i) Uv = 8.

i i i i i

i
Figure 4.4: Performance
i
WB and nominal plan fori values of Ui v . Horizontal axes depict i
the maximal plan modification Gv , v ∈ V2 , vertical axes the maximal crane capacity
reservation.

for the nominal plan and the WB-plan shift upwards. This makes sense, since the
wider the bounds on the arrivals grow, the more crane capacity reservations are
required.
i i i i i
• For Uv = 0, the WB-plan requires a slightly higher amount of crane capacity than
i i i i i
the nominal plan. As Uv increases from zero to two, the difference between the
performance decreases towards zero.

• As Uv increases from three to eight, the performance of the WB-plan with respect
to the nominal plan increases significantly.

• Apparently, the WB-plan based on a window size of Wv = 8, is more robust to

medium and wide arrival uncertainty bounds (Uv ≥ 3) than the nominal plan.
i i i i i

i i i i i
CHAPTER 4. GENERATION OF A ROBUST TIMETABLE 82

4.4 Conclusions and Recommendations

We considered a set of container vessels that has to be discharged and loaded in a container
port by a terminal operator on a periodic basis. Disturbances on travel times lead to
stochastic arrivals in the port. To cope with these disturbances, the terminal operator
agrees on a so-called arrival window for each vessel rather than a single arrival time. Only
if a vessel arrives within its window, the terminal operator has to process this vessel within
an agreed process time. If a vessel arrives out of its window, the terminal operator is not
bound to any process time.
We proposed an MILP to construct a window-based periodic berth plan (WB-plan)
with minimally required crane capacity in the worst case arrival scenario, i.e. an MILP
that minimizes the required crane capacity while the agreements for all scenarios where
vessels arrive within their windows are still satisfied. Experiments on a representative
berth plan provided by the terminal operator PSA HNN in Antwerp, Belgium, suggested
that with small modifications to the representative plan, already significant reductions in
the maximally required crane capacity reservation can be achieved.
As a particular case, our MILP approach constructs a window-ignoring periodic berth
plan (nominal plan) by setting the arrival window width to zero. This nominal plan is
similar to the one generated in the previous chapter. We investigated the performance
of the WB-plan and the nominal plan for deterministic arrivals and bounded arrivals
within the arrival window bounds. Results suggested that although the WB-plan requires
slightly more crane capacity than the nominal plan for narrow arrival uncertainty bounds,
it is much more sensitive to medium and wide arrival uncertainty bounds, which are still
within the windows.
In the first optimization step (see Figure 1.5 and Chapter 3), a nominal timetable is
constructed for a multi-terminal container operation. In this chapter, a method is pro-
posed that takes an arbitrary (e.g the nominal) berth plan as a starting point and makes
it more robust to the within window arrival agreements. In Chapter 5, the timetables are
assumed to be a given and accordingly, berth positions for the vessels and yard positions
for the containers have to be allocated. The objective is to minimize the total carrier
travel distance between vessels and yard. Finally in Chapter 6, the generated berth plan
and yard layout are considered to be a reference planning for the considered cycle. A
model predictive control approach is applied to act upon all kinds of disturbances in the
daily operations.

To better quantify the robustness of a timetable, actual arrival distributions of the vessel
lines are required. Having these, the expected amount of required crane capacity can be
computed. Moreover, with these distributions we can compute the percentage of time
during which the maximum capacity is required.
Another recommendation emerges from the observation that the proposed MILP de-
termines an upper bound on the maximally required crane capacity. Namely, it does not
explicitly evaluate each possible arrival scenario, but instead reserves sufficient crane ca-
pacity such that the window agreements are always met. This approximation significantly
reduces the number of scenarios to be evaluated. A recent study investigates for small
instances the deviation between our determined upper bound and the optimal value for
the maximally required crane capacity.
Chapter 5

Joint Berth Allocation and Yard

Design Problem

5.1 Introduction
In Chapter 3, a multi-terminal container port, run by one terminal operator, is considered
and terminals and berth times are allocated to a set of vessels assuming deterministic
parameters. Subsequently in Chapter 4, the constructed timetable of a terminal is slightly
modified to increase its robustness to stochastic arrivals. In this chapter, we consider the
(robust) timetable as given and accordingly determine appropriate berth positions for the
vessels at the quay.
After unloading and before loading, containers are temporarily stacked in capacitated
areas in the storage yard of a terminal. Several types of containers have designated stacks
in the yard, while others can be stacked wherever space is available. The allocation
of these containers in the stack is defined as the yard design problem. To guarantee
consistency for the carrier drivers and yard planners, the allocation of the containers from
a certain shipping line to a specific stack should not change too frequently. Decisions on
the reference berth positions and yard design are therefore only reconsidered once each
year or once each two years.
The decisions on the berth positions and the yard design determine the total distance
that has to be covered by carriers operating between quay and yard. In this chapter, we
address the joint problem of allocating i) vessels’ berth positions, taking into account no
overlapping of vessels, and ii) container stack positions, taking into account non-exceeding
of stack capacities, such that the total carrier distance is minimized.

5.1.1 Related Work

As mentioned in Section 3.1, the berth allocation problem (BAP) has been investigated
extensively over the last decades. The problem involves the allocation of container vessels
in time and space in order to minimize a certain objective function. In most of the
reported studies, the objective is to minimize each vessel’s turnaround time [15], [35],
[24], [8], [27]. A limited number of studies, considers a multi-objective problem, where
besides the turnaround times, the weighted deviations from predetermined berth positions
are minimized [16], [51]. The authors in these studies consider reference berth positions,
which are chosen closest to given positions of containers in the yard. In this way, the

83
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 84

distance that has to be covered by container carriers between quay and yard is tried to
be reduced.
Of particular interest is the study in [36]. The authors mention that the total travel
distance strongly depends on the stack positions of containers in the yard. They assume
that all containers for/from a certain vessel are stacked closest to the position where the
particular vessel berths and approximate the total travel distance by the travel distances
for transshipment containers between connecting vessels. In practice however, two main
reasons exists to contradict that all containers can be stacked close to a vessel’s berth
position. First of all, special container types (like reefers, IMCO’s and empty containers)
have designated areas in the yard and hence cannot be stacked arbitrarily. Second of all,
once a stack is filled up to its capacity, containers have to be allocated to surrounding
stacks, unavoidably inducing additional travel distances for carriers. In this chapter, we
take both these issues into account while solving the joint problem of allocating berth
positions to vessels and stack positions to containers to minimize the total carrier travel
distance.
The above mentioned studies all consider the allocation of time and space to vessels
within one terminal. Actually, in an increasing number of ports (e.g. Antwerp, Singapore
and Rotterdam), one terminal operator is responsible for a number of terminals. Along the
same line, we consider a multi-terminal container operation run by the terminal operator
PSA HNN in Antwerp, Belgium. As mentioned in Section 3.1, the problem then becomes
to i) allocate a vessel to a terminal (actually the loop to which a number of vessels
belong), ii) allocate a time interval to a vessel for berthing, iii) allocate a vessel to a
berthing position, and iv) allocate quay cranes to a vessel, taking into account the cyclic
nature of vessel calls.
The BAP in such ports can no longer be considered per terminal for two main reasons.
One is that it makes sense to avoid peaks and troughs in quay crane utilization and to
spread vessel calls evenly over the various terminals. The other is that transshipment
containers will unavoidably generate inter-terminal traffic, whose costs should be taken
into account. All possible flows of containers through such a port are depicted in Figure
1.6. Since the overall problem is too complex to be solved within satisfactory time, it is
cut into several subproblems. These subproblems are chosen such that they are practically
interesting and at the same time can be solved within the time allowed at the considered
level. We have chosen to cut the overall container operation problem into four subproblems
(see Figure 1.5).
The first is a strategic problem (Chapter 3), i.e. which terminal and which time win-
dow to allocate to each (vessel in the) loop. This subproblem is addressed occasionally
when contracts for new loops are negotiated or existing contracts are renegotiated. The
algorithm to solve this problem optimally may run for several hours, if not days. Ad-
ditionally, as a second problem, the robustness of the generated timetable per terminal
is increased by taking within arrival window agreements into account. The problem of
adding robustness to the timetable is addressed in Chapter 4 and can be solved within
minutes.
Given these (robust) timetables per terminal, the problem left on the tactical level is
where to berth the vessels along the quay. We have to emphasize that this problem is
typically different from the BAP studies mentioned above for two reasons. One is that the
berth interval at this level is no longer a variable of the problem, it is a given. Secondly,
the objective is not to minimize the make-span of each vessel, but to optimally use the
time allowed to service the vessel. In this chapter, we consider the constructed (robust)
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 85

timetables to be given and address the tactical problem of where to berth vessels taking
into account their lengths, and where to stack their containers taking into account the
various types of containers and the storage capacity of each stack. While doing so we try
to minimize the horizontal carrier travel distance between quay and yard by taking into
account fixed yard positions for some special container types, and by taking into account
the transshipment matrix of the vessels, i.e. which vessels carry how many containers
that are to be loaded onto connecting vessels.
The generated tactical timetable, berth positions and yard design derived in Chapters 3
through 5, are considered to be the reference allocation for a longer period of time (a year
or two). In the daily operation however, the system is disturbed (e.g. due to storm,
tailwind, technical problems) and the terminal operator faces the problem of rescheduling
the disrupted system to serve the shipping lines at minimal costs. A model predictive
control approach is developed in Chapter 6 to on-the-fly allocate actual i) start and end
times for processing, ii) berth positions and iii) quay cranes to vessels in a container
terminal under disturbance.

5.1.2 Contributions
As has been mentioned, the timetables per terminal are considered to be a given. Further-
more, we consider four types of containers (reefer, IMCO, empty and full). Additionally,
for each call we assume the expected number of each of these types to be given (both for
import and export). Usually, a terminal operator uses designated areas for reefer, ICMO,
and empty containers, which we assume to be given as well. The stack positions for the
full containers are still to be chosen and considered as variables in this problem.
Hence, we address the joint problem of allocating berth positions to vessels at the quay
and stack positions of vessels’ containers in the yard to minimize the total carrier travel
distance. While doing so we ensure that vessels that are berthing simultaneously do not
overlap and we ensure that for each moment in time stacks are never filled up above their
capacities.
We propose an appropriate mixed integer quadratic program, which turns out to be
non-convex and consequently complex from a computational point of view. The con-
straints however are convex and separable in the two decision variables being i) the ves-
sels’ berth positions and ii) the amount of containers flowing from a particular vessel to
a particular stack. This enables to construct two individual problems, an MILP and an
LP, being i) the vessel position allocation problem and ii) the container stack allocation
problem respectively, which are coupled in the objective function. A solution technique,
that continues alternating between both problems, appears to converge to a local opti-
mum very fast (the alternating procedure between two separable LP’s, coupled in the
objective, is proven to converge to a local minimum in [4]). The performance of the
alternating procedure however turns out to heavily depend on the initial condition. To
find a proper initial guess, we propose an MILP, which is an approximation of the earlier
proposed MIQP, since it allocates fixed-sized groups of containers to stack, rather than
variably-sized groups. Using this initial condition leads to a solution that outperforms
results from random initial conditions.
A case study on a representative data set, provided by the terminal operator PSA
HNN, learns that the proposed method is very efficient and yields a reduction of more than
20% in the total carrier travel distance compared to the distance for the provided berth
plan and yard lay out. In Section 5.2, the problem is formally phrased and the parameter
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 86

set is given. Additionally, the solution approach is discussed. Section 5.3 presents the case
study and shows results. We end with conclusions and recommendations in Section 5.4.

5.2 Approach
5.2.1 Problem Description
The relevant quay and yard operations for this problem are depicted in Figure 5.2, which
is a suitable representation of an actual container terminal in the port of Antwerp. The
terminal has (quay) length Q and (yard) width B. The yard is divided into N stacks, in
the figure N = 24 (8x3), where stack n has a capacity of Cn , n ∈ {1, ..., N } containers.
The position of stack n is defined by the (x, y) coordinates of its center (Xn , Yn ), where the
origin O is positioned at the lower left end of the terminal. A number of V vessels berths
at the quay periodically, where vessel v has length Lv and can be positioned anywhere
along the quay. Since the arrival and departure times of the vessels are given, the pairs of
vessels that berth simultaneously can be determined in advance. The vessel pair (i, j) is
added to the mathematical set S if vessel i and vessel j berth simultaneously according
to the given timetable. As an example, Figure 5.1 depicts the set S for an arbitrary
timetable of an arbitrary vessel set.
K

4
1
Time

2
3

5
6
7

Space Q

Figure 5.1: Set of pairs S={(1,2),(1,3),(1,4),(2,3),(2,4),(3,4),(4,7),(5,6),(6,7)} for given

time allocation. Note that vessel 4 arrives at the end of a cycle and departs at the
beginning of the next cycle.

We consider T container types to be transported into and out of the terminal area.
Each vessel imports a number of containers of each type for destinations {0, ..., V }, where
destination 0 represents the ”hinterland” and destination 1, ..., V represents vessel 1, ..., V ,
respectively. Additionally, containers of different types are brought in from the hinterland
with export destinations 1, ..., V . We assume proforma load compositions and destination
matrices of the weekly calls to be known. The given timetable then learns at what time,
which type of containers with a specific destination arrive in and leave from the port.
We assume that vessels are totally unloaded first, before the loading starts. Moreover,
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 87

the unloading and loading fractions of the given berthing time are divided proportionally
to the fractions of import and export containers, where the (un)loading of different con-
tainer types is distributed uniformly among the total (un)loading time. Additionally, we
assume all import containers to be transported out of the yard eight hours after they are
stacked and all export containers to be present in the yard eight hours in advance of the
corresponding vessel’s arrival.
Having these assumptions and considering discrete time k ∈ {1, ..., K} (where K is
the cycle length), the container flows into and out of the terminal as a function of time are
completely determined and captured in the following parameters: Iijt (k) is the amount
of containers of type t flowing into the yard from the hinterland i = 0 or from vessel
i ∈ {1, ..., V }, with final destination j, j ∈ {0, ..., V } during time slot [k, k + 1i. Ovt (k) is
the amount of containers of type t flowing to the hinterland v = 0 or to vessel v ∈ {1, ..., V }
during time slot [k, k + 1i. The parameters are arranged in Table 5.1.

Legend
Group of Containers
Straddle Carrier
Quay Crane
Vessel
Q

O x
1 2 3

L1 L2 L3

Figure 5.2: Straddle carriers operating between vessels and yard, which contains N=24
stacks.

We assume that each container is stacked somewhere in the yard in between the time
it arrives in the port and the time it leaves the port again. The import and export
container transport between the vessels/hinterland and the yard is performed by straddle
carriers. For the distance between a vessel/hinterland and a stack we assume the principle
of Manhattan distance, i.e. the distance between the vessel and the stack measured along
the x and y axes. The problem is to find i) berth positions for the vessels taking into
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 88

account non-overlapping, and ii) stack positions for the containers taking into account
the stack capacities, such that the total carrier travel distance is minimized.

Parameter Definition
V Number of vessels in the set
S Set of pairs of vessels that berth simultaneously
N Number of stacks in the yard
Lv Length of vessel v [m]
K Number of discrete time slots in the cycle
Q Terminal length [m]
B Terminal width [m]
Xn x-position of stack n
Yn y-position of stack n
Cn Capacity of stack n
Iijt (k) # containers with source i and destination j flowing into the yard at time [k, k + 1i
Ovt (k) # containers flowing from the yard to destination v at time [k, k + 1i

Table 5.1: Model parameters

In the next section, we propose a solution technique to find a suboptimal carrier

travel distance, taking into account non-overlapping of vessels and non-exceeding of stack
capacities. First in Section 5.2.2, a mixed integer quadratic program (MIQP) is proposed.
Since the corresponding objective is non-convex, the problem becomes complex from a
computational point of view.
Next in Section 5.2.3, we split the MIQP into an MILP and an LP, which are coupled
in the objective and can both be solved very fast. The problems are executed in an alter-
nating fashion, i.e. one continues solving the problems subsequently until the objective
has converged. This procedure turns out to be very fast, however experiments learn that
the performance heavily depends on the chosen initial condition.
To select an initial condition in a sophisticated way (rather than randomly), an ad-
ditional MILP is proposed in Section 5.2.4. Besides vessels’ berth positions, this MILP
allocates fix-sixed groups of containers to stacks rather than variably-sized groups, and
hence generates a rough yard layout. Applying the alternating procedure suggests that
starting from this initial condition, the carrier travel distance can be decreased below the
best result out of an extensive number of random instances. Moreover, it reduces the
total carrier travel distance in the provided allocation by more than 20%.

5.2.2 MIQP
Decision variables

pv= Position of the center of vessel v.

t
iijn (k)= # containers of type t with origin i and destination j flowing to stack n during.
time slot [k, k + 1i.
t
ovn (k) = # containers of type t flowing from stack n to destination v during time slot [k, k + 1i.
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 89

Auxiliary variables

1 if vessel i is positioned left of vessel j,
eij =
0 if vessel i is positioned right of vessel j.
zvn = Manhattan distance between the center of vessel v and stack n.
t
svn (k) = # containers of type t with destination v in stack n during time slot [k, k + 1i.

The objective is to find vessel positions at the quay and container positions in the
stack to minimize the total carrier travel distance. The carrier travel distance of a certain
amount of containers between their vessel and their stack is equal to the product of i)
the amount of containers and ii) the distance between the corresponding vessel and the
designated stack. Hence, we have to minimize the sum over all products of container
amounts and their related distances:
V X N X T X K V
X X
t t


t
min i vjn (k) + ovn (k) · zvn (5.1)
pv ,ivjn (k),otvn (k)
v=0 n=1 t=1 k=1 j=0

With respect to the vessel positions, the following constraints are valid:
Lv
pv ≥ 2
∀v
Lv
pv ≤ Q− 2
∀v
Li +Lj
pi − pj ≥ 2
− Q · eij ∀(i, j) ∈ S
Li +Lj
pj − pi ≥ 2
− Q · (1 − eij ) ∀(i, j) ∈ S (5.2)
zvn ≥ pv − Xn + Yn ∀n, v ∈ {1, ..., V }
zvn ≥ −pv + Xn + Yn ∀n, v ∈ {1, ..., V }
zvn = B − Yn ∀n, v = 0
The first two constraints ensure that vessel v is totally positioned within the quay. The
third and fourth constraints ensure the non-overlapping of vessels that berth simultane-
ously. The fifth and sixth constraints define the minimal Manhattan distance between
vessel v and stack n. The seventh constraint defines the minimal distance between the
hinterland and stack n. Note that this distance only depends on the y coordinate of stack
n, which implies that containers to and from the hinterland are respectively delivered
and collected at the x coordinate of their designated stack. Note that the former four
constraints can be obtained after substituting the, for this problem known, values of δijt ,
i, j ∈ {1, ..., V } and δijp , i, j ∈ {1, ..., V }, into the BAP formulation of [24].
With respect to the container stacking problem, the following constraints are valid:
N
itijn (k) = Iijt (k)
P
∀t, i, j, k
n=1
N
otvn (k) = Ovt (k)
P
∀t, v, k
n=1
V (5.3)
stvn (k stvn (k) itivn (k) otvn (k)
P
+ 1) = + − ∀t, v, n, k
i=0
T P
V T V PV
stvn (k) ≤ Cn − itivn (k)
P P P
∀n, k
t=1 v=0 t=1 v=0 i=0

The first two constraints depict conservation laws, i.e. the total number of containers
of type t with origin i and destination j flowing into the yard during time slot [k, k + 1i
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 90

should be distributed among stacks 1, ..., N , and the number of containers of type t flowing
from the (different stacks in the) yard to destination v during time slot [k, k + 1i should
sum up to the total number of containers of that type with destination v during that time
slot. Note that the special types of reefers, IMCO’s and empty containers (t = 2, 3, 43,
respectively) have designated stacks and hence for these containers itijn (k) and otvn (k) are
given parameters. So, the variables left in the container stacking problem are itijn (k) and
otvn (k) for t = 1 (full containers).

5.2.3 Alternating Optimization

The drawback of the described MIQP is that the objective function is non-convex, which
makes the high-dimensional problem hard to solve. However, the constraint set concern-
ing the vessel position allocation and the constraint set concerning the container stack
allocation are convex and separable in the variables of berth positions and container stack
positions, as the distinction between equations set (5.2) and (5.3) already shows. This
enables to construct an MILP and an LP, which are coupled in the objective function
and can be solved alternatingly. The alternating optimization procedure is depicted and
explained below.
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 91

V P
N P
T P
K V

itvjn (k)? + otvn (k)?
P P
min · zvn
pv v=0 n=1 t=1 k=1 j=0

Lv
s.t. pv ≥ 2
∀v
Lv
pv ≥ Q− 2
∀v
Li +Lj
pi − pj ≤ 2
− Q · eij ∀(i, j) ∈ S
Li +Lj
pj − pi ≥ 2
− Q · (1 − eij ) ∀(i, j) ∈ S
zvn ≥ pv − Xn + Yn ∀n, v ∈ {1, ..., V }
zvn ≥ −pv + Xn + Yn ∀n, v ∈ {1, ..., V }
zvn = B − Yn ∀n, v = 0

 x
 
?  t
zvn  ivjn (k)? , otvn (k)?


y 

V P
N P
T P
K V
 ?
itvjn (k) + otvn (k) · zvn
P P
min
itvjn (k),otvn (k) v=0 n=1 t=1 k=1 j=0

N
itijn (k) = Iijt (k)
P
s.t. ∀t, i, j, k
n=1
N
otvn (k) = Ovt (k)
P
∀t, v, k
n=1
V
stvn (k + 1) = stvn (k) + itivn (k) − otvn (k) ∀t, v, n, k
P
i=0
T P
V T P
V P
V
stvn (k) itivn (k)
P P
≤ Cn − ∀n, k
t=1 v=0 t=1 v=0 i=0

We start from the upper box, which presents the berth position problem, and select
initial conditions itvjn (k)? , the amount of container type t flowing from source v into stack
n during time slot [k, k + 1i having destination j, and otvn (k)? , the amount of container
type t flowing from stack n to destination v during time slot [k, k + 1i. For these initial
?
values, the berth position problem is solved and the suboptimal values zvn , following
?
directly from pv , are passed to the lower box problem. In this box, the container stacking
?
problem is solved for the delivered values zvn . The resulting suboptimal values itvjn (k)?
t ?
and ovn (k) are passed to the upper box again. This procedure is repeated until the
objective value does no longer decrease.
In Section 5.3, we show that the performance of the procedure heavily depends on
the chosen initial condition. In the next subsection, we derive an additional MILP to
construct a proper guess for the initial values itvjn (k)? and otvn (k)? . Results suggest that
the solution, generated by the alternating optimization for this initial guess, outperforms
all solutions, generated by the alternating optimization for an extensive number of random
initial conditions.
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 92

5.2.4 Initial Condition

In this section, an MILP is proposed to construct a proper initial condition for the al-
ternating procedure derived in the previous section. The main difference of the MILP
here and the original MIQP is that the MILP allocates fixed-sized groups of containers
rather than variably-sized groups of containers. Since the fixed-sized container groups are
parameters in the model proposed here, a linear objective function remains (rather than a
non-convex quadratic objective function). The total amount of containers of type t with
source i and destination j in the cycle is considered to be one group. The remaining prob-
lem is to allocate berth positions to vessels and stack positions to the container groups
such that the total carrier travel distance is minimized. Below, the MILP formulation is
discussed.

Decision variables

pv = Position of the center of vessel v.


 1 if the group of containers of type t from source i to destination j is
t
wijn = allocated to stack n,
0 iotherwise.


Auxiliary variables

1 if vessel i is positioned left of vessel j,
eij =
0 if vessel i is positioned right of vessel j.
zvn = Manhattan distance between the center of vessel v and stack n.
t
svn (k) = # containers of type t with destination v in stack n during time slot [k, k + 1i.

 1 the minimal distance between source i and stack n if the group of containers
atijn = of type t from source i to destination j is allocated to stack n
0 otherwise.


 1 the minimal distance between stack n and destination j if the group of
dtijn = containers of type t from source i to destination j is allocated to stack n
0 otherwise.


The objective is to find vessel positions at the quay and container group positions in
the stack to minimize the total carrier travel distance. The total carrier travel distance of
a certain group of containers between their vessel and their stack is equal to the product
of i) the number of containers in that particular group and ii) the distance between the
corresponding vessel and the designated stack. Hence, we have to minimize the sum over
all these products:
T X
X V X
V X
K N
X N
X
Iijt (k) atijn t
dtijn


min
t
· + Oij (k) · (5.4)
pv ,wijn
t=1 i=0 j=0 k=1 n=1 n=1

First, there is a set of constraints with respect to the vessel positions. Note that this set
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 93

is equal to the set of constraints in (5.2):

Lv
pv ≥ 2
∀v
Lv
pv ≤ Q− 2
∀v
Li +Lj
pi − pj ≥ 2
− Q · eij ∀(i, j) ∈ S
Li +Lj
pj − pi ≥ 2
− Q · (1 − eij ) ∀(i, j) ∈ S
zvn ≥ pv − Xn + Yn ∀n, v ∈ {1, ..., V }
zvn ≥ −pv + Xn + Yn ∀n, v ∈ {1, ..., V }
zvn = B − Yn ∀n, v = 0
Second, we have a set of constraints with respect to the group-to-stack allocation.
N
t
P
wijn = 1 ∀t, i, j
n=1
VP
+1 VP
+1
stvn (k + 1) = stvn (k) + t
Iiv t
(k) · wivn − t
Oiv t
(k) · wivn ∀t, v, n, k (5.5)
i=1 i=1
T VP
+1 T VP+1 VP
+1
stvn (k) ≤ Cn − t t
P P
Iiv (k) · wivn ∀n, k
t=1 v=1 t=1 v=1 i=1

The first constraint in the upper set enforces that each group of containers is allocated to
exactly one stack. The second constraint updates the number of containers of a certain
type t and destination v in stack n over time k. The third constraint ensures that the
capacity of stack n is never exceeded.
Finally, we have a number of dependent constraints, which link the vessel positioning
to the group stacking, and count travel distances for a container group between a vessel
and a stack only if that particular container group of that particular vessel is allocated to
that particular stack:
atvjn ≤ t
(Q + B) · wvjn ∀t, v, j, n
atvjn ≥ 0 ∀t, v, j, n
atvjn ≤ zvn ∀t, v, j, n
atvjn ≤ t
zvn − (Q + B) · (1 − wvjn ) ∀t, v, j, n
(5.6)
dtivn ≤ t
(Q + B) · wivn ∀t, i, v, n
dtivn ≥ 0 ∀t, i, v, n
dtivn ≥ 0 ∀t, i, v, n
dtivn ≤ t
zvn − (Q + B) · (1 − wivn ) ∀t, i, v, n

5.3 Case Study

The terminal operator PSA HNN provided us with a representative data set consisting
of a cyclic timetable, vessels’ load compositions and yard lay out. For this data set, the
total carrier travel distance can be computed for the weekly cycle. In this section, we aim
for reductions in this total carrier travel distance by applying the proposed alternating
optimization procedure for i) quasi-randomly generated initial conditions and ii) the initial
condition generated by the MILP in Section 5.2.4.

5.3.1 Random Initial Conditions

First, the performance dependency of the alternating optimization on the chosen initial
condition itijn (k)? and otvn (k)? is investigated. These initial conditions are constructed in
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 94

a quasi-random manner: we use the, in Section 5.2.4 introduced, fixed-sized container

groups, allocate the groups of special container types (t ∈ {2, 3, 4}) to their fixed desig-
nated stacking areas, and randomly allocate each group of full containers t = 1 to exactly
t
one stack, i.e. we fix the values for wijn , t ∈ {2, 3, 4} according to the provided data set,
N
1 1
P
and randomly generate values for wijn , such that wijn = 1, ∀i, j. Note that in this
n=1
generation, the stack capacities are neglected and due to the random allocation, a stack
t
might be filled above its capacity. From the generated values for wijn and the parameters
t t t ? t ?
Iij (k) and Ov (k), values for iijn (k) and ovn (k) follow directly and are fed into the the
alternating procedure.
Results for one hundred of these experiments are depicted in Figure 5.3a, where a tri-
angle represents the carrier travel distance for quasi-randomly chosen stacks and variable
berth positions (objective from the first step of the alternating procedure), and the circle
straight below represents the carrier travel distance after convergence of the alternating
procedure for that particular initial condition. The carrier travel distances are scaled to
the carrier travel distance corresponding to the provided data set (grey dotted line). From
Figure 5.3a we learn the following:

• In each experiment, the alternating procedure yields significant reductions in the

carrier travel distance starting from the initial condition.

• A good randomly selected initial conditions does not necessarily provide a good end
solution. One of the possible reasons might be that, while allocating the groups
the stacks’ capacities are not taken into consideration. Hence, several groups of
containers might be allocated to the same stack, inducing a small travel distance,
however exceeding the actual capacity. Since the alternating optimization does take
the stack capacities into account, the converged solution might end up not being
that good.

• The major part (about 80%) of the quasi-randomly generated initial conditions
leads to an allocation with a larger carrier travel distance than in the representative
allocation.

• For each of the hundred performed experiments, the alternating procedure yields a
solution that is at least about 10% better than the representative allocation.

• The best solution found (black marked circle and corresponding triangle) outper-
forms the current allocation by almost 20%.

5.3.2 Sophisticated Initial Conditions

In this section, we execute the alternating optimization for the initial condition generated
by the MILP in Section 5.2.4. Since the MILP consists of a relatively large number of
t
binary variables wijn , it takes forever to solve it to optimality. Hence, the following is
proposed: the carrier travel distance for the representative allocation is computed by
t
fixing all variables pv , and wijn accordingly and evaluating the corresponding objective
function (5.4). Next, the groups of full containers (type t = 1) are ordered according to
their contributions to the total carrier travel distance. From this ordered set, we take the
first G groups with the largest distance contributions, declare the corresponding values
 CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 95

Representative allocation
1.1 1.1 Initial condition
Alternating procedure

Carrier travel distance

Carrier travel distance

1.05 1.05

1 1

0.95 0.95

0.9 0.9

0.85 0.85

0.8 0.8
0 10 20 30 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8
Experiment number G
(a) Quasi-randomly generated initial condition. (b) Sophisticated initial condition.

Figure 5.3: Results of the alternating procedure. In each experiment (both triangles and
circles), the berth positions of all vessels are variable.

1
for wijn , ∀i, j, n and pv , ∀v to be variables again, and run the MILP of Section 5.2.4.
Subsequently, the generated solution is fed into the alternating optimization procedure
of Section 5.2.3 as an initial condition. Note that in the lower box of the alternating
optimization procedure i1ijn (k) is a variable ∀i, j and otvn (k) is a variable ∀v (and not only
for the first G groups with the largest contributions). Figure 5.3b shows the results of the
found initial condition and corresponding converged solution as a function of G. From
the experiments shown in this Figure 5.3b we conclude the following:
• In each experiment, the alternating procedure yields significant reductions in the
carrier travel distance starting from the initial condition. The reduction however
decreases as G increases. A reason for this might be that for relatively large values
of G, the initial guess is that good that less improvements can be obtained by the
alternating procedure.

• Each generated initial condition already outperforms the allocation in the data set
provided by PSA HNN.

• For G = 0 (zero groups can be modified) the initial condition already outperforms
the allocation in the data set provided by PSA HNN. Although the container stack
allocation is fixed and cannot be changed for G = 0 while generating this initial
condition, the berth positions of the vessels are variable. Apparently, a modification
of only the vessels’ berth positions already yields a reduction of about 3% in carrier
i i
travel distance. i

i
• As G increases, the found objective
i value for the initial conditions (triangles)i de-
crease. This makes sense since if more container groups are variable, no larger travel
distance will result from the optimization.

• A better initial condition (triangle) never yields a worse converged solution (circle).
This disagrees with the observation made for Figure 5.3a, where a better initial
condition not necessarily led to a better converged solution. Apparently, the reason
we gave for this observation in the first place is a crucial one. Namely, the method
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 96

to construct a proper initial guess does take the stacks’ capacities into consideration
right away.

• The modification of only the eight largest contributions (and possible all vessels’
berth positions) already leads to a better solution than the best found solution for a
hundred random initial conditions. The corresponding reduction in travel distance
with respect to the travel distance in the representative allocation is more than 20%.

From these results we conclude that it pays off to generate a proper initial condition
rather than executing an extensive number of experiments for random initial conditions.

5.4 Conclusions and Recommendations

In this chapter we considered a special form of the well-known berth allocation problem,
which allocates space and time to vessels. The special form results from specific cuts
chosen in this dissertation for the overall decision problems in a multi-terminal container
operation. Namely, due to the generation of terminals’ timetables at a higher decision
level, a berthing interval is already allocated and only the space allocation problem is
left at this level. This berth position problem is addressed jointly with the allocation of
containers to stacks in the yard to minimize the total travel distance of carrier operating
between vessels and yard.
First, an appropriate MIQP is formulated, which chooses berth positions for vessels
and amounts of containers for stacks to minimize the carrier travel distance. Since the
objective is non-convex and consequently real-life instances run forever, the problem is
separated into an MILP, which represents the berth position allocation, and an LP, which
presents the container amounts to stack allocation. These problems are coupled in the
objective function and solved in an alternating fashion. The method converges to a local
optimum very fast, however appears to be very sensitive to the initial condition.
Hence an MILP is proposed to find a proper initial condition by allocating berth posi-
tions to vessels and fixed-sized groups of containers to stacks. The solution found by this
MILP is passed to the alternating method as an initial condition. The alternating opti-
mization now finds a solution, which outperforms all solutions resulting from an extensive
number of randomly generated initial conditions. Applying the alternating optimization
on a representative data set provided by PSA HNN suggests that a reduction of more
than 20% in the carrier travel distance can be obtained.
These results suggest that the same amount of work can be done with less straddle
carriers or that the inter-arrival time of straddle carriers at the quay cranes can be reduced
while using the same number of straddle carriers. The latter might reduce the idle time
of the quay cranes and with that increase their utilization.
Chapters 3 through 5 generate a reference timetable, reference berth positions and a
reference yard lay out for the considered cycle. Since in daily operations disturbances
are present on e.g. vessel arrivals, call sizes and crane productivity, the daily and even
hourly reference allocation can often not be met. Hence, an online optimization method
is required to reschedule the disturbed system. A rolling horizon approach is developed
in Chapter 6 to serve the shipping lines as good as possible for minimal operational costs.

With respect to the research in this chapter, the following recommendations are given:
CHAPTER 5. JOINT BERTH ALLOCATION AND YARD DESIGN PROBLEM 97

• The study in this chapter considers the single objective of minimizing the total
carrier travel distance. As a result, blocks closest to the berth position of a loop are
typically stacked up to three containers high (in case of a straddle carrier operation)
with containers from/for this loop, while other blocks remain relatively empty. In
this way the total distance between vessel and containers indeed is minimized. A
drawback of stacking containers on top of each other however is that it may require
additional handling to retrieve a certain container. Namely, a container, which for
instance is stacked at the bottom of a pile cannot be picked up directly, but first the
top containers have to be shifted. Each individual container that has to be shifted
to retrieve another one is called a shifter. The probability of a shifter increases as
the stacking height increases.
Since each shifter requires time and money, an interesting recommendation is to
minimize the number of shifters as a second (conflicting) objective. In an ongoing
case study, a relation between the stacking height and the expected number of
shifters is derived and introduced in the model. Given the average time required
for a shifter and the average speed of a straddle carrier, the costs of a shifter are
translated into a ”lost” straddle carrier distance. In this way, the two objectives
can be traded off fairly.

• As mentioned in Chapter 1, one of the strategic decisions a terminal operator has

to make concerns the type of equipment used for the yard operations. Generally,
for this decision a terminal operator can choose from two alternatives: i) straddle
carriers, or ii) rubber-tyred gantry cranes (RTG’s) and trucks. In the latter case, the
gantry cranes can move above the container blocks and feed/are fed by the trucks
that deliver containers to/from the quay cranes. The advantage of a straddle carrier
is that it can pick up and put down a container by itself. The main disadvantage
is that it can only stack containers up to three high and hence the yard storage
capacity is restricted.
The other alternative requires a synchronization between a crane (quay crane or
gantry crane) and a truck for loading and unloading, which may cause a crane to
wait. Moreover, the container handling rate per block is very limited since maxi-
mally two gantry cranes can operate one block at a time. The main advantage of
this alternative however is that containers can be stacked up to six high and there-
fore the yard storage capacity is double the storage capacity of a straddle carrier
operation in the same terminal.
It is interesting to compare i) the tactical performance, i.e. the straddle carrier
distance and number of shifters and ii) the operational performance of both alter-
natives. The former of these two can be completed relatively easily by increasing
doubling the storage capacities in the blocks and limiting the container handling
rate per block in the approach in this chapter. The latter issue however requires a
new optimization approach and/or simulation model and is therefore added to the
more general recommendations in Chapter 7.
Chapter 6

Online Disruption Management in a

Container Terminal: a Rolling
Horizon Approach

6.1 Introduction
The solution approaches for the strategic and tactical decision making in a multi-terminal
container port, addressed in Chapters 3 through 5, result in i) a timetable, ii) berth
positions and iii) yard layout per terminal. This cyclic schedule is considered to be a
reference for the operational planning and could be repeated over and over again if all
parameters were deterministic. In practice however, a container operation is a highly
stochastic system: vessels are sometimes early and often late (e.g. due to tailwind and
storm, respectively), call sizes of one shipping line may heavily change from one cycle to
another, and quay cranes may break down for a certain period of time. Hence, a terminal
operator has to continuously adapt the reference schedule taking forecasted information
on stochastic parameters into consideration and facing sudden breakdown of resources.
In this chapter, we propose a rolling horizon approach to support a terminal operator
in his operational decision making. This online approach constructs operational decisions
in three sequential steps taking forecasts along a limited future horizon into consideration.
The former two sequential steps, i) berth time allocation and ii) berth position allocation,
are similar to the ones chosen for the strategic and tactical decision making (see also
Figure 1.5 and Chapters 3 through 5). The remaining step is the actual quay crane
scheduling for which we propose a heuristics in this chapter.
Solving the overall operational problem in three sequential steps, rather than at once,
enables to construct all considered decisions within a couple of minutes. Since in practice,
these decisions are reconsidered each hour, the proposed rolling horizon approach is very
suitable to be applied online. Although the found decisions are suboptimal and might
theoretically even turn out to be infeasible, an extensive number of experiments suggests
that found solutions are always feasible. Case studies are performed to evaluate the effect
of applying different policies. For instance, taking forecasts on actual load compositions
into account while finding vessels’ berth positions significantly reduces the actual carrier
travel distance.

98
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 99

6.1.1 Related Work

Many studies address the well-known berth allocation problem (BAP), which involves the
joint problem of allocating time and space to vessels. As mentioned in Section 3.1.1, a main
distinction is made between the static BAP and the dynamic BAP. In the static case [21],
all vessels are already in the port before the actual berth allocation is determined. In the
dynamic case [23], vessels arrive while work is in progress. Additional release constraints
are then required to ensure that vessels only berth after their arrival. These studies
however consider a deterministic problem (no disturbances on arrivals, load compositions,
crane capacity or whatsoever) and do not incorporate the crane scheduling problem, which
can influence the process times of vessels and with that the performance of the overall
system. The authors in [41] do consider the joint problem of berth allocation and crane
scheduling, however still neglect disturbances. A berth allocation (and crane scheduling)
found by above mentioned studies might be optimal with respect to a certain objective,
however the way of how to act in case of disturbances is not addressed.
To the best of our knowledge, only the study in [37] addresses the berth allocation
problem in a stochastic environment. The authors consider a set of vessels that call at a
terminal on a regular basis. Given arrival distributions and load compositions, a cyclic
berth template is constructed using a simulated annealing algorithm that searches through
the space of all possible sequence pairs. The objective is to minimize expected delays and
connectivity costs. The connectivity costs represent the total distance that should be
covered to transport containers between two connecting vessels. The performance of the
constructed berth template is evaluated in a rolling horizon simulation, in which the
arrivals are stochastic parameters. The objective in each iteration step of the rolling
approach is to minimize delays and deviations from berth positions determined in the
berth template. The study neglects the crane scheduling problem and assumes the process
times of vessels to be fixed. The authors recommend to incorporate the crane allocation
in the model, since it influences the process time of a vessel and probably has an impact
on the system’s performance.
In this chapter, we consider the berth allocation and crane scheduling problem in a
stochastic environment. The robust cyclic timetable from Chapter 4 and the tactical berth
position and yard lay out from Chapter 5 together are considered as the reference planning.
A rolling horizon approach is proposed to online construct intelligent operational decisions
taking the reference planning and disturbances on the system into account. As time
evolves, updated forecasts on vessels’ arrivals and load compositions along a limited future
horizon become available. In each iteration step of the rolling horizon approach, the
updated forecasts are taken into account while constructing the operational decisions
along the limited time horizon. Only the operational decisions of the current time instance
are actually executed.
The operational decisions are generated in three sequential steps: i) berth time allo-
cation, ii) berth position allocation and iii) quay crane scheduling. The first two steps
together determine the berth allocation, while the latter constructs a quay crane schedule.
The considered quay crane scheduling problem is quite similar to the one considered in
[41], however we develop an alternative heuristic to solve it. The authors in [41] apply a
dynamic programming technique, which considers a new stage if a new event occurs being
either the arrival of a vessel or a change in the number of cranes (both predetermined in
the berth allocation problem). Each stage can be represented by all possible sequences of
the left most crane at each vessel berthing at that stage. An optimal path through the
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 100

stages is found by minimizing the total number of crane setups.

In this chapter, an appropriate MILP is constructed for the crane scheduling problem,
having information on the number of cranes at each vessel over time. Since the MILP
turns out to be complex from a computational point of view, it is cut into subproblems,
which are solved sequentially. The end state of one subproblem is used as an initial
condition for the next one. Experiments learn that the problem can then be solved much
faster, while still quite accurate solutions are found.

6.1.2 Contributions
The strategic and tactical decisions generated in chapters 3 through 5 depict the reference
cyclic schedule. In this chapter, the operational decision making is addressed, which takes
care of adapting the schedule under disturbances. We take into account disturbances on
• arrival times,

• number and types of containers,

• quay crane productivity.

An online rolling horizon approach is proposed, which takes updated forecasts on these
stochastic parameters along a limited time horizon into consideration to determine deci-
sions for berthing and crane scheduling at the current time instance. Each iteration step
of the rolling horizon approach consists of three sequential optimization steps:
1. Berth time allocation:
All vessels that are forecasted to be (partly) within the time horizon are taken into
account. Considering vessel lengths and call sizes, and considering limited quay
capacity and crane capacity in the terminal, the vessels are tried to be scheduled
within their reference process time. Dependent on an (expected) arrival within or
out of the defined arrival window (see Chapter 4), costs are assigned for not meeting
this condition.

2. Berth position allocation:

Given the time allocation from the vessels in the previous step, the current yard
layout, the (forecasted) load compositions (types and number of containers), and
the (fixed) positions of vessels that are actually berthing, berth positions for all
other vessels within the horizon are chosen such that non-overlapping is guaranteed
and the actual carrier travel distance is minimized.

3. Quay crane scheduling:

Given the time allocation and the berth positions, cranes are allocated to vessels,
such that the average vessel delay and the additional amount of resources are min-
imized.
Note that the first two steps are similar to steps chosen for the strategic and tactical de-
cision making in previous chapters. Namely, the berth time allocation and berth position
allocation are also solved sequentially in chapters 3 and 4, and Chapter 5. The third
step involving the quay crane scheduling typically is an operational decision problem and
therefore has not yet been considered at the strategic or tactical level. For each of the
three steps, an appropriate MILP is constructed. The total time for solving all three
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 101

MILP’s turns out to be less than ten minutes. This enables to embed them in an online
rolling horizon simulation, which is used to evaluate the system’s performance.
Simulation experiments suggest that the rolling horizon approach is very suitable as an
online decision support tool for real-life operational planning. Furthermore, the simulation
tool enables to investigate the influence of different policies on the systems’s performance:
• The rolling horizon approach takes updated information on load compositions and
the actual yard lay out into account to find berth positions that minimize the actual
carrier travel distance. Since the actual load composition of a call often deviates
from the reference load composition, the optimal operational berth position often
deviates from the reference berth position. We depict a Pareto front of the deviations
from preferred berth positions versus the actual straddle carrier distance.
• A logistics provider always has to trade-off service level against resource investments.
We depict a Pareto front of the vessels’ delays versus the number of resources re-
quired additionally to the standard number of resources.

6.2 Approach
In this section, first the decision making faced at the operational level is formally phrased
and the chosen steps to solve the problem are proposed. Next, the three sequential steps
and corresponding MILP’s are discussed in sections 6.2.2 through 6.2.4.

6.2.1 Problem description

We consider one terminal where a number L of shipping lines have one of their vessels
calling exactly once a week according to a cyclic schedule as determined in chapters 3
and 4. Given i) this schedule and ii) the expected call sizes and compositions, suitable
berth positions and container stack position(s) for each vessel line have been determined
in Chapter 5. Together, this depicts a tactical timetable for the weekly cycle of L lines,
where time is discrete and the considered cycle consists of P time slots.
According to this timetable, line l, l ∈ {1, ..., L} has a scheduled arrival time Al .
Additionally, the expected call sizes and compositions of the lines are given by Cijt , t ∈
{1, ..., T }, i, j ∈ {0, ..., L}, the number of containers of type t from source i to destination
j, where source/destination 0 represents the hinterland and source/destination 1, ..., L
represents line 1, ..., L, respectively.
In this chapter, we address the decision making problems during the operational exe-
cution of N cycles for a container terminal under disturbance. We define vessel l + αL to
be the vessel of line l that is scheduled to arrive in cycle α. Accordingly, the scheduled
arrival time Āl+αL of vessel l + αL is given by

Āl+αL = Al + αP, (6.1)

t
Moreover, the expected value of C̄i+αL,j+αL , the number of containers of type t with
source i + αL and destination j + αL, is assumed to be independent of the cycle:
t
C̄i+αL,j+αL = Cijt , (6.2)
In this chapter, we construct operational decisions each time slot k, k ∈ {1, ..., K}
and K >> P , taking forecasts of arrivals and call sizes along a limited future horizon
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 102

{k, ..., k + H}, H < P into account, and facing unexpected crane breakdowns. With
respect to the forecasts, we assume Âl+αL (k) to be the forecast at the current time k on
t
the arrival of vessel l + αL, and Ĉi+αL,j+αL (k) to be the forecast at current time k on the
number of containers of type t with source i + αL and destination j + αL. Furthermore,
we assume that while a vessel approaches the port, the forecasts on its actual arrival time
and load composition become more accurate and are exactly known a couple of hours in
advance of the actual arrival. Due to different kind of disturbances, containers destined
for a certain vessel might not yet be in the port during the time this vessel is berthing. In
this case, the concerning containers are loaded onto a subsequent vessel of the same line.
t
The values of Ĉi+αL,j+αL (k) for both vessels are adapted accordingly. Finally, we assume
that a quay crane’s mean availability is 98%, i.e. on average a quay crane is down for 2%
of the time.
Given the forecasts along horizon {k, ..., k + H} and given the current state of cranes
in the terminal at time k, a berth and crane schedule is constructed along the horizon
{k, ..., k + H}. Only the found decisions of the current time step k are actually executed.
Then, the same procedure is applied for time step k + 1, and in this way the horizon
rolls forward. A rolling horizon procedure is proposed to construct operational decisions
on berth allocation and quay crane scheduling at a certain time k. Figure 6.1 depicts
the principle of the applied rolling horizon approach. At time k, the state of current
executions and forecasts on future arrivals and load compositions are fed into the rolling
horizon procedure. Based on these parameters, the berth time allocation, the berth
position allocation and the crane scheduling are determined sequentially along horizon
{k, ..., k + H}. Next, only the constructed decisions for time k are actually executed in
the container operation. The resulting state at time k + 1 together with updated forecasts
is then again fed into the rolling horizon procedure, and in this way the horizon rolls
forward. The definition of the parameters can be found in Table 6.1 and is clarified
throughout Section 6.2. The following operational decisions are constructed sequentially
at the current time slot k:

1. The berth interval of vessels forecasted to be (partly) within the interval {k, ...k+H}.
For a vessel that is actually berthing at time k, the departure time is determined
based on the remaining workload for that vessel. For a vessel that is forecasted
not to be processed before the horizon end, the ”artificial” departure time is set to
k + H. Additionally, the corresponding workload is based on the part of the vessel
forecasted to be within the horizon. Dependent on whether vessel v arrives within
or out of its arrival window in the tactical timetable, costs are assigned if that vessel
is not processed within Pvmax .

2. The berth position of vessels forecasted to be (partly) within the horizon {k, ...k +
H}. Dependent on the constructed time allocation, and the forecasted load com-
positions, non-overlapping berth positions are constructed. Conflicting objectives
are to minimize i) the deviations from positions in the tactical plan and ii) the
forecasted straddle carrier travel distance.

3. The crane schedule for all vessels that are currently berthing or have a forecasted
arrival within the time interval {k, ...k + H}. Again, for vessels that are (forecasted
to be) only partly within the horizon, the crane schedule is constructed for the part
that is within the horizon.
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 103

From the description of the three sequential steps, it becomes clear that the procedure
does not explicitly induce a convergence to the tactical timetable. Namely, although we
take arrival window agreements into account, the deviations from the tactically scheduled
departure time are not minimized. Hence, for zero disturbances, a time allocation might
be constructed where the process time of vessel v is less than Pvmax .

Parameter Definition
L Number of shipping lines in the cycle
N Number of cycles considered for the operational planning
P Number of time slots in the cycle
K Number of time slots considered for the operational planning (K = N · P )
H Horizon length
M Terminal length
Sv Maximal number of quay cranes that can process vessel v simultaneously
F max Total amount of quay cranes and shifts in the terminal
F stand Standard amount of quay cranes and shifts operating in the terminal
Al Scheduled arrival time of line l
Āv Scheduled arrival time of vessel v
Âv (k) Arrival time of vessel v forecasted at time k
Pvmax Agreed process time vessel v
D̂v (k) Agreed departure time of vessel v forecasted at time k
Cijt Scheduled amount of containers of type t from line i to line j
C̄ijt Scheduled amount of containers of type t from vessel i to vessel j
Ĉijt Amount of containers of type t from vessel i to vessel j forecasted at time k
Wv Length of arrival window of vessel v
Q̂v (k) Total amount of work on vessel v forecasted at time k
Q?v (k) Total amount of work on vessel v forecasted to be done within horizon at time k
Lv Length of vessel v
G Safety gap between quay cranes
R Number of pieces the crane scheduling problems is cut into
V max Maximal number of vessels in one piece of the crane scheduling problem
p̄v Reference berth position of vessel v
ctv Costs if vessel v departs after its agreed departure time
cv Costs if vessel v deviates from its reference berth position
cev Reward if vessel v departs before its agreed departure time
cρv Costs if vessel v berths before its scheduled predecessors have left
cf Costs for an additional quay crane and shift

Table 6.1: Model parameters

For ease of notation, we address the operational decision making at the current time
step k. At time k, we have to consider all vessels, that are forecasted to be partly or
completely within the time horizon {k, ..., k + H}. We define the following mathematical
sets to indicate differences in vessels’ states:

• V is the set of vessels, which at time k are (forecasted to be) (partly) within the
time horizon {k, ..., k + H},

• B, B ⊆ V is the set of vessels, which are actually berthing at time k.

 CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 104

• P, P ⊆ V is the set of vessels, which at time k have a forecasted arrival time Âv (k)
within the time horizon {k, ..., k + H} and are forecasted to depart after the horizon
end k + H,

• E, E ⊆ V is the set of vessels that is early, i.e. which have arrived or are forecasted
to arrive before the left end of their scheduled arrival window (Chapter 4),

• T , T ⊆ V is the set of vessels that is in time, i.e. which have arrived or are forecasted
to arrive within their scheduled arrival window.

• L, L ⊆ V is the set of vessels that is late, i.e. which have arrived or are forecasted
to arrive after the right end of their scheduled arrival window.

• E ∪ T ∪ L = V, E ∩ T = E ∩ L = T ∩ L = Ø.

• Wv , Wv ⊆ V is the set of vessels, which were direct predecessors of vessel v in the

berth allocation at time k − 1. In our definition, vessel i is a direct predecessor of
vessel v if vessel i and v have overlapping quay positions and vessel v berths after
vessel i without another vessel berthing in between.

In the next three subsections, subsequently the berth interval allocation, the berth
i i
position allocation and the crane scheduling problem at time k are addressed. A discrete
“rollinghorizon˙temp” — 2008/12/10 — 11:43 — page 1 — #1
i event tool is used to simulate the actual container terminal under disturbancesialong time
k ∈ {1, ..., K}. The below described steps are subsequently executed for each time step
k.
reference
planning disturbances

Rolling horizon horizon plan at Container actual execution

approach time k operations at time k

forecasts at
time k
actual execution
at time k

av* av*
av*
dv* dv*
Time Position dv* Crane
allocation qv(k,...,k+H|k)* allocation qv(k,...,k+H|k)*
scheduling qv(k,...,k+H|k)*
pv*
pv*
xfv(k,...,k+H|k)*

Figure 6.1: Illustration of applied approach.

CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 105

6.2.2 Berth time allocation

In this subsection, the time allocation problem at time step k is addressed. First, the
decision variables are defined and next an MILP is proposed and discussed.

Variables

av = First time slot that vessel v is actually berthing

dv = Last time slot that vessel v is actually berthing
qv (k) = Crane capacity allocated to vessel v during time slot [k, k + 1i

MILP
Considering the previously mentioned definitions of the parameters and mathematical
sets, the following MILP is proposed for the time allocation problem at time k:
X X X
ρ −
min ctv ∆+
v + c ρ
v v + cf δ + (6.3)
av ,dv ,qv (k)
v∈V v∈V v∈V
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 106

av = k ∀v ∈ B (6.4a)
av ≥ Âv (k) ∀v ∈V \B (6.4b)
av ≤k+H ∀v ∈V (6.4c)
dv =k+H ∀v ∈P (6.4d)
dv ≥k ∀v ∈V (6.4e)
dv ≤k+H ∀v ∈V (6.4f)
av ≥ dj ∀v ∈ V, j ∈ B ∩ Wv (6.4g)
ãv ≥ dj ∀v ∈ V, j ∈ Wv \ B (6.4h)
ρv = av − ãv ∀v ∈V (6.4i)
∆v = dv − D̂v (k) ∀v ∈ V ∪ B (6.4j)
i · bv (i) ≤ dv − 1 ∀v ∈ V, i ∈ {k, ..., k + H} (6.4k)


(k + H − i · bv (i) ≤ k + H − av ∀v ∈ V, i ∈ {k, ..., k + H} (6.4l)
k+H
X
bv (i) = dv − av ∀v ∈ V (6.4m)
i=k
mv (i) = Lv · bv (i) ∀v ∈ V, i ∈ {k, ..., k + H} (6.4n)
X
mv (i) ≤ M ∀i ∈ {k, ..., k + H} (6.4o)
v∈V
qv (i) ≥ 0 ∀v ∈ V, i ∈ {k, ..., k + H} (6.4p)
qv (i) ≤ Sv · bv (i) ∀v ∈ V, i ∈ {k, ..., k + H} (6.4q)
k+H
X
ηv λ̄ · qv (i) = Q∗v (k) ∀v ∈ V (6.4r)
i=k
X
qv (i) ≤ F max (k) ∀i ∈ {k, ..., k + H} (6.4s)
v∈V
X
δ≥ qv (i) − F stand ∀i ∈ {k, ..., k + H} (6.4t)
v∈V

Constraint (6.4a) ensures that vessels that are already berthing, continue berthing,
while constraint (6.4b) ensures that the vessels having a forecasted arrival within the win-
dow cannot start berthing before their forecasted arrival time. Constraint (6.4c) enforces
that vessels, which are forecasted to arrive within the horizon start berthing before the
end of the horizon. Vessels that are forecasted to arrive within the horizon, but have a
forecasted departure time (Âv (k)+Pvmax ) out of the horizon are treated as follows: an (ar-
tificial) departure time equal to the end of the horizon k+H is assigned (constraint (6.4d))
and a vessel’s workload is assigned based on the portion of the vessel that is forecasted
to be within the horizon (see explanation of constraint (6.4r) below). Constraints (6.4e)
and (6.4f) enforce that the departure time of vessel has to be within the time horizon.
Constraint (6.4g) ensures that a vessel only berths after its direct predecessors that
are actually berthing have left. This constraint is introduced to avoid infeasible solutions
in the next step (berth position allocation) due to sudden changes in arrival forecasts in
this step. Additionally, a vessel is always tried to be placed behind its current predeces-
sors that are not berthing yet. This is strived for to enable the allocation of a suitable
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 107

berth position (Section 6.2.3). Therefore, the additional variables ãv and ρv are intro-
duced in constraints (6.4h) and (6.4i) and accordingly costs cpv are added to the objective
function (6.3), where ρ−v represents max{0, −ρv }.
As addressed in Chapter 4, vessel v has an arrival window of length Wv , where the
left and right end of the window are given by Āv − W2v and Āv + W2v , and Wv is typically
8 hours. Agreements between the terminal operator and the vessel line of vessel v state
that if the actual arrival time Âv (k) of vessel v lies within the arrival window, vessel v
has to be processed within Pvmax time. In the model, we assume the following:
• if vessel v has arrived or is forecasted to arrive before its arrival window (Âv (k) <
Āv − W2v ), than a high cost (ctv = βv ) is assigned to departing later than D̂v (k) =
Âv (k) − W2v + Pvmax .
• if vessel v has arrived or is forecasted to arrive within its arrival window (Āv − W2v ≤
Âv (k) ≥ Āv + W2v ), than a high cost (ctv = βv ) is assigned to departing later than
D̂v (k) = Âv (k) + Pvmax .
• if vessel v has arrived or is forecasted to arrive after its arrival window (Âv (k) >
Āv + W2v ), than a low cost (ctv = βv ,  << 1) is assigned to departing later than
D̂v (k) = Âv (k) + Pvmax .
To assign above costs and rewards appropriately, constraint (6.4j) is introduced and costs
−
are added to the objective function (6.4g), where ∆+ v and ∆v represent max{0, ∆v } and
max{0, −∆v }, respectively. Since cpv << ctv , vessel v is never placed behind a direct
predecessor if this would lead to a late departure time ∆v > 0.
Constraints (6.4k), (6.4l) and (6.4m) ensure that a vessel is only berthing between
its start and end time of berthing. Constraint (6.4n) enforces that Lv quay meters are
occupied if vessel v is berthing and constraint (6.4o) guarantees that the total quay
capacity M is never exceeded.
Constraint (6.4p) and (6.4q) ensure that crane capacity can only be allocated to a
vessel if this vessel is berthing. Constraint (6.4r) enforces that sufficient capacity is
allocated to process the vessels in set V ∪ B. Originally, the forecast at time k of the total
amount of containers Q̂v (k) to be unloaded from and loaded onto vessel v is given by

X N ·L
T X N ·L
T X
X
t t
Q̂v (k) = Ĉiv (k) + Ĉvj (k) (6.5)
t=1 i=1 t=1 j=1

We distinguish the following cases for the amount of work Q∗v (k) planned at time k to
be performed on vessel v within horizon {k, ..., k + H}:
• For a vessel v in the set B, the work Q∗v (k) is equal to the total amount of work on
vessel v Q̂v (k) minus the work performed on vessel v before time instance k. The
work performed on vessel v before time instance k is obtained from the operations
executed in previous time slots.
• For a vessel v in the set V that has a forecasted departure time Dv within the time
horizon, the work Q∗v (k) is equal to the total amount of work Q̂v (k),
• For a vessel v in the set P that at time k has a forecasted departure time Dv (k)
outside the time horizon, Q∗v (k) is equal to total amount of work minus the portion
of the work that is expected to be outside of the window, i.e. Q̂v − DDv (k)−(k+H)
(k)−Â (k)
Qv (k).
v v
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 108

The total crane capacity allocated at time i, i ∈ {k, ..., k + H} can never exceed the
maximal crane capacity F max (k) at time k in the terminal (constraint (6.4s)). If a crane
cannot be used due to a breakdown or maintenance, F max (k) is updated accordingly.
In a terminal operation, a fixed number of full time employees is working to drive the
equipment. Sometimes, the workload exceeds the amount of work that can be done by
this standard number of people and equipment, which we define as F stand . In that case,
additional shifts of people have to be hired, which induces additional costs. Therefore,
we aim to minimize the amount of equipment and shifts δ + additional to the standard
amount of equipment and shifts F stand . This is achieved by including constraint (6.4t) and
assigning costs cf in the objective. The MILP can be solved within a couple of minutes
for real-life instances and is hence very suitable to be applied in this online procedure.
The optimal values a?v , d?v , qv (k, ..., k + H|k)? and the positions p?v , ∀v ∈ B (p?v , ∀v ∈
B has been determined in the previous iteration step of the rolling horizon approach
approach) are fed into the berth position allocation problem for time k. The proposed
MILP is discussed in the next subsection.

6.2.3 Berth position allocation

Given the time allocation for the vessels in the considered horizon at time step k, suitable
berth positions are generated in this step. Since disturbances on load compositions are
taken into account, the optimal berth position of a vessel might deviate from the one
determined in Chapter 5, which is defined as p̄v . The multi-objective then consists of
minimizing i) the total straddle carrier travel distance to transport the forecasted amount
of containers between vessels and stacks and ii) the weighted deviations from reference
berth positions p̄v as determined in Chapter 5.

Variables

pv = Berth position of vessel v

MILP
Since the arrival and departures of the vessels are predetermined in the previous step
and fixed in this step, we know the vessels that berth simultaneously. Vessel pair (i, j) is
added to the set S if vessel i and j are in V and berth simultaneously (see Chapter 5 for
a better explanation) according to the time allocation of the previous step. Considering
the forecasted load compositions of the vessels in the set V and considering the set S,
berth positions pv , v ∈ V are generated such that the carrier travel distance is minimized
and non-overlapping of vessels is guaranteed. The following MILP is proposed:
T X
X V X
V V
X
min Ĉijt (k) · zijt + cv |pv − p̄v | (6.6)
pv
t=1 i=0 j=0 v=1
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 109

Lv Lv
≤ pv ≤ M − ∀v ∈ V (6.7a)
2 2
pv = p∗v ∀v ∈ B (6.7b)
L i + Lj
pi − pj ≥ − M eij ∀(i, j) ∈ S (6.7c)
2
L i + Lj
pj − pi ≥ − M (1 − eij ) ∀(i, j) ∈ S (6.7d)
2
t
zv0 ≥ atn tn
v0 · pv + bv0 ∀v ∈ V (6.7e)
t
z0v ≥ atn tn
0v · pv + b0v ∀v ∈ V (6.7f)
zijt ≥ pi − pj ∀t, ∀i, j ∈ V (6.7g)
zijt ≥ pj − pi ∀t, ∀i, j ∈ V (6.7h)

Constraint (6.7a) enforces that each vessel with a forecasted arrival within the time
horizon {k, ...k + H} is totally positioned within the terminal. Constraint (6.7b) fixes the
berth positions of the vessels in set B, i.e. the vessels that are actually berthing do not
change position. Constraints (6.7c) and (6.7d) guarantee that simultaneously berthing
vessels do not overlap. Additional binary variables eij are required to achieve this.
In Chapter 5, the reference yard layout has been determined and hence the stack areas
for the containers of a certain vessel are known. Given the position of the containers
in the stack areas, the mean distance to transport an import container of type t from
vessel v to stack as a function of the berth position pv of vessel v (constraint (6.7e))
is determined. Similar functions are constructed for export containers from stack to
vessel. Next, piece-wise linear functions are constructed that approximate these distance
functions. Constraints (6.7e) and (6.7f) depict the lower bound on these distances, where
atn tn
v0 and bv0 represent the parameters of the n
th
linear piece to describe the mean travel
distance of an import container of type t from vessel v to stack. In a similar way, atn 0v and
tn th
b0v represent the parameters of the n linear piece to describe the mean travel distance
of an export container of type t from stack to vessel v.
With respect to transshipment containers, the travel distance of a carrier is assumed
to be equal to the absolute difference in berth position of the two connecting vessels
(constraints (6.7g) and (6.7h)). The objective consists of two conflicting parts: minimizing
i) the total carrier driving distance and ii) the amount of organizational changes, expressed
by the weighted deviations from the reference position p̄v as determined in Chapter 5.
The MILP can be solved within less than a second for real-life instances and is therefore
very suitable to be applied in an online optimization procedure.
The optimal values p?v together with the determined values a?v , d?v , qv (k, ..., k + H|k)?
are fed into the next step being the crane scheduling.

6.2.4 Crane scheduling

The two previous steps construct the berth time and the berth position allocation at time
k considering a future horizon of length H. In this subsection, we construct at time k
a crane schedule for the horizon {k, ..., k + H} for the found berth allocation along this
horizon. First, a suitable MILP is proposed, which turns out to be rather complex from a
computational point of view. Hence, a heuristics is developed, which solves the problem
within satisfactory time, while finding a quite accurate solution.
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 110

Decision variables
In the first step, the arrival and departure a?v and d?v , and a time-variant crane capacity
qv (k, ..., k + H|k)? have been determined. Due to the allocation of crane capacity rather
than actual cranes, the process time at the first step might turn out to be to short at this
third step. Moreover, due to the breakdown of a crane, a feasible non-crossing schedule
might not be found although theoretically sufficient capacity is available. Hence, in this
step we introduce decision variables that describe the start and end of processing required
for a non-crossing crane schedule:

av = First time slot that vessel v is actually berthing

dv = Last time slot that vessel v is actually berthing

Additional to these decision variables, the auxiliary binary variable bv is required:


1 if vessel v can be processed during time slot [k, k + 1i,
bv (k) =
0 otherwise.
(6.8)

A third decision variable is the activity of crane f , f ∈ {1, ..., F } on vessel v, v ∈ V ∪B

at time i, i ∈ {k, ..., k + H}:

1 if crane f processes vessel v during time slot [k, k + 1i,
xf v (k) =
0 otherwise.

MILP
The following MILP is proposed:
X
min ctv · ∆+ f
v + c · δ̂
+
(6.9)
xf v (k)
v∈C
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 111

av ≥ a?v ∀v ∈ V (6.10a)
∆ˆ v = dv − d? ∀v ∈ V (6.10b)
v
k · bv (i) ≤ dv − 1 ∀v ∈ V, ∀i (6.10c)


(k + H − i · bv (i) ≤ k + H − av ∀v ∈ V, ∀i (6.10d)
k+H
Xc
bv (i) = dv − av ∀v ∈ V (6.10e)
i=k
F
X
xf v (i) ≤ Sv · bv (i) ∀v ∈ V, ∀i (6.10f)
f =1
X
xf v (i) ≤ 1 ∀f, i (6.10g)
v∈V
F k+H
X X k+H
X
λf v · xf v (i) ≥ qv (i|k)? ∀v ∈ V (6.10h)
f =1 i=k i=k
Lv
lf (i) ≥ xf v (i) · (p?v − ) ∀f, i, v ∈ V (6.10i)
2
Lv
lf (i) ≤ xf v (i) · (p?v + ) + M · (1 − xf v (i)) ∀f, i, v ∈ V (6.10j)
2
lf (i) ≤ lf (i + 1) − G ∀f, i (6.10k)
F X
X
xf v (i) ≤ F max ∀i (6.10l)
f =1 v∈V
F X
X
δ̂ ≥ xf v (i) − F stand ∀i (6.10m)
f =1 v∈V

The start time of processing av of vessel v has to be at least the arrival time a?v
allocated in the first step (constraint (6.10a)). Furthermore, a costs is assigned if the end
time of processing dv is larger than the generated departure time d?v (constraint (6.10b)).
Only between the start and end time of processing, a number of cranes can operate the
vessel with a minimum of zero and a maximum of Sv (constraints (6.10c) through (6.10f)).
Constraint (6.10g) enforces that at time i, a crane can be operating at one vessel at most.
k+H
qv (i|k)?
P
Within the considered horizon {k, ..., k + H}, at least a total amount of work
i=k
forecasted at time k has to be performed on vessel v (constraint (6.10h)).
Constraints (6.10i) and (6.10j) ensure that at time i, crane f is positioned within the
vessel bounds of vessel v if at time i crane f is allocated to vessel v. Furthermore, the
non-crossing of quay cranes is modeled by constraint (6.10k), which enforces that crane
f is always left of crane f + 1, f ∈ {1, ..., F − 1}, where G is the safety gap between
two neighboring cranes. The total amount of cranes allocated at time i cannot exceed
the maximal number F max (k) of cranes that are not broken or in maintenance at time k
(constraint (6.10k)). Finally, costs are assigned for each crane and shift, which is required
additional to the standard number of cranes and shifts.
If the found start and end time a?v and d?v of vessel v deviate from the generated berth
interval a?v and d?v of vessel v, the values of a?v and d?v are set as respectively upper and
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 112

lower bounds in the first step of the next rolling horizon approach iteration. To be more
precise, the following constraints are added for vessel v to the time allocation problem at
time k + 1:

av ≤ a?v (6.11a)
dv ≥ d?v (6.11b)

The MILP cannot always be solved within a time that is required for an online appli-
cation. Hence, we propose a heuristics, which can be solved within satisfactory time and
still finds quite accurate solutions.

Heuristics
In each of the Sections 6.2.2, 6.2.3 and 6.2.4, a problem is solved considering a limited
time horizon. A vessel that is partly within the horizon is cut off and the amount of work
to be done for the part within the window is assigned proportionally. The principle of the
heuristics for the crane scheduling problem at time k is very similar:

1. The values p?v , a?v , d?v and qv (k, ..., k + H|k)? from the generated berth allocation are
considered along the horizon {k, ..., k + H}.

2. The considered berth allocation is cut into R time pieces, where V r is the set of
(parts of) vessels in piece r, r ∈ {1, ..., R} such that |V r | ≤ V max ∀r ∈ {1, ..., R}.

3. The proposed MILP is applied to sequentially construct crane schedules for pieces
r ∈ {1, ..., R}. The solution values at the end time interval of piece r is used as an
initial condition for the crane scheduling of piece r + 1.

The time required to solve the heuristics significantly decreases as V max decreases,
at the expense of a less accurate solution. An extensive number of experiments suggest
that for V max = 6 a solution is found within 5% of the global solution while the runtime
is satisfactory. Since the heuristics can be solved within a couple of minutes, it can be
applied in an online setting.
Experiments on real-life cases suggest that one rolling horizon approach iteration step
(involving the time allocation, position allocation and crane scheduling) requires less than
ten minutes. Since in practice, the operational decisions on berth allocation and crane
scheduling are updated each hour, this method is very suitable to be implemented into a
practical online container terminal application. It can serve as an online decision support
tool, where it is up to the terminal operator to which extent he adopts the generated
decisions.

6.3 Results
The above described rolling horizon approach method is applied in a simulation environ-
ment (see also Figure 6.1). A discrete event tool is used to mimic the real-life container
operation system under disturbances, by generating stochastic arrivals, load compositions
and crane breakdowns. These parameters together with the reference planning are input
for the rolling horizon approach at time k, k ∈ {1, ..., K}. The rolling horizon approach
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 113

executes the three sequential steps (sections 6.2.2, 6.2.3 and 6.2.4) and delivers the values
into the container operation simulation. The constructed decisions for time k are actually
simulated. Next, the state of the system under disturbance after these executed decisions
is again fed into the rolling horizon approach and subsequently the same procedure is
performed for time k + 1.
The constructed simulation tool is applied to investigate the system’s performance for
different policies. In the presented experiments we consider discrete time k, k ∈ {1, ..., K},
where each time slot has a width of one hour. Furthermore N = 10 and P = 168, and thus
K = 1680, i.e. we consider the operational decision making along ten weeks (ten cycles).
Furthermore, at each time k a future horizon of three days length H = 72 is taken into
consideration. Forecasts on arrivals are updated three times and become more accurate
as a vessel approaches the port according to the following procedure: If the right end
of the rolling horizon approaches the scheduled arrival time of vessel v, the first forecast
on the actual arrival of vessel v Âv (k) is drawn from a truncated normal distribution
N (Āv , σl2 ). For the next time steps, this first forecast is considered to be the actual
arrival time of vessel v. However, if the center of rolling horizon approaches this forecast,
a second sample from a truncated normal distribution N (Âv (k), 1 σl2 ), 1 < 1 is drawn.
This sample becomes the updated forecast of the actual arrival time of vessel v. Finally,
if the left end of the rolling horizon is only eight hours away from the updated forecast, a
third sample is drawn from a truncated normal distribution N (Âv (k), 2 σl2 ), 2 < 1 . This
latter forecast is assumed to be the actual arrival time of vessel v. The forecasts for the
load composition of vessel v are constructed according to a similar procedure. Means and
variances of the stochastic parameters are representative for real-life container operations,
however left out for confidentiality reasons.

6.3.1 Experiment 1
The berth positions for the various loops in the tactical plan are constructed based on
the expected load compositions of the loops (Chapter 5), i.e. we assume that each of the
vessels in one loop import and export exactly the same number and types of containers
each cycle. In practice however, the load composition of the vessels of the same loop can
differ from one cycle to another. We are interested in a reduction in carrier travel distance
if operational berth positioned are generated by taking forecasts on load compositions
into account, rather than trying to satisfy the berth positions in the tactical plan. This
can be investigated by running experiments for different values of cv in (6.6) while the
other parameter values remain the same. To obtain a statistically confident result, this
is repeated nine times for each value of cv . Although this number seems quite low, a
Lilliefors test (see [31]) suggests that the output data is normally distributed. Results are
shown in Figure 6.2a.
From this figure we learn that the carrier distance can be reduced by 4% by taking the
forecasts on actual load compositions into account while finding vessels’ berth positions.
In this case, on average a vessel berths about 45 meters away from its scheduled berth
position (Note that a terminal length is commonly larger than one kilometer). We think
it is worth to construct berth positions based on forecasts on load compositions at the
expense of relatively small deviations from the tactical berth positions.
Note that the average deviation from preferred (or tactical) berth position is never
zero. A reason for this is that delayed vessels block the preferred berth positions of other
subsequent vessels.
 CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 114

45 1

Deviation from pref. berth position [m]

40
0.95

Crane capacity required

35
30
0.9
25
20
0.85
15
10 0.8
5
0 0.75
0.96 0.965 0.97 0.975 0.98 0.985 0.99 0.995 1 0 0.5 1 1.5 2 2.5
Carrier distance Average delay [hours]
(a) Deviation from preference berth position vs. (b) Average delay vs. Additional cranes and
Carrier distance. shifts.

Figure 6.2: System performance for different policies.

6.3.2 Experiment 2
In a container terminal, commonly a standard number F stand of cranes and employees are
working on a regular basis. In busy periods, an additional number of people or actually
shifts is hired to guarantee a desired service level (of vessels departing in time).
We are interested in the average vessel delay as a function of the additional cranes and
t
shifts. This can be investigated by running experiments for different values of ccfv in (6.3)
and (6.9) while the values of all other parameters (forecasts included) remain the same.
Again a Liliefors test suggests that the output data (of nine experiments) is normally
distributed. Results are shown in Figure 6.2b.
This figure suggests that on average 5% of additional cranes and shifts are required to
reduce the average vessel delay below half an hour. We have to remark that in the sim-
ulation all vessels that have arrived before or within their arrival window were processed
within their maximum process time. Hence, the average delays as depicted in the figure
totally result from vessels that arrived late.

6.4 Conclusions and Recommendations

The strategic and tactical decisions constructed in Chapters 3 through 5, result in a
cyclic reference schedule for the berth allocation and yard layout. If a container operation
was a deterministic system, this reference schedule could be executed cycle after cycle.
i
i A container operation however is a highly stochastic system and hence the operational i

i
schedule will often deviate from the
i reference. i
In this chapter, we addressed the operational decisions of finding berth times, berth
positions and crane schedules for a container terminal under disturbances. We applied
a rolling horizon approach that takes forecasts on arrivals and load compositions along
a limited future horizon and sudden crane breakdowns into account while making the
operational decisions at a certain time step. Each time step, updated information be-
comes available, and decisions are adapted accordingly. This results in a rolling horizon
procedure, which can be used in an online setting.
CHAPTER 6. DISRUPTION MANAGEMENT IN A CONTAINER TERMINAL 115

Each step of the rolling horizon approach consists of three sequential steps, being i)
berth time allocation, ii) berth position allocation and iii) crane scheduling. For each
step, an MILP is proposed, which can be solved very efficiently. Solving the three steps
sequentially rather than solving the joint problem leads to a small computation time,
which is required to implement the approach in an online setting. Since the total compu-
tation time of the three sequential steps is less than ten minutes it is very suitable to be
applied in an online container operation. Namely, in practice the considered operational
decisions are only updated each hour of the day.
The rolling horizon approach approach was successfully embedded in a simulation
environment, where a discrete event tool is used to mimic the real-life container operation
under disturbances. Forecasts on arrivals and load compositions are generated by the
discrete event tool and fed into the rolling horizon approach. Given the forecasts and the
reference schedule, the rolling horizon approach constructs operational decisions along the
future time horizon. Only the decisions at the current time instance are actually executed
by the discrete event tool. The resulting state of the system and updates on arrivals and
load compositions are passed to the rolling horizon approach again.
The simulation tool enables to investigate the performance of the system for different
policies. A first experiment learned that the straddle carrier travel distance can be reduced
by 4% if the forecasts on load compositions are taken into account and deviations from
scheduled berth positions are allowed. In a second experiment, we depicted the average
vessel delay as a function of the additionally required amount of cranes and shifts. Results
suggested that on average one additional crane and shift are required to reduce the average
vessel delay below one hour.
The developed rolling horizon approach is very suitable to be applied in an online con-
tainer operation. Forecasts on stochastic parameters and resource breakdowns are taken
into account while constructing intelligent operational decisions.

For the conducted experiments with representative stochastic behavior, a horizon length
of three days appeared to be sufficient to yield a stable system. However, if due to a con-
fluence of events a large number of vessels would arrive at the port at the same time, the
question is whether could all be fit within this horizon. A first solution would be to adapt
the window length in case such a situation occurs. An interesting theoretical study would
be to determine the minimal horizon length required to have a rolling horizon approach
that is proven to be stable for given bounds on the stochastic parameters.
Chapter 7

Conclusions and Recommendations

The last decades, the number of people and goods that have to be transported from
one place to another has made an exponential growth. For the transport resources are
required. A set of these resources linked together is called a logistics network. Such a
logistics network is usually run by a logistics service provider, who aims to deliver the
right amount of people and/or goods at the right place at the right time. To fulfill these
objectives, a provider has to make decisions on different levels, classified according to
their extent of impact on the network performance and according to the order of time
scale of their actual execution.
We distinguish strategic, tactical and operational decision making levels on the large,
medium and short term time scale, respectively. Due to the tremendous sizes of today’s
logistics networks, intelligent decision support tools are required to make the combined
decisions so as to run the network efficiently. Since the combined decisions at the various
levels or even at one specific level are often too complex to be solved at once, the overall
decision making problem can be cut into several subproblems, which then are solved
sequentially or alternatingly.
This dissertation addresses the overall decision making problems in two specific logis-
tics networks, being i) a distribution network run by a logistics service provider and ii) a
multi-terminal container operation in a sea port. For both these logistics, the following
conceptual approach is applied: firstly, we reason about how to cut the overall decision
making problem into subproblems. This is done in such a way that each subproblem
can be solved within the time allowed (dependent on the decision level) and is still a
practically interesting problem in its own right. Secondly, for each of the subproblems,
a mathematical model is built to optimize the decisions concerned. Finally, case stud-
ies are performed to compare the performance of the currently applied decisions to the
performance of the decisions found by our approach.

7.1 Conclusions
Distribution network
In Chapter 2, the decision making problem in a distribution network from the viewpoint
of a logistics service provider (LSP) is considered. The tasks of an LSP in such a dis-
tribution network is to ship different product from suppliers to retailers, possibly using
intermediate warehouses for i) storage to buffer the variability in supply and demand and

116
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 117

ii) consolidation of the various product flows so as to leverage on the economy of scales
principle.
In this study, the strategic decisions of locating warehouses in between suppliers and
retailers are assumed to be given, and only the combined tactical and operational decisions
are addressed: the LSP faces the tactical problem of constructing an efficient and consis-
tent network topology (i.e. where to establish so-called line hauls, which enable shipments
between two facilities), that still performs satisfactorily on the operational level. An ef-
ficient topology contains a small number of line hauls so as to leverage on the economy
of scales principle. A consistent topology has fixed line hauls to provide constant and
with that familiar schedules and routings for personnel. We therefore aim to construct a
network topology with a small number of fixed line hauls that still performs well on the
operational level, where costs are assigned per truck, per stored item and for early and
late delivery.
Since solving the combined tactical and operational decisions at once requires too
much computational time, a procedure is proposed that alternates between the tactical
and operational decision making. This procedure iteratively drops least used line hauls
(starting from a fully connected network) at the higher tactical level, after constructing
the operational decisions for a medium scale time period (where each day forecasts on
supply and demand are available along a limited time horizon).
Experiments suggest that this procedure finds approximately similar topologies if lin-
ear transportation costs are used rather than stepwise (economy of scale) transportation
costs. The linear model however can be solved much faster and enables to construct
topologies for real-life network instances within minutes. An extensive number of exper-
iments suggests that the procedure always finds a topology with a very small number of
links that still performs satisfactorily at the operational level. Furthermore, the found
topology turns out to be robust to changes in second and higher order moments of sup-
ply and demand distributions, however appears to be sensitive to changes in the first
moments.

Multi-terminal container operations

A container terminal operator provides vessel lines with his services of unloading and
loading, transporting and storing containers. Dependent on his turnover, an operator
might consider the expansion of the number of terminals he is operating. Once a termi-
nal becomes available somewhere, an operator can take place in a competitive bidding
procedure to operate the concerning terminal in the future. Further strategic decisions
concern the type of yard operation (either a straddle carrier operation or an operation
with rubber-tyred gantry cranes and trucks), and the number of quay cranes that have
to be placed at the quay to unload and load the vessels.
Each vessel line owns a fleet of vessels to realize a number of loops around the world.
Commonly, the number and phasing of the vessels in one loop are such that one vessel
calls at a terminal at the same time each cycle (typically once a week). On the tactical
level, a terminal operator faces the problem of constructing a tactical timetable for the
loops that call on his terminal. Since the actual arrivals often deviate from the tactical
timetable due to bad weather conditions and breakdowns, a terminal operator and each
of the vessel lines agree on two types of arrivals: within and out of a so-called arrival
window. If a vessel arrives within its arrival window, the terminal operator has to process
the vessel within a predetermined process time. If not, the operator is not bound to any
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 118

process time. Of course he tries to operate these vessels as soon as possible, however
without sacrificing the within window agreements for other vessels. Hence, a terminal
operator faces the problem of constructing a tactical timetable, which is robust to all
scenarios where vessels arrive anywhere within their arrival windows.
A second tactical problem is to allocate a berth position to a loop at the quay and the
stack positions of its containers in the yard. Goal is to minimize the distance that has to
be covered to transport containers between quay and yard. To provide operational con-
sistency for its employees, decisions on berth positions and yard lay-out are reconsidered
at the medium term time scale.
The strategic and tactical decisions result in a periodic plan, that can theoretically
be repeated over and over again. However, in practice a container operation is a highly
stochastic system: arrivals, call sizes and compositions often deviate from the tactical plan
and resources may break down for unknown time. Hence, a terminal operator constantly
has to reschedule the operations to service the vessel lines as good as possible.
In this dissertation, we consider the terminal operator PSA HNN, which is respon-
sible for a number of terminals in the port of Antwerp, Belgium. The current policy of
PSA HNN is to satisfy the preferences of the vessels lines with respect to the terminal
and time of berthing. Although the strategic decisions of the number of terminals, the
kind of operations (straddle carriers), and the number of quay cranes are already given,
we question whether the same amount of vessel lines could be serviced with less quay
cranes if small modifications to the lines’ preferences were allowed. If so, the spare crane
capacity could be used to service additional vessel lines. Additionally, we are interested
in a reduction in the costs for inter-terminal transport induced by connecting vessels that
berth at different terminals (in the same port).

In Chapter 3, we address the combined strategic and tactical decision making problem
of determining a i) terminal and ii) time for berthing such that the required quay crane
capacity and the costs for inter-terminal transport are minimized. A mathematical model
is proposed to allocate a terminal and time for berthing to a number of loops that calls
periodically on the port of Antwerp. Conflicting objectives are spreading the workload
evenly over the terminals and minimizing the costs for inter-terminal transport. Although
sufficient quay and crane capacities are reserved, the actual berth position of vessels and
quay cranes operating them are left to be allocated in a subsequent problem. The model
is applied to a representative allocation provided by PSA HNN. Results suggest that with
small modifications to the representative allocation, the same number of vessel lines can
be operated with only 75% of currently used capacity while at the same time a reduction
of 3% in the costs for inter terminal transport can be achieved.
The decisions made in Chapter 3 result in a (periodic) timetable per terminal. In
Chapter 4, this timetable is slightly modified to increase its robustness to the within
arrival window agreements. In our definition, a berth plan is robust with respect to a
given set of arrival windows if a feasible solution exists for each arrival scenario where all
vessels arrive within their windows. The price for achieving this robustness is then the
additional crane capacity reservation that is required in the worst case arrival scenario
where all vessels arrive within their windows. The problem is hence to construct a window-
based berth plan that minimizes the maximally required crane capacity for all scenarios
where vessels arrive within their arrival windows.
An appropriate mathematical model is proposed that allocates a time for berthing and
reserves crane capacities for each vessel such that the within window arrival agreements
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 119

are satisfied. Objective is to minimize the maximally required crane capacity reservation.
Although sufficient quay and crane capacities are reserved, the actual berth position of
vessels and quay cranes operating them are left to be allocated in a subsequent problem.
As a special case, the model finds the nominal berth plan, i.e. the optimal deterministic
plan if disturbances were neglected. Results suggest that the window-based plan requires
slightly more crane capacity than the nominal plan if arrivals are only slightly disturbed.
However, the window-based plan is much more robust for relatively large disruptions that
are still within the arrival window bounds.
The allocations in chapters 3 and 4 result in a robust tactical timetable per terminal.
The exact positions of the vessels within the terminal however, are still to be determined.
In Chapter 5, the combined problem of berth position allocation and yard layout design
is addressed. Given the arrival and departure times per loop and their call sizes and call
compositions, the objective is to find berth positions at the quay and container positions
in the yard such that the total carrier travel distance is minimized.
A mathematical model for the combined problem is proposed, however cannot be
solved within reasonable time. Hence the problem is cut into the berth position problem
and the yard design problem. Since both problems can be decoupled completely, but
share the same objective function they are very suitable to be solved alternatingly. The
alternating optimization converges to a local minimum very fast, however heavily depends
on the initial condition. Therefore, an additional mathematical model is proposed that
finds a proper initial condition. Experiments suggest that starting from this initial con-
dition leads to a solution that outperforms the best of all solutions resulting from an
extensive number of random initial conditions. Experiments suggest that this method
finds an allocation that reduces the current carrier travel distance by over 20%.
The strategic and tactical decisions constructed in chapters 3,4 and 5 depict the pe-
riodic schedule and allocation per terminal. In Chapter 6, an online decision tool is
proposed to react upon disrupted vessel arrivals, call sizes and crane rates. A rolling
horizon approach is used that takes forecasts of these stochastic parameters along a lim-
ited time horizon into account while constructing the operational decisions for the current
time instance. The operational decisions are subdivided into three sequential steps. For
each of the steps a mathematical model is proposed: firstly, a time for berthing is allo-
cated taking the within window arrival agreements, and quay and crane capacities into
account. Secondly, given the time allocation, an appropriate berth position is found for
each vessel such that the carrier travel distance is minimized. Finally, given time and
position allocation, a feasible quay crane schedule is constructed.
Experiments for a representative data set suggest that the first and third steps can be
solved within a couple of minutes, while the second step can be solved within less than
a second. Since in practice these operational decisions are only reconsidered each hour,
the proposed procedure is very suitable to be applied in an online setting. Simulations
suggest that using this procedure, significant reductions in the carrier travel distance can
be achieved compared to the distance following from the current policy. Namely, the
procedure takes forecasts on actual call compositions into account while finding an appro-
priate berth position of a vessel, which may deviate from the tactical berth position. The
current policy however tries to satisfy these tactical berth positions as good as possible.
As mentioned before, the combined decisions in the considered logistics networks are
too complex to be solved within reasonable time and we choose to cut them into several
subproblems, which are then solved subsequently. It turns out that with the proposed
method each of the subproblems can be solved within the time allowed at the concerning
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 120

decision level. Although some managerial decisions or maybe even some physical con-
straints might not be incorporated, the proposed methods at least enable to quantify the
resulting additional costs, which can be very valuable for a logistics provider. The solution
methods derived in this dissertation are therefore very suitable to provide insights in the
costs of certain modifications. Moreover, the methods can be applied as decision tools,
which support a provider in his decision making. Next, it is up to the provider to which
extent he actually follows the decisions generated by the proposed methods. Below we
give a short summary of the main contributions and refer to the corresponding section(s)
where these contributions can be found.

Contributions
• Identification of the combined decision problems at strategic, tactical and opera-
tional levels in two logistics networks: a distribution network and a multi-terminal
container operation. This is addressed in Section 1.2.1 for the distribution network
and in Section 1.2.2 for the multi-terminal container operation.

• Division of the combined problems into subproblems that i) can be solved within
the time allowed and ii) are practically interesting in their own right. For the dis-
tribution network this is discussed in Section 1.4.1, for the multi-terminal container
operation in Section 1.4.2.

• Proposition and construction of approaches to solve the individual subproblems. In

Section 2.2, the bi-level approach for the combined decision making at the tactical
and operational levels in a distribution network is developed. Section 3.2 proposes
two mathematical formulations (mixed integer linear program) for a strategic de-
cision problem in a multi-terminal container operation. A mathematical problem
(mixed integer linear program) for the construction of a robust container terminal
timetable is derived in Section 4.2. An alternating optimization method is developed
in Section 5.2 to find tactical berth positions and a corresponding yard lay-out. In
Section 6.2, a rolling horizon approach is proposed and constructed for the online
disruption management of a container terminal.

• Validation of the performance and suitability of the methods. Section 2.3 suggests
that the proposed bi-level optimization method for the distribution network can be
solved very efficiently and hence can be very helpful to a logistics service provider.
In Section 3.3, a performance analysis of the two proposed models for the strategic
decision making problem in a multi-terminal container port learns that the alter-
native formulation can be solved much faster than the straightforward formulation.
The alternative method enable to solve real-life instances. The suitability of the
rolling horizon approach for an online application is shown in Section 6.3.

• Application of the methods in case studies with representative data provided by

logistics providers. A case study on the potential reductions in the number of quay
cranes and inter-terminal transport is conducted in Section 3.4. The case study in
Section 4.3 aims for robustness improvements on a representative terminal timetable.
In Section 5.3, a case study is performed on to reduce the total carrier travel distance
of a representative data set.
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 121

• Quantification of the improvements that can be achieved with respect to solutions as

currently applied by logistics providers. Although we are aware of the fact that our
models do not take managerial decisions or even physical constraints into account,
the performed case studies give a proper quantification of the resulting additional
costs. The case study in 3.4 suggests that with relatively small modifications to an
existing terminal allocation, the required crane capacity can be reduced by about
25%, while at the same time the costs for inter-terminal transport can be reduced
by 3%. Furthermore, results suggest that if all vessel lines would allow a change in
the terminal they call on, the costs for inter-terminal transport could be reduced by
about 40%. The results of the case study in Section 4.3 suggest that improvements
can be made on the robustness of a timetable. For a proper quantification, arrival
distributions are required. The case study in Section 5.3 suggests that the total
carrier driver distance can be reduced by over 20%.

7.2 Recommendations and Ongoing Work

At the end of each chapter, recommendations are given for studies that can be performed
relatively fast by slightly adapting or expanding the proposed models or methods for that
particular chapter. In this section, first recommendations for improvements on the total
chosen multi-step approach are given. Next, we phrase ongoing or subsequent studies
that require relatively large modifications or even totally new models. Finally, we discuss
the potential of the proposed models and methods for different applications.

Criticism
One drawback of the concept of solving the subproblems subsequently is that an alloca-
tion constructed in one subproblem might be infeasible in a subsequent step. Although
not one of an extensive number of representative experiments revealed this problem, a
recommendation is to build in feedback loops that in case of an infeasibility return to
a previous subproblem and generates additional constraints to fix it. For instance, in
the first optimization step of the container operations, a set of vessel loops is allocated
to a certain terminal and a corresponding timetable is constructed. Although sufficient
quay capacity is reserved at this first optimization step, an actual feasible berth position
(third optimization step) might not exist. A possible solution may be to track the time
instance(s) that the quay is overloaded and to decrease the total quay capacity on these
instances in the first optimization step and run it again.
Another drawback of the concept of solving subsequent subproblems is that we are no
longer guaranteed to find a global optimum. In this dissertation however little attention
is paid to determine the (expected) deviation between the found solutions and a global
optimum. The focus is more on the improvements that can be gained with respect to
the solutions as currently applied in real-life logistics networks. A recommendation is to
quantify the deviations between the solutions found be the proposed methods and a global
optimum. In Chapter 2, results from the heuristics are compared to a global optimum for
small instances. The same can be done for the combined decision problems in the multi-
terminal container operations. For instance, the performance of the initial model that
incorporates both the berth position allocation and the yard design problem (Chapter 5)
can be compared to the developed alternating procedure (Chapter 5) for small instances.
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 122

Additionally, the performance of the chosen subsequent procedure can be compared to

the performance of other heuristics (e.g. genetic algorithms).

Subsequent studies
In Chapter 2, it is assumed that the strategic decisions on the location of the warehouses
in the distribution network are given. A study that addresses the strategic problem of
finding proper locations and capacities of (new) warehouses is very interesting. As a
problem definition we think of the following: given the locations of suppliers, retailers
and existing warehouses and given the expected supplies and demands, find the loca-
tion(s) of new warehouses such that the expected driving distance is minimized. The
alternating optimization procedure as proposed in Chapter 5, would be very suitable to
solve this problem. Namely, the warehouses to be placed can be seen as vessels and the
existing facilities as the stacks. Variables are i) the (two-dimensional) position of the new
warehouses and ii) the number of trucks along a certain line haul between two facilities.
The number trucks along a line haul can be seen as the number of containers between a
vessel and a block. The goal is then to minimize the sum over all products of the number
of trucks along a line haul and the graphical length of this line haul.
In Chapter 3, a timetable for a multi-terminal container operation is constructed under
the assumption that each calling loop has a cycle length of one week. In practice however,
a limited number of loops have a deviating cycle length, by calling three times in four
weeks or once in two weeks. In this case, a cyclic timetable of four weeks instead of one
is required. The existing model only needs to be adapted slightly to incorporate these
multiple cycle lengths. For each loop four, three or two (dependent on the cycle length)
variables are introduced, which each represent subsequent calls of the same loop. The
time allocations of the calls however are dependent since their inter-arrival time (cycle
length) is known. Moreover, containers can now be distributed among the (four, three
or two) vessels of the destined loop and hence an additional constraint is required to
model this. By minimizing the number of stored containers over time, stacked containers
are induced to leave with the next vessel. Although the number of variables will grow
by about a factor four, the additional variables are all auxiliary. Hence, we expect the
complexity of the model not to increase much.
Along the same line, the method in Chapter 5 can be expanded to manage multiple
cycle lengths. In an ongoing case study for a terminal operated by PSA HNN, we have
already successfully adapted the model to incorporate multiple cycle length. The case
consists of a number of vessel line that calls once each week and one vessel line that calls
three times in four weeks. Hence, a cyclic plan for four weeks had to be constructed.
Variables were introduced for the berth positions of the four, and in the one case three,
calls of one loop. Since the arrivals, departures and load compositions of the calls are
given for this problem, we could predetermine which containers had to go with which call.
The four week plan resulting from the alternating optimization has recently been adopted
by PSA HNN.
The online operational decision making has been addressed in Chapter 6 of this disser-
tation. The latter of the three steps approach assigns quay cranes to vessels for a certain
time. However, the exact sequence of unloading and loading the bays of a vessel is still
to be determined. In an ongoing study, we aim to construct an unloading and loading
sequence for quay cranes such that the number of opposite (loading versus unloading)
movements is maximized. The underlying thought of this objective is to increase the
CHAPTER 7. CONCLUSIONS AND RECOMMENDATIONS 123

number of double moves by straddle carriers. Commonly, a number of straddle carriers is

assigned to a vessel. If the vessel is totally unloaded first, the concerning straddle carriers
bring containers from vessels to stack and return empty (single move). For the loading
this process is performed exactly the other way around. However, if part of the loading
and unloading processes of a vessel would be done simultaneously, straddle carriers could
transport import containers from vessel to stack and bring export containers on their way
back. We think this will significantly increase the efficiency of the straddle carriers. A
mixed integer linear program has been constructed that maximizes the opposite move-
ments given the unload and load times for each bay and guaranteeing non-crossing of
quay cranes. Since the model is quite slow, we have to find a heuristics to solve real-life
instances.
In Chapter 5, we shortly introduced the principle of a shifter, defined as the handling
of shifting a container to retrieve another one underneath. As has been mentioned, the
number of shifters increases with the stacking height and can thus be reduced by spreading
containers evenly over the blocks. However, another way to reduce the number of shifters
is by intelligent stacking. Containers that are expected to leave the stack shortly should
be stacked on top of containers that are expected to remain in the stack for a longer
period of time. With respect to the minimization of shifters one can consider i) the
tactical problem, i.e. construct a three-dimensional yard layout that is robust (with
respect to shifters) to stochastic behavior of vessel arrivals and load compositions, and ii)
the operational problem of how to stack containers in an online setting.

Potential applications
As discussed above, the method proposed in Chapter 5 of finding i) node (vessel) locations
and ii) flows between nodes to minimize the total expected travel distance can be applied
in the strategic distribution network design problem. Since the minimization of the total
(expected) travel distance is one of the core objectives in strategic network design, we
think this alternating method is suitable for the strategic design of many other applications
(railway, air and road networks) as well.
The same is valid for the tactical decisions of allocating vessels to the different ter-
minals in the same port. Railway and air applications for instance deal with similar
problems. One main difference is that in these networks the minimization of the inter-
terminal traffic (of people) is of major importance rather than spreading the workloads
evenly. For an airway network the workload for unloading and loading should still be
spread evenly over time, but is of minor importance. In public transportation, no effort
for loading and unloading is required at all.
In Chapter 6, the online disruption management tool is constructed for one single
terminal. This can be justified by the fact that in practice container vessels of one loop
have a fixed terminal (according to the tactical timetable) and never deviate from it.
In public transport and air applications, vehicles are often assigned a different terminal
just before arrival. The online tool should therefore be expanded to consider multiple
terminals simultaneously. Another difference is that the timescale in air and railway
applications is much more refined and decisions have to be reconsidered each minute or
even each second. The question is whether the proposed rolling horizon approach can
still be solved in satisfactory time. Probably, additional heuristics are required to update
decisions much faster.
Bibliography

[1] M. Ball, C. Barnhart, G. Nemhauser, and A. Odoni. Handbooks in Operations Re-

search and Management Science: Transportation, chapter 1: Air Transportation:
Irregular Operations and Control, pages 1–67. Elsevier B.V., 2007.

[2] L. Bertazzi, G. Paletta, and M.G. Speranza. Deterministic order-up-to level in an

inventory routing problem. Transport Sci, 36(1):119–132, 2002.

[3] D. Bertsimas and M. Sim. The price of robustness. Operations Research, 52(1):35–53,
2004.

[4] J.C. Bezdek and R.J. Hathaway. Lecture Notes in Computer Science, volume
2275/2002, chapter Some Notes on Alternating Optimization, pages 187–195.
Springer Berlin / Heidelberg, 2002.

[5] G. Caimi, M. Fuchsberger, M. Laumanns, and K.r Schüpbach. Periodic railway

timetabling with event flexibility. In C. Liebchen, R. Ahuja, and J. Mesa, edi-
tors, ATMOS - 7th Workshop on Algorithmic Approaches for Transportation Mod-
eling, Optimization, and Systems, pages 124–141. Internationales Begegnungs- und
Forschungszentrum für Informatik (IBFI), Schloss Dagstuhl, Germany, 2007.

[6] R.K.M. Cheung and W.B. Powell. Models and algorithms for distributions problems
with uncertain demands. Transport Sci, 30(1):43–59, 1996.

[7] J. Clausen. Disruption management in passenger transportation — from air to tracks.

In C. Liebchen, R. Ahuja, and J. Mesa, editors, ATMOS - 7th Workshop on Algo-
rithmic Approaches for Transportation Modeling, Optimization, and Systems, pages
30–47. Internationales Begegnungs- und Forschungszentrum für Informatik (IBFI),
Schloss Dagstuhl, Germany, 2007.

[8] Jean-François Cordeau, Gilbert Laporte, Pasquele Legato, and Luigi Motta. Models
and tabu search heuristics for the berth-allocation problem. Transportation Science,
39:526–538, 2005.

[9] J.F. Cordeau, G. Laporte, and F. Pasin. An iterated local search heuristic for the
logistics network design problem with single assignment. Int J Prod Econ, 2006.

[10] Jim Dai, Wuqin Lin, Rajeeva Moorthy, and Chung-Piaw Teo. Berth allocation plan-
ning optimization in container terminals. Working Paper, 2004.

[11] M.S. Daskin. Networks and Discrete Location: Models, Algorithms, and Applications.
John Wiley and Sons, Inc., 1995.

124
BIBLIOGRAPHY 125

[12] R. de Borst, S.J. Hulshoff, S. Lenz, E.A. Munts, H. van Brummelen, and W. A. Wall.
Multiscale methods in computational fluid and solid mechanics. In P. Wesseling, E. O
nate, and J. Périaux, editors, ECCOMAS CFD 2006, 2006.

[13] C.E. Garcia, D.M. Prett, and M. Morari. Model predictive control: theory and
practice. Automatica, 25(3):335–348, 1989.

[14] Michael R. Garey and D.S. Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. Freeman, New York, 1979.

[15] Y. Guan and R.K. Cheung. The berth allocation problem: Models and solutions
methods. OR Spectrum, 26:75–92, 2004.

[16] P. Hansen, C. Ŏguz, and N. Mladenović. Variable neighborhood search for minimum
cost berth allocation. European Journal of Operational Research, 191(3):636–649,
2008.

[17] Pierre Hansen, Ceyda Oğuz, and Nenad Mladenović. Variable neighborhood search
for minimum cost berth allocation. Article Press, Eur. J. Oper. Res, 2007, 2007.

[18] A.C. Hax and D. Candea. Production and Inventory Management. Prentice Hall,
Inc., 1984.

[19] Maarten Hendriks, Dieter Armbruster, Marco Laumanns, Erjen Lefeber, and Jan Tij-
men Udding. Strategic alloction of cyclically arriving container vessels to inter-related
terminals. Submitted, 2007.

[20] A. Imai, H.C. Chen, E. Nishimura, and S. Papadimitriou. The simultaneous berth
and quay crane allocation problem. Transportation Research Part E, 44:900920, 2008.

[21] A. Imai, K. Nagaima, and W.T. Chan. Efficient planning of berth allocation for
container terminals in asia. Journal of Advanced Transportation, 31:75–94, 1997.

[22] A. Imai, E. Nishimura, Masahiro Hattori, and S. Papadimitriou. Berth allocation at

indented berths for mega-containerships. Eur. J. Oper. Res, 179:579–593, 2007.

[23] A. Imai, E. Nishimura, and S. Papadimitriou. The dynamic berth allocation problem
for a container port. Transportation Research Part B, 35:401–417, 2001.

[24] A. Imai, E. Nishimura, and S. Papadimitriou. Berth allocation in a container port:

using a continuous location space approach. Transportation Research Part B, 39:199–
221, 2005.

[25] P. Jaillet, G. Song, and G. Yu. Airline network design and hub location problems.
Locat Sci, 4(3):195–212, 1996.

[26] W.C. Jordan and S.C. Graves. Principles on the benefits of manufacturing process
flexibility. Manage Sci, 41(4):577–594, 1995.

[27] Kap Wan Kim and Kyung Chan Moon. Berth scheduling by simulated annealing.
Transportation Research Part B, 37:541–560, 2003.
BIBLIOGRAPHY 126

[28] S. Kwon, K. Park, C. Lee, S.S. Kim, H.J. Kim, and Z. Liang. Supply chain network
design and transshipment hub location for third party logistics providers. In Lecture
notes in computer science, volume 3982, pages 928–933. Springer Berlin / Heidelberg,
2006.
[29] S. Lan, J.P. Clarke, and C. Barnhart. Planning for robust airline operations: Opti-
mizing aircraft routings and flight departure times to minimize passenger disruptions.
Transportation Science, 40(1):15–28, 2006.
[30] J.M.Y. Leung, T.L. Magnanti, and V. Singhal. Routing in point-to-point delivery
systems. Transport Sci, 24:245–261, 1990.
[31] H. Lilliefors. On the kolmogorov-smirnov test for normality with mean and variance
unknown. Journal of the American Statistical Association, 62:399–402, 1967.
[32] Andrew Lim. The berth planning problem. Operations Research Letters, 22:105–110,
1998.
[33] C.I. Liu, H. Jula, K. Vukadinovic, and P. Ioannou. Automated guided vehicle system
for two container yard layouts. Transportaion Research Part C, 12:349368, 2004.
[34] Y. Meepetchdee and N. Shah. Logistical network design with robustness and com-
plexity considerations. Int J Phys Distrib, 37(3):201–222, 2007.
[35] M.F. Monaco and M. Sammarra. The berth allocation problem: A strong formulation
solved by a lagrangean approach. Transportation Science, 41(2):265–280, May 2007.
[36] R. Moorthy and C.P. Teo. Berth management in container terminal: the template
design problem. OR Spectrum, 28(4):495–518, October 2006.
[37] Rajeeva Moorthy and Chung-Piaw Teo. Berth management in container terminal:
the template design problem. OR Spectrum, 28:495–518, 2006.
[38] E. Nishimura, A. Imai, and S. Papadimitriou. Berth allocation planning in the public
berth system by genetic algorithms. Eur. J. Oper. Res, 131:282–292, 2001.
[39] Jaap A. Ottjes, Hans P.M. Veeke, Mark B. Duinkerken, Joan C. Rijsenbrij, and
Gabriel Lodewijks. Simulation of a multiterminal system for container handling. OR
Spectrum, 28:447–468, 2006.
[40] Christos H. Papadimitriou and Kenneth Steiglitz. Combinatorial optimization: al-
gorithms and complexity. Englewood Cliffs: Prentice-Hall, 1982.
[41] Young-Man Park and Kap Hwan Kim. A scheduling method for berth and quay
cranes. OR Spectrum, 25:1–23, 2003.
[42] Michael Pinedo. Scheduling Theory, Algorithms, and Systems. Englewood Cliffs:
Prentice-Hall, 1995.
[43] Hassan Rashidi and Edward P.K. Tsang. Container terminals: Scheduling decisions,
their formulation and solutions. Submitted to the Journal of Scheduling, 2006.
[44] D. Simchi-Levi, X. Chen, and J. Bramel. The Logic of Logistics. Springer Sci-
ence+Business Media, Inc., 2nd edition, 2005.
BIBLIOGRAPHY 127

[45] R. Stahlbock and S. Voß. Operations research at container terminals: a literature

update. OR Spectrum, 30:1–52, 2008.

[46] Dirk Steenken, Stefan Voß, and Robert Stahlbock. Container terminal operation and
operations research-a classification and literature review. OR Spectrum, 26:3–49,
2004.

[47] H. Topaloglu. An approximate dynamic programming approach for a product distri-

bution problem. IIE Trans, 37:697–710, 2005.

[48] P. Tsiakis, N. Shah, and C.C. Pantelides. Design of multi-echelon supply chain
networks under demand uncertainty. Ind Eng Chem Res, 40:3585–3604, 2001.

[49] Iris F.A. Vis and René de Koster. Transshipment of containers at a container termi-
nal: An overview. European Journal of Opertional Research, 147:1–16, 2003.

[50] M.J.C.M. Vromans, R. Dekker, and L. Kroon. Cyclic railway timetabling: A stochas-
tic optimization approach. Lecture Notes in Computer Science, 4359:41–66, 2007.

[51] Fan Wang and Andrew Lim. A stochastic beam search for the berth allocation
problem. Decision Support Systems, 42:2186–2196, 2007.

[52] M. Wasner and G. Zapfel. An integrated multi-depot hub-location vehicle routing

model for network planning of parcel service. Int J Prod Eco, 90:403–419, 2004.

[53] G. Zapfel and M. Wasner. Planning and optimization of hub-and-spoke transporta-

tion networks of cooperative third-party logistics providers. Int J Prod Econ, 78:207–
220, 2002.