–A ship in the harbor is safe, but that is not what ships

are built for.

John A. Shedd

1
Introduction

h e European Commission (EC) defined a goal to achieve a 60%

T reduction of greenhouse gas (GHG) emission by 2050 in the trans-

port sector [15]. To achieve this goal, every transport sector will
have to make smarter and better decisions. The International Mar-
itime Organization (IMO) responsible for regulating shipping independently
announced that by 2025 all new ships will have to be 30% more energy ef-
ficient than those built in 2014 [35]. Furthermore in a recent press briefing
the IMO announced that the shipping industry should reduce the total an-
nual GHG emission by 50% by 2050 compared to 2008 [18]. To achieve the
goals set by the EC and the IMO, the new ships that are currently being
engineered will have to be optimized for minimum environmental impact. Of
course, the environmental impact is not the only objective to consider while
optimizing a ship. The ship owners also want their ship to be operationally
efficient and to have the lowest operational expenses possible. Additionally,
safety and comfort of crew and/or passengers should meet the criteria given
by the regulating authorities.
To achieve an optimal solution where all stakeholders are satisfied, typ-
ically different experts work together to optimize the ship. These experts
traditionally optimize the ship using heuristic methods learned over the
course of years, derived from knowledge gained through a process of trial
and error. In this optimization process, naval architects traditionally use
the design spiral [13] (Figure 1.0.1). As prescribed in the theory about the
design spiral, the components of a ship are optimized step by step, iteratively.

1
Research Question Chapter 1. Introduction

Figure 1.0.1: Classical ship design spiral.

This traditional, expert driven, iterative approach used to design a ship

can cause the design process to get stuck in local optima. This is the case
because, for a single naval architect or a group of experts, it is impossible to
consider the whole design space and all the relationships and dependencies
between the variables, constraints and objectives [26].
To make better design decisions in the future, the ship optimization
processes such as proposed by Papanikolau [26] could be used. This integrated
design approach brings together all key aspects of a design task at the same
time.

1.1 Research Question

In this thesis, it is shown that an integrated design approach in combination
with the proposed optimization algorithm results in significantly improved
ship design solutions. This is shown by giving answer to the following research
question and sub-questions:

How can state of the art optimization algorithms efficiently

support multidisciplinary ship design applications?

The answer to this question is not a simple one and starts with a better
understanding of state of the art optimization algorithms, multidisciplinary
ship design applications, and efficient multi-objective problem solving.

2
Outline Chapter 1. Introduction

Currently there already exist several algorithms that show great results
on artificially designed constrained multi-objective optimization problems.
To find out whether they are suitable for solving ship design applications the
following sub-question is defined: Which already existing state of the
art optimization algorithms and theories could be used to solve
multidisciplinary ship design optimization applications?
The algorithm should be able to optimize multidisciplinary ship design
optimization applications. Here, the term multidisciplinary refers to different
ship design problems. Some optimization problems will have for example six
decision variables, four objectives and fifteen inequality constraints, while
other optimization problems might have equality constraints or only two
objectives. To see how the developed algorithm deals with the multiple con-
straints and objectives, the following sub-question is defined: How flexible
are the state of the art optimization algorithms in terms of dealing
with multiple constraints and multiple objectives?
Furthermore, the algorithm has to be efficient, because optimizing a ship
design can be very time consuming. Consider for example the case where we
want to minimize the resistance. To minimize the resistance of a ship design,
multiple Computational Fluid Dynamics (CFD) simulations [36] will have
to be executed for different design variations. Only one CFD of one design
variation can take up to several hours of computation time. Therefor, the
next sub-question to be answered is: In terms of function evaluations
and solution quality, how efficient is the proposed optimization
algorithm compared to existing approaches?

1.2 Outline
The next chapters describe the research that has been conducted to answer
the questions. In Chapter 2 the related research and relevant algorithms are
discussed. The formulation of the problem statement is given in Chapter 3.
In Chapter 4 the proposed algorithm is discussed in detail. Next, it is shown
empirically in Chapter 5 that the proposed algorithm is efficient and is
able to find a good approximation of the Pareto frontier using a limited
evaluation budget. Finally, the results are discussed and conclusions are
drawn in Chapter 6.

3
–If I have seen further it is by standing on the shoulders
of Giants.
Isaac Newton

2
Earlier Research

l r e a dy q u i t e s o m e wo r k has been done in the domains

A of constraint problem solving, multi-objective problem solving,

and ship design optimization. Therefore, in this chapter the most
relevant approaches and algorithms to our problem are discussed.
This is organized by first discussing the constraint handling, then multi-
objective optimization, and finally some earlier research in ship design opti-
mization is addressed.

2.1 Constraint Handling

In the literature, several approaches for constraint handling in evolutionary
algorithms have been presented so far [1, 4, 5, 7, 8]. The most prominent
constraint handling techniques are the following:

1. Add a penalty to the objective functions for infeasible solutions [7],

2. Preferring feasible solutions above infeasible solutions in selection [8],
3. Multi-objective optimization with additional objectives for the con-
straint functions [5],
4. Repair algorithms that repair proposed infeasible solutions during the
search [4],
5. Machine learning/surrogate models to model the constraint functions [1].

4
Constraint Handling Chapter 2. Earlier Research

2.1.1 Penalty Functions

Adding a penalty to the objective function for infeasible solutions is a fre-
quently used approach to handle constraints [7]. The idea behind this proce-
dure is that infeasible solutions get a worse objective function value compared
to the feasible solutions. This way the constrained optimization problem
becomes an unconstrained one by adding the constraint violations to the
objective function as a penalty term.
The main drawback of this approach is that this penalty function should
include every constraint. Together with the penalty function the right balance
between the fitness function and penalty function should be found. To find
this right balance often new balance parameters need to be optimized. In
order to find this right balance, typically a lot of trail and error is needed
for effectively minimizing the penalty term to find good, feasible solutions.

2.1.2 Feasible Solution Preference

Feasible solution preference methods always prefer feasible solutions to infea-
sible solutions. This typically causes that during search, too little information
from infeasible solutions is used. Deb [8] improved this method by introduc-
ing a diversity mechanism where an infeasible solution is sometimes ranked
according to its fitness value among the feasible ones. This often shows good
results but typically a large number of function evaluations are needed to
find a feasible solution and even more to find a good feasible solution.

2.1.3 Objectives for Constraints

The general idea of optimizing an objective function that is subject to a
number of constraints is: deal with constraints as hard objectives and deal
with objectives as soft objectives. The hard objectives imply the following: if
the constraints are not satisfied, the solution is not feasible and the objective
value is irrelevant. One way to make sure that the constraints become satis-
fied in the optimization process is by treating the constraints as objectives
that need to be optimized as well [5]. The objective and constraints (repre-
sented by additional objective functions) become equally important this way.
The chosen optimization algorithm would now optimize the constraints and
objectives simultaneously. A disadvantage of this approach is that the opti-
mization focus lies to much on the constraints, and the problem gets much
more complex and thereby may slow down the efficiency of the algorithm
significantly.

2.1.4 Repair Infeasible Solutions

Another approach to deal with constraints is repairing infeasible solutions [4].
For some known linear boundaries of the decision variables this can be very

5
Multi-Objective Optimization Chapter 2. Earlier Research

easy by simply truncating the decision space, but for non-linear real-valued
functions of the decision variables it can be very hard. Therefore, it typically
takes a large number of function evaluations before an infeasible solution is
transformed into a feasible one.

2.1.5 Model Assisted Optimization

In the last decade, different models have been used to model the constraint
functions: Poloczek and Kramer for example used Support Vector Machines
as a classifier to predict the feasibility of solutions [29]. Powell invented a
direct search method, which models the objective and constraints using linear
approximations, better known as COBYLA [31], and recently Bagheri et al.
improved the Constrained Optimization by Radial Basis function Approxi-
mation (COBRA) of Regis [32], by making the parameters of COBRA Self
Adjusting (SACOBRA) [1]. Especialy the efficient SACOBRA solver that
makes use of the Radial Basis Function (RBF) interpolation to model the
constraints is very interesting since it is able to find high-quality results using
only few function evaluations, is precise, is fast and does not need additional
parameters. The only drawback known from SACOBRA and model assisted
optimization is that it is not very suitable for solving highly multimodal
functions in few function evaluations.

2.2 Multi-Objective Optimization

Simultaneously, optimizing objectives becomes extra challenging when the
objectives are conflicting. To deal with this problem the classical approach
would simply be: combine the objectives into one function which typically
gives rise to one optimal point. However, we want the whole Pareto optimal
set and not just one solution. A variety of approaches for finding this set, also
called the Pareto frontier, has been proposed by different researchers, and is
a very active field of research (e.g. see [6, 10, 40]). Fully automated methods
which do not need any interference of a human decision maker during the
optimization process, can generally be divided into three categories:

1. Enumerative
2. Deterministic
3. Stochastic

2.2.1 Enumerative Methods

The enumerative method is the most simple one [16]. This method simply
evaluates every possible solution in a finite design space. This method is not
very efficient since the number of possible solutions in a design space can be
very large. On the other hand, if capacity and time is available to enumerate

6
Multi-Objective Optimization Chapter 2. Earlier Research

the whole design space, this is the best solution since it can be guaranteed
that all the optimal solutions will be found.

2.2.2 Deterministic Methods

Deterministic methods are similar to the enumerative methods but they try
to include domain knowledge in such a manner that not the whole design
space needs to be evaluated [9]. Examples of a few deterministic strategies are:
greedy, hill-climbing, branch & bound, and depth and breadth-first search.
The problem remains that the search space is in most cases still to big to
evaluate all the solutions proposed by the deterministic methods.

2.2.3 Stochastic Methods

Stochastic optimization techniques are used as an alternative for enumerative
and deterministic strategies and are able to deal with complex search spaces
and multiple objectives. Stochastic search techniques typically have some
randomness involved which causes that the obtained solutions are typically
different for every run of a stochastic optimization technique. This also means
that most of the stochastic algorithms cannot guarantee to find the global
optimal solution. Examples of stochastic search strategies and algorithms are:
Non-dominated Sorting Genetic Algorithm version II [11], Stength Pareto
Evolutionary Algorithm version 2 [21], S-Metric Selection Multi-Objective
Optimization [30], and the Multi-Objective Genetic Algorithm [41].

Non-dominated Sorting Genetic Algorithm, version II

NSGA-II [11] is a classic multi-objective optimization algorithm. NSGA-II
uses a non-dominated sorting-based selection operator. This operator creates
a mating pool by combining the parent and child population to select the best
N solutions for the next generation. This selection operator makes sure that
the mating pool is well spread, the solutions in the pool have a high fitness,
and the emphasis lies on the non-dominated solutions. Finally crossover is
used to generate the new individuals from the selected parents.

Strength Pareto Evolutionary Algorithm 2

SPEA2 [21] is the second version of an evolutionary algorithm that optimizes
the Pareto frontier. This algorithm uses a fine-grained fitness assignment
strategy that is based on how many individuals each individual dominates
and is dominated by. Furthermore, a nearest neighbor density estimation
technique is incorporated which takes care of a more precise guidance of
the search process. The algorithm also makes sure that the boundaries are
guaranteed by truncation of the solutions that fall outside of the boundary.

7
Ship Design Applications Chapter 2. Earlier Research

S-Metric Selection Multi-Objective Optimization

SMS-EGO [30] is an efficient multi-objective optimization algorithm that
uses a Design and Analysis of Computer Experiments (DACE) to train
Kriging [20] surrogate models in order to efficiently optimize the objective
functions. This model assisted optimization technique utilizes the assump-
tion that close solutions are more likely to have similar objective values.
Furthermore, SMS-EGO uses the S-metric or (hyper)volume contribution [2]
to optimize the (hyper)volume between the current Pareto frontier and a
reference point. This optimization algorithm, however, does not offer a con-
straint handling technique and is again not very suitible for solving highly
multimodal functions.

Multi-Objective Genetic Algorithm

MOGA is based on the first version of the SPEA algorithm [41], where the
fitness value is again based on the number of dominated individuals. The
selection of the parents is done by tournament selection and the children are
generated by single-point crossover. Furthermore, the children have a chance
to get mutated by the so called creep mutation operator. This algorithm is
currently used in the widely used ship design software NAPA1 .

2.3 Ship Design Applications

As mentioned before in the introduction (Chapter 1), ships are tradition-
ally optimized step by step using the design spiral (Figure 1.0.1). Therefore,
most earlier research to ship designs focuses on only one part of the ship.
Consider for example the research on hull form optimization for reducing the
resistance [3, 37], or the research on scantling optimization of the midship
section [33], or the research on the dredging operations [39]. All of those typ-
ically focus on only one part of the ship. A combination of different scientific
studies is made by Papanikolaou in [26]. In this paper all components of a
ship are put together and optimized simultaneously.
A practical implementation of Papanikolaou’s method is developed by
C-Job Naval Architects2 under the name of Accelerated Concept Design
(ACD) and can be found in Figure 2.3.1. The figure contains eight crucial
components that need to be optimized and or should be designed in such a
manner that they meet criteria given by the stakeholders.

1
NAPA Oy, Release 2017.3-3 (2018), NAPA software, http://www.NAPA.fi/.
2
C-Job Naval Architects, (2018), https://c-job.com/.

8
Ship Design Applications Chapter 2. Earlier Research

Figure 2.3.1: C-Job Naval Architects Design Circle. Four different levels of accu-
racy, with eight different design aspects to consider in the optimization process.

2.3.1 Accelerated Concept Design

The ACD tool consists of several software components that are capable of
simultaneous evaluation of the ship objectives and constraints of a design
variation in an automated manner. This way, different design variations
can be computed sequentially without the interference of an engineer. The
objectives and constraints can be computed using different levels of precision.
The level of precision is typically set according to a client needs. The higher
the accuracy level, the more calculation time will be required to compute the
objective and or constraint. In the experiments conducted in this thesis, the
level of accuracy is set to level 3 except for weight and cost. The accuracy
of weight and cost is set to level 2.
A general overview of the objectives and constraints of a typical ship can
be found in the following descriptions:

Floating Position The floating position of a ship is checked to see if the

ship is in constant balance in still water and if the ship does not
heel or trim too much in a stable state. The maximum values for
these constraints are set by regulation authorities. Together with the
floating position, the maximum draught is checked to see if it meets
the stakeholders demands. Typically the maximum draught is set to a
number of meters so it can still moor in a port or sail through a sea
lock, river or canal.
Intact Stability When a ship is in normal operation configuration the ship
should meet various stability criteria. The ship, for instance, should

9
Ship Design Applications Chapter 2. Earlier Research

remain stable under all, even extreme loading conditions. Additionally,

the ship should also be capable of sailing with heavy weather conditions
without capsizing. Of course a transatlantic passenger ship has more
strict regulations compared to a simple bulk carrier that only sails
inland. Therefore, different regulations are defined per ship type.

Damage Stability After a collision, a ship should not immediately sink

or capsize. It should be capable of handling some damage to some
selected spaces/rooms of the ship. The guidelines for the locations of
the damage is defined by the IMO and other regulating authorities. Of
course the regulations differ again per type of ship.

Strength Typically when cargo gets loaded to a ship, the ship bends a bit
in the middle. This is the case because when a ship is fully loaded, the
water gives an upward pressure to the ship while the gravity on the
cargo now pushes the ship downwards in the middle, this phenomenon
is better known as sagging ((1) in Figure 2.3.2). The inverse of sagging
is named hogging ((2) in Figure 2.3.2) and mainly occurs when the
ship is empty. How much the ship is allowed to sag and hog is again
defined by the regulating authorities. The ship should be made strong
enough so it does not exceed the maximum stress limit.

Figure 2.3.2: Ship hull is sagging in (1) and hogging in (2). Note that the
bending is exaggerated.

Weight & Cost For a ship yard the weight goes hand in hand with the
cost of a ship. This is the case because the required amount of of steel
that needs to be bought and put together gives a rough indication of
how much the ship will cost to build. Therefore, the weight of a ship

10
Ship Design Applications Chapter 2. Earlier Research

is typically an objective that is to be minimized, because ship owners

want their ship to be cheap to build.

Space Reservation A space reservation should be made in such a manner

that the fuel tank, engine, pump, crew accommodation, steering hut,
and other equipment all fit in the hull or on deck. This space reservation
typically results in a lot of constraints and often changes a bit when
more details of the ship gets defined. A schematic space reservation
can be found in Figure 2.3.3.

Figure 2.3.3: Space reservation of a ship. Where every color defines a differ-
ent space.

Resistance Resistance is another objective that typically needs to be mini-

mized for a ship to be as operationally efficient as possible. This can
be achieved by changing the shape of the hull. Typically a long and
slender ship will lead to less hull resistance compared to a shorter wider
ship. Resistance is usually estimated using CFD simulations which can
be very expensive in terms of computation time.

Motions Motion deals with how the ship reacts to waves and wind. It should
not roll, pitch, yaw, heave, sway or surge (see Figure 2.3.4) too much
after a single or a sequential number of waves. Passengers, for example,
get seasick when there is too much heave/pitch and equipment can
break when there is too much rolling back and forth after a number of
sequential waves.

11
Ship Design Applications Chapter 2. Earlier Research

Figure 2.3.4: Roll, pitch, yaw, heave, sway, surge of a ship.

12
–The biggest room in the world is room for
improvement.
Helmut Schmidt

3
Ship Design Optimization
Problem

h i p d e s i g n o p t i m i z at i o n is a complex process requiring the

S successful coordination many disciplines [27]. This results from the

fact that a ship consist of a wide variety of systems and compo-
nents, for example cargo storage and handling, energy generation
and ship propulsion, accommodation of crew/passengers and ship navigation.
All these components need to be optimized in such a manner that all the
design stakeholders are satisfied, and the right balance is found between the
conflicting objectives. Before we can optimize all components, we first have
to describe the problem in more detail. In this chapter, the ship design stages
and the optimization problem are defined. For illustration purposes a case
study of a ship design optimization problem is presented.

3.1 Design Stages

Every ship starts with an initial idea from a client. After this idea has been
described in a design specification, the ship design process can begin. The
ship design process can roughly be divided into four design stages: Concept-,
basic-, functional- and detail design. In these four stages an idea of the client
is transformed into a realistic and complete design a shipyard can build. The
different stages are further described in the following subsections and are
visually presented in Figure 3.1.1.

13
Design Stages Chapter 3. Ship Design Optimization Problem

Figure 3.1.1: Four ship design stages: Concept-, basic-, functional- and detail
design.

3.1.1 Initial Idea

In the initial idea, the future ship owner defines the purpose and the physical
constraints of the ship. Different purposes of ships could for example be:
dredging, fossil fuel drilling, urban transport, cargo transport or sailing for
pleasure. Besides the purpose, the initial idea also contains the information
about whether the ship should be sea worthy or that it should only be capable
of sailing inland.
Depending on the purpose of the ship, different regulations are defined
by the regulating authorities. These regulation constraints are from now on
referenced to as domain constraints. Besides these regulations, a typical ship
also has some preference or physical constraints like: The ship should fit
trough the Suez Canal, the ship should be able to moor in the harbour of
Rotterdam, or the ship should be able to sail underneath the Tower Bridge.
These rough constraints typically can be transformed into relatively simple
box constraints for the design, such as a maximum/minimum draught, air
draught, width, length and deck hight.
Finally the future ship owner will probably also have some preferences
such as CO2 neutrality, a small building, and minimum operational costs
which are to be translated into either constraints or objectives.

14
Design Stages Chapter 3. Ship Design Optimization Problem

3.1.2 Concept Design

The concept design phase is the crucial part of the ship design process. In the
concept design phase the naval architects translate the initial idea into the
concept design of the ship. In the concept design, the following components
get defined, computed, and parameterized: the general arrangement, first
estimations regarding stability, strength, and the main cross section.
These components will define the ship’s future performance, safety and
cost. In this stage of the design process, all different components need to
be optimized and designed in such a manner that they meet all regulations
and safety criteria. This is not trivial for the following three reasons: 1) The
objectives are typically conflicting. 2) Computing the constraints and objec-
tives is very time consuming due to the required simulation time. 3) Only
little parallelism is possible due to a typically limited number of commer-
cial licences available to the ship design company. A more mathematical
representation of the optimization problem can be found in the problem in
description Section 3.2.
After the concept design phase, a ship yard can make an estimation of
how long it will take, and how expensive it will be to build the ship. A benefit
of the concept design phase of C-Job Naval Architects is that the owner of
the concept design can go to different ship yards to get the best offer.

3.1.3 Basic Design

This phase yields two things: The drawings of the design and a class approval
which proves that the ship meets the latest regulations and criteria.
In the basic design stage only a few last minute, very small design changes
take place. These small design changes lead to a ship design that meets all
regulations and safety criteria. To meet the regulations and safety criteria,
the strength, damage stability, and intact stability should for example be
within regulated bounds.

3.1.4 Functional Design

This phase is a phase in between the detail and the basic design. Some
shipyards in Asia and Southern Europe do not have enough capacity to go
from basic to detail design in one step. Therefore, some guidance to the
shipyard can be provided in the shape of a functional design phase. This
phase yields a information package to support the next phase.

3.1.5 Detail Design

In this phase, multiple block sections of the ship are created and drawn.
These workshop drawings are created so that the shipyard can follow the
instructions to create and put together all the parts of the final ship.

15
Problem Description Chapter 3. Ship Design Optimization Problem

3.2 Problem Description

Every ship design optimization problem depends on the decision variables,
constraints and the objectives of the problem. The constraints can often
be seen as hard objective functions, which need to be satisfied before the
minimization of the remaining, soft objective functions takes place [14]. A
general definition of a Multi-Objective Problem is given in Definition 1, the
definition of Pareto optimality is given in Definition 2, the definition of Pareto
dominance is given in Definition 3, and Pareto optimal sets are defined in
Definition 4, all from [6].

Definition 1.
(Multi-Objective Problem [6]) : A general Multi-Objective Problem
is defined as minimizing F (~x) = (f1 (~x), . . . , fk (~x)) subject to gi (~x) ≤ 0, i =
{1, . . . , m} and ~x ∈ Ω. Multi-Objective problem solutions minimize the
components of the vector F (~x) to find a set of Pareto optimal solutions,
where ~x is a n-dimensional decision variable vector ~x = (x1 , . . . , xn ) from
some search space Ω. It is noted that gi (~x) ≤ 0 represent constraints that
must be fulfilled while minimizing F (~x) and Ω contains all possible ~x that
can be used to satisfy an evaluation of F (~x).

Definition 2.
(Pareto Optimality [6]) : A feasible solution ~x ∈ Ω is said to be Pareto
Optimal with respect to Ω if and only if there is no x~2 for which ~v = F (x~2 ) =
[f1 (x~2 ), . . . , fk (x~2 )] dominates ~u = F (~x) = [f1 (~x), . . . , fk (~x)].

Definition 3.
(Pareto Dominance [6]) : A vector ~u = [u1 , . . . , uk ] is said to dominate
another vector ~v = [v1 , . . . , vk ] if and only if ~u is partially less than ~v , i.e.,
∀i ∈ {1, . . . , k}, ui ≤ vi ∧ ∃i ∈ {1, . . . , k} : ui < vi . The following symbol is
used to indicate Pareto Dominance: .

Definition 4.
(Pareto Optimal Set [6]) : For a given Multi-Objective Problem, F (~x),
the Pareto Optimal Set, P ∗ , is defined as:

P ∗ := {~x ∈ Ω | @~y ∈ Ω, F (~y )  F (~x)} (3.2.1)

To get some more insight in the symbols, variables, constraints, objectives,

a real world application, a dredger from C-Job Naval Architects, the largest
independent ship design and engineering company in the Netherlands, is
given in detail in the following subsections.

16
Problem Description Chapter 3. Ship Design Optimization Problem

3.2.1 Decision Variables

The decision variables of a ship design problem are the numerical quantities
for which values can be varied in the optimization process [6]. These quanti-
ties are denoted as xj , where j = 1, . . . , n, where xj represents one descision
variable. The vector ~x is then represented by:

x1
 
 .. 
~x =  .  (3.2.2)
xn
All the possible combinations in between a predefined lower and upper bound
of ~x together is called the search space denoted by Ω. Of course due to the
constraints, not every combination will lead to a feasible solution.
The dredger (Figure 3.2.1) has the following decision variables: ∆breadth ,
∆length , foreship length, hopper length extension, hopper breadth, hopper
height. Here ∆ means a change opposed to the original design.

Figure 3.2.1: Trailer Suction Hopper Dredger designed by C-Job Naval Architects,
with the design variables annotated.

The overall length and breadth of the hull can be transformed with the help
of Free Form Deformation (FFD) [34]. For this transformation a box is drawn
around the hull. Any point on the box can be moved in all directions and
the parent surface that is inside this box will be transformed accordingly.
This FFD can be achieved by changing the ∆breadth and ∆length parameter
which then applies the FFD on the original concept design.
The part of the ship from the most forward bulkhead to the front is called
the foreship. The location of this last bulkhead can be changed by varying
the foreship length decision variable.
The cargo space, where the dredged material is dumped in, is called the
hopper. Changes can be made to the height, the breadth, and to the length

17
Problem Description Chapter 3. Ship Design Optimization Problem

extension of the hopper.

The optimization problem therefore has in total 6 design variables which
is hard to visualize. To still be able to visualize the 6 dimensions, a par-
allel coordinate plot can be made. An example of such a plot is shown in
Figure 3.2.2.

Parallel Coordinate Plot

3.40 Infeasible
9.00 15.93 22.00 9.00 9.80
Feasible
2.42 Non-dominated
8.20 15.14 20.80 8.21 7.28

7.40 14.36 19.61 7.43 4.76 1.43

6.60 13.57 18.41 6.64 2.24 0.45

5.80 12.79 17.22 5.86 -0.28 -0.53

5.00 12.00 16.02 5.07 -2.80

-1.51
HOPPER HEIGHT
HOPPER EXTENSION

HOPPER WIDTH

FORESHIP LENGTH

SHIP LENGTH

SHIP BREADTH

Figure 3.2.2: Parallel coordinate plot with infeasible solutions (red), feasible so-
lutions (blue), Non-dominanated solutions (green), every vertical axes represents a
decision variable.

3.2.2 Constraints
The constraints can be expressed in terms of function inequalities, where
one function inequality represents one of the m constraints. All inequality
constraints can then be expressed the following way:

gi (~x) ≤ ci ∨ gi (~x) ≥ ci ∀ i ∈ {1, . . . , m1 } (3.2.3)

Where gi is a, generally non-linear, real-valued function of the decision vari-
ables ~x and ci is a constant value. Some special cases of constraints are
equality constraints. These equality constraints are transformed into two
approximate inequality constraints:

gi (~x) ≤ ci +  ∧ gi (~x) ≥ ci −  ∀ i ∈ {1, . . . , m2 } (3.2.4)

18
Problem Description Chapter 3. Ship Design Optimization Problem

Where  = 0.0001 and therefore so small that in practice it can be neglected.

Finally, all function inequalities are transformed without loss of generality in
such a manner that we always get the function inequalities that represents
all the constraints as presented in Equation 3.2.5.

gj (~x) ≤ 0 ∀ j ∈ {1, . . . , m} (3.2.5)

Note that it is assumed that there is at least one solution for ~x ∈ Ω that
satisfies all the constraints. If there is no such solution the optimization
process is doomed to fail.
The constraints of a ship design optimization problem can be divided
into two categories: domain constraints and practical constraints. Domain
constraints are the constraints given by the regulating authorities while the
practical constraints are typically physical constraints given by the client or
imposed by physics. In the dredger case, the constraints mainly make sure
that everything fits in the hull and that the safety constraints are taken into
account.
For the practical constraints, the space reservation for: payload, fuel tank,
engine, pump, and accommodation are checked to see if the design variation
at least meets the minimum space required.
The domain constraints: steel arrangement, hull formation, double bottom
check, location of foremost bulkhead, intact stability, draft when fully loaded,
trim, and heel are checked to see if the ship meets the recommended stability
criteria, and to see if it at least meets the prescribed rules from the regulating
authorities.
In total, after all constraints are transformed, the dredger case has sixteen
constraints, which are computed by subtracting the required minimum value
from the actual value of the design variation. When all values are negative
the ship is feasible.

3.2.3 Objectives
The objectives of a ship design optimization problem are typically conflicting
and non-commensurable. As a consequence there is usually not one unique,
perfect solution but a set of alternative, so called non-dominated solutions.
This non-dominated solution set contains good compromises between the
objective functions fj (~x), j = 1, . . . , k. Together they form the vector of
functions to be minimized:

F (~x) = [f1 (~x), . . . , fk (~x)] (3.2.6)

The non-dominated set of solutions together form the Pareto frontier, where
Pareto optimality is defined by Definition 2 and Pareto dominance is defined
by Definition 3.

19
Problem Description Chapter 3. Ship Design Optimization Problem

Note that all the objective functions are transformed to minimization

problems. A transformation of a maximization problem into a minimization
problem can easily be achieved by simply multiplying the objective function
by minus one.
The dredger case has two objectives: maximizing the performance and
minimizing the building costs. This can be achieved by minimizing the hull
resistance and steel weight. This sounds trivial, but the objectives are a
classical example of conflicting ones. A long and slender ship will lead to less
hull resistance and a higher steel weight, while a wide shorter ship will have
a higher hull resistance and a lower steel weight. These design variations of
the ship, if non-dominated by other solutions, are then design variations in
the Pareto Optimal set as described in Definition 4.
The resistance of the design variation can be estimated with a CFD
simulation. There are different types of CFD simulation methods. In the
concept phase of the dredger, a relatively simple potential flow solver [36]
is used. This approach does not take everything into account but it is very
suitable for comparing the resistance between different design variations.
An indication of the steel weight in the concept phase is calculated by first
creating the main frame scantlings for either port or starbord and then mirror-
ing it (example in Figure 3.2.3). The main frame is made strong enough such
that it does not exceed the maximum stress limit by changing the scantlings
(mainly plate thickness and profile heights). This way the maximum bending
moments can never be exceeded. The surface of the scantlings multiplied by
the length of the ship can then be used as an estimation of the steel weight.

Figure 3.2.3: Schematic cross-section view of main frame scantlings.

20
–The inspiration you seek is already within you. Be
silent and listen.
Jalāl ad-Dı̄n Muhammad Rūmı̄

4
Proposed Solution

n this chapter, a new Constrained Efficient Global Optimization(CEGO)

I algorithm is proposed, combining the strengths of both the S-metric

multi-objective optimization techniques from SMS-EGO and the con-
straint handling techniques from SACOBRA. This combination of the
two algorithms can then be used to efficiently find an approximation of the
feasible Pareto frontier of constrained multi-objective problems.

4.1 CEGO
The proposed algorithm consist of several crucial components: the initial sam-
pling, the training of surrogate models for both objectives and constraints,
defining the Pareto frontier, defining the S-Metric Selection criterion, opti-
mizing the infill criterion subject to the constraints, the evaluation of the
found local optima, and the adjustment of the allowed constraint violations.
A general overview of the algorithm is given in the CEGO psuedocode that
can be found in Algorithm 1. Additionally, a flowchart of the algorithm is
shown in Figure 4.1.1. Lastly the individual components of the algorithm
are described in detail in the description following.

21
CEGO Chapter 4. Proposed Solution

Algorithm 1 Pseudocode CEGO.

1: par := InitialSample
2: obj, con := Evaluate(par)
3: eps := 0.01
4: while eval < maxEval do
5: objectiveModels := Kriging(par, obj)
6: constraintModels := cubicRBF (P ar, Con, EP S)
7: paretoFront := paretoFrontFeasible(Obj, Con)
8: criterion := SMetricSelectionCriterion(objM odel, paretoF ront, ref )
9: x := findLocalOptima(objectiveModels, constraintModels, criterion, eps)
10: obj[evall], con[evall] := Evaluate(x)
11: eps = adjustMargins(con[evall], eps)
12: evall := evall + 1

Figure 4.1.1: CEGO flowchart.

Initial Sampling The proposed algorithm starts with an initial sampling

of the n decision variables using Latin Hypercube Sampling(LHS) [24].
The generated samples by LSH are as uncorrelated as possible. This way
the average amount of information that can be gained by computing
the objectives and constraints given by the LHS is maximized. The
recommended size of the LHS is 11 · n − 1 [22] but for problems with
a large number of parameters it can be smaller. Note that the inital
size of the LHS must at least be bigger than n + 1 to be able to build
the surrogate models [1]. The LSH samples then get evaluated for all
objectives (Eq. 3.2.6) and constraints.

22
CEGO Chapter 4. Proposed Solution

Objective Models The objective values are used to train the objective sur-
rogate models. The objective surrogate models used are Kriging [23]
(often also called Gaussian Process Regression models). For every ob-
jective function a separate Kriging model is fitted. Kriging treats every
unknown objective function f as the combination of a centered Gaus-
sian Process (x) of zero mean with an unknown constant trend µ. The
advantage of using Kriging is that in addition to the predicted mean
y(x), the predicted uncertainty, called the Kriging variance σ(x), is
provided. The Kriging variance can be exploited in the optimization
procedure (see description Criterion for optimization).
In Figure 4.1.2, 4.1.3, 4.1.4, and 4.1.5 a Kriging training process is
presented for the y = x2 function. As can be seen from the figures,
the more points we use to train the Kriging model, the smaller the
variance gets and the more precise the approximation of the kriging
model becomes.
Kriging approximation with Gaussian smoothning Kriging approximation with Gaussian smoothning
1.0 1.2

1.0
0.5
0.8
0.0
0.6
y
y

0.5
0.4
1.0 training points training points
predicted values 0.2 predicted values
actual values actual values
1.5 variance 0.0 variance
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x x

Figure 4.1.2: Kriging with Gaussian Figure 4.1.3: Kriging with Gaussian
smoothing function approximation of smoothing function approximation of
y = x2 . Training points used: x = 0.5, y = x2 . Training points used: x = −1,
and x = 1. x = 0.5, and x = 1.

Kriging approximation with Gaussian smoothning Kriging approximation with Gaussian smoothning
1.0
1.0
0.8
0.8
0.6
0.6
y

0.4
0.4
training points 0.2 training points
0.2 predicted values predicted values
actual values actual values
0.0 variance 0.0 variance
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x x

Figure 4.1.4: Kriging with Gaussian Figure 4.1.5: Kriging with Gaussian
smoothing function approximation of smoothing function approximation of
y = x2 . Training points used: x = −1, y = x2 . Training points used: x = −1,
x = 0.25 x = 0.5, and x = 1. x = 0, x = 0.25, x = 0.5, and x = 1.

23
CEGO Chapter 4. Proposed Solution

Constraint Models The corresponding constraint values are used to train

the constraint surrogate models. For the constraint surrogate models,
Cubic Radial Basis Functions (CRBF) are used. Usually, to model a
function with an CRBF model in a proper way, additional parameters
need to be tuned manually. In CEGO this is taken care of by using
the self adjusting parameter control proposed by Bagheri & co [1]. The
self adjusting parameter control is used to fit a CRBF model for every
constraint function. The steps taken to model the constraint functions
are:

1. Rescale the search space to an interval of [−1, 1],

2. Rescale the constraint functions so that they are equally impor-
tant,
3. Define the distance requirement factor (DRC) that defines how
close the solutions are allowed to be to each other, and alter it at
every iteration,
4. Adjust the margin () of allowed violation of the CRBF model at
every iteration (see description Adjust margins).

In the first few iterations, the CRBF model might not fit the constraint
function very well. Therefore, a violation of the constraints is allowed.
The magnitude of the allowed violation decreases as more feasible
solutions are found.
Now, lets assume that the objective function y = x2 is subject to
the constraint function g = −x3 − 14 where g should be smaller then
or equal to 0 to be feasible. The constraint now limits the available
ranges of values that the Kriging model is able to choose for x. In
Figure 4.1.6, 4.1.7, 4.1.8, and 4.1.9 the available choices in each step
that can be made by the Kriging model is visualized.

CRBF approximation CRBF approximation

0.75 training points 0.75 training points
actual values actual values
0.50 predicted values 0.50 predicted values
0.25 Infeasible area 0.25 Infeasible area
0.00 0.00
0.25 0.25
g

0.50 0.50
0.75 0.75
1.00 1.00
1.25 1.25
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x x

Figure 4.1.6: CRBF function approxi- Figure 4.1.7: CRBF function approxi-
mation of y = −x3 − 41 . Training points mation of y = −x3 − 41 . Training points
used: x = 0.5, and x = 1. used: x = −1, x = 0.5, and x = 1.

24
CEGO Chapter 4. Proposed Solution

CRBF approximation CRBF approximation

0.75 training points 0.75 training points
actual values actual values
0.50 predicted values 0.50 predicted values
0.25 Infeasible area 0.25 Infeasible area
0.00 0.00
0.25 0.25
g

g
0.50 0.50
0.75 0.75
1.00 1.00
1.25 1.25
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x x

Figure 4.1.8: CRBF function approxi- Figure 4.1.9: CRBF function approxi-
mation of y = −x3 − 41 . Training points mation of y = −x3 − 41 . Training points
used: x = −1, x = 0.25 x = 0.5, and used: x = −1, x = 0, x = 0.25, x = 0.5,
x = 1. and x = 1.

Pareto Frontier Feasible In the previous example y = x2 , there is only

one objective y. When more objectives and more constraints are present
we have to optimize the objectives simultaneously. We do this by search-
ing for non dominated feasible solutions. To do this, the determination
of the current feasible Pareto frontier is important. The feasible Pareto
~ is defined by the set of feasible solutions which
frontier (denoted as Λ)
are also Pareto dominant according to Definition 3.

Criterion for Optimization In the criterion for optimization, CEGO uses

the idea of Emmerich et al. to use S-metric or (hyper)volume contri-
bution [2] of a potential solution ŷpot to the current Pareto frontier
approximation. CEGO uses the S-metric extended as an infill crite-
rion [30]. The infill criterion function computes for a given input vector
~x, the expected improvement ~x contributes to the current Pareto fron-
tier approximation. The predicted objective scores ŷ are predicted with
the Kriging models together with their estimated uncertainties ŝ. If
the 95% lower confidence bound ŷpot = ŷ − α · ŝ is -dominant we
compute the additional (hyper)volume it adds to the Pareto frontier.
-dominance is applied to support a good distribution over the Pareto
frontier. The values of ~ are set every iteration:
~ − min(Λ)
max(Λ) ~
~ = . (4.1.1)
~ | − 1k · (maxEval − eval)
1+ | Λ 2

Here max(Λ) ~ is the maximum value per objective on the Pareto frontier,
min(Λ)~ is the minimum value per objective on the Pareto frontier, k is
the number of objectives, maxEval the maximum number of allowed
iterations, and eval the number of evaluations executed so far. The final
(hyper)volume multiplied by minus one that ŷpot adds to the Pareto
frontier is the score the S-metric criterion will return. If ŷpot is not

25
CEGO Chapter 4. Proposed Solution

expected to be -dominant, the infill criterion will return a penalty

value p as presented in the following Equation:
X m
Y
p= −1 + 1 + ŷpot,j − yji (4.1.2)
~
y i ∈Λ j=1

3 hypothetical cases of the S-Metric Selection infill criterion in a two

dimensional objective space is shown in Figure 4.1.10. In this figure the
blue triangles represent the non dominated solutions found in previous
iterations, the black star represents the reference point, the grey area
represents the the current (hyper)volume, and the area between the
dark grey line and the grey area represents the  area.
Suppose case 1, where we find the solution with the predicted objective
scores (1.65, 1.65). This point would be dominated by the previously
found non-dominant solutions. Therefore, the infill criterion will return
the penalty score: 0.3225. This score is computed with Equation 4.1.2.
Suppose case 2, where we find a solution with the predicted objective
scores (1.75, 1.05) which lies in the epsilon dominated area. This point
does not significantly contribute anything to the Pareto frontier so the
criterion will return 0.
Suppose case 3, where we find a solution with the predicted objective
scores (1.2, 1.2). This potential solution is not dominated by the earlier
found solutions and does not lie in the epsilon dominated area. There-
fore, the (hyper)volume that this point adds to the current Pareto
frontier is calculated. This is equal to the surface of the green area in
Figure 4.1.10. The score the infill criterion will return in in this case
−0.27 which is equal to the (hyper)volume multiplied by minus one.

S-Metric Selection infill criterion

2.0

1.8

1.6
Objective 2

1.4 Reference Point

Non Dominant Solutions
area
1.2 Dominated Solution
-Dominated Solution
1.0 Non- -Dominated Solution
S-Metric Selection Score
Current Hypervolume
1.0 1.2 1.4 1.6 1.8 2.0
Objective 1

Figure 4.1.10: S-Metric Selection infill criterion

26
CEGO Chapter 4. Proposed Solution

Find Local Optima The infill criterion is optimized in such a manner that
it searches for new optimal solutions that do not violate any of the
constraints. This can be done with an optimization algorithm that
is capable of dealing with: multiple decision variables (~x), single ob-
jective problem (S-metric criterion), and multiple constraints (CRBF
functions). In this research the Constrained Optimization by Linear
Approximation (COBYLA) algorithm [31] is used to optimize the ob-
jectives subject to the constraints.
The problem that COBYLA can solve is setup the following way: The
decision variables ~x can be changed within the predefined bounds.
The objective is the S-metric criterion that predicts the additional
(hyper)volume given ~x. The constraints can be interpolated with the
CRBF models given ~x and finally a the DRC constraint is added that
makes sure the solutions that get evaluated are not to close to the
previously found solutions.
COBYLA can deal wich such problems and therefore is able to optimize
the infill criterion under the condition that the constraints are satisfied.
The vector ~x that is expected to be feasible and expected to contribute
the most to the Pareto frontier approximation is then proposed as
new solution. If in the current design space no feasible solution can be
found, the vector ~x with the smallest expected constraint violation is
chosen.

Evaluate The minimum found by COBYLA is proposed as a new solution,

which is evaluated with the actual evaluation functions that are be-
ing optimized. This evaluation of ~x gives new objective values and
constraint values that can be added to the population. The surrogate
models are re-trained and the next iteration starts. This optimization
process goes on until the evaluation budget is exhausted. The evalua-
tion budget can be determined by the user. For simple problems the
budget could be small but it is recommended to make the budget larger
for complex problems with a lot of parameters. A guideline to take in
mind is: maxEval = 40 · n.

Adjust margins In the experiments reported in this thesis, the -value

√
used for the constrained CRBF models starts at 0.01. When b2 · nc
feasible solutions are found sequentially,  decreases by 50%. This
√
decrease of allowed violation can be justified by the fact that b2 · nc
consecutive feasible solutions are found and therefore the CRBF models
with the allowed violation have a better understanding of the constraint
boundaries so the allowed violation should decrease. Furthermore by
decreasing the allowed violation of the CRBF functions the CRBF
functions can be more exploited so solutions closer to the constrained
boundaries can be found.

27
CEGO Chapter 4. Proposed Solution

√
Alternatively,  increases by 100% when b2 · nc infeasible solutions
are found. This increase of the allowed violation can be explained by
the idea that the allowed violation of the CRBF is probably too tight
and the CRBF models are not yet capable of modelling the constraint
functions good enough yet. By increasing the allowed constraint viola-
tion margin of the CRBF models the constrained area can be explored
more thoroughly so that the fit of the CRBF models can increase. Of
course it does not make sense to explore the already known heavily
violated constrained area, therefore the maximum of  is set to 0.02 in
the experiments reported in this thesis.

28