11

Real-World Applications of Multiobjective

Optimization

Theodor Stewart1 , Oliver Bandte2 , Heinrich Braun3 , Nirupam Chakraborti4 ,

Matthias Ehrgott5 , Mathias Göbelt3 , Yaochu Jin6 , Hirotaka Nakayama7,
Silvia Poles8 , and Danilo Di Stefano8
1
University of Cape Town, Rondebosch 7701, South Africa
[email protected]
2
Icosystem Corporation, Cambridge, MA 02138 [email protected]
3
SAP AG, Walldorf, Germany [email protected]
4
Indian Institute of Technology, Kharagpur 721 302, India
[email protected]
5
The University of Auckland, Auckland 1142, New Zealand
[email protected]
6
Honda Research Institute Europe, 63073 Offenbach, Germany
[email protected]
7
Konan University, Higashinada, Kobe 658-8501, Japan [email protected]
8
Esteco Research Labs, 35129 Padova, Italy [email protected]

Abstract. This chapter presents a number of illustrative case studies of a wide

range of applications of multiobjective optimization methods, in areas ranging from
engineering design to medical treatments. The methods used include both conven-
tional mathematical programming and evolutionary optimization, and in one case
an integration of the two approaches. Although not a comprehensive review, the
case studies provide evidence of the extent of the potential for using classical and
modern multiobjective optimization in practice, and opens many opportunities for
further research.

11.1 Introduction

The intention with this chapter is to provide illustrations of real applications

of multiobjective optimization, covering both conventional mathematical pro-
gramming approaches and evolutionary multiobjective optimization. These
illustrations do cover a broad range of application, but do not attempt to
provide a comprehensive review of applications.

Reviewed by: Alexander Lotov, Russian Academy of Sciences, Russia

Tatsuya Okabe, Honda Research and Development Inc., Japan
Kalyanmoy Deb, Indian Institute of Technology Kanpur, India

J. Branke et al. (Eds.): Multiobjective Optimization, LNCS 5252, pp. 285–327, 2008.
c Springer-Verlag Berlin Heidelberg 2008
286 T. Stewart et al.

In examining the case studies presented here, it may be seen that the
applications may be distinguished along two primary dimensions, namely:
The number of objectives which may be:
• Few, i.e. 2 or 3 (impacts of which can be visualized graphically);
• Moderate, perhaps ranging from 4 to around 20;
• Large, up to hundreds of objectives
The level of interaction with decision makers, i.e. the involvement of policy
makers, stakeholders or advisers outside of the technical team. The level
of such interaction may be:
• Low, such as in many engineering design problems (but for an excep-
tion, see Section 11.7) where the analyst is part of the engineering
team concerned with identifying a few potentially good designs;
• Moderate, such as in operational management or interactive design
problems where solutions may need to be modified in the light of pro-
fessional or technical experience from other areas of expertise;
• Intensive, such as in public sector planning or strategic management
problems, where acceptable alternatives are constructed by interaction
between decision makers and other stakeholders, facilitated by the an-
alyst.
Not all combinations of number of objectives and level of interaction may
necessarily occur. For example, the public sector planning or strategic man-
agement problems which require intensive interactions, also tend often to be
associated with larger numbers of objectives. In the case studies reported in
this chapter, we have attempted to provide summaries of a number of real
case studies in which the authors have been involved, and which do between
them illustrate all three levels for each dimension identified above. Table 11.1
summarizes the placement of each of the cases along the above two dimensions.
These studies exemplify the wide range of problems to which multiobjec-
tive optimization methods can and have been applied. Half of the case studies
deal with engineering design problems, which is clearly an important area
of application, but even within this category there is a wide diversity. For

Table 11.1. Classification of case studies

Number of Level of Case study
Objectives Interaction
Few Low Aerodynamic design (Section 11.2)
Industrial neural network design (Section 11.3)
Molecular structures for drugs (Section 11.4)
Few Moderate Medical decision making (Section 11.5)
Supply chain management (Section 11.6)
Moderate Moderate Interactive aircraft design (Section 11.7)
Moderate Intensive Land use planning (Section 11.8)
Large Low Lens and bridge designs (Section 11.9)
 11 Real-World Applications of Multiobjective Optimization 287

example, we have two examples from aircraft design, but one (Section 11.2)
focuses on the trade-off between robustness and cost in aircraft design, while
the other (Section 11.7) deals with the need to provide a broad holistic in-
teractive decision support to aircraft designers. Although both applications
relate to aircraft design, the issues raised are substantially different so that
different sections in this chapter are devoted to each of them.
Other applications range over operational management of supply chains,
effective treatment of cancers and conflicts between environmental, social and
economic factors in regional planning.
A perhaps less usual application is that described in Section 11.4. Here
the multiobjective optimization methods are applied not directly to design,
operational or strategic decisions, but to the development of understanding of
molecular processes in synthesizing drugs.

11.2 Aerodynamic Design Optimization

11.2.1 Problem Description

Although the number of objectives are typically low (two or three) if the ge-
ometrical constraints are not counted, aerodynamic design optimization is a
challenging engineering task for a number of reasons. Firstly, aerodynamic
optimization often needs to deal with a large number of design parameters.
Secondly, no analytical function is available for evaluating the performance of
a given design, and as a result many gradient-based optimization techniques
are inapplicable. Thirdly, to evaluate the quality of designs, either computa-
tionally expensive computational fluid dynamics (CFD) simulations have to
be performed or costly experiments have to be conducted. Finally, aerody-
namic optimization involves multiple disciplines and more than one objective
must be considered.
In recent years, evolutionary algorithms have successfully been applied to
single and multiobjective aerodynamic optimization (Obayashi et al., 2000;
Olhofer et al., 2000; Hasenjäger et al., 2005). Despite the success that has
been achieved in evolutionary aerodynamic optimization, several issues must
be carefully addressed.

11.2.2 Methodology

Geometric Representation

Finding a proper representation scheme is the first and most important step
toward successful optimization of aerodynamic structures. A few general cri-
teria can be mentioned for choosing an appropriate geometric representation.
Firstly, the representation should be sufficiently flexible to describe highly
complex structures. An overly constrained representation will produce only
288 T. Stewart et al.

suboptimal results. Secondly, the representation should be efficient, which

means that the flexibility of representation can be achieved with a minimum
number of free parameters. Inefficient representations may result in an unnec-
essarily large search space, which reduces the search efficiency of evolutionary
algorithms. Finally, the representation should makes it possible to perform a
local search. This requirement is important for refining the performance and
for reducing search space.
Several methods are available for the geometric representation, such as
B-Splines, Bezier curves, and T-Splines. Furthermore, constrained deforma-
tion instead of global deformation techniques can be used. An example of
constrained deformation is free form deformation or simplified constrained
deformation.
Nevertheless, it can happen that no single representation is able to satisfy
all the above-mentioned properties. To solve this problem, adaptive represen-
tation techniques can be used. The basic idea of adaptive representation is to
start the optimization with a relatively compact representation, so that the
global search can be conducted first. After that, the number of search pa-
rameters can be increased, e.g., by inserting new control points in a B-Spline
based representation. Encouraging results have been reported where an adap-
tive representation for evolutionary optimization of turbine blades has been
adopted (Olhofer et al., 2001).
However, adaptive representation is not as straightforward as it appears.
On the one hand, it is not trivial to establish when a new point should be
inserted, or removed. On the other hand, the insertion or removal of a search
point should be neutral to the fitness value. Moreover, an adaptation in repre-
sentation may degrade the search performance of evolutionary algorithms, for
example, for evolutionary strategies with a small population size (Jin et al.,
2005).
In multiobjective optimization, even more complex situations can occur.
For example, it has been found that in micro heat-exchanger optimization
more than one representation is needed to achieve the whole Pareto front (Ok-
abe et al., 2003).

Reduction of Computational Cost

Evolutionary algorithms (EAs) acquire strong search power at the cost of

search efficiency. In contrast to gradient-based search methods, EAs often
need a large number of quality evaluations to achieve acceptable solutions.
This poses serious problems in aerodynamic optimization where each qual-
ity evaluation is costly. For example, a full three-dimensional CFD simulation
may takes several hours on a high-end computer. To reduce the computational
time for evolutionary optimization aerodynamic structures, the following ap-
proaches are adopted. Firstly, efficient and scalable evolutionary algorithms
need to be developed. An efficient and scalable EA should be able to converge
 11 Real-World Applications of Multiobjective Optimization 289

to an acceptable non-dominated front within a small number of fitness evalua-

tions, and should be insensitive to the increase in search dimension. Secondly,
both the CFD simulations and the fitness evaluations should be parallelized.
A further step is to take advantage the grid computing techniques that enable
us to use available computational resources as efficiently as possible. In addi-
tion, computational efficiency can further be improved if EAs are adapted to
the parallel or grid based computing hardware architecture. Finally, computa-
tionally expensive full simulations can be replaced partly by computationally
more efficient reduced simulations, or surrogates. In recent years, metamodel-
ing techniques have been extensively investigated to reduce the computational
cost in evolutionary optimization of expensive problems, both for single ob-
jective (Jin et al., 2002) and multiobjective optimization (Emmerich et al.,
2006). Refer to Chapter 10 on metamodeling for a more detailed discussion
of their use in multiobjective optimization.
Metamodels can introduce errors in quality evaluation, which may lead
to the convergence to a false minimum (Jin et al., 2000). Therefore, a major
concern in using surrogates in evolutionary optimization is to reduce compu-
tational cost as much as possible without misleading the evolutionary algo-
rithm. To this end, the meta-model should be properly interleaved with the
original fitness function, which is known as evolution control or model manage-
ment (Jin et al., 2002). A meta-model can be employed in different operations
of EAs, such as population initialization, crossover or mutation, evolutionary
fitness evaluation, or local search combined with evolutionary search. Different
model management techniques have been suggested. In the individual-based
techniques, all individuals in the current generation are evaluated with the
metamodel. Then, the most promising solutions according to the metamodel
are re-evaluated using the original fitness function. In a generation-based evo-
lution control framework, the meta-model is used for fitness evaluation in
some of the generations, and the frequency at which the meta-model is used
can be adjusted according to the fidelity of the model (Jin et al., 2002). If a
metamodel is employed in local search, the trust-region framework (Alexan-
drov et al., 1998) can be adopted. A comprehensive survey of techniques for
using meta-models (surrogates) in evolutionary optimization can be found
in Jin (2005).

Robustness Considerations

In aerodynamic optimization, uncertainties in the environment must be taken

into account. For example, the Mach number may deviate from the normal
condition during the flight. In this case, a robust optimal solution is very
much desired. By robustness, it is meant in general that the performance of
an optimal solution should be insensitive to small perturbations of the design
variables or environmental parameters. In multiobjective optimization, the
robustness of a solution can be an important factor for a user in choosing the
final solution. Robust solutions can be achieved in evolutionary optimization
290 T. Stewart et al.

by a number of means. One simple approach is to add perturbations to the

design variables or environmental parameters before the fitness is evaluated,
which is known as implicit averaging, see e.g., Tsutsui and Ghosh (1997). An
alternative to implicit averaging is explicit averaging, which means that the
fitness value of a given design is averaged over a number of designs gener-
ated by adding random perturbations to the original design. One drawback of
the explicit averaging method is the number of additional quality evaluations
needed, which may make the approach impractical. Partly to solve this prob-
lem, metamodeling techniques have been considered (Ong et al., 2006; Paenke
et al., 2006). A slightly different approach is to find the solution with the maxi-
mal allowed deviation given the allowed performance deterioration (Lim et al.,
2007). One potential advantage of this methods is that no assumptions need to
be made concerning the noise distribution (as needed in the averaging based
approaches).
Search for robust solutions can also be treated as a multiobjective task, i.e.,
to maximize the performance and the robustness simultaneously. These two
tasks are very likely conflicting, and therefore, Pareto-based multiobjective
methods can be employed to find a number of trade-off solutions. Refer to Jin
and Branke (2005) for a more detailed discussion on evolutionary search for
robust solutions.

11.2.3 An Example

We present here an example of evolutionary multiobjective optimization of

a three-dimensional (3D) turbine stator blade used in gas turbines. Two ob-
jectives are taken into account in the optimization. The first objective is the
average pressure loss, which indicates the energy efficiency of the blade. The
second objective, as suggested in Hasenjäger et al. (2005), is the variation of
the pitch-wise static outlet pressure.
The 3D geometry of the blade is represented by two sections of closed cubic
B-splines, namely, a hub section and a tip section, each consisting of 25 control
points, as illustrated in Fig. 11.1. In the representation, the first three and
the last three control points of the closed B-splines are overlapping, resulting
in 22 control points. In addition, since the hub and tip sections are supposed
to lie on a cylindrical surface, the z-coordinate of the control points is fixed
and not optimized during the evolution. As a result, 88 design parameters in
total (x and y coordinates of 22 control points) are to be optimized by the
evolutionary algorithm.
To evaluate the performance of a given design, 3D CFD simulations have to
be performed. In our work, a 3D Navier-Stokes flow solver, HSTAR3D (Arima
et al., 1999) is employed, which usually takes from two to four hours on
an AMD Opteron 2 GHz double processor depending on the convergence
speed of the fluid dynamics. To reduce computation time, a two-level parallel
computing architecture has been adopted. At the first level, fitness evaluations
for each individual in the population are parallelized using the master-slave
 11 Real-World Applications of Multiobjective Optimization 291

Fig. 11.1. Three-dimensional representation of a blade using B-splines.

model. At the second level, each CFD simulation is again parallelized on four
computers using the node-only model. Consequently, if the population size is
P , the needed number of computing nodes will be 4P +1.
An efficient model-based evolutionary multiobjective optimization algo-
rithm, the regularity modeling multiobjective estimation of distribution algo-
rithm (RM-MEDA) (Zhang et al., 2008), has been employed for the optimiza-
tion of the 3D turbine blade. RM-MEDA is in principle a variant of estimation
of distribution algorithms (EDA) (Larranaga and Lozano, 2001). Instead of
using Gaussian models, a first-order principle curve has been used to model
the regularly distributed Pareto-optimal solutions complemented by a Gaus-
sian model. As demonstrated in Jin et al. (2008), by modeling the regularity
in the distribution of Pareto-optimal solutions, the scalability of the EDA can
be greatly improved. Furthermore, unlike most EDAs, which require a large
population size, RM-MEDA performs well even with a small population size.
In this example, a population size of 20 has been used.
The optimization results from two independent runs are plotted in Fig.
11.2, in each of which the population has been evolved for 100 generations.
Note, however, that the population was initialized not randomly, but with
solutions from previous optimization runs using weighted aggregation ap-
proaches (Hasenjäger et al., 2005). Compared to the results reported in Hasen-
jäger et al. (2005), we see that the non-dominated solutions obtained by the
RM-MEDA are better in terms of both coverage and accuracy.
292 T. Stewart et al.

40.000

35.000

30.000
PS variation

25.000

20.000

15.000

10.000

5.000
9 9.5 10 10.5 11 11.5 12 12.5 13
pressure loss
Fig. 11.2. Solutions obtained from two independent evolutionary runs. The dia-
monds denote the solutions from the first run, and the squares from the second
run.

11.3 Design of a Neural Network for an Industrial Blast

Furnace
11.3.1 Problem Description

Intuitively one can conceive of a neural network that is simultaneously at-

tempting to satisfy two different requirements: it should be able to reproduce
the data in an accurate manner and simultaneously, it should engage a small
number of neural connections, primarily to avoid over-training of the data set.
These two objectives in many cases could be conflicting with each other. As
expected, as the number of nodes becomes smaller, the training error tends to
shoot up, and the converse usually remains true as well. In term of these two
objectives, one can thus think of working out a Pareto tradeoff, where each
solution in the Pareto frontier denotes a neural network of a unique archi-
tecture with a unique set of weights. A procedure for evolving such frontiers
through a Predator-Prey type multi-objective Genetic Algorithm (Li, 2003)
has been demonstrated in a recent article (Pettersson et al., 2007a) and was
elaborately tested on the highly nonlinear data from an industrial iron making
blast furnace shown schematically in Figure 11.3.
The aim of the study reported in Pettersson et al. (2007a) was to evolve
a neural network that would optimally predict the carbon, sulfur and silicon
contents of the hot metal produced in the blast furnace as a function of a
number of process parameters. What was attempted there was simultaneously
to minimize (i) the training error of the network (E) and (ii) the required
 11 Real-World Applications of Multiobjective Optimization 293

number of active connections in the lower part of it (N ). The idea was to

tinker with the architecture of the lower part of the network, and to treat
their corresponding weights as variables influencing the objective functions,
as further elaborated in Figure 11.4. The trade off situation between E and
N is expected to be represented as a Pareto-frontier.

11.3.2 Methodology

The evolutionary process: Here the crossing over was done between two enti-
ties both of which are essentially self-sustaining neural networks. This process
is further elaborated in Figure 11.5. A self-adaptive real-coded mutation was
performed on the weights, which draws its inspirations from Differential Evo-
lution (Price et al., 2005).
The multi-objective algorithm used in this study utilized a Moore neigh-
bourhood inhabited by two distinct species: the predators and the preys. The
preys are a family of sparse neural networks, initiated randomly as a popu-
lation, and they evolved in the usual genetic way. The members of the prey
population differed from each other both by the topology of the lower part
connections and the corresponding weight values. The predators in this algo-
rithm are a family of externally induced entities, which do not evolve, and the
major purpose of their presence is to prune the prey populations based upon
the fitness values. A two dimensional lattice was constructed as a computa-
tional space and both the predators and the prey were randomly introduced
there, where each of them would have its own neighbourhood. The basic idea
propagated in this algorithm inherits some of the concepts of cellular au-
tomata in Moore’s neighbourhood. However, unlike cellular automata, here
the lattice here does not denote the discretized physical space; it is just a
mathematical construction that facilitates a smooth implementation of this
algorithm. Further details are available in the original work (Pettersson et al.,
2007a).
The method seems to have worked better when the initial population is de-
liberately generated in the vicinity of the estimated nadir region. The progress
of the rank-one members is captured in Figure 11.6 and a computed Pareto
frontier is shown in Figure 11.7. Each discrete point in the frontier denotes a
neural net with a different ability of prediction than the others. Some typi-
cal examples are shown in Figure 11.8. As the ultimate choice between them
remains the task of the decision maker, the conservative middle ground ‘B’
shown in Figure 11.7 should be adequate for most applications.
This novel method of multi-objective analysis is not just to benefit the
steel industries: basically it is robust enough to handle noisy data irrespective
of their sources. Very recently this methodology has been augmented further
through the use of Kalman filters (Saxén et al., 2007), and it has also been
effectively utilized for identifying the most important in-signal in a very large
network (Pettersson et al., 2007b), rendering it of further interest to the soft
computing researchers at large.
294 T. Stewart et al.

Fig. 11.3. Schematic diagram of an iron making blast furnace.

Fig. 11.4. Multi-objective formulation of the neural network.

Fig. 11.5. The crossover scheme. The shaded regions are participating in the
crossover process.
 11 Real-World Applications of Multiobjective Optimization 295

Fig. 11.6. Movement of rank 1 population in different generations.

Fig. 11.7. A typical Pareto frontier presented in (Pettersson et al., 2007a).

296 T. Stewart et al.

Fig. 11.8. Data prediction through three networks A (top), B (middle) and C
(bottom). The lighter lines denote actual observations for a period of 200 days and
the darker lines are the predicted values provided in (Pettersson et al., 2007a).

11.4 Molecular Docking

11.4.1 Problem Description

The docking of a highly flexible small molecule (the ligand) to the active site of
a highly flexible macromolecule (the receptor) is described in this Section. See
Morris et al. (1998); MacKerell Jr. (2004) for a more detailed discussion of the
problem. The ability to predict the final docked structure of the intermolecular
complex has a great importance for the development of new drugs as docking
modifies the biological and chemical behavior of the receptor. Most of the
current docking methods account only for the ligand flexibility and consider
the receptor as a rigid body because the inclusion of receptor flexibility could
involve thousands of degrees of freedom. Current research in this field is faced
with this problem. The application described here focuses on a different aspect
of the docking procedure: the optimization methodology applied to find the
best docked structure. The application of a multi-objective approach to the
docking problem based on the Pareto optimization of different features of a
docked structure is proposed. It is shown that this approach allows for the
identification of the dominating interactions that drives the global process.
A drug performs its activity by binding itself to the receptor molecule,
usually a protein. In their bounded structure, the molecules gain complemen-
tary chemical and geometrical properties that are essential for the therapeutic
 11 Real-World Applications of Multiobjective Optimization 297

function of the drug. The computational process that searches for a ligand that
best fits the binding site of the receptor from a geometrical and chemical point
of view is known as molecular docking.
A molecule is represented by its atoms and the bonds connecting them.
Atoms are described mainly by their Van der Waals radius that roughly de-
fines their volume; bonds are described by their lengths (the distance between
atoms), by the angle between two consecutive bonds and by their conforma-
tional state (the dihedral angle between three consecutive bonds). Molecules
are not static systems. At room temperature they perform a variety of motions
each one having a characteristic time scale. Since the time scales of stretching
(changes in bond lengths) and bending (changes in bond angles) have greater
time scales than conformational motions (changes in dihedral angles), bond
lengths and bond angles can be considered fixed. Thus, from the docking point
of view, only conformational degrees of freedom are important.
Typically, ligands have from 3 to 15 conformational degrees of freedom;
their values define the conformational state of the ligand. Receptors have typ-
ically from 1000 to 3000 conformational degrees of freedom, so the dimension
of the complete search space for best docked conformation becomes compu-
tationally unaffordable even for routine cases. The most widely used simplifi-
cation is to consider only the ligand flexibility, so reducing the complexity of
the search space.
The different possible ligand conformations are ranked according to their
fitness with the receptor. What this fitness stands for is one of the key aspects
of molecular docking and differentiates various docking methodologies. Most
of the docking fitness functions are based on the calculation of the total en-
ergy of the docked structure. Energy based fitness functions are built starting
from force fields which represent a functional form of the potential energy of a
molecule. They are composed of a combination of different terms that can be
classified in bonded terms (regarding bond energies, bond angles, bond con-
formations) and non-bonded terms (Van der Waals and electrostatic). This
energy can be calculated in various ways, ranging from quantum mechanics
to empirical methods. Obviously, a more “exact” fitness function as derived
from quantum mechanical simulations strongly impacts on the computational
complexity and is applicable only for small systems on massive parallel com-
puters; the opportunity to use “rough” empirical models creates the possibility
of treating more realistic cases.
In summary, a docking procedure is composed by two main elements: a fit-
ness function to score different conformations for the molecular complex and
a search procedure to explore the space of possible conformations. In current
docking approaches, the bonded and non-bonded terms both contribute to
the fitness function and the optimization has a single objective equal to their
weighted sum. The weights are determined by statistical analysis of exper-
imental data. The proposed multi-objective optimization approach incorpo-
rates two conflicting objectives, i.e. the concurrent minimization of internal
298 T. Stewart et al.

and intermolecular energy terms, derived from a suitable scoring function,

each one corresponding to an objective for the optimization algorithm.

11.4.2 Methodology
MOGA-II
MOGA-II is an improved version of the MOGA (Multi-Objective Genetic Al-
gorithm) of Poloni (Poloni and Pediroda, 1997). It uses smart multi-search
elitism for robustness and directional crossover for fast convergence. The ef-
ficiency of MOGA-II is controlled by its operators (classical crossover, di-
rectional crossover, mutation and selection) and by the use of elitism. The
internal encoding of MOGA-II is implemented as in classical genetic algo-
rithms. Elitism plays a crucial role in multi-objective optimization because it
helps preserving the individuals that are closest to the Pareto front and the
ones that have the best dispersion. MOGA-II uses four different operators for
reproduction: one point crossover, directional crossover, mutation and selec-
tion. At each step of the reproduction process, one of the four operators is
chosen with regard to the predefined operator probabilities.
A strong characteristic of this algorithm is the directional crossover that
is slightly different from other crossover operators and assumes that a direc-
tion of improvement can be detected comparing the fitness of individuals. A
novel operator called evolutionary direction crossover is introduced and it is
shown that even in the case of a complex multi-modal function this operator
outperforms classical crossover. The direction of improvement is evaluated by
comparing the fitness of the individual Indi from generation t with the fit-
ness of its parents belonging to generation t − 1. The new individual is then
created by moving in a randomly weighted direction that lies within the ones
individuated by the given individual and his parents.

Multi-objective Ligand-Receptor Docking

In this example, the MOGA-II implementation in modeFRONTIER R is used
to optimize the docking towards each of the different contributions of the
docking program Autodock v. 3.05 (http://autodock.scripps.edu) scoring
function:

∆G = CCVDW ∆GGVDW + Chbond ∆Ghbond + Celec∆Gelec

+ Ctor Ntor + Csol ∆Gsol (11.1)
that tries to estimate the change in Gibbs free energy G involved in passing
from the system (receptor + ligand) to the docked system (receptor-ligand).
The coefficients C are parametrized from experimental data and set to proper
values. VDW stands for Van der Waals contribution, hbond for hydrogen
bonds contribution, elec for electrostatic contributions, tor for the entropy
change if Ntor rotatable bonds are connected with heavy atoms, and sol for
the change in solvation energy.
 11 Real-World Applications of Multiobjective Optimization 299

11.4.3 Results

A “bound docking” experiment was performed: on the basis of the x-ray struc-
ture of the complex, the receptor coordinates were separated from those of the
ligand, and then an attempt made to reconstruct the original x-ray structure
by docking the ligand to the receptor. Starting from the scoring function of
equation (11.1), Autodock gives the values for the internal energy of the lig-
and and for the intermolecular ligand-receptor interaction energy. These two
outputs were assigned as the objective of the optimization.
The tests were conducted on PDB code 1KV3 chain A co-crystallized with
GDP (http://www.rcsb.org/pdb). The resulting Pareto front is reported in
Figure 11.9 (in which the units for the axes are Kcal/mol).

Fig. 11.9. Pareto frontier for the molecular docking problem. Energies are in
Kcal/mol. Boxed values represent RMSD in Angstrom between the candidate solu-
tion and the original x-ray structure.

The squared values represent the root mean squared deviation (RMSD) in Å
(angstrom) between the candidate solution and the original x-ray structure.
Typically, RMSD values less than 1.5 Å are considered as good solutions.
It is possible to note that in this case the docking process is mainly driven
by the intermolecular energy. This information could be useful for a deeper
understanding of the effective relative influence of the contributions of the
scoring function to this particular docking process. From a practical point of
view, it could also be useful for the design of a tailored scoring function for the
300 T. Stewart et al.

docking of similar drug candidates. Also note the presence of a “knee” point
(RMSD=1.27 Å). This is a particularly interesting solution of the docking
problem in which a small improvement in the minimization of the ligand
energy leads to a large deterioration of the intermolecular energy.

11.5 Radiotherapy Treatment Planning

11.5.1 Problem Description

Cancer is one of the most significant health problems worldwide. In industri-

alised countries it is the second most common cause of death and more than 5
in every 1,000 people are diagnosed with some type of cancer every year. The
main treatment form besides surgery and chemotherapy is radiation therapy.
It is estimated that 50% of all patients diagnosed with cancer would currently
benefit from radiotherapy.
Ionising radiation is used to damage the DNA and interfere with division
and growth of cancer cells. Radiation therapy exploits the fact that cancerous
cells are more susceptible to radiation than healthy cells. The goal of radio-
therapy treatment planning is therefore to ensure that enough radiation is
delivered to the targeted region to kill the cancerous cells while surrounding
anatomical structures are spared.
In the past it was possible for a physician manually to design a treatment
that took full advantage of the available technology. Modern procedures use a
technique called Intensity Modulated Radiotherapy (IMRT). This technique
uses a multileaf collimator to shape the beam and control, or modulate, the
intensity that is delivered along a fixed beam direction. IMRT allows patients
to receive complicated treatments and the number of options that are avail-
able in IMRT places the optimal planning of a treatment outside the realm of
human awareness. Because of this complexity of the planning process, treat-
ment planning is segmented into a three-phase process that first selects beam
directions then decides an intensity map (exposure times, fluence) for the di-
rections selected in phase one, and finally chooses a delivery sequence that
efficiently administers the treatment. Computer assisted optimisation meth-
ods are needed in each phase and since the end of the 20th century these
problems have attracted the interest of the Operations Research community.
Surveys on optimisation methods for the three phases can be found in Ehrgott
et al. (2008a); Shao (2005); Ehrgott et al. (2008b), respectively. In the follow-
ing we will concentrate on the intensity optimisation problem and assume that
beam directions are given.
In 2000, we started a collaboration with the Physics Section of the Oncol-
ogy Department at Auckland City Hospital to work on treatment planning
problems. Treatment planners spend between 30 minutes and several hours on
one single case. This is because the available planning system (like almost all
commercially available ones worldwide) requires a trial and error approach.
 11 Real-World Applications of Multiobjective Optimization 301

Apart from desired dose levels in the tumour and surrounding critical struc-
tures, so called “importance factors” for these entities need to be specified as
input. The software then employs heuristics to find a good treatment plan,
which is presented to the planner. If it is unsatisfactory the importance factors
have to be changed and the process will be repeated. Treatment planners are
aware of the inefficiency of this approach. So the goal was to investigate the
possibility of a planning system that would calculate several plans right away
and provides decision support for choosing an appropriate one.

11.5.2 Methodology

Mathematical models for the intensity optimisation problem are based on the
discretisation of the body and the beams. The body is divided into volume
elements (voxels) represented by dose points. Voxels are cubic and their edge
length is defined by the slice thickness and resolution of the patient’s CAT
images and is in the range of a few mm. at most. Deposited dose is calculated
for one dose point in every voxel and assumed to be the same throughout
the voxel. A beam is discretised into beam elements (bixels). Their size is
defined by the number of leafs of the collimator and the number of stops for
each leaf. The number of voxels may be tens or hundreds of thousands and
the number of bixels can be up to 1,000 per beam. The relationship between
intensity and dose is linear, i.e., d = Ax where x is a vector of bixel intensities.
The entries aij of A represent the rate at which dose is deposited in voxel i
by bixel j. Finally, d is a dose vector that represents the discretised dose
distribution in the patient. The computation of the values aij is referred to
as dose calculation.
While most optimisation models in the medical physics literature have a
single objective, they do try to accommodate the conflicting goals of destroying
tumour cells and sparing healthy tissue. Almost all can be interpreted as
weighted sum scalarisations of multi-objective programming models, where
the weights are the importance factors mentioned above. Almost all of these
multi-objective models are convex problems, so that their efficient sets can be
mapped to one another. We decided to use a multi-objective version of the
model of Holder (2003), which has some nice mathematical properties. Here
A is decomposed by rows into AT , AC , and AN depending on whether a voxel
belongs to the tumour, critical structures, or normal tissue. Accordingly, T U B
and T LB are vectors of upper and lower bounds on the dose delivered to the
tumour voxels; CU B is a vector of upper bounds for the critical structure
voxels; and N U B a vector of upper bounds for the remaining normal tissue
voxels. The objectives of the model are to minimise the violation of any of the
lower and upper bounds and can be stated as shown in (11.2). αU B, βU B,
and γU B are parameters to restrict the deviations to clinically relevant values.
302 T. Stewart et al.

min{(α, β, γ) : T LB − αe ≦ AT x ≦ T U B, AC x ≦ CU B + βe,
AN x ≦ N U B + γe, 0 ≦ α ≦ αU B,
min CU Bi ≦ β ≦ βU B, 0 ≦ γ ≦ γU B, 0 ≦ x}, (11.2)
i

where e denotes a vector of ones of appropriate dimension.

We will denote the feasible set of (11.2) by X and its image in the objec-
tive space by Y . In (11.2) we have a multi-objective linear programme with
three objectives, a large number of variables (order of thousands), and a very
large number of constraints (order of hundred thousands). Under these cir-
cumstances we never did try to solve the problem with simplex methods, as
it is known that the number of efficient basic feasible solutions can be very
large, even for moderately sized problems. Moreover, treatment planners will
never use the intensity maps to decide on a treatment, but always look at the
dose distribution. The obvious choice was to try and solve the problem in the
three-dimensional objective space.
To that end Benson’s outer approximation algorithm (Benson, 1998) was
implemented. With this method 2D planning problems (i.e. on a single CAT
slice) could be solved, but the experiments indicated that 3D problems would
require prohibitive computation times. It was therefore necessary to consider
approximate solution of the problem. Discussions with physicists on whether
that would be acceptable from a radiotherapy point of view provided valu-
able insights. We discovered that dose calculation is imprecise because the
mathematical models to compute the entries of A are inexact since they can-
not exactly capture the specific tissue composition in individual patients. The
medical physicists assured us that it is acceptable to work with precision of
about 0.1 Gy (Gy, for Gray, is the physical unit for radiation dose).
This allowed us to consider ε-efficient solutions of (11.2). It was possible to
adapt Benson’s algorithm in such a way that it does guarantee the construc-
tion of an ε-nondominated set, the modified algorithm is described in Shao
and Ehrgott (2008). Solving the problems approximately reduced the compu-
tation times dramatically. Figure 11.10 (a) and (b) show the ε-nondominated
set of a 2D problem with 986 voxels and 1140 bixels. For ε = 0.1 the prob-
lem had 152 nondominated extreme points and was solved in 20 minutes. For
ε = 0.005 it took 9 hours to compute 1,989 nondominated extreme points.

11.5.3 Interactive Scheme

From the planners’ point of view the whole set of nondominated points is not
very useful, since it is infinite. Also, for the same reason of imprecision in dose
calculation mentioned above, planners would not distinguish between plans if
they differ only by very small amounts. It is necessary to select a finite set of
nondominated points (efficient solutions). The nondominated extreme points
and associated basic solutions have only mathematical relevance, but no clini-
 11 Real-World Applications of Multiobjective Optimization 303

100 100

90 90 100

80 80
80
γ

70 70

γ
60

γ
60 60

50 40
50
−20 0 0
5

0 10

15 −40
10
−40 −20
−20 0
α 20 20 0 20 20
β −40
α β α −20
20 0
20
β

(a) (b) (c)

Fig. 11.10. ε-nondominated set with ε = 0.1 (a) and ε = 0.005 (b) and set of
representative nondominated points (c).

cal meaning. The selection of plans should represent the whole nondominated
set, but guarantee a certain minimal difference between the points. We devel-
oped a method to determine a representative subset of nondominated points
in Shao and Ehrgott (2007). The method first constructs an equidistant lat-
tice of points placed with distance d on a simplex S (the reference plane) that
supports Y at the minimiser of eT y over Y and such that Y ⊂ S + R3≧ . For
each lattice point q an LP

min{t : q + te ∈ Y, t ≧ 0}

is solved. If the optimal value is t̂, the point q + t̂e is tested for nondominance.
It can be shown√that the distance between remaining nondominated points is
between d and 3d. Figure 11.10 (c) shows a representative set for the same
example shown in Figure 11.10 (a) and (b).
Since the representative points are all nondominated the planners now
have a choice between several plans. By the theory of linear programming,
we know that they are all optimal solutions of some weighted sum problem
using importance factors as used in current practice. Moreover, the whole
range of such solutions is represented. To support planners in the choice of
a plan, visual aids are necessary within a decision suppor system. Planners
are used to judging the quality of a plan by looking at isodose curves and
dose volume histograms (DVH). The former are colour-wash pictures showing
curves of equal dose superimposed on CAT pictures. The latter are plots
of the percentage of tumour and critical structures against dose levels, see
Figure 11.11.
The representative set of solutions (treatment plans) is stored in a database
and input to the software Carina (Ehrgott and Winz, 2008) which first pro-
poses a balanced solution of (11.2) (with as equal as possible values of α, β, γ)
displaying the corresponding DVH and isodose plots as well as some informa-
tion on available trade-offs. The planner can then specify changes (going to a
neighbouring solution, searching for solutions with specific values, or for solu-
304 T. Stewart et al.

Fig. 11.11. Isodose curves and dose-volume histogram for a brain tumour treatment
plan.

tions satisfying some thresholds). This process is continued until the planner
accepts a treatment plan.
The interaction with the treatment planner is therefore ex-post, allowing
the likely time consuming plan calculation to be decoupled from plan selection.
As a consequnce, plan selection becomes faster as it is based on information
retrieval from databse, a real-time operation. Moreover, the specification of
dose levels is more natural than the “guessing” of importance factors.

11.5.4 Remarks

It is interesting to note that the optimisation model (11.2) tries to characterise

dose distribution by a few numbers, whereas the quality is actually judged by
the whole DVH. This, of course, is not part of the model. Attempts to specify
some points on the DVH curves as constraints in optimisation models exist.
But they lead to mixed integer programming models that at this time cannot
be solved as multi-objective models.
Throughout the project radiotherapists have been involved in the project.
This had several advantages. We obtained valuable information on radio-
therapy practice and could ensure to develop usfeul tools that would be ac-
cepted by the actual users. Fears that we intend to replace people by software
could easily be laid to rest once the radiotherapists understood that we never
thought it is possible to replace their role in the treatment design, but that
we could improve the planning process. Finally, we made sure to use the tools
they are accustomed to work with every day.
The work on this project has thus far resulted in an academic software
that allows solution of the multi-objective linear programme (11.2) for 2D
and smaller sized 3D problems. The software has been developed in close con-
sultation with treatment planners at Auckland City Hospital. Further work
 11 Real-World Applications of Multiobjective Optimization 305

needs to address numerical issues arising in large 3D problems. Before actual

use in clinical practice a lengthy and costly approval process needs to be com-
pleted for which the support of a medical software company will be required.

11.6 Supply Chain Management

11.6.1 Problem Description

In supply chains often thousands of individual decisions need to be made

and coordinated. Due to the high degree of complexity successive planning
approaches are therefore often chosen in practice.
Figure 11.12 shows typical planning tasks that arise in supply chains. These
planning tasks are arranged in two dimensions. The first dimension is the “sup-
ply chain process”. In this dimension, planning tasks are arranged focusing on
the most important processes following the flow of goods in supply chains.
These are procurement, production, distribution and sales. The second di-
mension is the “planning horizon”. In this dimension the planning tasks are
distinguished by their temporal impact on the supply chain. These may be
strategic decisions with a long-term impact or operational decisions, which
have only an immediate impact in the near future (short-term).
The strategic network planning module covers the long-term decisions
across all supply chain processes. It supports the user to determine the struc-
ture of the supply network (plant location, distribution system) as well as
the product program. Although its results are important for the long-term

Fig. 11.12. Supply Chain Planning Matrix Fleischmann et al. (2005), p. 87.
306 T. Stewart et al.

profitability of supply chains, it is often not a core functionality of Advanced

Planning Systems (APS). This is because APS are primarily built to support
daily business, whereas strategic decisions are only reviewed periodically and
most often not within the regular organization, but rather on a project basis.
The master planning module coordinates procurement, production and
distribution on a mid-term level. Its major decision support is about sourcing:
Which product is produced at which location and when? Thus, in this module
the master production schedule is fixed. However, it is important to anticipate
the key characteristics of the lower (short-term) planning levels within this
module, because otherwise inconsistent plans (for procurement, production
and distribution) will result on the lower planning level.
In the area of distribution and transport planning, distribution related
planning tasks are addressed, the latter on a more detailed level (e.g., schedul-
ing of transports, vehicle loading and routing). Production planning and
scheduling on the other hand are the two modules that support production re-
lated issues in the short-term planning horizon. Finally, purchasing and mate-
rial requirements planning support the (short-term) procurement of materials
and components.
The capabilities offered within mySAP Supply Chain Management (SCM)
extend far beyond the scope of this article. The key functionalities we will
describe in the following are highlighted in Figure 11.13, which is based on
the generic supply chain planning matrix (Figure 11.12). They are part of the
SAP Advanced Planner and Optimizer (SAP APO), which is the advanced
planning component within the mySAP SCM solution. For more information
on SAP APO refer to Bartsch and Bickenbach (2001); Dickersbach (2005).

Fig. 11.13. Supply chain planning matrix using mySAP SCM terminology.
 11 Real-World Applications of Multiobjective Optimization 307

Solving a planning problem of this complexity in its entirety within one

planning step is neither feasible from an algorithmic perspective nor sensible
from a planning process point of view. A hierarchical decomposition of the
complete planning problem into a master planning and a production planning
and scheduling part addresses planning complexity as well as planning pro-
cess design issues. In the following, we describe how the planning problem is
partitioned into a master planning and a production planning part and which
business requirements are addressed on which planning level.

11.6.2 Methodology

Multi Location Optimization in Supply Network Planning (SNP)

In SAP APO, the master planning process is implemented in the Supply

Network Planning (SNP) module. SNP offers a multitude of functionalities,
not all of which can be described in the limited scope of this article. More
details on the SNP module can be found, among others, in Dickersbach (2005).
The SNP model contains all relevant locations, i.e. production plants and
distribution centres, in the supply network. SNP determines which of the
plants produces which quantities of which products in which time periods.
On a rough level, SNP also determines which production alternative is used
at a specific plant, for instance with regard to ingredients and general process
characteristics.
To reduce the complexity of the master planning model, not all products
are considered in the SNP optimization run. The selection is made by flagging
specific products as not relevant for SNP planning. SNP planning takes into
account all products produced in a location, all products for which a stock
transfer between locations is possible, externally procured active ingredients,
goods for resale and selected forming auxiliaries. Not relevant to SNP are
most raw materials, most forming auxiliaries as well as packaging materials.
A similar logic is used for resources. Only bottleneck resources are selected
for SNP planning.
The concentration on key products and bottleneck resources also results
in a significant simplification of the recipes1 used in SNP, which are derived
from the more detailed recipes used in Production Planning and Detailed
Scheduling (PP/DS) and the attached enterprise resource planning (ERP)
system. Furthermore, compared to the recipes used on the PP/DS level, not
all setup and cleaning operations are considered in SNP recipes. Small setup
operations are normally neglected while key setup activities which are relevant
for campaign planning on bottleneck resources due to their long duration or
high costs are considered. To account for the resource capacity consumed by
small setup and cleaning operations, a loss factor is applied to calculate the
resource capacity for SNP planning.
1
In APO, a recipe is commonly referred to as PPM (production process models)
or PDS (production data structure).
308 T. Stewart et al.

One of the main aspects of the SNP planning process is the cost-based
plan determination. The following cost types are used to build a cost model
which represents the business scenario of value base planning:
• Penalties for not meeting customer demand / forecast,
• Penalties for late satisfaction of customer demand / forecast (location
product specific)
• Penalties for not meeting safety stock / safety days’ supply requirements
(location product specific, linear or piecewise linear)
• Storage cost (location product specific)
• Penalty for exceeding maximum stock level / maximum coverage (location
product specific, linear or piecewise linear)
• External procurement cost (linear or piecewise linear, location product
specific)
• Handling in / out cost (location product specific)
• Transportation cost (transportation lane, product and means of transport
specific, linear or piecewise linear)
• Variable production cost (production process specific, linear or piecewise
linear)
• Fixed production cost / setup cost (production process specific)
• Resource utilization cost (resource specific)
• Costs for additional resource utilization (e.g. use of additional shifts, re-
source specific)
• Cost for falling below minimum resource utilization.
The definition of the cost model is of crucial importance for controlling the
behaviour of the SNP optimizer. One of the central questions is whether to
maximize service level – which usually means using high penalties for non
and late delivery – or to maximize profits – which requires use of realistic sale
prices. In the case study scenario, the non delivery cost levels reflect real sale
prices sufficiently close to enable a profit maximization logic.
Another important feature of the case study scenario and the resulting
cost model is inventory control. High seasonality effects and long campaign
durations necessitate considerable build-up of stocks. To avoid an unbalanced
build-up of stock, soft constraints for safety stock and maximum stock levels
are used. To achieve an even better inventory levelling across products and
locations, piecewise linear cost functions for falling below safety stock as well
as for exceeding maximum stock levels are employed. In SNP optimization all
revelevant constraints can be considered, including
• capacities for production, transportation, handling and storage resources,
• maximum location product specific storage quantities,
• minimum, maximum and fixed production lot sizes,
• minimum, maximum and fixed transportation lot sizes,
• minimum production campaign lot sizes.
 11 Real-World Applications of Multiobjective Optimization 309

An optimization model which considers all these constraints – especially those

which can only be modelled using binary or general integer variables – can be
highly complex.

Production Planning and Detailed Scheduling (PP/DS)

The short term planning process is dealt with in the Production Planning and
Detailed Scheduling module within SAP APO.
PP/DS focuses on determining an optimal production sequence on key
resources. In PP/DS, a more detailed modelling than on the SNP planning
level is chosen. On the basis of the results determined in SNP optimization,
a detailed schedule which considers additional resources and products is cre-
ated. This schedule is fully executable and there is no need for manual planner
intervention, even though manual re-planning and adjustments are fully sup-
ported within the PP/DS module. An executable plan can only be ensured
by considering additional complex constraints in PP/DS optimization. These
additional constraints include:
• Time-continuous planning
• Sequence-dependent setup and cleaning operations.
As the value based planning part is handled within SNP, the PP/DS optimizer
uses a different objective function than the SNP optimizer. The following goals
can be weighted in the objective function, which is subject to minimization:
• Sum of delays and maximum delay against given due dates
• Setup time and Setup cost
• Makespan (i.e. time interval between first and last activity for optimizing
the compactness of the plan)
• Resource cost (i.e. costs associated with the selection of alternative re-
sources)
The main objective of the PP/DS optimizer run in the scenario at hand is to
minimize setup times and costs on resources without incurring too much delay
against the order due dates. For some resource groups, resource costs are also
used to ensure that priority is given to the ‘best’ (i.e. fastest, cheapest, etc.)
resources.

11.6.3 Remarks

We have seen that both in Supply Network Planning and in Detailed Schedul-
ing there are a huge number of objectives to be minimized. However these
objectives can be mastered by forming a 4-level hierarchy.
On the root or top level, two dimensions of the second level can be differen-
tiated: Service degree and real costs. The objective of real costs differentiates
at the third level between for example:
310 T. Stewart et al.

• storage costs: the minimization of inventory by weighting the inventory of

each storage location by an estimated cost factor
• safety costs: the minimization of the risk of getting our of stock by weight-
ing the risk of each storage location by a cost factor
• setup costs: the minimization the overhead of change over for each resource
by weighting each change over by a cost factor.
The objective of service degree differentiates at the third level between for
example:
• delay penalties: the minimization of delay for each demand or customer
order by weighting the priority of the customer
• non delivery penalties: the minimization of non delivery for each demand
or customer order by weighting the priority of the customer.
Only for the top level are weighting factors not appropriate. The planner wants
to see several alternative solutions of the Pareto front of these two objectives:
By how much would costs increase if we wish to achieve a better service
level? A high service level is clearly an important objective, but there is no
direct cost measure for a delay. Summarizing in Advanced Planning for Supply
Chain Management we can focus on an optimization problem with just two
objectives maximizing service degree while minimizing the costs. In particular,
two dimensional visualization of the Pareto front and representative solutions
of this front are needed and sufficient.

11.7 Interactive Processes for Aircraft Design

11.7.1 Problem Description

Interactive Evolutionary Computation (IEC) has started to capture the fasci-

nation of researchers from fields as diverse as art, architecture, data mining,
geophysics, medicine, psychology, robotics, and sociology. Takagi (2001) out-
lined many of these applications in his overview paper. However, to this day
only very few researchers have applied IEC to the problem of engineering and
design of complicated artifacts. While the main reason for the slow pace of
adoption in engineering is mostly open for speculation, it is partially a result
of the field’s reluctance to accept new methods, like Genetic Algorithms, as
well as the field’s already heavy reliance on automated optimization processes
that leave decidedly little room for subjectivity. While the reliance of engi-
neers on analysis tools requires interactive evolutionary techniques to utilize
them in the fitness generation, it is also true that many design decisions in
practice are made through gut feel and intuition rather then analysis. Rec-
ognizing that fact, this example identifies an IEC approach to design that
allows for automatic fitness calculation through analysis as well as selection
and fitness assignment by the human designers directly.
 11 Real-World Applications of Multiobjective Optimization 311

There are few things humans build that are more complicated than air-
craft. Not only are the reliability requirements enormous, given the fatal conse-
quences of failure, but the system itself strides a multitude of areas in physics,
such as aerodynamics, thermodynamics, mechanics, and materials. This con-
volution of disciplines has led historically to a very sequential design process,
tackling the various disciplinary issues separately: aerodynamicists only tried
to maximize the performance of the wing (or even just an airfoil), propulsion
engineers tried to build the largest engines, structural engineers tried to build
the sturdiest airframe, while material scientists attempted to only utilize the
lightest and sturdiest materials. As a consequence, the design process itself was
a highly inefficient iterative process of ever changing airplane configurations,
only reconciled by rare, experienced individuals that were proficient in all (or
at least many) disciplines. As these people retired, and significant computa-
tional power became available, a new design process emerged, attempting to
satisfy the concerns of all disciplines concurrently: Multidisciplinary (Design)
Optimization. MDO is inherently a multicriteria optimization problem, since
each discipline contributes at least one objective function that potentially
conflicts with the objective(s) of the other disciplines. The following example
demonstrates the ability of one MCDM technique, Interactive Evolutionary
Design, to address the difficult task of balancing the different disciplinary
objectives when determining the preliminary design configuration of a Super-
sonic Business Jet.
Figure 11.14 outlines the interactive evolutionary design process employed
for this application example (see Bandte and Malinchik, 2004, for background
discussion). After the problem is set-up by defining design variables, objec-
tives and constraints and sufficient feasibility has been established, a GA is
being interrupted after several generations to display the current population
via spider-graphs and Pareto Frontier displays for objective values as well as
visualizations of the aircraft configurations. Based on this information the

Fig. 11.14. Integrated interactive evolutionary design process.

312 T. Stewart et al.

designer can make some choices regarding objective preferences and features
of interest, redirecting search and influencing selection respectively. To limit
the scope of this example, only the redirection of the search through objective
preferences is being implemented here. However, as exemplified later, designer
selection of features of interest is an important part of interactive evolutionary
design and should not be neglected in general. The following sections lay out
in detail all tasks performed over several iterations for this example.
As in any design problem, the first step is to define the independent param-
eters, objectives and constraints, as well as evaluation functions that describe
the objectives’ dependencies on the independent variables. For this interactive
evolutionary design environment, this step also identifies the genotype repre-
sentation of a design alternative, the fitness evaluation function, influenced by
the objectives, and how to handle design alternatives that violate constraints.
The supersonic business jet is described by five groups of design variables,
displayed in a screen shot presented in Figure 11.15. The first group, general,
consists of variables for the vehicle, some of which could be designated design
requirements. The other four groups contain geometric parameters for the
wing, fuselage, empennage, and engine. The engine group also entails propul-
sion performance parameters relevant to the design. All in all, the chromo-
some for this supersonic business jet contains 35 variables that can be varied
to identify the best solution.
A mix of economic, size, and vehicle performance parameters were chosen
as objectives in this example, with a special emphasis on noise generation,
since it is anticipated to be a primary concern for a supersonic aircraft. Hence,
for the initial loop the boom loudness and, as a counter weight, the acquisition
cost are given slightly higher importance of 20%, while all other objectives are
set at 10%. Furthermore, certain noise levels could be prohibitively large and
prevent the design from getting regulatory approval. Hence, some of the noise
objectives have to have constraint values imposed on them. In addition to
these constraints, the design has to fulfill certain FAA requirements regarding
take-off and landing distances as well as approach speed. Furthermore, the
amount of available fuel has to be more than what is needed for the design
mission. Finally, fitness is calculated via a weighted sum of normalized ob-
jective values, penalized by a 20% increase in value whenever at least one
constraint is violated. Note that the “best solution” is identified as the one
with the lowest fitness, i.e. objective function values. All constraints, objec-
tives, normalization values and preferences are also displayed in Figure 11.15.

11.7.2 Methodology and Interactive Scheme

Run GA

Since the initial objective preferences were already specified at problem defi-
nition, the GA can be executed next without requiring further input from the
designer. The GA chosen for this example is one of the most general found
 11 Real-World Applications of Multiobjective Optimization 313

Fig. 11.15. Screen shot: Display of information after 80 generations.

in the literature Holland (1975); Mitchell (1996); Haupt and Haupt (1998). It
has a population size of 20 and makes use of a real valued 35-gene represen-
tation, limited to the ranges selected at problem set-up. New generations are
created from a population with a two individuals elite pool and proportionate
probabilistic selection for crossover. The crossover algorithm utilizes a strat-
egy with one splice point that is selected at random within the chromosome.
Since the design variables are grouped in logical categories, this crossover al-
gorithm enables, for example, a complete swap of the engine or fuselage-engine
assembly between parents. Parent solutions are being replaced with offspring.
Each member of the new population has a 15% probability for mutation at
ten random genes, sampling a new value from a uniform distribution over the
entire range of design variable values. The GA in this example is used for
demonstration purposes only and therefore employs just a small population.
A population size of 50 to 100 seems more appropriate for a more elaborate
version of the presented interactive evolutionary design approach.

Display Information

Once the GA has executed 80 generations, it is interrupted for the first time to
display the population of design alternatives found to this point. The designers
are presented with information that is intended to provide maximum insight
314 T. Stewart et al.

into the search process and the solutions it is yielding. In order to allow for a
reasonable display size of the aircraft configuration, only the four best design
alternatives, based on fitness and highest diversity in geometrical features, are
presented in detail on the top of the left hand side of the display. A screen
shot of the displayed information is presented in Figure 11.15, highlighting the
individual with the best/lowest fitness, which is also enlarged to provide the
designers with a more detailed view of the selected configuration. The design
variable values for the highlighted alternative and their respective ranges are
presented below this larger image, completing the “chromosome information”
on the left hand side of the pane.
On the right hand side of the pane, the designers can find the objective
and constraint information pertaining to the population and the highlighted
individual. On the top, a simple table outlines the specific objective values for
the highlighted alternative, as well as the objective preferences and normal-
ization factors used to generate the fitness values for the current population.
Below the table, a spider graph compares the four presented alternatives on
the basis of their normalized objective values, while to the right four graphs
display the objective values for the entire population, including its Pareto
frontier (highlighted individual in black). Below the spider chart, a table lists
the constraint parameter values for the highlighted alternative as well as the
respective constraint values. A green font represents constraint parameter val-
ues near, orange font right around, and red font way beyond the constraint
value. Finally at the bottom, three graphs display the population with respect
to its member’s constraint parameter values as well as the infeasible region,
superimposed. These graphs in particular indicate the level of feasibility in
the current population.

Provide Input

This step represents the central interaction point of the human with the IEC
environment. Here they process the information displayed and communicate
preferences for objectives, features of interest in particular designs, whether
specific design variable values should be held constant in future iterations,
what parameter setting the GA should run with in the next iteration (e.g.
a condition that identifies the end of the GA iteration), or whether specific
design alternatives should serve as parents for the next generation.
Analyzing the data provided, it is noticeable that all objectives except
for the boom loudness are being satisfied well. Consequently, in an attempt
to achieve satisfactory levels for the boom loudness in the next iteration, its
preference is increased to 30%, reducing the acquisition cost’s importance to
10%. This feedback is provided to the GA via a pop-up screen (not displayed
here) that allows the designer to enter the new preference values for the next
iteration. With this new preference information the GA is executed for another
80 generations.
 11 Real-World Applications of Multiobjective Optimization 315

Second Iteration, 160 Generations

After the GA has executed an additional 80 generations, it is apparent from

the objective values that the last set of preferences did not emphasize the boom
loudness and sideline noise enough, since boom loudness did not improve (from
87.11 to 87.12 dB) and the sideline noise got worse (from 90.89 to 93.65 dB).
Consequently, for the next iteration the importance of both is increased to
35% while all other objectives are reduced to 5%.

Third Iteration, 240 Generations

Examining the population after another 80 generations yields that the last
set of preferences still did not emphasize the boom loudness enough, since
boom loudness improved only marginally (from 87.12 to 86.96 dB). On the
other hand, sideline noise did improve significantly (from 93.65 to 90.94 dB),
so that for the next iteration all emphasis can be given to boom loudness. To
keep the score even for all other objectives, they are being kept at 5% with
boom loudness at 65%.

Fourth Iteration, 400 Generations

For this iteration 160 more generations were executed to produce the popula-
tion displayed in the screen shot presented in Figure 11.16. In part due to the
longer GA run, a very good solution, #7172, is found after 400 generations
with largely improved values for almost all the objectives. This result is some-
what surprising, considering most objectives had only a 5% level of preference
and the one objective with 65%, boom loudness, improved only marginally.
This result can be attributed to an exemplified effect from summing corre-
lated objective function values that are caused by similar design alternatives
(with similar design variable settings) exhibiting similarly good (or bad) val-
ues for all objectives except boom loudness. The fact that boom loudness is
a conflicting objective, specifically with sideline noise, can also be observed
from the pronounced Pareto frontier in the second objective chart from the
top.
However, the presented solution after 400 generations seems to satisfy
the objectives better than the published solution in Buonanno et al. (2002),
generated by MATLAB c ’s Fmincon function (The MathWorks Inc. (2008)).
So it could be concluded that none of these objective values are dramatically
out of sync or range and the presented individual is the final solution.

11.8 Land Use Planning

11.8.1 Problem Description

The work described here was motivated by problems of land use allocation or
re-allocation in the Netherlands. Land which is already intensely developed
316 T. Stewart et al.

Fig. 11.16. Screen shot: Fourth iteration after 400 generations.

has often to be redeveloped to meet current needs for agriculture, residence,

recreation and industry, while at the same time recognizing conservation goals
(including possible restoration of some land to approximately pristine condi-
tions). The initial model development was based on a specific region near
Amsterdam (the Jisperveld), as briefly described in Section 5 of Stewart et al.
(2004). However, the longer term intention is to build the model into a general
land use planning decision support system (LUDSS). The function of such an
LUDSS would be to generate a small number of plans which can be evaluated
holistically by decision makers or planners. They would then indicate which
features of the plan they like or dislike, which would lead to a readjustment of
goals in the LUDSS and the generation of a new solution. This process may be
repeated until planners are satisfied that no substantial further improvements
are likely.
The model represents the region under consideration by a rectangular grid
of (say) R × C equal-sized cells. It is then assumed that one and only one
land use (from a set of Λ possible uses) is allocated to each grid. Formally,
we define binary variables xrcℓ , such that xrcℓ = 1 if land use ℓ is allocated
to cell (r, c) and xrcℓ = 0 otherwise. For ease of notation we shall denote
the three dimensional array of all xrcℓ values by x. Typical constraints on x
would include exclusions of certain land uses from certain cells (corresponding
 11 Real-World Applications of Multiobjective Optimization 317

xrcℓ set to zero) and upper and lower bounds on the total area allocated to a

R
C
particular land use (i.e. on r=1 c=1 xrcℓ ).
Some objectives relate to directly quantifiable costs and benefits, and tend
to be additive in nature. For example, if all such objectives (without loss in
generality) are expressed as costs then:
R C Λ
i
fi (x) = βrcℓ xrcℓ
r=1 c=1 ℓ=1

i
where βrcℓ is the cost in terms of objective i associated with allocating land
use ℓ to cell (r, c).
As initially described by Aerts et al. (2005), however, a critical manage-
ment objective is to ensure that land uses are sufficiently compact to allow
integrated planning and management. Aerts et al. (2005) introduce essentially
one measure of compactness, related to the numbers of cells adjacent to cells
of the same land use. This concept was extended in our work by means of a
more detailed evaluation of the fundamental underlying management objec-
tives. Defining a cluster of cells as a connected set of cells allocated to a single
land use, three measures of performance for each land use type were identified
as follows:
• Numbers of clusters for each land use, Cℓ : These measure the degree of
fragmentation of land uses, and minimization of the number of clusters
would seek to ensure that areas of the same land use are connected as far
as possible.
• Relative magnitude of the largest cluster for each land use: Maximization
of the ratio Lℓ = nL L
ℓ /Nℓ is sought, where nℓ and Nℓ are respectively
the number of cells in the largest cluster and the total numbers of cells
allocated for land use ℓ. If multiple clusters are formed, then it would
often be better to have at least one large consolidated cluster, than for all
clusters to be relatively small.
• Compactness of land uses, denoted by Rℓ , defined by a weighted average
across all clusters for land use ℓ of the ratio of the perimeter to the square
root of the area of the cluster. This measure should be minimized as a
compact area for one land use (e.g. a square or circular region) may be
easier to manage than a long thread-like cluster.
The above measures define an additional 3Λ objectives, as the compactness
goals need to be achieved for each land use individually. Furthermore, the
calculation of Cℓ , Lℓ and Rℓ require the execution of a clustering algorithm, so
that these additional objectives are non-linear and computationally expensive.
The total number of objectives is thus k = k0 + 3Λ, where k0 is the number
of additive objectives.
318 T. Stewart et al.

11.8.2 Methodology

In view of the large number of objectives, it is not practical to seek to represent

the efficient frontier in full. The approach adopted was thus based on sampling
the efficient frontier by optimizing an aggregate measure of performance sub-
ject to the constraints on x, for each of a number of different aggregations.
The aggregation chosen was that of the scalarizing function introduced by
Wierzbicki (1999) in the context of his reference point methodology, except
that we chose a smoother function than that based on the maximum operator.
Thus for any given reference point (which can be viewed as a set of goals or
aspiration levels for each objective), say g1 , g2 , . . . , gk , the scalarizing which is
to be minimized is defined by:
k  4
fi (x) − Ii
(11.3)
i=1
gi − Ii

where Ii is the ideal (best achievable measure of performance) for objective i.

Constrained minimization of (11.3) with respect to x is a non-linear com-
binatorial optimization problem, with the added complexity that most of the
functions cannot be evaluated explicitly in closed form (but are are derived as
outputs from a clustering algorithm). Stewart et al. (2004) describe a special
purpose genetic algorithm (GA), designed to exploit a number of special char-
acteristics of the land use planning problem models, both in the generation
of population elements and in the execution of cross-overs (see cited reference
for details).
It is interesting to emphasize at this point that the chosen methodology in-
cludes elements from conventional multiobjective optimization and from evo-
lutionary optimization, thus representing an integration of the two themes of
the present volume.

11.8.3 Interactive Scheme and Results

In implementing the algorithm within an LUDSS, the process starts by select-

ing one or more tentative reference points, perhaps a central reference point
(all goals positioned midway between ideals and worst performance levels)
and a selection of reference points which favour each individual goal in turn.
Each individual solution generated will be efficient (to the level of optimiza-
tion accuracy achieved by the GA), and so represents a point on the efficient
frontier. In response to assessments by the user as to the direction in which
it is desired to move the solution, the reference points are adjusted and the
optimizations repeated. In this way, the user is able incrementally to explore
the efficient frontier until such time as a satisfactory solution, or short list of
possible solutions, is found.
A detailed case study in the use of this system is given in Janssen et al.
(2007). An illustration of the manner in which the interactions may progress
 11 Real-World Applications of Multiobjective Optimization 319

is given in Figure 11.17, which presents three land use maps. The numbers
in the maps indicate nine potential land use types, namely: 1. Intensive agri-
culture; 2. Extensive agriculture; 3. Residence; 4. Industry; 5. Day recreation;
6. Overnight recreation; 7. Wet natural area; 8. Water (recreational); and 9.
Water (limited access).

Initial Map First Solution

Final Solution
Fig. 11.17. Three land use maps generated from the LUDSS.

The upper left hand map displays the original land use pattern. The upper
right hand map was generated in the first optimization step, and provides a
very compact allocation of land uses. However, the costs were deemed to be
too high, largely because of the extent of agricultural land reclaimed from wet
areas. For this reason, the priority on the cost attribute was increased. Some
more fragmentation of the land uses was then re-introduced, much of which
was acceptable except for that of the agricultural land. Also, the values as-
sociated with conservation goals were found to be unsatisfactory. Adjustment
320 T. Stewart et al.

of these priorities led after the 8th iteration to the lower map in Figure 11.17,
and this was found to represent a satisfactory compromise.

11.9 Engineering Design Problems with Large Numbers

of Objectives
11.9.1 Problem Description: Cable-Stayed Bridges
Cable-stayed bridges are gaining much popularity due to their beautiful shape.
During and after construction, this kind of bridge needs to have the cable
length adjusted in order to attain errors of cable tension and camber (the
configuration of the girders of the bridge) within some allowable range.
To this end, the following criteria are considered (Nakayama et al. (1995)):
• residual error in each cable tension,
• residual error in camber at each node,
• amount of shim adjustment for each cable,
• number of cables to be adjusted.
Since the change of cable rigidity is small enough to be neglected with respect
to cable length adjustment, both the residual error in each cable tension and
that in each camber are assumed to be linear functions of the amount of shim
adjustment.
Let us define n as the number of cables in use, ∆Ti (i = 1, . . . , n) as the
difference between the designed tension values and the measured ones, and
xik as the tension change of i-th cable caused from the change of the k-th
cable length by a unit. The residual error in cable tension caused by the shim
adjustment is given by
n
pi = |∆Ti − xik ∆lk | (i = 1, . . . , n)
k=1

Let m be the number of nodes, ∆zj (j = 1, . . . , m) the difference between the

designed camber values and the measured ones, and yjk the camber change at
j-th node caused from the change of the k-th cable length by a unit. Then the
residual error in the camber caused by the shim adjustments of ∆l1 , . . . , ∆ln
is given by
n
qj = |∆Zj − yjk ∆lk | (j = 1, . . . , m)
k=1
In addition, the amount of shim adjustment can be treated as objective func-
tions of
ri = |∆li | (i = 1, . . . , n)
The upper and lower bounds of shim adjustment inherent in the structure of
the cable anchorage are as follows:
∆lLi ≤ ∆li ≤ ∆lUi (i = 1, . . . , n). (11.4)
 11 Real-World Applications of Multiobjective Optimization 321

11.9.2 Methodology

Now we have a multi-objective optimization problem in which p1 , . . . , pn ,

q1 , . . . , qm and r1 , . . . , rn are to be minimized under the constraint (11.4).
Some large scale bridges have around 100 cables at each side, so that the
problem results easily in a very large number of objective functions. For this
multi-objective optimization problem, engineers in bridge construction have
tried to apply goal programming (Charnes and Cooper, 1961), in which they
want to get a desirable solution by adjusting weights imposed on criteria.
However, it has been pointed out in the literature (e.g. see Nakayama, 1995)
that this task is very difficult even for simple problems. In addition, the shim
adjustment must ususally be done during a relatively short period (say, 2:00
am to 8:00 am) with a stable temperature, because the cable length is greatly
affected by change of temperature. Therefore, the decision of cable length
adjustment must be made very quickly. Also, due to this reason, the tradi-
tional goal programming approach is not satisfactory for practical use in this
problem.
On the other hand, an interactive multi-objective programming technique
has been developed, called the satisficing trade-off method (Nakayama and
Sawaragi, 1984). The method is one of aspiration level approaches to multi-
objective optimization, which are observed to be effective in many practical
problems because they are very simple and easy to implement and do not re-
quire any mathematical consistency of decision makers’ judgment, and in ad-
dition take aspiration levels of decision makers as a probe rather than weights
imposed on criteria.
Figure 11.18 shows the graphical user interface (GUI) for the erection man-
agement system of a cable stayed bridge using the satisficing trade-off method.
The residual error of each criterion and the amount of shim adjustment are
represented by bar graphs. The aspiration level is inputted by a mouse on
the graph. After solving the auxiliary min-max problem, the Pareto solution
according to the aspiration level is represented by another bar graph in a sim-
ilar fashion. If the designer is not satisfied with the Pareto solution displayed,
he/she can revise the aspiration level by means of mouse operations, and the
process repeated.
This procedure is continued until the designer obtains a desirable shim-
adjustment. The interactive operation using the GUI is very easy for the de-
signer, and the visual information on trade-offs among criteria is user-friendly.
The software has been used for real bridge construction, for example tje
Tokiwa Swan Bridge (Ube City) and the Karasuo Harp Bridge (Kita-Kyusyu
City) in 1992.
One of the important aspects of such a problem with a large number of
objective functions is the graphical user interface. As can be easily seen in
Fig.11.18, it is not too difficult for designers to make a trade-off analysis on
the basis of the displayed visual information, even in cases with hundreds
of objective functions. However, it might be difficult or even impossible to
322 T. Stewart et al.

Fig. 11.18. Erection Management System of Cable-stayed Bridge.

grasp the total trade-off context on the basis of numerical information only,
for problems with such a large number of objective functions.

11.9.3 A Further Application: Lens Design

Another good example with a large number of objective functions is lens

design. There are many kinds of lenses such as copier, camera, medical instru-
ments and so on. Above all, lenses in semiconductor chip production are very
expensive (of the order of million dollars), and hence have to be designed very
carefully.
In lens design, there are around 200 design variables such as
• kinds of glass,
• number of lenses,
• diameter,
• curvature,
• distance between lenses,
and around 400 criteria such as
• cost
• weight
• criteria for images:
– aberration
– chromatic
– spherical
– astigmatism
– coma
 11 Real-World Applications of Multiobjective Optimization 323

– distortion
– curvature of field
– color balance
– resolution
– MTF
– CCI
In lens design, there is the further difficulty of nonlinear optimization in addi-
tion to the large number of objective functions: Scalarized optimization prob-
lems are usually highly nonlinear and highly multi-modal. Moreover, those
functional forms are not given explicitly in terms of design variables. Those
function values are evaluated on the basis of some kind of simulation (ray
trace). Therefore, it is difficult to obtain a global minimum for the objective
function.
So far, engineers use specific software in lens design. Their main attention
has been directed to how they obtain a global optimum for the scalarized
objective function, while the linearly weighted sum scalization function is
applied. It will be surely a good subject to investigate how interactive multi-
objective optimization techniques can work in lens design.

11.10 Concluding Comments

In what sense are the above applications different to other optimization stud-
ies? Clearly, the distinction lies in the multiplicity of objectives which are
central to the applications discussed here. In common with more general ap-
proaches to multiple criteria decision making (MCDM), those applying the
multiobjective methods start by careful problem structuring to identify the
underlying objectives, and to represent these in meaningful manner, much as
has been described in Chapter 3 of Belton and Stewart (2002).
Some of the approaches reported in the case studies do ultimately seek to
identify an overall mathematical objective function as a surrogate measure of
performance to be “maximized” or “minimized”, but:
• This is done only after careful attention to tradeoffs between objectives
and clear recognition that these tradeoffs may change as one explores the
decision space;
• The methods are applied interactively, with systematic changes in formula-
tion (revising goals or value tradeoffs) in the light of preference information
(as described in other Chapters of this book concerning interactive meth-
ods), and thus providing a means of implicit exploration of the Pareto
frontier.
Other reported approaches avoid the use of surrogate objective functions, by
seeking to identify the Pareto frontier explicitly, leaving the user or ultimate
client to explore the options visually before making the final selection. Unfor-
tunately, it is difficult to provide an unambiguous visualization of the frontier
324 T. Stewart et al.

for more more than two or three objectives, although the Chapter 9 in this
book seeks to extend such opportunities.
The clear challenge to future research lies precisely in the interface be-
tween these implicit and explicit methods of searching the Pareto Frontier.
An opportunity may lie in using the interactive methods using surrogate mea-
sures of performance for an initial exploration, but using the explicit search
methods (linked to appropriate visualization) to refine the exploration across
those objectives which are found most critical to the final decisions in the
most promising regions of the decision space.

References
Aerts, J.C.J.H., van Herwijnen, M., Janssen, R., Stewart, T.J.: Evaluating spatial
design techniques for solving land-use allocation problems. Journal of Environ-
mental Planning and Management 48(1), 121–142 (2005)
Alexandrov, N.M., Dennis, J.E., Lewis, R.M., Torczon, V.: A trust region framework
for managing use of approximation models in optimization. Journal on Structural
Optimization 15(1), 16–23 (1998)
Arima, T., Sonoda, T., Shirotori, M., Tamura, A., Kikuchi, K.: A numerical investi-
gation of transonic axial compressor rotor flow using a low-Reynolds-number k − ǫ
turbulence model. ASME Journal of Turbomachinery 121(1), 44–58 (1999)
Bandte, O., Malinchik, S.: A broad and narrow approach to interactive evolutionary
design – an aircraft design example. In: Deb, K., et al. (eds.) GECCO 2004. LNCS,
vol. 3103, pp. 883–895. Springer, Heidelberg (2004)
Bartsch, H., Bickenbach, P.: Supply Chain Management mit SAP APO. Galileo
Press, Bonn (2001)
Belton, V., Stewart, T.J.: Multiple Criteria Decision Analysis: An Integrated Ap-
proach. Kluwer Academic Publishers, Boston (2002)
Benson, H.: An outer approximation algorithm for generating all efficient extreme
points in the outcome set of a multiple objective linear programming problem.
Journal of Global Optimization 13(1), 1–24 (1998)
Buonanno, M., Lim, C., Mavris, D.N.: Impact of configuration and requirements on
the sonic boom of a quiet supersonic jet. Presented at World Aviation Congress,
Phoenix, AZ (2002)
Charnes, A., Cooper, W.: Management Models and Industrial Applications of Linear
Programming, vol. 1. John Wiley, New York (1961)
Dickersbach, J.T.: Supply Chain Management with APO, 2nd edn. Springer, Berlin
(2005)
Ehrgott, M., Winz, I.: Interactive decision support in radiation therapy treatment
planning. OR Spectrum 30, 311–329 (2008)
Ehrgott, M., Holder, A., Reese, J.: Beam selection in radiotherapy design. In: Linear
Algebra and Its Applications, vol. 428, pp. 1272–1312 (2008a)
Ehrgott, M., Hamacher, H.W., Nußbaum, M.: Decomposition of matrices and static
multileaf collimators: A survey. In: Alves, C.J.S., Pardalos, P.M., Vicente, L.N.
(eds.) Optimization in Medicine. Springer Series in Optimization and Its Applica-
tions, vol. 12, pp. 25–46. Springer Science & Business Media, New York (2008b)
 11 Real-World Applications of Multiobjective Optimization 325

Emmerich, M., Giannakoglou, K., Naujoks, B.: Single and multi-objective evolution-
ary optimization assisted by Gaussian random field meta-models. IEEE Transac-
tions on Evolutionary Computation 10(4), 421–439 (2006)
Fleischmann, B., Meyr, H., Wagner, M.: Advanced planning. In: Stadtler, H., Kilger,
C. (eds.) Supply Chain Management and Advanced Planning. Concepts, Models,
Software and Case Studies, 3rd edn., pp. 81–106. Springer, Berlin (2005)
Hasenjäger, M., Sendhoff, B., Sonoda, T., Arima, T.: Three dimensional evolution-
ary aerodynamic design optimization using single and multi-objective approaches.
In: Schilling, R., Haase, W., Periaux, J., Baier, H., Bugeda, G. (eds.) Evolutionary
and Deterministic Methods for Design, Optimization and Control with Applica-
tions to Industrial and Societal Problems EUROGEN 2005, Munich, FLM (2005)
Haupt, R.L., Haupt, S.E.: Practical Genetic Algorithms. John Wiley & Sons, New
York (1998)
Holder, A.: Designing radiotherapy plans with elastic constraints and interior point
methods. Health Care Management Science 6, 5–16 (2003)
Holland, J.: Adaptation in Natural and Artificial Systems. The University of Michi-
gan Press, Ann Arbor (1975)
Janssen, R., van Herwijnen, M., Stewart, T.J., Aerts, J.C.J.H.: Multiobjective deci-
sion support for land use planning. Environment and Planning B, Planning and
Design. To appear (2007)
Jin, Y.: A comprehensive survey of fitness approximation in evolutionary computa-
tion. Soft Computing 9(1), 3–12 (2005)
Jin, Y., Branke, J.: Evolutionary optimization in uncertain environments – A survey.
IEEE Transactions on Evolutionary Computation 9(3), 303–317 (2005)
Jin, Y., Olhofer, M., Sendhoff, B.: On evolutionary optimization with approximate
fitness functions. In: Genetic and Evolutionary Computation Conference, pp. 786–
792. Morgan Kaufmann, San Francisco (2000)
Jin, Y., Olhofer, M., Sendhoff, B.: A framework for evolutionary optimization with
approximate fitness functions. IEEE Transactions on Evolutionary Computa-
tion 6(5), 481–494 (2002)
Jin, Y., Olhofer, M., Sendhoff, B.: On evolutionary optimization of large problems
using small populations. In: Wang, L., Chen, K., Ong, Y.S. (eds.) ICNC 2005.
LNCS, vol. 3611, pp. 1145–1154. Springer, Heidelberg (2005)
Jin, Y., Zhou, A., Zhang, Q., Tsang, E.: Modeling regularity to improve scalabil-
ity of model-based multi-objective optimization algorithms. In: Multiobjective
Problem Solving from Nature. Natural Computing Series, pp. 331–356. Springer,
Heidelberg (2008)
Larranaga, P., Lozano, J.A. (eds.): Estimation of Distribution Algorithms: A New
Tool for Evolutionary Computation. Kluwer Academic Publishers, Dordrecht
(2001)
Li, X.-D.: A real-coded predator-prey genetic algorithm for multiobjective optimiza-
tion. In: Fonseca, C.M., Fleming, P.J., Zitzler, E., Deb, K., Thiele, L. (eds.) EMO
2003. LNCS, vol. 2632, pp. 207–221. Springer, Heidelberg (2003)
Lim, D., Ong, Y.-S., Jin, Y., Sendhoff, B., Lee, B.S.: Inverse multi-objective robust
evolutionary optimization. Genetic Programming and Evolvable Machines 7(4),
383–404 (2007)
MacKerell Jr., A.D.: Empirical force fields for biological macromolecules: Overview
and issues. Journal of Computational Chemistry 25(13), 1584–1604 (2004)
326 T. Stewart et al.

Mitchell, M.: An Introduction to Genetic Algorithms. The MIT Press, Cambridge

(1996)
Morris, G.M., Goodsell, D.S., Halliday, R.S., Huey, R., Hart, W.E., Belew, R.K., Ol-
son, A.J.: Automated docking using a Lamarckian genetic algorithm and an em-
pirical binding free energy function. Journal of Computational Chemistry 19(14),
1639–1662 (1998)
Nakayama, H.: Aspiration level approach to interactive multi-objective programming
and its applications. In: Pardalos, P.M., Siskos, Y., Zopounidis, C. (eds.) Advances
in Multicriteria Analysis, pp. 147–174. Kluwer Academic Publishers, Dordrecht
(1995)
Nakayama, H., Sawaragi, Y.: Satisficing trade-off method for interactive multiob-
jective programming methods. In: Grauer, M., Wierzbicki, A.P. (eds.) Interactive
Decision Analysis – Proceedings of an International Workshop on Interactive De-
cision Analysis and Interpretative Computer Intelligence, pp. 113–122. Springer,
Heidelberg (1984)
Nakayama, H., Kaneshige, S., Takemoto, S., Watada, Y.: An application of a multi-
objective programming technique to construction accuracy control of cable-stayed
bridges. European Journal of Operational Research 87, 731–738 (1995)
Obayashi, S., Sasaki, D., Takaguchi, Y., Hirose, N.: Multi-objective evolutionary
computation for supersonic wing-shape optimization. IEEE Transactions on Evo-
lutionary Computation 4(2), 182–187 (2000)
Okabe, T., Foli, K., Olhofer, M., Jin, Y., Sendhoff, B.: Comparative Studies on Micro
Heat Exchanger Optimisation. In: Proceedings of IEEE Congress on Evolution-
ary Computation (CEC-2003), pp. 647–654. IEEE Computer Society Press, Los
Alamitos (2003)
Olhofer, M., Arima, T., Sonoda, T., Sendhoff, B.: Optimization of a stator blade
used in a transonic compressor cascade with evolution strategies. In: Parmee,
I. (ed.) Adaptive Computing in Design and Manufacture, pp. 45–54. Springer,
Heidelberg (2000)
Olhofer, M., Jin, Y., Sendhoff, B.: Adaptive encoding for aerodynamic shape opti-
mization using evolution strategies. In: Congress on Evolutionary Computation
(CEC), Seoul, Korea, May 2001, vol. 2, pp. 576–583. IEEE Computer Society
Press, Los Alamitos (2001)
Ong, Y.-S., Nair, P.B., Lim, K.Y.: Max-min surrogate-assisted evolutionary algo-
rithms for robust design. IEEE Transactions on Evolutionary Computation 10(4),
392–404 (2006)
Paenke, I., Branke, J., Jin, Y.: Efficient search for robust solutions by means of
evolutionary algorithms and fitness approximation. IEEE Transactions on Evolu-
tionary Computation 10, 405–420 (2006)
Pettersson, F., Chakraborti, N., Saxén, H.: A genetic algorithms based multiobjec-
tive neural net applied to noisy blast furnace data. Applied Soft Computing 7,
387–397 (2007a)
Pettersson, F., Chakraborti, N., Singh, S.B.: Neural networks analysis of steel plate
processing augmented by multi-objective genetic algorithms. Steel Research In-
ternational 78, 890–898 (2007b)
Poloni, C., Pediroda, V.: GA coupled with computationally expensive simulations:
tools to improve efficiency. In: Genetic Algorithms and Evolution Strategies in
Engineering and Computer Science, pp. 267–288. John Wiley and Sons, Chichester
(1997)
 11 Real-World Applications of Multiobjective Optimization 327

Price, K., Storn, R.N., Lampinen, J.A. (eds.): Differential Evolution: A Practical
Approach to Global Optimizations. Springer, Berlin (2005)
Saxén, H., Pettersson, F., Gunturu, K.: Evolving nonlinear time-series models of
the hot metal silicon content in the blast furnace. Materials and Manufacturing
Processes 22, 577–584 (2007)
Shao, L.: A survey of beam intensity optimization in IMRT. In: Halliburton, T. (ed.)
Proceedings of the 40th Annual Conference of the Operational Research Society
of New Zealand, Wellington, 2-3 December 2005, pp. 255–264 (2005), Available
online at http://secure.orsnz.org.nz/conf40/content/paper/Shao.pdf
Shao, L., Ehrgott, M.: Finding representative nondominated points in multiobjec-
tive linear programming. In: IEEE Symposium on Computational Intelligence in
Multi-Criteria Decision Making, pp. 245–252. IEEE Computer Society Press, Los
Alamitos (2007)
Shao, L., Ehrgott, M.: Approximately solving multiobjective linear programmes in
objective space and an application in radiotherapy treatment planning. Mathe-
matical Methods of Operations Research (2008)
Stewart, T.J., Janssen, R., van Herwijnen, M.: A genetic algorithm approach to
multiobjective land use planning. Computers and Operations Research 32, 2293–
2313 (2004)
Takagi, H.: Interactive evolutionary computation: Fusion of the capacities of EC op-
timization and human evaluation. Proceedings of the IEEE 89, 1275–1296 (2001)
The MathWorks Inc. (2008)
Tsutsui, S., Ghosh, A.: Genetic algorithms with a robust solution searching scheme.
IEEE Transactions on Evolutionary Computation 1(3), 201–208 (1997)
Wierzbicki, A.P.: Reference point approaches. In: Gal, T., Stewart, T.J., Hanne, T.
(eds.) Multicriteria Decision Making: Advances in MCDM Models, Algorithms,
Theory, and Applications, Kluwer Academic Publishers, Boston (1999)
Zhang, Q., Zhou, A., Jin, Y.: RM-MEDA: A regularity model-based multi-objective
estimation of distribution algorithm. IEEE Transactions on Evolutionary Com-
putation 12(1), 41–63 (2008)