Computational Intelligence - August 2013 - 13

aware that much effort needs to be
expended in solving a number of issues
even before the actual optimization can
be conducted. The first issue is the problem formulation, including definition of
the objectives, constraints and the representation, which defines the decision
variables. Difficulties may arise from the
fact that many CEO problems consist of
a number of sub-systems or sub-processes that are inter-dependent, and optimization of the individual sub-systems
separately may not lead to a globally
optimal solution. At the same time, the
representation of the complete problem
at once is often prohibited by the very
high dimensionality. Problems from
aerodynamic shape optimization are
good examples, where a holistic representation of e.g. a racing car would easily
consist of thousands of decision variables.
Therefore, representations have to be
defined which are inherently incomplete
covering only part of the design space.
Methods to cope with this deficiency are
required, e.g. by choosing an optimal
spatio-temporal decomposition of the
problem or by adaptively changing the
representation during the search process
focusing on those design areas which are
highly sensitive.
In addition, CEO problems may not
be easily described by explicit mathematical models and are subject to a large
amount of uncertainty. Understanding
and properly formulating an optimization problem is itself part of the overall
problem-solving process and it typically
requires several iterations between the
optimization expert and the application
engineer often also involving a simulation specialist working on the particular
problem. Many simulation methods are
iterative approaches (e.g. CFD or FEM)
with a residual error and a problemdependent setup (mesh type and size)
that strongly interacts with the optimization methodology. Another typical
class of CEO problems are plant-wide
product process optimization [7], e.g.,
optimization of the global operation of
mineral processing, which is composed
of multiple coupled processes such as
ore crushing, grinding and regrinding,
and selection. No exact mathematical

models are available to describe such
processes and several objectives including product quality, energy efficiency
and productivity need to be satisfied.
One particularly interesting issue in such
process optimization problems is that
both optimization and control are
involved. An integrated control and
optimization strategy may not only lead
to global optimization, but may also
offer a new approach to deal with
uncertainties that optimally balance
dynamic optimization and robustness.
Another example is aircraft design,
where various parts of an aircraft, e.g.,
fuselage, wing and tail must be designed
in an integrated and holistic way to
ensure that each part is designed for the
optimization of the whole aircraft with
respect to multiple objectives including
energy efficiency, emission reduction
and safe operations. To deal with the
optimization of such CEO problems, a
systems engineering perspective, i.e. a
holistic problem view must be taken as
suggested in [7], [10].
Similar challenges can be identified
for the optimization of passenger cars,
where not just the optimization criteria
from several engineering disciplines
such as structural safety (crash), aerodynamics and thermodynamics have to be
integrated, but also issues of aesthetic
design, cost efficient manufacturing, and
product disposal (recycling) have to be
taken into account. The optimization
framework basically embraces and interrelates all segments in the product life
cycle management process that is the
backbone of most complex engineering
problems. In a sense, the optimization framework itself is hierarchically
organized consisting of many sub optimization problems that are allowed to
operate on different time scales from
minutes to months and that need to
interact with each other and with the
respective decision makers during the
complete development, procurement,
manufacturing and service processes. In
real-world challenges tuning the algorithm to be embeddable into such a complex framework is often more relevant
than providing optimization results that
are marginally better on a limited test

suite the state-of-the-art methods. This
shall not belittle the efforts that have
been taken and are taking to improve
numerical and combinatorial optimization methods, but shall emphasize that
often the needs in a practical CEO challenge are different. At the same time,
evolutionary algorithms (EAs) including
other meta-heuristics are very promising
candidates and approaches for fitting
into complex design frameworks,
because of their inherent robustness,
flexibility and adaptability.
Optimization algorithms developed
for solving CEO problems must be scalable to the number of decision variables
as well as to the number of objectives
and be able to deal with uncertainties,
by tracking the optimum or finding
robust optimal solutions or by identifying optimal (acceptable) solutions that
are robust over time when frequent
change of solution is prohibitive. Whilst
algorithms proposed for solving largescale and many-objective optimization
tasks are very helpful, it is equally desirable if the optimization algorithm is able
to identify a small number (two to three)
of the most critical objectives. To this
end, preference-based interactive search
may be more tractable than an uninformed search aiming to find all Paretooptimal solutions, if the number of
objectives cannot be reduced. This is one
example where engineering data mining, knowledge acquisition, see e.g. [13],
and visualization techniques become
more and more important to guide the
interactive search process, to formulate
the initial problem or to identify the
most appropriate problem representation, which is directly related to reducing the number of decision variables.
Here, we naturally come to another
important question that may arise in
dealing with CEO problems, i.e., how to
make sure that a developed optimization
algorithm is able to gain problem-specific knowledge during optimization so
that the search is more efficient, adaptable and well prepared for change. To this
end, hyper-heuristics [11] that systematically integrate optimization and learning
techniques can be promising approaches.
Incorporation of learning techniques to

August 2013 | IEEE ComputAtIonAl IntEllIgEnCE mAgAzInE

Table of Contents for the Digital Edition of Computational Intelligence - August 2013