Informatik Centrum Dortmund
Center for Applied Systems Analysis
Joseph-von-Fraunhofer-Str. 20
D-44227 Dortmund
Evolutionary algorithms are probabilistic search and optimization methods gleaned from the model of organic evolution. Using a population of search points, stochastic variation operators imitating recombination and mutation, and a selection method that favors better search points for survival and propagation into the next generation, these algorithms yield surprisingly good solutions for a variety of optimization problems in industrial as well as research applications.
This talk will provide a perspective on evolutionary algorithms that emphasizes their practical as well as theoretical aspects.
Concerning the practical applications, the talk will give an overview of some of the actual applications of evolutionary algorithms at the Center for Applied Systems Analysis, including the optimization of nuclear reactor core reload designs, of optical multilayer coatings, and of routing in telecommunication networks.
Concerning the theoretical aspects, the talk will focus on the
investigations of convergence velocity and reliability.
Moreover, the important parameter control mechanism of self-adapting
the strategy parameters, which facilitates the simultaneous
search on the object variable and strategy parameter levels,
will be discussed. After giving an introduction to this principle,
recent theoretical results concerning the robustness of this parameter
adjustment method are summarized.
Real-Parameter Genetic Algorithms in Engineering Design
Most real-world search and optimization problems involve object variables which are of different types: (i) Boolean (ii) Real-valued, and (iii) Discrete. Classical optimization methods are particularly not very efficient in solving such mixed-variable problems. Researchers resort to some penalty-based methods or iterative approximation methods to solve such problems.
Evolutionary algorithms (EAs) offer a better way to solve such problems, primarily because of their use of a flexible representation scheme and their use of operators which do not require any additional information other than objective and constraint values. Of different EA variants, genetic algorithms (GAs), evolution strategies (ESs), and evolutionary programming (EP) have been widely used for solving problems with mixed variables.
Although binary-coded GAs have been used in handling real-valued object variables, there exists a number of difficulties such as achieving arbitrary precision, Hamming cliff problem, and others, which make binary-coded GAs inconvenient to be used in problems having real-valued variables.
In this tutorial, we shall briefly review a number of real-parameter GAs, where object variables are used directly. Although ES and EP methods also use object variables directly, the main difference between real-parameter GAs and ES or EP is that GAs mainly rely on their crossover operator and ES or EP mainly rely their search on mutation operator.
We shall particularly discuss a real-parameter GA with an efficient crossover operator known as the Simulated Binary Crossover (SBX). Simulation results of GAs with SBX and a discrete version of SBX will be shown to solve a number of mechanical component design problems of mixed-variable type.
Finally, we shall demonstrate that the real-parameter GAs with SBX has
self-adaptive properties, which have been shown to exist in
evolution strategies and evolutionary programming. All these results will
demonstrate the efficacy of the real-parameter GA with the SBX operator and
will suggest their use in more complex engineering design problems.
Genetic algorithms and Game Theory tools for solving multi criteria
optimization CFD/CEM problems
Dassault Aviation
Genetic Algorithms (GAs) are inspired from the darwinian mechanisms of evolution.
Unlike classical deterministic optimization tools GAs work with a population of potential solutions represented by chromosomes and genes which evolve during the optimization process.
Selection, cross over and mutation are the semi stochastic genetic operators which allow GAs to explore very hilly search spaces of complex problems and capture a global solution.
GAs are run in the computer on a natural selection mode since the information technology of the genes of living organisms is digital.
Decisive advantages of GAs in complex industrial environment are their simplicity and robustness.
The tutorial will introduce step by step on simple mathematical functions the mechanisms of GAs in the context of Game Theory like Pareto,Nash and Stackelberg games.
Several numerical experiments of GAS on industrial problems related to CFD optimal shape design and CEM low observability control problems arising in aerospace engineering will be presented.
The results demonstrate the potential of these emergent evolutionary methods for solving new multidisciplinary designs of increasing complexity.