Local performance of evolution strategies in the presence of noise.

*(English)*Zbl 1140.90300
Dortmund: Univ. Dortmund, Fachbereich Informatik. 158 p. (2001).

Summary: Noise is a common factor in most real-world optimization problems. Sources of noise include, to name but a few, physical measurement limitations, stochastic simulation models, incomplete sampling of large spaces, and human-computer interaction. Evolutionary algorithms are general, nature-inspired heuristics for numerical search and optimization that are frequently observed to be particularly robust with regard to the effects of noise. However, their performance depends on a multitude of parameters, and in combination with fitness environments they form stochastic dynamical systems that are not easily analyzed and understood.

This dissertation contributes to the understanding of evolutionary optimization in the presence of noise by investigating the performance of evolution strategies, a type of evolutionary algorithm frequently employed for solving real-valued optimization problems. By considering simple noisy environments, it is possible to analytically obtain results that describe how the performance of the strategies scales with both parameters of the problem and of the strategies considered. Such scaling laws allow for comparisons of different variants of the strategies, for tuning evolution strategies for maximum performance, and they offer insights and an understanding of the behavior of the strategies that goes beyond what can be learned from mere experimentation.

Issues investigated in this text include the effects of systematic fitness overvaluation, the benefits of distributed populations, and the potential of genetic repair for optimization in the presence of noise. A complete evolution strategy including a mutation strength adaptation component is analyzed. The performance of the algorithm is compared with that of other common optimization strategies, confirming the frequently claimed relative robustness of evolutionary optimization in the presence of noise.

This dissertation contributes to the understanding of evolutionary optimization in the presence of noise by investigating the performance of evolution strategies, a type of evolutionary algorithm frequently employed for solving real-valued optimization problems. By considering simple noisy environments, it is possible to analytically obtain results that describe how the performance of the strategies scales with both parameters of the problem and of the strategies considered. Such scaling laws allow for comparisons of different variants of the strategies, for tuning evolution strategies for maximum performance, and they offer insights and an understanding of the behavior of the strategies that goes beyond what can be learned from mere experimentation.

Issues investigated in this text include the effects of systematic fitness overvaluation, the benefits of distributed populations, and the potential of genetic repair for optimization in the presence of noise. A complete evolution strategy including a mutation strength adaptation component is analyzed. The performance of the algorithm is compared with that of other common optimization strategies, confirming the frequently claimed relative robustness of evolutionary optimization in the presence of noise.

##### MSC:

90-02 | Research exposition (monographs, survey articles) pertaining to operations research and mathematical programming |