Complex Systems

Dynamic Neighborhood Structures in Parallel Evolution Strategies Download PDF

Klaus Weinert
Electronic mail address: weinert@isf.mb.uni-dortmund.de.
Department of Machining Technology,
University of Dortmund,
44221 Dortmund, Germany

Jörn Mehnen
Electronic mail address: mehnen@isf.mb.uni-dortmund.de.
Department of Machining Technology,
University of Dortmund,
44221 Dortmund, Germany

Günter Rudolph
Electronic mail address: rudolph@LS11.cs.uni-dortmund.de.
Department of Computer Science,
University of Dortmund,
44221 Dortmund, Germany

Abstract

Parallelizing is a straightforward approach to reduce the total computation time of evolutionary algorithms. Finding an appropriate communication network within spatially structured populations for improving convergence speed and convergence probability is a difficult task. A new method that uses a dynamic communication scheme in an evolution strategy will be compared with conventional static and dynamic approaches. The communication structure is based on a so-called diffusion model approach. The links between adjacent individuals are dynamically chosen according to deterministic or probabilistic rules. Due to self-organization effects, efficient and stable communication structures are established that perform robustly and quickly on a multimodal test function.