Parallel evolutionary optimisation with constraint propagation

Abstract

This paper describes a parallel model for a distributed memory architecture of a non traditional evolutionary computation method, which integrates constraint propagation and evolution programs. This integration provides a problem-independent optimisation strategy for large scale constrained combinatorial problems over finite integer domains. We have adopted a global parallelisation approach which preserves the properties, behaviour, and theoretical studies of the sequential algorithm. Moreover, high speedup is achieved since genetic operators are coarsegrained, as they perform a search in a discrete space carrying out constraint propagation. A global parallelisation implies a single population but, as we focus on distributed memory architectures, the single virtual population is physically distributed among the processors. Selection and mating consider all the individuals in the population, but the application of genetic operators is performed in parallel. The implementation of the model has been tested on a CRAY T3E multiprocessor using two complex constrained optimisation problems. Experiments have proved the efficiency of this approach since linear speedups have been obtained.

Supported by project TIC95-0433.