Cellular genetic algorithms as function optimizers: Locality effects

Abstract

In this paper, we compare the performance of several cellular genetic algorithms (CGA) to investigate their effectiveness in function optimization. Sometimes called "massively-parallel" GA's, CGA's assign one individual to each processor. Individuals are arranged in a torus and their mating is restricted to within demes (neighborhoods). Three models are considered: fixed topology, random walk, and island. We do not lay to obtain the fastest run times nor do we compare these results to other GA's. A test suite Of numerical functions and Traveling Salesman Problems (TSP) is used. We investigate the interaction between problem difficulty and measuremeats of the spatial locality in the CGA's. Initial results suggest that difficult numerical optimization problems require small demes (high spatial locality), but this is not evident for TSP problems.