Evolving high-posterior self-organizing maps

Abstract

Bayesian inference for neural networks has received a good deal of attention in recent years. Unlike standard methods, the bayesian approach provides the analyst with the richness (and complexity) of a probability distribution over the space of network weights (and possibly other quantities of interest). These posterior distributions prompt an optimization problem that may be suitable for evolutionary algorithms. This possibility is obviously of foremost interest when no alternative global functions are available for optimization. Some preliminary results related to one of such cases, namely, the self-organizing map, are presented in this paper. Specifically, a familiar "steady-state" diffusion genetic algorithm is described and tested. © Springer-Verlag Berlin Heidelberg 2001.