View Related Documents

Abstract

The fine grain, data-driven parallelism shown by neural models as the Boltzmann machine cannot be implemented in an entirely efficient way either in general-purpose multicomputers or in networks of computers, which are nowadays the most common parallel computer architectures.
In this paper we present a parallel implementation of a modified Boltzmann machine where the processors, with disjoint subsets of neurons allocated, asynchronously compute the evolution of their neurons by using values that might not be updated for the remaining neurons, thus reducing interprocessor communication requirements. An evolutionary algorithm is used to learn the rules that allow the processors to cooperate by interchanging the local optima that they find while concurrently exploring different zones of the Boltzmann machine state space. Thus, the way the processors interact changes dynamically during execution of the algorithm, adapted to the problem at hand. Good figures for speedup with respect to the Boltzmann machine computation in a uniprocessor computer have been experimentally obtained.

Key words  Boltzmann machines - combinatorial optimization - evolutionary computation - multicomputers and networks of computers - parallel processing

Fulltext Preview

Image of the first page of the fulltext document