Optimizing the Number of Learning Cycles in the Design of Radial Basis Neural Networks Using a Multi-Agent System

Abstract

Radial Basis Neural (RBN) network has the power of the universal approximation function and the convergence of those networks is very fast compared to multilayer feedforward neural networks. However, how to determine the architecture of the RBN networks to solve a given problem is not straightforward. In addition, the number of hidden units allocated in an RBN network seems to be a critical factor in the performance of these networks. In this work, the design of RBN network is based on the cooperation of n+m agents: n RBN agents and m manager agents. The n+m agents are organized in a Multi-agent System. The training process is distributed among the n RBN agents, each one with a different number of neurons. Each agent executes a number of training cycles, a stage, when the manager decides about that is the best RBN agent and sends it the corresponding message. The m manager agents have in charge to control the evolution of each problem. Each manager agent controls one problem. Manager agents govern the whole process; each one decides about the best RBN agent in each stage for each problem. The results show that the proposed method is able to find the most adequate RBN network architecture. In addition, a reduction in the number of training cycles is obtained with the proposed Multi-agent strategy instead of sequential strategy.