Agent-Based Approach to RBF Network Training with Floating Centroids

In this paper the agent-based population learning algorithm designed to train RBF networks (RBFN's) is proposed. The algorithm is used to network initialization and estimation of its output weights. The approach is based on the assumption that a location of the radial based function centroids can be modified during the training process. It is shown that such a floating centroids may help to find the optimal neural network structure. In the proposed implementation of the agent-based population learning algorithm, RBFN initialization and RBFN training based on the floating centroids are carried-out by a team of agents, which execute various local search procedures and cooperate to find-out a solution to the considered RBFN training problem. Two variants of the approach are suggested in the paper. The approaches are implemented and experimentally evaluated.