APPLICATION OF RECURSIVE ORTHOGONAL LEAST SQUARES ALGORITHMS TO TRAINING AND THE STRUCTURE OPTIMIZATION OF RADIAL BASIS PROBABILISTIC NEURAL NETWORKS

摘要

This paper introduces applying recursive orthogonal least squares algorithm (ROLSA) to training radial basis probabilistic neural networks (RBPNN) and selecting their hidden centers. Firstly ROLSA is used to solve the weights between the second layer and the output layer of RBPNN. Secondly we interpret the basic principle of selecting hidden centers and give a detailed selection procedure. In addition, we deduce the solution of orthogonal decomposition terms under the condition of varying centers. Finaly two-spirals problem is presented to testify the effectiveness and efficiency of our algorithms. The experimental results show that our algorithm is very effective and feasible.