Research on Structure Learning of Product Unit Neural Networks by Particle Swarm Optimization

摘要

In this study, we put forward a new method to learn structure of Product Unit Neutral Network (PUNN). The technique used in our research is based on Particle Swarm Optimization (PSO) algorithm. The technique can optimized collocate network structure and weight of the PUNN at the same time using PSO algorithm through standard data set. Moreover, the number of Hidden Layer units of PUNN is decided by training set, not prefixed by the designer's prior knowledge. Particles encoding scheme is simple and effective. The design of fitness function considers not only the mean square error between networks output and desired output, but also the number of hidden layer units. Therefore, the resulting network can alleviate the problem of over-fitting. The results of the experiment indicate that PSPUNN algorithm can achieve rational architecture for PUNN relying on standard data set and the resulting networks hence obtain strong generalization abilities.