Evolving Product Unit Neural Networks with Particle Swarm Optimization

Product unit neural network (PUNN) training is formulated as an optimization problem and then particle swarm optimization (PSO), an emerging evolutionary computation algorithm, is employed to resolve it. A simple and effective encoding scheme for particles is proposed by which PSO algorithm can configure the architecture and weight of PUNN simultaneously depending on training sets. Because the training algorithm takes into account not only network error but also the complexity of network, the resulting networks alleviate over-fitting. Experimental results show that proposed algorithm achieves rational architecture for PUNN networks and the resulting networks obtain strong generalization abilities.