MULTI-LAYER NEURAL NETWORK LEARNING ALGORITHM BASED ON RANDOM PATTERN SEARCH METHOD

A Random pattern search method (RPS) for learning multi-layer artificial neural network is proposed. As a derivative-free direct search algorithm, the proposed learning model provides a very simple and effective means of searching the. minima of art objective function directly without any knowledge of its derivatives. Furthermore, by incorporating a random mechanism, it also has some chance of escaping from local minima by permitting temporary error increases during learning. Thus the network may eventually reach the global minimum state or its best approximation with higher probability. We test this algorithm on several benchmark problems, such as exclusive-or (XOR), parity, alphabetic character learning, function approximation problems and a real world classification task. Simulation results show that the systems can be trained efficiently by our method for all problems.