A weight evolution algorithm for multi-layered network

In spite of the general success of backpropagation, it still has several weaknesses. First, it has the possibility of being trapped at local minima during learning. Second, the convergence rate is typically too slow even if learning can be achieved. In this paper, we present a weight evolution algorithm (WEAL) for multilayered network to overcome the problems of the back-propagation algorithm. The basic idea is to evolve the weights under suitable controls during the learning phase of back-propagation so as to bypass all the local minima and to improve the convergence rate. A mathematical framework of the new algorithm is also given to ensure that the perturbation of weight can achieve a better error performance. Simulation results are used to illustrate the fast learning behavior and the global search capability of the algorithm in improving the performance of back-propagated networks