Weight Values Optimization of Neural Network using Homotopy Algorithm

Abstract

To overcome the disadvantage of standard BP algorithm with slower learning rate and easily trapping into the local minima, a new method of weight values optimization based on homotopy algorithm is provided in this paper. Minimization of error function of BP neural networks was converted nonlinear equations about weight values vector firstly, and then fixed point homotopy equations were constructed by homotopy mapping and were calculated using adaptive step length Li-York algorithm. Simulation results show the proposed method has both better global convergence and faster learning rate by comparing with standard BP method and momentum method.