Research on Overcoming the Local Optimum of BPNN

It´s difficult to satisfy the practical applications for BPNN (back propagation neural network), which has some vital drawbacks including learning slowly and evolving into some local optimum solutions early etc. There are two fundamental reasons which lead to these drawbacks. One reason is that BPNN is too mechanical and elementary to generalize from training experience for lack of advanced intelligent characteristic. The other lies in that BP can not overcome the symptoms including fatigue and overload produced in the process of learning without any powerful mechanism. In order to generalize from training experience and detect-correct the symptomatic neurons, PEBP (phase evaluation BP) is proposed with the phase evaluation mechanism by dividing the process of training into a certain number of phases. At the end of each phase, PEBP evaluate some statistic of current phase and whole training process, then regulate current learning mode and rectify the symptomatic neurons. Compared with the traditional BP, two kinds of PEBP show higher performance from the result of simulated prediction on the same test data set