Study on initialization algorithm of BP neural network

CHI Kai-li， DONG Lin-xi
（College of Electronic Information Engineering， Hangzhou Dianzi University， Hangzhou 310018， China）

Abstract： In order to solve the problems in the applications of the BP neural network，the algorithm optimizing was investigated. After the analysis of many improving algorithms，the step-by-step initialization method was established. The relationship between initial value with the BP algorithm was focused， and linear equation of the BP neural network was established on the bas of the analysis of various BP optimization algorithm. Optimal initial weights of BP neural network were obtained from the equation. The experimental results show that the new algorithm not only reduces the training time from 13 s to less than 1，but also makes the error curve change into a straight line，avoiding the local minimum of BP algorithm.
Key words： back propagation（BP） neural network； decoupled initialization； sensitive arrears； linear equations； Gause-Jardon method of elimination