A new weight initialization method for the MLP with the BP in multiclass classification problems

Initial learning process of the BP, which can influence the performance of learning in multiclass classification problems, is analyzed. Also, the weights decreasing phenomena in the initial stage of learning are investigated. On the basis of this analysis, a new initialization method is proposed. The proposed method minimizes the initial objective function. It eliminates the phenomenon that weights decrease in the beginning of learning. Several simulation results show that the proposed initialization method performs much better than the conventional random initialization method in the batch mode and slightly better in the pattern mode. Since it requires only a little additional computation, it is a strong alternative to the conventional random initialization. It is expected that the proposed initialization method can be used with any accelerated learning algorithm to enhance the learning speed.