BP网络微观与宏观弊病的研究

Research on microcosmic and macroscopical defects in BP-NET

BP网络理论基础坚实,通用性强,是前向网络学习的主要模型,但是BP网络也存在一些公认的缺陷,这些缺陷一方面阻碍了BP网络的应用,另一方面也为研究BP网络指明了方向。实验发现这些缺陷产生于两个根本原因——微观的神经元病态和宏观的学习盲目性。经改进的BP网络划分学习阶段,在每个学习阶段结束后的阶段评估中检查和校正病态神经元,评估和调整BP网络的学习模式。实验表明改进的BP网络克服了部分缺陷,提高了性能。

BP-NET has solid theoretical foundation and extensive application, which is the main feedforward learning model. While BP-NET has some generally acknowledged defects, on one hand these defects hinder the application of BP-NET, on the other hand point out the direction of the study of it. The experimental show that these defects evolve from two basic reasons-the microcosmic morbidity of neuron. And the macroscopical blindness of learning this reformative BP-NET divided the whole learning process to many learningphases. It can detect and mend the morbidity of neuron and evaluate and regulate the learning mode in the phase-evaluation after every learning-phase. The experimental result show that these reform improve the performance of BP-NET.