为了研究BP神经网络改进学习算法的适用情况,通过对实际的4个应用运用BP神经网络的多种改进的学习算法进行训练,比较得到各学习算法的适用范围,并能根据所研究问题类型、网络大小和要求精度等来选择合适的学习算法.结果表明:LM算法逼近效果好,但不适合大规模网络,RPROP算法应用于模式识别收敛速度最快,但不太适合函数逼近,SCG算法对较大网络规模的性能很好,且逼近效果好.%In order to research the application of the BP neural network improved learning algorithm, four practical applications were trained by improved BP neural network learning algorithm, and through comprising these learning algorithms to obtain the adaptation scope of them. Thus, the actual type of research question, network size and precision to choose appropriate learning algorithm were accorded. The experimental results showed that LM algorithm was good to be the function approximation, but not for large-scale networks; The convergence speed of RPROP algorithm was fast when it was applied pattern recognition, but not so good for function approximation. SCG algorithm for larger-scale network performance was very good, and the effect of function approximation was good.
展开▼
