一种新型的前馈多层神经网络学习算法

A New Algorithm for Forward Neural Netwroks

利用矩阵ＭｏｏｒｅＰｅｎｒｏｓｅ逆的方法，提出了一种新型的前馈多层神经网络学习算法———ＭＢＰ（ＭａｔｒｉｘＢＰ）算法。该算法采用了群体搜索的策略，打破了ＢＰ算法一次一点的搜索方式，一次可搜索权空间中一个超平面。仿真结果表明，该算法在提高收敛速度和避免陷入局部极小点方面都有一定的进展。

A new algorithm for forward neural networks——MBP algorithm is put forward on the basis of Moore Penrose inverse,in which group searching method is used and a superplane in weight space,not a point,can be searched at a time.The simulation result shows that the algorithm has some progress on the fast convergence and efficient avoidance to be the local minimum.