前馈式神经网络的非梯度学习方法及在智能控制中的应用

The slow learning speed problem of multi-layer forward neural network has become the “bottleneck” which seriously influenses its wider application. The project deeply analyzes the characters of the structure of neural network and the disadvanges of BP algorithm. A new non-grads learing algrithm,named SPDS algorithm, is presented in order to increase the learning speed of.multi-layer forward neural network. The algorithm is an important breakthrough.in learing algorithm and the advantages of the method over the mostly applied.BP algorithm is theoretically proved. SPDS algorithm can overcome the network.paralysis. Its calculational complexity is only expressed as an exponent of the.parameters of the structure of neural network and it can easily come to the global minimum. Meanwhile, the multiplication of calculating the objective function and the times of using the activating function of working out error function are greatly redused. SPDS algorithm has obvious faster convergent speed than BP algorithm for the kind of practical problem with huge numbers of samples and characters. In fact, SPDS algorithm has been successfully applied in the many fields, such as the stable control of the electrode moments, the jointing handled by intelligent robots,the intelligent medical diagnostic system,etc. In due course, it will advance the application of neural network.

学习速度慢已成为“瓶颈”严重影响神经网络在智能控制领域中的推广和应用。本项目深入分析前馈式神经网络模型结构的特点，论证它对于BP算法带来的弊端。提出一种非梯度学习算法。它是学习算法的重要突破。本项目将从各种角度论证它性能上的优越性。对于规模稍大的实际问题，它收敛速度可上百地快于BP算法。从而为神经网络的应用推进一步。