可调参数前馈神经网络及其启发式模拟退火法

Adjustable Parameter FeedforwardNeutral Network and Heuristic Simulated Annealing

针对传统BP神经网络权值算法速度慢、易陷入局部极小等缺陷，在权值平衡算法的基础上，提出一种激励函数参数可调的前馈神经网络，并给出了相应的权值和参数快速学习算法；该算法运用文章提出的非单调启发式模拟退火搜索法实现网络权值和参数的快速搜索；实验表明，该算法不仅能明显提高网络的学习速度。而且可较好地避免学习过程陷入局部极小点而导致学习失败。

Aiming at traditional BP neutral network weight algorithm＇s defect such as slow convergence and easy to plunge into local extremum, an adjustable parameter feedforward neutral network and a fast algorithm to train it are proposed on the basis of weight balance algorithm. Based on the search technique of non--monotone heuristic simulated annealing proposed, the algorithm realizes fast search of network weight and parameter. Experiments show that this algorithm can not only speedup the learning process of network, but also solve the problem of local extremum in learning process in a certain extend.