自适应调节的隐节点神经网络结构优化算法

Adjustable hidden node algorithm for optimal neural network architecture

为解决现有神经网络算法易陷入局部最优化的问题,根据隐节点对整个网络的贡献值大小,提出一种自适应的动态隐节点调节策略。根据网络的节点容量与输出精度,建立一个资源分配机制,结合隐节点的删除与分裂策略,实现在网络学习过程中对隐节点自适应地增加或删除,形成一个节点利用率高且结构分布合理的神经网络系统。实验结果表明,在不同信号下,该算法能够将误差控制在一个较为合理的数值内,保持较快的收敛速度,有效解决局部最优化问题。

To solve the problem that the neural network algorithm is easy to fall into local optimization,an adaptive strategy of hidden node,according to the size of contribution of hidden nodes on the whole network,was proposed.A resource allocation mechanism was established according to fitting accuracy and output precision of the network.Hidden nodes deletion and split policy were combined,to form the neural networks in which node utilization was high and the structure was reasonably distributed.Experimental results show that this algorithm can get the error value at a more reasonable value and maintain at a rapid convergence rate under different signals.It can effectively solve the problem of local optimization.