变尺度搜索遗传算法在神经网络训练中的应用

为解决传统的BP学习算法因采用梯度搜索技术而具有的收敛速度慢、容易陷入局部极小等缺点，提出了一种基于多参数空间快速搜索遗传算法的网络权值优化方法.该方法采用误差均分的变尺度搜索技术来训练网络权值，将网络权值的训练转化为多参数空间的寻优问题.训练实例对比结果表明，由于采用了并行计算及变尺度搜索技术，该算法不仅收敛速度快，网络逼近精度高，而且能实现全局最优，克服了BP算法易于陷入局部极小的问题，说明了该方法的有效性.

Variable-scale searching genetic algorithm used in ANN training

To overcome the shortcomings of slow convergence speed and local optimization due to using grads searching technology in traditional BP learning algorithm, a new neural network weight optimizing method based on multi-parameters space rapid searching GA algorithm has been proposed. The algorithm adopts variable-scale searching and error averaging techniques to train the weight of the network, and converts weight training problem into optimization searching in multi parameters space. Experiment shows that the algorithm can not only accelerate convergent speed and achieve higher approximation accuracy of the network, but also can reach collectivity optimization.