基于PSO的BP训练算法

Approach for optimizing BackPropagation training based on PSO

在BP训练算法中,关于变权值、学习速率、步长的问题已被广泛地研究,几种基于启发式改进的技术也表明具有改善训练时间以及避免陷入局部最小的明显效果。这里BP训练过程由基于PSO同时优化log-Sigmoid函数与网络权值的新算法(PSO-GainBP)实现。实验结果表明,PSO-GainBP比传统基于PSO的BP算法在网络训练方面具有更好的性能。

The issue of variable weight, learning rate, step size and bias in the Backpropagation training algorithm has been widely investigated. Several techniques employing heuristic factors are suggested to improve training time and reduce convergence to local minima. Backpropagation training is based on a new approach in which variable gain of the log-sigmoid function and weights of the network are both optimized by using particle swarm optimization technique. The new approach is implemented and the training result demonstrates that the new approach is more efficient than usual BP algorithm based on PSO.