基于复合适应度微粒群算法的神经网络训练

Neural Network Training Based on Compound Fitness Particle Swarm Optimization

为提高神经网络的泛化能力,针对以均方误差为适应度的PSO算法在训练神经网络时会产生一定的过拟合问题,提出对均方误差和误差分布均匀度进行信息融合,构成复合适应度作为训练指标.实验结果表明,该方法可使网络的泛化能力得到明显的改善.

The over fitting arises if the PSO（particle swarm optimization） algorithm whose fitness is mean squared eviation is applied in training neural network. In order to improve generalization capacity of feedforward neural network. The compound fitness based on information merging of mean square deviation and error uniformity is proposed as the training index of PSO. The results show that the approach can improve the generalization capacity of feedforward neural networks remarkably.