基于改进的PSO算法的神经网络集成

提出了一种新的神经网络集成结论生成方法，即基于可重复采样技术(Bootstrap)的粒子群优化(PSO)算法——BPSO算法，通过限制组合权值的范围来减小“多维共线性”的影响，还利用采样技术构造不同的适应度函数，增加“粒子”的多样性从而便于在一定范围内灵活调节组合权值，并减小噪声对集成的影响．实验表明。BPSO算法是优化组合权值的有效方法，提高了神经网络集成的泛化能力．

Neural Network Ensembles Based on Improved PSO Algorithm

An improved PSO algorithm based on Bootstrap named with BPSO for neural network ensembles is proposed. On the one side, the effect of collinearity is decreased by restricting the range of combination weights. On the other hand, different fitness functions are constructed on different data by using bootstrap technique, which increases the diversity of particles. In consequence combination weights are easier to be adjusted in a definite range and the effects of noise are decreased. Experiments results show that the BPSO algorithm is an effect ensemble method and improves the generalization ability of neural network ensembles.