一种确定神经网络初始权值的新方法

A new method of determining the initial weights of a neural network

针对BP神经网络对易陷入局部极小的缺点,结合粒子群优化算法(PSO)在全局搜索上的良好性能,提出了一种新的算法———PSO-BP混合算法。该算法先用PSO算法将BP网络的初始权值优化到全局极小点附近,然后用传统BP神经网络学习算法进行进一步优化,仿真表明:该方法很好地解决了BP神经网络对初始值敏感、易局部收敛的问题。

A kind of neural network learning hybrid algorithm called PSO - BP is pressented, because the BP algorithm is apt to plunge into a local minimum and the convergence speed is very low, while the PSO algorithm can find a global optimal solution. In this algorithm, the neural network weights are optimized by using the PSO algorithm and then the accuracy is improved by using the BP algorithm. Simulation shows that the hybrid algorithm has higher accuracy than the BP one.