一种新的全局优化前馈神经网络

A New Global Optimization BP Neural Networks

将 L-M算法与填充函数法相结合,提出一种训练前向网络的混合型全局优化新算法.L-M算法的收敛速度快,利用它先得到一个局部极小点,然后利用填充函数算法跳出局部最小,得到一个更低的局部极小点.重复计算即可得到全局最优点.经实验验证,该算法收敛速度很快,避免局部收敛,而且性能稳定.

This paper proposes a global optimization technique which combines the filled function method and Levenberg-Marquardt algorithm for training feed forward neural networks.With the L-M algorithm whose astringency is good ,we can find one of local minimal points quickly.Afterwards ,the filled function method will find the point that is lower than the minimal point previously found.By repeating these processes,a global minimal point can be obtained at last.Practical examples indicate that the method has a higher accuracy in astringency and works well to avoiding sticking in local minima.