神经元网络鲁棒能量函数的构造原理

THE PRINCIPLE OF CONSTRUCTING A ROBUST ENERGY FUNCTION FOR NEURAL NETWORKS

本文分析了MLP网络常用的LS型能量函数和BP算法在随机非线性函数逼近中的局限性,由此提出了"学习强度"、"误差中心曲面"等新概念.基于一个结构性定理,本文给出了构造能抵抗噪声干扰的特殊能量函数的一般方法.仿真结果证实,加入噪声后,基于鲁棒能量函数的新算法,在收敛率、平稳性和鲁棒性方面,明显优于通常的BP算法.

By analysing the limitations of LS energy function and BP algorithm of MLP networks in stochastic nonlinear function approximation, some new concepts, such as ' learning intensity ' and ' residual centered surface ' , are proposed. Based upon a constructive theorem, a general method of designning a special energy function, which is resistant to the noise effects, is introduced. Some experimental results demonstrate that the new algorithm using the robust energy function is clearly superior to the popular BP algorithm in convergence rate, stability and robustness when a noise perturtion is exerted.