基于鲁棒成本函数的随机模糊神经网络参数学习算法

Robust Cost Function Based Parameter Learning Algorithm for Stochastic Fuzzy Neural Network

为了解决采用最小方差型的误差成本函数进行输入含噪系统参数学习时的随机模糊神经网络(SFNN)参数不能收敛至真值的问题,将包含噪声方差的误差成本函数推广到多入单出系统,并根据鲁棒统计学理论和目标函数在参数学习中的导向作用,对目标函数进行修正,使之对于不服从统计分布的粗大误差也能有效处理.在此基础上提出了SFNN的鲁棒参数学习算法,并且输入输出数据中的噪声方差也通过学习而得到,从而避免了需要多次测量的要求.结果表明,SFNN的鲁棒参数学习算法能抑制粗大误差和系统噪声.最后,通过仿真对比验证表明了该方法的有效性.

To solve the problem where the parameters of stochastic fuzzy neural network (SFNN) cannot strongly converge to the true values when using traditional least squares cost function with noisy input data, the cost function which contains the error variables is extended to multi-input single output system, According to robust statistics theory and the directional role of target function, the cost function is mended to treat the gross error and a new robust learning algorithm of SFNN is proposed. The error variables are obtained via learning algorithm to avoid repeating measurement. The simulation results show that the robust algorithm of the SFNN can eliminate the gross error and execute a filtering effect on noisy input.