Wang Junping,Chen Quanshi
(State Key Laboratory of Automobile Safety £¦ Energy Conservation, Tsinghua University,Beijing 100084,China)
Abstract:To solve the problem where the parameters of stochastic fuzzy neural network £¨SFNN£©can not 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 funcction 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 repeasted 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£®
Keywords:stochastic fuzzy neural network;error cost function;robust parameter learning algorithm