Abstract: The traditional stochastic gradient algorithm uses squared error cost function based on second order statistics．It is difficult to achieve higher precision because it contains less information．To solve the problem,a new minimum exponential squared error adaptive learning algorithm was put forward．It uses exponential squared error cost function based on high order statistics,and combines the nonlinear adaptive filter based on convex combination of two RBF networks．The simulation experimental results of nonlinear system identification and nonlinear channel equalization show that the convergence performance of the improved algorithm is superior to the traditional stochastic gradient algorithm．

Key words: Radial basis function neural network,Nonlinear adaptive filter,Stochastic gradient algorithm,Nonlinear system identification,Nonlinear channel equalization