A fast learning algorithm of neural network for approximation function was presented by analyzing different characters of parameters in network. This method can speed up the rate of network convergence. By approximating various functions, and comparing with other methods, this algorithm is shown to have satisfactory performances.