Abstract: Analyzes the principle of the basic algorithm of RBF neural network and the problem that the excitation function parameters and the number of hidden layer elements are both selected empirically. A new high-precision algorithm of RBF neural network is proposed according to RBF neural network structure where the threshold and weight values are taken as design variables with network error as objective function, and all of the dynamic variables are ranked reasonably with a computation program given. Compared with the basic algorithm of RBF neural network, the algorithm proposed is really an optimization process taking the threshold and weight values as unknown variables so as to implement the high-precision computation of the RBF neural network. According to the equation theory, a reasonable way is given to determine the hidden layer structure in network. The program analysis by examples indicated that the optimization algorithm has a high goodness of fit with samples and high-precision interpolation results, thus providing the foundation for further theoretic study and engineering applications.