A new algorithm-resolution of overlapping chromatographic peaks by radial basis function neural network(RBFNN) is presented. A two-phase genetic algorithm(GA) which has robustness and random globe optimization is used to train RBFNN so that it has the ability on the resolution of overlapping chromatographic peaks. The two-phase genetic algorithm involves two procedures: training structure and optimizing parameter. The first procedure uses GA to train the architectures of RBFNN, the second procedure uses gradient descent to train the center(tR) and the width(sigma) of RBFNN. The alternate use of these two procedures makes the network having the ability to learn structure, therefore makes itself adaptable to resolution of the chromatographic peaks with unknown number of components. The method proposed here needs no artificial interference, not only has it robustness and globalism, but also the ability of accurate resolution to completely overlapped chromatographic peaks. The simulation experiments show that this method is more accurate than other methods.