Abstract: An on-line radial basis function neural networks modeling method for nonlinear dynamics system, called resource-optimizing network (RON), is presented. RON introduces the ideal of sliding window and on-line structure optimization to the standard learning process of resource-allocating network. According to the newest error information, RON optimizes network structure on-line to adapt to the change of system dynamics, and maintains a compact network and satisfactory generalization. With the use of sliding window, RON is robust against the changing of learning parameters, and is easy to converge. Three benchmark examples demonstrate the effectiveness of the method.