Abstract: To model nonlinear systems, we propose a nonlinear modeling method based on an improved deep learning algorithm. First, we design a learning model of a deep belief network based on the Gaussian radial basis function. Second, we adjust the weight, center, and width of the radial basis functions by a contrastive divergence algorithm and optimize the weights of the deep belief network using a back-propagation algorithm. Finally, we apply the improved deep learning algorithm to model nonlinear systems. The experimental results verify the effectiveness and feasibility of the algorithm.