This paper describes an improved self-organizing CPN-based (Counter-Propagation Network) fuzzy system. Two self-organizing algorithms IUSOCPN and ISSOCPN, being unsupervised and supervised respectively, are introduced. The idea is to construct the neural-fuzzy system with a two-phase hybrid learning algorithm, which utilizes a CPN-based nearest-neighbor clustering scheme for both structure learning and initial parameters setting, and a gradient descent method with adaptive learning rate for 8ne tuning the parameters. The obtained network can be used in the same way as a CPN to model and control dynamic systems, while it has a faster learning speed than the original back-propagation algorithm. The comparative results on the examples suggest that the method is fairly e:cient in terms of simple structure, fast learning speed, and relatively high modeling accuracy.