Abstract: This paper proposes a novel neural-fuzzy inference network and learning algorithm for fuzzy identification of complex systems based on input-output data. The learning algorithm is used for both structure identification and parameter identification of the fuzzy model. In the process of structure identification, a new approach is introduced for rule extraction from input-output data directly. By combining both unsupervised and supervised learning, a hybrid learning algorithm is presented for initial adjustment and optimization of membership functions. Simulations illustrate good performance of the proposed network and learning algorithm in terms of accuracy, readability, number of rules and practicability.