In recent years, combination of soft computing with wavelet theory has eventuated new guidelines. Fuzzy wavelet neural networks which are a combination of fuzzy logic, neural network and wavelet theory, have been used in many researches. The ability of fuzzy wavelet network and extended usage of multi variable systems, have made the motivation of designing multi input multi output fuzzy wavelet network in this research. Also, a new algorithm which is called improved hybrid learning algorithm has been proposed in order to increase training speed of fuzzy wavelet neural network. The fuzzy wavelet neural network is constructed on the basis of fuzzy rules that incorporate wavelet functions in their consequent parts. The ability of fuzzy wavelet neural network in function approximation and system identification has been the subject of many researches. In addition to the inherent ability and property of fuzzy wavelet network, that’s learning process is one of the most important effective items. Many different learning algorithms have been proposed for fuzzy wavelet neural network, but the back propagation method is probably the most frequently used technique in order to train a fuzzy wavelet neural network. Although this algorithm has high ability in finding optimum points, it has some shortcomings. One of the most important of them is its slow convergence to a minimum that is the main topic focused in this study. In this research, an improved hybrid learning algorithm which is a combination of clustering method, recursive least square and accelerated back propagation algorithm is applied in order to train a fuzzy wavelet neural network. In this method, fuzzy wavelet neural network has been learned in three steps. These steps include initialization, optimization of linear parameters and optimization of nonlinear parameters. This proposed method gives the initial parameters by clustering algorithm then updates them with a combination of back propagation and recursive least square methods. The parameters are updated in the direction of steepest descent, but with a local adaptive learning rate which is different for each epoch and only depends on the sign of gradient error function. In order to accelerate the convergence speed, a new idea is applied to determine the learning rate without trial and error. That is iired from the halving method for finding function roots. The convergence condition of the algorithm has been obtained by expressing a theory. Even though the results are much satisfactory, the algorithm is much simpler than other reported. Also it does not include any excessive term in adapting formulation unlike most of researches in this area. Simulation results indicate a superior convergence speed in comparison to other researches Key Words Fuzzy wavelet neural network, Improved hybrid learning algorithm, Multi variable fuzzy wavelet neural network, System identification