Abstarct: One of the most important research fileds in the world is person identification. In this reserch mixture of expert network is suggested for writer classification and identification in Persian handwriting, which is a technique for person identification baxsed of biometrics features. It is suggested to combine three neural networks that are trained baxsed on three different text features extracted by Roberts, Sobel and Gabor filters. The numbers of extracted features by different approaches are reduced using principal component analysis algorithm, and k nearest neighbor algorithm is also used to find the optimal number of confident features. Three neural networks are trained baxsed on gradient descent and error back propagation procedures. Outputs of these networks are combined by using a set of weights which are obtained by four different methods. In the first experiment we combine neural network outputs by equal weights so each neural network which is trained by different feature sets has equal effect in the result of classification and writer identification. In the second experiment confidence of each network has been used as weight of that neural network. In the third experiment the PSO optimization algorithm is used to minimize the mean square error in training stage and provide a set of weights for combining the networks. In the last experiment the PSO is used again with this difference that we assign different weigths to outputs of each network so instead of 3 weights in third method we must find 120 weigths for output neurons of neural networks and combining corresponding neurons of each neural networks together to obtain finally 40 output neurons. Simulation results show the high accuracy of writer identification up to 97.42%, 97.75%, 97.59% and 98.17% respectively for different proposed methods which are evidently higher than accuracy obtained in each single neural network.