Abstarct: Due to the increasing use of neural networks in areas such as data classification, function approximation and time series prediction, many efforts have been made to overcome their deficiencies. One of the reasons that limits the application of neural networks is their training stage. To train neural networks, we need a large group of input data, while setting up the training parameters of the network is difficult and requires experience. Also, some problems such as local minimas and convergence to an inappropriate response are the obstacles of neural networks. In this thesis, training and prediction of the artificial neural network structure have been proposed by the Particle Swarm Optimization (PSO) algorithm for classifying several benchmark datasets and approximate some benchmark function. In the proposed methodology, multilxayer neural network with PSO algorithm for classifying and predicting the network structure and the combination of RBF and PSO are also used to approximate functions and predict their structure. Comparison of the results with similar tasks that have been performed and the results of the MATLAB software Toolbox indicate the excellence of proposed algorithm. For example, for several multi-class datasets, an accuracy of nearly 100% has been achieved, and in the approximation task, for the majority of continuous datasets, a complete estimation is made, while the proper structure of network is predicted.