Classification is important in data mining and machine learning. In this paper, a classification approach based on evolutionary neural networks (CABEN) is presented, which establishes classifiers by a group of three-layer feed-forward neural networks. The neural networks are trained by an improving algorithm synthesizing modified Evolutionary Strategy and Levenberg-Marquardt optimization method. The class label of the identifying data can first be evaluated by each neural network, and the final classification result is obtained according to the absolute-majority-voting rule. Experimental results show that the algorithm CABEN is effective for the classification, and has the better performance in classification precision, stability and fault-tolerance comparing with the traditional neural network methods, Bayesian classifiers and decision trees, especially for the complex classification problems with many classes.