Abstract: An improved evolutionary method based on real-number encoding is presented to optimize the connection weights and the topology of neural networks. The algorithm could adaptively adjust magnitude of mutation according to individual fitness, and mutation rate will increase with evolving generations as soon as evolution gets into stagnancy. Experiments show that the evolutionary artificial neural network is efficient to predict tool wear in electrical discharge milling machining and the prediction results are better than the standard BP neural networks. The proposed prediction model can be used for tool compensation on-line in electrical discharge milling machining.