Improving generalization performance of artificial neural networks with genetic algorithms
Jiansheng Wu, Mingzhe Liu · 2005
The focus on the study of artificial neural networks (ANN) is how to balance the trade-off of the goodness-of-fit in the training sample and the next-step-predictability in the testing sample. In this paper a novel optimization approach ERCNN (Evolving Regularization Coefficient and Neural Network) is proposed. The non-linear function approximation and sunspot time series forecasting problems are used to validate the network performance of our proposed approach. Numerical results show that both accuracy and generalization abilities of our proposed approach outperform the traditional back propagation (BP) algorithm and fixed regularization coefficient (RC) method. The examples demonstrate that our approach is feasible and valid.