A Hybrid Methodology for Improving Generalization Performance of Neural Networks

A new hybrid approach to improving generalization of feedforward neural networks is presented. The hybrid approach combines several efficient methods, such as feature extraction via wavelet transform, construction of optimal networks architecture via fast cascade-correlation, dynamic optimization of learning parameters via simultaneous determination, avoiding overfitting problem via fast cross-validation. The experimental results show that the hybrid approach can automatically design optimal neural networks with good generalization capability and small network size and short training time in comparison with other ways