A new regularization learning method for improving generalization capability of neural network

When the network structure has been determined, it is very effective when regulation methods are used to improve generalization capability. However, there are some obvious drawbacks: long computation time, and difficulty in parameter control. The paper proposes a novel method that dynamically tunes the regularization coefficient by fuzzy rule inference, effectively determining the fuzzy inference rules and membership functions. Furthermore, it compares the method with the traditional BP algorithm and the fixed regularization coefficients method. The proposed method has the merits of the highest precision, rapid convergence, and the best generalization capability. Finally, it indicates that the proposed method is a very effective method by several nonlinear function approximation and pattern classification examples.