A novel neural network learning method for dynamically tuning regularization coefficient

When network structure has been determined, it is very effective that regulation methods are used to improve generalization ability. However, there are some obvious drawbacks. Based on this, the paper has proposed a novel method that dynamically tune the regularization coefficient by fuzzy rules inference, effectively determined the fuzzy inference rules and membership functions, and implemented the method. Finally, it has compared the method with traditional BP algorithm and fixed regularization coefficient's method through several examples simulations. The results indicate that the proposed method is a very effective method. Compared with other two methods, the proposed method has the merits of the highest precision, rapid convergence, and the best generalization ability.