Based on bias variance model, a novel method of dynamically tuning the regularization coefficient by fuzzy rules inference was proposed. The fuzzy inference rules and membership functions were effectively determined. Furthermore, the method was compared with the traditional BP algorithm and fixed regularization coefficien's method. The result is that the proposed method has the merits of the highest precision, rapid convergence and best generalization capacity. The capacity proposed method is shown to be a very effective method by several examples simulation.