Based on the analysis of the internal relation between learning ability and generalization ability of the overfitting of BP neural network

by the introduction of mulit-correlative coefficient to manifest the complexity of the function

a quantitative uncertainty relation between the fitting relative error of training sample sets and the testing relative error of verifying sample sets

which describe the learning ability and the generalization ability of BP network

respectively

was revealed in the overfitting of BP neural network.Tests of numerical simulation for multi-kinds of different functions were carried out to determine the value distribution (0.007~0.07)of overfitting parameter

q

in the relation.Based on the quantitative relation

the training method for the improvement of generalization ability in the training process of sample sets using BP neu

ral network was given.