Abstract:

Aiming at the model’s reliability problem of a neural network, a trend examination method was presented to check the model’s reliability. It checked the model through the influence trend of evaluation index to evaluation grade. The process of the method was  incessantly adjusting the model’s parameter, training, and trend examination, until  the best model was obtained. This method  presented  a new idea and can be used in any problems of model’s reliability examination based on the foreknowable experience method. To the problem of the contribution difference of samples, the method of weighted samples was used to preprocess the samples and the samples weight was used in the objective function of the neural network. Finally, a genetic algorithm was adopted to optimize the parameter of the  neural network and a GAANN model based on the trend examination method was established. The improved model was applied to practical engineering about surrounding rock classification and the results showed that this method can improve the neural network generalization ability and prediction accuracy.

Key words:  neural network, reliability, samples, weights, surrounding rock classification