Abstract:

In view that the traditional wavelet neural network (WNN) was affected largely by the number of hidden layer nodes, easy to fall into local minimum and had unstable forecast results, a method of combining the GentleAdaBoost algorithm with WNN was put forward to improve the forecasting accuracy and generalization ability. First, this method performed the pretreatment for the historical data and initialized the distribution weights of test data. Second, different hidden layer nodes and wavelet basis functions were selected randomly to construct weak predictors of WNN and trained the sample data repeatedly. Finally, the multiple weak predictors of WNN were used to form a new strong predictor by GentleAdaBoost algorithm for regression forecasting. A simulation experiment using datasets from the UCI database was carried out. The results showed that this method had reduced the average error value by more than 40％ compared to the traditional WNN, improved the forecasting accuracy of neural network, and could provide references for the WNN forecasting.

Key words: basis functions, strong predictor, Gentle AdaBoost algorithm, wavelet neural network, iterative algorithm, regression forecasting