A genetic-based input variable selection algorithm using mutual information and wavelet network for time series prediction

In this paper we presented a genetic-based optimal input selection method. This method uses mutual information as similarity measure between variables and output. Based on mutual information the proper input variables, which describe the time series dynamics properly, will be selected. The selected inputs have a maximum relevance with output variable and there exists minimum redundancy between them. This algorithm prepares proper input for wavelet neural network (WNN) prediction model. The WNN prediction model utilized for time series prediction benchmark in NN3 competition and sunspot data. Presented result shows that selected input with GA outperform other input selection method like correlation analysis, gamma test and greedy alg. prediction result indicates that proper inputs have a great impact on prediction efficiency.