Input Variable Selection for Non-Parametric Regression, Classification, and TimeSeries Modeling

摘要

Variable selection is a critical step in constructing statistical regression,pattern classification, or time series models that are capable of optimum generalization performance. Since the project got started in February 1996, we have implemented the prototype K-test as proposed, carried out extensive testing on regression and time series problems, and developed a selection criterion based upon unsupervised clustering methods. The latter can be applied to both regression and classification type problems. Under ONR sponsorship, a number of criterion functions have been devised and tested for developing the variable selection methodologies. The work on this project has been conducted by Hong Pi and John Moody. Since Hong Pi has taken a job in industry, Howard Yang (from Amari's research group in Tokyo) will continue working on the project in place of Hong.