Extra tion of Rules from Arti ial

Networks for Nonlinear Regression Rudy Setiono S hool of Computing National University of Singapore Kent Ridge, Singapore 119260 Wee Kheng Leow S hool of Computing National University of Singapore Kent Ridge, Singapore 119260 Ja ek M. Zurada Dept. of Ele tri al and Computer Engineering University of Louisville Louisville, KY 40208, USA Abstra t Neural networks have been su essfully applied to solve a variety of appli ation problems in luding lassi ation and fun tion approximation. They are espe ially useful as fun tion approximators be ause they do not require prior knowledge of the input data distribution and they have been shown to be universal approximators. In many appli ations, it is desirable to extra t knowledge that an explain how the problems are solved by the networks. Most existing approa hes have fo used on extra ting symboli rules for lassi ation. Few methods have been devised to extra t rules from trained neural networks for regression. This arti le presents an approa h for extra ting rules from trained neural networks for regression. Ea h rule in the extra ted rule set orresponds to a subregion of the input spa e and a linear fun tion involving the relevant input attributes of the data approximates the network output for all data samples in this subregion. Extensive experimental results on 32 ben hmark data sets demonstrate the e e tiveness of the proposed approa h in generating a urate regression rules.