Non-Linear Variable Selection in a Regression Context

A Bayesian approach to variable selection in a regression context is presented. This aims to find which of a large number of input variables are the important ones in that they contribute to the given regression output. This approach is unlike many in the literature which focus more on features, and do not explicitly seek to include prior belief that many of the input variables do not contribute any information. The EM methodology presented enables this to be done in a nonlinear regression framework, in particular that of kernel regression. An initial experiment on a biscuit dough problem is presented.