Abstract
We implement genetic algorithm based predictive model building as an alternative to the traditional stepwise regression. We then employ the Information Complexity Measure (ICOMP) as a measure of model fitness instead of the commonly used measure of R-square. Furthermore, we propose some modifications to the genetic algorithm to increase the overall efficiency.
Highlights
Variable selection in predictive model building is known to be a difficult procedure
We propose the use of genetic algorithms (GAs) to determine the subset of variables with the highest goodness of fit for a multiple regression model
While model selection remains to be a difficult procedure in case of a large number of parameters, using a GA to find the least complex model can be quite helpful
Summary
Variable selection in predictive model building is known to be a difficult procedure. The probabilities for the selection procedure are chosen arbitrarily, which may lead to a poorly selected model Since these methods employ local search, it is unlikely that the global maximum set of variables will be found (Mantel, 1970; Hocking, 1976, 1983; Moses, 1986). We propose the use of genetic algorithms (GAs) to determine the subset of variables with the highest goodness of fit for a multiple regression model. Due to their global search capabilities, the GA based model building is not prone to the problems associated with local search method, is a wise choice for this procedure.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have