Dynamic training subset selection for supervised learning in Genetic Programming

Abstract

When using the Genetic Programming (GP) Algorithm on a difficult problem with a large set of training cases, a large population size is needed and a very large number of function-tree evaluations must be carried out. This paper describes how to reduce the number of such evaluations by selecting a small subset of the training data set on which to actually carry out the GP algorithm.

Three subset selection methods described in the paper are:

• Dynamic Subset Selection (DSS), using the current GP run to select ‘difficult’ and/or disused cases,

• Historical Subset Selection (HSS), using previous GP runs,

• Random Subset Selection (RSS).

Various runs have shown that GP+DSS can produce better results in less than 20% of the time taken by GP. GP+HSS can nearly match the results of GP, and, perhaps surprisingly, GP+RSS can occasionally approach the results of GP. GP+DSS also produced better, more general results than those reported in a paper for a variety of Neural Networks when used on a substantial problem, known as the Thyroid problem.