Learning to Represent Codons: A Challenge Problem for Constructive Induction

Abstract

The ability of an inductive learning system to find a good solution to a given problem is dependent upon the representation used for the features of the problem . Systems that perform constructive induction are able to change their representation by constructing new features . We describe an important, real-world problem - finding genes in DNA - that we believe offers an interesting challenge to constructiveinduction researchers. We report experiments that demonstrate that: (1) two different input representations for this task result in significantly different generalization performance for both neural networks and decision trees; and (2) both neural and symbolic methods for constructive induction fail to bridge the gap between these two representations. We believe that this real-world domain provides an interesting challenge problem for constructive induction because the relationship between the two representations is well known, and because th e representational shift involved in constructing the better representation is not imposing.