Constructive Induction on Domain Information

One obstacle to wider use of inductive learning algorithms in problem-solving systems is the sensitivity of the algorithms to the way in which examples of the concept are represented. Humans normally decide how the examples will be represented, so success in incorporating inductive learning algorithms varies from person to person. Constructive induction reduces, but does not eliminate, this sensitivity. An ideal solution would eliminate the need for any human intervention in determining how a problem-solving system and an inductive learning algorithm are integrated. This paper shows how a problem-solver can use its domain knowledge to automatically create a representation of examples that is adequate for learning search control knowledge. The resulting representation describes the examples in terms of how and how well they satisfy the problem-solver's goals. Experimental evidence from two domains is presented to support the claim that this approach is generally useful.