Use of Domain Knowledge in Constructive Induction

One of the important problems in integrating inductive learning algorithms with problem-solving systems is determining how they communicate. It is well-known that inductive algorithms are sensitive to the vocabulary with which examples of a concept are described. It is also known that a vocabulary can be acceptable for problem-solving but cause the inductive algorithm to learn slowly or inaccurately. It remains an open question how best to choose a language with which a problem-solving system and an inductive algorithm communicate. This thesis proposal considers the representation problem for a class of architectures in which the problem-solver is a search procedure and the inductive learning algorithm is a source of heuristic guidance. It presents a solution in which the problem-solver uses its domain knowledge to generate automatically a set of features suitable for learning search control knowledge. Each new feature describes the problem-solver's progress in achieving one of its goals. Experimental evidence from two domains supports the claim that this solution results in faster and more accurate learning. A program of additional research and experiments is described that is expected to provide further support for the solution.