Parsing to learn fine grained rules
Subrata Roy, Jack Mostow · National Conference on Artificial Intelligence · 1988
The grain size of rules acquired by explanation-based may vary widely depending on the size of the training examples. Such variation can cause redundancy among the learned rules and limit their range of applicability. In this paper, we study this problem in the context of LEAP, the learning apprentice component of the VEXED circuit design system. LEAP acquires circuit design rules by analyzing and generalizing design steps performed by the user. We show how to reduce the grain size of rules learned by LEAP by using synthetic parsing to extract parts of the manual design step not covered by existing design rules and then using LEAP to generalize the extracted parts. A prototype implementation of this technique yields finer grained rules with more coverage. We examine its effects on some problems associated with the explanation-based technique used in LEAP.