Acquiring General Iterative Concepts by Reformulating Explanations of Observed Examples.
Abstract
Most research in explanation-based learning involves relaxing constraints on the variables in the explanation of a specific example, rather than generalizing the structure of the explanation itself. However, this precludes the acquisition of concepts where an iterative process is implicitly represented in the explanation by a fixed number of applications. Such explanations must be reformulated during generalization. The fully-implemented BAGGER system analyzes explanation structures and detects extendible repeated, inter-dependent applications of rules. When any are found, the explanation is extended so that an arbitrary number of repeated applications of the original rule are supported. The final structure is then generalized and a new rule produced which embodies a crucial shift in representation. An important property of the extended rules is that their preconditions are expressed in terms of the initial state-they do not depend on the results of intermediate applications of the original rule. BAGGER's generalization algorithm is presented and empirical results that demonstrate the value of generalizing to N are reported. To illustrate the approach, the acquisition of a plan for building towers of arbitrary height is discussed in detail. Keywords: Artificial intelligence, Machine learning, Explanation-based learning, Empirical analysis.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1987
- Accession Number
- ADA190665
Entities
People
- Gerald F. Dejong
- Jude W. Shavlik
Organizations
- University of Illinois Urbana–Champaign