Refining Representations to Improve Problem Solving Quality

• Published in: Machine Learning Proceedings 1989 pp. 457 - 460
• Main Author: SCTILIMMER, JEFFREY C.
• Format: Book Chapter
• Language: English
• Published: Elsevier Inc 1989
• ISBN: 1483297403; 9781558600362; 1558600361; 9781483297408
• DOI: 10.1016/B978-1-55860-036-2.50116-8

This chapter describes a learning method involving the refining of representations to improve problem solving quality of systems. Declarative, domain-independent problem solving is a widespread and effective technique for addressing artificial intelligence problems. It relies on a domain-independent inference engine and a declarative representation of domain-specific knowledge. The flexibility afforded by this approach is frequently offset by limitations in efficiency and competence of the problem solver. An alternative approach for more efficiency in processing systems is to do away with macro-inference rules and use machine learning methods to build appropriate representations explicitly. Useful problem-solving distinctions are explicitly identified for the general inference engine and other methods sharing the same declarative knowledge, like planning, inductive learning, and analogical reasoning. The employment of Bumble's explanation constructor as a learning method improves the quality and speed of inference for the computations it refines. The bad news is that the trade-off between the rise in classification costs is not offset by the drop in specific inference costs. Classification times rise as Bumble adds additional structure to the hierarchy.