Concept Learning and Heuristic Classification in Weak-Theory Domains
Abstract
This paper describes a successful approach to the task of concept learning for heuristic classification. This task differs from the usual concept learning task in three ways. First, classifications must be explained, not simply reported. Second, a program for this task must accommodate incomplete case descriptions. Third, the program must learn domain-specific knowledge for inferring case features needed for classification. Section 2 summarizes this learning and classification task. The traditional approach to concept learning and classification relies on generalizations. It requires a strong domain theory both to summarize training cases with concept descriptions and to classify new cases using these descriptions. Section 3 argues that this approach is ineffective for many domains. An alternative approach relies on exemplars. concepts are learned by retaining exemplars, and new cases are classified by matching them with exemplars. Our learning and classification program, Protos, uses the exemplar-based approach. Section 4 describes Protos's design and its appropriateness for weak-theory domains. To evaluate the design, Protos was applied to the task in clinical audiology of identifying a patient's hearing disorder from symptoms, test results, and history. An expert clinician instructed Protos with 200 cases - a level of training comparable to that received by student clinicians.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1990
- Accession Number
- ADA248064
Entities
People
- Bruce W. Porter
- Ray Bareiss
- Robert C. Holte
Organizations
- University of Texas at Austin