Discovering Problem Solving Strategies: What Humans Do and Machines Don't (Yet)
Abstract
People can discover new problem solving strategies on their own, without help from a teacher, text or other source. Many machine learning programs exist that discover strategies under similar conditions. Do we now have a sufficient set of computational models for understanding human strategy discoveries? This paper presents a detailed analysis of a human problem solving protocol that uncovers 10 cases of strategies being discovered. It is argued that most cases are adequately modeled by existing machine learning techniques, and several are not, which suggests some interesting research problems for machine learning. The paper has five parts. After a brief discussion of the methods of the analysis and the protocol, the protocol analysis is presented in enough detail to allow evaluation of the accuracy of the empirical claims. A subsequent section classifies the cases of strategy discovery found in the data are classified according to standard machine learning concepts. The last section indicates which types of learning exhibited by the subject have not yet been exhibited by machine learning systems. This leads to the view that strategy acquisition by a component human is like scientific theory formation, with the attendant tasks hypothesis generation. Although current machine learning models of strategy acquisition seem pale by comparison, there seems to be nothing stopping us from building machine learning systems with human-level capabilities for strategy discovery. Keywords: Strategy discovery; Skill acquisition; Machine learning; Cognitive science; Impasses-driven learning; Artificial intelligence; Problem-solving; Human factor engineering. (JG)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 21, 1989
- Accession Number
- ADA219008
Entities
People
- Kurt VanLehn
Organizations
- Carnegie Mellon University