Discrimination Learning in Production Systems

Besides extending the Bacon research, my early post-dissertation work focused on adaptive production systems, specifically on an approach to finding the conditions on rules that I called discrimination learning. The approach was a variant on an early mechanism that John Anderson and colleagues developed in ACTF, but I extended the idea and applied it to new areas. It also bore a close resemblance to methods developed independently by Pavel Brazdil around the same time.

The basic idea involved storing with each rule the previous situation (the state of working memory and variable bindings) in which applied correctly and comparing this with a new situation in which it applied incorrectly. This comparison produced a set of differences, each of which led to a candidate rule that was more specific that the one that had fired. The method constructed all such variants, then used a strengthening process to collect statistical evidence and resolve conflicts at performance time. In hindsight, this approach was an incremental version of the separate-and-conquer methods that now enjoy some popularity in machine learning.

I incorporated this approach to rule learning into a number of systems, including the Amber model of syntax acquisition, the Sage system for learning search heuristics, and the first computational model of learning on the balance-scale task (done jointly with Stephanie Sage). The same basic method played a role in early versions of ACM, a system that automated cognitive modeling (developed with Stellan Ohlsson and Stephanie Sage). These systems were implemented within the PRISM architecture, which supported discrimination as one of its central learning mechanisms.


Related Publications

Langley, P. (1987). A general theory of discrimination learning. In D. Klahr, P. Langley, & R. Neches (Eds.), Production system models of learning and development. Cambridge, MA: MIT Press.

Langley, P. (1985). Learning to search: From weak methods to domain-specific heuristics. Cognitive Science, 9, 217-260.

Sage, S., & Langley, P. (1983). Modeling development on the balance scale task. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (pp. 94-96). Karlsruhe, West Germany: Morgan Kaufmann.

Langley, P. (1983). Learning search strategies through discrimination. International Journal of Man-Machine Studies, 18, 513-541.

Langley, P. (1983). Learning effective search heuristics. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (pp. 419-421). Karlsruhe, West Germany: Morgan Kaufmann.

Langley, P. (1982). Language acquisition through error recovery. Cognition and Brain Theory, 5, 211-255.

Langley, P. (1982). A model of early syntactic development. Proceedings of the 20th Annual Conference of the Society for Computational Linguistics (pp. 145-151). Toronto, Ontario.

Langley, P. (1980). A production system model of first language acquisition. Proceedings of the Eighth International Conference on Computational Linguistics (pp. 183-189). Tokyo, Japan.

For more information, send electronic mail to langley@isle.org


© 1997 Institute for the Study of Learning and Expertise. All rights reserved.