Reflexive Learning in Soar

Chunking in Soar does not occur deliberately but instead occurs automatically, on the resolution of any impasse. This allows the following four questions -- which must be answered by any system which learns -- to be directly answered by the architecture itself, requiring no extra deliberation.

• What should be learned?

  Soar's answer: The knowledge needed to avoid the impasse which caused the generation of a subgoal. A dependency analysis is performed to determine what knowledge at the creation of the impasse led to its resolution. This, and the result knowledge, are what is learned. Soar's commitment to learning at every impasse brings up the issue of the utility of this automatic learning.

• When should learning occur?

  Learning occurs in Soar whenever an impasse is resolved. Since impasses arise whenever there is a lack of knowledge about how to proceed toward a decision, the Soar architecture can be said to learn whenever a lack of knowledge is architecturally detected.

• What resources should be learned from?

  The dependency analysis procedure answers this question as described above: whatever led to the resolution of the impasse.

• When can learned knowledge be applied?

  Because chunking stores new knowledge as productions, in the same form as the other productions of long-term memory, learned knowledge is considered and applied in the same way as other long-term memory productions. Thus, learned information may be applied any time it is appropriate and this appropriateness is made inherent in the representation.

Go to a discussion of this property for multiple architectures.

Current Location: Soar - Properties - Reflexive Learning

Go to NEXT page.