CMSC 491M/691M - Spring 2003

Discussion Questions for Class #5, February 10

Reading: Lehman et al. (Soar)

Does chunking seem like a good universal model for learning? Why or why not? If it is, why are people still doing research in machine learning?

Laird et al. [Soar: An Architecture for General Intelligence], states that "The adoption of the problem space as the fundamental organization for all goal-oriented symbolic activity (called the Problem Space Hypothesis [and due to Newell]) is a principal feature of Soar." How does Soar's adoption of this hypothesis differ from other AI methods?
Do you think the Problem Space Hypothesis is a reasonable model of how humans think?
Is "generic search" a reasonable summation of Soar's model of reasoning? This mdoel allows it to be used to easily implement "weak" search methods. Are there tasks for which this model inadequate?
Here are two opposite views of the "generic" nature of Soar as an architecture for intelligence. Be prepared to argue for or against either of these viewpoints, and to offer other points of view:

The "generic" universal subgoaling of Soar makes it an exceptionally powerful architecture for reasoning.
The "generic" universal subgoaling of Soar makes it such a general model that it could cover any sort of behavior -- and therefore, it really says nothing about reasoning. In fact, all of the so-called "power" of Soar is embedded in the domain knowledge (production rules and heuristics).

Soar and ACT-R both have production rules. What are the similarities and differences in the nature or application of these production rules?
Soar and ACT-R both have chunks. What are the similarities and differences in the nature or application of these chunks?
Try to identify domains that would be (1) good domains to apply either Soar or ACT-R; (2) good domains to apply Soar but not ACT-R; (3) good domains to apply ACT-R but not Soar; and (4) good domains for neither ACT-R nor Soar applications.