CMSC 491M/691M - Spring 2003

Discussion Questions for Class #6, February 12

Reading: Veloso et al.

What planning method does PRODIGY use? What are the "head-plan" and the "tail-plan?" Why does PRODIGY bother with the "head-plan" at all?
The focus of PRODIGY is on learning in planning. What are the learning methods in PRODIGY?
What does it mean for a search-control decision to be a "learning opportunity?" What do the PRODIGY designers mean by having a "glass-box" planner?
EBL, STATIC, and the other algorithms in section 3.1 just restate existing knowledge. In that case, how can they be said to be "learning?" When is this type of learning valuable? What are the tradeoffs in this type of learning?
Does the QUALITY learning algorithm seem more like "learning to improve efficiency (section 3.1) or "learning domain knowledge" (section 3.2)? What are the limitations of this approach to learning about plan quality?
If you were going to extend PRODIGY for your master's thesis, what aspect would you focus on?

Is Soar's chunking more like the learning methods in 3.1 (learning to improve efficiency) or more like those in 3.2 (learning domain knowledge)?
One striking difference between PRODIGY and ACT-R/Soar is that the latter are very uniform, general architectures, whereas PRODIGY feels more like a framework for a "bag of tricks" (collection of various learning algorithms). What are the advantages and disadvantages of each of these architectural approaches?
What are the key differences (other than the above) between PRODIGY and ACT-R/Soar? Try to identify domains that would be good for the former but not the latter, and vice versa. What about domains that would not be particularly good for either PRODIGY or ACT-R/Soar?