From: Parag (parag_at_cs.washington.edu)
Date: Tue Apr 15 2003 - 11:19:23 PDT
Refinement Planning as a Unifying Framework for Plan Synthesis
- Subbarao Kambhapmpati
The paper presents refinmenet planning as a generalized
paradigm which subsumes most of the approaches to classcial planning as
a special case.
One of the main ideas in the paper was to clearly separate the
refinement of the current plan set and the solution extraction phase
in any planning algorithm. Although simplistic, it does seem to
represent the successful planners like graphplan and satplan as
illustrated by the paper. This seems to give useful insight what
different planners do and where one could improve their performances.
Also, the idea of splitting - as a tradeoff between time spent
for solution extraction and search time seemed quite insightful.
Another related idea was to use disjunctive plans for efficient
planning and the graphplan being a special case of it.
One of the problems that I felt was with the tradeoff section.
The section on asymptotic trade off seemed a little too
abstract. Some more elaborations could have been given
providing more insight into where would one refinement strategy
work over the others. And similarly, the section on empirical
trade-offs did not convey much. Also, the author did not
talk much about scaling up refinements - it was not clear
to me what he meant by saying scaling up refinement
planners - if the claim is that most of the planners
are essentially nothing but refinement planners - are
we talking about scaling up each and every planning
approach?
Some of the directions for the future work could be:
1. As already pointed out, exploring the learning techniques
to improve the performance of the planner.
2. Extending the idea of refinement planning to non-classical
domains.
This archive was generated by hypermail 2.1.6 : Tue Apr 15 2003 - 11:19:23 PDT