From: Parag (parag_at_cs.washington.edu)
Date: Tue May 06 2003 - 11:08:51 PDT
"Incremental Contingency Planning," by R.Dearden, N. Meuleau, S.
Ramikrishnan, D. Smith, and R. Washington.
The paper tries to address the problem of generating contingency plans
in a domain where sources of uncertainty involve continuous quantities
like time and resources.
The paper tries to address quite a difficult problem. And there are
some good ideas in the paper.
1. Instead of planning in advance for each possible outcome of an action,
(which is impossible to consider in case of continuous outcomes), they
propose building a seed plan and then adding contingency branches at the points
which are expected to give most benefit in terms of utility.
2. To find the utilities of a particular Branch at a branch point, they make
use of plangraph for back-propagating the utilities from the goal state, making
using of a greedy strategy.
It seems unclear which goals/subset of goals do they consider to be satisfied from
a particular branch point. Also, the aspect of different goals interacting with
each other has not been presented clearly - whether the approach handles
that to some extent or does it completely ignore it. The paper lacks a formal framework
which would have made the approach more convincing.
This seems an initial step in this direction and probably there is a lot more
to be done before the approach can be proved to be effective. First,
a set of experiments need to be performed to determine the efficacy of the
approach. Also, one could look presenting the approach in a formal manner,
which would help in understanding why the approach may/may not work, and close
to the optimal it might be.
This archive was generated by hypermail 2.1.6 : Tue May 06 2003 - 11:08:51 PDT