From: Parag (parag_at_cs.washington.edu)
Date: Fri Apr 11 2003 - 11:02:32 PDT
Planning as Heuristic Search: New Results
Blai Bonet and Hector Geffner
The paper presents a new approach for heuristic search planning
which significantly improvements over the HSP planner by introducing
the regression search in place of the forward search for the goal state.
It also puts forward the idea that well known planners like
Graphplan could be viewed as Heuristic search planners.
The paper clearly identifies the bottolneck in the heuristic
search planner that is the computation of the heuristic function
for every state which takes up most of the time. The authors
introduce the idea of regression(backward) search
which overcomes this problem. The results have been
given comparing the performance of two approaches which
clearly show significant improvement in the later approach.
Another important contribution of the paper is the idea
that well known planners like Graphplan could be viewed
as heuristic search planners. It could give a lot of insight
into the way graphplan(and possible other planners) work and
the way they could be improved.
Though above idea seems quite interesting, the prorblem
is that authors give a very hand waivy argument to the
proof which does not take us too far. It would have
been much better had they tried to do it formally. Also,
no experiments have been reported which compare the performance
of HSPr with graphplan, which would probably have thrown
more light on the above idea.
There could be several directions for future work for
example finding an admissible heuristic function which
is informative as well. But the key to imporvement might
lie in the idea of how one could view general planners
as heuristic planners giving a lot of insight
into how one could improve the HSPr and the as well
as planners in general.
This archive was generated by hypermail 2.1.6 : Fri Apr 11 2003 - 11:02:32 PDT