From: Sumit Sanghai (sanghai_at_cs.washington.edu)
Date: Fri Apr 11 2003 - 10:26:36 PDT
Planning as Heuristic Search : New Results
---B. Bonet and H. Geffner
Summary : The paper presents a new and improved method to do planning
using heuristic search.
Key Ideas : I think the most important idea used is to view GraphPlan as
heuristic search. Once one understands this, the rest of the paper follows
directly. Some of the ideas inspired from GraphPlan are to do backward
search
(which avoids calculating the heuristic for every state) instead of
forward and to calculate mutex pairs. Apart from this, the Greedy Best
First Search seems to be a neat idea.
Flaws : The paper just gives an intuition of GraphPLan being similar to
heuristic search but doesn't actually prove it which was frustrating. The
experiments although provided some ideawhere HSPr may succeed or fail,
they are not detailed and the numbers are for only those cases where
HSP/HSPr did well.
Future Work : The most immediate future work is to prove that
GrpahPlan and heuristic search are similar. This alone can turn out to be
one of the most significant contributions. The paper talks about various
opportunities where the algorithm can be improved such as finding better
heuristics, reducing memory consumption, faster node generation and
extending this to larger domains.
This archive was generated by hypermail 2.1.6 : Fri Apr 11 2003 - 10:26:37 PDT