From: Christophe Bisciglia (chrisrb_at_cs.washington.edu)
Date: Tue May 13 2003 - 11:08:56 PDT
Sven Koenig Agent-Centered Search
This paper formalizes the notion of agent-centered search . a methodology
that interleaves planning and execution with the goal of minimizing their
combined time.
One of the main ideas for the paper was the LRTA* algorithm. The authors
claim this algorithm unifies many common online planning algorithms in use
(in some form or another). Essentially, the algorithm uses the notion of a
local search space to update heuristic values, picks an action, executes
it, and repeats until the goal is reached.
One area of the paper that I found particularly interesting was the
comparison of online planning in nature to adversarial search (game
playing). The combination of LRTA* with mini-max allows the agent to
progress assuming the worst possible observations, and acting accordingly.
As Tal mentioned, the first part of the paper is well written, but as it
progresses, it gets very abstract and hard to follow. Although the ideas
presented are quite plausible, it would have been nice to see more support
that .And X said this works..
As far as research directions go, I found myself wondering. If we can
think of planning under uncertainty as an adversarial search problem,
could we apply similar pruning methods form game playing (alpha-beta), or
some variation thereof. In game playing, good move ordering can double
your depth . could this cut planning time in half . or make it
considerably better in the same amount of time?
This archive was generated by hypermail 2.1.6 : Tue May 13 2003 - 11:08:57 PDT