From: MAUSAM (mausam_at_cs.washington.edu)
Date: Tue May 13 2003 - 10:51:47 PDT
Agent-Centered Search
Sven Koenig
This paper illustrates a review of planning problems where real
time algorithms (interleaving planning and execution) can be used.
The paper calls such algorithms as agent centered search because the
execution and planning are done by an intelligent agent. The most distinct
advantage of such an agent centered search is that one could aim at
achieving minimization of sum of planning and execution cost, thereby
minimizing the total time taken by the agent to reach the goal. However,
this approach might lead to troubles especially if there are some
irreversible actions as the agent will execute some action without being
too confident about it.
The author presents his LRTA* algorithm, which performs a local search
updating the value function for the local space, then computes the best
action from the current state, executes the action and repeats the
process. The author mentions that most of the other algorithms that do
real time planning fall into some version of LRTA*, with some changes. So,
LRTA* in some sense unifies the domain of real time planning. Then the
author goes on to give various examples of different domains where such
methods work.
The major drawback of the paper is that the paper talks about a lot of
things but never gives any experimental justifications for any of it. The
author discusses various examples but does not compare his algorithm with
any other existing one. For cases where interleaving would have have
deteriorated the quality of results, the extent of that could have been
mentioned as experiments. Similarly, the speed up using minimization of
total time could be compared to classical planning approaches.
For future work, I couldn't exactly follow how does the agent decides that
in the step, it has done sufficient planning to do an action. Likewise
how does the agent determine what the local search space would be
considered as. I don't know if these questions have already been studied
but if they are not they should be the first things to be studied to make
LRTA* a practical solution.
This archive was generated by hypermail 2.1.6 : Tue May 13 2003 - 10:51:48 PDT