From: Alexander Yates (ayates_at_cs.washington.edu)
Date: Tue May 13 2003 - 10:34:02 PDT
Review of "Agent-Centered Search," by Sven Koenig.
This paper describes a planning paradigm called agent-centered search, in
which the planning phase and execution phase are interleaved, so the agent
can (and some times must) make decisions about what action to perform next
before finding a complete solution.
Instead of trying to minimize planning time alone, agent-centered search
tries to minimize the sum of planning and execution cost by interleaving
execution with planning, and it allows the agent to meet real-time
constraints (soft or hard) on when it must act next. In the worst case, I
don't think the sum of planning and execution costs for, say, LRTA* is any
better than A*, but for many practical domains like robot exploration and
localization agent-centered search performs better. In any situation where
there are real-time constraints, such as chess, the local search property is
vital.
The paper described "agent-centered search" in a vague or roundabout way.
There was no single concrete definition of the term, only some examples and
a description of its properties. It also provided a lot of examples and
sample applications of agent-centered search, and made several claims about
nice properties, but it never gave any kind of evidence to support the
claims, either theoretical or experimental.
One possible direction for research is figuring out a principled way for
having the agent decide at run-time on the best possible way to divide its
time between planning and execution. It seems like there would be cases
where the agent should figure out it needs to stop and think awhile before
acting next (like in a particularly dangerous and new chess position).
Alex
This archive was generated by hypermail 2.1.6 : Tue May 13 2003 - 10:34:09 PDT