From: Sandra B Fan (sbfan_at_cs.washington.edu)
Date: Mon Nov 24 2003 - 09:51:57 PST
Title:
Acting Optimally in Partially Observable Stochastic Domains
Authors:
Anthony R. Cassandra, Leslie Pack Kaelbling, and Michael Littman
One-line summary:
The authors describe a more efficient algorithm for solving POMDPs by
casting the problem as a completely observable continuous space MDP, over
the space of belief states.
Important ideas:
Their Witness Algorithm can be shown to be fairly efficient. It
handles their continuous belief-state space by using value iteration on
MDPs, and then apply this to POMDPs. Also, their use of the policy graph was interesting.
Flaws:
I thought the paper was not bad. No major flaws.
Open questions:
Some possible directions for future work, which the authors also mention,
are how to extend their algorithm to policy iteration, and also, to
possibly use their algorithm on larger problems.
This archive was generated by hypermail 2.1.6 : Mon Nov 24 2003 - 09:51:58 PST