Greedy Linear Value Function Approximation for Factored Markov Decision
Processes
Pascal Poupart
Department of Computer Science
University of Toronto
Toronto, ON M5S 3H5
email: ppoupart@cs.toronto.edu
Relu Patrascu
Dept. of Computer Science
University of Waterloo
Waterloo, ON, N2L 3G1
email: rpatrasc@cs.uwaterloo.ca
Dale Schuurmans
Dept. of Computer Science
University of Waterloo
Waterloo, ON, N2L 3G1
email: dale@cs.uwaterloo.ca
Craig Boutilier
Department of Computer Science
University of Toronto
Toronto, ON M5S 3H5
email: cebly@cs.toronto.edu
Carlos Guestrin
Dept. of Computer Science
Stanford University
Stanford, CA 94305-9010
email: guestrin@stanford.edu
Abstract
Significant recent work has focused on using linear representations to
approximate value functions for factored Markov decision processes (MDPs).
Current research has adopted linear programming as an effective
means to calculate approximations for a given set of basis functions,
tackling very large MDPs as a result.
However, a number of issues remain unresolved:
- How accurate are the approximations produced by linear programs?
- How hard is it to produce better approximations?
and
- Where do the basis functions come from?
To address these questions,
we first investigate the complexity of minimizing the Bellman error
of a linear value function approximation---showing that this is an
inherently hard problem.
Nevertheless, we provide a branch and bound method for
calculating Bellman error and performing
approximate policy iteration for general factored MDPs.
These methods are more accurate than linear programming, but more expensive.
We then consider linear programming itself and investigate methods for
automatically constructing sets of basis functions that allow
this approach to produce good approximations.
The techniques we develop are guaranteed to reduce L1 error,
but also empirically reduce Bellman error.
Although L1 is a weak error measure for value function approximation,
our results show that linear programming still yields competitive
approximations,
and our methods tend to build basis functions that reveal the underlying
decompositional structure of the domain.
To appear, AAAI-02
Return to
List of Papers