Model Acquisition for Markov Decision Problems
Abstract
In this research, we developed new models and techniques for representing stochastic processes that enabled us to compactly represent problems that couldn't be represented at all with previous techniques. We also developed new algorithms for efficiently solving such problems by directly exploiting the structure in the representations. Our models achieved efficiency in representation by factoring the state and action spaces of a dynamical system using a set of features (variously called state variables or fluents). We were able to explain the sources of combinatorial leverage in our and other structured models such as those of Boutilier et al. We found that the structure was due to certain symmetries in the dynamics that, in certain cases, could be exploited to significantly reduce computation time. Using these insights, we developed new algorithms that realized these reductions in computation time.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 31, 1998
- Accession Number
- ADA373795
Entities
People
- Thomas L. Dean
Organizations
- Brown University