A New Approach to Integrating Graphical Models in Decision-Theoretic Planning
Decision-theoretic planning is concerned with the problem of how to choose a sequence of actions to achieve a goal in situations characterized by uncertainty, imperfect information, and tradeoffs among competing objectives — a key challenge for Artificial Intelligence and related fields. There are two widely-used models for decision-theoretic planning: influence diagrams and partially observable Markov decision processes. This talk considers how to integrate these two models in a way that combines their complementary strengths.
How to integrate these models is a long-studied problem. A large body of previous work shows how to adapt techniques for influence diagrams (and related graphical models) to improve the scalability of algorithms for solving Markov decision processes. In this talk, I describe a complementary approach to integrating these models in which techniques for partially observable Markov decision processes are adapted to improve the scalability of algorithms for solving influence diagrams.This new approach leads to a much more scalable variable elimination algorithm for influence diagrams, an improved approach to solving non-Markovian problems, and a more compact and easier-to-understand representation of plans constructed by these algorithms.
Eric Hansen is an Associate Professor in the Department of Computer Science and Engineering at Mississippi State University. A graduate of Harvard University and the University of Massachusetts at Amherst, he has served on the editorial boards of the Artificial Intelligence Journal and the Journal of Artificial Intelligence Research, among other professional activities.