Knowledge representation for stochastic decision processes

  • Authors:

  • Craig Boutilier

  • Affiliations:

  • Dept. of Computer Science, University of British Columbia, Vancouver, BC, Canada

  • Venue:
  • Artificial intelligence today
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

Reasoning about stochastic dynamical systems and planning under uncertainty has come to play a fundamental role in AI research and applications. The representation of such systems, in particular, of actions with stochastic effects, has accordingly been given increasing attention in recent years. In this article, we survey a number of techniques for representing stochastic processes and actions with stochastic effects using dynamic Bayesian networks and influence diagrams, and briefly describe how these support effective inference for tasks such as monitoring, forecasting, explanation and decision making. We also compare these techniques to several action representations adopted in the classical reasoning about action and planning communities, describing how traditional problems such as the frame and ramification problems are dealt with in stochastic settings, and how these solutions compare to recent approaches to this problem in the classical (deterministic) literature. We argue that while stochastic dynamics introduce certain complications when it comes to such issues, for the most part, intuitions underlying classical models can be extended to the stochastic setting.