An Overview of Some Recent Developments in Bayesian Problem-Solving Techniques

The last five years have seen a surge in interest in the use of techniques from Bayesian decision theory to address problems in AI. Decision theory provides a normative framework for representing and reasoning about decision problems under uncertainty. Within the context of this framework, researchers in uncertainty in the AI community have been developing computational techniques for building rational agents and representations suited to engineering their knowledge bases. This special issue reviews recent research in Bayesian problem-solving techniques. The articles cover the topics of inference in Bayesian networks, decision-theoretic planning, and qualitative decision theory. Here, I provide a brief introduction to Bayesian networks and then cover applications of Bayesian problem-solving techniques, knowledge-based model construction and structured representations, and the learning of graphical probability models. The past five years or so have seen increased interest and tremendous progress in the development of Bayesian techniques for building problem solving systems. We have come a long way since the Uncertainty in AI Workshop was founded in 1985, an event precipitated in large part by the fact that the mainstream AI community at that time considered probabilistic approaches impractical for building intelligent systems. Since then the workshop has become the Conference on Uncertainty in AI, attracting high-quality contributions from researchers in a broad array of disciplines, including AI, statistics, operations research, and decision science. In the last several years, concepts from Bayesian Decision Theory, along with representational and computational techniques developed within the Uncertainty in AI community have found their way into mainstream AI and are appearing more and more routinely in papers not focused primarily on uncertainty. Areas include vision, natural language processing, robot navigation, planning, and machine learning. One sign of the importance now regarded Bayesian techniques in building and analyzing software systems is the fact that the Journal of the ACM recently introduced a track on Decisions, Uncertainty, and Computation. Bayesian Decision Theory, like much of AI, is concerned with the characterization of rational behavior. According to Bayesian Decision Theory [Savage, 1954], a choice situation is characterized by a set of possible acts A, a probability distribution P over the set of possible states of the world S, the outcome of each act in each possible state a(s), and a utility function u over the outcome space. The optimal act is the one that maximizes expected utility: Acts here can be physical actions, speech acts, deliberative acts, or complex plans composed of various kinds of actions. Decision Theory is interesting for AI because it provides a normative theory for designing agents capable of reasoning and acting under conditions of uncertainty. This normativity is expressed in the form of a representation theorem stating that if an agent’s preferences obey a set of intuitively appealing constraints, then there exists a probability function P and a utility function U, such that the most preferred action is the one that maximizes expected utility. But what Decision Theory does not provide are the computational mechanisms for building rational agents and the representations suited to engineering their knowledge bases. These issues have been the primary focus and contribution of the work conducted by uncertainty researchers in AI. In this introduction I will discuss issues in the elicitation, representation and manipulation of probability models. I provide a brief discussion of Bayesian networks and then cover applications of Bayesian techniques, knowledge-based model construction and structured representations, and learning of graphical probability models. More recently, researchers have begun to explore these same issues with regard to utility models. The article by Jon Doyle and Richmond Thomason and the one by Jim Blythe that follow provide some discussion of this work. In the interests of brevity, in this introduction I have omitted discussion of exciting and valuable research contributions in many other subareas. I point the interested reader to the proceedings of the Conference on Uncertainty in AI [UAI-98], the Workshop on AI and Statistics [AIStat-99], and to the web page of the Association for Uncertainty in AI (www.auai.org) for further reading.