Representing and Reasoning With Probabilistic Knowledge: A Bayesian Approach

PAGODA (Probabilistic Autonomous GOal­ Directed Agent) is a model for autonomous learning in probabilistic domains [desJ ardins, 1992) that incorporates innovative techniques for using the agent's existing knowledge to guide and constrain the learning process and for representing, reasoning with, and learn­ ing probabilistic knowledge. This paper de­ scribes the probabilistic representation and inf�rence mechanism used in PAGODA. PAGODA forms theories about the effects of its actions and the world state on the envi­ ronment over time. These theories are rep­ resented as conditional probability distribu­ tions. A restriction is imposed on the struc­ ture of the theories that allows the inference mechanism to find a unique predicted dis­ tribution for any action and world state de­ scription. These restricted theories are called uniquely predictive theories. The inference mechanism, Probability Combination using Independence (PCI), uses minimal indepen­ dence assumptions to combine the probabili­ ties in a theory to make probabilistic predic­ tions.