INTEGRATING REASONING ABOUT ACTIONS AND BAYESIAN NETWORKS

According to the paradigm of Cognitive Robotics (Reiter, 2001a), intelligent, autonomous agents interacting with an incompletely known world need to reason logically about the effects of their actions and sensor information they acquire over time. In realistic settings, both the effect of actions and sensor data are subject to errors. A cognitive agent can cope with these uncertainties by maintaining probabilistic beliefs about the state of world. In this paper, we show a formalism to represent probabilistic beliefs about states of the world and how these beliefs change in the course of actions. Additionally, we propose an extension to a logic programming framework, the agent programming language FLUX, to actually infer this probabilistic knowledge for agents. Using associated Bayesian networks allows the agents to maintain a single and compact probabilistic knowledge state throughout the execution of an action sequence.