Tim@CIS.UPENN.EDU (Tim Finin) (08/13/86)
Active Reduction of Uncertainty in Multi-sensor Systems Ph.D. Thesis Proposal Greg Hager (greg@upenn-grasp) General Robots and Active Sensory Perception Laboratory University of Pennsylvania Department of Computer and Information Sciences Philadelphia, PA 19104 10:00 AM, August 15, 1986 Room 554 Moore If robots are to perform tasks in unconstrained environments, they will have to rely on sensor information to make decisions. In general, sensor information has some uncertainty associated with it. The uncertainty can be conceptually divided into two types: statistical uncertainty due to signal noise, and incompleteness of information due to limitations of sensor scope. Inevitably, the information needed for proper action will be uncertain. In these cases, the robot will need to take action explicitly devoted to reducing uncertainty. The problem of reducing uncertainty can be studied within the theoretical framework of team decision theory. Team decision theory considers a number of decision makers observing the world via information structures, and taking action dictated by decision rules. Decision rules are evaluated relative to team and individual utility considerations. In this vocabulary, sensors are considered as controllable information structures whose behavior is determined by individual and group utilities. For the problem of reducing uncertainty, these utilities are based on the information expected as the result of taking action. In general, a robot does not only consider direct sensor observations, but also evaluates and combines that data over time relative to some model of the observed environment. In this proposal, information aggregation is modeled via belief systems as studied in philosophy. Reducing uncertainty corresponds to driving the belief system into one of a set of information states. Within this context, the issues that will be addressed are the specification of utilities in terms of belief states, the organization of a sensor system, and the evaluation of decision rules. These questions will first be studied through theory and simulation, and finally applied to an existing multi-sensor system. Advisor: Dr. Max Mintz Committee: Dr. Ruzena Bajcsy (Chairperson) Dr. Zolton Domotor (Philosophy Dept.) Dr. Richard Paul Dr. Stanley Rosenschein (SRI International and CSLI)