Conceptual Design of Goal Understanding Systems: Investigation of Temporal Reasoning Under Uncertainty

Greg Cooper, Eric Horvitz, Renwick Curry

Abstract:

We discuss techniques for inferring the intention or leading goal of a pilot through interpreting available information about flight status, airplane configuration, and pilot activity. The ability to identify a pilot's focus of attention at any moment during a flight can provide an essential link to the provision of effective decision support. In particular, understanding the current goals of a pilot decision maker can be applied to select the presentation of alternative systems and displays. We present alternative artificial-intelligence approaches to data interpretation, and review logical and probabilistic approaches to reasoning. After introducing the problem of representing and reasoning about evidential relationships in an atemporal setting, we introduce temporal reasoning issues. We delve into our problem-solving approach, dwelling upon the representation of probabilistic relationships among goals, and events or sequences of events. After reviewing the limitations of our approach, we describe a prototype implementation, and detail its behavior. Finally, we discuss future extensions and research directions. (42 pages)

Keywords: Goal understanding, probability, uncertainty management, Bayesian networks, control of display, temporal reasoning.

Technical Memorandum NAS2-12381, NASA-Ames Research Center, Mountain View, CA, February 1988.


Some background

As part of this work, we constructed a prototype system that reasoned under uncertainty over time about a pilot's goals, given phase of flight and additional observations, such as pilot activity and radar findings. Here are some snapshots from the prototype. The first graph shows the likelihood of three competing goals over time given the prior probabilities and information about the phase of flight.

The likelihood of alternate goals can change with information about the pilot's behavior or from such classes of evidence as communication with air traffic control and radar results.

The following screen shot shows how the probabilities of the goals are revised in an identical run except for the presence of two additional pieces of evidence, observed in adjacent periods. The first observation is present and has a duration of 1 between time periods 6 and 7, and the second is present between periods 7 and 8.

Related background:

G.F. Cooper, E.J. Horvitz, D.E. Heckerman, A Method for Temporal Probabilistic Reasoning, Technical Report 88-30. Knowledge Systems Laboratory, Stanford University, July 1988.