Abstract:
In the real world, one usually cannot gather enough evidence to completely determine the activity of a system. Typically, different pieces of evidence tell you about different aspects of the system with different certainties. Correlating these at a given point of time is a much studied problem. The problem becomes even less tractable when the evidence is acquired at different points in time. The system described in this paper uses frame-like objects called "models" that propagate the effects of a piece of evidence through time and uses Gordon and Shortliffe’s theory to combine the effects of the active models. These models do not require either a great deal of storage or that the evidence be processed in temporal order. Further, they seem to be a construct that the experts in our problems easily relate to. Results with test problems are consistent with the estimates of experts and run in not unreasonable time.