segall@CAIP.RUTGERS.EDU (Ed Segall) (12/17/86)
There will be an III talk Thursday morning (12/18) at 10 am. in Hill 250 (the normal Machine Learning meeting room and time). Mike Barley will be speaking. His abstract follows: In his chapter "Why are design derivations hard to replay" in MACHINE LEARNING: A Guide to Current Research, Jack Mostow identified two classes of problems that any intelligent replay mechanism will need to address: (1) missing preconditions; and (2) the reference problem. These problems deal with applying a rule in the replay plan in the current environment. During this past summer I developed a simple replay facility, called Legal Replay, for the Vexed (VLSI EXpert Editor) system which showed that the difficulty of these problems is determined, to a large extent, by the architecture of the problem-solver in which it is embedded. The abstract refinement with constraint propagation architecture of Vexed made certain aspects of these replay problems disappear. In the course of implementing Legal Replay another replay problem became apparent: the correspondence problem. The correspondence problem deals with controlling rule selection in the current environment based upon the replay plan. In this talk I will briefly describe the replay paradigm, the replay problems that Jack identified, the Vexed architecture, and the correspondence problem. I will concentrate on describing the Legal Replay architecture, how it solves some of the replay problems, how it handles the correspondence problem, and my current research into a more intelligent handling of the correspondence problem.