Learning preference rules for a VLSI design problem-solver

A description is given of a VLSI design problem-solving system, and a method of learning preference rules that enable the problem-solver to select the refinement rule that best meets its current design goals. This explanation-based learning technique uses a domain-independent theorem prover and domain-dependent models to explain why a particular design refinement is preferred over alternative refinements. Generalization of this explanation is accomplished by turning constants to variables, dropping common subexpressions from the explanation, and promoting lone subexpressions. The resulting preference rules are general enough to be correctly applied in several different situations. It is asserted that automating the learning of preference rules will make knowledge acquisition easier, reduce the cognitive load on the VLSI designer, and improve the performance of the design problem-solving system