Toward A Universal Inference Engine

Abstract

In the early days of AI some researchers proposed that intelligent problem solving could be reduced to the application of general purpose theorem provers to an axiomatization of commonsense knowledge. Although automated first-order theorem proving was unwieldy, general reasoning engines for propositional logic turned out to be surprisingly efficient for a wide variety of applications. Still many problems of interest to AI involve probabilities or quantification, and would seem to be beyond propositional methods. However, recent research has shown that the basic backtrack search algorithm for satisfiability generalizes to a strikingly efficient approach for broader classes of inference. We may be on the threshold of achieving the old dream of a universal inference engine.