A neural network architecture for the general problem solver

One of the difficulties of means-ends analysis, a general model of human problem solving, is having to symbolically express the evaluation function for the domain problem solving heuristics. In the present work, the authors propose a neural network architecture called NGPS (Neural General Problem Solver) to avoid this difficulty. Instead of explicitly and symbolically expressing the evaluation function, NGPS can be trained to acquire implicitly the problem solving heuristics. NGPS uses a two-level problem solving architecture: a meta-level controller and an object-level performer. It is shown how tasks of propositional logic theorem proving can be successfully performed by NGPS. In addition, NGPS apparently has the ability to perform structure sensitive operations, which J.A. Fodor and Z.W. Pylyshyn (1988) claimed connectionist models could not do