A neural propositional reasoner that is goal-driven and works without pre-compiled knowledge

Abstract

This work presents the propositional version of a neural engine for finding proofs by refutation using the resolution principle. Such a neural architecture does not require special arrangements or different modules in order to do forward or backward reasoning, driven by the goal posed to it. Also, the neural engine is capable of performing monotonic reasoning with both complete and incomplete knowledge in an integrated fashion. In order to do so, it was necessary to provide the system with the ability to create new sentences (clauses). The neural mechanism presented herein is the first to our knowledge that does not require that the clauses of the knowledge base be either pre-encoded as constraints or learnt via examples, although the addition of these features to the system is not an impossibility.