Preliminary study of theorem proving with RGoal hierarchical reinforcement learning architecture

Abstract

We aim at building a model of the information processing mechanism of the prefrontal cortex in the brain. For that purpose, we proposed an architecture of hierarchical reinforcement learning with unlimited recursive subroutine calls, RGoal. In this paper, we show that the slightly extended RGoal can execute a kind of symbolic inference, theorem proving. Moreover, we consider a mechanism of acquiring symbolic knowledge from agent's experience in the environment. These mechanisms are candidates of the models of symbolic inference and knowledge acquisition of human brain.