Title: Reinforcement Learning Applied to Natural Deduction
Speaker: Mark Chevallier
Abstract: Reinforcement learning is a set of methods for an agent to learn how to act in a given environment using “rewards” to encourage correct action. Natural deduction is a method to prove statements based on “natural” reasoning with limited axioms and a small set of logical rules that can be applied to a set of statements to deduce further statements.

I discuss how to apply the tools of reinforcement learning to natural deduction applied to propositional logic. I present a simple python based example and discuss its implementation. Early problems included how to limit the assumptions the reinforcement learning engine could make, related to what it could add to an OR statement, and finding a reasonable sense of “state” for the agent. I demonstrate the theoretical effectiveness of the approach by showing limited success in proving simple deductions.