Learning Lifted Operator Models with Logical Neural Networks

Abstract

We tackle the problem of relational model based reinforcement learning. Specifically, we are trying to learn lifted logical operator models from interacting with an environment whose states and actions are in a logic form. For this problem, we leverage the capability of the Logical Neural Network (LNN) which is designed for learning with logic statements. We show the feasibility of the LNN in this problem setting and discuss how this approach might be extended to handle contemporary RL environments which do not have logical states.