Browse Teach Curate

First-Order Logic Reward Machines for Reinforcement Learning

First-Order Reward Machines (FORMs) label reward-machine transitions with first-order logic instead of propositional formulae, letting a single compact machine abstract over object properties; the work contributes a method for learning FORMs and a multi-agent formulation for exploiting them, improving learnability and transfer over traditional reward machines.

Data Text

FORM: Learning Expressive and Transferable First-Order Logic Reward Machines

This paper introduces First-Order Reward Machines (FORMs), an extension of reward machines for reinforcement learning that addresses non-Markovian rewards. Traditional reward machines label the edges…