Specification-Guided Reinforcement Learning.

The problem of reinforcement learning (RL) is to generate an optimal policy w.r.t. a given task in an unknown environment. Traditionally, the task is encoded in the form of a reward function which becomes cumbersome for long-horizon goals. An appealing alternate is to use logical specifications, opening the direction of RL from logical specifications. This paper summarizes the trials and triumphs in developing highly performant algorithms and obtain theoretical guarantees in RL from logical specifications.