Abstract
Reinforcement Learning (RL) offers a versatile framework for achievinglong-term goals. Its generality allows us to formalize a wide range of problemsthat real-world intelligent systems encounter, such as dealing with delayedrewards, handling partial observability, addressing the exploration andexploitation dilemma, utilizing offline data to improve online performance, andensuring safety constraints are met. Despite considerable progress made by theRL research community in addressing these issues, existing open-source RLlibraries tend to focus on a narrow portion of the RL solution pipeline,leaving other aspects largely unattended. This paper introduces Pearl, aProduction-ready RL agent software package explicitly designed to embrace thesechallenges in a modular fashion. In addition to presenting preliminarybenchmark results, this paper highlights Pearl's industry adoptions todemonstrate its readiness for production usage. Pearl is open sourced on Githubat github.com/facebookresearch/pearl and its official website is located atpearlagent.github.io.