Abstract
In this tutorial article, we aim to provide the reader with the conceptualtools needed to get started on research on offline reinforcement learningalgorithms: reinforcement learning algorithms that utilize previously collecteddata, without additional online data collection. Offline reinforcement learningalgorithms hold tremendous promise for making it possible to turn largedatasets into powerful decision making engines. Effective offline reinforcementlearning methods would be able to extract policies with the maximum possibleutility out of the available data, thereby allowing automation of a wide rangeof decision-making domains, from healthcare and education to robotics. However,the limitations of current algorithms make this difficult. We will aim toprovide the reader with an understanding of these challenges, particularly inthe context of modern deep reinforcement learning methods, and describe somepotential solutions that have been explored in recent work to mitigate thesechallenges, along with recent applications, and a discussion of perspectives onopen problems in the field.