Offline Reinforcement Learning Workshop

Neural Information Processing Systems (NeurIPS)

December 12, 2020

@OfflineRL · #OFFLINERL2020

The remarkable success of deep learning has been driven by the availability of large and diverse datasets, such as ImageNet, that are challenging to collect, but can be reused repeatedly. In contrast, the common paradigm in reinforcement learning (RL) assumes that an agent frequently interacts with the environment and learns using its own collected experience. This mode of operation is prohibitive for many complex real-world problems, where repeatedly collecting diverse data is expensive (e.g., robotics or educational agents) and/or potentially unsafe (e.g., healthcare).

Alternatively, Offline RL, or pure batch RL, focuses on training agents with logged data in an offline fashion with no further environment interaction. Offline RL promises to bring forward a data-driven RL paradigm and carries the potential to scale up end-to-end learning approaches to real-world decision making tasks such as robotics, recommendation systems, dialogue generation, autonomous driving, healthcare systems and safety-critical applications. Recently, successful deep RL algorithms have been adapted to the offline RL setting and demonstrated a potential for success in a number of domains, however, significant algorithmic and practical challenges remain to be addressed. Within the past two years, performance on simple benchmarks has rapidly risen, so the community has also started developing standardized benchmarks (RLUnplugged, D4RL) specifically designed to stress-test offline RL algorithms. However, these are only initial proposals that could benefit from community input on design and evaluation protocol.

Goal of the workshop: Our goal is to bring attention to offline RL, both from within and from outside the RL community (e.g., causal inference, optimization, self-supervised learning), discuss algorithmic challenges that need to be addressed, discuss potential real-world applications, discuss limitations and challenges, and come up with concrete problem statements and evaluation protocols, inspired from real-world applications, for the research community to work on. In particular, we are interested in bringing together researchers and practitioners to discuss questions on theoretical, empirical, and practical aspects of offline RL, including but not limited to,

  • Algorithmic decisions and associated challenges in training RL agents offline

  • Properties of supervision needed to guarantee the success of offline RL methods

  • Relationship with learning under uncertainty: Bayesian inference and causal inference

  • Model selection, off-policy evaluation and theoretical limitations

  • Relationship and integration with the conventional online RL paradigms

  • Evaluation protocols and frameworks and real-world datasets and benchmarks

  • Connections to transfer learning, self-supervised learning and generative modelling


To contact the organizers, please send an email to

Thanks to Jessica Hamrick for allowing us to borrow this template.