The 13th Barbados Workshop on Reinforcement Learning will take place April 10-17, 2020 at McGill's Bellairs Institute in Holetown, Barbados. This year's theme is State Construction / Planning, organized by Doina Precup, Rich Sutton, Veronica Chelu, Brian Tanner, Adam White, and Joseph Modayil.
The 2020 Barbados Reinforcement Learning workshop will focus on two open problems of AI: state construction and planning with learned models.
The notion of the state of the environment is central to reinforcement learning. We contrast this environment state with the agent state: a compact summary of the agent’s current and historic inputs (i.e., camera image, physical location, or joint angles and velocities) -- supporting the agent’s decision of what to do now. Ideally, we would like learning algorithms to construct agent state autonomously and without human input. Constructing agent state in partially observable, non-stationary domains remains an open challenge. There are many directions to explore, including: recurrent network architectures, predictive state representations, generate and test, and meta-learning approaches to name a few. One aim of this workshop is to better understand the potential candidates solutions for state construction, flesh out new ideas!
Our second focus is on how to represent and learn models of the world and use them for planning. There are many open questions unique to model-based RL. What should be the form of the model and how it should be updated? How should planning work, and how should we handle inaccurate or partial models of the environment? How should the agent use the model to aid exploration? What can we say theoretically about the fixed points of model-based RL?
State construction and planning have been a central focus of AI and Machine learning research for decades. This workshop will explore the capabilities and limitations of our current algorithms and formalisms.