Open Source Summit + ELC North America 2020

Deep reinforcement learning (RL) is an exciting paradigm for solving real-world problems. Most research aims to improve how well RL performs, and much less effort focuses on understanding the products of deep RL -- often enormous neural network policies -- which is critical for establishing enough trust to deploy them in the real world. The open-source Atari Zoo project aims to catalyze research into understanding deep RL by reducing the barriers to entry for researchers and citizen scientists alike. A team at Uber AI trained popular RL algorithms at scale on simulated Atari 2600 games (the prototypical deep RL benchmark), and released models and code, making it easy to load, visualize, and analyze such pre-trained models. Joel Lehman describes the Atari Zoo project and highlights surprising results from initial explorations and the exciting potential for what is still left to be discovered.