Large-scale deep learning to augment production RL workloads

At Riot Games, we leverage large-scale deep reinforcement learning to build bots that play our games at various skill levels to provide additional signals to our designers and ensure that we are releasing the best experiences for our players. As part of this experience we've experimented with multiple methods for controlling and tuning game servers to get the most out of our training time while the bots are learning to play the game. With our game Team Fight Tactics, we were able to take a unique approach that leveraged large neural networks to predict outcomes rather than have to rely on the server. In this talk, we will discuss the supervised learning process that we took and how we simplified and scaled with Ray Data, Ray Train, and Ray Tune.