Agenda
How to Experiment with RL in the Heat of Gamedev
Alexander Zap (Data Scientist, Wargaming Group Limited)
Pass Type:
Academic Pass, Festival Pass, Game Changer Pass, Indie Pass, Start-Up Pass
Track:
Programming
Discipline:
Game AI Summit
Audience Level:
Intermediate
Format:
Lecture
Vault Recording: TBD
Audience Level: Intermediate
We are Wargaming, probably most known for the classic World of Tanks, with our new game WoT: HEAT just having released.
We provide insights into how we apply Reinforcement Learning (RL; learning through interaction) to train models that directly are facing players (e.g., bot steering). We give an exact overview on what is required to accelerate training time, how to have processes your researchers and applied data scientists will not hate, and how to make sure your models end up on production without killing the servers. Especially we cover the extra challenge of doing all of this when the game is in an active development process and there a multiple new builds every day.
We provide insights into how we apply Reinforcement Learning (RL; learning through interaction) to train models that directly are facing players (e.g., bot steering). We give an exact overview on what is required to accelerate training time, how to have processes your researchers and applied data scientists will not hate, and how to make sure your models end up on production without killing the servers. Especially we cover the extra challenge of doing all of this when the game is in an active development process and there a multiple new builds every day.
Takeaway
Listeners will get insights on how to set up an RL experimentation process that makes rapid iteration of models possible. We provide a detailed breakdown of our architecture, capabilities and tooling that make training launches seamless, that decrease training cycle duration, and that work for constantly changing game builds. We share our best practices and standardized processes for structured experimentation.
Intended Audience
This talk is for MLOps Engineers that want to provide games as bot training environments, for Data Scientists that are looking for efficient experimentation processes, and any teams that want to scale their model iteration cycles and are looking into interactive learning with RL. We provide insights into our applied research process that ends up on prod.