GDC + VRDC 2019 Session Scheduler

View, browse and sort the ever-growing list of sessions by pass type, track, and format. With this Session Scheduler, you can build your schedule in advance and access it during the show via export or with the GDC Mobile App, once live. Sessions do fill up so please arrive early to sessions that you would like to attend. Adding a session to your schedule does not guarantee you a seat.

ML Tutorial Day: Smart Bots for Better Games: Reinforcement Learning in Production

Olivier Delalleau (Data scientist, Ubisoft)

Pass Type: All Access, GDC Conference + Summits, GDC Summits - Get your pass now!

Topic: Programming

Format: Tutorial

Tutorials: ML Tutorial Day

Vault Recording: Video

Audience Level: Intermediate

This talk provides an overview of various reinforcement learning algorithms and how they may help game studios create better games, more efficiently. Besides AI development, the ability to train bots to play games during production opens up promising opportunities for automated testing and design assistance. But applying reinforcement learning to modern games brings up many challenges, illustrated here through several examples, with a focus on recent experiments within Ubisoft games. Whether you want to directly learn from pixels to minimize the integration burden, or entirely rewrite your engine to make it more "reinforcement learning-friendly", this presentation is packed with practical tips to help you reach your goal without (too many) tears.


This talk presents to attendees results and lessons learned from several use cases of reinforcement learning in game development, with in particular applications to automated testing of specific game components, as well as learning various AI behaviors (related to decision making, driving, navigation, animation, etc.).

Intended Audience

This talk is directed towards all game developers curious about the promises of reinforcement learning, and willing to know more about potential benefits and pitfalls. Previous knowledge of reinforcement learning is helpful to follow some technical points and practical recommendations, but is not mandatory for a high-level understanding.