Agenda

In this talk, we will explore the integration of a designer-first reinforcement learning approach into EA SPORTS FC 26. The approach focuses on creating human-like AI behaviors for the game's goalkeeper. Unlike traditional reinforcement learning methods that demand extensive resources and time, our new approach enables game developers to train AI agents overnight. This talk will highlight how we address common issues when applying reinforcement learning to game production, such as fixing undesired behaviors, and how to build a robust testing system featuring over 300 "unit-test" scenarios, ensuring continuous validation. Our new goalkeeper AI not only outperforms the legacy game's built-in AI with a 10% improvement in ball saving rate, but trains 50% faster than standard reinforcement learning methods, while being less robotic than other DRL or traditional AI approaches. During the talk we will show several videos comparing the performance of the goalkeeper trained with reinforcement learning and the old game's CPU AI. The videos will also showcase the difference in quality between the two approaches, demonstrating that our approach produces more human-like behavior. By sharing these insights, we aim to demonstrate how reinforcement learning can be successfully applied in video game production for player-facing gameplay to create realistic and engaging AI, targeting developers' needs.

Takeaway

Attendees will learn practical techniques for training human-like game agents with reinforcement learning for player-facing features. Moreover, attendees will discover how to build an efficient production pipeline that enables overnight training, semi-automated debugging, and robust regression testing to support rapid iteration. Finally, attendees will learn best practices for integrating tester feedback into the training loop to create more controllable and designer-aligned game AI.

Intended Audience

The lecture is aimed at game AI experts and practitioners, machine learning (especially reinforcement learning) practitioners for games, and machine learning researchers.