Скачать или смотреть Evolutionary Game Theory and Robustness in RL – Callum Lawson – PIBBSS Symposium '25

This video was recorded during the 2025 PIBBSS Symposium. Read more about it on our website:
https://pibbss.ai/symposium-25/

About the talk:
Adaptive training environments have been central to superhuman AI performance in games and are increasingly used to train agentic LLMs. They also offer a route to scaling alignment, through Unsupervised Environment Design (UED) algorithms that search for failure cases and redirect training to fix them. Yet such arms races between agents and environments can follow many paths, only some of which yield robustness. In this talk, we’ll explore how evolutionary game theory (EGT) could help anticipate these dynamics. We’ll link policy gradients to selection gradients and adaptive environments to resource competition, and present a first-step evolutionary UED algorithm that shows similar training dynamics to existing approaches. We’ll also highlight key challenges, such as quantifying trade-offs under the transient regimes typical of UED. Addressing these issues could establish EGT as a principled framework for steering adaptive environment design, helping direct training toward more robust and aligned agents.