Events

PILAF: A stable foundation for empirically effective multi-agent reinforcement learning

Speaker: Prof. Eugene Vinitsky

Location: TBA
Videoconference link: https://nyu.zoom.us/j/92570929898

Date: Wednesday, February 26, 2025

Reinforcement learning, particularly in multi-agent settings, is believed to be finicky and hard to deploy, requiring custom algorithms. We present and discuss a contrary conclusion, that single-agent reinforcement learning algorithms, with appropriate regularization, appear to outperform algorithms specifically formulated for multi-player settings. As measuring performance in multi-agent settings is challenging, we construct new benchmarks with computationally tractable solutions that we use to perform a large-scale algorithmic comparison. Finally, we demonstrate that applications of these algorithms at the scale of trillions of samples allows us to build decentralized control systems with surprising robustness, using them to build a simulated driver that has below human-level crash rates.