Geometry of Neural Reinforcement Learning in Continuous State and Action Spaces
Published in ICLR [ORAL: top 1.8% of submitted], 2025
We prove that the state space is a low dimensional manifold for reinforcement learning in the infinite width limit of two layer neural networks and utilise this to improve performance in dog and humanoid environments.
Recommended citation: Saket Tiwari, Omer Gottesman, & George Konidaris. (2025). "Geometry of Neural Reinforcement Learning in Continuous State and Action Spaces." ICLR 2025 https://openreview.net/pdf?id=AP0ndQloqR