Mean Field Langevin Actor-Critic: Faster Convergence and Global Optimality beyond Lazy Learning
Lists
Only you see this
Abstract
We study how deep reinforcement learning algorithms learn meaningful features when optimized for finding the optimal policy. In particular, we focus on a version of the neural actor-critic algorithm where both the actor and critic are represented by over-parameterized neural networks in the mean-field regime, and are updated via temporal-difference (TD) and policy gradient respectively. Specifically, for the critic neural network to perform policy evaluation, we propose $\textit{mean-field Langevin TD learning}$ method (MFLTD), an extension of the mean-field Langevin dynamics with proximal TD updates, and compare its effectiveness against existing methods through numerical experiments. In addition, for the actor neural network to perform policy updates, we propose $\textit{mean-field Langevin policy gradient}$ (MFLPG), which implements policy gradient in the policy space through a version of Wasserstein gradient flow in the space of network parameters. We prove that MFLTD finds the correct value function, and the sequence of actors created by MFLPG created by the algorithm converges linearly to the globally optimal policy of the Kullback Leibler divergence regularized objective. To our best knowledge, we provide the first linear convergence guarantee for neural actor-critic algorithms with $\textit{global optimality}$ and $\textit{feature learning}$.
BibTeX
@inproceedings{yamamoto2024mean,
author = {Kakei Yamamoto and Kazusato Oko and Zhuoran Yang and Taiji Suzuki},
title = {{Mean Field Langevin Actor-Critic: Faster Convergence and Global Optimality beyond Lazy Learning}},
booktitle = {International Conference on Machine Learning},
year = {2024},
url = {https://openreview.net/forum?id=FOJE1kRcHG},
}