Timezone: »
Temporal-difference and Q-learning play a key role in deep reinforcement learning, where they are empowered by expressive nonlinear function approximators such as neural networks. At the core of their empirical successes is the learned feature representation, which embeds rich observations, e.g., images and texts, into the latent space that encodes semantic structures. Meanwhile, the evolution of such a feature representation is crucial to the convergence of temporal-difference and Q-learning.
In particular, temporal-difference learning converges when the function approximator is linear in a feature representation, which is fixed throughout learning, and possibly diverges otherwise. We aim to answer the following questions: When the function approximator is a neural network, how does the associated feature representation evolve? If it converges, does it converge to the optimal one?
We prove that utilizing an overparameterized two-layer neural network, temporal-difference and Q-learning globally minimize the mean-squared projected Bellman error at a sublinear rate. Moreover, the associated feature representation converges to the optimal one, generalizing the previous analysis of Cai et al. (2019) in the neural tangent kernel regime, where the associated feature representation stabilizes at the initial one. The key to our analysis is a mean-field perspective, which connects the evolution of a finite-dimensional parameter to its limiting counterpart over an infinite-dimensional Wasserstein space. Our analysis generalizes to soft Q-learning, which is further connected to policy gradient.
Author Information
Yufeng Zhang (Northwestern University)
Qi Cai (Northwestern University)
Zhuoran Yang (Princeton)
Yongxin Chen (Georgia Institute of Technology)
Zhaoran Wang (Northwestern University)
Related Events (a corresponding poster, oral, or spotlight)
-
2020 Poster: Can Temporal-Difference and Q-Learning Learn Representation? A Mean-Field Theory »
Wed Dec 9th 05:00 -- 07:00 AM Room Poster Session 2
More from the Same Authors
-
2020 Poster: Pontryagin Differentiable Programming: An End-to-End Learning and Control Framework »
Wanxin Jin · Zhaoran Wang · Zhuoran Yang · Shaoshuai Mou -
2020 Poster: Provably Efficient Neural GTD for Off-Policy Learning »
Hoi-To Wai · Zhuoran Yang · Zhaoran Wang · Mingyi Hong -
2020 Poster: End-to-End Learning and Intervention in Games »
Jiayang Li · Jing Yu · Yu Nie · Zhaoran Wang -
2020 Poster: Provably Efficient Neural Estimation of Structural Equation Models: An Adversarial Approach »
Luofeng Liao · You-Lin Chen · Zhuoran Yang · Bo Dai · Mladen Kolar · Zhaoran Wang -
2020 Poster: Dynamic Regret of Policy Optimization in Non-Stationary Environments »
Yingjie Fei · Zhuoran Yang · Zhaoran Wang · Qiaomin Xie -
2020 Poster: On Function Approximation in Reinforcement Learning: Optimism in the Face of Large State Spaces »
Zhuoran Yang · Chi Jin · Zhaoran Wang · Mengdi Wang · Michael Jordan -
2020 Poster: Upper Confidence Primal-Dual Reinforcement Learning for CMDP with Adversarial Loss »
Shuang Qiu · Xiaohan Wei · Zhuoran Yang · Jieping Ye · Zhaoran Wang -
2020 Poster: Risk-Sensitive Reinforcement Learning: Near-Optimal Risk-Sample Tradeoff in Regret »
Yingjie Fei · Zhuoran Yang · Yudong Chen · Zhaoran Wang · Qiaomin Xie -
2020 Spotlight: Risk-Sensitive Reinforcement Learning: Near-Optimal Risk-Sample Tradeoff in Regret »
Yingjie Fei · Zhuoran Yang · Yudong Chen · Zhaoran Wang · Qiaomin Xie -
2019 Poster: Statistical-Computational Tradeoff in Single Index Models »
Lingxiao Wang · Zhuoran Yang · Zhaoran Wang -
2019 Poster: Provably Global Convergence of Actor-Critic: A Case for Linear Quadratic Regulator with Ergodic Cost »
Zhuoran Yang · Yongxin Chen · Mingyi Hong · Zhaoran Wang -
2019 Poster: Neural Proximal/Trust Region Policy Optimization Attains Globally Optimal Policy »
Boyi Liu · Qi Cai · Zhuoran Yang · Zhaoran Wang -
2019 Poster: Neural Temporal-Difference Learning Converges to Global Optima »
Qi Cai · Zhuoran Yang · Jason Lee · Zhaoran Wang -
2019 Poster: Variance Reduced Policy Evaluation with Smooth Function Approximation »
Hoi-To Wai · Mingyi Hong · Zhuoran Yang · Zhaoran Wang · Kexin Tang -
2019 Poster: Convergent Policy Optimization for Safe Reinforcement Learning »
Ming Yu · Zhuoran Yang · Mladen Kolar · Zhaoran Wang -
2018 Poster: Contrastive Learning from Pairwise Measurements »
Yi Chen · Zhuoran Yang · Yuchen Xie · Zhaoran Wang -
2018 Poster: Provable Gaussian Embedding with One Observation »
Ming Yu · Zhuoran Yang · Tuo Zhao · Mladen Kolar · Zhaoran Wang -
2018 Poster: Multi-Agent Reinforcement Learning via Double Averaging Primal-Dual Optimization »
Hoi-To Wai · Zhuoran Yang · Zhaoran Wang · Mingyi Hong -
2017 Poster: Estimating High-dimensional Non-Gaussian Multiple Index Models via Stein’s Lemma »
Zhuoran Yang · Krishnakumar Balasubramanian · Zhaoran Wang · Han Liu