Deep Metric Tensor Regularized Policy Gradient

Gang Chen, Victoria Huang

Keywords: Deep Reinforcement Learning, Policy Gradient, Metric Tensor Field, Close-to-zero Divergence

Abstract: In this paper, we propose a novel policy gradient algorithm for deep reinforcement learning. Unlike previous approaches, we focus on leveraging the Hessian trace information in the policy parametric space to enhance the performance of trained policy networks. Specifically, we introduce a metric tensor field that transforms the policy parametric space into a general Riemannian manifold. We further develop mathematical tools, deep learning algorithms, and metric tensor deep neural networks (DNNs) to learn a desirable metric tensor field, with the aim to achieve close-to-zero divergence on the policy gradient vector field of the Riemannian manifold. As an important regularization mechanism, zero divergence nullifies the principal differential components of the loss function used for training policy networks. It is expected to improve the effectiveness and sample efficiency of the policy network training process. Experimental results on multiple benchmark reinforcement learning problems demonstrate the advantages of our metric tensor regularized algorithms over the non-regularized counterparts. Moreover, our empirical analysis reveals that the trained metric tensor DNN can effectively reduce the absolute divergence towards zero on the Riemannian manifold.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 462