A Neural Tangent Kernel Approach for Constrained Policy Gradient Reinforcement Learning

Balazs Varga, Attila Lischka, Balazs Kulcsar, Morteza Haghir Chehreghani

Keywords: Reinforcement learning, Policy gradient methods, Constrained learning, Neural Tangent Kernel

Abstract:

This paper presents a constrained policy gradient method where we introduce constraints for safe learning, augmenting the traditional REINFORCE algorithm by taking the following steps. First, we analyze how the agent's policy changes if a new data batch is applied, leading to a nonlinear differential equation system in continuous time (gradient flow). This description of learning dynamics is connected to the neural tangent kernel (NTK) which enables us to evaluate the policy change at arbitrary states.
Next, we introduce constraints for action probabilities based on the assumption that there are some environment states where we know how the agent should behave, ensuring safety during learning. Then, we augment the training batch with these states and compute fictitious rewards for them, making the policy obey the constraints with the help of the NTK-based formulation. More specifically, exogenous discounted sum of future rewards (returns) are computed at these constrained state-action pairs such that the policy network satisfies the constraints. Computing the constraining returns is based on solving a system of linear equations (equality constraints) or a constrained quadratic program (inequality constraints). To tackle high-dimensional environments, a dynamic constraint selection methodology is proposed.
Simulation results demonstrate that adding constraints (external information) to the learning can improve learning in terms of speed and transparency reasonably if they are selected appropriately.

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: 7017