Physics Informed Model Based Reinforcement Learning for Controlling Synchronization of Weakly Coupled Kuramoto System

NeurIPS 2024 Workshop MusIML Submission20 Authors

Published: 30 Nov 2024, Last Modified: 01 Dec 2024MusIML PosterEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Reinforcement learning, physics informed machine learning, control system, Kuramoto network
Abstract: Kuramoto network as a representative of collective dynamics presents a challenging control task of affecting the synchronization of the interacting oscillators. As the dynamics become harder to estimate, making use of a learned model for controlling purposes is difficult. Learning through interactions with the environment enhanced by model-based reinforcement learning (MBRL) algorithms can alleviate the lack of sample efficiency involved with model-free reinforcement learning (MFRL) methods. Given prior knowledge of the underlying dynamics of the system, physics-informed MBRL can achieve even higher efficiency. In this study, we compare the performance of physics-informed MBRL, MBRL, and MFRL in synchronizing the Kuramoto network. We assess the scalability of these three reinforcement learning methods in a naturally chaotic or unsynchronized network.
Submission Number: 20
Loading