Learning Latent Symmetry for Sample-Efficient Dynamical Modeling

Data-driven dynamical modeling, fundamental to control and Reinforcement Learning in systems with unknown dynamics, faces challenges from data scarcity, such as low-resolution measurements. For example, in power systems, smart meter data may not capture fast load dynamics. This prevents us from training an accurate and robust Deep Learning model.

In this talk, I will address the problem by exploiting symmetry in dynamical systems. Symmetry, defined as a group of transformations that leave a system’s behavior or properties equivariant, is prevalent across various domains. Symmetry transformations enable a system state to represent a large set of equivalent states, reducing the amount of data needed for training. Based on this intuition, I demonstrate how to systematically design DL models that preserve symmetries. In addition, rigorous theoretical support is provided. This framework not only enhances our understanding of the intersection between dynamic modeling and geometric DL but also establishes a solid foundation for applying Model-Based Reinforcement Learning (MBRL) in power systems.

Speaker: Haoran Li, Massachusetts Institute of Technology