Fri, Sep 24, 2021 @ 11:00 AM - 01:00 PM
James A. Preiss
Gaurav S. Sukhatme
Nora Ayanian, Stefanos Nikolaidis, Heather Culbertson, Ashutosh Nayyar, Gaurav S. Sukhatme
Reasoning about sets of dynamical systems:
Identification, adaptive control, and suboptimal coverings
Robots often have physical dynamics that are unknown, but belong to a highly structured set. Knowledge of the set structure is useful for designing adaptive mechanisms, especially if synthesizing a good control policy for each system in the set is straightforward. However, when the systems in the set are complex enough to necessitate black-box methods like reinforcement learning, it is less clear how to use knowledge of the set structure. For example, neural network policies that take dynamics parameters as a secondary input fall short of the performance of single-system policies.
We propose both practical and theoretical inquiries into this difficulty. From the practical side, we propose an alternative to black-box learning methods that combines differentiable simulation, recurrent neural network models, and established methods from nonlinear control. From the theoretical side, we introduce the framework of suboptimal covering numbers to quantify "how much" a good policy must change with respect to dynamics parameters. We bound the suboptimal covering number for a simple class of linear-quadratic systems.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon