Safe Zero-Shot Model-Based Learning and Control: A Wasserstein Distributionally Robust Approach.
Year: 2020
Type: preprint
Abstract: This paper explores distributionally robust zero-shot model-based learning and control using Wasserstein ambiguity sets. Conventional model-based reinforcement learning algorithms struggle to guarantee feasibility throughout the online learning process. We address this open challenge with the following approach. Using a stochastic model-predictive control (MPC) strategy, we augment safety constraints with affine random variables corresponding to the instantaneous empirical distributions of modeling error. We obtain these distributions by evaluating model residuals in real time throughout the online learning process. By optimizing over the worst case modeling error distribution defined within a Wasserstein ambiguity set centered about our empirical distributions, we can approach the nominal constraint boundary in a provably safe way. We validate the performance of our approach using a case study of lithium-ion battery fast charging, a relevant and safety-critical energy systems control application. Our results demonstrate marked improvements in safety compared to a basic learning model-predictive controller, with constraints satisfied at every instance during online learning and control.
Source: arXiv (Cornell University)
Authors: Aaron Kandel, Scott Moura
Institution: University of California, Berkeley
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3014554550