Distributionally Robust Linear Quadratic Control

Classical LQG control assumes the distribution of the noise is known. We instead allow any distribution within a Wasserstein ball around a Gaussian estimate, including non-Gaussian distributions, and optimize against the worst case. Despite this ambiguity, a policy that is linear in the observations remains optimal. An efficient numerical method uses the Frank-Wolfe algorithm to find the least-favorable distributions and Kalman filtering to compute the controller.

Advances in Neural Information Processing Systems (NeurIPS)