This tutorial focuses on the new paradigm of non-stochastic control, which proposes regret minimization as a natural desideratum for control of dynamical systems with unknown and possibly adversarial disturbations.
The first half of our tutorial focuses on the control problem. We begin with a review of classical paradigms of robust and stochastic control, as well as their solutions for linear dynamical systems. We then introduce low regret as a desideratum that lies between, and asymptotically subsumes, these two regimes. From there, we introduce a family of adaptive control policies that leverage two ingredients: convex control parametrization, which date back to Youla '76, and the rich algorithmic toolkit of online convex optimization. We derive regret guarantees in a range of regimes - including partially observed, unknown dynamics, optimal rates for curved losses - as well as open problems.
In the second half, we focus on the problem of prediction in linear dynamical systems subject to adversarial disturbances - a natural model for the unpredictable dynamics of modern sequence modeling applications such as large language models. We propose a new family of methods based on the spectral filtering , and describe how to build on this technique for advance neural architectures and general nonlinear control.
The literature in the "References" section covers the most recent research results in this young field, that are the focus of the tutorial.
For more classical texts see "extra reading" at the bottom of the References page, as well as the lecture notes.