Nonstochastic Control

Nonstochastic control focuses on designing control algorithms for dynamical systems subject to unpredictable, non-stochastic disturbances, aiming to minimize regret—the difference in performance compared to an optimal strategy known only in hindsight. Current research emphasizes developing algorithms like Follow the Regularized Leader (FTRL) and its optimistic variants, often within a framework of online convex optimization, to handle these adversarial disturbances and achieve sublinear regret bounds. This field is significant because it provides robust control strategies for real-world scenarios with unpredictable noise, finding applications in areas such as robotics and hyperparameter optimization where stochastic assumptions are often unrealistic. The development of adaptive algorithms with regret bounds that scale with the complexity of the environment is a key focus.