Strongly Adaptive OCO with Memory

02/02/2021

∙

by   Zhiyu Zhang, et al.

∙

7

∙

share

Recent progress in online control has popularized online learning with memory, a variant of the standard online learning problem with loss functions dependent on the prediction history. In this paper, we propose the first strongly adaptive algorithm for this problem: on any interval ℐ⊂[1:T], the proposed algorithm achieves Õ(√(|ℐ|)) policy regret against the best fixed comparator for that interval. Combined with online control techniques, our algorithm results in a strongly adaptive regret bound for the control of linear time-varying systems.