Parameter-free Mirror Descent

@inproceedings{Jacobsen2022ParameterfreeMD,
  title={Parameter-free Mirror Descent},
  author={Andrew Jacobsen and Ashok Cutkosky},
  booktitle={COLT},
  year={2022}
}
We develop a modified online mirror descent framework that is suitable for building adaptive and parameter-free algorithms in unbounded domains. We leverage this technique to develop the first unconstrained online linear optimization algorithm achieving an optimal dynamic regret bound, and we further demonstrate that natural strategies based on Follow-the-Regularized-Leader are unable to achieve similar results. We also apply our mirror descent framework to build new parameter-free implicit… 
Optimal Dynamic Regret in LQR Control
TLDR
An efficient online algorithm is provided that achieves an optimal dynamic (policy) regret of Õ(n) dynamic regret on a family of “minibatched” quadratic losses, which could be of independent interest.
Optimal Parameter-free Online Learning with Switching Cost
TLDR
A simple yet powerful algorithm for Online Linear Optimization (OLO) with switching cost is proposed, which improves the existing suboptimal regret bound [ZCP22a] to the optimal rate.

References

SHOWING 1-10 OF 31 REFERENCES
A survey of Algorithms and Analysis for Adaptive Online Learning
TLDR
This approach strengthens pre-viously known FTRL analysis techniques to produce bounds as tight as those achieved by potential functions or primal-dual analysis, and proves regret bounds in the most general form.
A Modern Introduction to Online Learning
TLDR
This monograph introduces the basic concepts of Online Learning through a modern view of Online Convex Optimization, and presents first-order and second-order algorithms for online learning with convex losses, in Euclidean and non-Euclidean settings.
Online mirror descent and dual averaging: keeping pace in the dynamic case
TLDR
This work modifies the classical OMD convergence analysis in a careful and modular way that allows for straightforward and flexible proofs and shows that OMD with stabilization and DA enjoy the same performance guarantees in many applications---even under dynamic learning rates.
Implicit Parameter-free Online Learning with Truncated Linear Models
TLDR
New parameter-free algorithms that can take advantage of truncated linear models through a new update that has an “implicit” flavor are proposed that are efficient, efficient, requires only one gradient at each step, never overshoots the minimum of the truncated model, and retains the favorable parameter- free properties.
Scale-free online learning
Lipschitz and Comparator-Norm Adaptivity in Online Learning
TLDR
Two prior reductions to the unbounded setting are generalized; one to not need hints, and a second to deal with the range ratio problem (which already arises in prior work).
Parameter-free Stochastic Optimization of Variationally Coherent Functions
TLDR
This work designs and analyzes an algorithm for first-order stochastic optimization of a large class of functions on R which is an instance of the Follow The Regularized Leader algorithm with the added twist of using rescaled gradients and time-varying linearithmic regularizers.
Black-Box Reductions for Parameter-free Online Learning in Banach Spaces
We introduce several new black-box reductions that significantly improve the design of adaptive and parameter-free online learning algorithms by simplifying analysis, improving regret guarantees, and
Adaptive Subgradient Methods for Online Learning and Stochastic Optimization
TLDR
This work describes and analyze an apparatus for adaptively modifying the proximal function, which significantly simplifies setting a learning rate and results in regret guarantees that are provably as good as the best proximal functions that can be chosen in hindsight.
Combining Online Learning Guarantees
TLDR
It is shown how to take any two parameter-free online learning algorithms with different regret guarantees and obtain a single algorithm whose regret is the minimum of the two base algorithms, which yields the first optimistic regret guarantees in the unconstrained setting and generically increases adaptivity.
...
...