Wasserstein Distributionally Robust Control and Optimization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2428 papers571 conceptsfrozen 2026-10-04

In classical control and machine learning, we often assume that the uncertainty in our system follows a known probability distribution—usually a Gaussian. However, when the real world deviates from these assumptions, controllers designed for average-case performance ($H_2$ control) become fragile, while those designed for the absolute worst-case ($H_\infty$ control) become too conservative to be useful.

A growing body of research addresses this gap through Wasserstein Distributionally Robust Optimization (DRO). This framework allows practitioners to design systems that are robust not just to noise, but to uncertainty about the noise distribution itself.

The Ambiguity Set as a Middle Ground

The core intuition behind this research direction is the "ambiguity set." Instead of committing to a single nominal distribution $\hat{P}$ (derived from training data), we define a neighborhood of plausible distributions $\mathcal{P}$ and optimize against the worst-case member of that set.

Taylan Kargin, Babak Hassibi, and their colleagues in Wasserstein Distributionally Robust Regret-Optimal Control in the Infinite-Horizon describe this as an interpolation. By adjusting the radius of this neighborhood, a designer can move smoothly between the optimism of stochastic control and the pessimism of robust control.

The choice of the Wasserstein distance to define this neighborhood is critical. Unlike other metrics like Kullback-Leibler divergence, the Wasserstein metric accounts for the underlying geometry of the space. As Liang Pang and Xin Hong argue in Robust Reinforcement Learning with Wasserstein Constraint, this allows the model to consider "shifted" distributions that don't necessarily share the same support as the training data—a vital feature for robots encountering slightly different terrains or dynamics than those seen in simulation.

Why It Works: Robustness as Regularization

One of the most powerful insights in this field is the mathematical equivalence between robustness and regularization. Esther Derman and Shie Mannor demonstrate in Distributional Robustness and Regularization in Reinforcement Learning that preparing for the worst-case distribution within a Wasserstein ball is often mathematically identical to adding a specific penalty term (like a gradient norm) to the original objective function.

This builds intuition for why these controllers are so stable: they are effectively penalized for being overly sensitive to small changes in the input data. This "certificate" of performance ensures that the agent plans against a range of possible dynamics rather than overfitting to a single estimated model.

Achieving Computational Tractability

Historically, minimax problems (optimizing for the best performance under the worst-case scenario) were computationally daunting. Recent work has focused on reformulating these infinite-dimensional problems into tractable ones.

For example, Bahar Taskesen, Daniel Kuhn, and their team showed in Distributionally Robust Linear Quadratic Control that for the classic Linear-Quadratic-Gaussian (LQG) problem, the worst-case noise distribution is actually Gaussian. This surprising result allows them to maintain a linear control policy and use efficient algorithms like Frank-Wolfe to solve for the robust controller much faster than general-purpose solvers.

Similarly, Kehan Long and Nikolay Atanasov have developed reformulations into Second-Order Cone Programs (SOCP) for real-time safety. In Safe and Stable Control Synthesis for Uncertain System Models via Distributionally Robust Optimization, they show that these methods can ensure a robot remains safe even when the underlying uncertainty distribution shifts during execution.

Applications Across Domains

The versatility of the Wasserstein DRO framework is evident in its broad application:

By treating the uncertainty distribution as a variable within a geometric "ball" rather than a fixed assumption, this research direction provides a rigorous, tunable, and increasingly efficient way to build autonomous systems that can handle the unpredictability of the real world.

Papers in this version

Showing 60 of 428 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 449 papers
  2. Version 1 · 2026-10-04 · 428 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_wasserstein_ambiguity_sets_8e4b462b,
  title = {Wasserstein ambiguity sets},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/wasserstein-ambiguity-sets/rdr_8e4b462bc54d468580325564dcde8711},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_wasserstein_distributionally_robust_cont_8e4b462b_v1,
  title = {Wasserstein Distributionally Robust Control and Optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/wasserstein-distributionally-robust-control-and-optimization/rdr_8e4b462bc54d468580325564dcde8711/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}