Minimax Optimization for Robust Control under Uncertainty

22 concepts17 papers1 proposal

When designing controllers for physical systems—whether steering an autonomous vehicle, managing a smart power grid, or training a robot to manipulate objects—we must confront a fundamental truth: our mathematical models are imperfect. Real-world systems are plagued by measurement noise, unmodeled environmental dynamics, and hidden confounding variables.

Traditional control and machine learning strategies often optimize for the nominal or average case. They assume a single best-guess model or minimize expected error over a known noise distribution. However, as noted by Michael Eichelbeck et al. in CommonPower: A Framework for Safe Data-Driven Smart Grid Control, naive optimal control strategies can fail rapidly when faced with real-world uncertainties and disturbances. If the true system deviates even slightly from the assumed average, an average-case controller can experience catastrophic instability.

To build systems that are guaranteed to remain safe, researchers turn to minimax optimization and worst-case error minimization. Instead of asking "how do we perform best on average?", this paradigm asks: "How do we perform best when conditions are at their worst?"

The Intuition: A Game Against an Adversary

The core mechanism of worst-case optimization is best understood as a two-player, zero-sum game. Imagine you are designing a controller, and your opponent is an omniscient "adversary" representing nature, noise, or modeling errors.

  1. You choose a control action $u$ to minimize a cost function (such as tracking error or energy consumption).
  2. Simultaneously, the adversary chooses a disturbance $w$ from a bounded set of possible errors to maximize that same cost function.

Mathematically, this is expressed as a min-max problem:

$$\min_{u} \max_{w \in \mathcal{W}} J(u, w)$$

By solving this optimization, you identify a control strategy that minimizes the maximum possible damage the environment can inflict. If your system can survive the absolute worst-case disturbance within the uncertainty set $\mathcal{W}$, it is guaranteed to remain stable and satisfy constraints under any milder, real-world realization of that uncertainty.

Historically, this philosophy marked a major evolution in engineering. As detailed in the retrospective by J Pearson et al. (James Boyd Pearson [Obituary]), the transition to robust control frameworks prioritized system stability under model uncertainty over mere nominal performance, utilizing mathematical norms to bound the maximum possible instability.

Quantifying the Worst Case: The $H^\infty$ Criterion

To operationalize this min-max game, engineers need a metric to quantify how severely disturbances can amplify through a system. This is the role of the $H^\infty$ norm.

In system theory, the $H^\infty$ norm represents the maximum energy amplification (or worst-case gain) from an external disturbance to the system's internal error. Minimizing the $H^\infty$ norm ensures that even if an adversarial disturbance hits the system at its most vulnerable frequency, the resulting error remains strictly bounded.

This criterion provides powerful robustness guarantees across diverse fields:

Robust Model Predictive Control

In dynamic environments, controllers must make decisions sequentially over a future time horizon. Model Predictive Control (MPC) achieves this by solving an optimization problem at each time step. When infused with the minimax principle, it becomes Robust MPC.

Instead of predicting a single future trajectory, a Robust MPC controller evaluates how a sequence of control actions will perform against the worst-case sequence of future disturbances. Jia Kang et al. demonstrate this advantage in Parallel Solution of Robust Nonlinear Model Predictive Control Problems in Batch Crystallization, showing that min-max formulations yield superior constraint satisfaction across all possible uncertainty realizations compared to standard nominal MPC.

Overcoming Over-Cautiousness

A common critique of worst-case optimization is that it can lead to "over-cautious" behavior; if the uncertainty bound $\mathcal{W}$ is too large, the controller might act timidly to protect against a catastrophic scenario that is highly improbable.

To resolve this, researchers design adaptive uncertainty sets. For example, Lei Chen et al. introduced an adaptive bound estimator in Reducing energy consumption for buildings under system uncertainty through robust MPC with adaptive bound estimator. By using a neural network to shrink or expand the uncertainty bounds in real-time based on environmental feedback, the controller balances energy efficiency with worst-case reliability.

Handling Hidden Confounders

Minimax MPC is also uniquely suited for environments with unobserved variables. In Off-policy Predictive Control with Causal Sensitivity Analysis, Myrl G Marmarelis et al. introduced a controller that optimizes for the worst-case scenario admitted by hidden confounding constraints. By maximizing the lower bound of the expected reward, the system achieves exceptional robustness despite missing information.

To deploy these computationally heavy min-max optimizations in fast, real-time applications like autonomous driving, Carlos Hernández et al. (Explicit multiobjective model predictive control for nonlinear systems under uncertainty) developed methods to pre-compute libraries of minimax solutions offline, drastically reducing the online computational burden.

Safe Exploration and Data-Driven Control

The stakes of worst-case optimization are highest when a controller must learn from data in an unknown environment. In reinforcement learning, standard "optimistic" exploration strategies encourage an agent to try actions based on the best-case assumption of what might happen. On physical hardware, however, an optimistic gamble on an unknown parameter can cause a system to break.

Minimax optimization enables robust exploration. Instead of assuming the best, the agent identifies a controller that maintains stability for the worst-case realization of the system parameters consistent with the collected data.

Optimistic Exploration:  [Current Data] ──> Assume Best-Case Model  ──> Aggressive Action (Risk of Failure)
Robust Exploration:      [Current Data] ──> Bound Worst-Case Model ──> Safe, Targeted Action (Guaranteed Stability)

As shown by Sarah Dean et al. in Regret Bounds for Robust Adaptive Control of the Linear Quadratic Regulator, prioritizing worst-case performance within uncertainty bounds prevents catastrophic failures during online learning. To make this learning efficient, Jack Umenberger et al. (Robust exploration in linear quadratic reinforcement learning) designed an exploration strategy that actively excites the system to reduce uncertainty specifically in the parameters that most heavily impact the worst-case cost.

This worst-case framing extends seamlessly to purely data-driven, non-parametric control paradigms:

  • Noisy Data Streams: In Robust and Kernelized Data-Enabled Predictive Control for Nonlinear Systems, Linbin Huang et al. embedded a min-max framework within a kernelized predictor, finding control inputs that minimize costs while maximizing resilience against worst-case data disturbances.
  • Networked Communication Constraints: When controlling a system over a limited network, Wenjie Liu et al. (Data-driven Self-triggered Control via Trajectory Prediction) used a norm-maximization problem to calculate the most conservative future trajectory consistent with historical data. This worst-case analysis allows the system to safely maximize communication intervals without risking instability.

Alternative Formulations: Games and Variance

While the min-max formulation is the most direct way to implement worst-case optimization, the same underlying philosophy manifests in other mathematical structures.

Explicit Game Formulations

Rather than treating the adversary as an abstract optimization constraint, some frameworks model it explicitly as a dynamic player. Olalekan P. Ogunmolu et al. developed a minimax iterative dynamic game framework in Minimax Iterative Dynamic Game: Application to Nonlinear Robot Control Tasks to explicitly quantify policy brittleness in robotics. Similarly, Andrey Polyakov et al. leveraged a minimax framework in Sliding Mode Control Design for Linear Evolution Equations with Uncertain Measurements and Exogenous Perturbations to guarantee that a system's state reaches a desired sliding hyperplane despite incomplete and noisy measurements.

Multi-Objective Variance Maximization

Instead of optimizing against an explicit adversary, robustness can be achieved by treating sensitivity as an independent objective. In Goal-Image Conditioned Dynamic Cable Manipulation through Bayesian Inference and Multi-Objective Black-Box Optimization, Kuniyuki Takahashi utilized multi-objective optimization to simultaneously minimize tracking error and maximize a variance parameter. By explicitly optimizing for execution robustness, the system selects control actions that are inherently insensitive to small execution errors or model inaccuracies.

Conclusion

Worst-case error minimization works because it replaces optimism with mathematical guarantees. By framing control as a game against the most unfavorable realization of uncertainty, frameworks utilizing the $H^\infty$ norm, robust MPC, and minimax exploration ensure that physical systems remain stable, predictable, and safe—no matter what the real world throws at them.

Key Foundational Work

Zero-forcing precoding for frequency selective MIMO channels with H∞ criterion and causality constraint (2009) — contributed 1 concept elements to this direction.
Hoo Optimality Criteria for LMS and Backpropagation (1993) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_worst_case_error_minimization_1dc91a5d,
  title = {worst-case error minimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/worst-case-error-minimization/rdr_1dc91a5d292b4666bd29d2af29b4141e},
  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_worst_case_error_minimization_1dc91a5d_v1,
  title = {worst-case error minimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/worst-case-error-minimization/rdr_1dc91a5d292b4666bd29d2af29b4141e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake