Data-Driven Ambiguity Sets With Probabilistic Guarantees for Dynamic Processes

Dimitris Boskos, Jorge Cortés, Sonia Martı́nezIEEE Trans. Automatic Control 2020 · DOI · arXiv

Abstract

Distributional ambiguity sets provide quantifiable ways to characterize the uncertainty about the true probability distribution of random variables of interest. This makes them a key element in data-driven robust optimization by exploiting high-confidence guarantees to hedge against uncertainty. This article explores the construction of Wasserstein ambiguity sets in dynamic scenarios, where data are collected progressively and may only reveal partial information about the unknown random variable. For random variables evolving according to known dynamics, we leverage assimilated samples to make inferences about their unknown distribution at the end of the sampling horizon. Under exact knowledge of the flow map, we provide sufficient conditions that relate the growth of the trajectories with the sampling rate to establish a reduction of the ambiguity set size as the horizon increases. Furthermore, we characterize the exploitable sample history that results in a guaranteed reduction of ambiguity sets under errors in the computation of the flow and when the dynamics is subject to bounded unknown disturbances. Our treatment deals with both full- and partial-state measurements and, in the latter case, exploits the sampled-data observability properties of linear time-varying systems under irregular sampling. Simulations on an unmanned aerial vehicle detection application show the superior performance resulting from the proposed dynamic ambiguity sets.

Introduction

In Distributionally Robust Optimization (DRO), decision-makers seek to optimize for the worst-case scenario within an "ambiguity set"—a collection of probability distributions that likely contain the true distribution of a random variable. While standard DRO techniques are effective for static datasets, they struggle with dynamic processes where the underlying distribution evolves over time. Simply discarding historical data leads to weak, data-poor models, while using it without adjustment leads to inaccuracies. This paper introduces a framework to mathematically "translate" past data into the present, enabling the construction of tighter, more reliable Wasserstein ambiguity sets for dynamic systems.

Methodology

Pushforward Maps and Data Assimilation

The core innovation of this framework is the use of a system's flow map (its physical laws of motion) to transform past samples into "predicted" samples at the current time. By "pushing forward" historical data points through the known dynamics, we can assimilate them into a single, larger dataset representing the current state distribution.

Illustration of pushforward maps transforming historical samples to the current time horizon.

Error Characterization and Observability

Real-world applications rarely feature perfect models. The framework incorporates terms for numerical integration errors and bounded disturbances to maintain probabilistic guarantees. A key concept introduced is the "effective sampling horizon"—a temporal limit beyond which the accumulation of model errors outweighs the benefit of additional historical data. For systems with partial-state measurements (e.g., observing position but not velocity), the authors leverage linear system theory to reconstruct full state distributions from multiple partial measurements over time.

Key Insights: The Data Assimilation Trade-off

There is a fundamental trade-off between the number of samples and the accuracy of the pushforward operation. While more samples generally shrink the ambiguity set radius (increasing precision), samples from the distant past accumulate more pushforward error due to model uncertainty and disturbances.

Analysis of the trade-off between sample size and sampling rate on the ambiguity set guarantees.

As shown in the analysis of the ambiguity set bound $\psi_N$, there exists an optimal number of samples $N^*$ for a given sampling rate $\Delta$. If the system's dynamics do not grow too rapidly relative to the sampling rate, the ambiguity set is guaranteed to shrink as more data is collected, providing a significant advantage over static approaches.

Results and Applications

The framework was evaluated in a UAV detection scenario, where a "blue" UAV must navigate a grid monitored by "red" UAVs following unknown circular orbits. By building dynamic ambiguity sets from sparse position measurements, the blue UAV could plan a path that maximizes the worst-case distance from the red targets.

Key results include:

  • Radius Reduction: Using 160 assimilated samples reduced the ambiguity radius to 0.085, a nearly 4x improvement over the 0.3023 radius obtained using only the single latest sample.
  • Performance: The "Dynamic DRO" approach significantly outperformed static methods, maintaining 95% confidence guarantees even in the presence of disturbances and partial observability.

Conclusion

By bridging control theory and stochastic optimization, this research provides a robust method for handling time-varying uncertainty. While the approach requires a reasonably accurate model of system dynamics, it demonstrates that physics-informed data assimilation can turn sparse historical data into a powerful tool for high-confidence robust decision-making.

Summary updated

Abstract

Distributional ambiguity sets provide quantifiable ways to characterize the uncertainty about the true probability distribution of random variables of interest. This makes them a key element in data-driven robust optimization by exploiting high-confidence guarantees to hedge against uncertainty. This article explores the construction of Wasserstein ambiguity sets in dynamic scenarios, where data are collected progressively and may only reveal partial information about the unknown random variable. For random variables evolving according to known dynamics, we leverage assimilated samples to make inferences about their unknown distribution at the end of the sampling horizon. Under exact knowledge of the flow map, we provide sufficient conditions that relate the growth of the trajectories with the sampling rate to establish a reduction of the ambiguity set size as the horizon increases. Furthermore, we characterize the exploitable sample history that results in a guaranteed reduction of ambiguity sets under errors in the computation of the flow and when the dynamics is subject to bounded unknown disturbances. Our treatment deals with both full- and partial-state measurements and, in the latter case, exploits the sampled-data observability properties of linear time-varying systems under irregular sampling. Simulations on an unmanned aerial vehicle detection application show the superior performance resulting from the proposed dynamic ambiguity sets.
BibTeX
@article{boskos2020data,
  author = {Dimitris Boskos and Jorge Cortés and Sonia Martı́nez},
  title = {{Data-Driven Ambiguity Sets With Probabilistic Guarantees for Dynamic Processes}},
  journal = {IEEE Transactions on Automatic Control},
  year = {2020},
  doi = {10.1109/tac.2020.3014098},
}
Figure