Overview
This research investigates distributionally robust linear chance-constrained problems where underlying uncertainty is characterized by a Gaussian mixture model (GMM). The work introduces a novel approach to define an ambiguity set of distributions, leveraging a Wasserstein-2 metric that integrates the Bures-Wasserstein (BW) metric. This method is designed to address identified limitations in existing finite-support distributionally robust (FDR) formulations, particularly concerning potential structural misspecification of nominal mixture-support parameters.
Research Context
Traditional data-driven robust optimization frequently employs finite-support distributionally robust (FDR) formulations. These formulations typically robustify over empirical mixture support points. A key characteristic of FDR is its primary focus on stress-testing the nominal mixture as it has been fitted. However, this methodology may prove insufficient, especially in scenarios where the reliability of a service is susceptible to structural misspecification within the parameters of the nominal mixture's support.
Approach
To overcome the limitations of FDR, the study defines an ambiguity set for distributions through a newly formulated Wasserstein-2 metric. This metric incorporates the Bures-Wasserstein (BW) metric, which is applicable to probability measures possessing finite second moments. A distinction from FDR is that the proposed ambiguity set does not set a finite number of empirical support points a priori. Instead, it allows for the worst-case distribution to endogenously determine two critical aspects: the number of mixture components that receive mass, and the specific locations of their means and covariances within a continuous support space.
For this specific ambiguity set, the researchers prove strong duality for the inner worst-case chance-constraint problem. This proof holds under mild regularity conditions. Following this, they derive a semi-infinite reformulation of the problem. To solve this reformulated problem, an adaptive cutting-surface algorithm is developed. This algorithm endogenously determines the locations of mixture components receiving mass, as well as the means and covariances of the Gaussian distributions at these determined locations. The algorithm is designed to achieve any prescribed optimality gap within a finite number of iterations. A block-alternating local search mechanism is employed within the algorithm to identify new components.
Findings
The core findings relate to the theoretical properties and practical capabilities of the proposed framework:
- The defined ambiguity set, utilizing a continuous parameter space Wasserstein-2 metric, allows the worst-case distribution to endogenously specify both the quantity of mixture components receiving mass and the positioning of their means and covariances within a continuous support. This contrasts with FDR, which pre-sets empirical support points.
- Strong duality for the inner worst-case chance-constraint problem was proven for the resulting ambiguity set, contingent upon mild regularity conditions.
- A semi-infinite reformulation of the problem was derived.
- An adaptive cutting-surface algorithm was developed capable of endogenously determining mixture component locations, means, and covariances. This algorithm achieves a prescribed optimality gap in a finite number of iterations.
- A block-alternating local search procedure facilitates the identification of new components.
Why This Matters
A case study involving the electric-vehicle charging-station energy-allocation problem was conducted to evaluate the framework's practical utility. The study demonstrated that the proposed framework is effective in achieving specified reliability targets. Furthermore, the robust chance-constrained optimization (CDR) framework induced structural modifications in energy allocations, a behavior not observed in FDR, where allocations remained proximate to the nominal solution. This indicates a potential for more adaptive and robust decision-making in real-world applications where distribution uncertainty is modeled via Gaussian mixtures.