L2O-CCG: Adversarial Learning with Set Generalization for Adaptive Robust Optimization
New bi-level framework replaces complex adversarial search with a learned optimizer that generalizes across uncertainty sets.
A team of researchers including Zhiyi Zhou, Ján Drgoňa, and Yury Dvorkin has introduced L2O-CCG, a breakthrough framework that tackles a major bottleneck in adaptive robust optimization (ARO). In two-stage ARO problems—common in energy systems, finance, and logistics—finding the worst-case uncertainty realization (the adversarial subproblem) is computationally intensive, especially when the recourse function is non-concave or when the uncertainty set changes. Traditional methods degrade when their specific structural assumptions are violated at deployment. L2O-CCG addresses this by creating a bi-level framework that integrates structure-aware solvers within the established Constraint-and-Column Generation (CCG) algorithm.
The core innovation is an instantiation of the framework that replaces the traditional solver with a learned optimizer. An inner neural network approximates the complex recourse value function from offline data, while an outer pre-trained mapping generates precise, iteration-dependent step sizes for a proximal gradient scheme. This learned optimizer can generalize to new, unseen uncertainty set geometries without needing retraining, a significant leap in flexibility. The team provided rigorous theoretical backing, establishing out-of-distribution convergence bounds that quantify how the solution trajectory deviates in response to shifts in the uncertainty set parameters. They demonstrated the method's practical efficacy on a building HVAC management task, showing it can maintain robust performance where previous methods would fail or require costly re-engineering.
- Replaces traditional adversarial search solvers with a learned proximal gradient optimizer powered by neural networks.
- Achieves generalization across different uncertainty set geometries without retraining, overcoming a key deployment hurdle.
- Provides proven out-of-distribution convergence bounds, linking optimizer performance directly to the magnitude of set parameter shifts.
Why It Matters
Enables more resilient and efficient real-world systems in energy, finance, and logistics by making robust optimization adaptable and faster.