Trustworthy optimization learning: a brief overview

Pascal Van HentenryckOpenalex Unknown 2025 · DOI · OpenAlex

Abstract

This article introduces the concept of trustworthy optimization learning, a methodology to design optimization proxies that learn the input/output mapping of parametric optimization problems. These optimization proxies are trustworthy by design: they compute feasible solutions to the underlying optimization problems, provide quality guarantees on the returned solutions, and scale to large instances. Optimization proxies are differentiable programs that combine traditional deep learning technology with repair or completion layers to produce feasible solutions. The article shows that optimization proxies can be trained end-to-end in a self-supervised way. It presents methodologies to provide performance guarantees and to scale optimization proxies to large-scale optimization problems. The potential of optimization proxies is highlighted through applications in power systems and, in particular, real-time risk assessment and security-constrained optimal power flow.

Introduction: The Need for Trustworthy Proxies

In critical infrastructure management—such as power grid dispatch, supply chain logistics, and real-time risk assessment—operators frequently face parametric optimization problems. These problems require solving the same mathematical model repeatedly as inputs (e.g., demand, weather, or component status) vary. While traditional solvers offer high accuracy, their computational latency often precludes real-time application.

Deep learning offers a potential solution by approximating these mappings in milliseconds. However, standard neural networks typically treat optimization as a regression task, often producing outputs that violate hard physical constraints (e.g., Kirchhoff’s laws). In safety-critical systems, such "untrustworthy" approximations are unusable. Trustworthy optimization learning addresses this by embedding mathematical guarantees directly into the model architecture.

The Optimization Proxy Framework

The core innovation is the optimization proxy, a differentiable program that fuses deep learning with traditional optimization logic. Unlike a black-box regressor, a proxy consists of two primary stages: prediction and repair.

General architecture of an optimization proxy, mapping inputs to feasible outputs via prediction and repair stages.

The prediction stage uses standard neural network layers to generate an initial approximation. This approximation is then passed to a "repair" or "completion" layer. These layers are designed to project the prediction onto the feasible region of the optimization problem. Because these operations are differentiable, the entire system can be trained end-to-end using gradient-based methods.

Ensuring Feasibility through Differentiable Repair

Feasibility is treated as a structural requirement rather than a penalty in the loss function. By building "repair" operations directly into the network, the model is mathematically incapable of producing an infeasible solution for the targeted constraints.

Implementation of differentiable repair layers within a neural network to enforce physical constraints like power balance.

For example, in power systems, a repair layer might proportionally scale generator outputs to ensure they exactly match demand (power balance). As shown in Figure 2, these layers follow the sigmoid or hidden layers and precede the final loss calculation, allowing the backpropagation of error signals through the physical constraints themselves.

Training and Scalability Methodologies

Trustworthy optimization learning employs several advanced training strategies to ensure high performance and scalability:

  • Self-Supervised Learning: Proxies are trained to minimize the original optimization objective directly. This removes the need for massive, pre-computed "ground truth" datasets, which are often expensive to generate.
  • Primal-Dual Learning: To handle complex inequality constraints, the framework incorporates Augmented Lagrangian Methods. It jointly trains a primal network (to minimize the objective) and a dual network (to update Lagrangian multipliers), ensuring the model learns the sensitivity of the constraints.
  • Compact Optimization Learning: To scale to massive networks (e.g., the 6,515-bus RTE French grid), the framework utilizes dimensionality reduction. By learning to predict a small subset of Principal Components (PCA) of the output space, the model can efficiently handle high-dimensional engineering problems where the solution space is highly structured.

Key Results and Applications

The effectiveness of this approach has been demonstrated in several high-stakes scenarios. In real-time power system risk assessment, optimization proxies reduced computation time from 15 minutes to 5 seconds, enabling real-time dashboards. For Security-Constrained Optimal Power Flow (SCOPF) on the RTE network, a primal-dual proxy was trained in under 3 hours and achieved inference times of 10 milliseconds with an optimality gap of less than 1%. Crucially, these proxies eliminated false positives in adverse event detection by guaranteeing physical feasibility.

Future Directions

While optimization proxies provide instance-specific quality guarantees, future work aims to establish formal, worst-case safety guarantees that hold across all possible inputs. Additionally, extending these differentiable frameworks to handle discrete or combinatorial variables remains a significant research frontier, as gradients are not readily available for such problems.

Summary updated

Abstract

This article introduces the concept of trustworthy optimization learning, a methodology to design optimization proxies that learn the input/output mapping of parametric optimization problems.These optimization proxies are trustworthy by design: they compute feasible solutions to the underlying optimization problems, provide quality guarantees on the returned solutions, and scale to large instances.Optimization proxies are differentiable programs that combine traditional deep learning technology with repair or completion layers to produce feasible solutions.The article shows that optimization proxies can be trained end-to-end in a self-supervised way.It presents methodologies to provide performance guarantees and to scale optimization proxies to large-scale optimization problems.The potential of optimization proxies is highlighted through applications in power systems and, in particular, real-time risk assessment and security-constrained optimal power flow.
BibTeX
@inproceedings{hentenryck2025trustworthy,
  author = {Pascal Van Hentenryck},
  title = {{Trustworthy optimization learning: a brief overview}},
  booktitle = {Openalex Unknown},
  year = {2025},
  doi = {10.1515/9783111376776-009},
}
Figure