This paper, presented at the ROADEF 2026 conference, explores how to make optimization models more "human-friendly" when they fail. Specifically, it applies the concept of Counterfactual Explanations to radar coverage planning, transforming a "computer says no" situation into a constructive negotiation between the operator and the algorithm.
TL;DR
When radar coverage requirements clash with a tight energy budget, standard optimization models simply return "infeasible." This paper proposes a mathematical framework that instead suggests the minimal adjustments—such as reducing coverage in a specific area or slightly increasing the budget—needed to make the plan work.
What Problem Does This Paper Address?
In radar operations, planning is a balancing act. You need to select a set of "dwells" (specific waveforms and durations) to cover various cells in the sky. This is mathematically modeled as a Weighted Set Cover Problem (WSCP): you want to satisfy all coverage demands at the lowest cost (time or energy) while staying within a fixed budget.
The problem arises when the requirements are impossible to meet. Traditional solvers are binary; they either give you an optimal plan or tell you the problem is infeasible. For a radar operator in the field, "infeasible" is useless. They need to know why it failed and what they can change to fix it. Currently, this involves tedious manual trial-and-error. The authors aim to automate this "what-if" analysis.
Methods
The authors extend the classical Mixed Integer Linear Program (MILP) formulation of the Set Cover Problem by introducing Counterfactual Explainability, a concept borrowed from Explainable AI (XAI).
- Elastic Variables: They introduce perturbation variables ($\delta$) for every major parameter of the problem:
- $\delta^A$: Changes to the coverage matrix (e.g., can a radar dwell reach a bit further?).
- $\delta^b$: Changes to the coverage demand (e.g., can we accept less density in one sector?).
- $\delta^c$: Changes to the cost of dwells.
- $\delta^B$: Changes to the total available budget.
- The Counterfactual Objective: Instead of just minimizing cost, the new objective function minimizes the original cost plus a "penalty" for the modifications. This ensures the suggested changes are as small and few as possible.
- Linearization: Because multiplying decision variables by these new perturbations creates non-linear "bilinearities," the authors use auxiliary variables and "big-M" constraints to keep the problem solvable using standard linear programming tools.
- Domain Constraints: They emphasize that these changes can't be random. For example, a radar dwell can only "expand" its coverage if it's physically possible (correct orientation and range), ensuring the suggestions are realistic.
Datasets
The paper does not utilize a public benchmark dataset. Instead, it is based on a Thales radar use case. The authors focus on the theoretical formulation of the model rather than providing a large-scale empirical evaluation on standardized sets.
Key Insights and Ideas
- From Optimization to Negotiation: The core shift is moving from "finding the best solution" to "finding the closest feasible world." This turns the optimization tool into a decision-support system.
- Actionable Insights: The model is designed to produce human-readable advice. Instead of a matrix of numbers, it can output: "Reducing coverage in Area X by 5% or increasing budget by 10% restores feasibility."
- Physical Grounding: A major insight is that counterfactuals in optimization must respect physical laws. You can't just "tweak" a coverage bit if the radar isn't pointed in that direction. The model incorporates these operational rules into the $\delta^A$ perturbations.
Key Results
As this is a short "work-in-progress" style paper for a conference, it focuses on the mathematical formulation (Equation 4) rather than quantitative performance metrics.
The qualitative result is the demonstration of a framework that can successfully identify which parameters (coverage, cost, or budget) are the "bottlenecks" of a specific scenario. By quantifying the "minimal adjustment," the authors provide a path toward restoring feasibility that was previously a manual, intuitive process for engineers.
Limitations and Future Work
- Computational Complexity: Adding elastic variables and big-M constraints significantly increases the search space for the MILP solver. The authors will likely need to investigate heuristics or decomposition methods for very large-scale radar grids.
- Defining "Minimal": The "penalty" function is crucial. Future work will need to determine how to weight different types of changes—is a 5% budget increase "cheaper" than a 2% coverage drop? This likely requires user studies with actual radar operators.
- Multi-Objective Trade-offs: The current model combines cost and penalties into one objective. Exploring a Pareto front of different counterfactual suggestions could give operators multiple "fix-it" options to choose from.
Summary updated