Title: Position Paper: Post-Solve Robustness in Decision Engines: Feasible Regions and Smoothness Under Perturbations

Abstract:

Mixed-Integer Linear Programming (MILP) decision engines are typically deployed to generate nominally optimal plans for critical industrial systems. However, real-world implementation often diverges from the assumptions made during the solving phase. Minor fluctuations in costs, demand levels, or resource availability can render a plan infeasible or cause abrupt, qualitative shifts to entirely different solutions. This paper posits that the lack of post-solve robustness represents a significant oversight in current optimization workflows and serves as an under-evaluated metric for learning-enhanced decision systems.

Instead of supplanting established methods like robust optimization or stochastic programming, we propose an additional auditing layer. This layer examines a solved incumbent solution and provides solver-verified evidence regarding the extent to which the solution’s reliability can be trusted. We define two primary constructs: (i) an $\epsilon$-near-optimal feasible neighborhood within parameter space, which delineates the conditions under which an incumbent solution remains both feasible and near-optimal despite perturbations; and (ii) solution smoothness in decision space, which assesses whether alternative solutions requiring only minor combinatorial adjustments remain competitively viable.

By synthesizing key insights from sensitivity and stability analysis, robust optimization, neighborhood search, adversarial testing, and learning-based improvements, we outline a roadmap for a cohesive post-solve robustness framework. Specifically, we advocate for the integration of certified inner approximations surrounding the incumbent, probabilistic robustness estimates with calibrated uncertainty, adversarial robustness margins, and learning-driven prediction and explanation mechanisms that align with solver-backed verification. The paper concludes by presenting a streamlined reporting template and evaluation protocol designed to elevate robustness to a primary output of decision engines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC