Solver-Admitted Metamorphic Testing of Learned Network Optimization Surrogates
Summary
- Researchers introduce solver-admitted metamorphic testing with explicit eligibility conditions, output alignment, and numerical thresholds to audit learned optimization solver surrogates.
- The study evaluates alternating-current optimal power flow (AC-OPF), the traveling salesman problem (TSP), the capacitated vehicle routing problem (CVRP), and maximum independent set (MIS).
- Evaluations show that models like GridSFM-Open, POMO, and DIFUSCO experience significant metamorphic relation violations that persist across training states or scale with problem size.
- Post hoc interventions, input canonicalization, and source patches reduce certain discrepancies but fail to guarantee preserved accuracy and feasibility across all test grids.
Why it matters
This research highlights reliability risks in learned optimization surrogates and advocates for reporting metamorphic consistency alongside standard accuracy and optimality metrics.