WARP Benchmark Exposes Flaws in ML Warm-Starting for AC-OPF
A recent study published on arXiv (2605.05728) presents WARP, a benchmark aimed at assessing the primal-dual warm-starting capabilities of interior-point solvers in AC Optimal Power Flow (AC-OPF). The researchers reveal that earlier machine learning approaches, which reported iteration reductions of 30-46%, relied on an unsuitable baseline—the flat start—rather than the solver's actual default, the variable-bound midpoint. When evaluated against this revised baseline, no primal-only warm-start technique achieves a reduction in solver iterations. This inadequacy is linked to a geometric characteristic: the accuracy of primal predictions inversely affects convergence speed, and supplying the true optimal solution without dual variables leads to divergence. Oracle tests validate the necessity of dual information for successful warm-starting.
Key facts
- Paper arXiv:2605.05728 introduces WARP benchmark for warm-starting IPMs in AC-OPF.
- Prior ML methods reported 30-46% iteration reductions using flat start baseline.
- Correct baseline is variable-bound midpoint (l+u)/2, which is near-optimal for log-barrier centrality.
- No primal-only warm-start method reduces iterations against corrected baseline.
- Primal prediction accuracy is anticorrelated with convergence speed.
- Ground-truth optimal solution x* without dual variables causes solver divergence.
- Oracle experiments show dual information is necessary for effective warm-starting.
- WARP benchmark enables fair comparison of primal-dual warm-start methods.
Entities
Institutions
- arXiv