Published results
What the calendar costs, measured.
Optimality gap is the avoidable stress-weighted crop-water deficit, plus water cost, above the best schedule the model admits. Every instance is built from real 2024 NASA POWER weather, ISRIC SoilGrids hydraulics and documented canal rotation practice in Khorezm.
The full ladder
| Instance | n | State calendar | Trigger rule | Annealing | GA (10× budget) | Reference |
|---|---|---|---|---|---|---|
| 3 zones x 7 days | 11 | 0% | +48.7% | 0% | 0% | proven |
| 3 zones x 14 days | 21 | +35.2% | +110.8% | +2.1% | 0% | proven |
| 3 zones x 28 days | 40 | +171.1% | +98.2% | +62.9% | 0% | proven |
| 6 zones x 28 days | 77 | +197.5% | +97.3% | +178.3% | 0% | proven |
| 12 zones x 28 days | 150 | +200.4% | +84.3% | +265.5% | +2.8% | proven |
| 24 zones x 28 days | 295 | +165.7% | +53.9% | +387.5% | +5.2% | best known |
| 24 zones x 56 days | 584 | +97.9% | +9.7% | +606.7% | +7.3% | best known |
On the one-week instance the pro-rata calendar happens to coincide with the optimum — rigidity only starts costing once the horizon is long enough for conditions to diverge from the plan. Above n=150 the optimum is not proven; the reference is the best solution found, so those gaps are lower bounds on the real ones.
Provenance
These figures are not re-derived in the browser. They are the committed outputs of the research pipeline, vendored into this app so a change to them shows up as a reviewable diff.
- Source
- mcpeblocker/amuflow-web @ ecb589d
- Solver run
- Python 3.14.3 · numpy 2.4.2 · 2026-07-10
- Budgets
- 20000 evaluations × 20 seeds · MIQP capped at 900 s