Recorded experiment results
16 independent delivery routes, tested in 1,024 correlated simulated situations per method.
| Method | Average route time | Average late packages |
|---|---|---|
| Original inputs | 482.7 min | 0.43 |
| Simple correction | 475.1 min | 0.281 |
| CatBoost correction | 469.1 min | 0.034 |
Same planner and search budget. Route time includes driving, service and waiting.
CatBoost saved 6 min per simulated plan compared with the simple correction.
Saved 95% interval: 4.9 min to 7 min. This is an exploratory estimate from the 16 routes, after multiple model comparisons.
| Compared with | Over 5 min slower | Worst slowdown | More late packages |
|---|---|---|---|
| Original inputs | 24 / 1,024 | 17.5 min | 1 / 1,024 |
| Simple correction | 47 / 1,024 | 134.5 min | 1 / 1,024 |
Counts are simulated plans where CatBoost did worse than the named method. Smaller slowdowns are not included in the “over 5 min” count.
Round 8 development simulations on previously inspected routes. The displayed configuration has not passed all qualification checks or received a fresh confirmation or final test.