Skip to content

Results

Baseline distances

Notation Variant Solver limits No port With port
NP-MIP No-port directed TSP L_NP = 2 hrs 3562.57 N/A
C-MIP Capacity-aware MIP L_cap = 12 hrs N/A timeout
MH-NN MH with nearest-neighbor L_2seg = 10-600 s N/A 6911.67
MH-CI MH with cheapest-insertion L_2seg = 10-600 s 4377.20 5354.93
MH-GE MH with greedy-edge L_2seg = 10-600 s 5027.03 5104.34
MH-noport MH with NP-based initialization L_2seg = 10-600 s NP-MIP 4762.17

The four initialization strategies yield very different starting points. NN gives the weakest capacity-feasible baseline, while noport gives the strongest baseline at the cost of a separate no-port MIP solve.

Overall findings

Two broad patterns emerge:

  1. The experiment sweep uses capped two-segment limits only, up to 600 seconds.
  2. Moderate two-segment limits, around 2 to 3 minutes, tend to perform best for a given initialization.
  3. Larger limits substantially increase runtime, so the 600-second refinement is used as the detailed before/after/table example rather than running uncapped refinement.

The practical implication is that many short local MIP solves are more effective than fewer long ones.

MH-NN

L_2seg (s) Sweeps Final distance Stations moved mean/med/max Accepted/solves MIP gap min/mean/max/std (%) Runtime (min)
60 7 5842.279 6 / 6 / 12 16 / 62 0 / 13 / 99 / 16 145
120 7 5994.583 5.4 / 4 / 12 12 / 62 0 / 7 / 32 / 8 273
180 7 6171.916 5.6 / 4 / 11 15 / 61 0 / 6 / 18 / 6 351
240 8 6263.959 4.4 / 2.5 / 11 14 / 55 0 / 7 / 99 / 13 431
300 6 6112.102 6.2 / 7 / 12 11 / 54 0 / 5 / 19 / 5 517
360 5 6067.262 6.6 / 8 / 12 10 / 51 0 / 4 / 33 / 7 560
420 7 6116.561 5.1 / 6 / 12 13 / 56 0 / 3 / 14 / 4 680
480 6 5894.774 5.5 / 5 / 12 14 / 49 0 / 4 / 28 / 5 727

MH-CI

L_2seg (s) Sweeps Final distance Stations moved mean/med/max Accepted/solves MIP gap min/mean/max/std (%) Runtime (min)
60 2 5000.347 5.5 / 5.5 / 11 4 / 24 0 / 14 / 99 / 20 66
120 2 5108.418 4.5 / 4.5 / 9 3 / 22 0 / 6 / 18 / 6 86
180 3 5062.339 4.0 / 2 / 10 4 / 24 0 / 7 / 18 / 6 155
240 5 5012.521 3.8 / 3 / 10 7 / 33 0 / 6 / 17 / 5 267
300 3 5067.459 5.3 / 6 / 10 7 / 28 0 / 5 / 14 / 5 247
360 5 4886.220 3.4 / 2 / 9 8 / 32 0 / 3 / 14 / 4 365
420 3 5108.968 3.7 / 3 / 8 4 / 24 0 / 5 / 13 / 4 367
480 4 5025.292 3.5 / 3 / 8 7 / 27 0 / 5 / 13 / 4 485

MH-GE

L_2seg (s) Sweeps Final distance Stations moved mean/med/max Accepted/solves MIP gap min/mean/max/std (%) Runtime (min)
60 6 4799.879 3.5 / 2.5 / 9 12 / 39 0 / 12 / 37 / 9 109
120 5 4751.567 4.6 / 5 / 8 12 / 40 0 / 11 / 52 / 10 229
180 6 4771.586 4.7 / 3.5 / 10 16 / 44 0 / 8 / 25 / 7 313
240 3 4850.068 4 / 4 / 8 7 / 27 0 / 8 / 29 / 7 267
300 5 4773.525 3.8 / 3 / 9 11 / 37 0 / 7 / 36 / 7 455
360 6 4853.620 3.8 / 3.5 / 9 16 / 39 0 / 6 / 26 / 6 549
420 3 4813.738 5.7 / 6 / 8 8 / 31 0 / 5 / 26 / 5 504
480 6 4746.185 3.5 / 2 / 9 11 / 39 0 / 6 / 20 / 5 825

MH-noport

L_2seg (s) Sweeps Final distance Stations moved mean/med/max Accepted/solves MIP gap min/mean/max/std (%) Runtime (min)
60 3 4533.433 3.7 / 4 / 7 7 / 23 0 / 14 / 99 / 21 89
120 3 4592.349 3.0 / 2 / 7 5 / 21 0 / 11 / 26 / 7 117
180 6 4484.448 4.7 / 5.5 / 8 10 / 48 0 / 9 / 34 / 7 388
240 2 4600.363 2.5 / 2.5 / 5 3 / 16 0 / 7 / 17 / 6 171
300 4 4437.978 4.5 / 5 / 8 8 / 31 0 / 8 / 62 / 12 312
360 5 4500.298 3.8 / 3 / 8 8 / 33 0 / 7 / 19 / 7 396
420 2 4604.483 2.5 / 2.5 / 5 3 / 16 0 / 6 / 14 / 5 285
480 10 4390.774 3.5 / 3 / 7 10 / 64 0 / 8 / 30 / 6 1120

Best observed plan

Across all runs, the best final plan improves the baseline distance from 4762.172 nm to 4390.774 nm, a reduction of about 270 nm or 5.8%.

At realistic steaming speeds of 8 to 12 knots, that corresponds to about 23 to 34 hours saved, roughly one day at sea.