Skip to content

Comparison Benchmarks — vs. Exact Solvers

How close does max-div get to provably optimal solutions, and what does proving optimality cost? This page anchors the heuristic quality numbers to exact references (CP-SAT and SCIP), on the built-in benchmark problems.

Protocol

  • max-div runs the same wall-clock budget ladder as the heuristics comparison (2× steps from 1 ms, SMART preset, 3 seeds).
  • Exact max-min reference: CP-SAT via threshold binary search over the distinct pairwise distances (the max-min optimum is always one of them), one worker for determinism, 120 s cap. Gaps are only reported where optimality was certified.
  • Exact mean/geomean reference: the nearest-neighbor assignment model (Lei–Church style), on both SCIP and CP-SAT backends. As shown below, no backend certifies it at the generators' sizes, so no mean/geomean gap-to-optimum is published — an incumbent-at-budget comparison is shown instead, clearly labeled as uncertified.
  • Hardware: 16" MacBook Pro with M3-class CPU, single sequential run.
  • Reproduce with uv run --group benchmarks python -m benchmarks.tier1.full (results), then ... -m benchmarks.tier1.report (tables).

Max-min: gap to the proven optimum

CP-SAT certifies the max-min optimum up to n = 300 within the cap (the proof cliff sits around n = 400; constraints help — they prune the conflict graph). max-div's gap to the certified optimum, mean over 3 seeds:

problem n k m optimum proof time gap @0.016s gap @0.128s gap @1.024s gap @16.384s
U1 100 10 0 0.0997 0.24 s 5.7% 5.7% 5.6% 3.1%
U1 200 20 0 0.2306 0.32 s 10.7% 9.6% 7.6% 7.0%
U1 300 30 0 0.3549 10.46 s 15.1% 11.3% 10.2% 6.5%
C1 100 10 2 0.8815 0.12 s 11.0% 6.5% 0.0% 0.0%
C1 200 20 4 0.5127 0.46 s 11.8% 6.7% 6.7% 3.6%
C1 300 30 6 0.3970 1.34 s 17.1% 9.1% 9.1% 6.7%

Reading: max-div lands within roughly 10 % of the proven optimum in ~100 ms and ~3–8 % at the ladder top, occasionally hitting the optimum exactly — while the exact solver needs up to seconds-to-minutes to certify, and stops certifying entirely a factor ~2 above these sizes. Note max-min is the harshest gap measure: the objective is the single worst pair, so one suboptimal swap shows in full.

Mean/geomean: why there is no gap-to-optimum

The canonical exact formulation for mean/geomean-of-NN separation is the NN-assignment model. Its proofs stop far below the generators' smallest problem (n = 100), whichever backend runs it — time to certified optimality on a d=4 random family (k = n/10, geomean; each backend stops at its first timeout):

n k SCIP (1 thread) CP-SAT (1 worker) CP-SAT (8 workers)
40 4 36.9 s 1.2 s 0.8 s
50 5 66.3 s 5.8 s 2.2 s
60 6 38.2 s 124.5 s 10.5 s
70 7 176.2 s 3499.2 s 38.4 s
80 8 317.7 s timeout 159.4 s
90 9 timeout 886.3 s
100 10 timeout

CP-SAT with 8 parallel workers is the strongest backend and still times out at n = 100 (cap 1 h; the trend suggests ~2 h would be needed, and n = 110+ is out of reach). The model's closest-assignment constraints relax weakly, so cost grows ~4–5× per +10 items — this is a property of the formulation, not of any solver. Hence: no certified mean/geomean optimum exists at any published problem size.

Geomean: incumbent-at-budget comparison (uncertified)

What a practitioner can still ask: "what if I just gave a MIP/CP solver a big time budget and took its best solution?" Below, CP-SAT (8 workers) runs the assignment model at a generous cap on two shipped problems, and its incumbent — best solution found, no optimality certificate; the bound gap column shows how far the proof remained — is compared against max-div's 1-second ladder rung:

problem n k m CP-SAT cap CP-SAT incumbent bound gap max-div @~1 s (best seed)
U3 100 10 0 10800 s 0.9915 338% 0.9847
C4 150 10 3 900 s 0.4612 543% 0.4612

On the unconstrained problem (U3), max-div reaches within ~1 % of the exact solver's 3-hour incumbent in one second. On the constrained problem (C4), max-div's 1-second solution matches the one the exact solver reached in 15 minutes (neither certified optimal). These are point comparisons on two problems, not certified gaps — but they answer the budget-parity question the certified experiment cannot reach: given comparable or far greater time, the exact solver does not pull ahead on these instances.