DispatchAtlasSolver Recommender
The DispatchAtlas solver recommender — rank solvers from benchmark conditions, declared capabilities, and evidence, with transparent inclusion reasons.
The recommender ranks solvers from benchmark conditions, declared capabilities, scale, objective type, budget assumptions, and evidence tier. Set the six selection axes below; the ranking is computed in the browser from build-inlined solver metadata and explains every fit. It never claims a globally best solver.
Recommendations rank solvers from declared capabilities and benchmark characterization; the recommender never asserts a globally best solver.
8 solvers ranked by declared fit; never a globally best claim.
adpso · metaheuristic family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
Caveat: success-rate feedback is noisy on small swarms
age-moea-ii · metaheuristic family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
Caveat: geometry estimate is noisy on tiny first fronts
ant-colony · metaheuristic family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
apparent-tardiness-cost · dispatching family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
arithmetic-optimization · metaheuristic family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
Caveat: every member is regenerated around the best vector each iteration, so the returned order is the best observed rather than a monotone incumbent
artificial-bee-colony · metaheuristic family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
Caveat: the per-component neighbor move explores slowly on long priority vectors, so many iterations may be needed at large task counts
artificial-fish-swarm · metaheuristic family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
Caveat: each behaviour re-scores candidate positions, so an iteration costs several schedule evaluations per member
beam-search · constructive family · source: metadata · confidence: metadata-only
Fit: matches the current selection on declared capabilities.
Recommendation source: metadata, learned-model. Confidence is capped at metadata-only — a metadata recommendation is never reported as a learned-model one, and no solver is asserted as globally best.
Confidence Labels
| Label | Meaning |
|---|---|
| metadata-only | Recommendation is based on declared capabilities only. |
| smoke-backed | A small deterministic run supports the interface path. |
| pilot-backed | Pilot campaign evidence has passed review. |
| evidence-grade | Full-campaign evidence has passed statistical and evidence-tier gates. |
The current recommender bundle is metadata-only.
Recommendation Sources
Every recommendation declares its source, so a metadata recommendation is never mistaken for a learned-model one:
| Source | Basis | Confidence it may carry |
|---|---|---|
| metadata | Declared solver capabilities and benchmark characterization alone; no learned model. | metadata-only |
| learned-model | A supervised model trained on a labeled corpus of past best-solver outcomes. | metadata-only until a leakage-free held-out evaluation lifts it to pilot-backed or evidence-grade |
A learned-model recommendation stays metadata-only until held-out generalization is measured under the leakage-free protocol below; it never borrows confidence from training-set fit.
Selection Features
The selector ranks against a fixed set of benchmark-characterization features. Each feature names the characterization metric it reads:
| Feature | Metric of origin |
|---|---|
| difficulty | aggregate difficulty score (contention, overload, dependency depth, solver sensitivity) |
| heterogeneity | resource-kind heterogeneity |
| objective_conflict | objective-conflict metric |
| uncertainty | uncertainty metric |
| dynamism | release-span dynamism metric |
| solver_sensitivity | solver-sensitivity metric |
Leakage-Free Cross-Validation
The supervised selector reports held-out generalization, never training fit. The labeled corpus is split so that no instance, no benchmark family, and no characterization record appears in both the training and the test partition; the selector trains on the remaining families and is scored on the held-out family. The labeled corpus is sourced from the comparative campaigns, so until those campaigns supply labels the selector reports the deferral explicitly rather than claiming unvalidated performance.
Learning And Hybrid Interfaces
The recommender ships seven named learning and hybrid interfaces. Two are implemented here — supervised algorithm selection and the benchmark-only baseline — and five (surrogate-assisted search, reinforcement-learning hooks, hyper-heuristics, policy-guided repair, and learned initialization) are registered as named interfaces whose realization is gated on the campaigns that produce training data. Heavy backends are optional extras; when an extra is absent, the interface falls back to its deterministic default rather than importing anything by default. The interface catalog, its evidence policies, and the recommendation-source taxonomy ship in the solver capabilities bundle.