# Experiments behind the scenario-sampling section

These scripts produced the tables in `docs/optimization_model.pdf`, Section "Scenario sets:
sampling error, variance reduction and large sets". They run against the local data store
(`quantcore setup` first) and print their tables; they are not part of the package or the tests.

| script | what it measures |
|---|---|
| `scenario_noise.py N S SEEDS [cont\|int] [kinds] [beta]` | GBM model: plain MC vs antithetic / moment matching / LHS / Sobol' at equal `S`, continuous (weights) or integer problem, against the exact Gaussian optimum (cutting planes) — Table "Sampling noise" |
| `empirical_noise.py N S SEEDS BETA [kinds]` | empirical-innovation model: bootstrap vs moment-matched / balanced / rank-stratified resampling, evaluated on a 500 000-draw set |
| `is_noise.py N SIZES SEEDS BETAS` | importance sampling (two-stage tilt, defensive mixture) vs moment-matched Sobol' — Table "Importance sampling" |
| `scaling.py N SIZES highs\|cuopt` | solve time versus the number of scenarios (HiGHS locally / cuOpt on the Spark) — Table "Solve times" |
| `rowgen.py N SIZES [full]` | tail-row generation versus the full MILP |
| `gurobi_ref.py N BETA` | the exact Gaussian MISOCP solved directly with Gurobi (needs `gurobipy` and a licence large enough for `N`) — validation of the cutting-plane reference |

Example: `python scripts/experiments/scenario_noise.py 30 2048 8 cont mc,antithetic,mc+mm,sobol,sobol+mm 0.95`