Reproducing the DMB + TF-ACO paper

Target paper: Density-Modulated Boids and Stigmergic Coverage.

Versions pinned for the canonical results

Fill from the canonical experiment.json provenance block.

Environment

python3.11 -m venv .venv && source .venv/bin/activate
pip install -r maneuver-map/backend/requirements.txt
export ENGINE_MODE=inprocess
export PYTHONHASHSEED=0

The phase-diagram sweep

The canonical phase diagram is produced by a Sobol sweep over two axes (local density control d0 and noise sigma_v) with 16 samples per axis. Submit via POST /experiments/sweep:

{
  "name": "dmb_phase_diagram",
  "steps": 4000,
  "num_agents": 200000,
  "dt": 0.1,
  "generations": 1,
  "output_frequency": 40,
  "environment_bound": 500000.0,
  "algorithm": "dmb",
  "design": "sobol",
  "n_samples": 64,
  "ranges": {
    "algo_params.dmb.d0":              [1.0, 20.0],
    "algo_params.dmb.noise_sigma_v":   [0.0, 0.3]
  },
  "seed_base": 42
}

Criticality instrumentation

The paper’s critical-point estimates use the instruments in gossamer.metrics.criticality:

• Susceptibility and Binder cumulant from the rolling alignment samples recorded by run_generation.
• Velocity correlation length from the terminal snapshot.
• Finite-size scaling requires repeating the sweep at num_agents ∈ {50_000, 200_000, 500_000}.

Reproducible analysis notebook: notebooks/criticality.py . Paste each exp_id from the sweep response into EXP_IDS and run.

Expected wall-clock

Per-trial cost at 200k agents / 4000 steps: ~20 minutes on the 32-core reference node. A full 64-trial Sobol sweep: ~22 hours. For finite- size scaling at three sizes, budget 60–70 hours.