Automated Negotiation Experiments (NegMAS)

This repository implements bilateral and multilateral SAO (Stacked Alternating Offers) negotiation experiments in NegMAS. It compares classic time-based concession strategies with a custom opponent-aware aspiration strategy under deadline pressure, reservation values, asymmetric utility profiles, and both two-party and one-to-many procurement settings.

The project is designed for reproducible benchmarking: it includes fixed-domain experiments, a multilateral extension, a randomized appendix tournament, and statistical analysis outputs (bootstrap confidence intervals, permutation tests, FDR correction, and effect sizes).

Project Parts

• src/run_experiments.py: bilateral fixed-domain runner (A..F)
• src/run_multilateral_experiments.py: multilateral procurement runner (M1..M3)
• src/appendix_tournament.py: randomized robustness tournament
• src/domain.py, src/domain_multilateral.py: issue spaces and utility-profile generators
• src/negotiators.py: strategy registry, custom SmartAspirationNegotiator, and GeniusBOABaselineNegotiator
• src/metrics.py: frontier/reference metrics, welfare regret, CIs, and pairwise significance tests
• scripts/run_all.py: end-to-end sequential pipeline
• scripts/plot_stats.py: bootstrap/effect-size figures from bilateral summaries
• scripts/plot_analysis.py: presentation analysis figures from run CSVs
• tests/: unit and integration checks

Strategy Set

• BoulwareTBNegotiator
• LinearTBNegotiator
• ConcederTBNegotiator
• SmartAspirationNegotiator
• GeniusBOABaselineNegotiator

Profile Families

Bilateral (A..F) in src/domain.py

Issues: price, quantity, delivery_time, service_level.

• A: balanced baseline buyer/seller preferences
• B: seller is highly delivery-time sensitive
• C: nonlinear price-vs-quantity tension
• D: service-level dominant weighting
• E: delivery-pressure and price asymmetry
• F: strong specialization (buyer quantity vs seller price emphasis)

Multilateral (M1..M3) in src/domain_multilateral.py

Issues: unit_price, volume, delivery_window, quality_tier with one buyer and multiple suppliers.

• M1: balanced procurement setting
• M2: buyer emphasizes speed and quality
• M3: buyer is strongly cost-focused

Environment Setup

From repo root:

python3.11 -m venv .venv311
source .venv311/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Run Everything

python scripts/run_all.py

This pipeline runs smoke test, pytest, bilateral experiments, stats plots, appendix tournament, multilateral experiments, and analysis plots sequentially, then writes a run bundle with SUMMARY.md and manifest.json.

Quick verification mode:

python scripts/run_all.py --quick

Run Components Manually

Smoke test

python scripts/smoke_test.py

Tests

pytest -q

Bilateral main experiments

python -m src.run_experiments \
  --output-dir results/main_final \
  --n-seeds 40 \
  --profiles A B C D E F \
  --n-steps 12 24 36 \
  --buyer-reserved-value 0.35 \
  --seller-reserved-value 0.35 \
  --pend 0.04 \
  --pend-per-second 0.02 \
  --bootstrap-resamples 2000 \
  --permutation-resamples 2000 \
  --progress-every 50

Appendix tournament

python -m src.appendix_tournament \
  --output-dir results/appendix \
  --n-scenarios 6 \
  --n-repetitions 2 \
  --n-steps 50 \
  --seed 2026

Multilateral experiments

python -m src.run_multilateral_experiments \
  --output-dir results/multilateral \
  --n-seeds 20 \
  --profiles M1 M2 M3 \
  --n-steps 20 40 \
  --n-suppliers 2 \
  --buyer-reserved-value 0.35 \
  --supplier-reserved-value 0.35 \
  --pend 0.04 \
  --pend-per-second 0.02 \
  --bootstrap-resamples 1000 \
  --progress-every 50

Plotting from existing results

python scripts/plot_stats.py \
  --results-dir <RUN_DIR>/main \
  --output-dir <RUN_DIR>/figures/stats

python scripts/plot_analysis.py \
  --results-dir <RUN_DIR>/main \
  --output-dir <RUN_DIR>/figures/analysis \
  --multilateral-runs <RUN_DIR>/multilateral/multi_runs.csv

Output Layout

Default sequential bundles are written to:

results/presentation/run_YYYYMMDD_HHMMSS/

Each run bundle contains:

• main/: bilateral raw + aggregated outputs
• multilateral/: multilateral raw + aggregated outputs
• appendix/: tournament outputs
• figures/stats/: CI/effect-size figures
• figures/analysis/: analysis/presentation figures
• logs/: per-stage execution logs
• SUMMARY.md: concise results summary
• manifest.json: exact commands, timing, and run metadata

results/presentation/LATEST.txt points to the latest run bundle.

Current bundled results in this repo are under:

• results/presentation/run_full_presentable/