Data and code for "Dynamical Path Dependence in Multi-Agent Learning"
Description
Simulation scripts, verfied output data, and figures for the paper submitted to Physica D: Nonlinear Phenomena
Files
Steps to reproduce
Requirements: Python 3.8+, numpy, scipy, scikit-learn, matplotlib (see requirements.txt). pip install -r requirements.txt All scripts are deterministic given their seeds. Run from the src/ directory: # Main experiment: value-based learners (produces Table 1 value-based block + Figure 1) python value_based_learners.py # Transfer study: Q-learning, REINFORCE, MWU (produces Table 1 policy-based block) python transfer_fast.py # Section 7: within-game predictor, leakage control, Minority Game python attractor_selection_experiments.py # ~6-9 min # Payoff-scale sweep (Table 2 + Figure 2) python scale_sweep_50.py # Lemma 1 numerical check python verify_lemma.py # Disjoint-seed robustness re-run python vb_robust.py # Regenerate figures python make_fig_repr.py # Figure 1 python make_scale_fig.py # Figure 2 python make_fig4.py # Figure 4 Expected runtimes on a single CPU core: value_based_learners.py ~15 s, transfer_fast.py ~15 s, attractor_selection_experiments.py ~6–9 min, everything else < 2 min. Every number in the paper can be cross-checked against the corresponding file in data/. The results/results_summary.md has the full cross-reference table.