A study of Counterfactual Regret Minimization in multi-agent environments, from zero-sum to general-sum games.
This project investigates CFR behavior across different game types, focusing on scalability, social welfare, and equilibrium selection in multiplayer settings.
Six sequential experiments testing CFR and CFR+ variants:
- Kuhn Poker (2-player and 3-player)
- Leduc Poker (scalability analysis)
- General-sum games (Goofspiel, Prisoner's Dilemma)
- Coordination games (equilibrium selection)
- Cross-game convergence metrics analysis
CFR successfully converges in competitive zero-sum games but exhibits limitations in:
- Cooperation dilemmas (finds selfish equilibria)
- Coordination problems (fails to break symmetry)
- Large state spaces (requires abstraction)
Advanced metrics (regret, entropy, stability) reveal that CFR+ outperforms Vanilla CFR by 20-40% in complex games.
pip install open_spiel numpy matplotlib jupyter
jupyter notebook- Brown & Sandholm (2019) - Pluribus
- Zinkevich et al. (2007) - CFR Algorithm
Course: MIT 6.S890 Topics in Multiagent Learning