PokeringDumDum 🃏

A not-so-dumb poker game-theory lab implementing CFR and CFR+ for Kuhn poker, with exact best responses and exploitability measurement.

Why this project exists

Poker is a clean setting for studying decision-making under hidden information. PokeringDumDum focuses on a result that can be verified: does the learned average strategy approach a Nash equilibrium, and how quickly?

This is intentionally not a graphical poker game and not a claim of real-money profitability. The research contribution is the solver, exact evaluation, tests, and convergence study.

What it includes

• explicit Kuhn poker rules and terminal utilities;
• full-tree vanilla CFR and CFR+;
• reach-weighted average strategies;
• exact expected-value and best-response enumeration;
• exploitability and NashConv metrics;
• deterministic tests and a reproducible convergence experiment.

Run it

python -m pip install -e ".[dev]"
python -m pytest -q
python run_experiment.py

The experiment writes data/convergence.csv, data/final_strategy.csv, and docs/assets/convergence.png. Generated tables are ignored because they can be recreated; the figure is versioned so the result is visible on GitHub.

Interpreting the result

Kuhn poker's equilibrium value for player zero is -1/18 chips per hand. A sound implementation should approach that value while exploitability trends toward zero. Do not judge the solver from one displayed strategy probability: Kuhn poker has a family of equilibria, so exact policy mixtures may differ while remaining sound.

Read docs/METHODOLOGY.md for the game rules, regret update, exact best-response method, metric definitions, and limitations.