Counterfactual Regret Minimization

CFR Poker Bot · play the equilibrium

CFR and CFR+ solvers for Kuhn and Leduc poker, built from the original papers and checked against ground truth. I solved the strategies offline and exported them, so the bot here plays a near-Nash equilibrium. You can't beat it over the long run. Try anyway.

GitHub repo ↗ Play the bot ↓ How it's validated ↓

Play the near-Nash bot

You swap seats every hand to keep it fair. The bot never adapts; it plays the one solved strategy.

Bot
Board
You

Scoreboard

Hands0
Your net0.00
Chips / hand—
Your ceiling—

Validated against ground truth

Kuhn poker has a closed-form solution, so a correct solver must reproduce it.

I checked the solver three ways. It reproduces Kuhn's exact game value, −1/18. The recovered opening strategy lands in the analytic α-family: bet the King three times as often as the Jack, and never the Queen. A best-response evaluator measures exploitability, how much the sharpest counter-strategy could win, and drives it toward zero. CFR+ hits the same accuracy in about a tenth of vanilla CFR's iterations.

Kuhn poker convergence: CFR vs CFR+
Kuhn: exploitability vs iterations, CFR vs CFR+ (log–log).
Leduc Hold'em convergence: CFR vs CFR+
Leduc: 288 information sets. CFR+ needs far fewer iterations to match CFR.