regretless

Poker bots built from scratch using Counterfactual Regret Minimization (CFR), the algorithm family behind the AI systems that beat top human professionals at heads-up poker (Libratus, Pluribus).

Two solvers per game: vanilla CFR (full game-tree traversal every iteration) and external-sampling MCCFR (samples opponent/chance actions instead of fully exploring them, trading per-iteration accuracy for far cheaper iterations). See BENCHMARKS.md for a head-to-head comparison.

Two games implemented:

• Kuhn Poker (3 cards, 1 betting round) -- has a known, published closed-form Nash equilibrium value, used here to prove the implementation is actually correct. The trained strategy converges to a game value of -0.0555, matching the theoretical -1/18 to within 0.0001.
• Leduc Hold'em (6 cards, 2 betting rounds with raises, a community card) -- the standard benchmark game used in CFR research papers. No closed-form answer exists, so correctness is verified by independent brute-force re-evaluation of the converged strategy against the training-tracked value.

Play against it

Open play.html in any browser. Pick Kuhn or Leduc from the tabs, play hands against the trained bot, and watch its mixed-strategy probabilities at each decision point.

Run the solvers

python3 kuhn_cfr.py
python3 kuhn_mccfr.py
python3 leduc_cfr.py
python3 leduc_mccfr.py

Run the tests

python3 test_kuhn_cfr.py
python3 test_kuhn_mccfr.py
python3 test_leduc_game.py
python3 test_leduc_cfr.py
python3 test_leduc_mccfr.py

Files

• kuhn_cfr.py / leduc_cfr.py -- the CFR solvers
• leduc_game.py -- Leduc Hold'em rules (betting, raises, hand evaluation)
• test_*.py -- verification suites
• play.html -- interactive browser demo with trained strategies embedded
• strategies.json -- exported trained strategies (source data for the demo)