PokerML
Updated Nov 2025
PythonFastAPIphevaluatorCFR+
this is a heads-up no-limit hold'em solver built on CFR+ and external sampling monte carlo. it trains over 100K iterations with k-means preflop bucketing and percentile-based postflop abstraction, then serves the converged strategy through a fastapi web app where you can play against the bot and see its 13x13 range grid in real time.
CFR+ with external sampling: 100K iterations, regret clamping, bucket-cached traversal across all 4 streets
k-means preflop abstraction (10 buckets) + empirical CDF postflop bucketing with sampled board completions
fastapi web UI where you play against the bot with a live 13x13 range heatmap showing action probabilities per hand
what i built
- CFR+ trainer with external sampling MCCFR, immutable game states, and per-iteration bucket caching for speed
- two-level card abstraction: k-means on preflop equity (169 hands to 10 buckets) and percentile CDF on postflop hand strength
- action abstraction with 50% and 100% pot bet sizings, fold/check/call, all-in, and a 4-raise-per-street cap
- fastapi server with session management, bot inference, and a 13x13 range grid API for the frontend
how it works
- 1deal hands and board, pre-cache card buckets for both players across all 4 streets
- 2traverse the game tree: enumerate all actions for the traversing player, sample one for the opponent
- 3update cumulative regrets (clamped to zero for CFR+) and accumulate strategy sums
- 4after 100K iterations, compress the averaged strategy to JSON and serve it via fastapi for live play
results
- ✓converged strategy across ~1,173 lines of python with an 18.5 MB compressed strategy file
- ✓interactive web UI where you can play full hands and inspect the bot's action distribution for any of 169 canonical hands
what's next
- add variance reduction techniques like pruning and warm starting
- expand action abstraction with more bet sizings and raise depths