This poker AI runs on Python, self-play, and mild regret

What it does

poker_ai is a Python research toolkit for training Texas Hold’em agents via Monte Carlo Counterfactual Regret Minimisation. It bundles a deterministic poker engine, state-based game traversal, and clustering logic to compress the vast card-combination space into something tractable. Once trained, you can face the agent in a terminal interface or inspect its decisions through a live web visualisation server.

The interesting bit

The project treats the game as a recursive state machine where every legal action spawns a new ShortDeckPokerState, making self-play and regret calculation straightforward to implement. Because a single player’s river combinations can exceed 56 billion, the authors use lossy and lossless clustering to shrink information sets—though, for now, this is hard-wired to a 20-card deck.

Key highlights

• MCCFR training via iterative self-play that adjusts strategy based on accumulated regret.
• A deterministic poker engine (PokerEngine, PokerTable, Pot) for managing rounds and replaying scenarios.
• A terminal runner to play against trained agents and a web-based PokerPlot visualiser for debugging game states.
• A CLI that handles clustering lookup tables, training sessions, and resuming previous runs.
• Explicitly framed as a Python prototype, with a roadmap pointing toward a future C++ rewrite for performance.

Caveats

• Only a 20-card deck is currently supported without modification.
• The repo structure describes the ai module as “stub functions” and the general poker code as a work in progress.
• The web visualisation and several other components are still under active development.

Verdict

Researchers and hobbyists interested in imperfect-information game theory will find a useful sandbox here; anyone expecting a production-grade, full-deck poker bot ready to clean up online should look elsewhere.