What it does

Neuron Poker is an OpenAI Gym-compatible environment for no-limit Texas Hold’em. You instantiate a table, add agents (random, keyboard-controlled, equity-based, or deep Q-learning), and let them play. The environment handles the full game loop, renders a virtual table, and exposes the complete game state as a structured observation dictionary.

The interesting bit

The project ships three equity calculators with a 500× performance spread between them: a slow Python Monte Carlo, a broken NumPy variant (tests failing, help wanted), and a C++ implementation that requires Visual Studio 2019. It’s a rare honest admission of partial failure in a README — the NumPy version “is not yet working correctly” and contributors are explicitly invited to fix it.

Key highlights

• Deep Q-learning agent built on keras-rl with TensorBoard logging and configurable experience replay
• Observation space encodes full table state: community cards, betting history per stage, player positions, stack sizes, and precomputed Monte Carlo equity against 2- and 3-player ranges
• Genetic algorithm example that self-adjusts equity thresholds based on winning performance
• Command-line entry points via main.py using docopt; modern Python packaging with uv
• Explicit call for collaborative model contributions with detailed PyCharm-centric contribution workflow

Caveats

• The NumPy Monte Carlo equity calculator has failing tests and is effectively non-functional
• keras-rl dependency suggests legacy Keras; no mention of migration to tf-agents or modern equivalents
• Windows-centric C++ toolchain (Visual Studio 2019) with only a passing note about Cygwin/GCC

Verdict

Worth a look if you’re building poker AI and want a ready-made Gym environment with equity-aware observations. Skip it if you need production-grade Monte Carlo simulation or modern RL frameworks out of the box — you’ll be fixing someone else’s NumPy bugs or wrestling with decade-old dependencies.