Skip to content

Latest commit

 

History

History
102 lines (62 loc) · 3.62 KB

readme.md

File metadata and controls

102 lines (62 loc) · 3.62 KB
Raw

Alpha NL Holdem

logo

This is an implementation of a self-play non-limit texas holdem ai, using TensorFlow and ray. While heavily inspired by UCAS's work of Alpha Holdem, it's not a offical implementation of Alpha Holdem.

This is a proof of concept project, rlcard's nl-holdem env was used. It's a 50bb 1v1 env, not the standard 100bb ACPC one, and bet sizes are slightly different than ACPC.

I asked a few pro holdem players to play against this ai for some dozen games. They report that the ai's moves are all logical. They did not observe any significant mistakes.

Goal

  1. Provide a clean codebase for apply self-play model-free RL method in Holdem-like games.
  2. Try to reproduce the result of the AlphaHoldem.
  3. Provide All data, including checkpoints, training methods, evaluation metrics and more.

Getting Started

This project assumes you have the following:

  1. Conda environment (Anaconda /Miniconda)
  2. Python 3.7+

Install dependences:

pip3 install -r requirements.txt

Usage

Play against Neural Net

First go to gui directory, and run the python script:

cd gui
python3 play_against_ai_in_ui.py

And go to http://localhost:8000/ to play against the NeuralNet.

Yes, it's a small tool I write to play against AI, Yes, it looks bad. But it works.

By default you are playing against an NN opponent which has been trained for about a week.

play_against_ai

Trainning

Training 101

To start the training, you have to first change the config in confs/nl_holdem.py

By default it would require 1 GPU and 89 cpu to run this program.

Modify the line in the config file:

    'num_workers': 89,

change it to the cpu core that your machine have. However you still need at least one gpu to run this the training.

And then use command:

python3 train_league.py --conf confs/nl_holdem.py --sp 0.0 --upwin 1.0 --gap=500 --league_tracker_n 1000

Winrate against history agents will be displayed in the stdout log.

training

Restore training

If the training process is somehow killed or you want to start from the weights you downloaded, First put the downloaded league folder in this project's root. Then use command :

python3 train_league.py --conf confs/nl_holdem.py --sp 0.0 --upwin 1.0 --gap=500 --league_tracker_n 1000 --restore league/history_agents

It would auto load all training weights and continue training.

Released data

  1. Weights of all checkpoints in the process of a week of training, ~ 1 billion of selfplay games: Google Drive: https://drive.google.com/file/d/1G_GwTaVe4syCwW43DauwSQi6FqjRS3nj/view?usp=sharing Baidu Drive: https://pan.baidu.com/s/1PYNLKN2CExRntVvkvYyKkA?pwd=7jmn
  2. Evaluation metrics and part of results: see nl-evaluation.ipynb

Known Issues

  1. Rlcard environment sucks, 50bb pot, wrong pot sizes, wrong action order after flop, I don't know where to start. But it's the only environment I konw out there suitable for this task.
  2. Even after ~ 1 billion self-play, over 1000 checkpoints, the model seems still not converge, still improving itself, I really don't know when it will converge. It could be some bug, not sure.

License

GNU AGPL v3

Warning

It's illegal to use this code in any way to commercial purpose, including researching project inside a commercial entity.

Especially for Chinese company JJ world(竞技世界). You'd better look elseware.