code-thing | status |
---|---|
master | |
develop | |
maintainability | |
coverage | |
license |
This repository will contain a best effort open source implementation of a poker AI using the ideas of Counterfactual Regret.
Made with love from the developers Leon and Colin.
A special thank you to worldveil for originally writing this awesome hand evaluator python2 module, which was ported to python3 and maintained here.
This repository assumes Python 3.7 or newer is used.
Either install from pypi:
pip install poker_ai
Or if you want to dev on our code, install the Python package from source by cloning this repo and pip -e
installing it:
git clone https://github.com/fedden/poker_ai.git # Though really we should use ssh here!
cd /path/to/poker_ai
pip install .
We have a CLI that will be installed when you pip install the package. To get help on any option, just add the --help
flag when invoking the CLI.
How to get a list of commands that can be run:
poker_ai --help
You will need to produce some lookup tables that cluster the various information sets. Here is more information on that:
poker_ai cluster --help
How to get information on training an agent:
poker_ai train start --help
How to get information on resuming training:
poker_ai train resume --help
Once you have an agent, and want to play against it, you can do the following:
poker_ai play --help
In poker, the number of card combinations for one player on the river can exceed 56 billion combinations. In order to make this information tractable, we must group together strategically similar situations. We do this with two types of compression: lossy and lossless compression. Currently we only support a 20 card deck without modification.
poker_ai cluster
You'll save the combinations of public information in a file called card_info_lut.joblib located in your project directory.
We use MCCFR to learn strategies. The MCCFR algorithm uses iterative self-play to adjust strategy based on regret.
poker_ai train start
You'll create a folder in your project directory with the learned strategy and configuration files, in case you need to resume later.
Finally, you can play your bot with the following command:
poker_ai play
You'll create a results.yaml file in ~/.poker/. So be sure to see how you stack up against your bot.
We are working hard on testing our components, but contributions here are always welcome. You can run the tests by cloning the code, changing directory to this repositories root directory (i.e poker_ai/
) and call the python test library pytest
:
cd /path/to/poker_ai
pip install pytest
pytest
See below on how to run the tests from the docker image.
We use a custom docker image for our testing suite.
You'll need to have computed the pickled card information lookup tables first (the cluster command for poker_ai). We build the images like below, in this case the luts are in './research/blueprint_algo'. First we build the parent image, with all of the dependancies.
docker build --build-arg LUT_DIR=research/blueprint_algo -f ParentDockerfile -t pokerai .
Then we build the test image.
docker build -t pokeraitest .
We then can run the tests with:
docker run -it pokeraitest pytest
This is just a note for the developers, but we can push the parent image to the registry with the following (please ensure the version tag that comes after the colon is correct). We want to do this because we need various dependancies for the remote tests, and travis builds the pokeraitest
image with the current git commit that has just been pushed.
docker tag pokerai pokerai/pokerai:1.0.0rc1
docker push pokerai/pokerai:1.0.0rc1
Documentation is hosted, but you can build it yourself if you wish:
# Build the documentation.
cd /path/to/poker_ai/docs
make html
cd ./_build/html
# Run a webserver and navigate to localhost and the port (usually 8000) in your browser.
python -m http.server
Below is a rough structure of the codebase.
├── applications # Larger applications like the state visualiser sever.
├── paper # Main source of info and documentation :)
├── poker_ai # Main Python library.
│ ├── ai # Stub functions for ai algorithms.
│ ├── games # Implementations of poker games as node based objects that
│ │ # can be traversed in a depth-first recursive manner.
│ ├── poker # WIP general code for managing a hand of poker.
│ ├── terminal # Code to play against the AI from your console.
│ └── utils # Utility code like seed setting.
├── research # A directory for research/development scripts
│ # to help formulate understanding and ideas.
└── test # Python tests.
├── functional # Functional tests that test multiple components
│ # together.
└── unit # Individual tests for functions and objects.
Here are some assorted examples of things that are being built in this repo.
To perform MCCFR, the core algorithm of poker_ai, we need a class that encodes all of the poker rules, that we can apply an action to which then creates a new game state.
pot = Pot()
players = [
ShortDeckPokerPlayer(player_i=player_i, initial_chips=10000, pot=pot)
for player_i in range(n_players)
]
state = ShortDeckPokerState(players=players)
for action in state.legal_actions:
new_state: ShortDeckPokerState = state.apply_action(action)
We also have some code to play a round of poker against the AI agents, inside your terminal.
The characters are a little broken when captured in asciinema
, but you'll get the idea by watching this video below. Results should be better in your actual terminal!
cd /path/to/poker_ai/dir
python -m poker_ai.terminal.runner \
--agent offline \
--pickle_dir ./research/blueprint_algo \
--strategy_path ./research/blueprint_algo/offline_strategy_285800.gz
We are also working on code to visualise a given instance of the ShortDeckPokerState
, which looks like this:
It is so we can visualise the AI as it plays, and also debug particular situations visually. The idea as it stands, is a live web-visualisation server like TensorBoard, so you'll just push your current poker game state, and this will be reflected in the visualisations, so you can see what the agents are doing.
The frontend code is based on this codepen.
Here is an example of how you could plot the poker game state:
from plot import PokerPlot
from poker_ai.games.short_deck.player import ShortDeckPokerPlayer
from poker_ai.games.short_deck.state import ShortDeckPokerState
from poker_ai.poker.pot import Pot
def get_state() -> ShortDeckPokerState:
"""Gets a state to visualise"""
n_players = 6
pot = Pot()
players = [
ShortDeckPokerPlayer(player_i=player_i, initial_chips=10000, pot=pot)
for player_i in range(n_players)
]
return ShortDeckPokerState(
players=players,
pickle_dir="../../research/blueprint_algo/"
)
pp: PokerPlot = PokerPlot()
# If you visit http://localhost:5000/ now you will see an empty table.
# ... later on in the code, as proxy for some code that obtains a new state ...
# Obtain a new state.
state: ShortDeckPokerState = get_state()
# Update the state to be plotted, this is sent via websockets to the frontend.
pp.update_state(state)
# http://localhost:5000/ will now display a table with 6 players.
There are two parts to this repository, the code to manage a game of poker, and the code to train an AI algorithm to play the game of poker. A low level thing to first to is to implement a poker engine class that can manage a game of poker.
The reason the poker engine is implemented is because it is useful to have a well-integrated poker environment available during the development of the AI algorithm, incase there are tweaks that must be made to accomadate things like the history of state or the replay of a scenario during Monte Carlo Counterfactual Regret Minimisation.
The following code is how one might program a round of poker that is deterministic using the engine. This engine is now the first pass that will be used support self play.
from poker_ai import utils
from poker_ai.ai.dummy import RandomPlayer
from poker_ai.poker.table import PokerTable
from poker_ai.poker.engine import PokerEngine
from poker_ai.poker.pot import Pot
# Seed so things are deterministic.
utils.random.seed(42)
# Some settings for the amount of chips.
initial_chips_amount = 10000
small_blind_amount = 50
big_blind_amount = 100
# Create the pot.
pot = Pot()
# Instanciate six players that will make random moves, make sure
# they can reference the pot so they can add chips to it.
players = [
RandomPlayer(
name=f'player {player_i}',
initial_chips=initial_chips_amount,
pot=pot)
for player_i in range(6)
]
# Create the table with the players on it.
table = PokerTable(players=players, pot=pot)
# Create the engine that will manage the poker game lifecycle.
engine = PokerEngine(
table=table,
small_blind=small_blind_amount,
big_blind=big_blind_amount)
# Play a round of Texas Hold'em Poker!
engine.play_one_round()
The following todo will change dynamically as my understanding of the algorithms and the poker_ai project evolves.
At first, the goal is to prototype in Python as iteration will be much easier and quicker. Once there is a working prototype, write in a systems level language like C++ and optimise for performance.
Implement a multiplayer working heads up no limit poker game engine to support the self-play.
- Lay down the foundation of game objects (player, card etc).
- Add poker hand evaluation code to the engine.
- Support a player going all in during betting.
- Support a player going all in during payouts.
- Lots of testing for various scenarios to ensure logic is working as expected.
Iterate on the AI algorithms and the integration into the poker engine.
- Integrate the AI strategy to support self-play in the multiplayer poker game engine.
- In the game-engine, allow the replay of any round the current hand to support MCCFR.
- Implement the creation of the blueprint strategy using Monte Carlo CFR miminisation.
- Add the real-time search for better strategies during the game.
Strengthen the game engine with more tests and allow users to see live visualisation of game state.
- Start work on a visualisation server to allow a game state to be displayed.
- Triple check that the rules are implemented in the poker engine as described in the supplimentary material.
- Work through the coverage, adding more tests, can never have enough.
This is an open effort and help, criticisms and ideas are all welcome.
First of all, please check out the CONTRIBUTING guide.
Feel free to start a discussion on the github issues or to reach out to me at leonfedden at gmail dot com.
The code is provided under the copy-left GPL licence. If you need it under a more permissive license then please contact me at leonfedden at gmail dot com.
We appreciate you getting this far in the README file! If you like what we are doing, please give us a star and share with your friends!