RLCard: A Toolkit for Reinforcement Learning in Card Games

RLCard is a toolkit for Reinforcement Learning (RL) in card games. It supports multiple card environments with easy-to-use interfaces for implementing various reinforcement learning and searching algorithms. The goal of RLCard is to bridge reinforcement learning and imperfect information games. RLCard is developed by DATA Lab at Texas A&M University and community contributors.

Community:

  • Slack: Discuss in our #rlcard-project slack channel.

  • QQ Group: Join our QQ group 665647450. Password: rlcardqqgroup

Installation

Make sure that you have Python 3.6+ and pip installed. We recommend installing the stable version of rlcard with pip:

pip3 install rlcard

If you are in China and the above command is too slow, you can use the mirror provided by Tsinghua University:

pip3 install rlcard -i https://pypi.tuna.tsinghua.edu.cn/simple

The default installation will only include the card environments. To use PyTorch implementation of the training algorithms, run

pip3 install rlcard[torch]

Alternatively, you can clone the latest version with (if you are in China and Github is slow, you can use the mirror in Gitee):

git clone https://github.com/datamllab/rlcard.git

or only clone one branch to make it faster:

git clone -b master --single-branch --depth=1 https://github.com/datamllab/rlcard.git

Then install with

cd rlcard
pip3 install -e .
pip3 install -e .[torch]

We also provide conda installation method:

conda install -c toubun rlcard

Conda installation only provides the card environments, you need to manually install Pytorch on your demands.

Examples

A short example is as below.

import rlcard
from rlcard.agents import RandomAgent

print(env.num_actions) # 2
print(env.num_players) # 1
print(env.state_shape) # [[2]]
print(env.action_shape) # [None]

env = rlcard.make('blackjack')
env.set_agents([RandomAgent(num_actions=env.num_actions)])

trajectories, payoffs = env.run()

RLCard can be flexibly connected to various algorithms. See the following examples:

Demo

Run examples/human/leduc_holdem_human.py to play with the pre-trained Leduc Hold’em model. Leduc Hold’em is a simplified version of Texas Hold’em. Rules can be found here.

>> Leduc Hold'em pre-trained model

>> Start a new game!
>> Agent 1 chooses raise

=============== Community Card ===============
┌─────────┐
│░░░░░░░░░│
│░░░░░░░░░│
│░░░░░░░░░│
│░░░░░░░░░│
│░░░░░░░░░│
│░░░░░░░░░│
│░░░░░░░░░│
└─────────┘
===============   Your Hand    ===============
┌─────────┐
│J        │
│         │
│         │
│    ♥    │
│         │
│         │
│        J│
└─────────┘
===============     Chips      ===============
Yours:   +
Agent 1: +++
=========== Actions You Can Choose ===========
0: call, 1: raise, 2: fold

>> You choose action (integer):

We also provide a GUI for easy debugging. Please check here. Some demos:

doudizhu-replay leduc-replay

Available Environments

We provide a complexity estimation for the games on several aspects. InfoSet Number: the number of information sets; InfoSet Size: the average number of states in a single information set; Action Size: the size of the action space. Name: the name that should be passed to rlcard.make to create the game environment. We also provide the link to the documentation and the random example.

Game

InfoSet Number

InfoSet Size

Action Size

Name

Usage

Blackjack (wiki, baike)

10^3

10^1

10^0

blackjack

doc, example

Leduc Hold’em (paper)

10^2

10^2

10^0

leduc-holdem

doc, example

Limit Texas Hold’em (wiki, baike)

10^14

10^3

10^0

limit-holdem

doc, example

Dou Dizhu (wiki, baike)

10^53 ~ 10^83

10^23

10^4

doudizhu

doc, example

Mahjong (wiki, baike)

10^121

10^48

10^2

mahjong

doc, example

No-limit Texas Hold’em (wiki, baike)

10^162

10^3

10^4

no-limit-holdem

doc, example

UNO (wiki, baike)

10^163

10^10

10^1

uno

doc, example

Gin Rummy (wiki, baike)

10^52

-

-

gin-rummy

doc, example

Supported Algorithms

Algorithm

example

reference

Deep Monte-Carlo (DMC)

examples/run_dmc.py

[paper]

Deep Q-Learning (DQN)

examples/run_rl.py

[paper]

Neural Fictitious Self-Play (NFSP)

examples/run_rl.py

[paper]

Counterfactual Regret Minimization (CFR)

examples/run_cfr.py

[paper]

Pre-trained and Rule-based Models

We provide a model zoo to serve as the baselines.

Model

Explanation

leduc-holdem-cfr

Pre-trained CFR (chance sampling) model on Leduc Hold’em

leduc-holdem-rule-v1

Rule-based model for Leduc Hold’em, v1

leduc-holdem-rule-v2

Rule-based model for Leduc Hold’em, v2

uno-rule-v1

Rule-based model for UNO, v1

limit-holdem-rule-v1

Rule-based model for Limit Texas Hold’em, v1

doudizhu-rule-v1

Rule-based model for Dou Dizhu, v1

gin-rummy-novice-rule

Gin Rummy novice rule model

API Cheat Sheet

How to create an environment

You can use the the following interface to make an environment. You may optionally specify some configurations with a dictionary.

  • env = rlcard.make(env_id, config={}): Make an environment. env_id is a string of a environment; config is a dictionary that specifies some environment configurations, which are as follows.

    • seed: Default None. Set a environment local random seed for reproducing the results.

    • allow_step_back: Defualt False. True if allowing step_back function to traverse backward in the tree.

    • Game specific configurations: These fields start with game_. Currently, we only support game_num_players in Blackjack, .

Once the environemnt is made, we can access some information of the game.

  • env.num_actions: The number of actions.

  • env.num_players: The number of players.

  • env.state_shape: The shape of the state space of the observations.

  • env.action_shape: The shape of the action features (Dou Dizhu’s action can encoded as features)

What is state in RLCard

State is a Python dictionary. It consists of observation state['obs'], legal actions state['legal_actions'], raw observation state['raw_obs'] and raw legal actions state['raw_legal_actions'].

Basic interfaces

The following interfaces provide a basic usage. It is easy to use but it has assumtions on the agent. The agent must follow agent template.

  • env.set_agents(agents): agents is a list of Agent object. The length of the list should be equal to the number of the players in the game.

  • env.run(is_training=False): Run a complete game and return trajectories and payoffs. The function can be used after the set_agents is called. If is_training is True, it will use step function in the agent to play the game. If is_training is False, eval_step will be called instead.

Advanced interfaces

For advanced usage, the following interfaces allow flexible operations on the game tree. These interfaces do not make any assumtions on the agent.

  • env.reset(): Initialize a game. Return the state and the first player ID.

  • env.step(action, raw_action=False): Take one step in the environment. action can be raw action or integer; raw_action should be True if the action is raw action (string).

  • env.step_back(): Available only when allow_step_back is True. Take one step backward. This can be used for algorithms that operate on the game tree, such as CFR (chance sampling).

  • env.is_over(): Return True if the current game is over. Otherewise, return False.

  • env.get_player_id(): Return the Player ID of the current player.

  • env.get_state(player_id): Return the state that corresponds to player_id.

  • env.get_payoffs(): In the end of the game, return a list of payoffs for all the players.

  • env.get_perfect_information(): (Currently only support some of the games) Obtain the perfect information at the current state.

Library Structure

The purposes of the main modules are listed as below:

More Documents

For more documentation, please refer to the Documents for general introductions. API documents are available at our website.

Contributing

Contribution to this project is greatly appreciated! Please create an issue for feedbacks/bugs. If you want to contribute codes, please refer to Contributing Guide. If you have any questions, please contact Daochen Zha with daochen.zha@tamu.edu.

Acknowledgements

We would like to thank JJ World Network Technology Co.,LTD for the generous support and all the contributions from the community contributors.