https://github.com/jbradberry/mcts
Board game AI implementations using Monte Carlo Tree Search
https://github.com/jbradberry/mcts
artificial-intelligence board-game board-game-framework monte-carlo-tree-search python
Last synced: about 2 months ago
JSON representation
Board game AI implementations using Monte Carlo Tree Search
- Host: GitHub
- URL: https://github.com/jbradberry/mcts
- Owner: jbradberry
- License: mit
- Created: 2015-11-05T01:02:46.000Z (over 9 years ago)
- Default Branch: master
- Last Pushed: 2020-04-19T23:24:46.000Z (about 5 years ago)
- Last Synced: 2025-03-18T23:44:24.150Z (2 months ago)
- Topics: artificial-intelligence, board-game, board-game-framework, monte-carlo-tree-search, python
- Language: Python
- Size: 26.4 KB
- Stars: 183
- Watchers: 14
- Forks: 34
- Open Issues: 1
-
Metadata Files:
- Readme: README.rst
- License: LICENSE
Awesome Lists containing this project
README
Monte Carlo Tree Search
=======================
This is an implementation of an AI in Python using the UCT Monte Carlo
Tree Search algorithm.
The Monte Carlo Tree Search AIs included here are designed to work
with `jbradberry/boardgame-socketserver
`_ and
`jbradberry/boardgame-socketplayer
`_.
Requirements
------------
* Python 2.7, 3.5+; PyPy; PyPy3
* six
Getting Started
---------------
To set up your local environment you should create a virtualenv and
install everything into it. ::
$ mkvirtualenv mcts
Pip install this repo, either from a local copy, ::
$ pip install -e mcts
or from github, ::
$ pip install git+https://github.com/jbradberry/mcts#egg=mcts
Additionally, you will need to have `jbradberry/boardgame-socketplayer
`_ installed in
order to make use of the players.
This project currently comes with two different Monte Carlo Tree
Search players. The first, ``jrb.mcts.uct``, uses the count of the
number of wins for a node to make its decisions. The second,
``jrb.mcts.uctv`` instead keeps track of the evaluated value of the
board for the playouts from a given node ::
$ board-play.py t3 jrb.mcts.uct # number of wins metric
$ board-play.py t3 jrb.mcts.uctv # point value of the board metric
These AI players can also take additional arguments:
time (default: 30)
The amount of thinking time allowed for the AI to make its decision,
in seconds. Ex: ``$ board-play.py t3 jrb.mcts.uct -e time=5``
max_actions (default: 1000)
The maximum number of actions, or plays, to allow in one of the
simulated playouts before giving up. Ex: ``$ board-play.py t3
jrb.mcts.uct -e max_actions=500``
C (default: 1.4)
The exploration vs. exploitation coefficient at the heart of the UCT
algorithm. Larger values prioritize exploring inadequately covered
actions from a node, smaller values prioritize exploiting known
higher valued actions. Experimentation with this variable to find
reasonable values for a given game is recommended. Ex: ``$
board-play.py t3 jrb.mcts.uct -e C=3.5``
The ``-e`` flag may be used multiple times to set additional
variables.
Games
-----
Compatible games that have been implemented include:
* `Reversi `_
* `Connect Four `_
* `Ultimate (or 9x9) Tic Tac Toe
`_
* `Chong `_