Skip to content

Latest commit

 

History

History
82 lines (57 loc) · 2.56 KB

README.rst

File metadata and controls

82 lines (57 loc) · 2.56 KB
Raw

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: