https://github.com/elliotpenson/n-gram-chess

https://github.com/elliotpenson/n-gram-chess

Last synced: 5 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/elliotpenson/n-gram-chess
• Owner: ElliotPenson
• License: mit
• Created: 2016-06-06T23:13:41.000Z (over 8 years ago)
• Default Branch: master
• Last Pushed: 2016-06-11T02:19:41.000Z (over 8 years ago)
• Last Synced: 2024-11-07T20:52:19.439Z (about 2 months ago)
• Language: Common Lisp
• Size: 62.1 MB
• Stars: 5
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.org
  • License: LICENSE

Awesome Lists containing this project

README

        
#+TITLE: n-gram-chess

#+AUTHOR: Elliot Penson

#+OPTIONS: num:nil

A chess engine powered by an n-gram model. Inspired by [[https://news.ycombinator.com/item?id=8945231][this post]] on

Hacker News. The project depends on my chess library [[https://github.com/ElliotPenson/knight-owl][knight-owl]].

The n-gram level may be specified by the user. The model is trained on

a set of chess games in portable game notation (PGN). Such a

collection exists in the corpus directory. These files are taken from

the [[http://www.pgnmentor.com/files.html#players][PGN Mentor Database]]. During training, the engine establishes a

frequency map. Various functions use this map to find conditional

probabilities of chess moves (i.e. the likelihood of a move given n-1

previous ones).

** Usage

   First, train the model. The ~train~ function begins by searching

   the corpus directory for PGN files. The function evaluates to a

   ~hash-table~ of n-gram counts from the moves of these games. This

   table needs to be stored for future use:

   #+BEGIN_SRC lisp

     (defparameter *count-map* (train 3))

   #+END_SRC

   The required parameter (in this case ~3~) indicates the n of the

   model. 2 would be a bigram model, 3 trigram, and so on.

   The quality (or /probability/) of a move can now be

   tested. ~move-probability~ takes a move in algebraic chess

   notation, a list of previous moves, and the n-gram count map. For

   example:

   #+BEGIN_SRC lisp

     (move-probability "Nc6" '("e5" "Nf3") *count-map*)

     -> 0.5555556

   #+END_SRC

   The model also has the ability to select the best move given a

   chess board's state. Let's begin by setting up a board.

   #+BEGIN_SRC lisp

     (defparameter *test-board* (new-board))

   #+END_SRC

   

   Now let's do some moves.

   #+BEGIN_SRC lisp

     (loop for move in '("e4" "e5" "Nf3" "Nc6" "Bb5")

           for whitep = t then (not whitep)

           do (make-move move *test-board* whitep))

     (print-board *test-board*)

     -> 8 |♜||_||♝||♛||♚||♝||♞||♜|

        7 |♟||♟||♟||♟||_||♟||♟||♟|

        6 |_||_||♞||_||_||_||_||_|

        5 |_||♗||_||_||♟||_||_||_|

        4 |_||_||_||_||♙||_||_||_|

        3 |_||_||_||_||_||♘||_||_|

        2 |♙||♙||♙||♙||_||♙||♙||♙|

        1 |♖||♘||♗||♕||♔||_||_||♖|

           A  B  C  D  E  F  G  H

   #+END_SRC

   Looking good! Now for the most likely move. The ~best-move~

   function accepts a board (like *test-board*), if the player is

   white or black, a list of previous moves, and the n-gram count

   map.

   #+BEGIN_SRC lisp

     (best-move *test-board* nil '("Nc6" "Bb5") *count-map*)

     -> "a6"

        0.45833334

   #+END_SRC

   

   We receive the most probable move and it's associated

   probability. Note that we only handed ~best-move~ the last two

   moves. This is because we are working with a trigram model.

** TODO Future Improvements