ChessAI

C++ Windows Forms application to play Chess with self-training Artificial Intelligence on Monte Carlo Tree
DOWNLOAD
PREVIOUS VERSION

Application Data Folder Path is "%appdata%/Chess by MrAlexeiMK" contains langs, weights, settings.txt, mcts.txt (Monte-Carlo Tree) and log.txt

Was used:

1. https://www.nature.com/articles/nature24270.epdf?author_access_token=VJXbVjaSHxFoctQQ4p2k4tRgN0jAjWel9jnR3ZoTv0PVW4gB86EEpGqTRDtpIz-2rmo8-KG06gqVobU5NSCFeHILHcVFUeMsbvwS-lxjqQGg98faovwjxeTUgZAUMnRQ
2. https://web.stanford.edu/~surag/posts/alphazero.html
3. https://habr.com/ru/post/282522/
4. https://web.archive.org/web/20180623055344/http://mcts.ai/about/index.html

Instructions:
Main menu:

[Play] - Play by yourself
[W.csv] [Play against AI] [train] - Play against AI with weights from file "W.csv", select "train" if you want to change your weights during the game
[W.csv] [Start training] [count] - Train Atrificial Intelligence by playing with itself "count" times with weights from file "W.csv"

Settings menu:

[Language] - Select file with language
[First step] - Select who will go first in "Play against AI"
[Sound] - Play sound
[Max. Steps] - Maximum steps in the game (when reached - a draw)
[Layers] - Layers of your neural network (so far only a simple neural network is supported)
[Factor] - Learning rate of neural network
[Simulations] - Number of simulations from unknown position in Monte-Carlo Tree
[History Steps] - Last how many positions to remember? (if 3 - first layer of neural network should be 256•3=768)
[Depth] - Depth of simulation (if reached and game is not ended - draw)
[Max. Depth] - Maximum depth of Monte-Carlo Tree from start node (if average number of moves in a position is 40 and Max. Depth is 8, then 40^8 ≈ 6.5e12 - average count of nodes in Monte-Carlo Tree)

Documentation: (Sources)
[logManager.h]:
Static class to print in log.txt (saves after closing application)

#include "logManager.h"

logManager::writeln("debug info");

[settingsManager.h]:
Static class to get access to settings.txt

#include "settingsManager.h"

int depth = settingsManager::getInstance().depth;

[matrix.h/matrix.cpp]:
Class for working with matrices

#include "matrix.h"

matrix A({{1,2},{2,4},{0,1}});
matrix B = A.inverse();

[neuralNetwork.h/neuralNetwork.cpp]:
Class for working with neural network (was tested on MNIST dataset and accuracy was above 60%)

#include "settingsManager.h"
#include "neuralNetwork.h"

neuralNetwork AI(settingsManager::getInstance().layers, settingsManager::getInstance().lr);
AI.loadW(settingsManager::getInstance().path + "weights\\W.csv");

[mcts.h/mcts.cpp]:
Static class for working with Monte-Carlo Tree

#include "mcts.h"
#include "logManager.h"

mcts::getInstance().init("mcts.txt");
mcts::getInstance().log();
mcts& tree = mcts::getInstance();

mcts.txt will store mcts tree by format:

[hash of position] [hash of parent position] [count of visits] [count of wins the player whose turn it is now] [step in 4 numbers]  

[chess.h/chess.cpp]:
Class containing the logic of the game of chess

#include "chess.h"

chess game();
game.doStep({{6, 4},{4, 4}});
game.printPos();

[Main.h/play.h/settings.h]:
Windows forms classes of Main/Settings/Play menus