MET-CS-767

Term Project for MetCS 767 Machine Learning Fall 2017

Trains a chess bot to learn strategies for chess via a few simple examples while still using a neural net. This is different than DeepMind's chess model which uses billions of complex training examples.

Requirements

• scala 2.12.3 (or higher)
• sbt 0.13.8 (or higher)

Building and Testing the project

Navigate to root directory and type:

sbt clean compile
sbt test

Running the Experiment

Experimentation is broken down into 3 parts

• Generating training data
• Training the Neural Net
• Testing the Neural Net

Generating Training Data

Go to DataGenerator.scala and run the main method, takes no command line arguments

Modifiable parameters:

• Starting state (line 45). Can use any of the other test cases 1 through 4. Default is testCase4.
• Number of iterations (line 46). Default is 500
• Number of turns before ending a game (line 52). Default is 10

45    val pieces = testCase4
46    for(i <- 1 to 500) {
47      var gameOver = false
48     var turn: Color = White
49      println(i)
50      while (!gameOver && turn == White) {
51        val game = new ChessGame(pieces, White)
52        game.runGame(10)
53        turn = game.turn
54        gameOver = game.isGameOver
55      }
56    }

Training/Testing the Neural Net

Go to MainExperiment.scala and run the main method, takes no command line arguments.

Modifiable parameters:

• Starting state (line 11). Can use any of the other test cases 1 through 4. Defaults to testCase4
• Training set (line 9). Can use any csv file of your choice. Defaults to final_training_data.csv

7   def main(args: Array[String]) {
8     val model = new InferenceModel
9     model.train("final_training_data.csv")
10    val pieces = DataGenerator.testCase4
11    val game = new ChessGame(pieces, White)
12    game.runGame(10, Some(model))
13    println(game.numMoves)
14    val move = model.computeMoveVector(StateVector(pieces, White))
15    println(move.toReadableMove)
16  }