java-reinforcement-learning-tic-tac-toe

Demo of reinforcement learning using tic-tac-toe with the java-reinforcement-learning package.

Usage - GUI

The gui version run reinforcement learning in the interactive mode. it is implemented in com.github.chen0040.jrl.ttt.gui.Application

To run the gui version, git clone this project, run the make.ps1 or make.sh to build the reinforcement-learning-tic-tac-toe.jar into the bin folder.

Run:

java -jar bin/reinforcement-learning-tic-tac-toe.jar

Usage - Console

The console version runs the reinforcement learning against a naive bot, they are implemented in:

• com.github.chen0040.jrl.ttt.dojos.DojoQ
• com.github.chen0040.jrl.ttt.dojos.DojoSarsa
• com.github.chen0040.jrl.ttt.dojos.DojoR

Q-Learn

The following create two Q-Learning bots that plays against each other on the tic-tac-toe game to a train a Q-Learn model:

public static QLearner train(Board board, int episodes) {

    int stateCount = (int)Math.pow(3, board.size() * board.size());
    int actionCount = board.size() * board.size();

    QLearner learner = new QLearner(stateCount, actionCount);
    //learner.setActionSelection(SoftMaxActionSelectionStrategy.class.getCanonicalName());

    QBot bot1 = new QBot(1, board, learner);
    QBot bot2 =new QBot(2, board, learner);
    
    int wins = 0;

    for(int i=0; i < episodes; ++i) {
        bot1.clearHistory();
        bot2.clearHistory();

        logger.info("Iteration: {} / {}", (i+1), episodes);
        board.reset();
        while (board.canBePlayed()) {
            bot1.act();
            bot2.act();
        }
        logger.info("winner: {}", board.getWinner());
        bot1.updateStrategy();
        bot2.updateStrategy();
        logger.info("board: \n{}", board);
        
        wins += board.getWinner() == 1 ? 1 : 0;
        logger.info("success rate: {} %", (wins * 100) / (i+1));
    }

    return learner;
}

Alternaitvely, the following create one QBot that plays against a NaiveBot on the tic-tac-toe game to a train a Q-Learn model:

public static QLearner train(Board board, int episodes) {

    int stateCount = (int)Math.pow(3, board.size() * board.size());
    int actionCount = board.size() * board.size();

    QLearner learner = new QLearner(stateCount, actionCount);
    //learner.setActionSelection(SoftMaxActionSelectionStrategy.class.getCanonicalName());

    QBot bot1 = new SarsaBot(1, board, learner);
    NaiveBot bot2 =new SarsaBot(2, board);

    for(int i=0; i < episodes; ++i) {
        bot1.clearHistory();
        bot2.clearHistory();

        logger.info("Iteration: {} / {}", (i+1), episodes);
        board.reset();
        while (board.canBePlayed()) {
            bot1.act();
            bot2.act();
        }
        logger.info("winner: {}", board.getWinner());
        bot1.updateStrategy();
        logger.info("board: \n{}", board);
    }

    return learner;
}

The following code uses the trained Q-Learn model to create a QBot (bot1) which plays against bot2 (which is a naive bot that randomly takes a possible action each step):

public static double test(Board board, QLearner model, int episodes) {

    QBot bot1 = new QBot(1, board, model);
    NaiveBot bot2 =new NaiveBot(2, board);

    int wins = 0;
    int loses = 0;
    for(int i=0; i < episodes; ++i) {
        bot1.clearHistory();
        bot2.clearHistory();

        logger.info("Iteration: {} / {}", (i+1), episodes);
        board.reset();
        while (board.canBePlayed()) {
            bot1.act();
            bot2.act();
        }
        int winner = board.getWinner();
        logger.info("Winner: {}", winner);
        wins += winner == 1 ? 1 : 0;
        loses += winner == 2 ? 1 : 0;
    }

    return wins * 1.0 / episodes;
}

The following code shows how to train and then test a Q-Learn model:

Board board = new Board();
QLearner model = train(board, 30000);

double successRate = test(board, model, 1000);
logger.info("Q-Learn Bot beats Random Bot in {} % of the games being played", successRate * 100);

To save and load the Q-Learn model:

String model_json = model.toJson();
QLearner loaded_from_json = QLearner.fromJson(model_json);

This sample code can be found in the DojoQ.java file in the project.

SARSA (State-Action-Reward-State-Action)

The following create two SARSA bots that plays against each other on the tic-tac-toe game to a train a SARSA model:

public static QLearner train(Board board, int episodes) {

    int stateCount = (int)Math.pow(3, board.size() * board.size());
    int actionCount = board.size() * board.size();

    SarsaLearner learner = new SarsaLearner(stateCount, actionCount);
    //learner.setActionSelection(SoftMaxActionSelectionStrategy.class.getCanonicalName());

    SarsaBot bot1 = new SarsaBot(1, board, learner);
    SarsaBot bot2 =new SarsaBot(2, board, learner);

    for(int i=0; i < episodes; ++i) {
        bot1.clearHistory();
        bot2.clearHistory();

        logger.info("Iteration: {} / {}", (i+1), episodes);
        board.reset();
        while (board.canBePlayed()) {
            bot1.act();
            bot2.act();
        }
        logger.info("winner: {}", board.getWinner());
        bot1.updateStrategy();
        bot2.updateStrategy();
        logger.info("board: \n{}", board);
    }

    return learner;
}

Alternaitvely, the following create one SARSA bot that plays against a NaiveBot on the tic-tac-toe game to a train a SARSA model:

public static QLearner train(Board board, int episodes) {

    int stateCount = (int)Math.pow(3, board.size() * board.size());
    int actionCount = board.size() * board.size();

    SarsaLearner learner = new SarsaLearner(stateCount, actionCount);
    //learner.setActionSelection(SoftMaxActionSelectionStrategy.class.getCanonicalName());

    SarsaBot bot1 = new SarsaBot(1, board, learner);
    NaiveBot bot2 =new SarsaBot(2, board);
    
    int wins = 0;

    for(int i=0; i < episodes; ++i) {
        bot1.clearHistory();
        bot2.clearHistory();

        logger.info("Iteration: {} / {}", (i+1), episodes);
        board.reset();
        while (board.canBePlayed()) {
            bot1.act();
            bot2.act();
        }
        logger.info("winner: {}", board.getWinner());
        bot1.updateStrategy();
        logger.info("board: \n{}", board);
        
        wins += board.getWinner() == 1 ? 1 : 0;
        logger.info("success rate: {} %", (wins * 100) / (i+1));
    }

    return learner;
}

The following code uses the trained SARSA model to create a SarsaBot (bot1) which plays against bot2 (which is a naive bot that randomly takes a possible action each step):

public static double test(Board board, SarsaLearner model, int episodes) {

    SarsaBot bot1 = new SarsaBot(1, board, model);
    NaiveBot bot2 =new NaiveBot(2, board);

    int wins = 0;
    int loses = 0;
    for(int i=0; i < episodes; ++i) {
        bot1.clearHistory();
        bot2.clearHistory();

        logger.info("Iteration: {} / {}", (i+1), episodes);
        board.reset();
        while (board.canBePlayed()) {
            bot1.act();
            bot2.act();
        }
        int winner = board.getWinner();
        logger.info("Winner: {}", winner);
        wins += winner == 1 ? 1 : 0;
        loses += winner == 2 ? 1 : 0;
    }

    return wins * 1.0 / episodes;
}

The following code shows how to train and then test a SARSA model:

Board board = new Board();
SarsaLearner model = train(board, 30000);

double successRate = test(board, model, 1000);
logger.info("SARSA Bot beats Random Bot in {} % of the games being played", successRate * 100);

To save and load the SARSA model:

String model_json = model.toJson();
SarsaLearner loaded_from_json = SarsaLearner.fromJson(model_json);

This sample code can be found in the DojoSarsa.java file in the project.