source: java/ecj/games/scenarios/SelfPlayTDLScenario.java @ 1305

package games.scenarios;

import ec.util.MersenneTwisterFast;
import games.Board;
import games.BoardGame;
import games.GameMove;
import games.Player;

import java.util.List;

import cecj.app.othello.OthelloBoard;


public class SelfPlayTDLScenario implements GameScenario {

        private double prob;
        private Player player;
        private double learningRate;
        private MersenneTwisterFast random;

        public SelfPlayTDLScenario(MersenneTwisterFast random, Player player, double prob,
                        double learningRate) {
                this.prob = prob;
                this.player = player;
                this.random = random;
                this.learningRate = learningRate;
        }

        public int play(BoardGame game) {
                while (!game.endOfGame()) {
                        List<? extends GameMove> moves = game.findMoves();
                        if (!moves.isEmpty()) {
                                GameMove bestMove = null;
                                if (random.nextBoolean(prob)) {
                                        game.makeMove(moves.get(random.nextInt(moves.size())));
                                } else {
                                        double bestEval = Float.NEGATIVE_INFINITY;
                                        for (GameMove move : moves) {
                                                double eval = game.evalMove(player, move);
                                                if (eval > bestEval) {
                                                        bestEval = eval;
                                                        bestMove = move;
                                                }
                                        }

                                        Board previousBoard = game.getBoard().clone();
                                        game.makeMove(bestMove);
                                        updateEvaluationFunction(previousBoard, game);
                                }
                        }
                        game.switchPlayer();
                }

                return game.getOutcome();
        }

        private void updateEvaluationFunction(Board previousBoard, BoardGame game) {
                double evalBefore = tanh(previousBoard.evaluate(player));
                double derivative = (1 - (evalBefore * evalBefore));
                double error;

                if (game.endOfGame()) {
                        int result;
                        if (game.getOutcome() > 0) {
                                result = 1;
                        } else if (game.getOutcome() < 0) {
                                result = -1;
                        } else {
                                result = 0;
                        }
                        error = result - evalBefore;
                } else {
                        double evalAfter = tanh(game.getBoard().evaluate(player));
                        error = evalAfter - evalBefore;
                }

                double delta = learningRate * error * derivative;
                for (int row = 1; row <= OthelloBoard.size(); row++) {
                        for (int col = 1; col <= OthelloBoard.size(); col++) {
                                double w = player.getValue(row, col);
                                player.setValue(row, col, w + (delta * previousBoard.getValueAt(row, col)));
                        }
                }
        }

        private static double tanh(double x) {
                return 2 / (1 + Math.exp(-2 * x)) - 1;
        }
}