def generate_supervised_training_data(cls, games, labeling_strategy):
"""
Generates training data by applying random moves to a board and labeling each sample with the move that :param labeling_strategy would have taken given the board.
:param games: The number of games to be simulated
:param labeling_strategy: The strategy used to label each sample. The label equals labeling_strategy.get_move(board)
:return: a list of tuples(board_sample, move_label)
"""
labeling_strategy.color = cls.config.BLACK
generator = RandomPlayer()
color_iterator = TicTacToeBaseExperiment.AlternatingColorIterator()
start = datetime.now()
training_set = []
for game in range(games):
board = TicTacToeBoard()
for i in range(9):
# generate training pair
expert_move = labeling_strategy.get_move(board)
training_set.append((board.copy(), expert_move))
# prepare for next sample
move = generator.get_move(board)
board.apply_move(move, color_iterator.__next__())
print("Generated %s training pairs form %s games in %s" % (len(training_set), games, datetime.now() - start))
return training_set
def test_Board_ApplyIllegalMove(self):
board = TicTacToeBoard()
board.apply_move((1, 1), config.BLACK)
self.assertEqual(board.illegal_move, None)
board.apply_move((1, 1), config.BLACK)
self.assertEqual(board.illegal_move, config.BLACK)
def test_getAfterstates(self):
board = TicTacToeBoard()
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2),
(2, 0), (2, 1), (2, 2)],
[a[1] for a in board.get_afterstates(config.BLACK)])
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2),
(2, 0), (2, 1), (2, 2)],
[a[1] for a in board.get_afterstates(config.WHITE)])
board.apply_move((2, 2), config.BLACK)
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2),
(2, 0), (2, 1)],
[a[1] for a in board.get_afterstates(config.BLACK)])
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2),
(2, 0), (2, 1)],
[a[1] for a in board.get_afterstates(config.WHITE)])
board.apply_move((2, 1), config.WHITE)
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2),
(2, 0)],
[a[1] for a in board.get_afterstates(config.BLACK)])
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2),
(2, 0)],
[a[1] for a in board.get_afterstates(config.WHITE)])
board.apply_move((1, 1), config.BLACK)
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 2), (2, 0)],
[a[1] for a in board.get_afterstates(config.BLACK)])
self.assertEqual([(0, 0), (0, 1), (0, 2), (1, 0), (1, 2), (2, 0)],
[a[1] for a in board.get_afterstates(config.WHITE)])
def test_Board_ApplyValidMoves(self):
board = TicTacToeBoard()
if config.BOARD_SIZE == 3:
self.assertEqual(set(board.get_valid_moves(config.BLACK)),
set([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1),
(1, 2), (2, 0), (2, 1), (2, 2)]),
msg="Valid moves incorrect")
self.assertEqual(set(board.get_valid_moves(config.WHITE)),
set([(0, 0), (0, 1), (0, 2), (1, 0), (1, 1),
(1, 2), (2, 0), (2, 1), (2, 2)]),
msg="Valid moves incorrect")
else:
self.assertEqual(config.BOARD_SIZE**2,
len(board.get_valid_moves(config.BLACK)),
msg="Incorrect Number of valid moves")
board.apply_move((1, 1), config.BLACK)
if config.BOARD_SIZE == 3:
self.assertEqual(set(board.get_valid_moves(config.BLACK)),
set([(0, 0), (0, 1), (0, 2), (1, 0), (1, 2),
(2, 0), (2, 1), (2, 2)]),
msg="Valid moves incorrect")
self.assertEqual(set(board.get_valid_moves(config.WHITE)),
set([(0, 0), (0, 1), (0, 2), (1, 0), (1, 2),
(2, 0), (2, 1), (2, 2)]),
msg="Valid moves incorrect")
def __run_episode__(self, generator):
player = self.player
rewards = []
color_iterator = self.AlternatingColorIterator()
board = TicTacToeBoard()
for i in range(9):
player_move = player.get_move(board)
# Win if predicted move is legal, loss otherwise
reward = config.LABEL_WIN if player_move in board.get_valid_moves(
player.color) else config.LABEL_LOSS
rewards.append(reward)
# prepare for next sample
board.apply_move(generator.get_move(board),
color_iterator.__next__())
loss = player.strategy.update()
player.strategy.rewards = []
average_reward = np.mean(rewards)
del rewards[:]
self.add_results([("Losses", loss), ("Score", average_reward)])
return loss, average_reward
def test_Board_CountStones(self):
board = TicTacToeBoard()
board.apply_move((0, 0), config.BLACK)
board.apply_move((1, 1), config.WHITE)
board.apply_move((2, 2), config.BLACK)
board.apply_move((1, 2), config.WHITE)
board.apply_move((1, 0), config.BLACK)
self.assertEqual((3, 2), board.count_stones())
def test_Board_GameWon(self):
board = TicTacToeBoard()
self.assertFalse(board.game_won(), msg="Empty Board")
board.apply_move((0, 0), config.BLACK)
board.apply_move((1, 1), config.WHITE)
board.apply_move((1, 0), config.BLACK)
board.apply_move((2, 2), config.WHITE)
self.assertFalse(board.game_won(), msg="No Winner yet")
board.apply_move((2, 0), config.BLACK)
self.assertEqual(board.game_won(), config.BLACK, msg="Black Won")
def test_DummyUpdate(self):
board = TicTacToeBoard()
value_function = PGStrategy(lr=0.001, weight_decay=0.003)
value_function.evaluate(board.board, board.get_legal_moves_map(config.BLACK))
move = RandomPlayer.get_move(board)
board.apply_move(move, config.BLACK)
value_function.evaluate(board.board, board.get_legal_moves_map(config.BLACK))
move = RandomPlayer.get_move(board)
board.apply_move(move, config.WHITE)
value_function.evaluate(board.board, board.get_legal_moves_map(config.BLACK))
def test_Board_Representation(self):
random_player = ttt_players.RandomPlayer()
boards = []
inverses = []
for i in range(100):
board = TicTacToeBoard()
inverse_board = TicTacToeBoard()
for j in range(9):
move = random_player.get_move(board)
color = (config.BLACK, config.WHITE)
color = random.choice(color)
board.apply_move(move, color)
boards.append(board.copy())
inverse_board.apply_move(move, board.other_color(color))
inverses.append((inverse_board.copy()))
for i in range(len(boards)):
rep = boards[i].get_representation(config.WHITE)
self.assertTrue((rep == inverses[i].board).all(),
msg="Inverting board failed")
def __run__(self, player1, player2):
"""
Runs an episode of the game
:param player1:
:param player2:
:return: The original color of the winning player
"""
self.board = TicTacToeBoard()
players = player1, player2
while True:
move = players[0].get_move(self.board.copy())
self.board.apply_move(move, players[0].color)
winner = self.board.game_won()
if winner is not None:
return config.get_label_from_winner_color(
player1, player2, winner)
players = list(reversed(players))
def test_DummyForwardPass(self):
board = TicTacToeBoard()
value_function = PGStrategy(lr=0.001, weight_decay=0.003)
value_function.evaluate(board.board, board.get_legal_moves_map(config.BLACK))
def run(self, lr, silent=False):
EVALUATION_GAMES = 10
player = FCReinforcePlayer(lr=lr)
player.color = config.BLACK
expert = ExperiencedPlayer(deterministic=True, block_mid=True)
expert.color = config.BLACK
generator = RandomPlayer()
color_iterator = self.AlternatingColorIterator()
validation_set = self.generate_supervised_training_data(
EVALUATION_GAMES,
ExperiencedPlayer(deterministic=True, block_mid=True))
print("Training ReinforcedPlayer supervised continuously with LR: %s" %
lr)
start = datetime.now()
for game in range(self.games):
rewards = []
board = TicTacToeBoard()
for i in range(9):
expert_move = expert.get_move(board)
player_move = player.get_move(board)
reward = config.LABEL_WIN if expert_move == player_move else config.LABEL_LOSS
rewards.append(reward)
# prepare for next sample
move = generator.get_move(board)
board.apply_move(move, color_iterator.__next__())
average_reward = sum(rewards) / len(rewards)
player.strategy.rewards = rewards
loss = player.strategy.update()
del rewards[:]
self.add_results([("Losses", loss), ("Reward", average_reward)])
if game % self.evaluation_period == 0:
test_rewards = []
for board, expert_move in validation_set:
# Evaluation mode
player.strategy.train, player.strategy.model.training = False, False
strategy_move = player.get_move(board)
player.strategy.train, player.strategy.model.training = True, True
test_reward = config.BLACK if expert_move == strategy_move else config.WHITE
test_rewards.append(test_reward)
average_test_reward = sum(test_rewards) / len(test_rewards)
del test_rewards[:]
self.add_results(("Test reward", average_test_reward))
if not silent:
if Printer.print_episode(game + 1, self.games,
datetime.now() - start):
plot_name = "Supervised Continuous training of %s" % (
player)
plot_info = "%s Games - Final reward: %s \nTime: %s" % (
game + 1, average_reward, config.time_diff(start))
self.plot_and_save(plot_name, plot_name + "\n" + plot_info)
return average_reward