def genmove(self, player):
if winner(self.board) or winner(flip(self.board)):
raise GtpException('Game is over')
self.history.append(self.board)
if player == 'black':
# TODO: reuse previous calculations from the MCTS
predictor = TreeSearchPredictor(self.config.search_config,
self.model, self.board,
not self.history)
predictor.run(self.config.iterations)
value, probabilities = predictor.predict()
move = best_move(probabilities)
self.board = flip(make_move(self.board, move))
elif player == 'white':
predictor = TreeSearchPredictor(self.config.search_config,
self.model, flip(self.board),
not self.history)
predictor.run(self.config.iterations)
value, probabilities = predictor.predict()
move = best_move(probabilities)
self.board = make_move(flip(self.board), move)
move = flip_move(move)
else:
self.history.pop()
raise GtpException('Player is invalid')
print('Estimated value: %.2f' % value, file=sys.stderr)
return write_move(move)
def play(self, player, move):
move = read_move(move)
self.history.append(numpy.copy(self.board))
if player == 'black':
self.board = flip(make_move(self.board, move))
elif player == 'white':
self.board = make_move(flip(self.board), flip_move(move))
else:
self.history.pop()
raise GtpException('Player is invalid')
return ''
def winner(self):
if winner(self.board):
return 'white'
elif winner(flip(self.board)):
return 'black'
else:
return 'none'
def compare(config, model_file1, model_file2):
models = [load_model(model_file1), load_model(model_file2)]
games = 0
first_player_wins = 0
win_ratio, uncertainty = None, None
while True:
move_index = 0
predictors = [TreeSearchPredictor(config.search_config, model, new_board(config.size), True) for model in models]
while not winner(predictors[0].board):
if move_index == 0:
predictor = predictors[1]
else:
predictor = predictors[(games ^ move_index) & 1]
predictor.run(config.iterations)
value, probabilities = predictor.predict()
if move_index < 3:
move = sample_move(probabilities)
else:
move = best_move(probabilities)
for predictor in predictors:
predictor.make_move(move)
if games & 1 == move_index & 1:
print_board(flip(predictors[0].board), move, file=sys.stderr)
else:
print_board(predictors[0].board, flip_move(move), file=sys.stderr)
print('%s model win probability: %.2f' % (['First', 'Second'][((games ^ move_index) & 1)], (value + 1) / 2), file=sys.stderr)
if games > 0:
print('Win ratio %.2f ± %.2f (%d games)' % (win_ratio, uncertainty, games), file=sys.stderr)
move_index += 1
games += 1
if games & 1 == move_index & 1:
first_player_wins += 1
win_ratio = float(first_player_wins) / games
uncertainty = win_ratio * math.sqrt(win_ratio * (1 - win_ratio) / games)
def compare(config, model_file1, model_file2, temp, num_games):
models = [load_model(model_file1), load_model(model_file2)]
games = 0
first_player_wins = 0
win_ratio, uncertainty = None, None
ratios = []
# while True:
for i in range(num_games):
move_index = 0
predictors = [
TreeSearchPredictor(config.search_config, model,
new_board(config.size), True)
for model in models
]
# exp uct of model2 is 100
# predictors = ["avshalom", "shlomo"]
# config.search_config.uct_factor = 5.0
# predictors[0] = TreeSearchPredictor(config.search_config, models[0], new_board(config.size), True)
# config.search_config.uct_factor = 100.0
# predictors[1] = TreeSearchPredictor(config.search_config, models[1], new_board(config.size), True)
# exp uct of model2 is 100
# predictors = ["avshalom", "shlomo"]
# config.search_config.uct_factor = 5.0
# predictors[0] = TreeSearchPredictor(config.search_config, models[0], new_board(config.size), True)
# config.search_config.uct_factor = 100.0
# predictors[1] = TreeSearchPredictor(config.search_config, models[1], new_board(config.size), True)
# exp virtual loss of model2 is 100
# predictors = ["avshalom", "shlomo"]
# config.search_config.virtual_loss = 3.0
# predictors[0] = TreeSearchPredictor(config.search_config, models[0], new_board(config.size), True)
# config.search_config.virtual_loss = 100.0
# predictors[1] = TreeSearchPredictor(config.search_config, models[1], new_board(config.size), True)
# exp virtual loss of model2 is 0
# predictors = ["avshalom", "shlomo"]
# config.search_config.virtual_loss = 3.0
# predictors[0] = TreeSearchPredictor(config.search_config, models[0], new_board(config.size), True)
# config.search_config.virtual_loss = 0
# predictors[1] = TreeSearchPredictor(config.search_config, models[1], new_board(config.size), True)
while not winner(predictors[0].board):
if move_index == 0:
predictor = predictors[1]
else:
predictor = predictors[(games ^ move_index) & 1]
predictor.run(config.iterations)
value, probabilities = predictor.predict()
# exp uniform probs
#uprobs = [0.00826446] * 121
#if games & 1 == move_index & 1:
# probabilities = np.array(uprobs).reshape(11, -1)
# exp temperature
tprobs = temperature(probabilities, temp)
if games & 1 == move_index & 1:
probabilities = tprobs
#print(probabilities)
if move_index < 3:
move = sample_move(probabilities)
else:
#move = best_move(probabilities)
move = sample_move(probabilities)
for predictor in predictors:
predictor.make_move(move)
if games & 1 == move_index & 1:
print_board(flip(predictors[0].board), move, file=sys.stderr)
else:
print_board(predictors[0].board,
flip_move(move),
file=sys.stderr)
print('%s model win probability: %.2f' %
(['First', 'Second'][((games ^ move_index) & 1)],
(value + 1) / 2),
file=sys.stderr)
if games > 0:
print('Win ratio %.2f ± %.2f (%d games)' %
(win_ratio, uncertainty, games),
file=sys.stderr)
move_index += 1
games += 1
if games & 1 == move_index & 1:
first_player_wins += 1
win_ratio = float(first_player_wins) / games
uncertainty = win_ratio * math.sqrt(win_ratio *
(1 - win_ratio) / games)
ratios.append(win_ratio)
return ratios