def run(config=None):
if config == None:
config = load_config(file_name=root_data_file + 'resnet_6_6_4.model',
only_load_param=True)
try:
board = Board(width=config.board_width,
height=config.board_height,
n_in_row=config.n_in_row)
game = Game(board)
# --------------- human VS AI ----------------
best_policy = PolicyValueNet(
config.board_width,
config.board_height,
Network=config.network,
net_params=config.policy_param
) # setup which Network to use based on the net_params
mcts_player = AlphaZeroPlayer(
best_policy.predict,
c_puct=config.c_puct,
nplays=100,
add_noise=True) # set larger nplays for better performance
# uncomment the following line to play with pure MCTS
# mcts_player2 = RolloutPlayer(nplays=1000, c_puct=config.c_puct)
# human player, input your move in the format: 2,3
human = HumanPlayer()
# set who_first=0 for human first
game.start_game(human, mcts_player, who_first=1, is_shown=1)
except KeyboardInterrupt:
print('\n\rquit')
def __init__(self):
# params of the board and the game
self.board_width = 6
self.board_height = 6
self.n_in_row = 4
self.board = ShogiBoard()
# training params
self.learn_rate = 5e-3
self.lr_multiplier = 1.0 # adaptively adjust the learning rate based on KL
self.temp = 1.0 # the temperature param
self.n_playout = 400 # num of simulations for each move
self.c_puct = 5
self.buffer_size = 10000
self.batch_size = 512 # mini-batch size for training
self.data_buffer = deque(maxlen=self.buffer_size)
self.play_batch_size = 1
self.epochs = 5 # num of train_steps for each update
self.kl_targ = 0.025
self.check_freq = 50
self.game_batch_num = 3000
self.best_win_ratio = 0.0
# num of simulations used for the pure mcts, which is used as the opponent to evaluate the trained policy
self.pure_mcts_playout_num = 1000
# start training from a given policy-value net
# policy_param = pickle.load(open('current_policy.model', 'rb'))
# self.policy_value_net = PolicyValueNet(self.board_width, self.board_height, net_params = policy_param)
# start training from a new policy-value net
self.policy_value_net = PolicyValueNet()
self.mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout,
is_selfplay=1)
def __init__(self, config=None):
# params of the board and the game
self.config = config if config else Config()
# Network wrapper
self.policy_value_net = PolicyValueNet(self.config.board_width, self.config.board_height,
net_params=self.config.policy_param,
Network=self.config.network)
# 传入policy_value_net的predict方法，神经网络辅助MCTS搜索过程
self.mcts_player = AlphaZeroPlayer(self.policy_value_net.predict, c_puct=self.config.c_puct,
nplays=self.config.n_playout, is_selfplay=True)
def __init__(self, modelPath=None):
# 棋盘和游戏
self.boardWidth = 4
self.boardHeight = 4
self.game = Game()
# 训练参数
self.learningRate = 5e-3
self.learningRateMultiplier = 1.0 # 自适应
try:
self.learningRateMultiplier = float(Util.getNewestLearningRateMultiplier(type='from_db' if modelPath is None else 'from_self_play'))
except Exception as e:
print(str(e))
self.temperature = 1.0 # 温度, 含义见参考资料.txt第2条
self.playoutTimes = 500 # 模拟次数
self.polynomialUpperConfidenceTreesConstant = 5 # 论文中的c_puct, 含义见参考资料.txt第2条
self.dataDequeSize = 10000
self.trainBatchSize = 512 # 训练批次尺寸,原本为512,先使用50用于调试
self.dataDeque = deque(maxlen=self.dataDequeSize) # 超出maxlen会自动删除另一边的元素
self.playBatchSize = 1
self.epochs = 5 # 单次训练拟合多少次
self.KLDParam = 0.025
self.checkFrequency = 1000 # 之前为100,改为1000,因为评测太浪费时间
self.gameBatchSize = 200000 # 5000时对mcts1500的胜率为0.9,所以再升到10000
self.maxWinRatio = 0.0
self.pureMctsPlayoutTimes = 1500 # 初始为500,现在已到1500s
self.pureMctsPlayoutTimesAddend = 500
self.maxPureMctsPlayoutTimes = 3000
self.modelPath = modelPath
self.trainedGameCountInDB = Util.readGameCount(type='train')
self.lossDataCount = 12184 # 被黑客删掉的棋谱数量,这些棋谱数据已无法恢复
if modelPath is not None:
self.policyValueNet = PolicyValueNet(self.boardWidth, self.boardHeight, logPath=Util.getTrainLogPath(isFromDB=False), modelPath=modelPath)
self.trainedGameCount = self.trainedGameCountInDB + self.lossDataCount
else:
self.policyValueNet = PolicyValueNet(self.boardWidth, self.boardHeight, logPath=Util.getTrainLogPath(isFromDB=True))
self.trainedGameCount = 0 + self.lossDataCount
self.zeroPlayer = ZeroPlayer(self.policyValueNet.policyValueFunction,
polynomialUpperConfidenceTreesConstant=self.polynomialUpperConfidenceTreesConstant,
playoutTimes=self.playoutTimes, isSelfPlay=1)
def __init__(self, config=None):
# params of the board and the game
self.config = config if config else Config()
if not hasattr(self.config, "use_gpu"):
setattr(config, "use_gpu",
False) # compatible with old version config
# Network wrapper
self.policy_value_net = PolicyValueNet(
self.config.board_width,
self.config.board_height,
net_params=self.config.policy_param,
Network=self.config.network,
use_gpu=self.config.use_gpu)
# forward the reference of policy_value_net'predict function，for MCTS simulation
self.mcts_player = AlphaZeroPlayer(self.policy_value_net.predict,
c_puct=self.config.c_puct,
nplays=self.config.n_playout,
is_selfplay=True)
def run(config=None):
if config == None:
config = load_config(file_name=root_data_file + 'resnet_6_6_4.model',
only_load_param=True)
try:
board = Board(width=config.board_width,
height=config.board_height,
n_in_row=config.n_in_row)
#--------------------1.set player:alphazero VS human---------------------#
best_policy = PolicyValueNet(
config.board_width,
config.board_height,
Network=config.network,
net_params=config.policy_param
) # setup which Network to use based on the net_params
player1 = AlphaZeroPlayer(
best_policy.predict, c_puct=config.c_puct,
nplays=1000) #set larger nplays for better performance
# uncomment the following line to play with pure MCTS
#player2 = RolloutPlayer(nplays=1000, c_puct=config.c_puct)
player2 = HumanPlayer()
# --------------------2.set order---------------------#
who_first = 0 # 0 means player1 first, otherwise player2 first
# --------------------3.start game--------------------#
game = Game(board, is_visualize=True)
t = threading.Thread(target=game.start_game,
args=(player1, player2, who_first))
t.start()
game.show()
except:
print('\n\rquit')