def action(self, board):
env = GameEnv().update(board)
key = self.counter_key(env)
#No_White = True
#White = ['W', 'X', 'Y']
#for i in range(3):
# for j in range(3):
# if env.board[i][j] in White:
# No_White = False
#if No_White: action = 999
#else:
for tl in range(self.play_config.thinking_loop):
self.search_moves(board)
policy = self.calc_policy(board)
action = int(np.random.choice(range(self.labels_n), p=policy))
action_by_value = int(
np.argmax(self.var_q[key] + (self.var_n[key] > 0) * 100))
if action == action_by_value or env.turn < self.play_config.change_tau_turn:
break
# this is for play_gui, not necessary when training.
self.thinking_history[env.observation] = HistoryItem(
action, policy, list(self.var_q[key]), list(self.var_n[key]))
self.moves.append([env.observation, list(policy)])
return action
def play_game(self, best_model, ng_model):
env = GameEnv().reset()
best_player = GamePlayer(self.config,
best_model,
play_config=self.config.eval.play_config)
ng_player = GamePlayer(self.config,
ng_model,
play_config=self.config.eval.play_config)
best_is_white = random() < 0.5
if not best_is_white:
black, white = best_player, ng_player
else:
black, white = ng_player, best_player
env.reset()
while not env.done:
if env.player_turn() == Player.black:
action = black.action(env.board)
else:
action = white.action(env.board)
env.step(action)
ng_win = None
if env.winner == Winner.white:
if best_is_white:
ng_win = 0
else:
ng_win = 1
elif env.winner == Winner.black:
if best_is_white:
ng_win = 1
else:
ng_win = 0
return ng_win, best_is_white
def start(config: Config):
PlayWithHumanConfig().update_play_config(config.play)
game_model = PlayWithHuman(config)
while True:
env = GameEnv().reset()
human_is_black = random() < 0.5
game_model.start_game(human_is_black)
while not env.done:
if env.player_turn() == Player.black:
if not human_is_black:
action = game_model.move_by_ai(env)
print("IA moves to: " + str(action))
else:
action = game_model.move_by_human(env)
print("You move to: " + str(action))
else:
if human_is_black:
action = game_model.move_by_ai(env)
print("IA moves to: " + str(action))
else:
action = game_model.move_by_human(env)
print("You move to: " + str(action))
env.step(action)
env.render()
print("\nEnd of the game.")
print("Game result:")
if env.winner == Winner.white:
print("X wins")
elif env.winner == Winner.black:
print("O wins")
else:
print("Game was a draw")
async def start_search_my_move(self, board):
self.running_simulation_num += 1
with await self.sem: # reduce parallel search number
env = GameEnv().update(board)
#print(env.board)
#input()
leaf_v = await self.search_my_move(env, is_root_node=True)
self.running_simulation_num -= 1
return leaf_v
def calc_policy(self, board):
#calc π(a|s0)
pc = self.play_config
env = GameEnv().update(board)
key = self.counter_key(env)
if env.turn < pc.change_tau_turn:
#print(key)
#print('self.var_n[key]',self.var_n[key])
#print('np.sum(self.var_n[key]',np.sum(self.var_n[key]))
#input()
return self.var_n[key] / np.sum(self.var_n[key]) # tau = 1
else:
action = np.argmax(self.var_n[key]) # tau = 0
ret = np.zeros(self.labels_n)
ret[action] = 1
return ret
def __init__(self, config: Config):
self.config = config
self.model = None # type: GameModel
self.loaded_filenames = set()
self.loaded_data = {}
self.dataset = None
self.optimizer = None
self.dbx = None
self.version = 0 # Change to dynamic lookup from Drop Box Files
self.env = GameEnv()
self.raw_timestamp = None
self.best_is_white = True
self.play_files_per_generation = 15 # each file includes 25 games so each generation adds 375 games
self.max_play_files = 300
# at final there are alawys 7500 games to look at from previous 20 generations
self.min_play_files_to_learn = 0
self.play_files_on_dropbox = 0
def convert_to_training_data(data):
state_list = []
policy_list = []
z_list = []
for state, policy, z in data:
board = list(state)
board = np.reshape(board, (3, 3))
env = GameEnv().update(board)
black_ary, white_ary = env.black_and_white_plane()
state = [
black_ary, white_ary
] if env.player_turn() == Player.black else [white_ary, black_ary]
state_list.append(state)
policy_list.append(policy)
z_list.append(z)
return np.array(state_list), np.array(policy_list), np.array(z_list)
def play_game(self, best_model, ng_model):
env = GameEnv().reset()
if (self.raw_timestamp != self.dbx.files_get_metadata(
'/model/model_best_weight.h5').client_modified):
print('A newer model version is available - giving up this match')
ng_win = 0
self.best_is_white = True
return ng_win, self.best_is_white
best_player = GamePlayer(self.config,
best_model,
play_config=self.config.eval.play_config)
ng_player = GamePlayer(self.config,
ng_model,
play_config=self.config.eval.play_config)
self.best_is_white = not self.best_is_white
if not self.best_is_white:
black, white = best_player, ng_player
else:
black, white = ng_player, best_player
env.reset()
while not env.done:
if env.player_turn() == Player.black:
action = black.action(env.board)
else:
action = white.action(env.board)
env.step(action)
ng_win = None
if env.winner == Winner.white:
if self.best_is_white:
ng_win = 0
else:
ng_win = 1
elif env.winner == Winner.black:
if self.best_is_white:
ng_win = 1
else:
ng_win = 0
return ng_win, self.best_is_white
def start(config: Config):
tf_util.set_session_config(per_process_gpu_memory_fraction=0.2)
return SelfPlayWorker(config, env=GameEnv()).start()