def decide_pong(self, player, new_tile, neighbors, game): self.begin_decision() fixed_hand, hand = player.fixed_hand, player.hand if self.display_step: self.print_game_board(fixed_hand, hand, neighbors, game, new_tile) self.print_msg("Someone just discarded a %s." % new_tile.symbol) q_network = get_Network(self.q_network_path) state = utils.dnn_encode_state(player, neighbors) if not self.skip_history and self.history_waiting: self.update_transition(state, REWARD_NON_TERMINAL) valid_actions = [34 + decisions_.index("%s_pong" % new_tile.suit), 34 + decisions_.index("no_action")] action_filter = np.zeros(n_decisions) action_filter[valid_actions] = 1 action = None while True: if action is not None and not self.skip_history: self.update_history(state, action, action_filter) self.update_transition(state, REWARD_INVALID_DECISION) action, value = q_network.choose_action(state, action_filter=action_filter, eps_greedy=self.is_train, return_value=True, strict_filter=not self.is_train) if action in valid_actions: break elif not self.is_train: action = random.choice(valid_actions) break if not self.skip_history: self.update_history(state, action, action_filter) self.end_decision() if action == 34 + decisions_.index("no_action"): self.print_msg("%s [%s] chooses to form a Pong %s%s%s. [%.2f]" % ( self.player_name, display_name, new_tile.symbol, new_tile.symbol, new_tile.symbol, value)) if game.lang_code is not None: game.add_notification(get_text(game.lang_code, "NOTI_CHOOSE_PONG") % ( self.player_name, new_tile.get_display_name(game.lang_code, is_short=False))) return True else: self.print_msg("%s [%s] chooses not to form a Pong %s%s%s. [%.2f]" % ( self.player_name, display_name, new_tile.symbol, new_tile.symbol, new_tile.symbol, value)) return False
def update_transition(self, state_, reward=0, action_filter_=None): if not self.is_train: return if not self.history_waiting: raise Exception("the network is NOT waiting for a transition") if type(state_) == str and state_ == "terminal": state_ = self.q_network_history["state"] self.history_waiting = False q_network = get_Network(self.q_network_path) q_network.store_transition(self.q_network_history["state"], self.q_network_history["action"], reward, state_, self.q_network_history["action_filter"])
def main(): global game_record_count trainer_models["deepq"]["parameters"]["q_network_path"] = deep_q_model_dir model = get_Network(deep_q_model_dir, **deep_q_model_paras) players = [] i = 0 for model_tag in trainer_conf: player = Player.Player(trainer_models[model_tag]["class"], player_name=names[i], **trainer_models[model_tag]["parameters"]) players.append(player) i += 1 deepq_player = Player.Player(Generator, player_name=names[i], q_network_path=deep_q_model_dir, skip_history=False, is_train=True, display_step=False) players.append(deepq_player) signal.signal(signal.SIGINT, signal_handler) game, shuffled_players, last_saved = None, None, -1 for i in range(n_epochs): if EXIT_FLAG: break if i % freq_shuffle_players == 0: shuffled_players = random.sample(players, k=4) game = Game.Game(shuffled_players) winner, losers, penalty = game.start_game() model.learn(display_cost=(i + 1) % game_record_size == 0) index = game_record_count % game_record_size game_record[index, :, :] = np.zeros((4, 2)) game_record_count += 1 if winner is not None: winner_id = players.index(winner) game_record[index, winner_id, 0] = 1 for loser in losers: loser_id = players.index(loser) game_record[index, loser_id, 1] = 1 if (i + 1) % game_record_size == 0: print("#%5d: %.2f%%/%.2f%%\t%.2f%%/%.2f%%\t%.2f%%/%.2f%%\t%.2f%%/%.2f%%" % ( i + 1, game_record[:, 0, 0].mean() * 100, game_record[:, 0, 1].mean() * 100, game_record[:, 1, 0].mean() * 100, game_record[:, 1, 1].mean() * 100, game_record[:, 2, 0].mean() * 100, game_record[:, 2, 1].mean() * 100, game_record[:, 3, 0].mean()* 100, game_record[:, 3, 1].mean()* 100)) if last_saved < n_epochs - 1: path = save_name.rstrip("/") + "_%d" % n_epochs utils.makesure_dir_exists(path) model.save(path)
def decide_drop_tile(self, player, new_tile, neighbors, game): self.begin_decision() fixed_hand, hand = player.fixed_hand, player.hand state = utils.dnn_encode_state(player, neighbors) if not self.skip_history and self.history_waiting: self.update_transition(state, REWARD_NON_TERMINAL) if self.display_step: self.print_game_board(fixed_hand, hand, neighbors, game, new_tile) q_network = get_Network(self.q_network_path) valid_actions = [] tiles = player.hand if new_tile is None else player.hand + [new_tile] for tile in tiles: valid_actions.append(Tile.convert_tile_index(tile)) action_filter = np.zeros(n_decisions) action_filter[valid_actions] = 1 action = None while True: if action is not None and not self.skip_history: self.update_history(state, action, action_filter) self.update_transition(state, REWARD_INVALID_DECISION) action, value = q_network.choose_action(state, action_filter=action_filter, eps_greedy=self.is_train, return_value=True, strict_filter=not self.is_train) if action in valid_actions: break elif not self.is_train: action = random.choice(valid_actions) break if not self.skip_history: self.update_history(state, action, action_filter) drop_tile = Tile.convert_tile_index(action) self.print_msg("%s [%s] chooses to drop %s. [%.2f]" % (self.player_name, display_name, drop_tile.symbol, value)) self.end_decision(True) if game.lang_code is not None: game.add_notification(get_text(game.lang_code, "NOTI_CHOOSE_DISCARD") % ( self.player_name, drop_tile.get_display_name(game.lang_code, is_short=False))) return drop_tile
def decide_chow(self, player, new_tile, choices, neighbors, game): self.begin_decision() fixed_hand, hand = player.fixed_hand, player.hand if self.display_step: self.print_game_board(fixed_hand, hand, neighbors, game) self.print_msg("Someone just discarded a %s." % new_tile.symbol) q_network = get_Network(self.q_network_path) state = utils.dnn_encode_state(player, neighbors) # store the transition to the network (state(t-1), action(t-1), reward(not terminated), state(t) if not self.skip_history and self.history_waiting: self.update_transition(state, REWARD_NON_TERMINAL) valid_actions = [34 + decisions_.index("%s_chow" % new_tile.suit), 34 + decisions_.index("no_action")] action_filter = np.zeros(n_decisions) action_filter[valid_actions] = 1 action = None # choose the action while True: if action is not None and not self.skip_history: self.update_history(state, action, action_filter) self.update_transition(state, REWARD_INVALID_DECISION) action, value = q_network.choose_action(state, action_filter=action_filter, eps_greedy=self.is_train, return_value=True, strict_filter=not self.is_train) if action in valid_actions: break elif not self.is_train: action = random.choice(valid_actions) break if not self.skip_history: self.update_history(state, action, action_filter) self.end_decision() # print the msg of the choice taken if action == 34 + decisions_.index("no_action"): self.print_msg("%s chooses not to Chow %s [%.2f]." % (self.player_name, new_tile.symbol, value)) return False, None else: chow_tiles_tgstrs = [] chow_tiles_str = "" choice = random.choice(choices) for i in range(choice - 1, choice + 2): neighbor_tile = new_tile.generate_neighbor_tile(i) chow_tiles_str += neighbor_tile.symbol chow_tiles_tgstrs.append(neighbor_tile.get_display_name(game.lang_code, is_short=False)) self.print_msg("%s chooses to Chow %s [%.2f]." % (self.player_name, chow_tiles_str, value)) if game.lang_code is not None: game.add_notification( get_text(game.lang_code, "NOTI_CHOOSE_CHOW") % (self.player_name, ",".join(chow_tiles_tgstrs))) return True, choice