def play_round(self,
num_reads: int) -> Tuple[Optional[str], List[np.ndarray]]:
"""
Evaluate the trained network by playing matches between the current and the previous NN
@param num_reads: see args
"""
print("Starting game round...")
# randomly choose starting player
if np.random.uniform(0, 1) <= 0.5:
white = self.current
black = self.best
w = "current"
b = "best"
else:
white = self.best
black = self.current
w = "best"
b = "current"
# initializing
current_board = Board()
game_won = False
dataset = []
value = 0
temperature = 0.1 # exploration vs exploitation factor (smaller -> more exploitation)
while not game_won and current_board.is_playable():
dataset.append(copy.deepcopy(current_board.encode()))
# get Policy
if current_board.player == PLAYER_1:
root = UCT_search(current_board, num_reads, white)
policy = get_policy(root, temperature)
print("Policy: ", policy, "white = %s" % (str(w)))
elif current_board.player == PLAYER_2:
root = UCT_search(current_board, num_reads, black)
policy = get_policy(root, temperature)
print("Policy: ", policy, "black = %s" % (str(b)))
else:
raise AssertionError("Invalid player.")
# Chose a Column with given policy
col_choice = np.random.choice(np.array([0, 1, 2, 3, 4, 5, 6]),
p=policy)
current_board.drop_piece(col_choice) # move piece
print(current_board)
if current_board.check_winner(): # someone wins
if current_board.player == PLAYER_1: # black wins
value = -1
elif current_board.player == PLAYER_2: # white wins
value = 1
game_won = True
# Append new board to the dataset encoded in one-hot-encoding manner
dataset.append(current_board.encode())
if value == -1:
dataset.append(f"{b} as black wins")
return b, dataset
elif value == 1:
dataset.append(f"{w} as white wins")
return w, dataset
else:
dataset.append("Nobody wins")
return None, dataset
def self_play(net: Connect4Network, start_index: np.int, cpu_index: np.int,
num_games: np.int, args: AlphaZeroArgs, iteration: np.int):
"""
Self Play of AlphaZero, generating and saving Datasets for the training of the Neural Network
@param net:
@param start_index: Start index of Self Play games
@param cpu_index:
@param num_games:
@param args:
@param iteration: current Iteration
"""
# number of more random moves, before lowering temp
n_max_moves = 11
print(f"CPU={cpu_index}: Starting MCTS")
iteration_dir = f"./datasets/iter_{iteration}"
if not os.path.isdir(iteration_dir):
os.makedirs(iteration_dir)
# Play self play games
for idx in range(start_index, num_games + start_index):
print(f"Game {idx}")
current_board = Board()
game_won = False # indicates that a game is won
dataset = []
states = []
value = 0
move_count = 0
while not game_won and current_board.is_playable():
t = 0.1
# less random further into the game
if move_count < n_max_moves:
t = args.temperature_mcts
# save current board state (encoded and unencoded)
states.append(current_board.current_board.copy())
board_state = current_board.encode().copy()
root = UCT_search(current_board, args.num_reads_mcts, net)
policy = get_policy(root, t)
print(f"Game {idx} policy: {policy}")
col_choice = np.random.choice(np.array([0, 1, 2, 3, 4, 5, 6]),
p=policy)
current_board.drop_piece(col_choice) # move piece
dataset.append([board_state, policy])
print(f"[Iteration: {iteration}]: Game {idx} CURRENT BOARD:\n",
current_board)
move_count += 1
if current_board.check_winner(): # if somebody won
if current_board.player == PLAYER_1: # black wins
print("Black wins")
value = -1
elif current_board.player == PLAYER_2: # white wins
print("White wins")
value = 1
game_won = True
dataset_p = []
for idx, data in enumerate(dataset):
s, p = data
if idx == 0:
dataset_p.append([s, p, 0])
else:
dataset_p.append([s, p, value])
# Save the dataset
time_string = datetime.datetime.today().strftime("%Y-%m-%d")
pickle_file = f"iter_{iteration}/dataset_iter{iteration}_cpu{cpu_index}_{idx}_{time_string}"
util.pickle_save(pickle_file, dataset_p)