class TestCreatePositionsSet(TestCase):
def setUp(self):
self._game_spec = TicTacToeGameSpec()
def test_create_positions(self):
number_of_positions = 100
positions = create_positions_set(
self._game_spec, number_of_positions,
self._game_spec.get_random_player_func())
self.assertGreater(len(positions), number_of_positions - 1)
if not args.hidden_layers:
args.hidden_layers = (100, 100, 100)
# create_network_func = functools.partial(create_network, game_spec.board_squares(), (100, 100, 100))
create_network_func = functools.partial(create_network,
game_spec.board_squares(),
args.hidden_layers)
network_file_path = 'current_network'
for n in args.hidden_layers:
network_file_path = network_file_path + ("_%05d" % n)
network_file_path = network_file_path + ".p"
random_opponent = game_spec.get_random_player_func()
perfect_opponent = game_spec.get_perfect_player()
def mixed_opponent(*args, **kwds):
opponent = random.choice([random_opponent, perfect_opponent])
return opponent(*args, **kwds)
if args.opponent == "random":
opponent_func = random_opponent
elif args.opponent == "perfect":
opponent_func = perfect_opponent
elif args.opponent == "mixed":
opponent_func = mixed_opponent
else:
mini_batch_board_states, mini_batch_moves, mini_batch_rewards = [], [], []
results = collections.deque(maxlen=PRINT_RESULTS_EVERY_X)
def make_training_move(board_state, side):
mini_batch_board_states.append(np.ravel(board_state) * side)
move = get_stochastic_network_move(session, input_layer, output_layer,
board_state, side)
mini_batch_moves.append(move)
return game_spec.flat_move_to_tuple(move.argmax())
for episode_number in range(1, NUMBER_OF_GAMES_TO_RUN):
# randomize if going first or second
if bool(random.getrandbits(1)):
reward = game_spec.play_game(make_training_move,
game_spec.get_random_player_func())
else:
reward = -game_spec.play_game(game_spec.get_random_player_func(),
make_training_move)
results.append(reward)
last_game_length = len(mini_batch_board_states) - len(
mini_batch_rewards)
# we scale here so winning quickly is better winning slowly and loosing slowly better than loosing quick
reward /= float(last_game_length)
mini_batch_rewards += ([reward] * last_game_length)
if episode_number % BATCH_SIZE == 0:
