def test_tic_tac_toe(self):
game_spec = TicTacToeGameSpec()
create_model_func = functools.partial(create_network, game_spec.board_squares(), (100, 100, 100,))
variables, win_rate = train_policy_gradients(game_spec, create_model_func, None,
learn_rate=1e-4,
number_of_games=60000,
print_results_every=1000,
batch_size=100,
randomize_first_player=False)
self.assertGreater(win_rate, 0.4)
"""
import functools
from common.network_helpers import create_network
from games.tic_tac_toe import TicTacToeGameSpec
from techniques.train_policy_gradient import train_policy_gradients
BATCH_SIZE = 100 # every how many games to do a parameter update?
LEARN_RATE = 1e-4
PRINT_RESULTS_EVERY_X = 100 # every how many games to print the results
NETWORK_FILE_PATH = 'current_network.p' # path to save the network to
NUMBER_OF_GAMES_TO_RUN = 1000
# to play a different game change this to another spec, e.g TicTacToeXGameSpec or ConnectXGameSpec, to get these to run
# well may require tuning the hyper parameters a bit
game_spec = TicTacToeGameSpec()
create_network_func = functools.partial(create_network,
game_spec.board_squares(),
(100, 100, 100))
train_policy_gradients(game_spec,
create_network_func,
NETWORK_FILE_PATH,
number_of_games=NUMBER_OF_GAMES_TO_RUN,
batch_size=BATCH_SIZE,
learn_rate=LEARN_RATE,
print_results_every=PRINT_RESULTS_EVERY_X)
def second_player_move(board_state, side):
first_unvisited_node = False
current_side = -current_side
result = game_spec.has_winner(current_board_state)
for path_board_state, path_side in rollout_path:
state_samples[path_board_state] += 1.
result *= path_side
# normalize results to be between 0 and 1 before this it between -1 and 1
result /= 2.
result += .5
state_results[path_board_state] += result
move_states = {move: game_spec.apply_move(board_state, move, side) for move in game_spec.available_moves(board_state)}
move = max(move_states, key=lambda x: state_results[move_states[x]] / state_samples[move_states[x]])
return state_results[move_states[move]] / state_samples[move_states[move]], move
if __name__ == '__main__':
from games.tic_tac_toe import TicTacToeGameSpec
sample_board_state = ((1, 0, -1),
(1, 0, 0),
(0, -1, 0))
print(monte_carlo_tree_search_uct(TicTacToeGameSpec(), sample_board_state, -1, 10000))
© 2017 GitHub, Inc.
def setUp(self):
self._game_spec = TicTacToeGameSpec()