def test_learn_simple_game(self):
game_spec = _VerySimpleGameSpec()
create_model_func = functools.partial(create_network, 2, (4,))
variables, win_rate = train_policy_gradients(game_spec, create_model_func, None,
learn_rate=0.1,
number_of_games=1000, print_results_every=100,
batch_size=20,
randomize_first_player=False)
self.assertGreater(win_rate, 0.9)
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)
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):
return game_spec.flat_move_to_tuple(int(input("Next Move:")))
'''
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,
opponent_func=second_player_move,
number_of_games=config['number_of_games'],
update_opponent_winrate=config['update_opponent_winrate'],
print_results_every=config['print_results_every'],
learn_rate=config['learn_rate'],
batch_size=config['batch_size'],
cnn_on=config['cnn_on'],
eps=config['eps'],
deterministic=config['deterministic'],
mcts=config['mcts'],
min_win_ticks=config['min_win_ticks'],
beta=config['beta'])
else:
res = train_policy_gradients(
game_spec,
create_network_func,
load_network_file_path=config['load_network_file_path'],
number_of_games=config['number_of_games'],
batch_size=config['batch_size'],
learn_rate=config['learn_rate'],
print_results_every=config['print_results_every'],
save_network_file_path=config['save_network_file_path'],
cnn_on=config['cnn_on'],
eps=config['eps'],
deterministic=config['deterministic'],
mcts=config['mcts'],
beta=config['beta'])
config["results"] = res[2]
plt.save(config)
# pdb.set_trace()
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:
raise Exception, "Invalid value for --opponent"
train_policy_gradients(game_spec,
create_network_func,
network_file_path,
opponent_func=opponent_func,
number_of_games=args.num_games,
batch_size=args.batch_size,
learn_rate=args.learning_rate,
print_results_every=args.print_freq,
draw_reward=args.draw_reward)
from techniques.min_max import min_max_alpha_beta
from techniques.train_policy_gradient import train_policy_gradients
from connect_4.network import connect_4_game_spec, create_convolutional_network
def min_max_move_func(board_state, side):
return min_max_alpha_beta(connect_4_game_spec, board_state, side, 3)[1]
train_policy_gradients(
connect_4_game_spec,
create_convolutional_network,
'convolutional_net_5_4_l_c_4_f_1_other_after.p',
opponent_func=min_max_move_func,
save_network_file_path=
'convolutional_net_5_4_l_c_4_f_1_other_after_vs_depth_3.p',
number_of_games=5000,
print_results_every=100)