def test_deep_cfr_runs(self, game_name):
game = pyspiel.load_game(game_name)
deep_cfr_solver = deep_cfr.DeepCFRSolver(game,
policy_network_layers=(8, 4),
advantage_network_layers=(4,
2),
num_iterations=2,
num_traversals=2,
learning_rate=1e-3,
batch_size_advantage=None,
batch_size_strategy=None,
memory_capacity=1e7)
deep_cfr_solver.solve()
def test_matching_pennies_3p(self):
game = pyspiel.load_game_as_turn_based('matching_pennies_3p')
deep_cfr_solver = deep_cfr.DeepCFRSolver(game,
policy_network_layers=(16, 8),
advantage_network_layers=(32,
16),
num_iterations=2,
num_traversals=2,
learning_rate=1e-3,
batch_size_advantage=None,
batch_size_strategy=None,
memory_capacity=1e7)
deep_cfr_solver.solve()
conv = pyspiel.nash_conv(
game,
policy.python_policy_to_pyspiel_policy(
policy.tabular_policy_from_callable(
game, deep_cfr_solver.action_probabilities)))
logging.info('Deep CFR in Matching Pennies 3p. NashConv: %.2f', conv)
def main(unused_argv):
logging.info("Loading %s", FLAGS.game_name)
game = pyspiel.load_game(FLAGS.game_name)
deep_cfr_solver = deep_cfr.DeepCFRSolver(
game,
policy_network_layers=(32, 32),
advantage_network_layers=(16, 16),
num_iterations=FLAGS.num_iterations,
num_traversals=FLAGS.num_traversals,
learning_rate=1e-3,
batch_size_advantage=None,
batch_size_strategy=None,
memory_capacity=int(1e7))
_, advantage_losses, policy_loss = deep_cfr_solver.solve()
for player, losses in six.iteritems(advantage_losses):
logging.info("Advantage for player %d: %s", player,
losses[:2] + ["..."] + losses[-2:])
logging.info("Advantage Buffer Size for player %s: '%s'", player,
len(deep_cfr_solver.advantage_buffers[player]))
logging.info("Strategy Buffer Size: '%s'",
len(deep_cfr_solver.strategy_buffer))
logging.info("Final policy loss: '%s'", policy_loss)
average_policy = policy.tabular_policy_from_callable(
game, deep_cfr_solver.action_probabilities)
pyspiel_policy = policy.python_policy_to_pyspiel_policy(average_policy)
conv = pyspiel.nash_conv(game, pyspiel_policy)
logging.info("Deep CFR in '%s' - NashConv: %s", FLAGS.game_name, conv)
average_policy_values = expected_game_score.policy_value(
game.new_initial_state(), [average_policy] * 2)
logging.info("Computed player 0 value: %.2f (expected: %.2f).",
average_policy_values[0], -1 / 18)
logging.info("Computed player 1 value: %.2f (expected: %.2f).",
average_policy_values[1], 1 / 18)