def test_net_vs_random(policy_fn, game_name, **kwargs):
game = pyspiel.load_game(game_name)
# Alphazero first
zero_bot = NeuralNetBot(game, 0, policy_fn)
random_bot = pyspiel.make_uniform_random_bot(game, 1, np.random.randint(0, 1000))
score1 = play_game(game, zero_bot, random_bot)
# Random bot first
zero_bot = NeuralNetBot(game, 1, policy_fn)
random_bot = pyspiel.make_uniform_random_bot(game, 0, np.random.randint(0, 1000))
score2 = -play_game(game, random_bot, zero_bot)
return score1, score2
def test_zero_vs_random(policy_fn):
game = pyspiel.load_game('connect_four')
# Alphazero first
zero_bot = AlphaZeroBot(game, 0, policy_fn=policy_fn, use_dirichlet=False)
random_bot = pyspiel.make_uniform_random_bot(game, 1, np.random.randint(0, 1000))
score1 = play_game(game, zero_bot, random_bot)
# Random bot first
zero_bot = AlphaZeroBot(game, 1, policy_fn=policy_fn, use_dirichlet=False)
random_bot = pyspiel.make_uniform_random_bot(game, 0, np.random.randint(0, 1000))
score2 = -play_game(game, random_bot, zero_bot)
return score1, score2, None
def main(unused_argv):
uct_c = 2
game = pyspiel.load_game(FLAGS.game)
state = game.new_initial_state()
print("Initial state: ", str(state))
# Check that the games satisfies the conditions for the implemented MCTS
# algorithm
if game.num_players() not in (1, 2):
raise ValueError("Game must be a 1-player game or 2-player zero-sum game")
if (game.num_players() == 2 and
game.get_type().utility != pyspiel.GameType.Utility.ZERO_SUM):
raise ValueError("Game must be a 1-player game or 2-player zero-sum game")
# Create MCTS bot
evaluator = mcts.RandomRolloutEvaluator(FLAGS.rollout_count)
mcts_bot = mcts.MCTSBot(game, FLAGS.mcts_player, uct_c,
FLAGS.max_search_nodes, evaluator)
# Create random bot
random_bot = pyspiel.make_uniform_random_bot(game, 1 - FLAGS.mcts_player, 123)
if FLAGS.mcts_player == 0:
bots = [mcts_bot, random_bot]
else:
bots = [random_bot, mcts_bot]
while not state.is_terminal():
# The state can be three different types: chance node,
# simultaneous node, or decision node
if state.is_chance_node():
# Chance node: sample an outcome
outcomes = state.chance_outcomes()
num_actions = len(outcomes)
print("Chance node, got " + str(num_actions) + " outcomes")
action_list, prob_list = zip(*outcomes)
action = np.random.choice(action_list, p=prob_list)
print("Sampled outcome: ",
state.action_to_string(state.current_player(), action))
state.apply_action(action)
elif state.is_simultaneous_node():
raise ValueError("Game cannot have simultaneous nodes.")
else:
# Decision node: sample action for the single current player
_, action = bots[state.current_player()].step(state)
print("Player ", state.current_player(), ", randomly sampled action: ",
state.action_to_string(state.current_player(), action))
state.apply_action(action)
print("Next state: ", str(state))
# Game is now done. Print return for each player
returns = state.returns()
for pid in range(game.num_players()):
print("Return for player {} is {}".format(pid, returns[pid]))
def test_python_and_cpp_bot(self):
game = pyspiel.load_game("kuhn_poker")
bots = [
pyspiel.make_uniform_random_bot(0, 1234),
uniform_random.UniformRandomBot(1, np.random.RandomState(4321)),
]
results = np.array([
pyspiel.evaluate_bots(game.new_initial_state(), bots, iteration)
for iteration in range(10000)
])
average_results = np.mean(results, axis=0)
np.testing.assert_allclose(average_results, [0.125, -0.125], atol=0.1)
class EvaluateBotsTest(parameterized.TestCase):
@parameterized.parameters([([
pyspiel.make_uniform_random_bot(0, 1234),
uniform_random.UniformRandomBot(1, np.random.RandomState(4321))
], ), (policy_bots(), )])
def test_cpp_vs_python(self, bots):
results = np.array([
evaluate_bots.evaluate_bots(GAME.new_initial_state(), bots,
np.random) for _ in range(10000)
])
average_results = np.mean(results, axis=0)
np.testing.assert_allclose(average_results, [0.125, -0.125], atol=0.1)
def test_random_vs_stateful(self):
game = pyspiel.load_game("tic_tac_toe")
bots = [
pyspiel.make_stateful_random_bot(game, 0, 1234),
uniform_random.UniformRandomBot(1, np.random.RandomState(4321))
]
for _ in range(1000):
evaluate_bots.evaluate_bots(game.new_initial_state(), bots,
np.random)