def test_cpp_python_best_response_oracle(self, game_name, num_players):
# Tests that these best responses interface well with Best Response Oracle
game = pyspiel.load_game(
game_name, {"players": pyspiel.GameParameter(num_players)})
all_states, _ = best_response.compute_states_and_info_states_if_none(
game, all_states=None, state_to_information_state=None)
current_best = [[policy.TabularPolicy(game).__copy__()]
for _ in range(num_players)]
probabilities_of_playing_policies = [[1.] for _ in range(num_players)]
# Construct the python oracle
py_oracle = best_response_oracle.BestResponseOracle(
best_response_backend="py")
# Construct the cpp oracle. Note that in this regime, BestResponseOracle
# uses base_policy to construct and cache TabularBestResponse internally.
cpp_oracle = best_response_oracle.BestResponseOracle(
game=game, best_response_backend="cpp")
# Prepare the computation of the best responses with each backend
# pylint:disable=g-complex-comprehension
training_params = [[{
"total_policies":
current_best,
"current_player":
i,
"probabilities_of_playing_policies":
probabilities_of_playing_policies
}] for i in range(num_players)]
# pylint:enable=g-complex-comprehension
py_best_rep = py_oracle(game, training_params)
cpp_best_rep = cpp_oracle(game, training_params)
# Compare the policies
for state in all_states.values():
i_player = state.current_player()
py_dict = py_best_rep[i_player][0].action_probabilities(state)
cpp_dict = cpp_best_rep[i_player][0].action_probabilities(state)
for action in py_dict.keys():
self.assertEqual(py_dict.get(action, 0.0),
cpp_dict.get(action, 0.0))
for action in cpp_dict.keys():
self.assertEqual(py_dict.get(action, 0.0),
cpp_dict.get(action, 0.0))
def init_br_responder(env):
"""Initializes the tabular best-response based responder and agents."""
random_policy = policy.TabularPolicy(env.game)
oracle = best_response_oracle.BestResponseOracle(game=env.game,
policy=random_policy)
agents = [random_policy.__copy__() for _ in range(FLAGS.n_players)]
return oracle, agents