def test_can_keep_track_of_window_of_winrate_for_learning_policy(RPS_task):
psro = PSRONashResponse(task=RPS_task, match_outcome_rolling_window_size=3)
training_agent_indeces = [1, 1, 0, 1]
expected_rolling_window = [1, 0, 1]
# TODO this is very ugly. It always chooses player 2 (1-index) as winner
# We should really find a way of mocking this.
sample_trajectory = [([], [], [0, 1], [])] # (s, a, r, s')
for i in training_agent_indeces:
psro.update_rolling_winrates(episode_trajectory=sample_trajectory,
training_agent_index=i)
np.testing.assert_array_equal(expected_rolling_window,
psro.match_outcome_rolling_window)
def test_can_keep_track_of_window_of_winrate_for_learning_policy(RPS_task):
psro = PSRONashResponse(task=RPS_task, match_outcome_rolling_window_size=3)
training_agent_indeces = [1, 1, 0, 1]
expected_rolling_window = [1, 0, 1]
# TODO this is very ugly. It always chooses player 2 (1-index) as winner
# We should really find a way of mocking this.
sample_trajectory = Trajectory(
env_type=EnvType.MULTIAGENT_SIMULTANEOUS_ACTION, num_agents=2)
sample_trajectory.add_timestep(None, None, [0, 1], None, True)
for i in training_agent_indeces:
psro.update_rolling_winrates(episode_trajectory=sample_trajectory,
training_agent_index=i)
np.testing.assert_array_equal(expected_rolling_window,
psro.match_outcome_rolling_window)