def evaluate_st(num_episodes=1000):
aggregate_stats = control.evaluate_control(
do_st_control,
num_episodes=num_episodes,
state_function=prediction.HighwayState.from_sumo,
verbose=True)
aggregate_stats.print_stats()
def evaluate_st_and_dump_crash(num_episodes=1000):
aggregate_stats = control.evaluate_control(
do_st_control,
num_episodes,
state_function=prediction.HighwayState.from_sumo,
crash_callback=plot_crash,
verbose=True,
save_state_on_crash=True)
aggregate_stats.print_stats()
def evaluate(self, num_episodes):
self._setup()
output = control.evaluate_control(
control_function=self.do_control,
state_function=prediction.HighwayState.from_sumo,
crash_callback=None,
num_episodes=num_episodes,
end_episode_callback=self.end_episode_callback,
verbose=True)
output.print_stats()
self._cleanup()
def evaluate(model_name, num_iterations):
dqn = DQN.load(model_name)
control_function = partial(do_dqn_control, dqn=dqn, epsilon=0.0)
reward_function = get_reward_function()
evaluation_stats = control.evaluate_control(
control_function=control_function,
state_function=prediction.HighwayState.from_sumo,
custom_stats_function=partial(rl.get_rl_custom_episode_stats,
reward_function=reward_function),
max_episode_length=Settings.EVALUATION_EPISODE_LENGTH,
num_episodes=num_iterations,
wait_before_start=20)
evaluation_stats.print_stats()
def evaluate_combined(self, num_episodes):
self._setup()
output = control.evaluate_control(
control_function=self.do_combined_control,
state_function=prediction.HighwayState.from_sumo,
crash_callback=None,
end_episode_callback=self.end_episode_callback,
num_episodes=num_episodes,
custom_stats_function=self.combined_stats_callback,
verbose=True)
self._cleanup()
output.print_stats()
self.plot_st_proportion()
def evaluate_q_model_and_log_metrics(dqn, iteration, writer, reward_function):
sumo.change_step_size(Settings.EVALUATION_TICK_LENGTH)
control_function = partial(do_dqn_control, dqn=dqn, epsilon=0.0)
evaluation_stats = control.evaluate_control(
control_function=control_function,
state_function=prediction.HighwayState.from_sumo,
custom_stats_function=partial(rl.get_rl_custom_episode_stats,
reward_function=reward_function),
max_episode_length=Settings.EVALUATION_EPISODE_LENGTH,
num_episodes=Settings.NUM_EVALUATION_EPISODES,
wait_before_start=20)
metrics = evaluation_stats.get_stat_averages()
for metric in metrics:
writer.add_scalar(metric, metrics[metric], iteration)
logging.info(metrics)
sumo.change_step_size(Settings.TRAINING_TICK_LENGTH)
def evaluate_q_model_and_log_metrics(Q,
iteration,
writer,
visits,
reward_function=slotted_reward):
sumo.change_step_size(control.Settings.EVALUATION_TICK_LENGTH)
control_function = partial(get_control, Q=Q, visits=visits, epsilon=0.0)
evaluation_stats = control.evaluate_control(
control_function=control_function,
state_function=get_state_tuple,
custom_stats_function=partial(get_rl_custom_episode_stats,
reward_function=reward_function),
max_episode_length=control.Settings.EVALUATION_EPISODE_LENGTH,
num_episodes=control.Settings.NUM_EVALUATION_EPISODES,
wait_before_start=20)
metrics = evaluation_stats.get_stat_averages()
metrics["unvisited_states"] = visits.size - np.count_nonzero(visits)
for metric in metrics:
writer.add_scalar(metric, metrics[metric], iteration)
logging.info(metrics)
sumo.change_step_size(control.Settings.TRAINING_TICK_LENGTH)