def get_ground_truth(self):
self.ground_truth_path = self.local_models_path / 'safety_ground_truth.npz'
load = self.ground_truth_path.exists()
if load:
try:
ground_truth = SafetyTruth.load(self.ground_truth_path,
self.env)
except ValueError:
load = False
if not load:
ground_truth = SafetyTruth(self.env)
ground_truth.compute()
ground_truth.save(self.ground_truth_path)
return ground_truth
def __init__(self, output_directory, name, envname, aname, envparams,
aparams, n_episodes, glie_start, safety_parameters_update_end,
reset_in_safe_state, metrics_sampling_frequency,
n_episodes_in_measurement, plot_every, seed):
self.env = ENV_CONSTRUCTOR[envname](**envparams)
self.agent = AGENT_CONSTRUCTOR[aname](env=self.env, **aparams)
safety_truth_path = SAFETY_TRUTH_PATH[envname]
if envname in SAFETY_TRUTH_FROM_VIBLY:
self.safety_truth = SafetyTruth(self.env)
self.safety_truth.from_vibly_file(safety_truth_path)
else:
self.safety_truth = SafetyTruth.load(safety_truth_path, self.env)
self.n_episodes = n_episodes
self.glie_start = glie_start if not isinstance(glie_start, float) else \
int(glie_start * self.n_episodes)
if safety_parameters_update_end is not None:
if isinstance(safety_parameters_update_end, float):
update_end = int(safety_parameters_update_end * n_episodes)
self.safety_parameters_update_end = update_end
else:
self.safety_parameters_update_end = safety_parameters_update_end
else:
self.safety_parameters_update_end = n_episodes
self.reset_in_safe_state = reset_in_safe_state
self.metrics_sampling_frequency = metrics_sampling_frequency
self.n_episodes_in_measurement = n_episodes_in_measurement
self.plot_every = plot_every
self.agent_has_safety_model = aname in HAS_SAFETY_MODEL
self.METRICS_NAMES = BenchmarkSingleSimulation.METRICS_BASE_NAMES
if self.agent_has_safety_model:
self.METRICS_NAMES += [
BenchmarkSingleSimulation.Q_C_Q_V_MNAME,
BenchmarkSingleSimulation.Q_V_Q_C_MNAME
]
plotters = {}
if envname in PLOTTABLE_Q:
if self.agent_has_safety_model:
plotters.update({
'Q-Values_Safety':
QValueAndSafetyPlotter(
self.agent,
self.safety_truth,
# ensure_in_dataset=True
)
})
else:
plotters.update({
'Q-Values':
QValuePlotter(
self.agent,
self.safety_truth,
write_values=False,
plot_samples=True,
)
})
super(BenchmarkSingleSimulation,
self).__init__(output_directory, name, plotters)
self.set_seed(value=seed)
self.metrics_path = self.output_directory / 'metrics'
self.metrics = AgentMetrics(*self.METRICS_NAMES)
simparams = {
'output_directory': output_directory,
'name': name,
'n_episodes': n_episodes,
'glie_start': glie_start,
'safety_parameters_update_end': safety_parameters_update_end,
'reset_in_safe_state': reset_in_safe_state,
'metrics_sampling_frequency': metrics_sampling_frequency,
'n_episodes_in_measurement': n_episodes_in_measurement,
'plot_every': plot_every,
}
logger.info(config_msg(f"Setting up simulation {name}"))
logger.info(config_msg(f"ENVIRONMENT: {envname}"))
logger.info(config_msg(str(envparams)))
logger.info(config_msg(f"AGENT: {aname}"))
logger.info(config_msg(str(aparams)))
logger.info(config_msg("SIMULATION:"))
logger.info(config_msg(str(simparams)))
