def __init__(self,
args,
observation_size,
action_size,
task_q,
result_q,
actor_id,
monitor,
name_scope='policy-actor',
init_filter_parameters=None):
# the base agent
super(rollout_agent, self).__init__(args=args,
observation_size=observation_size,
action_size=action_size,
task_q=task_q,
result_q=result_q,
name_scope=name_scope)
self.allow_monitor = monitor
self.actor_id = actor_id
self._npr = np.random.RandomState(args.seed + actor_id)
if init_filter_parameters is not None:
self.ob_normalizer = ob_normalizer.normalizer(
mean=init_filter_parameters['mean'],
variance=init_filter_parameters['variance'],
num_steps=init_filter_parameters['step'])
else:
self.ob_normalizer = ob_normalizer.normalizer()
logger.info('The sampler {} is online'.format(self.actor_id))
def __init__(self,
args,
task_q,
result_q,
agent_id,
name_scope='evolutionary_agent'):
# the base agent
self.agent_id = agent_id
base_agent.__init__(self,
args=args,
observation_size=-1,
action_size=-1,
task_q=task_q,
result_q=result_q,
name_scope=name_scope)
self.reset_running_mean_info()
# evolutionary agent will never load from ckpt_name
# self.load_running_means()
self.ob_normalizer = ob_normalizer.normalizer()
# the variables and networks to be used, init them before use them
self.baseline_network = None
self.env_info = None
self.error_count = 0
# used to save the checkpoint files
self.best_reward = -np.inf
self.timesteps_so_far = 0
self._npr = np.random.RandomState(args.seed + self.agent_id)
self.debug = 0
# self.last_save_iteration = 0
self.last_average_reward = np.nan
self.end_average_reward = np.nan
self.start_time = None
self.is_dead_species = False
self.brute_search_reward = []
if self.args.nervenetplus:
if self.args.fc_pruning:
assert self.args.use_nervenet
else:
assert self.args.use_nervenet and self.args.use_gnn_as_policy