def __init__(self,
config,
policy_params,
env_creator,
noise,
min_task_runtime=0.2):
self.min_task_runtime = min_task_runtime
self.config = config
self.policy_params = policy_params
self.noise = SharedNoiseTable(noise)
self.env = env_creator(config["env_config"])
from ray.rllib import models
self.preprocessor = models.ModelCatalog.get_preprocessor(self.env)
self.sess = utils.make_session(single_threaded=True)
if config["policy_type"] == "LinearPolicy":
self.policy = policies.LinearPolicy(
self.sess, self.env.action_space, self.preprocessor,
config["observation_filter"], **policy_params)
else:
self.policy = policies.MLPPolicy(
self.sess, self.env.action_space, self.preprocessor,
config["observation_filter"], config["fcnet_hiddens"],
**policy_params)
def _init(self):
policy_params = {"action_noise_std": 0.0}
# register the linear network
utils.register_linear_network()
env = self.env_creator(self.config["env_config"])
from ray.rllib import models
preprocessor = models.ModelCatalog.get_preprocessor(env)
self.sess = utils.make_session(single_threaded=False)
if self.config["policy_type"] == "LinearPolicy":
self.policy = policies.LinearPolicy(
self.sess, env.action_space, preprocessor,
self.config["observation_filter"], self.config["model"],
**policy_params)
else:
self.policy = policies.MLPPolicy(self.sess, env.action_space,
preprocessor,
self.config["observation_filter"],
self.config["model"],
self.config["fcnet_hiddens"],
**policy_params)
self.optimizer = optimizers.SGD(self.policy,
self.config["sgd_stepsize"])
self.rollouts_used = self.config["rollouts_used"]
self.num_rollouts = self.config["num_rollouts"]
self.report_length = self.config["report_length"]
# Create the shared noise table.
print("Creating shared noise table.")
noise_id = create_shared_noise.remote(self.config["noise_size"])
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
print("Creating actors.")
self.workers = [
Worker.remote(self.config, policy_params, self.env_creator,
noise_id) for _ in range(self.config["num_workers"])
]
self.episodes_so_far = 0
self.timesteps_so_far = 0
self.reward_list = []
self.tstart = time.time()