def __init__(self, env_name, config, upload_dir=None):
config.update({"alg": "EvolutionStrategies"})
Algorithm.__init__(self, env_name, config, upload_dir=upload_dir)
policy_params = {
"ac_noise_std": 0.01
}
env = gym.make(env_name)
utils.make_session(single_threaded=False)
self.policy = policies.GenericPolicy(
env.observation_space, env.action_space, **policy_params)
tf_util.initialize()
self.optimizer = optimizers.Adam(self.policy, config["stepsize"])
self.ob_stat = utils.RunningStat(env.observation_space.shape, eps=1e-2)
# Create the shared noise table.
print("Creating shared noise table.")
noise_id = create_shared_noise.remote()
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
print("Creating actors.")
self.workers = [
Worker.remote(config, policy_params, env_name, noise_id)
for _ in range(config["num_workers"])]
self.episodes_so_far = 0
self.timesteps_so_far = 0
self.tstart = time.time()
self.iteration = 0
def __init__(self, env_name, config):
Algorithm.__init__(self, env_name, config)
policy_params = {
"ac_bins": "continuous:",
"ac_noise_std": 0.01,
"nonlin_type": "tanh",
"hidden_dims": [256, 256],
"connection_type": "ff"
}
# Create the shared noise table.
print("Creating shared noise table.")
noise_id = create_shared_noise.remote()
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
print("Creating actors.")
self.workers = [
Worker.remote(config, policy_params, env_name, noise_id)
for _ in range(config.num_workers)
]
env = gym.make(env_name)
utils.make_session(single_threaded=False)
self.policy = policies.MujocoPolicy(env.observation_space,
env.action_space, **policy_params)
tf_util.initialize()
self.optimizer = optimizers.Adam(self.policy, config.stepsize)
self.ob_stat = utils.RunningStat(env.observation_space.shape, eps=1e-2)
self.episodes_so_far = 0
self.timesteps_so_far = 0
self.tstart = time.time()
self.iteration = 0
def __init__(self,
config,
policy_params,
env_name,
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 = gym.make(env_name)
self.preprocessor = ModelCatalog.get_preprocessor(
env_name, self.env.observation_space.shape)
self.preprocessor_shape = self.preprocessor.transform_shape(
self.env.observation_space.shape)
self.sess = utils.make_session(single_threaded=True)
self.policy = policies.GenericPolicy(self.env.observation_space,
self.env.action_space,
self.preprocessor,
**policy_params)
tf_util.initialize()
self.rs = np.random.RandomState()
assert (self.policy.needs_ob_stat == (self.config["calc_obstat_prob"]
!= 0))
def __init__(self,
config,
policy_params,
env_name,
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 = gym.make(env_name)
self.sess = utils.make_session(single_threaded=True)
self.policy = policies.MujocoPolicy(self.env.observation_space,
self.env.action_space,
**policy_params)
tf_util.initialize()
self.rs = np.random.RandomState()
assert self.policy.needs_ob_stat == (self.config.calc_obstat_prob != 0)