def step_attempt(self):
config = self.config
theta = self.policy.get_flat_weights()
assert theta.dtype == np.float32
assert len(theta.shape) == 1
# Put the current policy weights in the object store.
theta_id = ray.put(theta)
# Use the actors to do rollouts, note that we pass in the ID of the
# policy weights.
results, num_episodes, num_timesteps = self._collect_results(
theta_id, config["num_rollouts"]
)
# Update our sample steps counters.
self._counters[NUM_AGENT_STEPS_SAMPLED] += num_timesteps
self._counters[NUM_ENV_STEPS_SAMPLED] += num_timesteps
all_noise_indices = []
all_training_returns = []
all_training_lengths = []
all_eval_returns = []
all_eval_lengths = []
# Loop over the results.
for result in results:
all_eval_returns += result.eval_returns
all_eval_lengths += result.eval_lengths
all_noise_indices += result.noise_indices
all_training_returns += result.noisy_returns
all_training_lengths += result.noisy_lengths
assert len(all_eval_returns) == len(all_eval_lengths)
assert (
len(all_noise_indices)
== len(all_training_returns)
== len(all_training_lengths)
)
self.episodes_so_far += num_episodes
# Assemble the results.
eval_returns = np.array(all_eval_returns)
eval_lengths = np.array(all_eval_lengths)
noise_indices = np.array(all_noise_indices)
noisy_returns = np.array(all_training_returns)
noisy_lengths = np.array(all_training_lengths)
# keep only the best returns
# select top performing directions if rollouts_used < num_rollouts
max_rewards = np.max(noisy_returns, axis=1)
if self.rollouts_used > self.num_rollouts:
self.rollouts_used = self.num_rollouts
percentile = 100 * (1 - (self.rollouts_used / self.num_rollouts))
idx = np.arange(max_rewards.size)[
max_rewards >= np.percentile(max_rewards, percentile)
]
noise_idx = noise_indices[idx]
noisy_returns = noisy_returns[idx, :]
# Compute and take a step.
g, count = utils.batched_weighted_sum(
noisy_returns[:, 0] - noisy_returns[:, 1],
(self.noise.get(index, self.policy.num_params) for index in noise_idx),
batch_size=min(500, noisy_returns[:, 0].size),
)
g /= noise_idx.size
# scale the returns by their standard deviation
if not np.isclose(np.std(noisy_returns), 0.0):
g /= np.std(noisy_returns)
assert g.shape == (self.policy.num_params,) and g.dtype == np.float32
# Compute the new weights theta.
theta, update_ratio = self.optimizer.update(-g)
# Update our train steps counters.
self._counters[NUM_AGENT_STEPS_TRAINED] += num_timesteps
self._counters[NUM_ENV_STEPS_TRAINED] += num_timesteps
# Set the new weights in the local copy of the policy.
self.policy.set_flat_weights(theta)
# update the reward list
if len(all_eval_returns) > 0:
self.reward_list.append(eval_returns.mean())
# Now sync the filters
FilterManager.synchronize(
{DEFAULT_POLICY_ID: self.policy.observation_filter}, self.workers
)
info = {
"weights_norm": np.square(theta).sum(),
"weights_std": np.std(theta),
"grad_norm": np.square(g).sum(),
"update_ratio": update_ratio,
"episodes_this_iter": noisy_lengths.size,
"episodes_so_far": self.episodes_so_far,
}
result = dict(
episode_reward_mean=np.mean(self.reward_list[-self.report_length :]),
episode_len_mean=eval_lengths.mean(),
timesteps_this_iter=noisy_lengths.sum(),
info=info,
)
return result
def step_attempt(self):
config = self.config
theta = self.policy.get_flat_weights()
assert theta.dtype == np.float32
assert len(theta.shape) == 1
# Put the current policy weights in the object store.
theta_id = ray.put(theta)
# Use the actors to do rollouts. Note that we pass in the ID of the
# policy weights as these are shared.
results, num_episodes, num_timesteps = self._collect_results(
theta_id, config["episodes_per_batch"], config["train_batch_size"])
# Update our sample steps counters.
self._counters[NUM_AGENT_STEPS_SAMPLED] += num_timesteps
self._counters[NUM_ENV_STEPS_SAMPLED] += num_timesteps
all_noise_indices = []
all_training_returns = []
all_training_lengths = []
all_eval_returns = []
all_eval_lengths = []
# Loop over the results.
for result in results:
all_eval_returns += result.eval_returns
all_eval_lengths += result.eval_lengths
all_noise_indices += result.noise_indices
all_training_returns += result.noisy_returns
all_training_lengths += result.noisy_lengths
assert len(all_eval_returns) == len(all_eval_lengths)
assert (len(all_noise_indices) == len(all_training_returns) ==
len(all_training_lengths))
self.episodes_so_far += num_episodes
# Assemble the results.
eval_returns = np.array(all_eval_returns)
eval_lengths = np.array(all_eval_lengths)
noise_indices = np.array(all_noise_indices)
noisy_returns = np.array(all_training_returns)
noisy_lengths = np.array(all_training_lengths)
# Process the returns.
proc_noisy_returns = utils.compute_centered_ranks(noisy_returns)
# Compute and take a step.
g, count = utils.batched_weighted_sum(
proc_noisy_returns[:, 0] - proc_noisy_returns[:, 1],
(self.noise.get(index, self.policy.num_params)
for index in noise_indices),
batch_size=500,
)
g /= noisy_returns.size
assert (g.shape == (self.policy.num_params, ) and g.dtype == np.float32
and count == len(noise_indices))
# Compute the new weights theta.
theta, update_ratio = self.optimizer.update(-g +
config["l2_coeff"] * theta)
# Update our train steps counters.
self._counters[NUM_AGENT_STEPS_TRAINED] += num_timesteps
self._counters[NUM_ENV_STEPS_TRAINED] += num_timesteps
# Set the new weights in the local copy of the policy.
self.policy.set_flat_weights(theta)
# Store the rewards
if len(all_eval_returns) > 0:
self.reward_list.append(np.mean(eval_returns))
# Now sync the filters
FilterManager.synchronize(
{DEFAULT_POLICY_ID: self.policy.observation_filter}, self.workers)
info = {
"weights_norm": np.square(theta).sum(),
"grad_norm": np.square(g).sum(),
"update_ratio": update_ratio,
"episodes_this_iter": noisy_lengths.size,
"episodes_so_far": self.episodes_so_far,
}
reward_mean = np.mean(self.reward_list[-self.report_length:])
result = dict(
episode_reward_mean=reward_mean,
episode_len_mean=eval_lengths.mean(),
timesteps_this_iter=noisy_lengths.sum(),
info=info,
)
return result