class PPOTrainer(Trainer):
"""The PPOTrainer is an implementation of the PPO algorithm."""
def __init__(self, sess, brain, reward_buff_cap, trainer_parameters,
training, seed, run_id):
"""
Responsible for collecting experiences and training PPO model.
:param sess: Tensorflow session.
:param trainer_parameters: The parameters for the trainer (dictionary).
:param training: Whether the trainer is set for training.
"""
super(PPOTrainer, self).__init__(sess, brain.brain_name,
trainer_parameters, training, run_id)
self.param_keys = [
'batch_size', 'beta', 'buffer_size', 'epsilon', 'gamma',
'hidden_units', 'lambd', 'learning_rate', 'max_steps', 'normalize',
'num_epoch', 'num_layers', 'time_horizon', 'sequence_length',
'summary_freq', 'use_recurrent', 'graph_scope', 'summary_path',
'memory_size', 'use_curiosity', 'curiosity_strength',
'curiosity_enc_size'
]
for k in self.param_keys:
if k not in trainer_parameters:
raise UnityTrainerException(
"The hyperparameter {0} could not be found for the PPO trainer of "
"brain {1}.".format(k, brain.brain_name))
self.use_curiosity = bool(trainer_parameters['use_curiosity'])
self.step = 0
self.policy = PPOPolicy(seed, brain, trainer_parameters, sess,
self.is_training)
stats = {
'cumulative_reward': [],
'episode_length': [],
'value_estimate': [],
'entropy': [],
'value_loss': [],
'policy_loss': [],
'learning_rate': []
}
if self.use_curiosity:
stats['forward_loss'] = []
stats['inverse_loss'] = []
stats['intrinsic_reward'] = []
self.intrinsic_rewards = {}
self.stats = stats
self.training_buffer = Buffer()
self.training_buffer2 = Buffer()
self.cumulative_rewards = {}
self._reward_buffer = deque(maxlen=reward_buff_cap)
self.episode_steps = {}
self.summary_path = trainer_parameters['summary_path']
if not os.path.exists(self.summary_path):
os.makedirs(self.summary_path)
self.summary_writer = tf.summary.FileWriter(self.summary_path)
def __str__(self):
return '''Hyperparameters for the PPO Trainer of brain {0}: \n{1}'''.format(
self.brain_name, '\n'.join([
'\t{0}:\t{1}'.format(x, self.trainer_parameters[x])
for x in self.param_keys
]))
@property
def parameters(self):
"""
Returns the trainer parameters of the trainer.
"""
return self.trainer_parameters
@property
def get_max_steps(self):
"""
Returns the maximum number of steps. Is used to know when the trainer should be stopped.
:return: The maximum number of steps of the trainer
"""
return float(self.trainer_parameters['max_steps'])
@property
def get_step(self):
"""
Returns the number of steps the trainer has performed
:return: the step count of the trainer
"""
return self.step
@property
def reward_buffer(self):
"""
Returns the reward buffer. The reward buffer contains the cumulative
rewards of the most recent episodes completed by agents using this
trainer.
:return: the reward buffer.
"""
return self._reward_buffer
def increment_step_and_update_last_reward(self):
"""
Increment the step count of the trainer and Updates the last reward
"""
if len(self.stats['cumulative_reward']) > 0:
mean_reward = np.mean(self.stats['cumulative_reward'])
self.policy.update_reward(mean_reward)
self.policy.increment_step()
self.step = self.policy.get_current_step()
def take_action(self, all_brain_info: AllBrainInfo):
"""
Decides actions given observations information, and takes them in environment.
:param all_brain_info: A dictionary of brain names and BrainInfo from environment.
:return: a tuple containing action, memories, values and an object
to be passed to add experiences
"""
curr_brain_info = all_brain_info[self.brain_name]
if len(curr_brain_info.agents) == 0:
return [], [], [], None, None
run_out = self.policy.evaluate(curr_brain_info)
self.stats['value_estimate'].append(run_out['value'].mean())
self.stats['entropy'].append(run_out['entropy'].mean())
self.stats['learning_rate'].append(run_out['learning_rate'])
if self.policy.use_recurrent:
return run_out['action'], run_out['memory_out'], None, \
run_out['value'], run_out
else:
return run_out['action'], None, None, run_out['value'], run_out
def construct_curr_info(self, next_info: BrainInfo) -> BrainInfo:
"""
Constructs a BrainInfo which contains the most recent previous experiences for all agents info
which correspond to the agents in a provided next_info.
:BrainInfo next_info: A t+1 BrainInfo.
:return: curr_info: Reconstructed BrainInfo to match agents of next_info.
"""
visual_observations = [[]]
vector_observations = []
text_observations = []
memories = []
rewards = []
local_dones = []
max_reacheds = []
agents = []
prev_vector_actions = []
prev_text_actions = []
for agent_id in next_info.agents:
agent_brain_info = self.training_buffer[agent_id].last_brain_info
if agent_brain_info is None:
agent_brain_info = next_info
agent_index = agent_brain_info.agents.index(agent_id)
for i in range(len(next_info.visual_observations)):
visual_observations[i].append(
agent_brain_info.visual_observations[i][agent_index])
vector_observations.append(
agent_brain_info.vector_observations[agent_index])
text_observations.append(
agent_brain_info.text_observations[agent_index])
if self.policy.use_recurrent:
if len(agent_brain_info.memories > 0):
memories.append(agent_brain_info.memories[agent_index])
else:
memories.append(self.policy.make_empty_memory(1))
rewards.append(agent_brain_info.rewards[agent_index])
local_dones.append(agent_brain_info.local_done[agent_index])
max_reacheds.append(agent_brain_info.max_reached[agent_index])
agents.append(agent_brain_info.agents[agent_index])
prev_vector_actions.append(
agent_brain_info.previous_vector_actions[agent_index])
prev_text_actions.append(
agent_brain_info.previous_text_actions[agent_index])
if self.policy.use_recurrent:
memories = np.vstack(memories)
curr_info = BrainInfo(visual_observations, vector_observations,
text_observations, memories, rewards, agents,
local_dones, prev_vector_actions,
prev_text_actions, max_reacheds)
return curr_info
def construct_curr_info2(self, next_info: BrainInfo) -> BrainInfo:
"""
Constructs a BrainInfo which contains the most recent previous experiences for all agents info
which correspond to the agents in a provided next_info.
:BrainInfo next_info: A t+1 BrainInfo.
:return: curr_info: Reconstructed BrainInfo to match agents of next_info.
"""
visual_observations = [[]]
vector_observations2 = []
text_observations = []
memories = []
rewards = []
local_dones = []
max_reacheds = []
agents = []
prev_vector_actions = []
prev_text_actions = []
for agent_id in next_info.agents:
agent_brain_info = self.training_buffer[agent_id].last_brain_info
agent_brain_info2 = next_info
agent_index1 = agent_brain_info.agents.index(agent_id)
agent_index2 = agent_brain_info2.agents.index(agent_id)
for i in range(len(next_info.visual_observations)):
visual_observations[i].append(
agent_brain_info.visual_observations[i][agent_index])
vector_observations2.append(
agent_brain_info.vector_observations[agent_index1],
agent_brain_info2.vector_observations[agent_index2])
text_observations.append(
agent_brain_info.text_observations[agent_index])
if self.policy.use_recurrent:
if len(agent_brain_info.memories > 0):
memories.append(agent_brain_info.memories[agent_index])
else:
memories.append(self.policy.make_empty_memory(1))
rewards.append(agent_brain_info.rewards[agent_index])
local_dones.append(agent_brain_info.local_done[agent_index])
max_reacheds.append(agent_brain_info.max_reached[agent_index])
agents.append(agent_brain_info.agents[agent_index])
prev_vector_actions.append(
agent_brain_info.previous_vector_actions[agent_index])
prev_text_actions.append(
agent_brain_info.previous_text_actions[agent_index])
if self.policy.use_recurrent:
memories = np.vstack(memories)
curr_info2 = BrainInfo(visual_observations, vector_observations2,
text_observations, memories, rewards, agents,
local_dones, prev_vector_actions,
prev_text_actions, max_reacheds)
return curr_info2
def add_experiences(self, curr_all_info: AllBrainInfo,
next_all_info: AllBrainInfo, take_action_outputs):
"""
Adds experiences to each agent's experience history.
:param curr_all_info: Dictionary of all current brains and corresponding BrainInfo.
:param next_all_info: Dictionary of all current brains and corresponding BrainInfo.
:param take_action_outputs: The outputs of the take action method.
"""
curr_info = curr_all_info[self.brain_name]
curr_info2 = curr_all_info[self.brain_name]
next_info = next_all_info[self.brain_name]
for agent_id in curr_info.agents:
self.training_buffer[agent_id].last_brain_info = curr_info
self.training_buffer[
agent_id].last_take_action_outputs = take_action_outputs
for agent_id in curr_info2.agents:
self.training_buffer2[agent_id].last_brain_info = curr_info2
self.training_buffer2[
agent_id].last_take_action_outputs = take_action_outputs
if curr_info.agents != next_info.agents:
curr_to_use = self.construct_curr_info(next_info)
else:
curr_to_use = curr_info
intrinsic_rewards = self.policy.get_intrinsic_rewards(
curr_to_use, next_info)
for agent_id in next_info.agents:
stored_info = self.training_buffer[agent_id].last_brain_info
stored_take_action_outputs = self.training_buffer[
agent_id].last_take_action_outputs
stored_take_action_outputs2 = self.training_buffer2[
agent_id].last_take_action_outputs
if stored_info is not None:
idx = stored_info.agents.index(agent_id)
next_idx = next_info.agents.index(agent_id)
if not stored_info.local_done[idx]:
for i, _ in enumerate(stored_info.visual_observations):
self.training_buffer[agent_id][
'visual_obs%d' % i].append(
stored_info.visual_observations[i][idx])
self.training_buffer[agent_id][
'next_visual_obs%d' % i].append(
next_info.visual_observations[i][next_idx])
if self.policy.use_vec_obs:
self.training_buffer[agent_id]['vector_obs'].append(
stored_info.vector_observations[idx])
self.training_buffer[agent_id][
'next_vector_in'].append(
next_info.vector_observations[next_idx])
if self.policy.use_recurrent:
if stored_info.memories.shape[1] == 0:
stored_info.memories = np.zeros(
(len(stored_info.agents), self.policy.m_size))
self.training_buffer[agent_id]['memory'].append(
stored_info.memories[idx])
actions = stored_take_action_outputs['action']
if self.policy.use_continuous_act:
actions_pre = stored_take_action_outputs['pre_action']
self.training_buffer[agent_id]['actions_pre'].append(
actions_pre[idx])
else:
self.training_buffer[agent_id]['action_mask'].append(
stored_info.action_masks[idx])
a_dist = stored_take_action_outputs['log_probs']
value1 = stored_take_action_outputs['value']
value2 = stored_take_action_outputs2['value']
if sum(value1) > sum(value2):
value = value1
else:
value = value2
self.training_buffer[agent_id]['actions'].append(
actions[idx])
self.training_buffer[agent_id]['prev_action'].append(
stored_info.previous_vector_actions[idx])
self.training_buffer[agent_id]['masks'].append(1.0)
if self.use_curiosity:
self.training_buffer[agent_id]['rewards'].append(
next_info.rewards[next_idx] +
intrinsic_rewards[next_idx])
else:
self.training_buffer[agent_id]['rewards'].append(
next_info.rewards[next_idx])
self.training_buffer[agent_id]['action_probs'].append(
a_dist[idx])
self.training_buffer[agent_id]['value_estimates'].append(
value[idx][0])
if agent_id not in self.cumulative_rewards:
self.cumulative_rewards[agent_id] = 0
self.cumulative_rewards[agent_id] += next_info.rewards[
next_idx]
if self.use_curiosity:
if agent_id not in self.intrinsic_rewards:
self.intrinsic_rewards[agent_id] = 0
self.intrinsic_rewards[agent_id] += intrinsic_rewards[
next_idx]
if not next_info.local_done[next_idx]:
if agent_id not in self.episode_steps:
self.episode_steps[agent_id] = 0
self.episode_steps[agent_id] += 1
def process_experiences(self, current_info: AllBrainInfo,
new_info: AllBrainInfo):
"""
Checks agent histories for processing condition, and processes them as necessary.
Processing involves calculating value and advantage targets for model updating step.
:param current_info: Dictionary of all current brains and corresponding BrainInfo.
:param new_info: Dictionary of all next brains and corresponding BrainInfo.
"""
info = new_info[self.brain_name]
for l in range(len(info.agents)):
agent_actions = self.training_buffer[info.agents[l]]['actions']
if ((info.local_done[l] or
len(agent_actions) > self.trainer_parameters['time_horizon'])
and len(agent_actions) > 0):
agent_id = info.agents[l]
if info.local_done[l] and not info.max_reached[l]:
value_next = 0.0
else:
if info.max_reached[l]:
bootstrapping_info = self.training_buffer[
agent_id].last_brain_info
idx = bootstrapping_info.agents.index(agent_id)
else:
bootstrapping_info = info
idx = l
value_next = self.policy.get_value_estimate(
bootstrapping_info, idx)
self.training_buffer[agent_id]['advantages'].set(
get_gae(rewards=self.training_buffer[agent_id]
['rewards'].get_batch(),
value_estimates=self.training_buffer[agent_id]
['value_estimates'].get_batch(),
value_next=value_next,
gamma=self.trainer_parameters['gamma'],
lambd=self.trainer_parameters['lambd']))
self.training_buffer[agent_id]['discounted_returns'].set(
self.training_buffer[agent_id]['advantages'].get_batch() +
self.training_buffer[agent_id]
['value_estimates'].get_batch())
self.training_buffer.append_update_buffer(
agent_id,
batch_size=None,
training_length=self.policy.sequence_length)
self.training_buffer[agent_id].reset_agent()
if info.local_done[l]:
self.stats['cumulative_reward'].append(
self.cumulative_rewards.get(agent_id, 0))
self.reward_buffer.appendleft(
self.cumulative_rewards.get(agent_id, 0))
self.stats['episode_length'].append(
self.episode_steps.get(agent_id, 0))
self.cumulative_rewards[agent_id] = 0
self.episode_steps[agent_id] = 0
if self.use_curiosity:
self.stats['intrinsic_reward'].append(
self.intrinsic_rewards.get(agent_id, 0))
self.intrinsic_rewards[agent_id] = 0
def end_episode(self):
"""
A signal that the Episode has ended. The buffer must be reset.
Get only called when the academy resets.
"""
self.training_buffer.reset_all()
for agent_id in self.cumulative_rewards:
self.cumulative_rewards[agent_id] = 0
for agent_id in self.episode_steps:
self.episode_steps[agent_id] = 0
if self.use_curiosity:
for agent_id in self.intrinsic_rewards:
self.intrinsic_rewards[agent_id] = 0
def is_ready_update(self):
"""
Returns whether or not the trainer has enough elements to run update model
:return: A boolean corresponding to whether or not update_model() can be run
"""
size_of_buffer = len(self.training_buffer.update_buffer['actions'])
return size_of_buffer > max(
int(self.trainer_parameters['buffer_size'] /
self.policy.sequence_length), 1)
def update_policy(self):
"""
Uses training_buffer to update the policy.
"""
n_sequences = max(
int(self.trainer_parameters['batch_size'] /
self.policy.sequence_length), 1)
value_total, policy_total, forward_total, inverse_total = [], [], [], []
advantages = self.training_buffer.update_buffer[
'advantages'].get_batch()
self.training_buffer.update_buffer['advantages'].set(
(advantages - advantages.mean()) / (advantages.std() + 1e-10))
num_epoch = self.trainer_parameters['num_epoch']
for k in range(num_epoch):
self.training_buffer.update_buffer.shuffle()
buffer = self.training_buffer.update_buffer
for l in range(
len(self.training_buffer.update_buffer['actions']) //
n_sequences):
start = l * n_sequences
end = (l + 1) * n_sequences
run_out = self.policy.update(
buffer.make_mini_batch(start, end), n_sequences)
value_total.append(run_out['value_loss'])
policy_total.append(np.abs(run_out['policy_loss']))
if self.use_curiosity:
inverse_total.append(run_out['inverse_loss'])
forward_total.append(run_out['forward_loss'])
self.stats['value_loss'].append(np.mean(value_total))
self.stats['policy_loss'].append(np.mean(policy_total))
if self.use_curiosity:
self.stats['forward_loss'].append(np.mean(forward_total))
self.stats['inverse_loss'].append(np.mean(inverse_total))
self.training_buffer.reset_update_buffer()
class BehavioralCloningTrainer(Trainer):
"""The ImitationTrainer is an implementation of the imitation learning."""
def __init__(self, sess, brain, trainer_parameters, training, seed, run_id):
"""
Responsible for collecting experiences and training PPO model.
:param sess: Tensorflow session.
:param trainer_parameters: The parameters for the trainer (dictionary).
:param training: Whether the trainer is set for training.
"""
super(BehavioralCloningTrainer, self).__init__(sess, brain, trainer_parameters, training, run_id)
self.param_keys = ['brain_to_imitate', 'batch_size', 'time_horizon',
'graph_scope', 'summary_freq', 'max_steps',
'batches_per_epoch', 'use_recurrent',
'hidden_units','learning_rate', 'num_layers',
'sequence_length', 'memory_size']
for k in self.param_keys:
if k not in trainer_parameters:
raise UnityTrainerException("The hyperparameter {0} could not be found for the Imitation trainer of "
"brain {1}.".format(k, brain.brain_name))
self.policy = BCPolicy(seed, brain, trainer_parameters, sess)
self.brain_name = brain.brain_name
self.brain_to_imitate = trainer_parameters['brain_to_imitate']
self.batches_per_epoch = trainer_parameters['batches_per_epoch']
self.n_sequences = max(int(trainer_parameters['batch_size'] / self.policy.sequence_length), 1)
self.cumulative_rewards = {}
self.episode_steps = {}
self.stats = {'losses': [], 'episode_length': [], 'cumulative_reward': []}
self.training_buffer = Buffer()
self.summary_path = trainer_parameters['summary_path']
if not os.path.exists(self.summary_path):
os.makedirs(self.summary_path)
self.summary_writer = tf.summary.FileWriter(self.summary_path)
def __str__(self):
return '''Hyperparameters for the Imitation Trainer of brain {0}: \n{1}'''.format(
self.brain_name, '\n'.join(['\t{0}:\t{1}'.format(x, self.trainer_parameters[x]) for x in self.param_keys]))
@property
def parameters(self):
"""
Returns the trainer parameters of the trainer.
"""
return self.trainer_parameters
@property
def get_max_steps(self):
"""
Returns the maximum number of steps. Is used to know when the trainer should be stopped.
:return: The maximum number of steps of the trainer
"""
return float(self.trainer_parameters['max_steps'])
@property
def get_step(self):
"""
Returns the number of steps the trainer has performed
:return: the step count of the trainer
"""
return self.policy.get_current_step()
@property
def get_last_reward(self):
"""
Returns the last reward the trainer has had
:return: the new last reward
"""
if len(self.stats['cumulative_reward']) > 0:
return np.mean(self.stats['cumulative_reward'])
else:
return 0
def increment_step_and_update_last_reward(self):
"""
Increment the step count of the trainer and Updates the last reward
"""
self.policy.increment_step()
return
def take_action(self, all_brain_info: AllBrainInfo):
"""
Decides actions using policy given current brain info.
:param all_brain_info: AllBrainInfo from environment.
:return: a tuple containing action, memories, values and an object
to be passed to add experiences
"""
if len(all_brain_info[self.brain_name].agents) == 0:
return [], [], [], None, None
agent_brain = all_brain_info[self.brain_name]
run_out = self.policy.evaluate(agent_brain)
if self.policy.use_recurrent:
return run_out['action'], run_out['memory_out'], None, None, None
else:
return run_out['action'], None, None, None, None
def add_experiences(self, curr_info: AllBrainInfo, next_info: AllBrainInfo, take_action_outputs):
"""
Adds experiences to each agent's experience history.
:param curr_info: Current AllBrainInfo (Dictionary of all current brains and corresponding BrainInfo).
:param next_info: Next AllBrainInfo (Dictionary of all current brains and corresponding BrainInfo).
:param take_action_outputs: The outputs of the take action method.
"""
# Used to collect teacher experience into training buffer
info_teacher = curr_info[self.brain_to_imitate]
next_info_teacher = next_info[self.brain_to_imitate]
for agent_id in info_teacher.agents:
self.training_buffer[agent_id].last_brain_info = info_teacher
for agent_id in next_info_teacher.agents:
stored_info_teacher = self.training_buffer[agent_id].last_brain_info
if stored_info_teacher is None:
continue
else:
idx = stored_info_teacher.agents.index(agent_id)
next_idx = next_info_teacher.agents.index(agent_id)
if stored_info_teacher.text_observations[idx] != "":
info_teacher_record, info_teacher_reset = \
stored_info_teacher.text_observations[idx].lower().split(",")
next_info_teacher_record, next_info_teacher_reset = next_info_teacher.text_observations[idx].\
lower().split(",")
if next_info_teacher_reset == "true":
self.training_buffer.reset_update_buffer()
else:
info_teacher_record, next_info_teacher_record = "true", "true"
if info_teacher_record == "true" and next_info_teacher_record == "true":
if not stored_info_teacher.local_done[idx]:
for i in range(self.policy.vis_obs_size):
self.training_buffer[agent_id]['visual_obs%d' % i]\
.append(stored_info_teacher.visual_observations[i][idx])
if self.policy.use_vec_obs:
self.training_buffer[agent_id]['vector_obs']\
.append(stored_info_teacher.vector_observations[idx])
if self.policy.use_recurrent:
if stored_info_teacher.memories.shape[1] == 0:
stored_info_teacher.memories = np.zeros((len(stored_info_teacher.agents),
self.policy.m_size))
self.training_buffer[agent_id]['memory'].append(stored_info_teacher.memories[idx])
self.training_buffer[agent_id]['actions'].append(next_info_teacher.
previous_vector_actions[next_idx])
info_student = curr_info[self.brain_name]
next_info_student = next_info[self.brain_name]
for agent_id in info_student.agents:
self.training_buffer[agent_id].last_brain_info = info_student
# Used to collect information about student performance.
for agent_id in next_info_student.agents:
stored_info_student = self.training_buffer[agent_id].last_brain_info
if stored_info_student is None:
continue
else:
next_idx = next_info_student.agents.index(agent_id)
if agent_id not in self.cumulative_rewards:
self.cumulative_rewards[agent_id] = 0
self.cumulative_rewards[agent_id] += next_info_student.rewards[next_idx]
if not next_info_student.local_done[next_idx]:
if agent_id not in self.episode_steps:
self.episode_steps[agent_id] = 0
self.episode_steps[agent_id] += 1
def process_experiences(self, current_info: AllBrainInfo, next_info: AllBrainInfo):
"""
Checks agent histories for processing condition, and processes them as necessary.
Processing involves calculating value and advantage targets for model updating step.
:param current_info: Current AllBrainInfo
:param next_info: Next AllBrainInfo
"""
info_teacher = next_info[self.brain_to_imitate]
for l in range(len(info_teacher.agents)):
teacher_action_list = len(self.training_buffer[info_teacher.agents[l]]['actions'])
horizon_reached = teacher_action_list > self.trainer_parameters['time_horizon']
teacher_filled = len(self.training_buffer[info_teacher.agents[l]]['actions']) > 0
if ((info_teacher.local_done[l] or horizon_reached) and teacher_filled):
agent_id = info_teacher.agents[l]
self.training_buffer.append_update_buffer(
agent_id, batch_size=None, training_length=self.policy.sequence_length)
self.training_buffer[agent_id].reset_agent()
info_student = next_info[self.brain_name]
for l in range(len(info_student.agents)):
if info_student.local_done[l]:
agent_id = info_student.agents[l]
self.stats['cumulative_reward'].append(
self.cumulative_rewards.get(agent_id, 0))
self.stats['episode_length'].append(
self.episode_steps.get(agent_id, 0))
self.cumulative_rewards[agent_id] = 0
self.episode_steps[agent_id] = 0
def end_episode(self):
"""
A signal that the Episode has ended. The buffer must be reset.
Get only called when the academy resets.
"""
self.training_buffer.reset_all()
for agent_id in self.cumulative_rewards:
self.cumulative_rewards[agent_id] = 0
for agent_id in self.episode_steps:
self.episode_steps[agent_id] = 0
def is_ready_update(self):
"""
Returns whether or not the trainer has enough elements to run update model
:return: A boolean corresponding to whether or not update_model() can be run
"""
return len(self.training_buffer.update_buffer['actions']) > self.n_sequences
def update_policy(self):
"""
Updates the policy.
"""
self.training_buffer.update_buffer.shuffle()
batch_losses = []
num_batches = min(len(self.training_buffer.update_buffer['actions']) //
self.n_sequences, self.batches_per_epoch)
for i in range(num_batches):
buffer = self.training_buffer.update_buffer
start = i * self.n_sequences
end = (i + 1) * self.n_sequences
mini_batch = buffer.make_mini_batch(start, end)
run_out = self.policy.update(mini_batch, self.n_sequences)
loss = run_out['policy_loss']
batch_losses.append(loss)
if len(batch_losses) > 0:
self.stats['losses'].append(np.mean(batch_losses))
else:
self.stats['losses'].append(0)