def __init__(self,
environment=None,
params=None,
bark_behavior=None,
observer=None):
BehaviorModel.__init__(self, params)
self._params = params
self._observer = observer
self._environment = environment
self._wrapped_env = tf_py_environment.TFPyEnvironment(
PyBARKEnvironment(self._environment))
self._ckpt = tf.train.Checkpoint(step=tf.Variable(0, dtype=tf.int64))
self._agent = self.GetAgent(self._wrapped_env, params)
self._ckpt = tf.train.Checkpoint(step=tf.Variable(0, dtype=tf.int64),
agent=self._agent)
ckpt_path = self._params["ML"]["BehaviorTFAAgents"]["CheckpointPath",
"", ""]
self._ckpt_manager = self.GetCheckpointer(
ckpt_path,
self._params["ML"]["BehaviorTFAAgents"]["NumCheckpointsToKeep", "",
3])
self._best_ckpt_manager = self.GetCheckpointer(
ckpt_path + "best_checkpoint/", 1)
self._logger = logging.getLogger()
# NOTE: by default we do not want the action to be set externally
# as this enables the agents to be plug and played in BARK.
self._set_action_externally = False
self._bark_behavior_model = bark_behavior or BehaviorContinuousML(
params)
def __init__(self, environment=None, params=None):
self._sac_params = params
BehaviorTFAAgent.__init__(self, environment=environment, params=params)
BehaviorModel.__init__(self, params)
self._replay_buffer = self.GetReplayBuffer()
self._dataset = self.GetDataset()
self._collect_policy = self.GetCollectionPolicy()
self._eval_policy = self.GetEvalPolicy()
def __init__(self, environment=None, params=None, bark_behavior=None):
BehaviorModel.__init__(self, params)
self._params = params
self._environment = environment
self._wrapped_env = tf_py_environment.TFPyEnvironment(
TFAWrapper(self._environment))
self._ckpt = tf.train.Checkpoint(step=tf.Variable(0, dtype=tf.int64))
self._agent = self.GetAgent(self._wrapped_env, params)
self._ckpt = tf.train.Checkpoint(step=tf.Variable(0, dtype=tf.int64),
agent=self._agent)
self._ckpt_manager = self.GetCheckpointer()
self._logger = logging.getLogger()
# NOTE: by default we do not want the action to be set externally
# as this enables the agents to be plug and played in BARK.
self._set_action_externally = False
self._bark_behavior_model = bark_behavior or BehaviorContinuousML(
params)
def __init__(self, params=None):
BehaviorModel.__init__(self, params)
self._dynamic_behavior_model = BehaviorDynamicModel(params)
def __init__(self, dynamic_model=None, params=None):
# BehaviorDynamicModel.__init__(self, dynamic_model, params)
BehaviorModel.__init__(self, params)
self._dynamic_model = dynamic_model
self._params = params
def __init__(self, env, test_env, params, bark_behavior=None):
BehaviorModel.__init__(self, params)
self._params = params
self.env = env
self.test_env = test_env
self._bark_behavior_model = bark_behavior or BehaviorDiscreteMacroActionsML(
params)
self.device = torch.device("cuda" if self._params["ML"]["BaseAgent"][
"Cuda", "", True] and torch.cuda.is_available() else "cpu")
self.online_net = None
self.target_net = None
self.steps = 0
self.learning_steps = 0
self.episodes = 0
self.best_eval_score = -np.inf
self.num_actions = self.env.action_space.n
self.num_steps = self._params["ML"]["BaseAgent"]["NumSteps", "",
5000000]
self.batch_size = self._params["ML"]["BaseAgent"]["BatchSize", "", 32]
self.double_q_learning = self._params["ML"]["BaseAgent"][
"Double_q_learning", "", False]
self.dueling_net = self._params["ML"]["BaseAgent"]["DuelingNet", "",
False]
self.noisy_net = self._params["ML"]["BaseAgent"]["NoisyNet", "", False]
self.use_per = self._params["ML"]["BaseAgent"]["Use_per", "", False]
self.reward_log_interval = self._params["ML"]["BaseAgent"][
"RewardLogInterval", "", 5]
self.summary_log_interval = self._params["ML"]["BaseAgent"][
"SummaryLogInterval", "", 100]
self.eval_interval = self._params["ML"]["BaseAgent"]["EvalInterval",
"", 25000]
self.num_eval_steps = self._params["ML"]["BaseAgent"]["NumEvalSteps",
"", 12500]
self.gamma_n = \
self._params["ML"]["BaseAgent"]["Gamma", "", 0.99] ** \
self._params["ML"]["BaseAgent"]["Multi_step", "", 1]
self.start_steps = self._params["ML"]["BaseAgent"]["StartSteps", "",
5000]
self.epsilon_train = LinearAnneaer(
1.0, self._params["ML"]["BaseAgent"]["EpsilonTrain", "", 0.01],
self._params["ML"]["BaseAgent"]["EpsilonDecaySteps", "", 25000])
self.epsilon_eval = self._params["ML"]["BaseAgent"]["EpsilonEval", "",
0.001]
self.update_interval = \
self._params["ML"]["BaseAgent"]["Update_interval", "", 4]
self.target_update_interval = self._params["ML"]["BaseAgent"][
"TargetUpdateInterval", "", 5000]
self.max_episode_steps = \
self._params["ML"]["BaseAgent"]["MaxEpisodeSteps", "", 10000]
self.grad_cliping = self._params["ML"]["BaseAgent"]["GradCliping", "",
5.0]
self.summary_dir = \
self._params["ML"]["BaseAgent"]["SummaryPath", "", ""]
self.model_dir = \
self._params["ML"]["BaseAgent"]["CheckpointPath", "", ""]
if not os.path.exists(self.model_dir) and self.model_dir:
os.makedirs(self.model_dir)
if not os.path.exists(self.summary_dir) and self.summary_dir:
os.makedirs(self.summary_dir)
self.writer = SummaryWriter(log_dir=self.summary_dir)
self.train_return = RunningMeanStats(self.summary_log_interval)
# NOTE: by default we do not want the action to be set externally
# as this enables the agents to be plug and played in BARK.
self._set_action_externally = False
# Replay memory which is memory-efficient to store stacked frames.
if self.use_per:
beta_steps = (self.num_steps - self.start_steps) / \
self.update_interval
self.memory = LazyPrioritizedMultiStepMemory(
self._params["ML"]["BaseAgent"]["MemorySize", "", 10**6],
self.env.observation_space.shape,
self.device,
self._params["ML"]["BaseAgent"]["Gamma", "", 0.99],
self._params["ML"]["BaseAgent"]["Multi_step", "", 1],
beta_steps=beta_steps)
else:
self.memory = LazyMultiStepMemory(
self._params["ML"]["BaseAgent"]["MemorySize", "", 10**6],
self.env.observation_space.shape, self.device,
self._params["ML"]["BaseAgent"]["Gamma", "", 0.99],
self._params["ML"]["BaseAgent"]["Multi_step", "", 1])