def __init__(self, config=default()):
super().__init__(config)
self.experience = namedtuple(
"Experience",
field_names=["state", "action", "reward", "next_state", "done"])
logger.info(describe(self))
self.clear()
def __init__(self,
config,
input_state_size,
output_action_size,
fc1_units=400,
fc2_units=300):
super(DDPG_Networks, self).__init__(config, input_state_size,
output_action_size)
self.actor_local = LowDimActor(config, self.input_state_size,
self.output_action_size, fc1_units,
fc2_units)
self.actor_target = LowDimActor(config, self.input_state_size,
self.output_action_size, fc1_units,
fc2_units)
self.critic_local = LowDimCritic(config, self.input_state_size,
self.output_action_size, fc1_units,
fc2_units)
self.critic_target = LowDimCritic(config, self.input_state_size,
self.output_action_size, fc1_units,
fc2_units)
logger.warn('DDPG_Networks is initialized!')
logger.info(utils.describe(self))
def __init__(self, config):
super(GymEnv, self).__init__(config)
self._env = gym.make(self.name)
self._set_attr_from_u_env(self._env)
self._env.seed(self.seed)
logger.info(utils.describe(self))
def __init__(self, config):
super(UnityEnv, self).__init__(config)
self._env = UnityEnvironment(file_name=get_env_path(self.name), seed=self.seed)
self.patch_gym_spaces(self._env)
self._set_attr_from_u_env(self._env)
# TODO: Logging
print(utils.describe(self))
def patch_gym_spaces(self, env):
r"""
For standardization, use gym spaces to represent observation and action spaces for Unity.
This method iterates through the multiple brains (multiagent) then constructs and returns lists of observation_spaces and action_spaces
:param env:
:return:
"""
observation_spaces = []
action_spaces = []
for brain_index in range(len(env.brain_names)):
brain = self._get_brain(env, brain_index)
# TODO: Logging
utils.describe(brain)
observation_shape = (brain.get_observable_dim()['state'],)
action_dim = (brain.get_action_dim(),)
if brain.is_discrete():
dtype = np.int32
action_space = spaces.Discrete(brain.get_action_dim())
else:
dtype = np.float32
action_space = spaces.Box(low=0.0, high=1.0, shape=action_dim, dtype=dtype)
observation_space = spaces.Box(low=0, high=1, shape=observation_shape, dtype=dtype)
utils.set_gym_space_attr(observation_space)
utils.set_gym_space_attr(action_space)
observation_spaces.append(observation_space)
action_spaces.append(action_space)
# set for singleton
env.observation_space = observation_spaces[0]
env.action_space = action_spaces[0]
return observation_spaces, action_spaces
def __init__(self, env: BaseEnv, models: tuple, optims: tuple, noise: OUNoise, configs: Configs = default()):
super(DDPGAgent, self).__init__(env, configs)
self.actor, self.actor_target, self.critic, self.critic_target = models
self.actor_optimizer, self.critic_optimizer = optims
self.curr_state = np.zeros((1,))
self.noise = noise
self.total_step = 0
self.episode_step = 0
self.i_episode = 0
if configs.glob.load_from is not None and os.path.exists(configs.glob.load_from):
self.load_params(configs.glob.load_from)
self._initialize()
logger.info(describe(self))