def __init__(self, dt: float, gamma: float, lr: float, tau: float,
optimizer: str, val_function: ParametricFunction,
adv_function: ParametricFunction) -> None:
CompoundStateful.__init__(self)
self._reference_obs: Tensor = None
self._val_function = val_function
self._adv_function = adv_function
self._target_val_function = copy.deepcopy(val_function)
self._target_adv_function = copy.deepcopy(adv_function)
self._adv_optimizer = \
setup_optimizer(self._adv_function.parameters(),
opt_name=optimizer, lr=lr, dt=dt,
inverse_gradient_magnitude=1,
weight_decay=0)
self._val_optimizer = \
setup_optimizer(self._val_function.parameters(),
opt_name=optimizer, lr=lr, dt=dt,
inverse_gradient_magnitude=dt,
weight_decay=0)
self._dt = dt
self._gamma = gamma
self._tau = tau
info(
f"setup> using AdvantageCritic, the provided gamma and rewards are scaled,"
f" actual values: gamma={gamma ** dt}, rewards=original_rewards * {dt}"
)
self._device = 'cpu'
def __init__(self, dt: float, gamma: float, lr: float, optimizer: str,
q_function: ParametricFunction, tau: float,
noscale: bool) -> None:
CompoundStateful.__init__(self)
self._reference_obs: Tensor = None
self._q_function = q_function
self._target_q_function = copy.deepcopy(self._q_function)
self._tau = tau
ref_dt = 0.02
if noscale:
self._gamma = gamma**(dt / ref_dt)
else:
self._gamma = gamma
if noscale:
dt = ref_dt
self._dt = ref_dt
else:
self._dt = dt
info(
f"setup> using ValueCritic, the provided gamma and rewards are scaled,"
f" actual values: gamma={gamma ** self._dt},"
f" rewards=original_rewards * {self._dt}")
self._q_optimizer = \
setup_optimizer(self._q_function.parameters(),
opt_name=optimizer, lr=lr, dt=self._dt,
inverse_gradient_magnitude=self._dt,
weight_decay=0)
self._device = 'cpu'
def __init__(self, gamma: float, dt: float,
v_function: ParametricFunction) -> None:
CompoundStateful.__init__(self)
self._reference_obs: Tensor = None
self._v_function = v_function
self._gamma = gamma**dt
self._device = 'cpu'
self._dt = dt
info(
f"setup> using OnlineCritic, the provided gamma and rewards are scaled,"
f" actual values: gamma={gamma ** dt}, rewards=original_rewards * {dt}"
)
def __init__(self, policy_function: ParametricFunction, noise: Noise,
lr: float, tau: float, opt_name: str, dt: float,
weight_decay: float) -> None:
CompoundStateful.__init__(self)
self._policy_function = policy_function
self._target_policy_function = copy.deepcopy(self._policy_function)
self._optimizer = setup_optimizer(self._policy_function.parameters(),
opt_name=opt_name,
lr=lr,
dt=dt,
inverse_gradient_magnitude=1,
weight_decay=weight_decay)
self._noise = noise
self._tau = tau
def __init__(self, T: int, actor: OnlineActor,
critic: OnlineCritic) -> None:
CompoundStateful.__init__(self)
# reset and set in train mode
self.train()
# define learning components
self._actor = actor
self._critic = critic
self._count = 0
self._T = T
self._device = "cpu"
self.reset()
# init _nb_train_env and _current_trajectories to None
self._nb_train_env: Optional[int] = None
self._current_trajectories: List[Trajectory] = []
def __init__(self, memory_size: int, batch_size: int, steps_btw_train: int,
learn_per_step: int, alpha: Optional[float],
beta: Optional[float], actor: Actor, critic: Critic) -> None:
CompoundStateful.__init__(self)
# reset and set in train mode
self.reset()
self.train()
# define learning components
self._actor = actor
self._critic = critic
self._sampler = setup_memory(alpha=alpha,
beta=beta,
memory_size=memory_size,
batch_size=batch_size)
# counter and parameters
self._count = 0
self._warm_up = 10 # prevents learning from a near empty buffer
self._steps_btw_train = steps_btw_train
self._learn_per_step = learn_per_step
def __init__(self, policy_function: ParametricFunction,
dt: float, c_entropy: float) -> None:
CompoundStateful.__init__(self)
self._policy_function = policy_function
self._c_entropy = c_entropy