b900fdf6f2
Closes #947 This removes all kwargs from all policy constructors. While doing that, I also improved several names and added a whole lot of TODOs. ## Functional changes: 1. Added possibility to pass None as `critic2` and `critic2_optim`. In fact, the default behavior then should cover the absolute majority of cases 2. Added a function called `clone_optimizer` as a temporary measure to support passing `critic2_optim=None` ## Breaking changes: 1. `action_space` is no longer optional. In fact, it already was non-optional, as there was a ValueError in BasePolicy.init. So now several examples were fixed to reflect that 2. `reward_normalization` removed from DDPG and children. It was never allowed to pass it as `True` there, an error would have been raised in `compute_n_step_reward`. Now I removed it from the interface 3. renamed `critic1` and similar to `critic`, in order to have uniform interfaces. Note that the `critic` in DDPG was optional for the sole reason that child classes used `critic1`. I removed this optionality (DDPG can't do anything with `critic=None`) 4. Several renamings of fields (mostly private to public, so backwards compatible) ## Additional changes: 1. Removed type and default declaration from docstring. This kind of duplication is really not necessary 2. Policy constructors are now only called using named arguments, not a fragile mixture of positional and named as before 5. Minor beautifications in typing and code 6. Generally shortened docstrings and made them uniform across all policies (hopefully) ## Comment: With these changes, several problems in tianshou's inheritance hierarchy become more apparent. I tried highlighting them for future work. --------- Co-authored-by: Dominik Jain <d.jain@appliedai.de>
83 lines
2.3 KiB
Python
83 lines
2.3 KiB
Python
from abc import ABC, abstractmethod
|
|
from collections.abc import Sequence
|
|
|
|
import numpy as np
|
|
|
|
|
|
class BaseNoise(ABC):
|
|
"""The action noise base class."""
|
|
|
|
@abstractmethod
|
|
def reset(self) -> None:
|
|
"""Reset to the initial state."""
|
|
|
|
@abstractmethod
|
|
def __call__(self, size: Sequence[int]) -> np.ndarray:
|
|
"""Generate new noise."""
|
|
raise NotImplementedError
|
|
|
|
|
|
class GaussianNoise(BaseNoise):
|
|
"""The vanilla Gaussian process, for exploration in DDPG by default."""
|
|
|
|
def __init__(self, mu: float = 0.0, sigma: float = 1.0) -> None:
|
|
self._mu = mu
|
|
assert sigma >= 0, "Noise std should not be negative."
|
|
self._sigma = sigma
|
|
|
|
def __call__(self, size: Sequence[int]) -> np.ndarray:
|
|
return np.random.normal(self._mu, self._sigma, size)
|
|
|
|
def reset(self) -> None:
|
|
pass
|
|
|
|
|
|
class OUNoise(BaseNoise):
|
|
"""Class for Ornstein-Uhlenbeck process, as used for exploration in DDPG.
|
|
|
|
Usage:
|
|
::
|
|
|
|
# init
|
|
self.noise = OUNoise()
|
|
# generate noise
|
|
noise = self.noise(logits.shape, eps)
|
|
|
|
For required parameters, you can refer to the stackoverflow page. However,
|
|
our experiment result shows that (similar to OpenAI SpinningUp) using
|
|
vanilla Gaussian process has little difference from using the
|
|
Ornstein-Uhlenbeck process.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
mu: float = 0.0,
|
|
sigma: float = 0.3,
|
|
theta: float = 0.15,
|
|
dt: float = 1e-2,
|
|
x0: float | np.ndarray | None = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self._mu = mu
|
|
self._alpha = theta * dt
|
|
self._beta = sigma * np.sqrt(dt)
|
|
self._x0 = x0
|
|
self.reset()
|
|
|
|
def reset(self) -> None:
|
|
"""Reset to the initial state."""
|
|
self._x = self._x0
|
|
|
|
def __call__(self, size: Sequence[int], mu: float | None = None) -> np.ndarray:
|
|
"""Generate new noise.
|
|
|
|
Return an numpy array which size is equal to ``size``.
|
|
"""
|
|
if self._x is None or isinstance(self._x, np.ndarray) and self._x.shape != size:
|
|
self._x = 0.0
|
|
if mu is None:
|
|
mu = self._mu
|
|
r = self._beta * np.random.normal(size=size)
|
|
self._x = self._x + self._alpha * (mu - self._x) + r
|
|
return self._x # type: ignore
|