3a1bc18add
This PR adds a new method for getting actions from an env's observation and info. This is useful for standard inference and stands in contrast to batch-based methods that are currently used in training and evaluation. Without this, users have to do some kind of gymnastics to actually perform inference with a trained policy. I have also added a test for the new method. In future PRs, this method should be included in the examples (in the the "watch" section). To add this required improving multiple typing things and, importantly, _simplifying the signature of `forward` in many policies!_ This is a **breaking change**, but it will likely affect no users. The `input` parameter of forward was a rather hacky mechanism, I believe it is good that it's gone now. It will also help with #948 . The main functional change is the addition of `compute_action` to `BasePolicy`. Other minor changes: - improvements in typing - updated PR and Issue templates - Improved handling of `max_action_num` Closes #981
243 lines
9.0 KiB
Python
243 lines
9.0 KiB
Python
from typing import Any, cast
|
|
|
|
import gymnasium as gym
|
|
import numpy as np
|
|
import torch
|
|
|
|
from tianshou.data import Batch
|
|
from tianshou.data.batch import BatchProtocol
|
|
from tianshou.data.types import ActBatchProtocol, ObsBatchProtocol, RolloutBatchProtocol
|
|
from tianshou.policy import BasePolicy
|
|
from tianshou.policy.base import TLearningRateScheduler
|
|
|
|
|
|
class PSRLModel:
|
|
"""Implementation of Posterior Sampling Reinforcement Learning Model.
|
|
|
|
:param trans_count_prior: dirichlet prior (alphas), with shape
|
|
(n_state, n_action, n_state).
|
|
:param rew_mean_prior: means of the normal priors of rewards,
|
|
with shape (n_state, n_action).
|
|
:param rew_std_prior: standard deviations of the normal priors
|
|
of rewards, with shape (n_state, n_action).
|
|
:param discount_factor: in [0, 1].
|
|
:param epsilon: for precision control in value iteration.
|
|
:param lr_scheduler: a learning rate scheduler that adjusts the learning rate in
|
|
optimizer in each policy.update(). Default to None (no lr_scheduler).
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
trans_count_prior: np.ndarray,
|
|
rew_mean_prior: np.ndarray,
|
|
rew_std_prior: np.ndarray,
|
|
discount_factor: float,
|
|
epsilon: float,
|
|
) -> None:
|
|
self.trans_count = trans_count_prior
|
|
self.n_state, self.n_action = rew_mean_prior.shape
|
|
self.rew_mean = rew_mean_prior
|
|
self.rew_std = rew_std_prior
|
|
self.rew_square_sum = np.zeros_like(rew_mean_prior)
|
|
self.rew_std_prior = rew_std_prior
|
|
self.discount_factor = discount_factor
|
|
self.rew_count = np.full(rew_mean_prior.shape, epsilon) # no weight
|
|
self.eps = epsilon
|
|
self.policy: np.ndarray
|
|
self.value = np.zeros(self.n_state)
|
|
self.updated = False
|
|
self.__eps = np.finfo(np.float32).eps.item()
|
|
|
|
def observe(
|
|
self,
|
|
trans_count: np.ndarray,
|
|
rew_sum: np.ndarray,
|
|
rew_square_sum: np.ndarray,
|
|
rew_count: np.ndarray,
|
|
) -> None:
|
|
"""Add data into memory pool.
|
|
|
|
For rewards, we have a normal prior at first. After we observed a
|
|
reward for a given state-action pair, we use the mean value of our
|
|
observations instead of the prior mean as the posterior mean. The
|
|
standard deviations are in inverse proportion to the number of the
|
|
corresponding observations.
|
|
|
|
:param trans_count: the number of observations, with shape
|
|
(n_state, n_action, n_state).
|
|
:param rew_sum: total rewards, with shape
|
|
(n_state, n_action).
|
|
:param rew_square_sum: total rewards' squares, with shape
|
|
(n_state, n_action).
|
|
:param rew_count: the number of rewards, with shape
|
|
(n_state, n_action).
|
|
"""
|
|
self.updated = False
|
|
self.trans_count += trans_count
|
|
sum_count = self.rew_count + rew_count
|
|
self.rew_mean = (self.rew_mean * self.rew_count + rew_sum) / sum_count
|
|
self.rew_square_sum += rew_square_sum
|
|
raw_std2 = self.rew_square_sum / sum_count - self.rew_mean**2
|
|
self.rew_std = np.sqrt(
|
|
1 / (sum_count / (raw_std2 + self.__eps) + 1 / self.rew_std_prior**2),
|
|
)
|
|
self.rew_count = sum_count
|
|
|
|
def sample_trans_prob(self) -> np.ndarray:
|
|
return torch.distributions.Dirichlet(torch.from_numpy(self.trans_count)).sample().numpy()
|
|
|
|
def sample_reward(self) -> np.ndarray:
|
|
return np.random.normal(self.rew_mean, self.rew_std)
|
|
|
|
def solve_policy(self) -> None:
|
|
self.updated = True
|
|
self.policy, self.value = self.value_iteration(
|
|
self.sample_trans_prob(),
|
|
self.sample_reward(),
|
|
self.discount_factor,
|
|
self.eps,
|
|
self.value,
|
|
)
|
|
|
|
@staticmethod
|
|
def value_iteration(
|
|
trans_prob: np.ndarray,
|
|
rew: np.ndarray,
|
|
discount_factor: float,
|
|
eps: float,
|
|
value: np.ndarray,
|
|
) -> tuple[np.ndarray, np.ndarray]:
|
|
"""Value iteration solver for MDPs.
|
|
|
|
:param trans_prob: transition probabilities, with shape
|
|
(n_state, n_action, n_state).
|
|
:param rew: rewards, with shape (n_state, n_action).
|
|
:param eps: for precision control.
|
|
:param discount_factor: in [0, 1].
|
|
:param value: the initialize value of value array, with
|
|
shape (n_state, ).
|
|
|
|
:return: the optimal policy with shape (n_state, ).
|
|
"""
|
|
Q = rew + discount_factor * trans_prob.dot(value)
|
|
new_value = Q.max(axis=1)
|
|
while not np.allclose(new_value, value, eps):
|
|
value = new_value
|
|
Q = rew + discount_factor * trans_prob.dot(value)
|
|
new_value = Q.max(axis=1)
|
|
# this is to make sure if Q(s, a1) == Q(s, a2) -> choose a1/a2 randomly
|
|
Q += eps * np.random.randn(*Q.shape)
|
|
return Q.argmax(axis=1), new_value
|
|
|
|
def __call__(
|
|
self,
|
|
obs: np.ndarray,
|
|
state: Any = None,
|
|
info: Any = None,
|
|
) -> np.ndarray:
|
|
if not self.updated:
|
|
self.solve_policy()
|
|
return self.policy[obs]
|
|
|
|
|
|
class PSRLPolicy(BasePolicy):
|
|
"""Implementation of Posterior Sampling Reinforcement Learning.
|
|
|
|
Reference: Strens M. A Bayesian framework for reinforcement learning [C]
|
|
//ICML. 2000, 2000: 943-950.
|
|
|
|
:param trans_count_prior: dirichlet prior (alphas), with shape
|
|
(n_state, n_action, n_state).
|
|
:param rew_mean_prior: means of the normal priors of rewards,
|
|
with shape (n_state, n_action).
|
|
:param rew_std_prior: standard deviations of the normal priors
|
|
of rewards, with shape (n_state, n_action).
|
|
:param action_space: Env's action_space.
|
|
:param discount_factor: in [0, 1].
|
|
:param epsilon: for precision control in value iteration.
|
|
:param add_done_loop: whether to add an extra self-loop for the
|
|
terminal state in MDP. Default to False.
|
|
:param observation_space: Env's observation space.
|
|
:param lr_scheduler: if not None, will be called in `policy.update()`.
|
|
|
|
.. seealso::
|
|
|
|
Please refer to :class:`~tianshou.policy.BasePolicy` for more detailed
|
|
explanation.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
*,
|
|
trans_count_prior: np.ndarray,
|
|
rew_mean_prior: np.ndarray,
|
|
rew_std_prior: np.ndarray,
|
|
action_space: gym.spaces.Discrete,
|
|
discount_factor: float = 0.99,
|
|
epsilon: float = 0.01,
|
|
add_done_loop: bool = False,
|
|
observation_space: gym.Space | None = None,
|
|
lr_scheduler: TLearningRateScheduler | None = None,
|
|
) -> None:
|
|
super().__init__(
|
|
action_space=action_space,
|
|
observation_space=observation_space,
|
|
action_scaling=False,
|
|
action_bound_method=None,
|
|
lr_scheduler=lr_scheduler,
|
|
)
|
|
assert 0.0 <= discount_factor <= 1.0, "discount factor should be in [0, 1]"
|
|
self.model = PSRLModel(
|
|
trans_count_prior,
|
|
rew_mean_prior,
|
|
rew_std_prior,
|
|
discount_factor,
|
|
epsilon,
|
|
)
|
|
self._add_done_loop = add_done_loop
|
|
|
|
def forward(
|
|
self,
|
|
batch: ObsBatchProtocol,
|
|
state: dict | BatchProtocol | np.ndarray | None = None,
|
|
**kwargs: Any,
|
|
) -> ActBatchProtocol:
|
|
"""Compute action over the given batch data with PSRL model.
|
|
|
|
:return: A :class:`~tianshou.data.Batch` with "act" key containing
|
|
the action.
|
|
|
|
.. seealso::
|
|
|
|
Please refer to :meth:`~tianshou.policy.BasePolicy.forward` for
|
|
more detailed explanation.
|
|
"""
|
|
assert isinstance(batch.obs, np.ndarray), "only support np.ndarray observation"
|
|
# TODO: shouldn't the model output a state as well if state is passed (i.e. RNNs are involved)?
|
|
act = self.model(batch.obs, state=state, info=batch.info)
|
|
return cast(ActBatchProtocol, Batch(act=act))
|
|
|
|
def learn(self, batch: RolloutBatchProtocol, *args: Any, **kwargs: Any) -> dict[str, float]:
|
|
n_s, n_a = self.model.n_state, self.model.n_action
|
|
trans_count = np.zeros((n_s, n_a, n_s))
|
|
rew_sum = np.zeros((n_s, n_a))
|
|
rew_square_sum = np.zeros((n_s, n_a))
|
|
rew_count = np.zeros((n_s, n_a))
|
|
for minibatch in batch.split(size=1):
|
|
obs, act, obs_next = minibatch.obs, minibatch.act, minibatch.obs_next
|
|
obs_next = cast(np.ndarray, obs_next)
|
|
assert not isinstance(obs, BatchProtocol), "Observations cannot be Batches here"
|
|
trans_count[obs, act, obs_next] += 1
|
|
rew_sum[obs, act] += minibatch.rew
|
|
rew_square_sum[obs, act] += minibatch.rew**2
|
|
rew_count[obs, act] += 1
|
|
if self._add_done_loop and minibatch.done:
|
|
# special operation for terminal states: add a self-loop
|
|
trans_count[obs_next, :, obs_next] += 1
|
|
rew_count[obs_next, :] += 1
|
|
self.model.observe(trans_count, rew_sum, rew_square_sum, rew_count)
|
|
return {
|
|
"psrl/rew_mean": float(self.model.rew_mean.mean()),
|
|
"psrl/rew_std": float(self.model.rew_std.mean()),
|
|
}
|