from dataclasses import dataclass
from typing import Literal
import numpy as np
import torch
import torch.nn.functional as F
from torch.distributions import Categorical
from torch.nn import ModuleList
from tianshou.algorithm.algorithm_base import (
LaggedNetworkFullUpdateAlgorithmMixin,
OfflineAlgorithm,
)
from tianshou.algorithm.modelfree.reinforce import (
DiscountedReturnComputation,
DiscreteActorPolicy,
SimpleLossTrainingStats,
)
from tianshou.algorithm.optim import OptimizerFactory
from tianshou.data import ReplayBuffer, to_torch, to_torch_as
from tianshou.data.types import BatchWithReturnsProtocol, RolloutBatchProtocol
from tianshou.utils.lagged_network import EvalModeModuleWrapper
from tianshou.utils.net.discrete import DiscreteCritic
[docs]
@dataclass
class DiscreteCRRTrainingStats(SimpleLossTrainingStats):
actor_loss: float
critic_loss: float
cql_loss: float
[docs]
class DiscreteCRR(
OfflineAlgorithm[DiscreteActorPolicy],
LaggedNetworkFullUpdateAlgorithmMixin,
):
r"""Implementation of discrete Critic Regularized Regression. arXiv:2006.15134."""
def __init__(
self,
*,
policy: DiscreteActorPolicy,
critic: torch.nn.Module | DiscreteCritic,
optim: OptimizerFactory,
gamma: float = 0.99,
policy_improvement_mode: Literal["exp", "binary", "all"] = "exp",
ratio_upper_bound: float = 20.0,
beta: float = 1.0,
min_q_weight: float = 10.0,
target_update_freq: int = 0,
return_standardization: bool = False,
) -> None:
r"""
:param policy: the policy
:param critic: the action-value critic (i.e., Q function)
network. (s -> Q(s, \*))
:param optim: the optimizer factory for the policy's actor network and the critic networks.
:param gamma: the discount factor in [0, 1] for future rewards.
This determines how much future rewards are valued compared to immediate ones.
Lower values (closer to 0) make the agent focus on immediate rewards, creating "myopic"
behavior. Higher values (closer to 1) make the agent value long-term rewards more,
potentially improving performance in tasks where delayed rewards are important but
increasing training variance by incorporating more environmental stochasticity.
Typically set between 0.9 and 0.99 for most reinforcement learning tasks
:param str policy_improvement_mode: type of the weight function f. Possible
values: "binary"/"exp"/"all".
:param ratio_upper_bound: when policy_improvement_mode is "exp", the value
of the exp function is upper-bounded by this parameter.
:param beta: when policy_improvement_mode is "exp", this is the denominator
of the exp function.
:param min_q_weight: weight for CQL loss/regularizer. Default to 10.
:param target_update_freq: the number of training iterations between each complete update of
the target network.
Controls how frequently the target Q-network parameters are updated with the current
Q-network values.
A value of 0 disables the target network entirely, using only a single network for both
action selection and bootstrap targets.
Higher values provide more stable learning targets but slow down the propagation of new
value estimates. Lower positive values allow faster learning but may lead to instability
due to rapidly changing targets.
Typically set between 100-10000 for DQN variants, with exact values depending on environment
complexity.
:param return_standardization: whether to standardize episode returns
by subtracting the running mean and dividing by the running standard deviation.
Note that this is known to be detrimental to performance in many cases!
"""
super().__init__(
policy=policy,
)
LaggedNetworkFullUpdateAlgorithmMixin.__init__(self)
self.discounted_return_computation = DiscountedReturnComputation(
gamma=gamma,
return_standardization=return_standardization,
)
self.critic = critic
self.optim = self._create_optimizer(ModuleList([self.policy, self.critic]), optim)
self._target = target_update_freq > 0
self._freq = target_update_freq
self._iter = 0
self.actor_old: torch.nn.Module | EvalModeModuleWrapper
self.critic_old: torch.nn.Module | EvalModeModuleWrapper
if self._target:
self.actor_old = self._add_lagged_network(self.policy.actor)
self.critic_old = self._add_lagged_network(self.critic)
else:
self.actor_old = self.policy.actor
self.critic_old = self.critic
self._policy_improvement_mode = policy_improvement_mode
self._ratio_upper_bound = ratio_upper_bound
self._beta = beta
self._min_q_weight = min_q_weight
def _preprocess_batch(
self,
batch: RolloutBatchProtocol,
buffer: ReplayBuffer,
indices: np.ndarray,
) -> BatchWithReturnsProtocol:
return self.discounted_return_computation.add_discounted_returns(
batch,
buffer,
indices,
)
def _update_with_batch( # type: ignore[override]
self,
batch: BatchWithReturnsProtocol,
) -> DiscreteCRRTrainingStats:
if self._target and self._iter % self._freq == 0:
self._update_lagged_network_weights()
q_t = self.critic(batch.obs)
act = to_torch(batch.act, dtype=torch.long, device=q_t.device)
qa_t = q_t.gather(1, act.unsqueeze(1))
# Critic loss
with torch.no_grad():
target_a_t, _ = self.actor_old(batch.obs_next)
target_m = Categorical(logits=target_a_t)
q_t_target = self.critic_old(batch.obs_next)
rew = to_torch_as(batch.rew, q_t_target)
expected_target_q = (q_t_target * target_m.probs).sum(-1, keepdim=True)
expected_target_q[batch.done > 0] = 0.0
target = rew.unsqueeze(1) + self.discounted_return_computation.gamma * expected_target_q
critic_loss = 0.5 * F.mse_loss(qa_t, target)
# Actor loss
act_target, _ = self.policy.actor(batch.obs)
dist = Categorical(logits=act_target)
expected_policy_q = (q_t * dist.probs).sum(-1, keepdim=True)
advantage = qa_t - expected_policy_q
if self._policy_improvement_mode == "binary":
actor_loss_coef = (advantage > 0).float()
elif self._policy_improvement_mode == "exp":
actor_loss_coef = (advantage / self._beta).exp().clamp(0, self._ratio_upper_bound)
else:
actor_loss_coef = 1.0 # effectively behavior cloning
actor_loss = (-dist.log_prob(act) * actor_loss_coef).mean()
# CQL loss/regularizer
min_q_loss = (q_t.logsumexp(1) - qa_t).mean()
loss = actor_loss + critic_loss + self._min_q_weight * min_q_loss
self.optim.step(loss)
self._iter += 1
return DiscreteCRRTrainingStats(
loss=loss.item(),
actor_loss=actor_loss.item(),
critic_loss=critic_loss.item(),
cql_loss=min_q_loss.item(),
)
