SoftUpdate¶
- class torchrl.objectives.SoftUpdate(loss_module: Union['DQNLoss', 'DDPGLoss', 'SACLoss', 'REDQLoss', 'TD3Loss'], *, eps: float = None, tau: Optional[float] = None)[source]¶
A soft-update class for target network update in Double DQN/DDPG.
This was proposed in “CONTINUOUS CONTROL WITH DEEP REINFORCEMENT LEARNING”, https://arxiv.org/pdf/1509.02971.pdf
One and only one decay factor (tau or eps) must be specified.
- Parameters:
loss_module (DQNLoss or DDPGLoss) – loss module where the target network should be updated.
eps (scalar) –
epsilon in the update equation: .. math:
\theta_t = \theta_{t-1} * \epsilon + \theta_t * (1-\epsilon)
Exclusive with
tau
.tau (scalar) – Polyak tau. It is equal to
1-eps
, and exclusive with it.