combine_dueling_outputs

combine_dueling_outputs(*, advantage: torch.Tensor, value: torch.Tensor, action_shape: int, num_atoms: int, advantage_centering: DuelingAdvantageCentering | None, categorical: bool) -> torch.Tensor

Combine advantage and value streams for duelling DQN.

Parameters:

• advantage (Tensor) – Advantage stream output.
• value (Tensor) – Value stream output.
• action_shape (int) – Number of actions.
• num_atoms (int) – Number of categorical atoms.
• advantage_centering (DuelingAdvantageCentering | None) – Optional masked-mean centring for the advantage stream.
• categorical (bool) – Whether categorical output is enabled.

Returns:

• Tensor – Combined logits or Q-values.

Source