[Docs] Clarify DQN hiddens docstring as dueling-only (#43051)

[dstrodtman]

Description

The hiddens parameter in rllib/algorithms/dqn/dqn.py only applies to the advantage and value branches of a dueling DQN, but the docstring did not say so. As a result, users could confuse it with model.fcnet_hiddens (see linked issue). This PR appends "in a dueling architecture" to the docstring to make the scope explicit.

Related issues

Closes #43051

[DOC-981]

Additional information

Target file: rllib/algorithms/dqn/dqn.py (the hiddens entry in the training argument list).

Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the docstring for the hiddens parameter in the DQN training method. The reviewer noted that the new description is misleading because the parameter is used regardless of whether a dueling architecture is enabled. Furthermore, the feedback identifies a typo in the docstring and an incorrect type hint in the method signature that requires a fix.

[gemini-code-assist]

The addition "in a dueling architecture" is technically misleading. In the current DQN implementation (both for PyTorch and TensorFlow), the hiddens parameter is also used to define the hidden layers of the single Q-network head when dueling=False. Specifying it as dueling-only might lead users to believe the parameter is ignored in non-dueling configurations. Additionally, there is a typo in the preceding line (307): "each the" should be "each of the" or just "the". Finally, please note that the type hint for hiddens in the training method signature (line 231) is Optional[int], but it should be Optional[List[int]] to match its usage and initialization.