Dueling and Distributional Q Learning on the top of Custom Policy

RLlib
1

I am wandering how the definition of the custom policy and the usage of the dueling and/or Distributional Q Learning will interact?
If I am using hyperparameter optimization on the parameters like “dueling” and “distributional”, with my custom policy and/or value function, then how the things will work. Will new layers be added and/or removed?

1 Like
2

Hi Moiz, thanks for your question.
If your custom policy is extending the existing dqn policy class properly, I don’t see why it wouldn’t honor those config bits. I.e., if they return the same model class, things should work.
In case you find otherwise, file an issue with your script, and we are happy to take a look.

1 Like