rlworkgroup / garage

A toolkit for reproducible reinforcement learning research.
MIT License
1.84k stars 309 forks source link

PEARLWorker #2310

Open braunjon opened 2 years ago

braunjon commented 2 years ago

Hi, I have two issues that I do not quite understand why they are part of the code

  1. The code overwrites the action a in the deterministic case: https://github.com/rlworkgroup/garage/blob/b4abe07f0fa9bac2cb70e4a3e315c2e7e5b08507/src/garage/torch/algos/pearl.py#L743-L746 There is an open pull request about this here: https://github.com/rlworkgroup/garage/pull/2275.

  2. I was wondering if the context is ever used in self.agent. As far as I understand in the pearl.py file we never use the context of self._policy and it is also not used within the class ContextConditionedPolicy. https://github.com/rlworkgroup/garage/blob/b4abe07f0fa9bac2cb70e4a3e315c2e7e5b08507/src/garage/torch/algos/pearl.py#L754-L759

Some hints are appreciated Thanks