DQN eval_epsilon overwrite with epsilon on 0.0

google-deepmind / acme

A library of reinforcement learning components and agents

Apache License 2.0

3.52k stars 426 forks source link

DQN eval_epsilon overwrite with epsilon on 0.0 #291

Open tseyde opened 1 year ago

tseyde commented 1 year ago

Hi Acme team, I think JAX DQN might set the evaluation epsilon to the exploration epsilon if deterministic evaluation is requested (eps=0.0, here). Replacing this with self._config.eval_epsilon is not None fixed it for me - does this also occur on your end?

Thanks so much for checking!