Closed pseudo-rnd-thoughts closed 2 years ago
Hi @pseudo-rnd-thoughts,
A frame-skip of four was used in the original DQN paper. Note that the ALE pre-dates DQN and a frame-skip of five was common amongst prior works. In the latest methodology review (Revisiting the Arcade Learning Environment) a frameskip of 5 was used when generating the empirical results.
With that being said, we would recommend you use a frame-skip of 4 if you want to align with the prevailing consensus in
the Deep-RL literature. The v5
default of 5 is due to a miscommunication on my part and the default will be changed to 4 in the next release to align with the post-DQN literature. Sorry, for the confusion.
Let me know if you have any other questions regarding the most recent methodology.
Hi Jesse,
That makes sense, thanks for the update
Mark
I'm implementing a wrapper of the V5 environment that includes frame skipping and stacking, etc Looking at the default constructor, frame skipping is set to 5. Why? Am I missing a paper that explains the advantage of 5 compared to 4 which to my knowledge was the default (https://arxiv.org/pdf/1312.5602.pdf)
Thanks