search

openai / evolution-strategies-starter

Code for the paper "Evolution Strategies as a Scalable Alternative to Reinforcement Learning"
https://arxiv.org/abs/1703.03864
MIT License
1.57k stars 278 forks source link

Virtual batch Normalization #8

Open sahiliitm opened 7 years ago

sahiliitm commented 7 years ago

If I understand the code correctly, it uses virtual batch normalization only for the inputs and not for the intermediate layers.

Was this done in the Atari context for getting the results stated in the paper?

Also, what was the network architecture used for the Atari domain?

joyousrabbit commented 7 years ago

@sahiliitm Hello, in the paper, it is stated " We used the same preprocessing and feedforward CNN architecture used by (Mnih et al., 2016)". So it should be the traditional two layers FF.

PatrykChrabaszcz commented 7 years ago

Hi @sahiliitm

Could you point me to the code where you see virtual batch normalisation implementation? I thought it is:

@property
def needs_ref_batch(self):
     return False

Which is currently not implemented

louiskirsch commented 4 years ago

In the code here we have just z-normalization for the inputs, no virtual batch norm. Also no hyperparameters for Atari. OpenAI, please be more reproducible! :-)