Any reasons for not using batch normalize in your model?

[YuhangSong]

Hi, I was doing a research that shares the same direction as yours. I am think may be include batch normalize in the model for predicting states in RL is good for it stabilizing training, but I am also concerned that the batch normalize could bring some disadvantages since the model is not working in a mini-batch.

Do you have any insight on this issue? Or did you do a experiment on whether batch normalize is good for state prediction in RL?

Your work is definitely of huge interests for the community and I appreciate that you release the code and open for discussion.