A3C_continuous_action - reward normalisation

[Anjum48]

Hi - thank you very much for a great set of tutorials.

I noticed this line where the reward is rescaled:
https://github.com/MorvanZhou/Reinforcement-learning-with-tensorflow/blob/master/contents/10_A3C/A3C_continuous_action.py#L134

Removing this line seems to break the Pendulum example, so therefore it is clearly sensitive to this parameter. I understand why this was done, but are you aware of any other way of dealing with this issue to make the algorithm more general?

I came across this which sounds interesting https://arxiv.org/pdf/1602.07714.pdf

[MorvanZhou]

Hi @Anjum48 ,
In my personal experience, the normalized reward does matter for a general RL setup. I also heard about the paper you linked at the bottom. It generalize the reward effect but I haven't tried this algorithm yet.
The issue wrt the reward range is the issue of neural nets themselves. A neural net would have difficulty in learning a larger range of predicted value. I don't have a better solution at the moment. For me, the best way to address this issue seems just follow that paper. haha.