Reinforcement Learning, how to apply normalization and for which parts?

Hi,

I have a model in Simulink and I am trying to solve a scheduling problem for control. Since we also use neural networks (I am using actor critic because of the continuous action and state space) for the agent, I believe normalization should help to find at least local optima for the problem.

However, it does not work (at least with the R2021a) if we change the command from 'none' to any other option for creating the state and action.

Here is an example of the code line:

--> featureInputLayer(numObservations,'Normalization','none','Name','State')

That is why I am applying the normalization in my simulink file.

I was wondering what should be the best way to implement normalization in .slx file. Currently, I am applying for my observations and reward function.

Should I apply for another component in addition to this or not apply for any of these parts?

Any comments would be helpful, thanks.

[H. M.]

hi @Aysegul Kahraman

I faced the same problem, if you find a way to solve this could you share it.

many thanks

[Aysegul Kahraman]

Hi, it seems there is no such an easy and direct way to normalize (you can not write like ('Normalization','zscore') or any other method at least according to my findings).

My solution was applying normalization seperately for observations, reward and so on. Hope that helps.

[Sourabh]

can you please elaborate like how did u applied normalization to rewards plzzzz