Input normalization using a reinforcement learning DQN agent

Hi all

I have built a DQN agent to solv a custom reinforcement problem.

Following mathworks examples I have no used any kind of normalization applied to the critic input.

In fact, as far as I could check, all examples of RL that use a DNN to create an actor or a critic especify 'Normalization', 'none' at the input layers of the Actor and Critic.

My question is, is it possible to use a normalization as for instance "zscore" at the input layers of a critic or of an actor when these are based on a DNN?

I'have tried to applied zscore normalization, but then, the agent does not work.

thanks

Answer 1

[Emmanouil Tzorakoleftherakis]

Hello,

Normalization through the input layers is not supported for RL training. As a workaround, you can scale the observations rewards on the environment side.

[Margarita Cabrera]

Ok, thanks

[H. M.]

@Emmanouil Tzorakoleftherakis

Could you explain more, the way you mentioned about normalization. I want to do it, but I can't figure it out.

Regards