Why my RL Agent action still passing the upper and lower limit ?

I am using Policy Gradient Agent, I want that my action only in range 0 - 100 and i already set up my UpperLimit to 100, and LowerLimit to 0. But as you can see -scope display 3-, my action still can passing the limit. How can i fix that ?

[Emmanouil Tzorakoleftherakis]

which one is the action here? How does your actor network look like?

[denny]

I have solve my similar problem.

actInfo =rlNumericSpec([ 1],'UpperLimit',0.0771,'LowerLimit',-0.0405)

it means the minimum value is -0.0405, the maximum value is -0.0405+0.0771*2.

but your output is -1000 to 1000, I also donot know it.

Answer 1

[denny]

I have the same problems. How to solve it?

Answer 2

[Azmi Yagli]

If you look at rlNumericSpec, you can see this on LoweLimit or UpperLimit section.

DDPG, TD3 and SAC agents use this property to enforce lower limits on the action. When using other agents, if you need to enforce constraints on the action, you must do so within the environment.

So if you use other algorithms you can use saturation, but it didn't work for me.

You can try discretize actions of your agent so it can have boundaries.

Or you can give negative reward, if your agent exceeds limits for action.