Tune PI Controller Using Reinforcement Learning

How is the initial value of the weight of this neural network determined? If I want to change my PI controller to a PID controller, do I just add another weight to this row that is initialGain = single([1e-3 2])?

This code is from the demo "Tune PI Controller Using Reinforcement Learning."

initialGain = single([1e-3 2]);

actorNet = [

featureInputLayer(numObs)

fullyConnectedPILayer(initialGain,'ActOutLyr')

];

actorNet = dlnetwork(actorNet);

actor = rlContinuousDeterministicActor(actorNet,obsInfo,actInfo);

Can my network be changed to look like the following：

actorNet= [

featureInputLayer(numObs)

fullyConnectedPILayer(randi([-60,60],1,3), 'Action')]

[Sam Chak]

Hi @嘻嘻

Could you please explain how the proposed code mathematically represents the PID control law ?

actorNet= [

featureInputLayer(numObs)

fullyConnectedPILayer(randi([-60,60],1,3), 'Action')]

[嘻嘻]

I want the weights of the network to represent the controller parameters, the input of the network to represent the error and the error integral and its first derivative, and the final output of the network to be the control instructions

[嘻嘻]

I'm not really sure. What do you think of this scheme?

Answer 1

[Emmanouil Tzorakoleftherakis]

I also replied to the other thread. The fullyConnectedPILayer is a custom layer provided in the example - you can open it and see how it's implemented. So you can certainly add a third weight for the D term, but you will most likely run into other issues (e.g. how to approximate the error derivative)