I want to convert the PI controller in this instance to a PID controller, so can my network be changed to something like this?

Respected all,

The following example shows the PI controller design using TD3 reinforcement learning algorithm. I want to convert the PI controller in this instance to a PID controller, so can my actornetwork be changed to something like the following?

actorNetwork = [

featureInputLayer(numObservations3,'Normalization','none','Name','state')

fullyConnectedPILayer(randi([-50,50],1,3), 'Action')];

Tune PI Controller using Reinforcement Learning - MATLAB & Simulink - MathWorks India

[轩]

Hello, I am trying the same method in the current time, can you leave a contact information for conmunication ?

你好，可以留一个邮箱或微信交流吗？

Answer 1

[Emmanouil Tzorakoleftherakis]

That should work - you are essentially adding one more weight to the PI layer for the D term. Potential issues may arise when you provide the error derivative as an observation (may be hard to approximate due to the trial and error nature of RL) but give it a try

[嘻嘻]

I am very grateful for your answer, which is very helpful to me. I would like to ask you another question, I would like to know how the initial value of the network weight in the example is determined, and what impact will it have on the result? Thank you again!

[Emmanouil Tzorakoleftherakis]

It looks like it's just a couple of values close to zero. RL solves an optimization problem under the hood so the better the initial guess, the more likely to get a better solution