This is very like the idea of hill climbing with an elegant use of sinusoidal perturbations and phase shift parameter, I guess. BTW, thank you very much, Prof. Brunton, for all nice lecture series (Control bootcamp, Data-Driven science, engineering, dynamical systems and control). They help a lot for learning and establishing a comprehensive understanding on control and the emerging data-driven control area, which can definitely boost the research in my own major and areas of interest (smart energy systems).
hopelessly i was wondering if someone could teach about ESC on youtube...thank you very much steve, your work here helps me to quickly understand or at least strengthen understanding of the different methods in the control area.
Great description! One small little error when describing what happens at the peak, where I looks like you got the sign of multiplying J or rho by u the opposite of what it is, but a really excellent description none the less. Thank you!
Professor, your sessions are mind-blowing 😇 Thanks for all your efforts to support us understand the very basics at such conveniently easy and encouraging ways 😀 Feeling short of words to Thank you 😊
Excellent and very didactic video from Professor Brunton, as always! I was wondering if this type of controller works well with systems that are very sensitive and if ESC could provide some robustness to that. Thank you Professor Brunton for providing such high-quality videos available to everybody!
How would I use this in motion control? I have up to 4 close loop gains and 3 feed forward gains. I can jiggle the gains a bit and look at the sum of errors squared between my target and actual position. After all the gains have been "jiggled", I can find a slope or Jacobian that reduces the sum of squared errors. That would be the same as your u except this is a cost function. I don't like the way u is used because that is normally a simple that is reserved for the control output. This jiggle the parameters method would take a few moves to jiggle all the gains. Then I must also keep track of which way I am moving. Sometimes the load is different depending on the direction. This means I have a different set of gains for extending and retracting or going positive, then negative. If I am doing normal production work then things don't change much but if I am moving logs then the system changes depending on the size/mass. It is hard to get enough "jiggles" under the same conditions. My example is only trying to minimize J or sum of squared errors. I don't see how this suggested method converges. The BFGS, L-M or Nelder-Mead would take hundreds of iterations to minimize the sum of squared errors. Basically, what is suggested is like executing a BFGS or L-M optimization on-the-fly instead of a one time "batch" mode.
This is very like the idea of hill climbing with an elegant use of sinusoidal perturbations and phase shift parameter, I guess.
BTW, thank you very much, Prof. Brunton, for all nice lecture series (Control bootcamp, Data-Driven science, engineering, dynamical systems and control). They help a lot for learning and establishing a comprehensive understanding on control and the emerging data-driven control area, which can definitely boost the research in my own major and areas of interest (smart energy systems).
hopelessly i was wondering if someone could teach about ESC on youtube...thank you very much steve, your work here helps me to quickly understand or at least strengthen understanding of the different methods in the control area.
Great description! One small little error when describing what happens at the peak, where I looks like you got the sign of multiplying J or rho by u the opposite of what it is, but a really excellent description none the less. Thank you!
Very interesting, thank you. What’s the difference between ESC and computing partial derivatives of J vs U?
Professor, your sessions are mind-blowing 😇 Thanks for all your efforts to support us understand the very basics at such conveniently easy and encouraging ways 😀 Feeling short of words to Thank you 😊
Excellent and very didactic video from Professor Brunton, as always! I was wondering if this type of controller works well with systems that are very sensitive and if ESC could provide some robustness to that. Thank you Professor Brunton for providing such high-quality videos available to everybody!
Just a brilliant explanation, love it
Extremum seeking control is extremely useful. 😊
Thanks for this great lecture
Im a big fan of your videos. It is an elegant method to find local optimization. Does it work also efficiently with MIMO system?
How would I use this in motion control? I have up to 4 close loop gains and 3 feed forward gains. I can jiggle the gains a bit and look at the sum of errors squared between my target and actual position. After all the gains have been "jiggled", I can find a slope or Jacobian that reduces the sum of squared errors. That would be the same as your u except this is a cost function. I don't like the way u is used because that is normally a simple that is reserved for the control output. This jiggle the parameters method would take a few moves to jiggle all the gains. Then I must also keep track of which way I am moving. Sometimes the load is different depending on the direction. This means I have a different set of gains for extending and retracting or going positive, then negative. If I am doing normal production work then things don't change much but if I am moving logs then the system changes depending on the size/mass. It is hard to get enough "jiggles" under the same conditions. My example is only trying to minimize J or sum of squared errors. I don't see how this suggested method converges. The BFGS, L-M or Nelder-Mead would take hundreds of iterations to minimize the sum of squared errors.
Basically, what is suggested is like executing a BFGS or L-M optimization on-the-fly instead of a one time "batch" mode.
Thank you very much Professor.
Thank you for the explanation 👍
What a great demonstration! Thank you! I guess you need to address the osculation issue in finding the u*... how do you handle that?
very clear ,thank you
Is this an implicit online gradient de/ascent?
My question, as you are exciting the system with a single freq aren't you constraining and maybe missing the effect ?
MPPT!