Robust Learning Control Based on Gaussian Process Regression
Вставка
- Опубліковано 30 лис 2014
- This video demonstrates a new approach to improving controllers based on data obtained during operation. The approach combines robust control theory and machine learning, and guarantees safe, online learning.
More details in our recent ECC paper: www.dynsyslab.org/wp-content/p...
Work by Felix Berkenkamp and Angela P. Schoellig, Dynamic Systems Lab at the University of Toronto Institute for Aerospace Studies (UTIAS), Canada.
See also:
www.dynsyslab.org
www.schoellig.name
berkenkamp.me
Music: Christian Bjoerklund - Hallon
Source: freemusicarchive.org/music/Chr...
License: creativecommons.org/licenses/b... - Наука та технологія
Angela Schoellig, hi, I read your paper "Safe and Robust Learning Control with Gaussian Process", it's really interesting and I have some questions., I want to know how do you split the priori model f(x,u); and unknown model g(x,u); with the system states and inputs x & u. and also which parameter you considered for training in GP. please.
Great work! My question is the GP already takes into account the uncertainty, so what is the benefit of robust framework? Just stability? I think both stability and uncertainty can be considered in GP model. Would you please give some clarification in this regard?
Ali Baheri The GP learns a model with uncertainty bounds (i.e. it learns the model behavior and also gives us information about how certain it is about what it learned). What do we do with this model then? We need a control framework that guarantees stability for this uncertain model, this is robust control
Is there a conection betwenn machine learning and control systems? I am really interested in control systems as a whole but also like the math and the coding for training neural networks and dealing with machine learning in Matlab for example. What would you recommend? This video is awesome by the way! Thanks for posting it
Santiago Infantino Moreno In my opinion, there is a nice connection between machine learning and control. Machine learning is always useful, and useful in control, when it is difficult or impossible to describe the system behavior in terms of mathematical equations based on our prior knowledge (e.g., based on Newton's equations).
which parameter you considered for training in GP. for example the (Inputs and Targets).