search

whitewolfcheng / tlmpc

1 stars 2 forks source link

dimension #1

Open lucattycord opened 3 years ago

lucattycord commented 3 years ago

Hi，I just started reading ‘Learning model predictive control for iterative tasks. A data-driven control framework’ recently. I have some questions. 1.Is this method suitable for systems with large amount state?Such as 5 even 8？ 2.Is this method suitable for LTV? Thanks!

whitewolfcheng commented 3 years ago

hello and welcome to study together.

about your first question,in that project, it is a system with 6 state: 3 speed state and 3 position state. so i think it's suitable for large amount state such as 5 even 8.

as for the second question,LMPC is a method combine machine learning and mpc. the status update equation of mpc is just a ltv model use linear regression to identification speed state.so it is suitable for LTV,too.

by the way , the project i uploaded in github is an immature code for me to sync data between my school lab and my private computer.The original code was about the car iterating through LMPC algorithms on the same track by Ugo Rosolia. My work is to migrate the safe set of LMPC control so that the car can inherit the experience of previous iterations when the track changes.so most of the code I upload is directly using their original code to reduce the my workload.

My migration algorithm for LMPC is still being tested for modification. so if you want to study the LMPC algorithm ，I suggest you learn their code directly in “https://github.com/urosolia/RacingLMPC/tree/devel-ugo‘ and their paper .

lucattycord commented 3 years ago

thanks a lot！ I thought that the ‘safe set’ was similar to 'invariant set' in the tube MPC .Because in the paper it uses polynomial function to calculate the safe set. So I afraid that as dimension larger, the computation will explode. I am now looking for how to use terminal constraints to make the controller converge.There are few such methods for data-driven mpc.Do you know other methods besides this iterative convergence?

lucattycord commented 3 years ago

In addition to this method, I saw that Julian Berberich also has research on this. But that method has limitations on dimensions.

whitewolfcheng commented 3 years ago

I'm sorry I only got in touch with this method，And I don't know much about Julian Berberich's research . But I think since the dimension of the original program is 6, then the dimension is 5 or 8 should also be able . But if the dimensions are higher, Like dozens or hundreds，I don't know.

whitewolfcheng commented 3 years ago

And in his program, the state variable is an adjustable super parameter, in some calculations can choose to combine the state variable and the control variable for calculation, in that case the dimension is 6 + 2 = 8, so I think the appropriate addition dimension should be allowed

whitewolfcheng commented 3 years ago

And I look at his paper, there are also the algorithm used in the inverted pendulum. The state variables for the cart and the inverted pendulum are different dimensions, so the dimensions should be adjustable

lucattycord commented 3 years ago

Thank you very much! I encountered the problem that the cost function did not converge for the first time and I am going to try this method. I will talk to you if there are any progress.