Advanced

Study of early termination of MPC Algorithms

Henriks, Gustav (2016)
Department of Automatic Control
Abstract
With a steady development of technology, the use of Model Predictive Control (MPC) has become more and more popular since the computation time has gone down. With this increase, a need for determining which MPC algorithm is good for solving a certain type of MPC problem has occurred which would facilitate the choice of algorithm and also could increase the performance. One of the determining aspects is if a solver has the possibility to terminate early (stop the algorithm before it has performed all of its iterations), for example if it is placed on an embedded system with strict real-time bounds. With the use of an MPC Benchmarking suite available at ABB Corporate Research Switzerland, the early termination of MPC algorithms has been... (More)
With a steady development of technology, the use of Model Predictive Control (MPC) has become more and more popular since the computation time has gone down. With this increase, a need for determining which MPC algorithm is good for solving a certain type of MPC problem has occurred which would facilitate the choice of algorithm and also could increase the performance. One of the determining aspects is if a solver has the possibility to terminate early (stop the algorithm before it has performed all of its iterations), for example if it is placed on an embedded system with strict real-time bounds. With the use of an MPC Benchmarking suite available at ABB Corporate Research Switzerland, the early termination of MPC algorithms has been investigated. With the usage of 19 benchmarks and 3 different solvers that uses Interior point method, Gradient descent and Active-set method a large number of results have been looked through with the help of different Machine Learning methods. The result has been classified as good or bad performance when terminated early and different models have been fitted to predict this data. From this a group of key-features have been attempted to get extracted to see if there is a possibility on beforehand to tell if a control problem and a certain solver can be early terminated. Important features that were found were mostly concerning whether the control input u was under constraint or not. Good results were especially achieved for a machine learning model based on the Active-set solver qpOASES which could give good indications on whether a certain problem could get early terminated or not. (Less)
Please use this url to cite or link to this publication:
author
Henriks, Gustav
supervisor
organization
year
type
H3 - Professional qualifications (4 Years - )
subject
report number
TFRT-6023
ISSN
0280-5316
language
English
id
8896087
date added to LUP
2016-12-16 13:58:41
date last changed
2016-12-16 13:58:41
@misc{8896087,
  abstract     = {With a steady development of technology, the use of Model Predictive Control (MPC) has become more and more popular since the computation time has gone down. With this increase, a need for determining which MPC algorithm is good for solving a certain type of MPC problem has occurred which would facilitate the choice of algorithm and also could increase the performance. One of the determining aspects is if a solver has the possibility to terminate early (stop the algorithm before it has performed all of its iterations), for example if it is placed on an embedded system with strict real-time bounds. With the use of an MPC Benchmarking suite available at ABB Corporate Research Switzerland, the early termination of MPC algorithms has been investigated. With the usage of 19 benchmarks and 3 different solvers that uses Interior point method, Gradient descent and Active-set method a large number of results have been looked through with the help of different Machine Learning methods. The result has been classified as good or bad performance when terminated early and different models have been fitted to predict this data. From this a group of key-features have been attempted to get extracted to see if there is a possibility on beforehand to tell if a control problem and a certain solver can be early terminated. Important features that were found were mostly concerning whether the control input u was under constraint or not. Good results were especially achieved for a machine learning model based on the Active-set solver qpOASES which could give good indications on whether a certain problem could get early terminated or not.},
  author       = {Henriks, Gustav},
  issn         = {0280-5316},
  language     = {eng},
  note         = {Student Paper},
  title        = {Study of early termination of MPC Algorithms},
  year         = {2016},
}