Automatic Control

Faculty of Engineering, LTH

Denna sida på svenska This page in English

FRTN75 - Learning-Based Control

Inlärningsbaserad reglering, 7.5 hp

General Information

Elective for: BME4, C4, D4-ssr, D4-mai, E4-ra, F4, F4-r, F4-mai, Pi4-ssr, MMSR1
Language of instruction: The course will be given in English


The course provides fundamental theory and methodology for developing control laws based on measured input and output signal data. The aim of the course is that the students should learn the important principles within the area of learning-based control, and to understand their limitations.

Learning outcomes

Knowledge and understanding
For a passing grade the student must

  • be able to understand algorithms for multivariable system identification, including estimations of model errors
  • understand the importance of excitation
  • uinderstand the basic principle of dynamic programming
  • understand the assumptions and approximations behind common variants of reinforcement learning
  • understand the fundamental properties and limitations of components based on machine learning, such as neural networks and classifiers.

Competences and skills
For a passing grade the student must

  • be able to implement multivariable system identification procedures and perform model choices and determine how to analyse a given dataset
  • be able to simulate and use identified models in a multivariable control synthesis
  • be able to implement simple adaptive controllers
  • be able to apply some variants of reinforcement learning
  • be able to implement path-following algorithms based on dynamic programming.

Judgement and approach
For a passing grade the student must

  • understand the achievable confidence when using data-driven control methods
  • show ability for teamwork and group collaboration during laboratory experiments.


The development of suitable models for describing dynamical systems is a central problem within automatic control, and it is critical for the development of robust and high performance control laws. When relationships between physical quantities are not fully known, then models and the control laws may instead be generated by measurement data, through system identification, machine learning, or adaptive control. The purpose of the course is to teach the basic principles of how this is done.

The first part of the course is devoted to adaptive control and system identification for systems with several input and output signals. The focus is on state-space models and methods for generating these, including greybox identification. We describe iterative methods for learning, as well as model reduction for the purpose of reducing the dimension of the state space.

The second part of the course is devoted to reinforcement learning. This includes the theory of dynamic programming and various approximate methods thereof. Policy iteration is explained, as well as discrete and continuous path planning.

The third part of the course deals with the usage of complete components for the purpose of control, for instance sensors, that have been developed using machine learning.

Examination details

Grading scale: TH - (U,3,4,5) - (Fail, Three, Four, Five)
Assessment: Written exam (5 hours), three laboratory exercises. In the case of less than 5 registered students, the retake exams may be given in oral form.


Assumed prior knowledge: FRTF05 Automatic Control, Basic Course.
The number of participants is limited to: 60
Selection: Completed university credits within the programme. Priority is given to students enrolled on programmes that include the course in their curriculum.

Reading list

  • Lennart Ljung: System Identification: Theory for the user. Pearson Education, 1998, ISBN: 0136566952.
  • Lecture slides, exercise material and laboratory manuals are available on the course homepage.

Contact and other information

Course coordinator: Anders Rantzer,
Director of studies: Anton Cervin,
Teacher: Bo Bernhardsson,
Further information: Replaces FRTN15 Predictive Control.

Page Manager: