Knowledge Based Control Systems
Responsible Instructor:Dr.-Ing. J. Kober (Jens)
Contact Hours / Week x/x/x/x:0/0/4/0
Course Contents:Theory and applications of knowledge-based and intelligent control systems, including fuzzy logic control and artificial neural networks:
* Introduction to intelligent control
* Fuzzy sets and systems
* Intelligent data analysis and system identification
* Knowledge based fuzzy control (direct and supervisory)
* Artificial neural networks, learning algorithms
* Control based on fuzzy and neural models
* Reinforcement learning
* Examples of real-world applications
Study Goals:Main objective: understand and be able to apply 'intelligent control' techniques, namely fuzzy logic and artificial neural networks to both adaptive and non-adaptive control.
After successfully completing the course, the student is able to:
* Name the limitations of traditional linear control methods and state the motivation for intelligent control. Give examples of intelligent control techniques and their applications.
* Formulate the mathematical definitions of a fuzzy set and the associated concepts and properties (alpha-cut, support, convexity, normality, etc.), basic fuzzy set-theoretic operators, fuzzy relations and relational composition.
* Explain the notion of a fuzzy system and define the Mamdani, Takagi-Sugeno and singleton fuzzy model. State and apply the compositional rule of inference and the Mamdani algorithm. Define and apply the center of gravity and the mean of maxima defuzzification method.
* Describe how fuzzy models can be constructed from data, give examples of techniques for antecedent and consequent parameter estimation. Compute consequent parameters in Takagi-Sugeno fuzzy model by using the least-squares method.
* Explain the difference between model-based and model-free fuzzy control design. Give the basic steps in knowledge-based fuzzy control design. Define a low-level and a high-level (supervisory) fuzzy controller, explain the differences.
* Explain the concept of an artificial neural network and a neuro-fuzzy network, give some examples and explain the differences. Define and apply the back-propagation training algorithm. Explain the difference between first-order and second-order gradient methods.
* Show how dynamics are incorporated into fuzzy models and neural networks, give examples. Discuss how dynamic models can be identified from data.
* Give block diagrams and explain the notions of inverse-model control, predictive control, internal model control, direct and indirect adaptive control. Explain the meaning of the variables and parameters in recursive least-squares estimation.
* Explain the motivation and the basic elements of reinforcement learning. Define and explain the concepts of value function, Bellman equation, value iteration, Q-iteration, on-line reinforcement learning algorithms, actor-critic control scheme.
* Define hard, fuzzy and possibilistic partitions, explain the fuzzy c-means algorithm and its parameters.
* Implement and apply the above concepts to a simulated nonlinear process or a given data set, using Matlab and Simulink.
Education Method:Lectures and two assignments - literature assignment and practical Matlab / Simulink assignment.
Literature and Study Materials:Lecture notes: R. Babuska. Knowledge-Based Control Systems. Slides and other course material (software, demos) can be downloaded from the course Website (www.dcsc.tudelft.nl/~sc42050).
Assessment:Written exam, closed book.
* SC42050 (TOETS-01) The exam constitute 60% of the final mark
* SC42050 (TOETS-02) Literature assignment 20% of the final mark
* SC42050 (TOETS-03) Practical Matlab / Simulink assignment 20% of the final mark.
A mini-symposium is organized in order for the students to present the results of the literature assignment.