Student Project: Integration of Evolutionary Algorithms to Optimize a Neuro-Fuzzy-Controller

HiWi- und Praktikantenstellen, Studienarbeiten

14.08.2010, Studentische Hilfskräfte, Praktikantenstellen, Studienarbeiten

Student Project: Integration of Evolutionary Algorithms to Optimize a Neuro-Fuzzy-Controller

Patients in the intensive care, especially after cardiac surgery, need to be monitored all the time. To keep the patients in a stable state, a physician is taking the current physiological signals into account to prescribe certain medication combinations. Programmable syringe pumps are available for this purpose, which can run at different speeds and intervals to infuse different drugs to the patient. In a joint project of TU München and Deutsches Herzzentrum München, a Neuro-Fuzzy controller is developed to analyse the patient's signals (e.g. ECG, arterial pressure, heart rate) and status (e.g. age, preconditions, medical condition) automatically and trigger the syringe pumps accordingly.
To be able to analyse a wide variety of patient status, a learning algorithm is used to adapt the experience and knowledge of the clinical personal. For the learning algorithm of a Neuro-Fuzzy inference system different approaches can be used. The simplest approach is the gradient decent. Disadvantages of gradient decent are on the one hand the slow performance in some cases and on the other hand the local minima, which can affect the optimization. That is why genetic algorithms shall be tested in the optimization of the Neuro-Fuzzy controller.

Description

The student will be part of the interdisciplinary research team at the German Heart Centre in Munich. The goal of this project will be, to develop a genetic algorithm to train a Neuro-Fuzzy inference system with test data. The implementation needs to be tested and compared to the gradient descent learning algorithm.
- Analysis of the Neuro-Fuzzy inference system used
- Development and integration of a genetic algorithm optimization in the learning process of the Neuro-Fuzzy inference system
- With training data the quality of the genetic algorithm shall be tested and compared with the performance and results of the gradient descent learning
The student has to analyse the Neuro-Fuzzy controller, its tasks and through which parameters an existing controller can be optimized. The genetic algorithm needs to be integrated to the C++ code of the existing learning algorithm and tested with training data. The implementation as well as the usage of the interface itself are to be fully documented.

Kontakt: n.sprunk@tum.de

Mehr Information

http://gsish.tum.edu/events-jobs/openings-for-student-projects/