Bitte benutzen Sie diese Kennung, um auf die Ressource zu verweisen:
http://dx.doi.org/10.18419/opus-11990
Autor(en): | Lefarov, Maksym |
Titel: | Model-based policy search for learning mulitvariate PID gain scheduling control |
Erscheinungsdatum: | 2018 |
Dokumentart: | Abschlussarbeit (Master) |
Seiten: | 74 |
URI: | http://nbn-resolving.de/urn:nbn:de:bsz:93-opus-ds-120074 http://elib.uni-stuttgart.de/handle/11682/12007 http://dx.doi.org/10.18419/opus-11990 |
Zusammenfassung: | Due to its simplicity and demonstrated performance, Proportional Integral and Derivative (PID) controller remains one of the most widely-used closed-loop control mechanisms in industrial applications. For the unknown model of a system, however, a PID design can become a significantly complex task especially for a Multiple Input Multiple Output (MIMO) case. For the efficient control of a nonlinear and non-stationary systems, a scheduled PID controller can be designed. The classical approach to gain scheduling is a system linearization and the design of controllers at different operating points with a subsequent application of interpolation. This thesis continues on the recent advances in application of Reinforcement Learning (RL) to a multivariate PID tuning. In this work we extend the multivariate PID tuning framework based on the Probabilistic Inference for Learning Control (PILCO) algorithm to tune a scheduled PID controllers. The developed method does not require the linear model of a system dynamics and is not restricted to the low-order or Single Input Single Output (SISO) systems. The algorithm is evaluated using the Noisy Cart-Pole and Non-stationary Mass-Damper systems. Additionally, the proposed method is applied to the tuning of a scheduled PID controllers of autonomous Remote Control (RC) car. |
Enthalten in den Sammlungen: | 05 Fakultät Informatik, Elektrotechnik und Informationstechnik |
Dateien zu dieser Ressource:
Datei | Beschreibung | Größe | Format | |
---|---|---|---|---|
18-lefarov-MSc.pdf | 3,3 MB | Adobe PDF | Öffnen/Anzeigen |
Alle Ressourcen in diesem Repositorium sind urheberrechtlich geschützt.