Ponavljajuće neuronske mreže i Gaussovi procesi za modeliranje nelinearne dinamike za modelsko prediktivno upravljanje

master's thesis

• Title (english): Recurrent neural networks and Gaussian processes for modeling of nonlinear dynamics for model predictive control
• Author: Ante Marić
• Mentor: Ivan Petrović (mentor)
• Committee member: Ivan Petrović (član povjerenstva)
• Committee member: Ivan Marković (član povjerenstva)
• Committee member: Luka Petrović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-02-07, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering

Abstract

Modelsko prediktivno upravljanje (MPC) moćna je metoda optimalnog upravljanja koja
se oslanja na model upravljanoga sustava. Za nelinearne dinamičke sustave, konvencionalne
tehnike modeliranja mogu postati neprimjenjive zbog složenog ponašanja i visokodimenzionalnog prostora stanja. Recentna istraživanja zbog toga su usmjerena na uključivanje modela strojnog učenja i dubokog učenja u MPC. Ovaj rad implementira metodu modeliranja dinamičkih sustava pomoću Gaussovih procesa i uspoređuje ju s Ponavljajućim neuronskim mrežama. Modeli su implementirani s regulatorom baziranim na modelskom prediktivnom integralu puta. Upravljačke sposobnosti uspoređene su s drugim regulatorima na simuliranom i fizičkom probnom sustavu. Dodatno, demonstrirana je mogućnost upravljanja fizičkim sustavom u stvarnom vremenu koristeći implementirani GP model treniran na maloj količini simuliranih podataka.

Abstract (english)

Model predictive control (MPC) is a powerful optimal control method that relies on the
model of the system that is being controlled. For nonlinear dynamical systems, conventional modeling techniques can become infeasible due to complex behavior and high-dimensional
state-space. Because of this, recent research has been focused on incorporating machine
learning and deep learning models with MPC. This work implements a Gaussian process
(GP) modeling method and compares it to Recurrent neural networks (RNNs). Both models
are implemented with a Model predictive path integral (MPPI) controller and tested against
other controllers on a simulated and physical example system. Furthermore, it is shown that
the implemented GP method is viable for real-time MPPI control of a physical system when
trained on small amounts of simulated data.

• Keywords: Modelsko prediktivno upravljanje; Modelski prediktivni integral puta; Gaussovi procesi; Ponavljajuće neuronske mreže
• Keywords (english): Model predictive control; Model predictive path integral; Gaussian processes; Recurrent neural networks
• Language: english
• URN:NBN: urn:nbn:hr:168:634326
• Study programme: Title: Electrical Engineering and Information Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije (magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-07-13 10:18:36