Year: 2009
Type: dissertation
Abstract: Practical use of optimization, for MPC or parameter estimation, requires a large number of gradient calculations.These gradients are used to compute search directions, for instance in a SQP algorithm.Computing gradients is time-consuming and limits the use of for instance MPC to small and medium-sized systems.Gradients are usually computed by finite difference methods.Alternatives include forward-and adjoint-based methods.The latter is efficient for problems with many decision variables and few outputs.Efficiency may however deteriorate in cases with output constraints which typically are present in MPC.In this project which continues earlier work the use of adjoints as a means to increase efficiency of optimization in MPC will be further studied.Tasks:1. Present adjoint-based methods and review central literature.The presentation shall focus on MPC which apply a sequential approach.2. Study efficient ways to compute the impulse response matrix (including changes in the MPC formulation), in particular how adjoints can aid this.3. Output constraints can be detrimental to the efficiency of adjoint-based methods.Hence, optimization algorithms where explicit output constraints can be removed, as for instance in barrier methods, is an interesting option to exploit the efficiency of adjoint-based methods.Propose methods in which the efficiency of adjoints can be exploited.4. Evaluate the methods above by comparing them with forward methods.This should be done by identifying suitable examples as well as test scenarios for evaluating the methods.The thesis report may include a paper to a selected conference with the main results of this work.
Author: Ruben Køste Ringset
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1553359573