Fixed-Size kernel models in System Identification under a reduced complexity perspective

Ricardo Castro; Siamak Mehrkanoon; Anna Marconato; Joannes Schoukens; Johan Suykens

Samenvatting

The goal of this work is to explore alternatives that allow the reduction of complexity of kernel models like Fixed Size Ordinary Least Squares (FS-OLS) and Fixed Size Ridge Regression (FS-RR). In general, black box approaches like these can achieve a good performance when undertaking system identification tasks, however, the complexity of the resulting models can be quite high. Under the conception that the Prediction Error is composed by a training error plus a complexity term, it is clear that a tradeoff between performance and complexity can be indeed a viable option. The focus of this work shifts from the traditional intention to maximize the performance to that of maximizing both: the performance and the complexity. This can result on a small loss of performance and a huge reduction of the complexity of the models used. Results found using Fixed Size Ordinary Least Squares and Fixed Size Ridge Regression on the Wiener-Hammerstein and Silverbox data set are presented.

Originele taal-2	English
Titel	Presentation of poster at ERNSI 2014, European Research Network on System Identification, Oostende, Belgium, September 21-24, 2014
Status	Published - 21 sep. 2014
Evenement	ERNSI 2014 - Thermae Palace Hotel, Ostend, Belgium Duur: 21 sep. 2014 → 24 sep. 2014

Workshop

Workshop	ERNSI 2014
Periode	21/09/14 → 24/09/14
Ander	Modelling of dynamical systems is fundamental in almost all disciplines of science and engineering, ranging from life science to plant-wide process control. Engineering uses models for the design and analysis of complex technical systems. System identification concerns the construction, estimation and validation of mathematical models of dynamical physical or engineering phenomena from experimental data. This is the 23rd version of the European Workshop on System Identification, the first one being held in Saint-Malo in 1992. All through these years the workshop has maintained the scope of bringing together European researchers in the area of System Identification, in an informal setting that gives ample opportunities for participants to meet. The workshop program is composed of lectures from invited speakers, lectures from members of the ERNSI community, and poster presentations by -particularly- the PhD students and postdocs that are active in the network.

Citeer dit

@inbook{dee2efe5270d4d8699b161030ab5ca7d,

title = "Fixed-Size kernel models in System Identification under a reduced complexity perspective",

abstract = "The goal of this work is to explore alternatives that allow the reduction of complexity of kernel models like Fixed Size Ordinary Least Squares (FS-OLS) and Fixed Size Ridge Regression (FS-RR). In general, black box approaches like these can achieve a good performance when undertaking system identification tasks, however, the complexity of the resulting models can be quite high. Under the conception that the Prediction Error is composed by a training error plus a complexity term, it is clear that a tradeoff between performance and complexity can be indeed a viable option. The focus of this work shifts from the traditional intention to maximize the performance to that of maximizing both: the performance and the complexity. This can result on a small loss of performance and a huge reduction of the complexity of the models used. Results found using Fixed Size Ordinary Least Squares and Fixed Size Ridge Regression on the Wiener-Hammerstein and Silverbox data set are presented.",

keywords = "Fixed Size Ordinary Least Squares (FS-OLS), Fixed Size Ridge Regression (FS-RR)",

author = "Ricardo Castro and Siamak Mehrkanoon and Anna Marconato and Joannes Schoukens and Johan Suykens",

year = "2014",

month = sep,

day = "21",

language = "English",

booktitle = "Presentation of poster at ERNSI 2014, European Research Network on System Identification, Oostende, Belgium, September 21-24, 2014",

note = "ERNSI 2014 ; Conference date: 21-09-2014 Through 24-09-2014",

}

TY - CHAP

T1 - Fixed-Size kernel models in System Identification under a reduced complexity perspective

AU - Castro, Ricardo

AU - Mehrkanoon, Siamak

AU - Marconato, Anna

AU - Schoukens, Joannes

AU - Suykens, Johan

PY - 2014/9/21

Y1 - 2014/9/21

N2 - The goal of this work is to explore alternatives that allow the reduction of complexity of kernel models like Fixed Size Ordinary Least Squares (FS-OLS) and Fixed Size Ridge Regression (FS-RR). In general, black box approaches like these can achieve a good performance when undertaking system identification tasks, however, the complexity of the resulting models can be quite high. Under the conception that the Prediction Error is composed by a training error plus a complexity term, it is clear that a tradeoff between performance and complexity can be indeed a viable option. The focus of this work shifts from the traditional intention to maximize the performance to that of maximizing both: the performance and the complexity. This can result on a small loss of performance and a huge reduction of the complexity of the models used. Results found using Fixed Size Ordinary Least Squares and Fixed Size Ridge Regression on the Wiener-Hammerstein and Silverbox data set are presented.

AB - The goal of this work is to explore alternatives that allow the reduction of complexity of kernel models like Fixed Size Ordinary Least Squares (FS-OLS) and Fixed Size Ridge Regression (FS-RR). In general, black box approaches like these can achieve a good performance when undertaking system identification tasks, however, the complexity of the resulting models can be quite high. Under the conception that the Prediction Error is composed by a training error plus a complexity term, it is clear that a tradeoff between performance and complexity can be indeed a viable option. The focus of this work shifts from the traditional intention to maximize the performance to that of maximizing both: the performance and the complexity. This can result on a small loss of performance and a huge reduction of the complexity of the models used. Results found using Fixed Size Ordinary Least Squares and Fixed Size Ridge Regression on the Wiener-Hammerstein and Silverbox data set are presented.

KW - Fixed Size Ordinary Least Squares (FS-OLS)

KW - Fixed Size Ridge Regression (FS-RR)

M3 - Meeting abstract (Book)

BT - Presentation of poster at ERNSI 2014, European Research Network on System Identification, Oostende, Belgium, September 21-24, 2014

T2 - ERNSI 2014

Y2 - 21 September 2014 through 24 September 2014

ER -

Fixed-Size kernel models in System Identification under a reduced complexity perspective

Samenvatting

Workshop

Vingerafdruk

Citeer dit