Weight Initialization on Neural Network for Neuro PID Controller: Case study

Theertham Akilesh Sai; HeeHyol Lee

doi:10.1109/ICT-ROBOT.2018.8549904

Weight Initialization on Neural Network for Neuro PID Controller: Case study

Theertham Akilesh Sai, HeeHyol Lee

Graduate School of Information, Production and Systems

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

5 Citations (Scopus)

Abstract

Neuro PID controller has been widely used in control field in recent times. Random weight initialization is used in the Neuro PID controller. The impact of various weight initialization has not been studied in the Neuro PID controller. The weight initialization methods such as Xavier initialization and He initialization have been proven to be effective in faster convergence in neural network. This paper investigated a weight initialization concept in Neuro PID controller by case studying with zero initialization, constant initialization, Gaussian distributed initialization, uniform distributed initialization, He initialization, and Xavier initialization in typical first-order lag elements, integrator elements, and dead time elements to obtain suitable initialization of weight coefficients, which reduces settling time for the neural network.

Original language	English
Title of host publication	2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018
Publisher	Institute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)	9781728119960
DOIs	https://doi.org/10.1109/ICT-ROBOT.2018.8549904
Publication status	Published - 2018 Nov 27
Event	2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018 - Busan, Korea, Republic of Duration: 2018 Sept 6 → 2018 Sept 8

Other

Other	2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018
Country/Territory	Korea, Republic of
City	Busan
Period	18/9/6 → 18/9/8

Keywords

Gaussian distributed initialization
He initialization
uniform distributed initialization
Xavier initialization
zero initialization

ASJC Scopus subject areas

Artificial Intelligence
Human-Computer Interaction
Electrical and Electronic Engineering
Control and Optimization
Communication

Access to Document

10.1109/ICT-ROBOT.2018.8549904

Cite this

Weight Initialization on Neural Network for Neuro PID Controller: Case study. / Sai, Theertham Akilesh; Lee, HeeHyol.
2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018. Institute of Electrical and Electronics Engineers Inc., 2018. 8549904.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Sai, TA & Lee, H 2018, Weight Initialization on Neural Network for Neuro PID Controller: Case study. in 2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018., 8549904, Institute of Electrical and Electronics Engineers Inc., 2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018, Busan, Korea, Republic of, 18/9/6. https://doi.org/10.1109/ICT-ROBOT.2018.8549904

@inproceedings{6a54557e7cde4712b96f10f782e5ee6b,

title = "Weight Initialization on Neural Network for Neuro PID Controller: Case study",

abstract = "Neuro PID controller has been widely used in control field in recent times. Random weight initialization is used in the Neuro PID controller. The impact of various weight initialization has not been studied in the Neuro PID controller. The weight initialization methods such as Xavier initialization and He initialization have been proven to be effective in faster convergence in neural network. This paper investigated a weight initialization concept in Neuro PID controller by case studying with zero initialization, constant initialization, Gaussian distributed initialization, uniform distributed initialization, He initialization, and Xavier initialization in typical first-order lag elements, integrator elements, and dead time elements to obtain suitable initialization of weight coefficients, which reduces settling time for the neural network.",

keywords = "Gaussian distributed initialization, He initialization, uniform distributed initialization, Xavier initialization, zero initialization",

author = "Sai, {Theertham Akilesh} and HeeHyol Lee",

year = "2018",

month = nov,

day = "27",

doi = "10.1109/ICT-ROBOT.2018.8549904",

language = "English",

booktitle = "2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

note = "2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018 ; Conference date: 06-09-2018 Through 08-09-2018",

}

TY - GEN

T1 - Weight Initialization on Neural Network for Neuro PID Controller

T2 - 2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018

AU - Sai, Theertham Akilesh

AU - Lee, HeeHyol

PY - 2018/11/27

Y1 - 2018/11/27

N2 - Neuro PID controller has been widely used in control field in recent times. Random weight initialization is used in the Neuro PID controller. The impact of various weight initialization has not been studied in the Neuro PID controller. The weight initialization methods such as Xavier initialization and He initialization have been proven to be effective in faster convergence in neural network. This paper investigated a weight initialization concept in Neuro PID controller by case studying with zero initialization, constant initialization, Gaussian distributed initialization, uniform distributed initialization, He initialization, and Xavier initialization in typical first-order lag elements, integrator elements, and dead time elements to obtain suitable initialization of weight coefficients, which reduces settling time for the neural network.

AB - Neuro PID controller has been widely used in control field in recent times. Random weight initialization is used in the Neuro PID controller. The impact of various weight initialization has not been studied in the Neuro PID controller. The weight initialization methods such as Xavier initialization and He initialization have been proven to be effective in faster convergence in neural network. This paper investigated a weight initialization concept in Neuro PID controller by case studying with zero initialization, constant initialization, Gaussian distributed initialization, uniform distributed initialization, He initialization, and Xavier initialization in typical first-order lag elements, integrator elements, and dead time elements to obtain suitable initialization of weight coefficients, which reduces settling time for the neural network.

KW - Gaussian distributed initialization

KW - He initialization

KW - uniform distributed initialization

KW - Xavier initialization

KW - zero initialization

UR - http://www.scopus.com/inward/record.url?scp=85060024091&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85060024091&partnerID=8YFLogxK

U2 - 10.1109/ICT-ROBOT.2018.8549904

DO - 10.1109/ICT-ROBOT.2018.8549904

M3 - Conference contribution

AN - SCOPUS:85060024091

BT - 2018 International Conference on Information and Communication Technology Robotics, ICT-ROBOT 2018

PB - Institute of Electrical and Electronics Engineers Inc.

Y2 - 6 September 2018 through 8 September 2018

ER -

Weight Initialization on Neural Network for Neuro PID Controller: Case study

Abstract

Other

Keywords

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this