Randomized algorithms for nonlinear system identification with deep learning modification

Both randomized algorithms and deep learning techniques have been successfully used for regression and classification problems. However, the random hidden weights of the randomized algorithms require suitable distributions in advance, and the deep learning methods do not use the output information i...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Information sciences Jg. 364-365; S. 197 - 212
Hauptverfasser: de la Rosa, Erick, Yu, Wen
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Inc 01.10.2016
Schlagworte:
Algorithms
Benchmarking
Classification
Deep learning
Dynamical systems
Learning
Nonlinear system modeling
Randomized algorithms
Regression
System identification
Training
Deep learning
Nonlinear system modeling
Randomized algorithms
ISSN:0020-0255, 1872-6291
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Both randomized algorithms and deep learning techniques have been successfully used for regression and classification problems. However, the random hidden weights of the randomized algorithms require suitable distributions in advance, and the deep learning methods do not use the output information in system identification. In this paper, the distributions of the hidden weights are obtained by the restricted Boltzmann machines. This deep learning method uses input data to construct the statistical features of the hidden weights. The output weights of the neural model are trained by normal randomized algorithms. So we successfully combine the unsupervised training (deep learning) and the supervised learning method (randomized algorithm), and take advantages from both of them. The proposed randomized algorithms with deep learning modification are validated with three benchmark problems.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:0020-0255
1872-6291
DOI:10.1016/j.ins.2015.09.048