UPT. PERPUSTAKAAN
Institut Teknologi Sepuluh Nopember Surabaya
Kampus ITS Sukolilo - Surabaya 60111
Phone
:
031-5921733 , 5923623
Fax
:
031-5937774
E-mail
:
libits@its.ac.id
Website
:
http://library.its.ac.id
Support (Customer Service) :
timit_perpus@its.ac.id
Welcome..guys!
Have a problem with your access?
Please, contact our technical support below:
LIVE SUPPORT
Davi Wahyuni
Tondo Indra Nyata
Anis Wulandari
Ansi Aflacha
ITS » Master Theses » Matematika - S2 Posted by aprill@is.its.ac.id at 29/09/2015 19:09:30 • 2052 Views
COMPARISON OF EXTENDED KALMAN FILTER AND GRADIENT DESCENT METHODS IN THE TRAINING OF RADIAL BASIS FUNCTION NETWORKS FOR IDENTIFICATION OF JAVANESE GAMELAN INSTRUMENT
Author : RISKI, ABDUH ( 1212 201 007 )
ABSTRAK
Pada instrumen gamelan tidak ada frekuensi standar dalam sistem tangga nadanya seperti pada instrumen musik modern. Gamelan dibuat oleh empu pembuat
gamelan berdasarkan pada perasaan dan pendengarannya begitu juga dalam hal
perawatan gamelan. Perawatan gamelan akan menjadi lebih efisien jika dapat dilakukan oleh selain empu pembuat gamelan. Untuk mengatasi hal tersebut dalam penelitian ini akan dilakukan identifikasi instrumen gamelan Jawa menggunakan jaringan fungsi basis radial. Jaringan fungsi basis radial merupakan jaringan multilayer feed-forward yang pelatihannya bersifat hybrid. Jaringan fungsi basis radial telah sering digunakan untuk pengklasifikasian identifikasi pola atau pengolahan sinyal karena proses pelatihannya yang cepat dibanding jaringan lain. Proses pelatihan jaringan fungsi basis radial sering dilakukan dengan metode gradient descent. Meskipun metode ini dikenal
lebih baik dari metode pelatihan konvensional lainnya tetapi pelatihan dengan gradient descent masih memerlukan waktu komputasi yang cukup lama. Sehingga dalam penelitian ini akan digunakan extended Kalman filter untuk mengoptimalkan akurasi dan waktu komputasi hasil pelatihan jaringan fungsi basis radial. Pelatihan
jaringan saraf dengan menggunakan extended Kalman filter dilakukan dengan
memformulasikan jaringan saraf sebagai konsep variabel keadaan yang mirip dengan sistem dinamik tak-linier.
Berdasarkan simulasi penggunaan metode pelatihan extended Kalman filter
untuk identifikasi jenis instrumen menghasilkan akurasi sebesar 100 sedangkan gradient descent hanya 8899. Untuk identifikasi nada instrumen extended Kalman filter menghasilkan akurasi sebesar 9904 dan gradient descent hanya menghasilkan akurasi sebesar 3726. Selain itu waktu komputasi pelatihan yang
diperlukan oleh extended Kalman filter lebih cepat dari pada waktu komputasi
pelatihan yang diperlukan oleh gradient descent. Sehingga extended Kalman filter lebih baik dari pada gradient descent sebagai metode pelatihan jaringan fungsi basis radial untuk identifikasi instrumen gamelan Jawa.
ABSTRACT
In gamelan instrument there is no standard frequency tone ladder system as in modern music instrument. Gamelan is made by master of gamelan instrument maker based on feeling and hearing as well as in terms of gamelan treatment. Gamelan treatment will be more efficient if can be done by other than master of gamelan instrument maker. To solve this problem this research will be identifying Javanese gamelan instrument using radial basis function networks. Radial basis function networks is a multilayer feed-forward networks whose training is hybrid. Radial basis function networks has been frequently used for classification identified patterns or signal processing because the training process
faster than the other networks. The training process of a radial basis function networks often done by gradient descent method. Although this method is known better than other conventional training methods the training using gradient descent still require considerable computational time. Thus in this research will be using extended Kalman filter for optimizing the accuracy and computational time of radial basis function networks training result. Neural networks training using extended Kalman filter done by formulating neural networks as a concept of state variable similar to the non-linear dynamic systems. Based on the simulation the using of extended Kalman filter training method
for identifying type of instrument produces an accuracy about 100 while gradient descent is only 8899. For instrument tone identification extended Kalman filter produces accuracy of 9904 and the gradient descent only produces accuracy of 3726. Beside that computational training time that required by extended
Kalman filter faster than the computational training time that required by gradient descent. Thus extended Kalman filter is better than the gradient descent as a radial basis function networks training method for identification of Javanese gamelan
instrument.
Keywords:
Extended Kalman Filter; Gamelan; Gradient Descent; Identifikasi; Jaringan Fungsi Basis Radial
Subject
: Kalman Filter
Contributor
Prof. Dr. Mohammad Isa Irawan, M.T.
Dr. Erna Apriliani, M.Si.
Date Create
: 29/09/2015
Type
: Text
Format
: pdf
Language
: Indonesian
Identifier
: ITS-Master-12103150001481
Collection ID
: 12103150001481
Call Number
: RTMa 518.1 Ris k
Source Master Theses of Mathematics, RTMa 518.1 Ris k, 2015
Coverage ITS Community
Rights Copyright @2015 by ITS Library. This publication is protected by copyright and per obtained from the ITS Library prior to any prohibited reproduction, storage in a re transmission in any form or by any means, electronic, mechanical, photocopying, reco For information regarding permission(s), write to ITS Library