REDUKSI DIMENSI INPUT PADA JARINGAN SYARAF PCA-RBF DENGAN SINGULAR VALUE DECOMPOSITION

Abstract

Artificial neural network is an information processing system that has characteristics similar to biological neural networks. Artificial neural networks are divided into single layers and multiple layers. One of the multiple layer neural networks is Radial Basis Function (RBF). RBF is known to have high computing speed. However, the performance of RBF decreases when it involves the input space with high dimension so it requires simplification of the network. One method of simplifying RBF with respect to the dimension of input space is to use Principal Component Analysis (PCA). When the number of data variables is greater than the number of observations, the ability of PCA to be less effective then required Singular Value Decomposition (SVD) to solve the problem. The purpose of this research is to apply Singular Value Decomposition (SVD) process on PCA-RBF neural network. This study discusses the neural network PCA-RBF. PCA serves to reduce the input dimension of RBF. This dimension of input is known as the principal component (PC). PC determination process is done using PCA method combined with SVD. Furthermore, the PC is used as a new input to the RBF and a clustering process is performed on the PC using the K-means method for the initialization of the RBF center. Inisisalisasi center is the first step RBF in classification. The classification process in RBF consists of two processes namely training and testing. The result of this research is the SVD process on PCA to reduce the dimension of input data consisting of the process of determining the right singular matrix (V) ie calculating the ATA matrix, finding the eigenvalues (λ) and eigenvectors of the ATA matrix, conducting Gram-Schmidt and normalization , and the process of forming Principal Component (PC) is by multiplying the matrix of training data with right singular matrix (V), so that PC is used as new input to RBF. In this research is given example of classification data that is Landsat satellite data. After repeating 30 times the average success of classification in Landsat training data is 79,889% with mean error 20,111%, while for data testing Landsat obtained average success equal to 93,333% with error percentage is 6,667%.