Investigation of Feedforward Backpropagation Artificial Neural Networks
Abstract
Yapay sinir ağları, biyolojik sinir sisteminden esinlenerek, sinirlerin farklı şekillerde birbirlerine bağlanmasıyla oluşan ağlardır. Bu ağlar, bilgiyi saklama, öğrenme ve kullanma özelliğine sahiptir. Böylece sınıflandırma, tahmin, ilişkilendirme ve en uygun şekle getirme işlemlerinde kullanılabilir. Çalışmada, İleri beslemeli geri yayılımlı yapay sinir ağları ile ilgili algoritmalar tanıtılarak, bir veri seti ile uygulama yapılmıştır. Uygulamada, serbest erişimli Machine Learning Repository internet sitesinden (UCI) sağlanan ve 81 açıklayıcı değişken içeren Süper İletken veri setinden 1111 veri kullanılmıştır. Ara katmandaki nöron sayısı 25, 50 ve 75 alınarak, algoritmalar için MATLAB paket programı ile toplam 13 uygulama yapılmıştır. Uygulamalarda, veri setinin % 70'i eğitim verisi, % 30'u ise test ve geçerlilik verisi olarak alınmış ve her uygulama için 1000 tekrar (iterasyon) yapılmıştır. Uygulamalar sonucunda gerçek değerler ile çıktı değerleri arasındaki korelasyonlar değerlendirilmiştir. Eğitim, geçerlilik ve test veri setleri birlikte incelendiğinde en yüksek korelasyon; % 95.46 ile trainbr algoritmasında gerçekleşirken, traingd, traingda, traingdm ve trainrp algoritmaları dışında en düşük korelasyon % 70.28 ile traingdx algoritmasında gözlenmiştir. Sonuç olarak, korelasyonların, algoritmalara ve ara katmandaki nöron sayısına göre belirgin değişiklik gösterdiği gözlenmiştir. Anahtar kelimeler: Algoritma, ara katman, iterasyon, aktivasyon fonksiyonu, eğitim verisi
Artificial neural networks are networks that are inspired by the biological nervous system and connected to each other in different ways. These networks have the ability for storage, learning and using of information. Thus, artificial neural networks can be used for classification, estimation, and association as well as providing in the most appropriate form. In this study, the algorithms related to feedforward back propagation artificial neural networks were introduced and performed an application with a data set. By 81 explanatory variables, 1111 data of Super-Conductor dataset that achieved from free-access Machine Learning Repository website (UCI) were used in the application. By considering number of neurons in the interlayer as 25, 50 and 75, a total of 13 applications were performed with MATLAB package program for the algorithms. 1000 iterations were performed for each application by taking 70% as training and 30% as validation and test of the data set. In the applications, correlations between the original and output values were evaluated. By excepting traingd, traingda, traingdm and trainrp algoritms, the highest correlation was found as 95.46% in trainbr algorithm while the lowest one 70.28% in traingdx algorithm for the training, validation and test data sets. Key words: Algorithm, interlayer, iteration, activation function, training data
Artificial neural networks are networks that are inspired by the biological nervous system and connected to each other in different ways. These networks have the ability for storage, learning and using of information. Thus, artificial neural networks can be used for classification, estimation, and association as well as providing in the most appropriate form. In this study, the algorithms related to feedforward back propagation artificial neural networks were introduced and performed an application with a data set. By 81 explanatory variables, 1111 data of Super-Conductor dataset that achieved from free-access Machine Learning Repository website (UCI) were used in the application. By considering number of neurons in the interlayer as 25, 50 and 75, a total of 13 applications were performed with MATLAB package program for the algorithms. 1000 iterations were performed for each application by taking 70% as training and 30% as validation and test of the data set. In the applications, correlations between the original and output values were evaluated. By excepting traingd, traingda, traingdm and trainrp algoritms, the highest correlation was found as 95.46% in trainbr algorithm while the lowest one 70.28% in traingdx algorithm for the training, validation and test data sets. Key words: Algorithm, interlayer, iteration, activation function, training data
Description
Keywords
Biyoistatistik, Biostatistics
Turkish CoHE Thesis Center URL
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
94