Application of Neural Network Variations for Determining the Best
Architecture for Data Prediction

Application of Neural Network Variations for Determining the Best
Architecture for Data Prediction

Data Prediction, Backpropagation, Resilent Backpropagation, Conjugate Gradient, Fletcher Reeves, Powell Beal.

This study focuses on the application and comparison of the epoch, time, performance/MSE training, and performance/MSE
testing of variations of the Backpropagation algorithm. The main problem in this study is that the Backpropagation algorithm
tends to be slow to reach convergence in obtaining optimum accuracy, requires extensive training data, and the optimization
used is less efficient and has performance/MSE which can still be improved to produce better performance/MSE in this
research—data prediction process. Determination of the best model for data prediction is seen from the performance/MSE
test. This data prediction uses five variations of the Backpropagation algorithm: standard Backpropagation, Resistant
Backpropagation, Conjugate Gradient, Fletcher Reeves, and Powell Beale. The research stage begins with processing the
avocado production dataset in Indonesia by province from 2016 to 2021. The dataset is first normalized to a value between 0
to 1. The test in this study was carried out using Matlab 2011a. The dataset is divided into two, namely training data and test
data. This research's benefit is producing the best model of the Backpropagation algorithm in predicting data with five methods
in the Backpropagation algorithm. The test results show that the Resilient Backpropagation method is the best model with a
test performance of 0.00543829, training epochs of 1000, training time of 12 seconds, and training performance of 0.00012667

Mochamad Wahyudi1
, Firmansyah2
, Lise Pujiastuti3
, Solikhun4

Universitas Nusa Mandiri

Universitas Nusa Mandiri

PDG

Indonesia

Text