Optimasi Algoritma Back Propagation Neural Network Menggunakan Particle Swarm Optimization untuk Prediksi Harga Emas

Abstract

Backpropagation Neural Network (BPNN) algorithm is often used by researchers to predict prices or classification and pattern recognition. BPNN has a problem in random initialization of weights and continuous weight updates which make the weights not optimal. This problem provides research into optimization efforts such as combining (hybrid) BPNN with other optimization algorithms such as genetic algorithms, Particle Swarm Optimization (PSO) and others. This research was conducted on monthly USA gold price predictions. Optimizing the BPNN algorithm with PSO changes the backward process with the PSO algorithm and represents the weights into particles and the learning rate into c1 and c2 values. Optimized BPNN gave good results by reducing errors by 20.64% and reducing training time by 37.11%. The optimized algorithm has the ability to reduce errors better than BPNN which is stuck on its error value. An optimized algorithm does not necessarily get better results than before optimization. As many as 16.67% of the trials reduced error and training time. It is necessary to choose the right value of w and learning rate as well as architecture to get good results. Determining the number of particles also has an influence on training time and model performance. More particles will make training longer. This hybrid algorithm give a small error value as 0,101273 RMSE that is good for prediction.