Bayesian Approach for Neural Network

Abstract

Global warming gets some attention from many countries in the world because it is feared give negative impacts. Neural Network (NN) model can capture nonlinear relationship, so this method widely used in weather research but over-fitting may happen. Therefore to prevent over-fitting of NN model, then use a Bayesian approach. Bayesian approach with Markov Chain Monte Carlo (MCMC) used to estimate parameters of NN model. Feed Forward Neural Network (FFNN) and Bayesian Neural Network were proposed to model and predict monthly temperature in Surabaya, Indonesia. The best model selection for two models based on minimum error value. Criterion of minimum error value used to evaluate fit model and forecasting, which is AIC, SBC and RMSE. The result showed that AIC and SBC of Bayesian Neural Network model smaller than NN model for training data. Based on testing data, for a small lead better used Bayesian Neural Network.