GRADIENT BASED OPTIMIZATION IN CASCADE FORWARD NEURAL NETWORK MODEL FOR SEASONAL DATA

摘要

Optimization technique is an important part in neural network modeling for obtaining the network weights. The choosing a certain optimization method would influenced the prediction result. Many gradient based optimization methods have been proposed. In this research, we applied the three optimization techniques for obtaining the weights of Cascade Forward Neural Network (CFNN), they were Levenberg-Marquardt, Conjugate Gradient and Quasi Newton BFGS. In CFNN, there are direct connection between input layer and output layer, beside the indirect connection via the hidden layer. The advantage is that this architecture allows the nonlinear relationship between input layer and output layer by not disappear the linear relationship between the two. The proposed model was applied in the time series data with the seasonal pattern. The two data types were used to select the most appropriate optimization method for seasonal series. The first type was the generated data from seasonal ARIMA model and the second was the rainfall data of ZOM 145 at Jumantono Ngadirojo Wonogiri. After processing the proposed methods by using Matlab routine we recommended to use the Levenberg Marquardt as the chosen one.