Optimization of Climate Dowscaling Using Gradient Descent with Momentum and Quasi-Newton Methods

บทคัดย่อ

Abstract

This paper presents two optimization methods in training algorithm to minimize error of the feed forword neural network.Gradient descent with momentum and Quasi-Newton methods are applied to optimize weights in iteration of a training network model. Data from a global model are downscaled to four provinces in Thailand namely: Chaingmai, Bangkok, Ubonratchathani and Phuket. The results of experiments show that the Quasi-Newton method can minimize the error better than Gradient descent with the momentum method. Moreover, the number of hidden nodes of the network structure also affected the regression between the output and observed data.

Keywords: component, Downscaling, Artificial Neural Network, Gradient Descent, Quasi-Newton