Optimizing the architecture of artificial neural networks by genetic algorithm in time series data forecasting

Abstract

Artificial neural networks (or neural networks) are known to be effective forecasting methods for time series data. However, a challenge of this method is how to determine an optimal set of parameters, such as number of inputs, number of hidden neurons, learning rate, matrix of link weights, etc. so that the neural network can achieve the most accurate prediction results. Given the requirement to incorporate many different parameters when designing, there is obviously an explosion in the number of different neural network architectures that need to be considered, so combining with another intelligent technique, like genetic algorithm, is necessary to be able to find the best neural network architecture. This paper will analyze the integration of neural networks with genetic algorithms (GANN model) in which each neural network architecture is a candidate solution in the search space and the genetic algorithm tries to identify the best architecture for forecasting. Time series data on tourist arrivals to Thua Thien Hue was used to evaluate the effectiveness of the GANN model. The research results show that the genetic algorithm has determined that the architecture ANN(12:14:1) is the best neural network, with the prediction error achieved MSE = 3.34E+08. The GANN model also improved by 22.24% compared to ANN(12:6:1), which is the best pure ANN model in a recent study.