IMPROVING THE EFFICIENCY OF TIME SERIES FORECASTING BY INTEGRATING ARTIFICIAL NEURAL NETWORKS AND GENETIC ALGORITHMS

Abstract

Artificial neural networks (ANN) are known to be an effective forecasting method for time series data. However, a challenge of this method is how to determine an optimal set of parameters for ANN, such as number of inputs, number of hidden neurons, learning rate, set of weights, etc. so that ANN can achieve the most accurate forecasting results. Given the possible combinations of parameters, there is an explosion in the number of ANN architectures that need to be considered, so it is necessary to incorporate another intelligent technique, such as genetic algorithms (GA), to be able to find the best ANN. This paper analyzes the integration of ANN with GA, called the GANN model, in which ANNs are candidate solutions in the search space and GA tries to determine the best architecture for forecasting. Time series data on tourism demand is used to evaluate the effectiveness of the GANN model. Experimental results show that, ANN(12:14:1) is the best ANN architecture, with the minimum forecasting error (MAPE = 5.50). A comparison between ANN(12:14:1) and other architectures with the same inputs, but varied hidden neurons was also investigated and the results show that ANN(12:14:1) is the global optimal architecture.