Data Envelopment Analysis for Selection of the Fitness Function in Evolutionary Algorithms Applied to Time Series Forecasting Problem

Abstract

Artificial Neural Networks (ANN) have been widely used in time series forecasting problem. However, a more promising approach is the combination of ANN with other intelligent techniques, such as genetic algorithms, evolutionary strategies, etc, where these evolutionary algorithms have the objective of train and adjust all parameter of the ANN. In the evolutionary process is necessary define a fitness function to guide the evolve process. Thus, for a set of possibles fitness function, how to determine the function more efficient? This paper aims to select the efficient fitness functions, through the use of Data Envelopment Analysis. This tool determines the relative efficiency of each unit under review, comparing it with each other and considering the relationship between inputs and outputs. Two different times series were used to benchmark the set of twenty fitness functions. The preliminary results show the proposed method is a promising approach for efficient selecting the fitness function.