DIFFERENTIAL EVOLUTION - SIMULATED ANNEALING (DESA) ALGORITHM FOR FITTING AUTOREGRESSIVE MODELS TO DATA

Abstract

In this paper, we propose Differential Evolution - Simulated Annealing (DESA), a hybrid optimization algorithm for fitting autoregressive models to data. The addition of a new strategy based on parabolic estimation to Differential Evolution (DE) algorithm and the incorporation of the Simulated Annealing (SA) algorithm for a selection strategy makes DESA a robust optimization algorithm. The proposed hybrid algorithm obtained acceptable solutions particularly for AR(1) processes with unknown drift and additive outliers. Experiments on the parameter estimation of autoregressive models showed that the proposed hybrid algorithm, DESA has shown reliability in finding global minimum of the reference problem sets. Moreover, we compared the performance of DESA algorithm with those of DE, SA, maximum likelihood estimator (MLE) and ordinary least squares (OLS) on the fitting problem. Simulation results have shown that the proposed algorithm, DESA, provides MSE lower than those of MLE and/or OLS for almost all situations. Using 10-minute average wind speed data, DESA also obtained a better model fit on the actual series.