ARMA MODEL SELECTION USING PARTICLE SWARM OPTIMIZATION AND AIC CRITERIA

Abstract

This paper presents a new method for determining ARMA model parameters using Particle Swarm Optimization (PSO). PSO is a new optimization method that is based on a social-psychological metaphor. Each ARMA model is represented as a particle in the particle swarm. Particles in a swarm move in discrete steps based on their current velocity, memory of where they found their personal best fitness value, and a desire to move toward where the best fitness value that was found so far by all of the particles during a previous iteration. PSO is applied for determining the ARMA parameters for the Wolfer Sunspot Data. The method is extended using Akaike's Information Criterion (AIC). PSO is used to simultaneously optimize and select an estimated “best approximating ARMA model” based on AIC. Several plots are included to illustrate how the method converges for various PSO parameter settings.