Particle Swarm Optimization for NARX structure selection — Application on DC motor model

This paper explores the application of the Binary Particle Swarm Optimization (BPSO) by (Kennedy and Eberhart, 1997) to perform model structure selection of a Nonlinear Auto-Regressive model with Exogenous Inputs (NARX) identification of a Direct Current (DC) motor. We describe the application of BPSO for model structure selection, by representing its particles´ solutions as probabilities of change (bit flip) of a binary string. The binary string was then used to select a set of regressors from the regressor matrix, then estimate the coefficients (linear least squares solution) of the reduced regressor matrix using QR decomposition. Tests performed on a simulated DC motor dataset showed that the BPSO-based selection method has the potential to become an effective method to determine parsimonious NARX model structure in the system identification model.