An information theory approach for recursive LPV-ARX model identification via LS-SVM

When modeling a dynamical system, linear models are always the first choice due to their simplicity. However, many times the system is too complex so that employing linear models results in poor performances. In this context, Linear Parameter Varying (LPV) models allow to represent non-linear input/output relationships while preserving the simple structure of linear models. The method of Least Squares Support Vector Machines (LS-SVM) is one of the most common approaches to identify a LPV model in an ARX formulation (LPV-ARX). However, due to its computational cost, it is difficult to identify a LPV-ARX model using LS-SVM in online applications, where the model must be updated every time new data are collected. An efficient update algorithm has been presented for online identification of such models, where the idea is to update the model only upon certain data that are considered informative. However, this approach requires the tuning of some hyperparameters and in certain conditions can stop updating the model even when data are informative. This paper proposes an information-based algorithm to overcome these drawbacks. Evaluation on simulated and experimental data show the effectiveness of the proposed approach on both identification and computational sides.