An Outlier-Robust Growing Local Model Network for Recursive System Identification

In this paper, we develop a self-growing variant of the local model network (LMN) for recursive dynamical system identification. The proposed model has the following features: growing online structure, fast recursive updating rules, better memory use (no storage of covariance matrices is required), and outlier-robustness. In this regard, efficiency in performance and simplicity of implementation are the essential qualities of the proposed approach. The proposed growing version of the LMN model results from a synergistic amalgamation of two simple but powerful ideas. For this purpose, we adapt the neuron insertion strategy of the resource-allocating network to LMN model, and replaces the standard OLS rule for parameter estimation with outlier-robust recursive rules. A comprehensive evaluation involving three SISO and one MIMO benchmarking data sets corroborates the proposed approach's superior predictive performance in outlier-contaminated scenarios compared to fixed-size LMN-based models.