Hybrid system identification using a mixture of NARX experts with LASSO-based feature selection

The availability of advanced hybrid system identifi-cation techniques is fundamental to extract knowledge in formof models from data streams. Starting from the current stateof the art, we propose an approach based on a specializedarchitecture, conceived to address the peculiar integration ofnonlinear dynamics and finite state switching behavior of hy-brid systems. Following the Mixtures of Experts concept, wecombine a set of Neural Network ARX (NNARX) models witha Gated Recurrent Units network with softmax output. Theformer are exploited to map specific nonlinear dynamical modelsrepresenting the behavior of the system in each discrete mode ofoperation. The latter, operating as a neural switching machine,infers the unobserved active mode and learns the state-transitionlogic, conditioned on input-output data sequences. Besides, weintegrate a LASSO based input features and model selectionmechanism, aimed to extract the most informative lags overthe sequences for each NNARX and calibrate the modes to beemployed. The overall system is trained end-to-end. Experimentshave been performed on a benchmark hybrid automata withnonlinear dynamics and transitions, showing the capability toachieve improved performances than conventional architectures.