ALVEN: Algebraic learning via elastic net for static and dynamic nonlinear model identification

Abstract

An algorithm is proposed that combines nonlinear feature generation and sparse regression to learn interpretable nonlinear models from noisy and limited data. This Algebraic Learning Via Elastic Net for Static and Dynamic Nonlinear Model Identification algorithm employs automated feature generation including families of ubiquitous chemical and biological nonlinear transformations. ALVEN balances model complexity and prediction accuracy through a two-step feature selection procedure, to produce an interpretable model useful for process applications while avoiding overfitting. The generalization to nonlinear dynamical systems, Dynamic ALVEN, is then described. The model accuracy of the algorithms is compared to well-established machine learning methods for a 3D printer and a chemical reactor.