An Artificial-Intelligence-Based Method to Automatically Create Interpretable Models from Data Targeting Embedded Control Applications

Abstract

The development of new automotive drivetrain layouts requires modeling of the involved components to allow for ideal control strategies. The creation of these models is both costly and challenging, specifically because interpretability, accuracy, and computational effort need to be balanced. In this study, a method is put forward which supports experts in the modeling process and in making an educated choice to balance these constraints. The method is based on the artificial intelligence technique of genetic programming. By solving a symbolic regression problem, it automatically identifies equation-based models from data. To address possible system complexities, data-based expressions like curves and maps can additionally be employed for the model identification. The performance of the method is demonstrated based on two examples: 1. Identification of a pure equation based model, demonstrating the benefit of interpretability. 2. Creation of a hybrid-model, combining a base equation with data-based expressions. Possible applications of the method are model creation, system identification, structural optimization, and model reduction. The existing implementation in ETAS ASCMO-MOCA also offers a high efficiency increase by combining and automating the two procedural steps of embedded function engineering and calibration.