On Ensemble Learning Models for Autoencoder-Based Identification of Nonlinear Dynamical Systems

Identifying the dynamic model of a system has been a subject of extensive research for decades. However, dealing with highly complex, possibly nonlinear, and noisy systems remains an open problem. In recent years, the use of data-driven techniques based on machine learning for system identification has represented a promising solution for addressing the challenges of complex system identification. Among these methods, auto-encoders have been successfully applied to compactly represent the system state in a latent space used for estimating the system dynamics. In this work, we propose leveraging the ensemble paradigm for system identification, and in particular to exploit an ensemble of autoencoders, to reduce the uncertainty compared to individual autoencoders. To this end, we propose an algorithm for selecting autoencoders based on a combination of an accuracy metric and a diversity index. We validate this approach using traditional benchmarks in the field of nonlinear system identification.