Hybrid System Identification through Optimization and Active Learning

We present a method to identify state-dependent switched nonlinear dynamical systems with polynomial ODEs through optimization and active learning. Our approach extends and incorporates segmentation into a previous optimization-based approach for identifying SARX models. We use logistic regression to find the polynomial or linear mode boundaries of the system. Additionally, we provide a way to refine the result of the classifier through active learning and equivalence queries, assuming the correct identification of continuous dynamics. We provide results of our approach on multiple experiments, including a parametric experiment with increasing number of modes. We also compare our results with a different approach that deals with a similar class of problems and show that our method performs better. Copyright (c) 2024 The Authors.