Computation-Aware Data-Driven Model Discrimination with Application to Driver Intent Identification

In this paper, we consider the problem of designing a model discrimination algorithm for partially known systems, where only sampled data of the unknown dynamics are available. Leveraging data-driven abstraction methods to over-approximate the unknown dynamics and an incremental abstraction approach, we propose a method to find a pair of piecewise affine functions that "includes"all possible trajectories of the original unknown dynamics, which further simplify the data-driven abstraction and would scale better for high dimensional systems. Then, using the models from the abstraction method, we analyze the detectability of these models from noisy, finite data as well as design a model discrimination algorithm to rule out models that are inconsistent with a newly observed output trajectory, by checking the feasibility of mixed-integer linear programs. Moreover, we investigate the trade-off among the accuracy of abstraction models, the computational cost for obtaining reduced models and the guaranteed detection time T for distinguishing the models. Finally, we evaluate the effectiveness of our approach on a vehicle intent estimation example using the highD data set of naturalistic vehicle trajectories recorded on German highways.