Not All Actions Are Created Equal: Bayesian Optimal Experimental Design for Safe and Optimal Nonlinear System Identification

TL;DR

This paper introduces a Bayesian optimal experimental design method for nonlinear system identification that efficiently gathers maximum information while ensuring safety, applicable to complex systems like vehicle models.

Contribution

The paper presents a novel Bayesian approach to experimental design that is safe, efficient, and applicable to nonlinear, non-Gaussian systems, reducing sample requirements significantly.

Findings

Requires orders of magnitude fewer samples than existing methods

Provides safety guarantees during system identification

Effective on high-fidelity vehicle models

Abstract

Uncertainty in state or model parameters is common in robotics and typically handled by acquiring system measurements that yield information about the uncertain quantities of interest. Inputs to a nonlinear dynamical system yield outcomes that produce varying amounts of information about the underlying uncertain parameters of the system. To maximize information gained with respect to these uncertain parameters we present a Bayesian approach to data collection for system identification called Bayesian Optimal Experimental Design (BOED). The formulation uses parameterized trajectories and cubature to compute maximally informative system trajectories which obtain as much information as possible about unknown system parameters while also ensuring safety under mild assumptions. The proposed method is applicable to non-linear and non-Gaussian systems and is applied to a high-fidelity vehicle model from the literature. It is shown the proposed approach requires orders of magnitude fewer samples compared to state-of-the-art BOED algorithms from the literature while simultaneously providing safety guarantees.