Reproduce “Sparse Kalman Identification for Partially Observable Systems via Adaptive Bayesian Learning” on OpenResearch

About this paper

Jilan Mei, Tengjie Zheng, Lin Cheng, Shengping Gong, Xu Huang

Researchers from Beihang University developed Sparse Kalman Identification (SKI), an online method for real-time dynamics identification in partially observable systems. SKI uses an Augmented Kalman Filter integrated with Automatic Relevance Determination to achieve sparse model discovery, demonstrating an 84.21% improvement in parameter estimation accuracy on a WingRock system and successfully identifying quadrotor thrust dynamics in real-world flight.

Key insights

• Integrating ARD within the AKF framework provides a principled Bayesian approach to induce sparsity by adaptively adjusting prior variances of basis function weights, allowing for online learning of model structure.
• Reinterpreting the ARD update as a Kalman-Filter-like correction step ensures that estimated states and model parameters dynamically adapt to the updated selection of basis functions.
• Leveraging the Unscented Kalman Filter enhances robustness and accuracy in nonlinear, partially observable, and noisy conditions by providing a more precise approximation of Gaussian distributions through nonlinear transformations.

Summary by alphaXiv.