Spatiotemporal Learning via Mixture Importance Gaussian Filtering With Sparse Regularization

Abstract

This letter proposes a novel spatiotemporal learning model via mixture importance Gaussian filtering (MIGF). In the MIGF, we explore the causal mapping between the target's true parameters and the latest measurement, and present an approach for elegantly combining the deterministic Gaussian-Hermite integral with the stochastic importance sampling method. This formulation allows for the use of soft-constrained and sparse regularization to reduce the truncation error and improve the weight adaptivity. We also provide an effective causal invariant updating rule to learn the parameters of this constrained dynamic model with a convergence guarantee. Experimental results show that MIGF can effectively eliminate the constraint uncertainty for the constrained dynamics.