We propose an Adaptive Kernel Bayes Filter (AKBF) that integrates the kernel Bayes rule with explicit dynamical models. Particles are evolved in data space while their kernel mean embeddings are propagated in reproducing kernel Hilbert space (RKHS). This hybrid approach avoids offline training, adapts to nonlinear dynamics, and mitigates instability issues of traditional KBR. Experiments on nonlinear systems demonstrate superior performance over particle filters and kernel Kalman variants, especially with fewer particles.
Adaptive Kernel Bayes Filter
Read on IEEE Xplore
