Learning by Simplified Cost-Reference Particle Filtering using Biased Data

In this paper we address the problem of online learning by cost-reference particle filtering combined with Kalman filtering. We propose an efficient learning scheme applicable to problems where some of the unknowns of a dynamic system of interest are linear given the remaining unknowns, which are nonlinear. To that end, we exploit a concept that is analogous to Rao-Blackwellization, and we implement it by using only one Kalman filter. The resulting algorithm is tested and compared to standard particle filtering for the problem of target tracking using bearings-only measurements acquired by two sensors.