Monty Hall Particle filter: A new method to tackle predictive model uncertainties

Abstract

This paper proposes a simple adaptive weight computing method for particle filters that utilizes knowledge about predictive model uncertainty. In each time step the particles are assigned into subsets based on the corresponding uncertainty estimates. The weights are then updated based on accumulated subset-inclusion and likelihood information using a discrete set of measurement likelihood functions. By controlling the aggressiveness of the weight computing, the method strives to achieve faster convergence without losing robustness to model errors. Two localization experiments are conducted to verify that the method has a clear advantage over particle filters with single likelihood function. In the first experiment we use synthetic Gaussian Process data. In the second experiment real indoor magnetic field data with very coarse interpolation and uncertainty approximation is used to verify the method's effectiveness in real-world scenarios. One of the main advantages of the proposed method is that despite its flexibility, it adds only little implementational or computational overhead to conventional particle filters.