Adaptive importance sampling in particle filtering

Computational efficiency of the particle filter, as a method based on importance sampling, depends on the choice of the proposal density. Various default schemes, such as the bootstrap proposal, can be very inefficient in demanding applications. Adaptive particle filtering is a general class of algorithms that adapt the proposal function using the observed data. Adaptive importance sampling is a technique based on parametrization of the proposal and recursive estimation of the parameters. In this paper, we investigate the use of the adaptive importance sampling in the context of particle filtering. Specifically, we propose and test several options of parameter initialization and particle association. The technique is applied in a demanding scenario of tracking an atmospheric release of radiation. In this scenario, the likelihood of the observations is rather sharp and its evaluation is computationally expensive. Hence, the overhead of the adaptation procedure is negligible and the proposed adaptive technique clearly improves over non-adaptive methods.