Approximate Online Inference for Dynamic Markov Logic Networks

We examine the problem of filtering for dynamic probabilistic systems using Markov Logic Networks. We propose a method to approximately compute the marginal probabilities for the current state variables that is suitable for online inference. Contrary to existing algorithms, our approach does not work on the level of belief propagation, but can be used with every algorithm suitable for inference in Markov Logic Networks, such as MCSAT. We present an evaluation of its performance on two dynamic domains.