A particle filter for probabilistic dynamic relational domains

We propose a probabilistic logic programming framework for the state estimation problem in dynamic relational domains such as Russell and Norvig's wumpus world and its probabilistic variants. The framework is based on the recently introduced notion of distributional clauses, an extension of Sato's distribution semantics with continuous distributions. The key contribution of this paper is that we introduce a particle filter for use with distributional clauses in dynamic relational domains and an unknown number of objects. The particles represent (partial) interpretations or possible worlds (with discrete and/or continuous variables) and the filter recursively updates its beliefs about the current state. Probabilistic background knowledge can be used to determine which variables must be included in the partial interpretations and magic sets or backward reasoning are employed to compute these.