Exploiting sparsity and sharing in probabilistic sensor data models

Probabilistic sensor models defined as dynamic Bayesian networks can possess an inherent sparsity that is not reflected in the structure of the network. Classical inference algorithms like variable elimination and junction tree propagation cannot exploit this sparsity. Also, they do not exploit the opportunities for sharing calculations among different time slices of the model. We show that, using a relational representation, inference expressions for these sensor models can be rewritten to make efficient use of sparsity and sharing.