Spatial Operators for Evolving Dynamic Bayesian Networks from Spatio-temporal Data

Learning Bayesian networks from data has been studied extensively in the evolutionary algorithm communities [Larranaga96, Wong99]. We have previously explored extending some of these search methods to temporal Bayesian networks [Tucker01]. A characteristic of many datasets from medical to geographical data is the spatial arrangement of variables. In this paper we investigate a set of operators that have been designed to exploit the spatial nature of such data in order to learn dynamic Bayesian networks more efficiently. We test these operators on synthetic data generated from a Gaussian network where the architecture is based upon a Cartesian coordinate system, and real-world medical data taken from visual field tests of patients suffering from ocular hypertension.