Improving Bayesian Network computational and modeling methods, with applications in forensics
Within the Bayesian paradigm for statistics, posterior probability distributions for variables of interest are computed based on fully specified stochastics models, which may be described in the form of Bayesian network. For some simple networks, exact inference is possible, but in many cases, numerical methods must be used, or one must resort to MCMC simulation. However, exact bounds for the accuracy of results from MCMC simulations are often not avaliable. In forensic applications of Bayesian networks, this can be a particular problem.
In this project, we will develop inference methods for ILDI (Inference with Low Dimensional Intergation) networks, using numerical integration in such a way taht precise bounds for the accuracy of results are obtained. ILDI networks contain many of the types of models we see in forensi sciences. In a cooperation between the Swedish National Laboratory for Forensic Science, the National Veterinary Institute in Sweden, and Chalmers University, we will also work with a number of exampel forensic applications, aiming to study and support the general process of Bayesian network model building and utilization.