UCLA

Probabilistic graphical models, e.g. Bayesian Networks, have been traditionally introduced to model and reason with uncertainty. A graph structure is crafted to capture knowledge of conditional independence relationships among random variables, which can enhance the computational complexity of reasoning. To generate such a graph, one sometimes has to provide vast and detailed knowledge about how variables interacts, which may not be readily available. In some cases, although a graph structure can be obtained from available knowledge, it can be too dense to be useful computationally. In this dissertation, we propose a new type of probabilistic graphical models called a Structured Bayesian network (SBN) that requires less detailed knowledge about conditional independences. The new model can also leverage other types of knowledge, including logical constraints and conditional independencies that are not visible in the graph structure. Using SBNs, different types of knowledge act in harmony to facilitate reasoning and learning from a stochastic world. We study SBNs across the dimensions of modeling, inference and learning. We also demonstrate some of their applications in the domain of traffic modeling.