Large scale reasoning under uncertainty

点击翻译按钮获取中文摘要

Individuals and society are increasingly faced with making decisions based on enormous amounts of information and they need to have a principled way to judge that information. Inspired by applications in geology, the first thrust of this proposal is to build the foundations of what we have called "semantic science" (in analogy with the semantic web, but for scientific data and hypotheses), to allow for ontologies that enable semantic interoperability, observational data that provides evidence, and hypotheses that make (probabilistic) predictions on data and can be used to form models to make predictions for specific cases. Making these all work together is a major research challenge. Spatial planning under uncertainty, such as arise in forestry applications, where effects of actions and utilities are not local is the second thrust. We are investigating policy gradient methods that iteratively improve relational representations of policies. A third thrust is in efficient inference for probabilistic relational models. In 2003 I first proposed lifted inference, where we do not distinguish individuals about which we have the same information. There has been considerable advances, but we still have not reached the "holy grail" where we can do inference exponentially faster than grounding as the number of indistinguishable individuals increases, although it seems plausible that there is such a method. This promises to form a foundation of the next generation of probabilistic modelling and programming languages. A fourth area is in reasoning about existence and identity uncertainty. Typically models refer to roles of individuals, but the observations of the world do not specify which individuals, if any, fulfill the roles of the models. This work requires advances in representations as well as reasoning and learning algorithms.