EPiC Series in Computing Volume 50, 2017, Pages 233–238 GCAI 2017. 3rd Global Con- ference on Artificial Intelligence Abduction for Learning Smart City Rules Nikolaj Bjørner 1 , Maria-Cristina Marinescu 2 , and Mooly Sagiv 34 1 Microsoft Research, Redmond, Washington, USA nbjorner@microsoft.com 2 Barcelona Supercomputing Center, Barcelona, Spain maria.marinescu@bsc.es 3 Tel Aviv University, Tel Aviv, Israel msagiv@acm.org 4 VMWare Research, Palo Alto, California, USA Abstract We propose using abduction for inferring implicit rules for Smart City ontologies. We show how we can use Z3 to extract candidate abducers from partial ontologies and leverage them in an iterative process of evolving an ontology by refining relations and restrictions, and populating relations. Our starting point is a Smart City initiative of the city of Barcelona, where a substantial ontology is being developed to support processes such as city planning, social services, or improving the quality of the data concerning (for instance) legal entities, whose incompleteness may sometimes hide fraudulent behavior. In our scenario we are supporting semantic queries over heterogeneous and noisy data. The approach we develop would allow evolving ontologies in an iterative fashion as new relations and restrictions are discovered. 1 Introduction Cities are complex systems of interrelated domains that produce massive amounts of data from many sources and in many different formats. Some of this is open data, which is seldom cured. Data produced by citizens, sensors, or mobile devices, is likewise not always accurate, and may be inconsistent with other data sources. Ontologies are a flexible way to model a city and integrate the many heterogeneous data sources at a semantic level without modifying the data itself. Semantic integration via ontologies balances the ease of customization that domain specialists need with the definition of a stable, well-organized set of concepts that an inference engine can reason about. This approach makes it easier to integrate heterogeneous data sources, but these come at varying quality and granularity levels. Arguably the main entry barrier for a semantic integration approach is the lengthy process of defining an ontology in the first place. A tool that can recommend extensions to an existing model would make ontology definition more palatable. At the same time, in our experience it is often unrealistic to define constraints over the data when this is of poor quality, as it may excessively restrict how much of it could be successfully integrated. Increasing the quality of the data would thus positively affect integration at a global level. C. Benzm¨ uller, C. Lisetti and M. Theobald (eds.), GCAI 2017 (EPiC Series in Computing, vol. 50), pp. 233–238