Machine learning for modelling urban dynamics

We live in the age of cities. More than half of the world’s population live in cities and this urbanisation trend is only forecasted to continue. To understand cities now and in the foreseeable future, we need to take seriously the idea that it is not enough to study cities as sets of locations as we have done in the past. Instead, we need to switch our traditional focus from locations to interactions and in doing so, invoke novel approaches to modelling cities. Cities are becoming “smart” recording their daily interactions via various sensors and yielding up their secrets in large databases. We are faced with an unprecedented opportunity to reason about them directly from such secondary data. In this thesis, we propose model-based machine learning as a flexible framework for reasoning about cities at micro and macro scales. We use model-based machine learning to encode our knowledge about cities and then to automatically learn about them from urban tracking data. Driven by questions about urban dynamics, we develop novel Bayesian inference algorithms that improve our ability to learn from highly complex, temporal data feeds, such as tracks of vehicles in cities. Overall, the thesis proposes a novel machine learning toolkit, which, when applied to urban data, can challenge how we can think about cities now and about how to make them ”smarter”