The ADAIT team conducts research on a range of topics in the intersection of data analytics and transportation science. Our broad interest is in addressing transportation problems via data-driven models that exploit multimodal data sources which are now increasingly available in the era of Big Data. Our long-term vision is that these models will aid transport network planners to better design the network and respond to faults, and aid transport network users to be better informed of their available choices in navigating the network.
Our research has been end-use driven, and has broadly concentrated on the following key topics relevant to network planners and users:
Network planners must know the demand profile at various times in the day: for example, for a road network, they would like to know the destinations that experience the most demand during rush hour, as well as the usage patterns of each individual road segment. This helps make decisions about where to invest effort when attempting to improve the network, e.g. if we want to reduce travel times to CBD, which road segments should we widen?
Traditionally, demand analysis has been done by conducting surveys, which are expensive to conduct and not always reliable. The ADAIT team has conducted research on automatically inferring the demand profile by exploiting the plentiful set of traffic flow data. This has involved using advanced machine learning techniques to learn from sparse data, as well those that incorporate any available exogenous information such as census data, prior demand profiles, and so on.
Unplanned traffic hazards or other anomalous incidents can have serious consequences for network usage. It is thus of interest to automatically detect that an incident occurred, and predict the impact of an incident on the network in terms of the congestion on the surrounding roads.
The ADAIT team has conducted research on both these problems, leveraging recent advances in deep learning as well as time-series anomaly detection. To better understand the propagation of congestion in the network, the team has also studied the problem of identifying the most frequent congestion spillbacks, which can help identify flaws in network design.
Suppose one knows that part of a road is closed for maintenance, or that one becomes aware of an incident via Twitter. What impact will this have on traffic on surrounding roads? Intuitively, by understanding the demand profile of the network, one should be able to simulate the flow under modifications to the network.
The ADAIT team has been developing a large-scale traffic simulation model, with the eventual aim of being able to predict fine-grained traffic patterns for the entire Sydney network. This has involved extending the rich literature on dynamic traffic assignment models. We have also been exploring novel traffic flow prediction models using the theory of random processes, in particular the use of self- and external-exciting processes, which have seen great success in finance applications.
Users of the network often have many choices as to how to navigate it. For example, one can either can a bus, or take a car; if going by car, one typically has several possible routes to choose from. Understanding why users make the choices they do can help improve the network, for example by determining how best to incentivise public transport use.
The ADAIT team has been exploring latent modelling approaches to choice analysis, which attempt to infer each user’s hidden decision-making profile. Similar approaches form the backbone of related choice modelling applications, such as book or movie recommendation.
Often, one can collect information about a transport network from multiple sources. For example, travel times for each road segment may be available from a mobile phone operator, as well as a taxi cab fleet. Intuitively, one expects each source of data to provide an imperfect but complementary window into the true state of the network. How does one fuse these disparate sources of information to a reliable aggregated summary of network performance?
The ADAIT team has explored the use of Bayesian fusion methods to solve precisely such problems, with the intent of further applications to fusion of data-driven and classical transportation models.