Advanced Data Analytics in Transport (ADAIT)
Data61 (CSIRO) and Transport for NSW
Data61 and Transport for NSW jointly developed a prototype AI engine for congestion management. The AI engine trains machine-learning models that complement traffic simulation for network prediction. The development is the first of its kind in Australia to make use of the power of deeplearning and big-data analytics to provide at realtime the foresight of transport network situations and the insight on the impact of major disruptions to multi-modal, end-to-end journeys in NSW.
AI is transforming the way of future transport. Under the NSW Premier’s Innovation Initiative, now an innovation pilot of the Intelligent Congestion Management Program, the AI engine complements the existing solutions by integrating machine learning with traffic simulation for short-term network prediction. The AI engine is generic and agnostic to transport networks. It ingests transport datasets for the training of special purpose predictive models. The models are used for predicting network demand and speed for the next 30-minutes and initiating action in 5-minutes. The AI engine forms a closed-loop with simulation models by continuously learning the causal relationship and calibrating the simulation model for improved modeling fidelity.
The engine is currently being refined and is scheduled for deployment to the Advanced Data Analytics in Transport (ADAIT) cloud platform by mid-2019. The first user is the NSW Traffic Management Centre (TMC). The use cases at TMC can be highlighted as: a) continuous improvement in predictive analytics by using the AI engine to build and train predictive models for changes in networks; b) real-time rolling prediction of the network demand and speed; c) autonomous network anomaly detection and verification; d) modeling of multi-modal journeys and the impact of major network disruptions.
Further expansion of the AI engine may include the feedback control to traffic signal systems, VMS and smart motorway systems.
Advanced Data Analytics in Transport (ADAIT)
This research provides a new approach to one of the central problems in urban transportation: reliable estimation of origin-destination traffic flows (that is, flows that specify where traffic comes from, and where traffic goes to). The research operates on two key insights. First, in most large-scale urban networks, most traffic is concentrated around a few dominant origins and destinations. This allows one to automatically select a parsimonious, interpretable set of origin-destination flows that explain the observed traffic on a network. Second, it is possible to automatically determine a finegrained zoning of the traffic network purely from observed traffic data. This allows one to zone the transport network in a way that facilitates tractable analyses, without undue sacrifice in accuracy.
According to Data61, the above insights are novel in the analysis of transportation networks. Further, the research conducted is the first to present a unified analysis of both the traffic zoning and origin-destination estimation problems. A range of experiments and analyses on a real-world traffic network have validated the efficacy of the proposed approach.
The likely first users of the research are transportation scientists and engineers interested in modelling and simulation of traffic flows in large-scale urban networks, as it allows them to reliably analyse real world traffic patterns with minimal manual intervention. The research has already proven useful in the modelling of a large motorway in Sydney with the proposed algorithm predicting traffic flows on the motorway to a very high (>99%) degree of accuracy. Efforts are currently underway within Data61 to integrate the research into the existing workflow of largescale simulation.
Machine Learning Research Group
NICTA (National ICT Australia)
Dr Fang Chen
The main focus of the Advanced Data Analytics in Transportation (ADAIT) Project is to apply modern machine learning techniques to transport science, specifically transport modelling. By utilising everincreasing volume of transport data in an advanced computing framework it will be possible to produce transport models that are more accurate and accountable than ever before. ADAIT team members: Fang Chen, Chen Cai, Hoang Nguyen, Weihong Wang, Aditya Menon, Tao Wen (Student), Tuo Mao (student).
The scientific underpinning is Machine Learning: a branch of computer science and mathematics that uses computer algorithms to discover recurrent patterns in data. By 'training' the algorithms on historical data these algorithms are able to rapidly identify 'learnt' patterns in previously unseen data. By combing (fusing) data from multiple sources in a machine learning framework it is possible to greatly improve the data inputs required for fundamental transport models, increasing their reliability and robustness enormously, and making them even more useful for designing and planning transport systems.
When used in a high-performance distributed computing environment the same techniques may be used to make real-time predictions thereby bridging the gap between planning and operations, a long sought-after goal in computational transportation science.
Transport authorities in Australia will be the first users of the research. The ideas embodied in this application have already been applied to assist NSW Roads and Maritime Services identify and address mobility issues that will arise in the Sydney CBD during the construction of the South East Light Rail Project. The machine learning techniques established event-to-impact causality from traffic data, and used it to predict the impact on traffic during the proposed construction.