Introduction

A city that moves more efficiently can do more for the people in it. Addressing this challenge, the Turing has joined forces with the Toyota Mobility Foundation to apply machine learning and data science to transform how traffic managers see their city – and create the mobility tools of the future.

Metropolitan areas are complex entities with a diverse set of unique mobility challenges that continue to grow. Increasing urban populations challenge the capacity of transport infrastructure, while modern phenomena such as ridesharing platforms like Uber and same-day delivery fleets affect traffic patterns in novel, complex ways.

In the short to medium term, getting smarter about traffic control offers a clear win, and pioneering data science, led by the Turing and its partners, aims to provide tools to do just that. “Optimising flow within mobility systems with AI” is an initial 18-month project between the Turing and the Toyota Mobility Foundation (TMF) that concludes in December 2019.

The project is led by Turing Fellows Damon Wischik at the University of Cambridge and Neil Walton at the University of Manchester as part of the Institute’s Urban Analytics programme. One of its goals is to transform conventional traffic-management systems into optimised, AI-powered systems that will enable cities to manage in real-time across many types of mobility. To do that means building relationships with a variety of key transport organisations, including Transport for London, Transport for Greater Manchester (TfGM) and the City of Los Angeles, and also innovative businesses such as Vivacity Labs, which is developing AI-based traffic solutions.  

One of the collaboration’s key outputs is a prototype Mobility Data Toolkit for assessing and presenting a wide range of traffic data, from traffic flow and congestion to levels of harmful pollutants, at virtually any scale and slice of time. Produced by the Turing’s Research Engineering Group, this visualisation toolkit, in conjunction with the latest in traffic-simulation algorithms and machine learning, has the potential to transform how urban planners design transport systems and how traffic controllers implement interventions to optimise traffic flow and reduce pollution.   

Higher-level impact

The Turing’s Programme Manager for Artificial Intelligence, Aida Mehonic, says Toyota Mobility Foundation sought out the Turing because of its unique position at the interface of academia, industry and government. “TMF’s metric is real-world value; they're interested in functional tools, collaborating with transport authorities on their adoption, and overall in translating research into impact at a higher level of technological readiness that goes beyond what universities normally do.” William Chernicoff, Senior Manager, Global Research & Innovation at TMF, agrees: “We have valued the Turing’s openness and desire to work with our mobility partners to anchor this research in real-world contexts at the local level.”

“We have valued the Turing’s openness and desire to work with our mobility partners to anchor this research in real-world contexts at the local level.”

William Chernicoff, Senior Manager, Global Research & Innovation at Toyota Mobility Foundation

The project started with three goals in mind. One was to devise new ways for fleet operators and city planners to work together, for example, by pricing appropriately and sharing data about congestion. The other two goals, the subjects of this blog, are closely linked: to create a prototype of the mobility toolkit, and to begin the development of a top-tier reinforcement-learning system for traffic-light optimisation, and capacity management across modes.

“The mobility toolkit we are making will enable planners – who are typically not trained in data science – to easily access, visualise and manage their transport data,” Walton explains. Such a system could help city planners prepare for the future and manage current conditions, optimising energy consumption and reducing pollution in sensitive areas, such as around schools at drop-off and pick-up times.

Traffic data is a mix of spatial and temporal information: a vehicle is travelling at velocity x (which has implications for its emissions), in location y, at time z. Simulating the traffic systems of cities such as London is a complex, big data scenario. Senior Research Software Engineer May Yong and Research Data Scientist Louise Bowler, both members of the Turing’s Research Engineering Group, have developed a prototype of the toolkit to wrangle and present this data. “Damon Wischik wanted us to develop ‘Excel for spatial data’,” says Yong. In other words, the sort of powerful, intuitive interface that allows traffic controllers with little or no coding experience to quickly get to grips with their data (Fig 1).

Fig 1. The upper map shows traffic building-up at a junction with ‘untrained’ traffic lights. The lower map shows the same junction, but with traffic lights trained using a machine-learning algorithm, resulting in a smoother flow of traffic. The top-right chart shows predicted emissions of CO2, while the bottom chart shows predicted emissions of harmful nitrous oxides. Emissions of both pollutants are expected to fall as a result of training the traffic lights – shown by the blue lines
Fig 1. The upper map shows traffic building-up at a junction with ‘untrained’ traffic lights. The lower map shows the same junction, but with traffic lights trained using a machine-learning algorithm, resulting in a smoother flow of traffic. The top-right chart shows predicted emissions of CO2, while the bottom chart shows predicted emissions of harmful nitrous oxides. Emissions of both pollutants are expected to fall as a result of training the traffic lights – shown by the blue lines

 

Avoiding logjams

The mobility toolkit is currently a prototype, but how might the final product be used? Let’s say you are a traffic controller, and there’s a risk that a critical section of your road network will need to be closed down during rush hour to deal with an emergency. How will the traffic respond, and how might you alleviate the inevitable problems? With the toolkit’s intuitive interface, you could draw out the region of interest on the interactive map, select the time period you are most concerned about, then run a traffic simulation before and after those key roads are closed off. This will give you a picture of what to expect.

With the data from your emergency-scenario at hand, you could also return to the simulator and use machine-learning algorithms to adjust the traffic-light patterns around the entire area to better manage the disrupted traffic flow, reducing the risk of gridlock. The toolkit will make it straightforward to compare simulations of before and after this virtual intervention. “It’s something that a transport planner, without any experience in data science, could then use and answer their own questions,” says Walton.

Ultimately, the toolkit will offer the ability to quickly simulate the effects on traffic of an additional bus lane here, a road closed for maintenance there, stricter emissions limits here, AI-powered traffic-light configurations there. “The toolkit is a way to predict clearly the effect of interventions,” says Bowler. “It's very convincing when you see things side by side.”

 

AI-augmented traffic lights

The management of traffic data is one thing, but what about the management of the traffic itself? The Turing’s collaboration with TMF is taking a pioneering approach to urban traffic control (UTC). The old approach is based on the use of fixed signal timings. Even now, in the US, only about one-tenth of signal-bearing junctions make use of traffic sensing or adaptive UTC systems. In Europe, there is much greater uptake of modern adaptive systems, but we believe they remain behind the curve of what is technologically possible, as we will demonstrate below.

Recent years have seen enormous investment in AI technologies for self-driving cars, which typically benefits the owner, but there has been much less investment in AI for adaptive traffic-signal control, which would benefit all road users by optimising traffic flow, reducing delays at junctions, cutting congestion, reducing emissions, and saving fuel. Here, AI—specifically, deep reinforcement learning (DRL)—is a powerful tool.

Alvaro Cabrejas Egea of the University of Warwick and Raymond Zhang of ENS Paris Saclay are developing the traffic-light training algorithms, using the latest DRL techniques. Although still in its early stages, the work is already revealing remarkable potential when its results are compared with the latest commercial traffic-control systems (see Fig 2).

Fig 2. This graph shows the accumulated time that vehicles are delayed at a simulated four-way intersection, with three lanes coming in and out of each branch, over the course of an hour. A standard commercial adaptive traffic control system, called MOVA, clocks up more than 70 hours of driver waiting time (green line). In the same scenario, several of our algorithms (the orange and blue lines), both based on “Duelling Deep Q Networks,” cut driver waiting time significantly, the best one by about 50%. See the simulated junction in action.
Fig 2. This graph shows the accumulated time that vehicles are delayed at a simulated four-way intersection, with three lanes coming in and out of each branch, over the course of an hour. A standard commercial adaptive traffic control system, called MOVA, clocks up more than 70 hours of driver waiting time (green line). In the same scenario, several of our algorithms (the orange and blue lines), both based on “Duelling Deep Q Networks,” cut driver waiting time significantly, the best one by about 50%. See the simulated junction in action below.

“The reason the sort of system we’re using can work better than modern commercial systems in some situations is that commercial packages typically consist of a ‘hard-wired’ model of how given types of junctions work, and aims to optimise traffic flow according to that model,” says Walton. “In contrast, our system uses a neural network to learn the characteristics of each junction, rather than deciding in advance how each specific junction works.”

For example, the relative angles of the branches of the junction, the placement of slip roads, driver speed-up/slow-down profiles can all affect how to optimise the lights. In other words, the system the Turing/TMF is developing is trying to deal accurately with the subtleties of reality, rather than using a rigid approximation of reality common in commercial systems.

"Our system uses a neural network to learn the characteristics of each junction, rather than deciding in advance how each specific junction works.”

Neil Walton, Turing Fellow, University of Manchester

The Turing is developing relationships with the Transport Research Laboratory, which produces the MOVA (Microprocessor Optimised Vehicle Actuation) signal-control system, and also AI start-ups such as Vivacity Labs. Both organisations are looking to develop new advanced signal control systems. Indeed, Vivacity Labs will be taking control of some traffic signals in the city of Manchester, starting in the next few months, to trial smart solutions to the traffic-control challenge.

“Transport for Greater Manchester is committed to supporting innovation across the city region,” says Hannah Tune, an Intelligent Transport Systems Engineer at TfGM. “The introduction of AI in adaptive traffic control has the potential to increase the capacity on the highway network and assist how we manage and integrate our highway network with systems such as city management platforms and traffic simulation solutions. Working with the Turing allows us to explore new technologies that offer the opportunity to gather operationally significant data that will enable local authorities and TfGM to improve the journeys of the travelling public.”

These multifaceted partnerships are exactly what will be needed to bring the benefits of AI and data science to the travelling public. “We're aiming to get everyone in the same room to talk about what a future traffic control system should look like and, from a scientific standpoint, how they should work,” says Walton. “We're hoping that the Turing can act as an open forum where these traffic control organisations can come in, with the shared aim of furthering the state-of-the-art of UTC signalling systems.”

 

Versatile applications

“Over the course of our collaboration, the Turing has become a trusted mobility partner to develop AI for the social good,” says Chernicoff. “We have worked to combine the Turing’s expertise and ability to deliver solutions with TMF’s focus and experience deploying tailored mobility solutions that benefit all sectors of society.”

The Turing and TMF are now scoping out the possibility of a much larger mobility project, which would bring these technologies to fruition and place the technologies with TMF’s existing mobility partners and projects. Within that, the AI-based traffic light control system would be built up, enabling the systematic training of large sets of road junctions, and incorporate human interactions and interventions. A future iteration of the mobility toolkit, meanwhile could integrate real-world traffic and emissions data, collected by agencies such as Transport for London, and be validated with data collected from real traffic scenarios.

The toolkit’s ability to slice and dice spatial and time-series data means it may prove useful well beyond the manipulation and presentation of traffic data. “We've built the framework in such a way that programmers can build software plugins that allow it to use data from different domains,” says Yong. This versatility is a key aspect of the Turing’s Research Engineering group, which creates solutions to the challenges at hand by developing overarching systems that can also boost advances in other data science domains.

What is more, the mobility toolkit is also an open-source endeavour, meaning researchers all over the world can help to build on it or adapt it to their own sphere of interest. As a result, the impact of this aspect of the Turing’s collaboration with the TMF will be multiplied.

This is the kind of big-impact data science that keeps on giving. Ultimately, the hope is to see the fruits of these mobility innovations benefitting cities all around the world, making urban mobility smoother, more efficient and healthier for many millions of people.