Giving smart cities the green light: how AI can make traffic flow

Research insights: Dr Ivana Dusparic, Ussher assistant professor in Future Cities and the Internet of Things, TCD

We hear a lot about how connected devices can support smarter cities. What are you working on?

“I look at how to co-ordinate connected devices using artificial intelligence, in order to make complex systems work more efficiently. In particular, I use machine learning on linked systems such as traffic lights to help keep transport and pedestrians moving in cities, and on household appliances to use electricity more sustainably.”

How do you apply artificial intelligence to traffic lights?

"I'm working on this in the Enable programme, in the Science Foundation Ireland Connect Centre, where we have made computer simulations of traffic junctions. Within those simulations, we set down high-level rules – we want the system to keep cars progressing well through the junctions, to give public transport priority and to ensure pedestrians are not left waiting for long periods. Then we use a machine learning and reinforcement approach. That means instead of figuring out possible combinations, the system tries things out in the simulation, and the junctions learn what results in the best outcome for a given set of conditions."


Where do you get the information to model the traffic junctions?

"To date, we have simulated around 40 junctions in Dublin city centre and along Western Road in Cork and we allow the machine to learn under all kinds of conditions. Initially we start with just one junction that is fed by others, and then we add more junctions to the simulation to widen it out and the artificial intelligence system manages the simulated traffic in more complex environments."

What kinds of challenges are there to teaching AI about traffic junctions?

"There are several! For one thing, pedestrians in Ireland can be a bit unpredictable, lots of people cross without waiting for the lights. That's something the AI needs to get to grips with."

How would you see an AI system like this being used to control traffic?

“In the shorter term, it could help to advise traffic engineers. Before a system like this could be unleashed to control traffic directly, we would need to ensure there was more explainability in the system.”

Explainability? Can you explain?

“This is a big issue in AI. It means understanding more about the process by which AI makes decisions. That way, if it starts to make decisions that we would see as undesirable or less than optimal, we could trace why that is happening. The learning process can be a bit of a black box, so explainability is important for real-world applications.”

You mentioned co-ordinating electrical appliances too, how does that work?

“We apply a similar machine-learning approach, such that household electrical appliances and systems would learn how to use electricity efficiently. In the future, the expectation is the electricity rates will change on an hourly basis, and consumers could identify what they want to happen by when – say you want to have clothes dry and laptops charged by the morning – then the AI system figures out the optimal time to use the energy in the cheapest configuration possible.”

What would you like people to know about machine learning?

“In the case of my research on using machine learning to teach systems, it’s that the software is just one part of it; we also use lots of insights from other fields too, such as psychology and education and ethics. We hope those insights will help to ensure we build a system that works efficiently and that we know how it is making decisions.”

In conversation with Claire O’Connell