Whilst traffic lights do use sensors to try and make slightly more intelligent decisions than perhaps they once did, they are still fairly dumb tools for regulating the flow of traffic.

A recent Chinese study explores whether machine learning can do a better job.

The Elephant & Castle roundabout near where I live is notorious for it’s complexity, with rush hour traffic bustling onto it from several directions.  It’s perhaps understandable therefore that humans struggle to program traffic lights to function effectively.

Automating traffic flow

Of course, automating the process is no mean feat either, both because it requires an accurate model of the traffic flow, and then the challenges inherent in optimizing the flow.

The modeling side of the equation is something modern AI can do relatively easily but optimizing the flow is somewhat trickier as traditional approaches struggle to optimize what they’re modeling.  The researchers believe reinforcement learning provides just such a capability however.

Video games have been enthusiastic adopters of reinforcement learning in recent years, with algorithms used to figure out the most beneficial action at a given moment, and therefore to award the player the highest score.

In traffic management, the top score is awarded when drivers are kept waiting for the shortest period of time and the shortness of queues.  Of course, in complex urban environments, the data sets generated to produce this optimum can quickly become vast, and so deep learning was brought into play to make it more manageable.

Like reinforcement learning, deep learning also takes inspiration from the human brain.  It uses neural networks to hunt out patterns that may lie hidden in huge data sets.  When it’s combined with reinforcement learning, it therefore makes the search for the best solution more efficient.

Proving its mettle

The system was put through its paces on a virtual four way intersection, with four lanes going north-south, and another four going east-west.  The system out performed another algorithm that just used reinforcement learning with both shorter queues and a better balance of traffic in both directions.

Indeed, as the simulation played out over a full day, roughly 1,000 fewer cars came to a complete stop, with the average delay reduced by 14% and vehicles spending 13 less seconds at the intersection.

Suffice to say, that isn’t to say that the system is ready to hit the streets just yet, and with Ford recently announcing that they will release automated vehicles within five years, it’s possibly a race between this and an automated network that figures these things out on a vehicle to vehicle level.

It will certainly be an interesting trend to follow though.