California traffic has drawn the attention of city planners and researchers for decades. But Forrest Jehlik, research engineer and one-time California resident, noticed that “it never seemed to get much better.”

Despite traffic troubleshooting measures, like Los Angeles’s rail system, most citizens still drive personal vehicles.

This trend, along with California’s dense and complex traffic, makes the city ideal for energy-saving, vehicle-centered traffic management studies.

In this way, computer scientist, Tanwi Mallick, Prasanna Balaprakash, and Jehlik, of Argonne National Laboratory, designed an AI-enabled traffic-forecasting software for large highway networks. In our interview with them, they discuss some of the challenges of developing AI models to understand the complexities and underlying physics of large traffic systems.

DCRNNs in traffic prediction

In the last few years, scientists began applying Diffusion Convolutional Recurrent Neural Networks (DCRNNs) to traffic prediction problems. These models take the network of roads as a graph and traffic signals as time series on the graph. In this way, the models account for the interconnectivity of the roads and the corresponding traffic correlations.

DCRNNs improved the accuracy of traffic predictions but did not overcome the field’s extant scalability issue. Because traffic data is large and complex – taking both spatial and temporal inputs – it can be time-consuming to train large DCRNN models.

To overcome this, Jehlik and Balaprakash segmented their graph-based neural network – which is representative of California’s highway system – into smaller portions.

In real life, however, drivers do not operate in portioned-off segments of road. The researchers account for this with another novel feature: overlapping regions. These can be thought of as closely related roads, or rather, roads where the traffic of one affects the traffic of the other.

They iteratively evaluated the efficacy of training their models via supercomputers on various amounts of data, identifying the minimum for accuracy and the maximum for time-efficiency.

In the end, they were able to train their AI models on California’s massive amount of traffic data, collected from the state’s 11,960 active sensors, in just three hours. 

Their methodology gives a glimpse of the complex and interconnected nature of traffic data.

“We are not just taking an image of the traffic and... trying to predict the traffic for the next step,” says Balaprakash.

An image might capture traffic density, and a series of images might capture speed, but these images would not capture traffic flow or how the drivers affect one another and their progress along the network.

In fact, images do not capture the arrangement of the road system at all, nor its effects on speed and flow.

However, by mapping the roads as the neural network itself, they do capture these effects. And by inputting time-series data into this network, they are able to predict how traffic will diffuse across it, given the day, time, and different, random variables – such as accidents, driver decision-making, and driver alertness.

Implementation and on

The scalability Jehlik and Balaprakash achieved allows them to work with large highway networks and it also compensates for the noisy nature of sensor data.

“If you have noisy data and a really small volume of data, it's very hard to do anything,” explained Balaprakash.

But because they are working with data from so many sensors, their software naturally infers when data is faulty by comparing it to that from nearby sensors. 

Following the training session, their software can be downloaded onto regular computers, where it predicts traffic based on incoming data streams – such as those collected by traffic sensors – for the next sixty minutes in a mere five.

In the background, Jehlik and Balaprakash are also looking at optimizing traffic based on the unique energy profiles of vehicles (from trucks to hybrids), data which their AI software is already capable of processing.

“Roughly, pre-Covid, you're looking at about 20 million barrels of oil a day,” said Jehlik. “70 percent of that is transportation, and then another 70 percent of that is basically light-duty vehicle transportation.”

Although they are excited about what their research portends for the fields of energy and resource optimization, we are excited about the significant contribution they have already made to AI methodology and scalability.

Looking ahead, Jehlik and Balaprakash commented on the possibility of running the AI software on very tiny computers – our cell phones – in a Siri-like app which would offer drivers informed and adaptive traffic guidance. 

If this sounds too good to be true, you might be pleased to know that, as Balaprakash pointed out, “...even bigger models than [ours] – the voice (like Siri) and image recognition models, for example – are already in your cell phone.” So keep your chin up, Californians, lighter traffic ahead.