Humans Cause Traffic Jams, AI Can Fix Them

- [host, Joe Hanson] There are a bunch of causes of traffic jams that make sense like weather, special events, road work.

Then there's the kind of traffic that occurs out of nowhere.

It's surprising.

Some might even call it spooky.

Well they probably wouldn't call it that, but scientists actually do call that type of traffic a phantom traffic jam.

And while the experience is more annoying than spooky the larger impacts of traffic are actually terrifying.

- [Alex] Traffic induces a lot of problems, emissions, pollution.

There's a clear correlation between the number of fatalities and accidents and congestion.

Traffic has become one of the major problems plaguing cities all around the world these days.

- [Joe] So what if I told you our cars themselves might help us solve traffic?

[ambient music] Meet Dr. Alex Bayen, he's kind of a traffic whisperer.

Dr. Bayen is the Director of the Institute of Transportation Studies at UC Berkeley and the Director of Aerospace Programs.

So it doesn't take a rocket scientist to solve our traffic woes, but, like, it helps.

[ambient music] - [Alex] Traffic congestion in cities started to become a real problem after World War II with the acceleration of the automobile.

- [Joe] In the nearly 100 years since the transportation engineering community started trying to understand it, no one has yet been able to solve traffic with science.

Generally, we understand that traffic occurs when the cars on the road require more space than what the roadway can accommodate.

There's a critical density after which small behaviors from drivers can start to seriously jam up the road.

And when that critical density is surpassed, the flow of the roadway becomes unstable, resulting in waves of stop-and-go traffic.

- [Alex] In the early ages of traffic engineering and traffic theory, scientists modeled traffic flow of vehicles as a fluid in pipes.

If you observe the propagation of waves in a river, what you will see is shock waves propagating just like on the freeways.

- [Joe] Traffic scientists have a grasp of the catalyzing events that set off a phantom traffic jam and theoretically what it would take to prevent them.

In order for traffic to flow smoothly, drivers have to maintain a consistent speed and keep an even distance from the cars around them.

Seems simple, right?

Wrong.

Turns out we are awful at this.

- [Alex] Humans are very eager to catch up with the car in front of them, but this is actually the cause of a lot of inefficiency in traffic.

- [Joe] So traffic can actually be caused by our own selfishness, the desire to get your car to its destination at the expense of all others.

Then when we realize the roadway is full, and we have to brake suddenly to stop from hitting the car in front of us, we can send ripples of traffic for miles behind us.

- [Alex] There was an experiment in 2008 by a group of Japanese researchers.

People were asked to drive in a circle at a given pace and maintain their distances with the vehicle in front of them.

It's a simple task, you would think.

- [Joe] But phantom jams started to occur in under a minute.

Now check out this experiment that researchers at Vanderbilt University and the University of Arizona did about 10 years later.

This time they inserted just one automated vehicle in the group.

See that arrow turn red.

That is where the driver let the car drive itself.

You can see as soon as a human driver is taken out of the equation, fuel consumption is reduced by 42% and that stop and go traffic is smoothed out.

Dr. Bayen believes that if as few as 1 out of 20 vehicles on a roadway were connected and automated they could guide human operated vehicles to create a steady flow of traffic instead of phantom jams.

[ambient music] Bayen thinks the automated vehicles can induce the whole population to behave collectively better.

Maybe this all sounds a bit alarming, machines acting to improve human behavior?

Dr. Bayen doesn't think it's that far of a leap from a feature that many drivers use today.

- [Alex] If you have a car that has cruise control, at some point, when you feel comfortable, you might engage it because you don't wanna be adjusting the speed back and forth.

The notion of autonomy on demand is similar.

Once these features start to also encompass lane changes, or longitudinal control, or keeping up with the person in front, the experience would be nearly identical.

At some point, you would engage these new features and then the car would do it for you, just like it's doing it for you now when you do cruise control and you probably don't even think about it.

- [Joe] Bayen's team has created a variety of digital models to simulate traffic scenarios that capitalize on various levels of autonomous participation.

The tool is aptly named Flow, and it's the first application of deep reinforcement learning with traffic microsimulation tools.

The team has tested their tool on loops, figure eights, and ramps.

In a figure eight simulation, just 1 automated car in 14, doubled the speed of all the cars in the course.

And they were able to get twice as many cars through a ramp with just 1 in 10 automated vehicles.

In a real world scenario, the connected cars would be able to anticipate slowdowns ahead by communicating with cars in front of them, and maybe the roadway itself.

They would then match their speed to the average speed of the entire roadway, which would prevent sudden slowdowns from forming in the first place.

In instances where a phantom jam has already occurred, the connected vehicle could dampen the shockwave by maintaining a steady flow and even distance from the car in front of them.

So what is this technology that allows a car to react in real time to traffic scenarios?

[ambient music] Bayen and his team are diving into the world of deep reinforcement learning - [Alex] Reinforcement learning algorithms have the ability to learn by optimizing a reward function.

It's an iterative process.

Every attempt the algorithm does to do things better, it gets a reward.

And if the reward keeps improving, then the algorithms will keep searching in the proper directions.

Now how can that be used for traffic?

If you wanna try to smooth traffic the reward would try to penalize the waves.

- [Joe] And there are decisions that the AI is making in these simulations that are a surprise to Bayen himself.

[ambient music] - [Alex] We already see a lot of cases where artificial intelligence has exceeded the ability of the humans, but we can't really explain why.

- [Joe] Learning from algorithmic simulations is one thing, but these systems are also learning from real world traffic.

- [Alex] One of the threads that is particularly exciting these days in reinforcement learning is pixel learning.

Pixel learning is the ability to learn from pictures.

In the past, a lot of algorithms have learned from modeling the car, and based on that, learning what to do with it.

But recently, with pixel learning, algorithms have learned to learn from the rendering of the car.

There's a lot of different images from which algorithms can learn.

And this has deep implications because that means that in order to learn how to improve traffic, maybe you don't need to measure as much as before.

Maybe all you need to do is to have enough video footage so that you can learn from it.

- [Joe] And cameras are coming to our roadways and our cars in increasing numbers.

As of 2018, every car that is sold in the US is required to have at least one camera, the backup camera.

And transportation departments across the country are busy installing cameras on roadways.

So how close are we to seeing the positive impacts of this research on our roadways today?

[ambient music] - [Alex] What's missing in order to enable this multiple vehicle coordination, in the spirit of achieving the greater good for traffic, is the ability to create a communication paradigm by which the vehicles start to collaborate.

- [Joe] Currently there is no one group that is bringing the private sector, scientists, and the public sector to the table to talk about the future of our roadways and driverless cars.

So Bayen and his team are working to build partnerships with everyone involved to do just that.

They've been supported by the Department of Energy, the National Science Foundation, and some private car manufacturers.

- [Alex] Travel is such an important part of human life.

I really believe we can make travel better.

Where better is not just a better experience but really something that is better for our planet.

- [Joe] To gain back our wasted time in traffic, increase safety for everyone on the road, and mitigate some of the environmental impacts of our transportation systems, we're going to have to learn to act collaboratively.

That spirit of cooperation and altruism, along with our increasingly intelligent cars, might just help us finally solve traffic.

[ambient music] ♪