Jun 06 2022

AI Can Help Traffic Jams

Traffic jams cost the US economy an estimated $179 billion per year. That’s a pretty good incentive to invest in strategies to reduce the problem. Slow traffic is also a huge pain, causing significant stress and wasted time. (Remember the linear bias I wrote about recently, improving traffic speeds at the low end by avoiding jams has a huge effect on average driving speeds.)  The causes of jams are various, including accidents and road construction, but some causes are “soft” in that they just happen due to driving patterns. Phantom traffic jams is one phenomenon, caused mainly by tailgating. When one driver slows even a little, the car behind them has to slow a bit more, and this continues until traffic stops for no apparent reason. Traffic light patterns is another cause. Hard causes need hard fixes, like expanding lanes and having adequate accident response infrastructure. But the soft causes can be mitigated by smarter driving and traffic control.

This is where artificial intelligence (AI) enters the picture. Both of these significant causes of slowed traffic, better driving behavior and smarter traffic light control, can be improved with the application of AI. With respect to phantom jams, the solution seems to be to apply what researchers are calling bilateral control – this means that drivers should space themselves out so that they are equal distance between the car in front and behind them, rather than riding the tail of the car in front of them. Simulations show that adopting this driving pattern can reduce phantom traffic jams by 50%.

How can we implement this strategy? Driver education may help a little, but it is difficult to get a large number of people to make a behavior change, and the benefits really only occur when high percentages of drivers follow this pattern. Some level of driver assist is therefore needed. Short of full self-driving capability, it’s possible to enact bilateral control with a driver assist tweak to cruise control. Even just providing feedback to the driver to back off to an optimal distance from the car in front of them could work. This is a relatively simple fix that can be rapidly implemented.

Traffic control is a trickier problem requiring more advanced AI, but researchers are making progress there as well. Using deep reinforcement learning (DRL), the researchers created a traffic simulation in order to train the AI to improve traffic throughput. DRL essentially provides a “reward” for positive outcomes and a negative reward for unwanted outcomes. The AI was therefore reinforced toward strategies that maximize cars passing through an intersection. The researchers did not determine for the AI how to accomplish this, only gave it the goal and reinforcement. The AI is programmed to try new strategies and then avoid those that don’t work and build on those that do. The simulation used video feedback, since this is available at many traffic intersections. (The other option is embedded sensors in the road, which some locations also have.)

The main goal of the research was to see if the simulation training would translate to the real world, and they found that it did. That’s good, because it’s a lot easier to train the AI on a simulation than in a real-world scenario. Once highly trained it can then be adapted and further trained in real-world implementation. This solution is also easy to implement, because in many locations we already have computer-controlled traffic light control. In many locations this would just be a software upgrade, or perhaps need a hardware upgrade too if the existing systems are too old. That sounds like something that should probably happen anyway.

Improving these two soft causes of slow traffic, phantom jams and suboptimal traffic light control, can take a huge bite out of traffic jams. This is an infrastructure issue, and seems like a worth investment, given the cost and inconvenience of traffic jams. These kinds of solutions may also be able to hold us over until a more aggressive solution can be implemented, namely when most cars on the road are self-driving and traffic is controlled by a network of these cars working together to manage flow. That is the ultimate solution to traffic jams, removing the human variable out of the equation and having AI optimize traffic flow.

This story also reinforces the notion that narrow AI is increasingly being embedded in our society, running in the background to make everything run more smoothly. Narrow AI (programmed to handle a specific task but lacking general intelligence or anything resembling self-awareness) is extremely powerful – it’s proving to be so powerful that one could argue (as I have) that the incentive to develop general AI is significantly reduced. We may find, as we increasingly are, that we don’t need it to accomplish everything we want AI to do. Even if we want a humanoid robot to act completely human, we can just combine a bunch of narrow AI algorithms together to simulate human behavior. Then again, some neuroscientists and philosophers would argue that this is exactly what humans are, a bunch of narrow AI algorithms massively networked together with “consciousness” emerging from the aggregate.

Either way, narrow AI will increasingly be running our society. Traffic is among the low-hanging fruit that will increasingly be driven by AI.

No responses yet