Hiding in Plain Sight – How can AI help us identify non-recurring congestion?
Post by Shai Maron, Waycare VP of R&D
I’m a planner. I like to know when things are going to happen, how long I need to prepare, and – most importantly – exactly how long it’s going to take to get there. Needless to say, as a Tel Aviv resident, my careful planning is often thwarted by unpredictable traffic. A 15-minute drive becomes a two-hour ordeal at the drop of a hat. This unpredictability seems random, however transportation experts have named the cause: non-recurring congestion.
What is non-recurring congestion? In the simplest terms, it’s a departure from the norm. Alex Smolyak, Waycare’s Data Science Team Lead, defines recurring traffic as ‘the typical traffic you’d expect day in, day out.’ On the flipside, non-recurring congestion is the type of traffic that comes out of nowhere and has you saying, “Hey, where did this come from?” According to a study conducted by the University of Cincinnati, non-recurring congestion is most often caused by construction, inclement weather, accidents, and special events. In addition, disabled vehicles, law enforcement activities, and heavy merging traffic, are often major causes of non-recurring congestion.
But how can we tell the difference between normal and abnormal traffic build-ups? That’s where AI comes in. Algorithms can assist us in identifying what is actually an outlier, versus what is expected traffic like during the holidays or rush hour. We spoke more with Alex Smolyak to find out how it works.
An interview with Alex Smolyak, Data Science Team Lead
What is the point of identifying the difference between recurring and non-recurring congestion?
“Knowing the difference between the two is important. When we identify a congestion event as ‘irregular’ or ‘non-recurring,’ that gives transportation agencies the information they need to investigate it further. Was there a crash that caused it? Was there construction taking place that we weren’t aware of? Is there an object in the road that needs to be removed?
It’s also important because commuters can plan around the more ‘regular’ traffic congestion because it’s predictable and it’s expected. To put it simply – if you have an appointment at 9am, you know you have to leave the house at 8am because there will be extra time spent in morning traffic. However, irregular congestion is unexpected and can cause you to run late. Therefore, irregular congestion is often the most detrimental to commuters.“
How is Waycare using AI to define the difference between recurring and non-recurring congestion?
“Where you draw the line between the two is tricky. Being stuck in traffic in the middle of the night on an otherwise empty road – that’s an easy call. But what if congestion starts along your daily commute half an hour earlier, or half a mile upstream of where you typically meet it?
The way we answer this question is by applying several metrics along different dimensions (time, space, speed, quantity and a few others) to estimate, based on historical traffic patterns, how unlikely it is to see the state of events we’re seeing at any given time slice. We’re typically working with 5-minute intervals. An event (i.e. set of measurements along each of the above dimensions) scoring high enough on a weighted combination of the metrics is marked by us as irregular.
Then there are a few finer points about how exactly to properly construct history to balance recent events vs older ones, how to properly weigh everything and adjust weights for various locations and conditions, how to process abrupt changes (construction etc). It’s important to know that these are events that are taken into consideration.“
Can you expand on what dimensions you are looking at to determine these metrics? For example, what data or information are you referring to?
“We try to utilize every factor available. For some data sources it’s fairly straightforward – is speed abnormally low? Then other factors are added. Can we say something about the quantity or density of vehicles on the road, either on their own or in relation to the measured speeds. Assuming we can identify congestion of any kind, can we measure its parameters, such as location or duration, in comparison to what we’d expect? Further questions could be asked about short-term temporal behavior, that is, given not only what we’ve recently recorded – half an hour, an hour ago – do the current measurements make sense? Once we’re able to transform all these questions into numbers, we can start assigning observations to typical or atypical traffic behavior, and if needed, to measure the magnitude of the irregularities.“
As Alex mentioned, the implementation of anomaly detection on the municipal level can give transportation agencies a better understanding of what’s actually happening on the roads. Joseph Williams is the TMC Operations Manager at the Central Texas Regional Mobility Authority (CTRMA). According to Joseph, “Being able to see irregular congestion is very beneficial because it prompts me to view local cameras for possible incidents in the area. It’s an additional tool that will help us in the Traffic Management Center when it comes to monitoring traffic.”
AI is a crucial resource for traffic management agencies to process extensive information and be able to pinpoint traffic irregularities like non-recurring congestion. Giving agencies these tools will not only facilitate quicker remediation, but will enable better communication to the public so that drivers can plan their trips accordingly. As I await the day that my carefully planned commute goes as scheduled, I hope that agencies take advantage of the invaluable tool that is AI.