Machine Learning and Work Zones

Last July we talked about the phenomenon of Poke’mon Go and augmented reality and considered ways in which that new technology might be applied to work zones. Today let’s consider another new technology. This one is not really new but is enjoying renewed interest thanks to Big Data and an abundance of inexpensive computing power. Using machine learning we can now process our data in ways we could not before.

In its simplest form, machine learning is an algorithm or model that goes through data looking for patterns. Then using those patterns it can make predictions for another, similar set of data. Machine learning works well with different types of data. So, for example, it could crunch through traffic speed data and compare it to social media sites focused on local traffic problems. In this way it can find trends or commonalities that don’t come through in conventional data analysis.

Machine learning is a faster and more accurate way of adapting to human factors issues related to work zones. We set up a system before the project starts and configure it for conditions at that time. But conditions change in work zones. So what worked early may not work as well later. Now don’t misunderstand. The system still works. It still warns of queuing ahead. But if the timing of the messages varied depending on changing conditions throughout the day, coming on sooner, or turning off faster, it could improve driver compliance. And the only way to learn that is through machine learning.

Now there is one problem with machine learning, and it is a big problem. It acts as a black box taking in data and spitting out recommendations. But the process is not something you can review after the fact to understand what influenced the outcome. So agencies could not tell their governor why they have made changes. But if you are more interested in being able to adapt to changing conditions quickly and accurately than in being able to explain it, machine learning is the tool for you.

Let’s consider a few possible applications:

  • Algorithms used to trigger messages could learn to de-emphasize input at certain sensors at certain times of the day when that sensors data is not in line with conditions elsewhere.
  • Models could make better recommendations about the best times and days of the week for lane closures while still meeting goals for minimal level of service.
  • In time, machine learning could establish baselines for crashes, fatalities, etc. resulting from a project both with and without work zone ITS systems in place. The system would be justifying itself! Of course, the predictions would have to prove true over time.
  • Predictive analytics reviewing probe and spot speed data may see locations of higher crash frequency or recommend a different geometry or staging in that location.

Is this complicated? Absolutely. But tools are already available to make it easy to get started, to choose the best algorithm, and to scale up and down in size as the situation demands. And Iowa DOT has already doing it at their Center for Transportation Research and Education.

Until now we could only design a system for a project to get the best overall performance. Now we can save even more lives through machine learning. Once our systems begin to learn from their own data, they will perform even better than they are today.

Medical Wearable’s and Autonomous Vehicles (and Work Zone ITS)

screenshot-www-linkedin-com-2017-01-10-10-12-32Marty Weed, who recently retired from WsDOT, is a good friend and still very much involved in work zone ITS. He ran across a talk on LinkedIn comparing medical wearable technology with autonomous vehicle technology. In the video Randy Hamlin, a VP and engineer at Phillips, said that both technologies are further ahead than policy or behavior. The technology won’t slow down adoption. Instead it will be policies or behaviors. I believe we can add Work Zone ITS to this comparison as well.

You can view his presentation HERE.

Mr. Hamlin stated that both industries have the opportunity to meet a very large and growing need. For medical wearable’s it is the opportunity to reduce chronic disease. For the autonomous vehicle industry it is the opportunity to reduce the chronic roadway fatalities of 30,000+ per year. But before either industry can make an impact, each must first achieve three key factors:

  1. Access
  2. Integration
  3. Adoption

Under access he said that for a technology to achieve widespread adoption it must be accessible. One easy way to make that happen today is to take advantage of the personal cell phone market. These devices have become very powerful, very inexpensive, and everyone owns one. So instead of expensive, stand alone systems, create hardware that interfaces with your phone using an app. This is already taking place in all three industries.

His second factor, integration, revolves around data. Medical wearable’s generate a great deal of data. So do autonomous vehicles. And so do work zone ITS systems. The data quickly overwhelms the practitioners. They quickly begin to ignore it and go back to doing what they did before. Because they just don’t know where to start in making use of that data. As he pointed out, data must be relevant.

Autonomous vehicle manufacturers are addressing this by keeping most of the data to themselves. They use it for product improvement and verification. They only release data to other users that is relevant. Data packages are customized for each group of users. This requires a good understanding of each user’s needs and habits, but results in faster and broader adoption.

Work zone ITS can learn from this. We must understand what our customers need from data. Traffic operations will want one package. Construction may want something else. And systems operations folks may want yet another set.

This will vary from state to state and even from one district office to another. A more urban district may watch volumes while a more rural district may be more interested in speeds or queue lengths.

As Mr. Hamlin said, for our systems to impact roadway work zone fatalities we must first achieve access and then integration. Only by packaging our data so that it is relevant can we hope to achieve the third factor, adoption. Once users become accustomed to receiving timely and useful data, they will come to depend on it. And once they depend on it, our systems will see far more common use.

Improving the Effectiveness of Smart Work Zone Technologies, Part 2

illinoisstudyIn our last post we discussed a brilliant new paper published in November by the Illinois Center for Transportation Studies. Today let’s look at their conclusions regarding work zone travel time systems. The writers point out that, “Two critical components for the success of a smart work zone deployment are the quality of the traffic data collected by sensor networks and the algorithms used for data processing.” We examined sensor types last time. Today we look at algorithms.

They conclude that, “The travel time estimation is consistently poor for all algorithms and sensor networks investigated in this study. The main reason is that the instantaneous travel time calculation is a poor estimator of the true travel time in a dynamic traffic environment. In addition, the use of Bluetooth sensors can only provide the travel time of vehicles that just exited the work zone. Consequently, the travel time estimation even using Bluetooth sensors is not likely to improve the accuracy of the travel time estimates when the traffic conditions are quickly changing.”

This makes perfect sense. In a work zone you are more likely to see frequent and dynamic queuing. And that is the kryptonite for every algorithm superman. It’s too bad, because we would all like to see accurate travel time estimates, especially for work zones with significant impacts. But, ironically, it is those impacts that make estimation so difficult.

They also discussed the potential use of more advanced algorithms. This is a subject for which I have only a very limited understanding. So I am not able to examine the relative advantages and disadvantages of popular methods. But for work zones, they really aren’t practical anyway. Unless it is a very long term project, one lasting several years, the work required ahead of time to test and adjust the algorithms is expensive and still won’t make much of a difference in the travel time accuracy.

As an industry, we have worked for years to make our systems faster and easier to set up. This, to my mind, would be moving backwards. Instead, let’s work to make our travel time estimates more useful to travelers. Perhaps it makes more sense to talk about delay times. Drivers seem to expect predicted travel times to match their experience perfectly. But when it comes to delay times, they are more likely to be relieved when the delay they encounter is slightly less than predicted.