By Eric Murphy
State transportation agencies are cautiously dipping their toes into the waters of “artificial intelligence” and “machine learning” to find applications in the transportation field. There are many potential uses, according to a new report, including opportunities to track assets like crosswalks, and to clear traffic incidents faster, which could lessen the need for major capacity investments. Agencies have also identified some lessons and pitfalls of the technology as they pilot new tools.
While the terms artificial intelligence and machine learning can give a false impression of intelligent computers that produce results on their own as if by magic, the tools process and analyze large amounts of data that people select to find patterns and insights that people can then interpret and use. In Nebraska, for example, the DOT used machine learning tools to analyze large amounts of video data to identify the location of crosswalks and create an inventory of those assets.
The National Cooperative Highway Research Program gathered several other examples of state DOTs using machine learning in their transportation systems. In Missouri, machine learning was used to create predictive analytics based on traffic volumes, weather, and other factors to determine where crashes were more likely to occur, which could lead to faster response times. Under Missouri’s current system of incident identification, congestion and secondary crashes have sometimes already occurred by the time the agency becomes aware of the incident. About half of all congestion is due to non-recurring factors like crashes.
But machine learning was not a silver bullet for these agencies, and the newer technology often comes with some bumps in the road. MoDOT operators, for example, found that having several data sources feeding into the machine learning platform created many duplicate incidents, requiring more work from operators. Upon first deployment, the tool was less accurate, with accuracy improving over time as it processed more data.
Because machine learning tools operate based on historical data, they may have trouble dealing with novel and unprecedented circumstances and reproduce past inequities, as has occurred in tests in other fields. Another potential pitfall is choosing the right data to “train” on which to train these tools and making sure the data is as clean as possible. The large amount of electrical and computing power needed by some machine learning or AI tools to process vast amounts of data can make their use unfeasible, very expensive, or counterproductive to state or agency climate goals.
State DOT leaders can educate themselves on the costs and benefits of machine learning techniques, guard against overconfidence in their results, and understand where they could be most beneficial and appropriate. Starting with small pilot programs to test the potential uses of this new technology can allow agencies to experiment and better understand its uses while minimizing risk.
As the Nebraska DOT found:
The project team focused specifically on low-cost and low-risk applications with clear foundational benefits. They intentionally avoided more complex ML applications with large project budgets, e.g., inventorying every asset and spending millions of dollars. The team suggests caution in adding complexity to the project as it increases the risk of not achieving the desired success.