AI alone is not enough for real-time traffic management
Paolo Rinelli
Global Head of Product Management
Aimsun Live is the ONLY predictive traffic management tool that combines AI techniques with simulation in real time. This unique flexibility provides traffic control rooms with the highest levels of accuracy both under recurrent and non-recurrent conditions, and the ability to compare and rank multiple response plans.
AI is a huge buzzword right now and a panacea for everything, and of course there are companies out there that market the use of AI techniques for real-time traffic management. The suggestion is that AI is “the future” and that those who use simulation are behind the times or doomed to fail because of the higher cost and effort.
At Aimsun, we place a high value on AI, but we insist on using it in conjunction with simulation because it’s the only way you can predict the outcome of previously unseen events and compare the effectiveness of multiple possible response plans.
A strength of AI is that it can find correlations between the data it processes, however, it cannot understand causalities or contexts that are not given explicitly as data input. AI doesn’t invent or make hypotheses; it just observes and finds; it only knows what you tell it.
For example, if you train AI to recognize dogs and cats, when you show it a horse it will tell you it’s either a dog or a cat. It will give you an answer, but you cannot trust it.
All this means that AI is only successful when it can access data that contains the relationships that you want to exploit; you generally want big datasets so that the training is less sensitive to errors and factors that are irrelevant; in other words, so that it can generalize what it finds without understanding causalities.
When using the Aimsun Live decision support tool for traffic management, AI and data science are employed extensively, but – and this is the important part – ONLY for the tasks that they are suitable for. Some of these are:
- Cleaning sensor data, i.e., detecting outliers and filling the gaps with reasonable values
- Classifying historical data, e.g., identifying groups of days that should go together because the trend of their data looks similar
- Extracting from clusters of historical data a band representing the trend of what you measure at that location in days belonging to that cluster
- Identifying a similar-looking cluster from the real-time data
- Fusing a trend observed in real time and a trend that corresponds to that time series of the type of day that has been identified, coming out with a single prediction of how the real-time data will evolve in the short term, over the next hour
- Making a longer-term prediction purely based on contextual factors such as date or weather, based on the correlations between the typical historical trends observed in the past under those different contextual factors
- Comparing over different locations real time data with typical patterns for the type of day that has been identified for today, spot anomalies that are likely symptom of an incident
If you build a traffic management system just with AI, it will ONLY be capable of predicting traffic conditions for areas where you have sensors that have collected historical data for training and that are sending the real-time data; so, for example, if you only have sensors on the beltway around a city, you won’t be able to know anything about traffic conditions inside the city.
Pure AI without simulation is also unable to give you values that you haven’t measured, so if you have sensors that only give you the traffic count, you won’t be able to get travel time.
AI-based traffic management is also limited to predicting traffic under conditions that have been observed in the past, so for example, if a crash closes a road for the first time, you won’t be able to predict its impact.
In Aimsun Live, by using simulation, we can formulate predictions all over the network, even where detectors are not present, for performance indicators that are not directly measured such as emissions, to estimate the impact of events that haven’t been experienced in the past, and to compare multiple response plans.
