The use of artificial intelligence (AI) in air traffic control (ATC) is promising. There are mixed opinions. In an ideal world, we’d have developed AI controllers that are capable of instructing aircraft safely and expeditiously. AI experts who don’t know a lot about ATC make predictions that may not seem realistic. ATC experts who don’t know much about AI naturally tend to be resistant to adopting it into their operations. Vendors continue to promote their solutions, as they always do.
ATC is highly complex. Controllers modify each aircraft’s trajectory to keep a minimum distance between them at all times. It is necessary to avoid “loss of separation” (LoS) between aircraft while ensuring they safely reach their destinations. Parallelly, they must also maintain a range of backup plans to ensure safety in the face of unexpected situations. Other considerations like the orderly transfer of planes between sectors, fuel efficiency, environmental improvements, landing sequence optimization, etc., need to be considered as well.
To optimize these elements, each flight’s potential future, predicting possible conflicts, and issuing timely instructions, need to be looked into. And there are many sources of uncertainty, from aircraft mass to individual pilot behavior, airline preferences, and weather conditions variations. Any decision-making system applied in air traffic control accounts for these uncertainties while optimizing for the other objectives. In other words, this is an exceptionally complex arena that offers rich opportunities for high-impact developments in machine learning.
For instance, let’s consider a self-driving car. It can sense its environment and navigate without human input. Each self-driving car is usually outfitted with a GPS unit, an inertial navigation system, and a range of sensors, including laser rangefinders, radar, and video. The car uses positional information from the GPS and an inertial navigation system to localize itself. It uses sensor data to refine its position estimate. It builds a 3D image of its environment.
Now, data from each sensor is filtered to remove noise and often fused with other data sources to augment the original image. Pretty cool, huh? How the car uses this data to make those accurate navigation decisions is determined by its control system.
The majority of self-driving vehicle control systems implement a deliberative architecture1, meaning that they are capable of making intelligent decisions by:
1) maintaining an internal map of their world and
2) using that map to find an optimal path to their destination that avoids obstacles (e.g., road structures, pedestrians, and other vehicles) from a set of possible paths.
Once the vehicle determines the best path to take, the decision is dissected into commands fed to the vehicle’s actuators. These actuators control the vehicle’s steering, braking, and throttle.
This localization process, mapping, obstacle avoidance, and path planning are repeated multiple times each second on powerful on-board processors until the vehicle reaches its destination. And this is pretty much how ATC infused with AI works.
When AI is used to replace or enhance some of the controller functions, it is trained on a set of goals. For example, feed aircraft to the active runway’s final approach while maintaining prescribed separation. Aircraft and ATC have access to conceptually similar sensors as those in the self-driving car.
Once you train the AI system with aircraft track data, it allows the software to determine what the aircraft does when an instruction is given. For example, it learns the relationship between speed and turn rate. It learns the performance profile of aircraft types and even airline profiles etc. The model allows the AI to make control decisions based on each aircraft’s characteristics so that it can achieve the prescribed objectives.
In other words, it knows the outcome of the instruction that it is about to give, very precisely. Additionally, given that it can be aware of much more traffic simultaneously, it can determine much like playing chess. The impact on the traffic situation will be that many more moves ahead than a human can. The further along you can see and predict, the fewer deconfliction instructions you will need to give.
When using neural networks, we do not know why the model may make the decision it makes. Another challenge is how the model will behave if procedures change in a manner that is new to the model. It can be a real problem when dealing with aircraft emergencies.
One of the things ATCs are good at is to adapt to rapid changes. For example, a closed runway, airport, chunk of airspace, or weather can have a massive impact on en-route controllers when aircraft have to divert to avoid such surprises. Such factors can affect air traffic flows massively and rapidly. How good would then an AI be if the situation dramatically diverges from the training dataset? Or will it be trained for such cases as well? If so, how to get sufficient significant data?
These are all very valid questions.
AI is always going to adapt much faster than humans. The model used for controlling traffic cab be trained to handle such situations. The good news is that with Brainalyzed Insight’s swarm AI technology and vast amounts of data, you should have enough data to create models for all situations.
As much as we might think air traffic is chaotic, it is mostly consistent for each airport. Of course, external events always interrupt that consistency, but if AI is given goals and objectives, then such external events can be handled easily.