61ST ANNUAL CONFERENCE, 23-27 May 2022
WP No. 68
Artificial Intelligence and Machine Learning in ATC
Presented by PLC & TOC
There are many sources and publications available on Artificial Intelligence (AI) and Machine Learning (ML). The objective of this paper is to introduce the reader to the complex yet fascinating world of AI and ML, and give a general insight into the reach and impact of AI and ML implementations. AI and the use of AI is changing and evolving rapidly in many industries including ATM. The information in this paper is backed up by the collaboration between the authors and academic and expert figures in AI.
1.1 In a society that is being controlled by data and information, IFATCA wants to inform its members on the application of Artificial Intelligence (AI) and Machine Learning (ML) in ATM. In 2018 and 2019, IFATCA organised two DigitATMisation conferences  on this subject and published articles and education material. The reason behind these activities, as well as this working paper, is to educate and prepare the ATCO community for an incremental modernisation that will take place in the coming years.
1.2 The reason behind all this attention is what is sometimes called “the 4th industrial revolution” driven by the exponential growth of accessible data stocks that has only just started to emerge. It is expected that by 2030, around 200 billion devices, sensors and people will be interconnected, expanding from 11 billion connections today . The possibilities for the use of this data are numerous, and many if not all industries will be affected. More information on big data can be found in TOC’s working paper on big data , WP No. 91, Singapore 2020.
1.3 Demand will exceed capacity in more parts of the global ATM system in the near future. According to IATA “people want to fly, demand for air travel in the next two decades is set to double. IATA expects 7.2 billion passengers to travel in 2035, a near doubling of the 3.8 billion air travellers in 2016” . More capacity is needed in the air and at airports. To relieve congestion and improve efficiency we might be able to use AI and ML to our advantage. Modernisation is needed and AI and ML could provide us with solutions to the current challenges.
Definitions of AI and ML
2.1 There are many definitions for AI and ML, which each vary slightly. For the purpose of this document the following definitions are used.
2.2 AI (Artificial Intelligence) is defined as “intelligence exhibited by an artificial entity to solve complex problems and such a system is generally assumed to be a computer or machine” . It is intended to create a machine or a software based on available data or impulses that would be endowed with human-like behaviour and/or intelligence. AI is designed to mimic human thinking and would allow a computer or a computer-controlled machine to perform tasks commonly associated with sentient and intelligent beings. Amongst the most important human intellectual processes that AI should perform are the ability to reason, discover meaning, generalise, extrapolate, predict outcomes, make complex decisions or learn from past experience.
2.3 ML (Machine Learning) is seen as a subset of AI. It is the scientific study of algorithms and statistical models that computer systems use to perform a specific task without using explicit instructions, relying on patterns and inference instead. Simplified: an algorithm is the language a system uses. The basic process of machine learning is to give training data to a learning algorithm. The learning algorithm then generates a new set of rules, based on inferences from the data. This is in essence generating a new algorithm, formally referred to as the machine learning model . It is important to understand that in machine learning, it is the learning algorithms that create the rules, and not the computer programmers .
2.3.1 For example, what number should be placed on the ?:
- 3 – 9
- 4 – 16
- 8 – 64
- 9 – ?
How did you come to 81?
That’s exactly the kind of behaviour ML is trying to teach to machines. ML is trying to teach machines to “learn from experience”. ML algorithms use computational methods to “learn” information directly from data without relying on a predetermined equation as a model. The algorithms adaptively improve their performance as the number of samples available for learning increases. ML algorithms find natural patterns within the data, get insights and predict the unknown for better decisions .
2.4 The differences between AI and ML
Simply put, AI applies rules to data, whereas ML looks at implicit rules and conclusions that evolve out of data. When discussing the application of AI in this paper it often also involves the application of ML.
Artificial intelligence is rule-based: build a computer program with rules, algorithms and knowledge explicitly embedded, based on previously acquired knowledge and high-speed computation of options.
Machine learning is data-based: learn bottom-up, from data, by observing and detecting patterns and statistical regularities in historical data to anticipate future events. Machine learning is based on data and input from artificial networks.
One cannot simply discuss AI and ML without mentioning deep learning as well. Deep learning is “a subfield of machine learning concerned with algorithms inspired by the structure and function of the brain called artificial neural networks” . It is designed to mimic the functions and the structure of human neurons, as the perfect basic structure of interconnectivity and functionality. Deep learning is a concatenation of mathematical equations and can handle (big amounts of) unstructured data to discover patterns by itself. See below the hierarchical relations between the three.
Fig. 1 Cousins of AI 
At the very foundation of every functional AI based system we will find at least two, if not all three of the elements described above.
2.5 Types of Machine Learning 
Machine learning has various forms: supervised learning, reinforcement learning and unsupervised learning.
2.5.1 Supervised learning
Is a form of controlled learning where the programme is given examples. For instance, if a system has the task of separating dogs and cats, it will first be shown many different examples of these animals. Then the system will formulate rules to determine why an image shows a dog or a cat. Through practice and by receiving feedback, the system further tightens the rules.
2.5.2 Reinforcement learning
In contrast to supervised learning, the system only receives feedback after the work has been completed. The system first decides whether the image shows a dog or a cat, then the system gets feedback whether the choice was right or not.
2.5.3 Unsupervised learning
The system itself searches for patterns in the available data. This is an uncontrolled way of learning which means there is no steering by means of correct examples. During learning, the system divides the dogs and the cats into categories itself.
2.6 The connection between AI & big data
Big data is a field that treats ways to analyse, systematically extract information from, or otherwise deals with data sets that are too large or complex to be dealt with by traditional data-processing application software . There’s a reciprocal relationship between big data and AI: AI depends heavily on big data for success . In order to make the best use of the information provided, AI needs an appropriate information architecture, this means that information needs to be structured. It all starts with an enormous legacy of data that is being processed (through cleansing, training, testing), optimised and only then “fed” to AI. As an AI system uses data it will also learn from it and provide feedback, hence closing the circle, and adding more data to the existing data . Where big data comes with the raw material, AI and ML will learn from it, categorise it, analyse it and organise it, connecting the dots between different types of information for specific, predetermined outcomes.
The practical applications of AI and ML
2.7 It may still sound like science fiction, but smartphones, cars, banks and houses make daily use of AI. Sometimes striking: if Siri leads you to the nearest gas station or suggests your route when it detects you leaving home to drive to work. Sometimes less obvious: when you get a fraud notification about an abnormal purchase with your credit card. Intelligent computers are everywhere and have a big impact on our daily lives. There are multiple applications of AI and ML in today’s society.
2.8 One of the most used application of AI in the airline industry  is in the field of airline customer service, from the moment a potential passenger looks for travel destinations, up to the end of the journey. Through chat bots AI permits airlines to better connect with customers, through automated servicing, using message applications (e.g. t-Social Chatbot powered by IBM Watson). ML allows customer segmentation, using the data at hand (reference data, reservations, ticketing, departure control) that the algorithm interprets as instructed (“crunching numbers”) thus providing with an adequate interpretation of resulting data.
2.9 In May 2019, EUROCONTROL organised the Aviation and Artificial Intelligence conference. It became clear that while it is only the beginning of the possible use of AI and ML in ATM, there are already many interesting developments and lessons learned for the future . Subjects at this conference were optimisation of operations, pilotless or single pilot operations, single manned controller positions and full automation of ATC .
2.10 EUROCONTROL is researching ways in which AI can be used to assist and support the controller. AI could be used in the planning phase for more accurate predictions and better traffic load forecasts. In close cooperation with several universities, EUROCONTROL is developing AI for predicting controller workload and sector capacity, aircraft trajectories and runway capacity, and they are also looking at some safety critical tasks for which AI may be used .
2.11 EUROCONTROL’s integrated Flow Management Position (iFMP) at Maastricht Upper Area Control Centre analyses the routes that ATCOs give to aircraft, to find out if paths can be identified to make them more standardised. Via ML the program is trained on historical aircraft trajectories and a set of predictors (e.g. entry time coordination point, entry Flight level, Requested Flight level, Exit time coordination point, Aerodrome of departure, Aerodrome of destination) & time/day of the week & status of military areas. ML enables the system to predict the statistically most likely route through the airspace and keeps up with route and procedures changes in the system. MUAC uses this system on about 10% of their traffic, and already announced that 65% of the predictions of the neural network stay within 6 NM of the route eventually flown (and 94% within 15 NM). For the original EFD (ETFMS flight data) predicted route from the FMP network manager, these percentages are only 10% (6 NM) and 60% (15 NM) .
2.12 KLM test robot for aircraft engine maintenance 
Aircraft engine maintenance company KLM Engine Services has carried out tests with a robot that performs visual inspections for aircraft engine maintenance. The robot inspects the engines with the help of a camera and image analysis. It is expected that the system will eventually replace the visual inspection by technicians.
2.13 AAATCS Automated Airport ATC Service  is designed by IATAS (International Air Traffic Automation Systems) and claims to be able to deliver next-generation solutions for ANSPs and airports to manage busy airside environments. Their product is claimed to be a highly modular and configurable system, as a controller tool, semi-automated solution, or as a fully automated tower control system. The system supports aircraft with CPDLC and IATAS’s dynamic airport moving map. IATAS aims to provide ANSPs with a solution “drastically lower ATM related costs while providing highest possibly safety and capacity” and reducing “ATC or pilot reduced runway incursions or excursions” .
2.14 Training of ATCOs
Adaptive learning, also known as adaptive teaching, is an educational method which uses computer algorithms to orchestrate the interaction with the learner and deliver customised resources and learning activities to address the unique training needs of individuals. In professional learning contexts, individuals may “try out” some training to ensure they engage with it. Smart interfaces can adapt content and the method of lesson delivery at a personal level for each student based on personal data and motivation, therefore enhancing learner engagement . This way ATCOs can be trained via e-learning and simulator training using AI to make an individual program.
Consider the following: “Turning final from downwind” as training exercise. AI could calculate winds, wake turbulence separation, speeds, and so on and thereby calculate the most effective landing sequence, faster than the trainee air traffic controller. If the trainee decides to clear an aircraft for the turn, AI will visualise on the display the outcome of this instruction and might also provide a proposed “best” or most effective solution. AI will be able to explain the difference between different solutions as it’s based on a mathematic and rule-based formula just like playing chess (where, for now, AI is world champion). Additionally, AI may replay the same exercise from the point where the learning can be improved. The student can train again and again, and an individual learning curve can be followed.
2.15 Recognise given clearance using audio in the simulator for training purposes 
The Adacel Maxsim simulator of US Air Force has 90 plus simulators with no simulator pilots. The FAA Tower simulator also has 50 plus simulators without simulator pilots. These simulators work rather simply: the objective is determined by the instructor and the desired scenarios uploaded. When AI is applied, objectives can be added to the simulator and workload can be managed (increase or decrease of workload is possible) based on the number of corrections in a transmission, number of calls, silence/pauses etc.
2.16 A-FOD (Automated Foreign Object Detection System)
This system is currently installed at Gazipaşa International Airport, Turkey. A-FOD is an airport-based system designed to detect the presence of FOD (foreign objects and debris) and wildlife on the manoeuvring area. Among the most common detected FOD there are: aircraft loose parts (metal bolts, strips, lose hardware), pavement fragments (concrete, asphalt bits and pieces), sand, rocks, pieces of luggage. As advertised by its developer, the automated system is supported by an AI central unit, providing information on FOD, wildlife and runway deposits (snow, ice) to the tower controllers. It is capable to identify and classify FOD, issue alerts, indicate the position of the FOD, store information gathered, keeping the false alarms rate within tolerable levels.
2.17 Linking ATC to AI and ML
In real traffic, there are two basic operations that an air traffic controller does: controlling the aircraft and planning the traffic. As mentioned in an article from Journal of Theoretical and Applied Information Technology , AI could be implemented in air traffic control in various methods. Starting with the complex and highly specialised nature of the work, adding the many variables that need be taken into consideration like aircraft performance, airspace design, spacing, weather, and so on, an AI driven system can be generated. The system would have AI and ML capabilities, as it would learn by itself and make decisions based on assumptions, justifications  and possibly from previous experience. In order to obtain a healthy and valid decision-making logic e.g. for collision avoidance systems, there are various artificial intelligence techniques available. In the mentioned study there are a few of the path finding AI techniques currently employed by researchers: A* algorithm, the Markov decision process, Neural networks, Reinforced Learning and Support Vectors Techniques . As reinforced throughout the said paper this is a system designed to remain human centred.
2.18 AI at Heathrow
According to Andrew Taylor, Chief Solutions Officer for NATS, there is an AI-based airfield monitoring system being designed by NATS Digital Tower Laboratory, currently for Heathrow. It aims to increase the capacity of the airport that nowadays is lost due weather (about 20%) . A combination of high-sensitivity sensors, cameras, and the AIMEE platform (AI and machine-learning platform developed by Searidge) provide controllers with an ultra-high-definition image, with an extra layer of data to accompany it.
2.19 When considering the definitions as stated above one should bear in mind that despite the continuous advances in computer processing speed and memory capacity, there are as yet no programs that can match human flexibility over wider domains or in tasks requiring much everyday knowledge. On the other hand, some programs have attained the performance levels of human experts and professionals in performing certain specific tasks, so that artificial intelligence in this limited sense is found in applications as diverse as healthcare diagnosis, targeted treatment, computer search engines, service robots, and voice or handwriting recognition .
2.20 An incremental modernisation is expected in the coming years with the implementation of AI and ML in ATM. However, not all changes are incremental (in the beginning). They might be in the future, but research, testing, developing and maturing is necessary.
2.21 Some people consider AI as the almost limitless solution to problems arising in operations in any field but there are also those who regard AI with scepticism or fear that the decisions made by AI are sometimes not understood and therefore cannot be checked by human operators. Especially when it comes to evolving AI in a field where safety is one of the top priorities it is clear that the decisions made by AI should be fully understood and controllable. There should always be human supervision on the applications and the data used. As well as a thorough understanding of how AI and ML decisions are made. AI and ML are now often implemented as a black box solution where auditing the results of decisions made is not possible.
2.22 Amidst the controversies regarding AI, one report  suggests that overall productivity will benefit from human-machine collaboration. It will improve as “human capability is wide and shallow. We can do many tasks but none perfectly. AI, on the other hand, is narrow and deep: great at specific tasks, but within a limited range”. These complementary attributes provide the basis for an efficient and productive human-machine collaboration.
2.22.1 Joint human-machine systems entails the complementarity of human and machine. A human–machine system is a system in which the functions of a human operator (or a group of operators) and a machine are integrated . The aim is to design a synergic system that focusses on and uses what people are good at and what technology is good at, and design them to complement and possibly reinforce each other . What most experts agree on is that AI will bring changes and disruption in many ways and will impact the way humans and machines interact. So far automation has proved it can increase employee capacity by allowing for more complex and higher-level services .
2.23 The complexity of the human-machine collaboration needs to be fully understood in order to design safe, functional and efficient systems. One way to cope with disruptive innovation in the future is the approach that the UK Civil Aviation Authority has with regard of the implementation of new technologies. The sandbox-approach is a separate and safe environment, with clear boundaries, that allows experiments with new techniques while the consequences of failure can be contained.
2.24 AI and ML are systems made possible by humans – they are developed, monitored and managed by humans. Major developments and changes in ATM are just emerging and this is the time to give input and shape the future. Controllers should be aware of the applications that are possible and the ways they can be used. This is the time to have an influence on the ways systems are developed to best suit our needs.
2.25 ATCOs should be trained to use the new AI and ML systems. Education and training strategies should be in place to properly prepare the current and next generation to use AI for the sake of continuity and standardisation.
Conditions for implementation of AI and ML
2.26 While data is widely available in ATM, the usability is not always obvious. To be able to use data for AI or ML, the data needs to be trustworthy.
2.27 In 2018 the European Commission set up an independent High-Level Group on Artificial Intelligence (AI HLEG) that was tasked to develop a set of solid rules for a safe development and implementation of AI. The result was the development of the Ethics Guidelines for Trustworthy AI. The guidelines are an overall framework for trustworthy AI, naming the foundation, the requirements, the technical and non-technical methods to realise trustworthy AI, as well as the opportunities and the critical concerns related to AI.
A trustworthy AI has three components :
- lawfulness, complying with all applicable laws and regulations;
- ethicality, ensuring adherence to ethical principles and values; and
- robustness, both from a technical and social perspective, since, even with good intentions, AI systems can cause unintentional harm.
The human brain is capable of extrapolation and generalisation, something that AI cannot do. It takes a huge amount of work and capacity to make unstructured data useable.
2.28 The AI HLEG also  created a list with seven solid requirements to achieve a trustworthy AI, that should include systemic, individual and societal aspects as the following:
- Human agency and oversight. Including fundamental rights, human agency and human oversight.
- Technical robustness and safety. Including resilience to attack and security, fall back plan and general safety, accuracy, reliability and reproducibility.
- Privacy and data governance. Including respect for privacy, quality and integrity of data, and access to data.
- Transparency. Including traceability, explainability and communication.
- Diversity, non-discrimination and fairness. Including the avoidance of unfair bias, accessibility and universal design, and stakeholders’ participation.
- Societal and environmental wellbeing. Including sustainability and environmental friendliness, social impact, society and democracy.
- Accountability. Including auditability, minimisation and reporting of negative impact, trade-offs and redress.
These seven requirements should be continuously evaluated and addressed throughout the AI system’s life cycle.
2.29 Regulatory compliance
Considering the commitment to safe skies, the future of AI in aviation needs to adhere to the same sound principles of standardisation, harmonisation and lawfulness. As previously shown, there are expert groups already setting ground-based rules. Regulatory aspects need to be addressed from the very beginning. In Europe the AI4EU platform was initiated. EUROCONTROL acknowledges and desires to participate in the AI revolution, foreseeing the benefits that AI will bring to the ATM system. EUROCONTROL plans to promote the use of an AI on demand platform in the SESAR ATM modernisation programme and accelerate the adoption of AI-based solutions . Specifically the UK, NATS, Searidge and UK CAA are currently developing a new regulatory framework, in order to develop specific standards for AI airport operations, technologies and procedures .
Basic SWOT (Strengths, Weaknesses, opportunities and threats) ANALYSIS [38, 39]
2.30 Previously, the most important initiatives of AI/ML within ATM that could actually influence ATCO’s daily work were identified. An analysis is required to see in what way the ATM system is suitable for applications of AI, and what the possible threats, weaknesses, strengths and opportunities are.
2.31.1 Weakness/threats Human-machine interaction
When it comes to human machine interaction there are some challenges to be tackled. One of them is determining the best interface between human and machine that transmits intent and awareness of current and pending condition. There is always the debate on what would be the optimal structure of shared human-machine responsibility and authority. Another challenge is how should humans and machines be coordinated to maximise the probability of desired outcomes and minimise the risk of adverse outcomes for a given decision or mission objective. Humans are poor monitors of complex automated systems  and as systems become more automated, the role of humans transitions further from that of operator to monitor as system complexity can exceed the supervisory abilities of the human. Another problem could arise due to the inability for the ATCO to take over when a system fails or when there is a malfunction. There is an argument that the controller needs to be more rather than less skilled, and less rather than more loaded than average. Humans have advantages in creativity and the ease with which they draw on previous experiences to reason and solve new and complex problems, but AI and ML aren’t capable of this (yet) . Lastly the composition of the skills required to use AI and ML are expected to evolve faster than the skills of the available human resources causing a widening of the labour market skills gap .
2.31.2 Weaknesses/Threats of AI/ML
One of the down sides of AI/ML is that it has trouble when coping with uncertainties such as weather, equipment failure, conflict alert, emergencies and flexible use of airspace. Humanists could argue that, since it’s a “machine” AI/ML is not self-aware, can’t extrapolate and has no compassion. Use of automation reduces the workload during normal operations but may add complexity and workload during demanding situations . Critical for the safety of operations is the labelling of data, determining what is good and what is not (e.g. training of ATCOs). There is a certain amount of uncertainty regarding how the system will make its decisions based on ML as it has been previously shown that ML can produce inexplicable data, untraceable data, as well as the uncertainty on how the system will behave when procedures change in a manner that is new to the model. Constant change is another issue: does AI and ML need a re-learning phase to adapt to change or should it be re-trained? It may be useful to apply simulator training before the planned change.
2.31.3 As shown before AI requires large amounts of data. Is there enough data for AI/ML available to feed the system? Do we have enough emergency data to train the model for every situation and airport in the world? One of the vulnerabilities of AI is the “Byzantine fault” described as “a condition of a computer system, particularly distributed computing systems, where components may fail and there is imperfect (or incomplete) information on whether a component has failed”.
One could argue that there could also be geopolitical or economical risks when a global AI system would be owned by one supplier or governing state. The responsibility of all stakeholders needs to be clear. It is of utmost importance that thorough testing is made available for end users (ATCOs). Finally, cybersecurity and bias from data are a threat to the wellbeing of an AI-based system.
2.32.1 Strengths of human-machine interaction
Humans have advantages in problem solving in certain scenarios and are good at adapting to rapid unexpected changes whereas machines can perform better in maths-intensive problems and when working with large databases. They can supplement and strengthen each other.
2.32.2 Strengths of AI/M
AI/ML performance can easily be replicated, is broadly more consistent and easily transferable. It can reduce the human workload while providing more consistent results during operation. Humans are limited to a specific workload, area of responsibility and number of calls, when applying AI and ML the workload can be monitored more accurately and the area of responsibility can be significantly larger.
Another aspect is that AI/ML can use multimodal interfaces (utilising two or more means of input), which is more reliable than unimodal interfaces. For humans, multimodal communication increases the demands on the working memory, two simultaneous tasks require separate perceptual areas, such as verbal and visual, they are performed as doing the tasks individually. Alternatively, when someone carries out two tasks simultaneously that use one perceptual domain, their performance declines . Multimodal interfaces are viewed by their users as a more human method of working with automated systems. One big advantage of AI and ML is that it can find correlations and patterns between even seemingly unrelated things.
As a machine AI is not susceptible to short term memory, information overload, sleep deprivation or distractions, it will never get tired, is pro-active, and will respond promptly to the input given. AI will increase reliability and reduce costs and of course it is also able to process an enormous amount of data and to learn, abstract and reason . An AI based system will strictly follow protocols and instructions and will be able to optimise existing processes. It can show operational decisions with accurate predictions of outcome and impact.
2.33 Opportunities of human-machine interaction
As already shown, human and machine present the possibility of working together more collaboratively to exploit each other’s strengths and minimise their respective weaknesses. This will maximise the probability of desired outcomes and minimise the risk of adverse outcomes for a given decision point or mission objective. Developers should acknowledge the fundamental differences of humans and machines and design systems that complement each other while minimising their limitations.
Machines need to be able to determine the state of the human operators monitoring or adjusting their settings. They could then adapt themselves in a manner making them more useful for the critical tasks at hand . Adaptive automation was tested in operators of simulated UAVs and showed a significantly improved performance. Parameters could be triggered to reduce the workload when automation detects an increased mental workload . Errors might be reduced if multi-modal interfaces are built in such a way as to help the human understand the interaction of the system. The interface must present data in such a way that the user can easily interpret and process the information while also considering the user’s abilities and preferences .
Multiple fallback scenarios should be in place so as not to create a single point of failure. Humans should understand their inherent limitations and build even better platforms and intelligent systems. We need to prepare the next generation with upgrades in computer science instructions in elementary education and at secondary level.
2.34 Opportunities of AI/ML
ML can provide us with new insights regarding the data abstracted and processed. Blending different data sources is a great opportunity to identify patterns and use these better analysed data to predict future behaviour . This list allows for improvement. There is an ongoing development for AI and ML and there is an abundance of information and examples available.
IFATCA policies on automation
2.35 Currently, there is no IFATCA policy on AI or ML. There are however multiple IFATCA policies on automation, from which future policies may be derived.
|IFATCA Policy is:
Automation must improve and enhance the data exchange for controllers. Automated systems must be fail-safe and provide accurate and incorruptible data. These systems must be built with an integrity factor to review and crosscheck the information being received.
The Human Factors aspects of Automation must be fully considered when developing automated systems.
Automation must assist and support ATCOs in the execution of their duties.
The controller must remain the key element of the ATC system.
Total workload should not be increased without proof that the combined automated/human systems can operate safely at the levels of workload predicted, and to be able to satisfactorily manage normal and abnormal occurrences. Automated tools or systems that support the control function must enable the controller to retain complete control of the control task in such a way so as to enable the controller to support timely interventions when situations occur that are outside the normal compass of the system design, or when abnormal situations occur which require non-compliance or variation to normal procedures.
Automation should be designed to enhance controller job satisfaction.
The legal aspects of a controller’s responsibilities must be clearly identified when working with automated systems.
A Controller shall not be held liable for incidents that may occur due to the use of inaccurate data if he is unable to check the integrity of the information received.
A Controller shall not be held liable for incidents in which a loss of separation occurs due to a resolution advisory issued by an automated system.
Guidelines and procedures shall be established in order to prevent incidents occurring from the use of false or misleading information provided to the controller .
3.1 AI and ML are made possible by the humans that develop, monitor and manage them. There should always be human supervision of the applications and the data used, as well as a thorough understanding of how AI and ML decisions are made.
3.2 AI and ML are being applied in many industries around the world. AI and ML can be applied in many ways, and they will change the way the air traffic controllers work. Nonetheless, this does not mean that AI or ML will take over the work of a controller any time soon. AI and ML can support the decision-making process of the controller, leading to improved effectiveness and efficiency of the ATM system as issues of capacity arise in more places around the world.
3.3 Human-machine interaction should make the “best of both worlds” – humans with their ability to imagine new solutions and adapt to unforeseeable situations; machines with their learning abilities, with the capacity to handle enormous amounts of information and providing insight into patterns in data. The key to achieving well-guided AI is to establish a human-machine coexisting environment, where the machine becomes a “sidekick” that supports humans, instead of being a prospective “rival”.
3.4 One crucial aspect is trustworthiness and how responsibility will be shared between AI developers (programmers/software developers) and AI users (ATCOs)/buyers (ANSPs). Unwanted outcomes should be foreseen and avoided. Thorough testing should be made available for end users (ATCOs).
3.5 The complexity of AI and ML is a point of concern as currently programmers cannot always reproduce or explain the results of ML. This creates one of the biggest challenges for ML: to ensure that it is understandable and controllable.
3.6 Education and training strategies should be in place to properly prepare the current and next generation of controllers to use AI for the sake of continuity and standardisation.
3.7 Safety is key when it comes to the aviation industry. However, technology will progress, there will always be a justifiable need to thoroughly test and validate machine-intelligence based systems. AI will have to be proven safe first in a testing environment before being gradually incorporated into the ATM system.
3.8 AI and ML are potential game-changers for the ATM system, and they should be seen as the next logical step into a sustainable future. IFATCA will have to stay alert and scan the technical revolution that is going on, with the intent to self-disrupt before other intruders disrupt us. It is best to be ahead of the game instead of being a mere follower.
3.9 Automation has brought positive changes in the work environment of air traffic controllers in the past. For the benefit of safety and in order to accommodate the expected increase in air traffic in the years to come, consider the benefits of AI and accept them as a step-up of automation.
4.1 It is recommended that IFATCA policy is:
Artificial Intelligence and/or Machine Learning based systems should only be implemented as Decision Support Systems and shall not replace the decision of the ATCO.
Where an ATCO is responsible for decision making, and in the event that system tools fail or are not available, the ATCO should always have the capacity to safely manage their area of responsibility.
And is inserted in the IFATCA Technical and Professional Manual.
References http://www.ifatca.org/digitalisation-in-air-traffic-management/ by Marc Baumgartner, SESAR expert.  IFATCA Digit ATMisation newsletter: Pass your message #2, Second IFATCA workshop on Digitalisation in ATM, February 2019, Geneva.  Developed by IFATCA, TOC in 2019 “Big Data”.  IATA – 20 Year Passenger Forecast, http://www.iata.org/publications, Oct 2018.  George F. Ludger (2009). Artificial Intelligence – Structures and strategies for complex problem solving. 5th Edition, Pearson.  “Artificial Intelligence and Machine Learning: Policy Paper”- April 2017, Internet Society, https://www.internetsociety.org/resources/doc/2017/artificial-intelligence-and-machine-learning-policy-paper/  “Artificial Intelligence and Machine Learning: Policy Paper”- April 2017, Internet Society,
https://www.internetsociety.org/resources/doc/2017/artificial-intelligence-and-machine-learning-policy-paper/  https://becominghuman.ai/machine-learning-for-dummies-explained-in-2-mins-e83fbc55ac6d  https://machinelearningmastery.com  https://towardsdatascience.com  https://app.ai-cursus.nl/home  IFATCA, “Big Data” WP 91 Paper, 2020.  https://enterprisersproject.com/article/2019/10/how-big-data-and-ai-work-together  Big Data Information Architecture for AI by Toby Woolfe, presentation during Big Data and AI Conference in Vilnius, 26-28 Nov 2019.  “Artificial Intelligence and Machine Learning in Aviation”- Aviation Data Symposium 2018- OpenJaw- IATA.  IFATCA Meeting Report: AI and Aviation conference, EUROCONTROL 23/05/2019, by Marc Baumgartner, SESAR Coordinator.  https://www.eurocontrol.int/event/high-level-conference-artificial-intelligence  Skyway magazine, EUROCONTROL, autumn/winter 2019.  EUROCONTROL newsletter (2018). Predicting Flight Routes (iFMP), February 2018.  http://www.computable.nl  http://iatas.aero/products/aaatcs.html  http://iatas.aero/products/aaatcs.html  Artificial Intelligence and the Future of Work- ICAO Training Report No 3 2017, Kristina Schneider, p. 1.
https://www.icao.int/training/Documents/A.I.pg.14.15.16.icao_training_report_vol7_No3.pdf  AI in ATC by Vincent Lambercy published by FoxATM.  “Embedding Machine Learning in Air Traffic Systems to Generate Effective Route Plans For Aircrafts In Order to Avoid Collision”- Journal of Theoretical and Applied Information Technology 31st January 2019 31st January 2019.  Embedding Machine Learning in Air Traffic Systems to Generate Effective Route Plans For Aircrafts In Order to Avoid Collision”- Journal of Theoretical and Applied Information Technology 31st January 2019 31st January 2019 (Analysis and Design of machine Learning System for ATC).  Ibid 31,32.  Skyway magazine, EUROCONTROL, autumn/winter 2019.  https://www.britannica.com/technology/artificial-intelligence  “The Intelligence Revolution: Future Proofing Canada’s Workforce”. Human Resources Professionals Association and Deloitte Canada. https://www2.deloitte.com/ca/en/pages/human-capital/articles/intelligence-revolution.html  https://en.wikipedia.org/wiki/Human%E2%80%93machine_system  The concept of MABA-MABA (Men Are Better At-Machines Are Better At).  Artificial Intelligence and the Future of Work- ICAO Training Report no 3 2017, Kristina Schneider.  Based on the High-level expert group on artificial intelligence (AI HLEG), set up by the European commission.  Based on the High-level expert group on artificial intelligence (AI HLEG), set up by the European Commission.  Skyway magazine – EUROCONTROL, autumn/winter 2019, p. 33.  Skyway magazine, EUROCONTROL, autumn/winter 2019.  AI in ATC by Vincent Lambercy.  Multimodal interfaces (utilizing 2 or more means of input) are more reliable than unimodal interfaces.  Ironies of automation, Lisanne Bainbridge, Automatica Vol 19, No 6, 1983.  “Artificial Intelligence: What Everyone Needs to Know”, Jerry Kaplan, 2018.  AI in aviation White paper, IATA, 2018.  Hindsight20, winter 2014, Some lessons learned about pilots and flight deck automated systems, K. Abbot, 2014.  “Human Machine Interaction”, Dennis Lalanne & Juerg Kohlas, 2009.  According to DARPA (US Defence Advanced Research Projects Agency), these capabilities are related to the so called “Second wave of AI” https://www.darpa.mil/  “Human-Machine Intelligence Interaction in Aviation”, A. Taylor, T.S. Cotter, Proceedings of the American Society for Engineering Management 2014 International Annual Conference S. Long, E-H. Ng, and C. Downing 2014.  “Human-Machine Intelligence Interaction in Aviation”, A. Taylor, T.S. Cotter, Proceedings of the American Society for Engineering Management 2014 International Annual Conference S. Long, E-H. Ng, and C. Downing 2014.  “Human-Machine Intelligence Interaction in Aviation”, A. Taylor, T.S. Cotter, Proceedings of the American Society for Engineering Management 2014 International Annual Conference S. Long, E-H. Ng, and C. Downing 2014.  AI in aviation white paper IATA, June 2018.  IFATCA TPM 2019, WC 8.2.5 Automation/Human factors, pages 193-194.  “Robotics, Automation & Data Analytics 2019”, Plenary Speech on “Intelligent Aviation” Dr.Assel Abdel Hamied Moussa.
Last Update: August 2, 2022