Toward Individual-Sensitive Automation for Air Traffic Control Using Convolutional Neural Networks

Abstract

Lack of trust and acceptance caused by strategic mismatches in problem-solving have been identified as obstacles in the introduction of workload-alleviating automation in air traffic control. One possible way to overcome these obstacles is by creating automation capable of providing personalized advisories conformal to the individual controller. This paper focuses on performing an exploratory investigation into the tools and methodology required for creating conformal automation. Central in the creation of individualized prediction models is the combination of a visual feature, capturing traffic situations, and a tailored convolutional neural network model trained on individual controller data recorded from a human-in-the-loop simulation. The main advantage of using a visual feature is that it could facilitate 'transparency' of the machine learning model. Results show that the trained models can reasonably predict command type, direction and magnitude. Furthermore, a correlation is found between controller consistency and achieved prediction performance. A comparison between individual-sensitive and general models showed a benefit of individually trained models, confirming the strategy heterogeneity of the population, which is a critical assumption for personalized automation. Future research should be done in refining the model architecture, finding richer visual features that capture the breadth of human decision-making behavior and feedback model outputs back to individuals for measuring controller agreement.

An algorithmic application with flexible airspace approach

“…Tran et al (2020) developed a digital assistant to assist the ATCo and manage the workload during heavy air traffic. Van Rooijen et al (2020) researched the critical elements to create an automation that offers advice to the ATCo to reduce the workload. Mercado Velasco et al (2020) developed an interface that provides decision support to reduce the workload of the ATCo in the Route Merging Task.…”

Data-driven prediction of Air Traffic Controllers reactions to resolving conflicts

“…Authors in [26] use Multi-Agent Deep Deterministic Policy Gradient (MADPG) to resolve conflicts, also considering time, fuel consumption and airspace complexity. Closer to our approach are methods that somehow consider the ATCO preferences, either in a data-driven way as in [27], [28] and [29], or by using rules and procedures derived from human experts as in [30].…”

“…The method proposed in [29] aims to provide personalized advisories to controllers. Authors train a convolutional NN on individual controller's data recorded from a human-in-the-loop simulation to predict conflict resolution actions.…”

From human-human trust to human-autonomous system trust

“…Indeed, deviations from perception and reality could be the reason for uncalibrated misuse and disuse. This also explains in operator preference for particular strategies suggested by autonomous tools in operation, regardless of solution accuracy (Rooijen, Ellerbroek, Borst, & Kampen, 2020;Westin, Borst, & Hilburn, 2016). As such, for efficient operations, the gap between perceived system performance and actual system performance needs to be reduced, if not eliminated.…”

“…Indeed, the enormous potential afforded by advanced autonomous systems in assisting future ATCO operations has seen significant investment and research conducted in this area (Rooijen et al, 2020). However, the pace of developments in the technical aspects of autonomous system research far outstripping that of trust research, and other human factors research for that matter, can be concerning (Hancock, 2017).…”