Explainable AI for Emergency Landing Decisions: A Comparative Study of Learning Classifier Systems and Neural Networks

Kurzfassung

During mid-flight emergencies, pilots must rapidly assess multiple operational and environmental factors to select a safe alternate landing airport. This Dynamic Alternate Airport Selection (DAAS) process requires fast, reliable decision-making under high cognitive load. Although established cockpit procedures such as those in the QRH provide essential guidance, additional data-driven support tools could further help pilots manage complex information under time pressure. Different Artificial Intelligence (AI) methods offer promising opportunities in this regard, however, for aviation applications, it is necessary that the applied methods are both, accurate and transparent, and that their decision logic is explainable to pilots. Accordingly, this paper investigates which AI methods are most suitable for modelling pilot behaviour in emergency airport-selection tasks while maintaining a high degree of explainability to foster trust in the system. Using a dataset derived from an online survey of professional pilots capturing their preferences across emergency diversion scenarios, and expanded through structured data augmentation to generate 7,140 labelled decision scenarios, the study evaluates two variants of interpretable Learning Classifier Systems (LCS), using Hyperellipsoid and Hyperrectangle conditions, whose decision-making is encoded in explicit, human readable IF–THEN rules that enable direct inspection of how inputs lead to decisions. These models were contrasted with a more modern, non-interpretable baseline: a Feedforward Neural Network (FNN). The models were designed for single-instance classification using a scoring framework. The scores were used to label the augmented dataset by combining the scenario scores with the Euclidean distance between the original decision scenarios and the unique airport combinations generated within each scenario. Model performance was evaluated using accuracy and interpretability considerations, key factors for integration into cockpit decision-support systems. The Hyperellipsoid LCS achieved the highest accuracy (86.34%), demonstrating strong adaptation to multidimensional feature interactions. The Hyperrectangular LCS offered greater rule-level transparency but lower accuracy (78.33%), while the FNN achieved intermediate accuracy (82.20%) with limited inherent interpretability. Results show that the Hyperellipsoid LCS provides the best overall balance between predictive performance and transparency, outperforming the Hyperrectangle LCS and the FNN. These findings indicate that ellipsoid-based LCS models offer a promising foundation for trustworthy AI components in future pilot decision-support systems