Log in


2022 Symposium Technical Program

AI applications in Crew Scheduling and Recovery problems

In post-COVID era, where operations tend to be less regular, flexibility and performance of scheduling and recovery become more important factors. CAE is actively developing solutions that utilize AI and machine learning models to improve efficiency of crew scheduling, manpower planning and demand forecasting. We will present AI-based approach to legality rules evaluation, resulting in optimizers performance improvement up to 80% while retaining solution quality. In the recovery problem, we improved overall performance and objective value by applying AI models to support selection of candidates being able to help efficiently recover disrupted schedules or pairings. We also are developing  an AI approach that will estimate runtimes for selected parameters and let users require a particular performance/quality ratio. Finally, we will present our approaches to combined pairing-rostering problem, where AI models are utilized to support schedule generation.

Jaroslaw Pyzik, CAE

Aircraft landing gear failure prediction using a machine learning model

In Vueling, landing gear failure is one of top-three cause of AOGs due to aircraft maintenance. The main reason is that  daily checks of the landing gear involve inspection of multiple components by engineers and in some cases it is difficult to identify issues with particular components. Hence, a machine learning (ML) model is developed to identify the critical gears and report them to the engineers so for specific checks. The proposed classification ML model considers different source of historical data of each gear including sensor data, weather data, part removal maintenance data, and operational data until one day before the removal date. Having tested this model in a period of two months (May and June of 2022), it has been observed that the model would be able to identify more than 72% of the AOGs caused by the landing gear one day before occurrence.

Mohammad Kouhi, Vueling

Airport Capacity Imbalance Problem

Connecting airports in a network can lead to delays because they are dependent on one another.   When one of the airports of the network is unable to accept the appropriate number of operations in time, it causes propagation of disturbances throughout the network. A sudden and unexpected drop in airport capacity may occur under conditions where operating minima do not allow operations at that airport. This causes flight diversions to other airports and makes it impossible to make efficient transit connections at the hub airport. The undertaken research topic focuses on the airport capacity imbalance problem.  A lack of capacity symmetry may lead to throughput imbalance. The presentation shows a methodology that allows to assess the potential lack of capacity along with the probability of its occurrence, what is of key importance when planning airport layout and resources for air connection network.

Janek Malawko, ICM

An approach for condition monitoring and diagnostics of business aircrafts in practice

This presentation details a data driven approach for automated health monitoring and diagnostics for business jet powered aircrafts using the onboard aircraft health monitoring system dataset. The approach is tuned for efficiency of operation, providing value to maintenance personnel, and leveraging existing safety related infrastructure. Using classical machine learning within an anomaly detection framework, health monitoring is reframed as a multi-variate time series anomaly detection problem accounting for sensor signal interdependence and flight-phase effects. Diagnostics are implemented as an anomaly-to-component mapping problem, the solution to which leverages existing failure mode and effects documents, maintenance history, and subject matter expertise. Together, the proposed approach provides maintenance personnel a convenient tool to detect potential component and operational issues in the aircraft with minimal manual intervention.

Ashwin Rai, Ivado Labs

An optimization framework for long-term fleet forecasting with integrated emissions targets

The Airbus Global Market Forecast has historically used a 'liberal unconstrained' model, whereby past drivers of growth were assumed to continue unchallenged into the future. In the face of increasingly stringent emissions targets and the implied consequences - such as rising energy prices and supply constraints - it is clear that this assumption no longer holds, and our models now need to explicitly anticipate these drivers when constructing fleet plans. In this talk, we introduce the latest revamp of our delivery forecast, a MIP model that takes the long view and aims to optimize each airline's fleet and energy mix subject to a number of constraints over a multiyear period. In particular, we will focus on the practical implementation of column generation (Dantzig-Wolfe decomposition) in order to impose global constraints that would otherwise have resulted in impractical runtimes.

Rohan Nanda, Airbus

Automatic adjustment of dynamic pricing under turbulent market conditions

Rapidly changing market conditions such as those seen during COVID19, represent a challenge for any forecasting system in revenue management. Such models require extensive transaction history to provide accurate results, while during the pandemic one faced unstable market conditions and data sparsity in periods of reduced operation. In the presented approach, instead of attempting new forecasts, pre-pandemic forecasts are used as a baseline. An additional module then monitors the sales performance in the last weeks and compares them against the ones expected from the baseline. Out of this comparison, the baseline is adjusted, and a new optimal dynamic pricing policy is devised. The effectiveness of this approach was assessed via automated A/B live testing, yielding  a 7% increase in revenue. This modular approach enables the enrichment of forecasts with additional and potentially unstructured data sources, increasing accuracy also under normal market conditions.

Giacomo Bonciolini, SWISS

Data-Driven Aircraft Assignment to Minimize Delay Propagation

Propagated delays due to late arriving aircraft contribute to 40% of all US flight delays as reported by the Bureau of Transportation Statistics (BTS). The objective of the aircraft assignment problem is to assign tail numbers on scheduled arriving flights at an airport to scheduled departing flights at the same airport with the objective of minimizing propagated delays. We propose a new data-driven approach for the aircraft assignment problem by formulating it as a balanced assignment problem. The assignment costs are estimated by using empirical observations of arrival delays from prior years' flight records. A clustering method is used to account for factors such as originating airport, time of day, and aircraft type that affect the arrival delay distribution. For an out of sample data set from 2018 for Delta Air Lines at Atlanta airport, we show that the data-driven policy yields a 18% improvement in total propagated delay, thus potentially saving approximately 6.5 million dollars in annual operating costs.

Vinayak Deshpande, University of North Carolina

Data-driven models for predictive aircraft maintenance and sustainable aviation

Modern aircraft are equipped with multiple sensors that continuously monitor the health of systems. The engine on a Boeing 737 generates 20 terabytes of condition monitoring data per flight hour. The entire world-wide fleet of aircraft generates approximately 2 million terabytes of condition monitoring data per year. The availability of such large volumes of data has incentivized the development of data-driven RUL prognostics and predictive maintenance planning. In this presentation we discuss a roadmap to go from sensor measurements to prognostics to predictive aircraft maintenance.

We also discuss how the lessons learned from predictive aircraft maintenance are of relevance for a more sustainable aviation. For example, data-driven predictive models can contribute to robust electric aircraft taxiing, with our aim to achieve the EU Green Deal target of aviation-neutrality by 2050.

Mihaela Mitici, Utrecht University

Deep Learning Based RM: A Quantitative Study on Behavior and Benefits of Deep Learning in Airline RM

Traditional RM methods are heavily dependent on demand forecasts based on historical patterns, this is a fragile assumption in real-world environments. Few airlines are now flying the same routes or frequencies, there are new entrants, and existing competitors are restructuring their operations. RM systems need to learn and adapt to such volatile market conditions automatically. Other industries have successfully adopted ML/DL techniques to navigate volatile environments. FLYR’s technology has been optimizing airline revenue for the past three years, using DL-based agents that predict optimal flight pricing policies and accurate LF&RASK forecasts. In this presentation, we will share some of the behaviors that our DL-based agents were able to learn by observing and managing millions of flights. We will also present some of the benefits these agents have been able to deliver in terms of pricing effectiveness, forecast accuracy, and revenue uplift.

Cole Wrightson, FLYR Labs

Deploying an Integrated Airline Disruption Management Solver at Lufthansa Group

Disruptions affecting aircraft rotations, passengers, and crew are part of the daily business in airline operations control. To assist operations controllers in their decisions, the Operations Decision Support Suite (OPSD) at Lufthansa Group and Google's Airline Disruption Management API are a central component to enable better decision making. To solve this problem, we adopt a framework in which the main operational decisions are decomposed into three subproblems, namely, rotations, passengers, and crew; and a coordinator problem selects a solution from each subproblem among potentially many feasible solutions. In this talk we focus on the engineering challenges of a large-scale recovery operations solver such as data preparation, distributed computing, reliability, and robustness. We present the impact of various features we have implemented in our decomposition framework on the quality of solutions found.

Daniel Bogado, Toby Davies, Daniel Duque, Google & Lufthansa

Direct and Distributed Demand Forecast for Revenue Management

Demand forecasting is a crucial step in an airline’s revenue management practice. One of the major factors contributing to the forecast error is historical schedule changes that disturb the signal and make it difficult mapping it into the future schedule. This was particularly evident during COVID-19 period when airlines dynamically changed schedules to react to demand shifts and regulations. We present a Distributed Demand Forecasting method that mitigates some of these pain points. It predicts demand on a market level and then distributes it to the individual alternatives using a modified MNL model. This approach is more stable, better represents natural traveler’s behavior and can be naturally extended into competitive-aware RM practice. We evaluate the performance of this approach on real airline data and compare the results with traditional Direct Forecasting used in practice.

Sergey Shebalov, Sabre

dispatching a fleet of electric towing vehicles for aircraft taxiing with conflict avoidance and efficient battery charging

This is one of the finalists in the Anna Valicek student paper competition.


Simon van Oosterom, Delft University 

Estimating Price Sensitivity via Machine Learning with Causal Inference

To improve pricing airlines have been experimenting with dynamically adjusting prices of their products based on unique product features and other relevant information available at the time of request. This requires estimates of price-sensitivity based on transactional data. Estimation of price-sensitivity parameters via demand response modeling lies in the realm of Causal Inference (and not prediction) and construction of estimators robust to confounders and model misspecifications remains a challenge. Modern machine learning (ML) despite its high predictive power, does not easily lend itself to constructing an interpretable framework for price elasticity estimation. In this talk we propose a hybrid framework that builds on the synergy between predictive power of modern Machine Learning (ML) approaches and interpretable semi-parametric models for robust price-sensitivity estimation. We show performance of our hybrid approach methods via simulation studies.

Dr. Darius Walczak, PROS

Fuel Prices have Decoupled from Airline Profits

The business press often claims that airlines’ profits are largely driven by the cost of fuel. We conducted a simple regression analysis to test this claim. We used FAA form 41 data for all substantial US carriers and separated the data into two periods: 1990-1999 and 2000 through 2019.  For the period before 2000  the correlation between fuel prices and profits was very strong, with periods of lower fuel prices being periods of higher profitability. However, since 2000 there was no significant linear correlation and the weak correlation that did exist implied higher profitability in periods of higher fuel prices. Fitting a quadratic equation to the data provided a better fit, but still only explained a small portion of the variability of the data.  All periods were analyzed using actual US dollars as reported by the FAA, and constant dollars to mitigate any potential impact of inflation.

William Farrell, Vaughn College

Inter Airline Slot Trade Opportunities Provider (ISTOP)

ISTOP is a mechanism that allows airlines to trade ATFM slots of a single regulation to mitigate their delay costs. It combines (a) a simple and practical collaborative decision-making framework, (b) no need for airlines to disclose their real costs, and (c) fairness since benefits are “equitably” distributed among airlines. In ISTOP the network manager (NM) plays the role of the airlines’ broker and finds slot swap offers which can be evaluated, accepted or refused by the parties. The NM computes the offers based on encrypted cost information and by solving an integer programming model that minimises an ad-hoc fairness metric guaranteeing the convenience of the offers and ensuring no penalisation to any airline. Our 200 data-driven random instances show that ISTOP provides an average delay cost reduction (w.r.t. the FCFS allocation) per instance of 12% for regulations affecting from 20 to 49 flights, 24% from 50 to 99 flights, and 31% from 100 to 150 flights.

Andrea Gasparin, University of Trieste

Lessons Learned from a live application of deep learning model based optimization on a low-cost carrier pricing systems

While developments in deep learning (DL) and reinforcement learning (RL) architectures enabled successful and profitable trading decision-making in volatile capital markets, until now, such technologies have not been widely adopted in traditional industries. The post-Covid volatility and uncertainties revealed the critical need for DL-based optimization solutions, resulting in continuous, dynamic real time pricing of both seats and ancillaries. Here, we present a system based on AI Native algo-trading technologies applied to live pricing optimization of a LCC, together with an analysis of the results observed. We describe the granular nature of such a system, the accuracy of the predictions and the nature of the optimization landscape generated. Finally, we discuss the business outcomes of such systems moving from the traditional, rule-based mechanisms to a goal-based engine, and how such systems can be integrated into an airline’s legacy infrastructure.

Dr. Uri Yerushalmi, Fetcherr

Line-training optimal scheduling using an innovative math-heuristic approach

The Crew Rostering Problem (CRP), consists in determining an optimal sequencing of a given set of crew pairings to create rosters, satisfying a series of operational constraints. This problem, considered to be NP-hard, has been largely studied by the operational research community.

The line-training rostering problem (LTRP) is an extension of the classical CRP where the objective is to determine an optimal sequencing of crew pairings to create simultaneously compatible rosters for both trainees and instructors, while taking into account specific pedagogical requirements for each trainee. The LTRP is also NP-Hard and its effective resolution using exact methods can be prohibitive in an industrial context.

This work presents the approach developed at Air France to handle the LTRP, allowing to create cost-effective line-training rosters in under an hour of computing time.

Solene Richard, Air France


Passenger-Centric Integrated Airline Schedule and Aircraft Recovery

Some passenger rights regulations impose monetary compensations under disruptions and may catalyze passengers’ response to recovered schedules. We embed passenger response into integrated schedule, aircraft and passenger recovery optimization, and develop a solution approach involving exact linearization of non-linear passenger costs, delayed constraint generation for aircraft maintenance feasibility and an acceleration technique penalizing deviations from planned schedules. Instances from two major European airlines with delayed flights, airport closures and unexpected grounding of aircraft show that our approach is tractable and scalable, with solutions superior to airline’s actual decisions, and highly robust in the face of passenger response uncertainty. Simulation results find that accounting for passengers’ disruption response behaviors – even in a highly approximate manner – yields significant benefits to the airline, than not accounting for them at all.

Vikrant Vaze, Dartmouth

Quantifying the Performance Gain of Using Shopping Data in Fare Pricing

Although shopping data is a good proxy for people’s intent to travel, making use of this data in practice is hard. We will explore use cases where shopping data’s demand forecasting potential can be utilized. We quantify the theoretical performance gain of using shopping data to augment demand forecasting. Our initial results show an RMSE reduction of up to ~40%. Although this gain is highly inflated due to the low denominator effect of using a test period during COVID, the pre-COVID test timeframe still yields ~20% improvement. Subsequent studies led to a surprising finding that simply aggregating across the granular dimensions can reduce forecast error significantly. When quantifying the forecast accuracy gain under the most granular dimension, results from two independent methods converge at ~8%. Although this is still likely inflated, it shows the value of using shopping data in demand forecasting, revenue management, and more generally, in airfare pricing.

Michael Wu (PROS); Angela Lombardi (Emirates)

Single-dimensional leg-level dynamic programming with booking-time dependent cancellation probabilities for revenue management

In this paper, an optimisation method is introduced that accounts for cancellations. We do so by estimating the opportunity cost of a booking between the time of booking and the expected time of cancellation, determine whether a product is available for sale, is adjusted by the risk the airline faces. We introduce an example which shows that there may be cases where it is optimal to reject a higher-priced product if the risk of cancellation is high, while accepting a lower-priced product. Through simulation studies, we show increases in revenues against a traditional dynamic programming formulation that does not explicitly models cancellations. Next, we show that the optimisation method is robust against choice of heuristic, misjudgment of cancellation probabilities and forecasting errors.

Daniel Hopman, VU Amsterdam

Smart Gauging (Demand driven re-fleeting)

Smart gauging or demand-driven re-fleeting is an approach that is used by airlines in response to demand uncertainty, operational disruptions, maintenance, and crew changes. The proposed model integrates the fleet assignment problem (FAP) and the aircraft maintenance routing problem (AMRP) with additional restrictions related to the crew pairings. The model needs a initial feasible plan and considers multiple operational restrictions related to aircraft maintenance stops, buffers and crew pairings and maximizes the total benefits that are calculated using the revenue forecast and airport and flight costs. The problem is formulated as a ILP model based on the arc and node graph where each node is defined as a task (flight, maintenance stops and buffers) and the arcs are the possible connections between tasks.

Rubén Jiménez Moreno, Vueling

Towards Rapid Prototyping of Artificial Intelligence Algorithms in Air Transportation

Advanced air mobility (AAM) has the potential to revolutionize travel by reducing ground traffic and emissions by leveraging new types of aircraft such as electric vertical take-off and landing (eVTOL) aircraft and new advanced artificial intelligence (AI) algorithms. Validation of AI algorithms require representative scenarios, as well as a fast time simulation to evaluate performance. Until now, there has been no such testbed available for AAM to enable a common research platform for individuals in government, industry, or academia. MIT Lincoln Laboratory has developed AAM-Gym to provide an ecosystem to develop, train, and validate AI algorithms across a wide variety of use-cases. In this work, we demonstrate the use of AAM-Gym to study the efficacy of AI applied to autonomous separation assurance in AAM corridors. We then provide recommendations for how such a testbed can be adapted to support AI research in traditional air transportation.

Marc Brittain, MIT

AIRLINE DYNAMIC OFFER CREATION USING A MARKOV CHAIN CHOICE MODEL

This is one of two finalist papers for the Anna Valicek student paper competition.

Kevin Wang, MIT

Two-way substitution through multiple flexible products

Because of the higher value passengers in business class, airlines do not overbook business class although the no-show rate in business class is higher, and rely on free or paid upgrades to salvage the empty seats in business class. Such a practice will cease to be revenue maximizing if the business class demand is greater than its capacity, as each upgraded passenger will displace a business passenger. We propose to sell intentional and informed downgradables (product that gives the airline the right to downgrade the ticket holder) that in-turn allows airlines to overbook business class and enhances much-needed flexibility during allocation. We use a Markov decision process to model the request arrivals and prove the existence of the nested optimal booking limits in threshold form. We identify and explain the instances wherein the downgradables is a benefactor and wherein it is not, thereby extracting novel insights for better understanding the downgradables.

Dhandabani Srinivasan, IITM

Powered by Wild Apricot Membership Software