| Flight Timetabling under Context-dependent Choice Model |
Keji Wei |
Tongji University |
Flight timetabling impacts an airline's operating profit. We consider a joint flight timetabling and revenue management problem with passengers choosing among multiple itineraries. The classic multinomial logit (MNL) choice model does not well depict the interaction between parallel itineraries. So, we introduce a context-dependent MNL model where the attraction value of one product is contingent on the offer set. The context-dependent MNL model extends common choice models and provides alternative explanations to several empirical observations. We formulate the joint flight timetabling and revenue management problem as a mixed integer programming model. We present a two-stage decomposition framework to compute the network timetable within a practical time budget. Numerical results suggest that the combination of the formulation and the solution approach leads to profit improvement for airlines. |
| Optimizing Cost and Enhancing Consistency in GenAI Chatbots with Semantic Caching |
Mourad Boudia |
Amadeus |
Chatbots are widely used in the travel industry to answer questions. Generative AI (GenAI) advancements offer significant potential to enhance these chatbots with less human involvement. However, current large language models (LLMs) face high costs, inaccuracies and inconsistencies.
This work introduces a novel caching system to reduce costs and improve response consistency within the Retrieval Augmented Generation (RAG) context. We discuss the challenges of building such systems and our strategies, including word embedding control, models fine-tuning and models combination, and benchmarking various pre-trained LLMs. Numerical results from simulated real-life FAQs demonstrate our approach's effectiveness. |
| Enhancing airport mobility through machine learning solutions |
Rim Jabri |
Groupe ADP |
This abstract explores the application of machine learning techniques to optimize the shuttle frequency of Paris-Charles-de-Gaulle airport's Automated People Mover (APM), known as the CDGVAL. Through the use of supervised learning algorithms, a model was created to predict passenger demand at each station. Key features of this model are the time of day, scheduled flights and passenger volumes.
The predictions from the model facilitated the creation of better optimized shuttle schedules along with dynamic adjustments to the frequency of shuttles. The shuttle schedules derived from the model's predictions improved service efficiency while also reducing energy consumption, highlighting the effectiveness of machine learning based optimisation in transportation.
|
| Transforming Measurement with LATAM’s Advanced Experimentation Platform |
Felipe Bahamonde (LATAM),
Joël Gastelaars (acmetric),
Sebastian Orellana (LATAM) |
LATAM Airlines; ACmetric |
Discover the ongoing journey of building LATAM's advanced experimentation platform, focused on detecting causal relationships to accurately estimate the impact of different treatment policies and projects. This platform integrates cutting-edge causal inference and experimentation methods, empowering teams across departments to adopt a scientific, data-driven approach to decision-making. While we’ve made significant progress, embedding a culture of experimentation within LATAM is still in progress. In this session, we’ll explore the challenges, learnings, and successes in fostering this cultural shift, with a key emphasis on understanding how different interventions drive outcomes. Join us as we discuss how this evolving platform is transforming resource allocation, driving continuous improvement, and positioning LATAM to lead in both physical and digital experimentation. |
| Reinforcement Learning for Advance Seat Reservation Pricing: Insights from Real-World Deployment |
David Lopez |
FLYR |
A-la-carte advance seat reservation has become a vital revenue stream as airlines increasingly unbundle fares. In our previous work, we introduced a reinforcement learning (RL) approach to dynamically price these seat reservations. Since then we have observed more than 10% increase in seat revenue per passenger with the first implementation and our model has evolved, informed by real-world deployment data. To address the challenge posed by zero-inflated training datasets, we implemented a two-step regression technique. In response to real-world feedback, we developed metrics that monitor and communicate the model's behavior, providing transparency into the pricing model’s state, particularly its balance between exploration and exploitation. |
| Exploring the Endogeneity of Seats and Fares in Airport Origin-Destination Demand Forecasting: Insights from a Large Airline Network |
Ahmed Abdelghany |
|
This paper explores the role of seat capacity and fares in forecasting airport origin-destination (O-D) demand across a large airline network. By investigating whether these variables should be modeled as endogenous (influencing demand directly) or exogenous (responding to external factors), the study assesses their impact on demand forecasting accuracy. Using a generalized model that balances specificity and scalability, the research aims to improve forecasting across diverse O-D pairs while accommodating key demand drivers. Various time series modeling approaches are compared to determine the most effective modeling structure. The findings offer insights into developing a scalable, accurate forecasting platform for airline network planning and decision-making, together with exploring the impact of seat capacity and fare on O-D demand. |
| Enhancing Flight Cancellation Decisions with AI-Powered Passenger Rebooking Estimates |
Claudia Bongiovanni, Georgia Lazaridou |
SWISS Intl. Air Lines |
In the event of disruptions like aircraft breakdowns, our network operations control (NOC) experts must swiftly decide which flights to cancel from a set of options. However, they lack detailed data on passengers’ origins and destinations, complicating the assessment of rebooking costs, which are a major post-cancellation expense. This evaluation is typically handled later by the passenger control center using tools like Amadeus OPR and with manual rework to meet passengers’ preferences and needs.
This project aims to equip NOC experts with quick estimates of passenger rebooking costs to support better flight cancellation decisions. We develop a combinatorial optimization (CO) model that rebooks passengers on alternative routes and compartments, accounting for seat availability. The CO model is enhanced by a machine learning layer that learns from historical data on flown itineraries and is then able to infer passengers’ rules, preferences, and needs. |
| Integrating Macroeconomic Forecasts into Airline Demand and Cost Estimation |
Soheil Sibdari |
University of Massachusetts Dartmouth |
This research explores the integration of key macroeconomic indicators, such as GDP growth, exchange rate, energy cost, consumer spending index, and the middle-class growth, into demand and cost estimation models for airlines. By linking industry-specific factors to broader economic trends, airlines can better anticipate fluctuations in passenger demand, fuel expenses, and other operational costs. This integrated approach enhances decision-making in areas such as pricing, route planning, and capacity management. This research introduces a mathematical model to illustrate the results and uses empirical analysis to demonstrates how incorporating macroeconomic forecasts into airline forecasting models improves predictive accuracy, enabling airlines to adapt swiftly to economic changes and market volatility. |
| AI-driven pricing with RSP: Evidence from a large-scale online field experiment |
Giacomo Bonciolini (LHG),
Darius Walczak (PROS) |
Lufthansa Group, PROS Inc. |
Estimating price elasticity is crucial for implementing effective dynamic pricing systems, yet it remains a complex task due to challenges such as data sparsity and price endogeneity. This talk presents an innovative modeling paradigm, RSP (Request Specific Pricing), which integrates modularly in the existing Revenue Management System. RSP employs a novel two-stage process combining causal inference and advanced machine learning techniques, such as Deep Neural Networks, to produce robust price elasticity estimates. The effectiveness of this method was tested through a comprehensive automated A/B live field experiment, demonstrating its substantial impact on increasing airline revenues through data-driven price optimization. Moreover, RSP paves the way for the calculation of tailored offers, by supporting the inclusion of more booking attributes in the optimal price determination. |
| Classless Revenue Management |
Sergey Shebalov |
Sabre |
Transition from class availability to optimal pricing is one of the major trends in the airline industry. It provides an opportunity to increase flexibility of the revenue management practice and improve revenue performance. It also allows to streamline the architecture of the revenue management system and leverage more intuitive algorithmic approach. We present several models we developed to implement the classless RM concept, discuss practical challenges we had to overcome and share some early results we obtained on the real-world data instances. |
| Predicting a Schedule's OTP and How that Impacts Schedule Build |
Pascale Batchoun |
Air Canada |
The first step toward establishing a resilient flight schedule is to accurately predict its On-Time-Performance (OTP) and to evaluate the impact that both scheduling and operational changes could have on OTP. Air Canada has developed and deployed to production a machine learning-driven methodology that predicts system-level OTP key performance indicators (KPIs) while providing low level estimates for the block and turn durations at the flight level. Augmented with a simulation engine, the “OTP Schedule Optimizer” evaluates various scenarios to simulate disruptive and cascading delays, and their impact on our aircraft performance, passenger connections and crew flow. The system uses an optimization engine that leverages various outputs from the machine learning and simulation models, to recommend schedule changes to improve its OTP while minimizing the impact on passengers’ misconnections. We will shed the light on how we designed the system and share insights on performance KPIs, including simulation of operational changes, and what are some of the pain points highlighted in the schedule, and what actions to take in the Planning, Scheduling and operational windows. |
| Unlocking Revenue Potential: The Impact of Modern RMS on Airlines |
Balaji Davangave |
Air India |
Revenue Management Systems (RMS) are crucial for airlines to optimize revenue, but evaluating the impact of new RMS compared to older systems is complex due to multiple influencing factors. This study introduces a robust framework for assessing the benefits of advanced RMS, using time series forecasting to analyze Revenue per Available Seat-Kilometer (RASK). By isolating the effect of RMS on performance, we provide insights into their potential to enhance operational efficiency and profitability. This methodology offers a scientific approach to understanding the true impact of RMS changes in the airline industry. |
| Efficient Warm Start Strategies for the Aircraft Routing Problem |
Samhita Vadrevu |
United Airlines |
The Aircraft Routing Problem (ARP) is a complex optimization problem that constructs feasible routes for aircraft while adhering to maintenance rules. However, with growing fleet sizes, the time required for computing feasible solutions increases significantly.
In this work, we present an innovative warm start method that serves to expedite the optimization process. Utilizing agent-based modeling, the technique constructs initial routes that are introduced to the model as a warm start. These routes are designed to promote maintenance adherence, both primary and secondary. By providing the model with these strategically devised initial routes, the proposed method helps in a quicker progression to the final solution. The application of this warm start strategy has led to a remarkable 50% decrease in runtime, proving its effectiveness in speeding up the ARP. |
| Unlocking revenue with customer search behaviour modelling |
Himanshu Jindal |
Air India |
The airline industry relies heavily on data. Using direct channels and GDS’ search data can unlock significant revenue opportunities. By analysing traveller search patterns, airlines gain insights into customer preferences and travel intentions. This predictive capability helps optimize marketing and sales campaigns, ensuring targeted and timed promotions and enhancing ROI. Integrating PoS search data allows sophisticated algorithms to align availability and inventory optimization. Understanding high-demand routes enables dynamic adjustments in inventory and pricing strategies, maximizing revenue and customer satisfaction. This data-driven approach streamlines operations and provides insights into network planning.
In summary, using search data from websites and GDS enhances operational efficiency, boosts marketing effectiveness, ancillary upsells and optimizes inventory management, leading to a more responsive and profitable business model. |
| Applications of Customer Lifetime Value (CLV) In Tactical Decision-making |
John Moore,
Alex Cosmas |
McKinsey & Co. |
Customer Lifetime Value (CLV), traditionally a tool for marketers, is proving to be a highly effective tool for tactical decision making. This study introduces a CLV model to assess the impact of operational mishaps — bag issues, delays, and cancellations — on CLV at a full-service airline.
The model measures customer behavior via propensity to book future trips (across 10 years of data, 15 tables, and 75 customer-level and flight-level features). Such CLV estimates can be used to assess the future cost of today’s operational disruptions (and could also measure uplift associated with positive service interventions).
We found operational mishaps have a statistically significant impact on customer booking patterns and overall CLV, highlighting a potential way for airlines to prioritize operational issues at the flight and route level (and create clear business cases to resolve them), rather than resorting to blunt schedule changes. |
| AI enabled innovations in workforce excellence |
Taylor Cornwall |
McKinsey & Co. |
In a cost-focused environment, the largest controllable cost for most airlines is the frontline workers. COVID recovery, employee turnover, and new contracts have put an increased imperative on workforce planning efficiency. Achieving productivity requires a coordinated approach to forecast demand under uncertainty, optimize resource planning, manage unpredictable daily operations, and drive individual performance. Fortunately, AI offers a wide range of tools to manage this complexity – applicable to many workforce planning domains in airlines. As an example, with uncertain demand for reserve pilots, creating an optimal crew schedule has posed a complex challenge for airlines. Conventional approaches involve using point estimate forecasts that do not account for variability in demand. To improve this, we followed a 3-pronged approach to lower pilot shortage and overage costs: Demand Forecast (incl. variability), Optimization Under Uncertainty, and Cost /Robustness Simulation. These technical capabilities must also be complemented by large-scale change management. This session will talk through both the technical approaches and the adoption process to drive workplace efficiency. |
| Digging into On Time Performance with Generative AI |
Kevin Hightower |
Cirium |
On Time Performance is critical for airlines for passenger satisfaction, managing costs (shorter flight times / taxi times means less fuel, less crew costs, etc.), reducing emissions, and even helps aircraft meet their maintenance slots.
We'll dig into some examples by hand, and then do the same with Generative AI (GAI), showing that with proper methods, GAI can assist professionals to drive faster and better decisions in Aviation. |
| Condition-Based Maintenance scheduling of an aircraft fleet under partial observability: A Deep Reinforcement Learning approach |
Iordanis Tseremoglou (Anna Valicek Finalist) |
Delft University of Technology |
In the Condition-Based Maintenance (CBM) context, the definition of optimal maintenance plans for an aircraft fleet depends on an efficient integration of : (i) the probabilistic predictions of the health condition of the components and (ii) the stochastic arrival of the corrective maintenance tasks, together with consideration of the preventive maintenance tasks as defined in the Maintenance Planning Document (MPD) . To this end, in this paper, we present a two-stage dynamic scheduling framework to solve the aircraft fleet maintenance scheduling problem under a CBM strategy in a disruptive environment. In the first stage of the framework, we address the uncertainty in the predicted health state of the monitored components by planning the optimal maintenance policy based upon the belief state-space of the health of the components. The decision making process is formulated as a Partially Observable Markov Decision Process (POMDP) and is solved using the Partially Observable Monte Carlo Planning (POMCP) algorithm, considering the aircraft maintenance scheduling problem requirements. In the second stage, a Deep Q-Network (DQN) is developed, that integrates the defined maintenance policy of the monitored components within the scheduling of the aircraft fleet’s preventive and corrective maintenance tasks. Our model, through a rolling horizon approach, continuously creates and adjusts the maintenance schedule, reacting to new updated task information, where the availability of maintenance resources constraints the execution of each task. The proposed framework was tested on a case study from a large airline and the performance was evaluated against the current state practice of the airline. The results show that our model can schedule 96.4% of monitored components on-time. As a consequence of this, a 46.2% maintenance cost reduction is achieved for the considered monitored components relative to a corrective maintenance approach.
|
| Aircraft maintenance scheduling under uncertain task processing time |
Matıas Villafranca (Anna Valicek Finalist) |
Pontificia Universidad Católica de Chile |
Unexpected delays in executing aircraft maintenance tasks can result in costly operational
disruptions for airlines, including flight delays, which can significantly impact their operations and
expenses. In this study, we deal with uncertainty in maintenance task processing times and design
cost-efficient aircraft maintenance schedules via two-stage stochastic programming. In the first
stage, we determine the daily tasks to execute in each aircraft, specifying their start time and
assigned maintenance base and technician. In the second stage, the start time of each task and
the departure time of each flight are adjusted based on realized task processing times. We aim to
minimize the expected costs incurred for outsourced maintenance tasks, technician overtime, and
flight delays. To solve our model, we design an ad-hoc Adaptive Iterated Local Search heuristic
that explores first-stage solutions via an efficient evaluation of the second-stage cost. We also test
our approach in a set of computationally simulated instances. Our proposed model and solution
yield 77% and 45% average cost savings compared to a deterministic approach assuming expected
task processing times and to a conservative solution assuming maximum possible processing times, respectively. Moreover, we obtain 12% average cost savings compared to a benchmark solution, which plans maintenance tasks with an optimized time buffer. Furthermore, we study the cost impact of varying structural parameters, such as task granularity, processing time variability,
workload, and cost structure.
|
| A New Era of ML-based Aircraft Performance Modeling |
Shinji Tanaka |
NABLA Mobility Inc. |
In order to optimize flight operations beyond the current plateau, it is a well-known consensus that airlines need more detailed and accurate predictions. However, the necessity for a highly granular aircraft performance model is often overlooked in real-world operations. We have conducted two studies regarding the aircraft performance models: (1) Assessing the accuracy of existing models, and (2) Exploring enhancements using ML technologies. In summary, our findings reveal that the aircraft performance models currently used in flight planning software has an error rate of 14% when estimating fuel burn, which can be reduced to 3% by employing our ML technologies. Having developed these technologies through extensive use of QAR data, this also yields significant opportunities to improve efficiency as well. By integrating it into the flight optimization process, airlines can unlock additional optimization opportunities.
|
| Elevating cities: The rise of regional and urban air mobility |
Laurie Garrow |
Georgia Tech |
This presentation covers recent advancements in the development of electric and hybrid-electric aircraft for advanced air mobility (AAM) applications, presents resources that track the development of these aircraft, and discusses major research and policy challenges associated with introducing these novel aircraft types into the air transportation system. The presentation ends with an update of ATL@GT Center activities and describes how a new competitive analysis tool, PassengerSim, can be used to evaluate the profitability of introducing these new aircraft into markets and determing if aircraft can remain profitable if competitors enter the market.
|