AGIFORS 59th Annual Symposium Program


Agenda (tentative)

AGIFORS 59th Annual Symposium Program


Technical program

A Reinforced-Learning Approach for Calibrating Airlines Itinerary Choice Models with Constrained Demand

Ahmed Abdelghany - Embry-Riddle Aeronautical University


We present a reinforced-learning approach to calibrate airlines itinerary choice models, while considering constrained demand. A simulation-optimization search algorithm is developed to find the optimal parameters of the itinerary choice models. The algorithm minimizes error in market shares and load factor of flights. The calibration is performed at the global network with more than 500,000 city-pairs. Thus, the demand-supply interaction is considered in the different markets.



The role of AI in the passenger journey

Rodrigo Acuna-Agost - Amadeus


In this presentation we describe how AI is transforming the travel industry by impacting different steps of the passenger journey. We will discuss the emerging applications and challenges of AI on travel with examples of different uses cases explored by our AI research team.



Modeling flight disruptions: Delays and Cancellations

Javier Avella-Gonzalez - Data Scientist - Air Canada


Finding optimal ways to manage flight disruptions, such as delays or cancellations, is a major concern for airlines and aeronautic authorities worldwide. Here, we focus on exploring, identifying, and understanding existing correlations between weather related conditions and the occurrence of disruptions, to further build a predictive model for flight delays or cancellations. Trained on typical airline data (provided by IATA - STEADES) with information such as Wind direction/speed, temperature, pressure, visibility, etc., the model prediction has an accuracy close to 90% in the cross-validation set, and we currently work on including other features that help improve its power. With this presentation we aim to leverage ML models as powerful AI tools capable to solve complex problems in the industry.



I know what you loaded last summer

Felix Brandt - FZI Research Center for Information Technology


Aircraft weight is a major determining factor of flight characteristics. In this talk we analyse if the reverse is also true, i.e., to which extend it is possible to derive the payload of cargo aircraft from externally observable flight characteristics. From a business point of view, gathering payload information of aircraft of other carriers is highly valuable for demand forecasting and future network planning. Therefore, we analyse the relationship between payload and a wide variety of flight attributes, which can be derived from publicly available tracking data. Based on our findings, we introduce a supervised machine learning approach to estimate the aircraft payload. To train our approach we use a set of all-cargo flights with given payload (provided by our airline partner), weather data, and ADS-B tracking messages of the respective flights. Finally, we present the results of our experiments that allow us to estimate payloads with good accuracy.



How much to invest in block times? – An Interactive Discussion

Cristián Carrizo - LATAM Airlines Gruop


Most of the time, when someone attends a presentation about Block Times will probably be expecting to develop two subjects about them: 1) How important are for OTP and 2) How to save fuel with it. But the interactive and open discussion we want to propose is much more than that: Is about a whole new vision of opportunities the block times could help the industry when being used in a non-conventional way. We hope this can be a chance for the airlines, all together in a great benchmark, to prove there is a world of business cases against what always have told us no to do. “Dare to challenge” is the spirit of this presentation.



Reinforcement Learning Approach for Customer Choice-based Network Revenue Management Problem

Neda Etebarialamdari - MILA (Montreal Institute for Learning Algorithms)


One of the critical issues in transportation industries is to efficiently manage the available inventories in a customer choice-based environment in order to maximize the total revenue of the company and satisfy arriving demands. In this research, we tackle this problem by taking the advantages of Artificial Intelligence techniques in order to develop a reinforcement learning-based method to improve the performance of existing inventory control methods. More specifically, we address this issue using a customized Deep Q-learning model which can be applied to both leg-based and network-based airline revenue management problems. Our primary results indicate a fast convergence to solutions close to upper-bound values obtained by exact approaches.



Medium Haul Fleet Segmentation - Airbus A320neo & A321neo Weight Variant Selection

João André Ferreira Abreu - TAP Air Portugal


The present paper reveals a novel aircraft Weight Variant (WV) selection method. The implemented method makes use different aircraft Zero-Fuel Weight and Trip Fuel values to create polynomial relations. After polynomial relations are computed for every considered route and time periods, quadratic coefficients as well as WV specifications become inputs of the designed model. The model reveals the total payload available and if any, operational limitations. Afterwards, available maximum pax number and cargo is computed for every WV. The method aims to mitigate an airline direct costs through the navigation and landing fees reduction as well as the aircraft acquisition value. Therefore, this method is capable of reducing airline's direct costs without compromising the operational feasibility, and thus, assuring a competitive market position.



Towards a Competitor-Aware RMS (CARMS)

Thomas Fiig - Amadeus


State-of-the-art revenue management systems (RMS) use customer choice models to optimize revenue in semi-restricted and fare-family fare structures. These choice models rely on estimates of customer willingness to pay (WTP). Current WTP estimation does not explicitly consider competitor prices. Instead, competition is indirectly accounted for by its effect on the airline’s booking activity. However, this approach can lead to biased estimates of WTP. While this issue has long been known, it has never been systematically corrected. Rather, it is managed by users manually adjusting their selling fares. In this talk, we describe how it is possible to design a competitor-aware RMS. Such a system requires forecasts of competitor pricing behavior at a granular level. We will discuss how this can be achieved in practice based on airfare analytics. Finally, we demonstrate how CARMS can increase revenue over traditional choice-based RMS.



Stochastic Tankering

Semi Gabteni - OpenAirlines


While Fuel efficiency programs are gradually spreading across the industry, the variety of Fuel saving best practices is not well known and the opportunities for improvement through ""analytical methods"" are many. In this talk, we will focus on tankering, which basic concept is to take advantage of fuel price variations across an airline network to buy fuel where it is cheapest. Obviously, the cost of carrying fuel, or cost of weight, is a key element, which airlines consider for optimal tankering strategies. While current practice is determinisitic with average fuel consumption on routes and seasons, we have evaluated the benefit of a stochastic approach to account for fuel consumption variations depending on operating conditions. Our approach relies on Monte-Carlo simulation and is tested on real data.



How Copa Airlines applies Machine Learning to increase Operational Efficiency: Visualizations, Block Times, Wind Components, Overflights and beyond

Miguel Gaitan - Copa Airlines


In increasingly complex operations with ever changing situations there is always a huge backlog of analytics to be done. With the help of programming and specifically some machine learning magic we can automate and increase accuracy of the complex analyses that run operations in the backstage of an airline. Mainly focusing on K Means Clustering we calculate the block times for our current flights, reducing the total amount of block hours through our network but increasing block performance by intelligently taking out or putting time without padding. In a second part of the presentation we can focus on how different simple analytics allow us to present Fuel Efficiency Programs, Delay and On Time Performance opportunities and how this approach helped Copa Airliens be the most punctual Airline in the world in 2018.



Quantum computing for airline planning at Jeppesen

Mattias Grönkvist - Jeppesen


Quantum computing is a potential disruptive technology for high performance computing such as large-scale airline planning. Jeppesen recently entered as an industrial partner in the Wallenberg Center for Quantum Technology (WACQT). The WAQCT project involves multiple Swedish universities and industry partners and aims to build a 100 qubit general purpose quantum computer within 10 years, along with practical industrial and academic use cases. We will present Jeppesen's involvement in the project, why we think quantum computing is important and what we are currently doing.



Weight and balance extensions to minimize ground handling times and efforts

Anna Hess - FZI Research Center for Information Technology


Weight and balance planning of cargo aircraft is subject to a significant number of planning criteria concerning safety and cost efficiency of load plans, which can be solved by OR-based weight and balance models. We present solutions for planning requirements relevant in practice that received little attention in scientific literature so far. In this talk, we present two new model extensions. First, we investigate the positioning of priority ULDs to minimize their unloading times. Second, we introduce an efficient alternative to temporarily unloading ULDs at intermediate stops for rearrangement. We shift ULDs inside the aircraft while minimizing the total number of movements and loading operations. We evaluate our approach with real-world test instances and show that both extensions can be integrated into existing weight and balance models. Further, we present practical applications on how to benefit from the new features.



Data-driven strategies for optimizing revenue from both air travel and holiday packages

Aime Kamgaingkuiteing - Air Transat


Many studied have been performed to deal with flight revenue optimization. Despite the presence of business units which manage holiday package in airline companies, few studies have been performed on optimizing conjointly air travel and holiday packages. The old fashioned for airline to manage holiday package by splitting flight inventory into two sets: one set reserved for air travel and another set reserved for holiday packages could result in missed revenue opportunities. This presentation defines the holiday package optimization problem, then introduces a model which combines with data-driven strategies conjointly optimize air travel and holiday packages.



A New Airline Seat Reassignment Module

Xiaodong Luo - Sabre


Schedule Change or Equipment Swap triggers re-accommodations of customers into a new flight or aircraft, and the seat reassignments is a key step in this process. The current seat reassignment process is inefficient and often results in customer dissatisfaction as well as revenue leakage. The underlying business problem is highly nonlinear and combinatorial in nature. Modeling and solving the problem exactly is not an option as the solution time could be prohibitive (several minutes per flight). In this talk, we present an ""optimize"" and ""fix"" approach for seat reassignment. We embed a highly efficient Hungarian Algorithm into the solution process and combine it with pre-processing and post-processing heuristics to generate quickly near optimal, legal seat assignments. We will present some preliminary computational results based on real data. This is a joint work between Sabre and AA.



Anomaly detection and machine learning for airplane maintenance

Wannes Meert - KU Leuven


Over the past few years KU Leuven, TUI fly and Boeing have looked at applications of machine learning and anomaly detection for airplane health management. Since airplanes collect a plethora of flight parameters there is ample opportunity to apply advanced data analytics methods. There are, however, a number of challenges which make that a naive machine learning approach is not adequate. For instance, the majority of the data is unlabeled, maintenance actions can change patterns present in the data, and each airplane might show unique behavior. In this talk we will detail the challenges we encountered and results we obtained in a number of use cases focussing on the cabin pressure controller, the auxiliary power unit, and the airplane spoilers.



Multi-Agent Systemic approach to support Dynamic Airline Operations based on Cloud Computing.

Frank Morales - The Boeing Co.


Air transportation is a system of systems, with many interdependent, manageable but challenging sub-problems involving complex subsystems such as airlines, airports, and airspace. Even with the optimized planning, operational irregularities happen. They propagate quickly across the entire network, and a flight change may create an avalanche effect by disrupting many aircraft, crews, and passengers. With the current technological advances and industry focus on NextGen and SESAR, we are able now to resolve this problem. We present a cloud environment based on collaborative intelligent agents combining data, optimization services, and analytics solutions through an open architecture from a holistic perspective. With our presentation, the attendees will learn about how this solution can bring values to Airline Operation Contol team using interoperability solutions to solve their challenges and improve their productivity and efficiency in their operations.



The die is cast

Tim Nickel - Lufthansa Systems


As Julius Cesar knew he would benefit from crossing into unknown territories when he called out his famous quote, as well as he knew that the consequences of his fate were inevitable, airlines must understand the impact of the upcoming deadlines of SESAR ATM solution European Regulation 716/2014 (the pilot common project). They must also understand the chances and benefits of utilizing previously ""unknown territory (Airspace). European Regulation 716/2014 deadlines are approaching. The SESAR functionalities that are being deployed place additional requirements on the OCCs in terms of processes, operational decision-making and workflows and structures. Those who won't be capable of properly interacting with the other stakeholders could undergo an inefficient access to ATM resources and jeopardize the system-wide benefits expected by the upcoming ATM Paradigm.



Comparing the performance of AutoML against custom models for automated incident report classification

Reuben Pereira - Air Canada


Aggregating and classifying safety incident reports is a largely manual and time-consuming process. In addition, the manual nature of the process can lead to classification errors and inconsistencies. The goal of this study was to leverage machine learning approaches in natural language processing to develop an accurate, automated and cost-effective incident report classification solution. For this study, we compared the performance of a SOTA deep learning based NLP model against an off-the-shelf (OTS) automated “AutoML” text classification solution. When assessed on a dataset of 20K de-identified airline incident reports (GADM) corresponding to 20 incident categories, the OTS solution performed comparably to the custom SOTA model (both had F1 score of approximately .80 and accuracies over 90%) but was more feasible and cost-effective to implement. This proposed solution was selected as the winner at the 2019 IATA ADS Datathon.



Multi-objective schedule evaluation

Sergey Shebalov - Sabre


Airline operations planning process is typically decomposed into multiple stages and involves optimization of several key resources, such as network, aircraft fleet, crew, airport equipment, etc. Airlines also need to continuously adjust these plans to account for both external and internal factors. Technology can provide significant gains in both efficiency and optimality of this process if systems used by different planning departments are interconnected and synchronized. We describe this idea on the example of multi-objective schedule evaluation that is performed automatically without human intervention. We also discuss opportunities this approach creates for modeling airline planning problem as a multi-agent system and leveraging recent advances in AI-based algorithms for solving it.



Applying Reinforcement Learning within the Flight OPS Environment

Manuel van Esch - zeroG


Nowadays, flight ops controllers are faced with the task of stabilising a flight schedule that is regularly disrupted by micro-changes. They base their decisions on information gathered from an ever growing number of sources, yet not receiving any insights on how their decisions turned out by the end of the day. Creating an actionable understanding of cause and effect requires technological support. Reinforcement Learning (RL), an AI technology that received lately major research breakthroughs, is well suited to learn cause and effect from data. Using this technology, we are currently investigating how we can build a RL assistant that provides concrete suggestions for stabilising the schedule. In our talk we will report from our progress achieved in a research cooperation with a major European airline and present insights and challenges encountered.



The 3 Success Criteria to Use Optimization in Highly Human-dependent Environment

Valentin Weber - Amadeus IT PAcific


Human is a key factor in the decision-making process inside the Operations Control Center. The operators are usually very experienced having held various positions inside the company that gave them a good understanding on how to react in case of disruption. In addition, the interactions between the different departments are still mostly done by humans through phone calls. In such highly human-dependent environment, it can be challenging to bring optimization in. The scope of the optimization can be quite difficult to capture as the decision-making process can vary a lot depending on the situation and the operator. The result of the optimization needs to be understandable by a human that can validate it and coordinate its application with other departments. We will present how, through several experiences, we have matured our methodology to develop practical optimization feature used to solve real disruption.



Individualized Dynamic Pricing of Airline Ancillaries

Kartik Yellepeddi - deepair


For next generation travel suppliers, ancillaries are not only a growing source of revenue but also a mission critical tool to compete and meet evolving customer expectations through personalization. Conventional pricing strategies for ancillaries are based on poorly optimized business rules that do not respond to changing market conditions or trip context. We present an individualized dynamic pricing approach. Embedded in it are three models for dynamic pricing of ancillaries, with increasing levels of sophistication. In an outer loop, we introduce an adaptive model-selection framework that intelligently routes live customer requests to the above models. Offline experiments show that deep learning algorithms outperform traditional machine learning techniques for this problem. In online testing, on a live airline’s website, our AI-driven pricing outperforms human rule-based approaches, improving conversion by 17% and revenue per offer by 25%.



A novel load factor progression model based on OTA search and transaction data

Siyuan Zhao - Ctrip


Load factor measures a flight’s capacity utilization. For a route with multiple flights, load factor is more relevant and has less fluctuation. In this presentation, we proposed a novel machine learning approach to predict load factor progression at route level based on search and transaction data from Ctrip, China’s leading Online Travel Agency (OTA). Huge volumes of visitor-initiated inquiries are received on OTA’s platform, which with its distinctive ability to tag searches to visitors, provide insights on market demand. Features such as route popularity are derived from search data. Historical load factor and real-time sales are available from transaction records. These information are then put together in the model which predicts the load factor at each future date up until departure. We found that without using price as input, our model achieves high accuracy with an average mean absolute error of 2%~6% depending on future dates to departure.

Contact us

Address: AGIFORS President, 931 Monroe Drive,  Ste 102 #613, Atlanta, GA 30308

Phone: +1 (404) 385-6634

Email: president@agifors.org

Powered by Wild Apricot Membership Software