FLYR was selected to present at the 2021 AGIFORS Annual Symposium to share our perspective of how new technology can be practically applied to continue to advance the practice of Revenue Management. In this virtual presentation, learn about the value of using novel Revenue Management (RM) strategies with Reinforcement Learning (RL) to successfully evaluate model performance prior to implementation and without risking revenue impact.

Speakers: Jon Ham, Head of Analytics; Cameron Yousng, Data Scientist; Austin Rochon, Data Scientist

Date Presented: September 20, 2021

Duration: 20:49

00:00 – Introduction:
“Know Your Worth: Valuing Novel RM Strategies with Reinforcement Learning,” is a virtual presentation by Jon Ham, Cameron Young, and Austin Rochon of FLYR Labs, the pioneer of The Revenue Operating System^® for airlines and transportation leaders.

00:30 – About Cirrus:
FLYR’s Cirrus Revenue Operating System includes four critical components: a single repository and unified data model to integrate all airline data, ultra-accurate revenue and load factor forecasts utilizing advanced deep learning AI technology, continuous class-agnostic pricing algorithms, and end-to-end reporting, analytics, and controls for airline decision-makers.

01:20 – Motivation:
Attempting to deploy changes to the Revenue Management System (RMS) or implement new Revenue Management (RM) strategies can put significant revenue at risk. Forecast accuracy and simulation approaches can pose challenges as well. At FLYR we apply a rigorous model validation process with a suite of tools, data, and metrics dedicated to evaluating model quality. We also utilized Reinforcement Learning (RL) techniques to estimate the value of novel RM strategies in order to deploy the best-performing models without risking real-world revenue.

03:02 – Reinforcement Learning (RL) Background:
RL is a subfield of machine learning, allowing models to make the optimal decision in dynamic environments. When applied to revenue management, various data points such as origin, destination, seat availability, departure dates, and days before departure are used to make real-time, revenue-optimized selling decisions.

12:16 – Approach:
In order to validate this approach, FLYR’s team of data scientists train two separate Off-Policy Evaluation (OPE) networks; one to learn the value of the historical RMS control policy, and another to learn the novel FLYR pricing model. The results are then compared against a baseline to confirm the OPE results.

16:43 – Results:
The FLYR team performed thorough testing of both OPE networks. The results from the validation period confirmed that FLYR’s model outperformed the Legacy RM controls.

18:51 – Discussion:
If a policy generates actions and states that don’t include historical data points, they can result in an overestimation of Q values. However, despite the risks of overestimation, these techniques can still help to determine which RM strategy will outperform the other. At FLYR, the team is working to further mitigate the overestimation of Q using batch-constrained learning methods to make the actions more similar to historical occurrences.

About AGIFORS:
The Airline Group of the International Federation of Operational Research Societies (AGIFORS) is a professional society within the airline industry dedicated to the advancement and application of Operational Research. For more information about AGIFORS, visit https://agifors.org.