Applying Reinforcement Learning to dynamic pricing for smarter airline pricing strategies

At its core, Reinforcement Learning (RL) is about learning through experience. An RL agent navigates complex environments step-by-step, refining its actions based on reward signals. Through iterative trial and error, it combs through different decision pathways to identify the most effective strategy to maximize overall rewards. RL has been successfully applied in autonomous vehicles, robots, and financial trading. When applied to dynamic pricing, RL helps airlines optimize pricing for more profitable revenue management.

Traditional revenue management relies on applying inventory management techniques to open and close 24-fare classes (RBDs) based on predefined rules covering availability, demand, and other factors like seasonality, day of week, competition, etc. However, rigid fare rules and the price jumps between the 24 RBDs fail to capture a customer´s actual willingness to pay (WTP) and real-time market changes, leaving incremental revenue off the table.

Advancements in hardware and computation have enabled airlines to use advanced methods like Bayesian forecasting and dynamic programming to derive optimum demand forecasts and booking limits. However, Artificial Intelligence (AI)-assisted techniques offer unprecedented opportunities to optimize pricing further and capture greater revenue.

Authentic dynamic pricing combines data-driven demand segmentation, fare adjustments based on true WTP, and fare strategies that optimize revenue cumulatively. Machine Learning (ML) models help identify true demand segments from historical data, eliminating manual processes, streamlining fare calculations, and discovering more price points that can be offered to passengers. As evidenced by our tests with a major airline, ML models can boost revenue by up to 6%.

So, how can we determine a passenger's actual WTP to optimize overall revenue?

RL enables airlines to conduct fare experiments through an agent in simulated booking environments that model customer behavior patterns and incorporate elements of volatility and uncertainty. Using a simulated environment eliminates the need to conduct fare experiments in a live booking environment, which would be unfair to passengers. The agent arrives at an optimal set of fares for each conceivable business context, cumulatively maximizing revenue for a given flight. Simulated customer behavior patterns give the agent data-driven insights into passengers' true WTP. Exploration and exploitation parameters prompt the agent to test untried prices and gauge passenger responses.

Optimizing revenue for all flights across an airline´s network involves considering countless business factors. For high-demand and competitive routes, fare experimentation considering basic business factors (like advance purchase and available seats) in various pricing contexts would involve millions of fare combinations. It would become even more complex if other factors like seasonality and weekend/weekday are added to the mix. The computational load for training an agent in such an environment would be immense, making the task seem infeasible. So, how could we overcome this challenge? More on that in our next blog.