AI, ChatGPT and Airline Revenue Management

We’ve all read the headlines. ChatGPT is well on its way to reshaping our world, including how we do business. For years, experts have discussed how AI has and will continue to impact the airline industry as well—everything from customer experiences to airline operations, and even how to better predict the weather to prevent flight delays and cancellations.

PROS is no stranger to the impact—having first used data and science to pioneer the first iterations of forecasting for revenue management (RM). It was back in 1988 that, along with some of our very first customers, we established what would become our first generation of RM science.

Since then, PROS has developed more than 5 generations of RM forecasting advancements, based in artificial intelligence and machine learning (ML), to help airlines accurately forecast demand and increase revenue. More recently, innovations have focused on helping airlines optimize offers both inside and outside current PSS and class-based fares with continuous pricing, as well as forecast based on price sensitivity versus class availability with willingness-to-pay forecasting.

With all the buzz around ChatGPT (a popular generative AI) and neural networks, you might be asking yourself, what kind of AI does PROS use? And how does that differ from generative AI and neural networks? What AI is the best when it comes to airline revenue management?

Generative AI

Generative AI has become ubiquitous with the launch and rapid rise of ChatGPT. This type of AI generates new content based on the input prompts. The underlying model of ChatGPT (i.e. GPT3.5 and more recently GPT4) was trained on over 45 terabytes of text data from a massive dataset known as the WebText corpus.

The GPT model underlying ChatGPT is a kind of neural network called Transformer, and hence the name Generative Pre-trained Transformer (GPT). It was initially pre-trained via supervised learning. During training, the model was fed a large amount of data and then trained to predict the next word based on sequence and context provided by the preceding words, similar to language modeling. This pre-trained large language model (LLM) is then fine-tuned using transfer learning to produce responses that mimics human dialogue. It is further fine-tuned via human-in-the-loop reinforcement learning to output the most desirable response according to the rewards specified by human rankings. As more people use ChatGPT, the model can collect more context from which it learns, and it will be able to recognize more complex prompts and generate in depth response, all without explicit human guidance.

Neural Networks

You might be asking, does PROS use neural network AI? We do!

Our Generation 4 price optimization AI, used primarily in our B2B portfolio, uses a neural network. We also use a neural network as part of our capacity-aware optimization solution which generates opportunity costs with fewer historical data elements. We like to refer to it as “demand forecast-free” AI and are widely promoting it in the air cargo industry for revenue management (RM) where historical data is harder to come by and the market is highly volatile.

Would you need more than a neural network for RM?

When used to predict demand, neural networks are trained on historical data to learn the relationship between different factors that affect demand, such as time of year, route, and pricing. Once trained, the network can be used to forecast demand for future periods. But airlines need more than just a predictive system. And airlines can go beyond just prediction and into prescription.

Learn more about PROS proprietary AI from Dr. Michael Wu, PROS Chief AI Strategist here.

Model-based Reinforcement Learning

Airline revenue management should not be just predicting what will happen but prescribing actions or policy that optimizes future outcomes. This lies at the core of RM – forecasting AND optimization.

That’s why PROS relies on model-based reinforcement learning, which is a type of AI that learns optimal behavior by learning which actions or policies maximize the reward. In the case of airline RM, that means, learning which pricing decisions return the best revenue in order to further optimize these decisions in the future. The PROS AI specifically seeks to prescribe optimal pricing to maximize future revenues.

These prescriptive systems create policies and actions that generate additional revenue. That implies these prescribed policies will be slightly different than policies implemented in the past. It is important that the AI is no black box, so analysts can understand, interpret, and trust these prescriptions. Transparency and explainability of these models are crucial for trust. Also, unique situations may occur where objectives or additional information is available to the analyst. It is important that the AI system allows analysts to influence the system to account for these changes. Having transparency and explainable models allows the analysts to know how to work with the AI system and make the most out it.

What’s next for PROS, airlines, and AI?

We are lucky to have some of the foremost experts in AI and ML working on our AI teams. In fact, we have more than two dozen full-time data scientists with PhDs working on our AI and algorithms every day. Things like willingness-to-pay, dynamic ancillary pricing, dynamically pricing bundles (seat + ancillaries), and request-specific pricing, which analyzes passenger attributes and market conditions to adjust pricing at the time of request to optimize revenue. In fact, our airline customers are actively collaborating as we develop and launch retailing innovations: