Is Machine Learning a Game-changer for Airlines?

Why has airline forecasting been so difficult? 

Airlines have worked for a generation with explicitly programmed optimizers and forecast algorithms to obtain some measure of predictability in their planning. On paper, optimization appears to be a panacea, if you have very few constraints and a fairly simple business model. In practice, the complexity of operational constraints, slots and limited data have made this type of program fairly impractical. Given the huge effort to get the data ready and a calibration process that is very expensive and time consuming, most airlines have not found the results they get to be worth the effort. Some airlines don’t even get the legacy systems to work at all for their situation.

Is Machine Learning a Game-changer?

The question in front of data analytics teams today is whether the new generation of machine learning solutions is a significant improvement over optimizers and forecast tools of the past. Does the new approach yield better insights and enable more confident decision-making? Machine learning is relatively new to network planning, but its ability to explore the variety of historic data and new information is showing a lot of potential compared to the expectation that real world data would work within a rigid black box program.

Some results of machine learning have been a pleasant surprise. Imagine showing a computer a few examples of new markets that proved to be successes and a few that were failures, then feeding it 4 years of historic data to seek predictive insights. The results of doing this retrospectively indicated that failures would have been predicted months before the airline actually did anything about it. On the flip side, some carriers actually got out of a few markets too early. The learning model predicted they would have had a longer-term success. In the actual data history, competitors eventually did enter those markets and make a success of them.

A Word of Caution

Machine learning enables innovative solutions, but it is not a piece of software you can just buy and use. It requires a development process with the collaboration of three key areas of expertise. The first is the customer’s domain knowledge of their business model and how their operation works. Combine this with a data scientist who can translate the logic of the domain expert into guidance the machine can recognize and you are nearly there. The final piece is “training” the machine in a series of iterative human-machine learning trials on a limited set of data and then allowing it to explore the large data set for similar patterns.

Experienced planners almost intuitively know how to solve complex problems, but they are limited in the scope of data that they can keep in mind. The value of machine learning solutions is to leverage the computing capacity to handle huge amounts of data, but to do it in an intelligent way. This is the human-machine nature of the learning approach. There are a lot of algorithms available for exploring data, but machine learning solutions require deep domain knowledge about the logic of a solution in order to yield high value insights for decision making. Each airline is somewhat unique, so the logic and data of each solution is also unique. Each airline will tailor machine learning solutions using sophisticated algorithms together with the domain knowledge of their unique business model.

Having a production software platform for integrating the machine learning solutions into the planning and scheduling workflow is the final piece of the system. A planning solution may appear optimal, but you need to check all of the rules you have for operational, crew and slot constraints so that a scenario can be both valuable and operationally feasible. This integration of workflow enables faster and more flexible capture of revenue opportunities and operational efficiencies for a higher value network.

Zulu has the tools and expertise to work with you when you are ready!