Aviation & Explainable AI
What can airline commercial organisations learn from self-driving cars?
Self-driving cars (SDC) are apparently going to change the world.
First up is the efficient use of urban space. Parking currently takes up huge amounts of land – more than 14 square kilometres in London alone according to the Centre for London – because most cars spend most of the day parked. Where Oliver lives, roads are effectively one-way as cars take up all the space on either side leaving just a little bit in the middle for actual driving.
SDC will allow people to order their vehicles, which will arrive on demand and make the journey before heading to the next customer. A few depots will take care of maintenance, repair and overhaul. The spaces currently covered in cars will be able to be used more productively, for gardens, shops, offices, wider pavements for pedestrians and even just better views.
Then there is safety. The UK’s Department for Transport reckons that 80% of accidents are caused by human error so in principle SDC could be subject to lower accident rates.
At Airline Revenue Economics we have our doubts about the viability of SDC technology, at least in the short- to medium-term. In any event, Oliver would rather take a train or bus than a taxi. While we were in Montreal he even dragged me onto the metro rather than taking a cab. Something about not getting stuck in traffic. Londoners, eh…
But the way that SDC are being developed has important lessons for airlines today. Engineers use a technique called Explainable Artificial Intelligence (XAI) to help make the cars drive better and safer.
The idea is that XAI is built into self-driving car prototypes so that when they are being tested they generate data about how the car decided what to do, when to go straight, when to turn and so forth.
Sometimes the SDC does the right thing, like driving in a straight line down a straight road with no obstacle, or taking the correct turn. But sometimes it gets things wrong, like turning for no reason or colliding with an obstacle. XAI tries to turn all the information that the SDC’s on-board controller used to reach each and every decision into insights that are meaningful for humans like text, charts and images.
In essence, XAI is about using colour and space to link pieces of information and show how an end model gets to it’s result. This model consists of many layers of models stacked together. It is a bit like taking a jigsaw and figuring out which pieces go together just from looking at a pile of pieces. Fundamentally, XAI tries to create a process that the eyes and people can cognitively follow and understand. It is about more logical steps, not opaque and tough equations.
To some extent, XAI is an extension of the principles airlines already use in accident investigations. When something goes wrong, investigators try to understand what happened and why. This approach, rather than saying who is to blame, is a vital part of airline safety culture.
XAI will surely be a part of all future black boxes, the part of the plane which records everything happening in a near indestructible casing, so if an accident occurs inspectors can find out what happened.
But XAI will have implications for airline revenue too.
Commercial departments are currently highly structured and linear, or stacked, while buyers, passengers and supporting technologies are not (see article). When airlines attempt to update their technology and processes they struggle. If airlines want to become more sophisticated, they need to simplify their commercial models first. They also need to become goal-centric, like the self-driving car.
At the heart of the problem is control. Buyers and passengers want to decide what their journey should look like. Instead they are ‘fed’ offers from airlines. Airline commercial products are often the wrong way round. For example, scheduling focuses on departure time, not arrival, and it can be better to buy a non-refundable return ticket and throw part of it away than pay an extremely expensive refundable fare (see article).
Abandoning the old principles that tickets are non-transferable or that airlines should consider offers from passengers can generate revenue growth (see article). Yet concepts like this are controversial not because their economics are unsound, but because they challenge the principle that airlines having full control of everything in their business is always and everywhere a good idea.
So how can XAI help?
There are three challenges which XAI will help airlines overcome – the difficulty in ensuring fairness and minimising bias, the fact that an insufficient number of models make it to production and the inability to use sophisticated but opaque algorithms.
XAI helps an airline’s mangers see clearly what is happening in their business in a way that traditional management reporting does not. While traditional reports show outcomes and relationships between KPIs that have been pre-defined, XAI also displays clearly the linkages that human eyes have missed.
While planes have their black boxes, XAI turns airline organisations into “glass boxes” and operates across the whole organisation – at the so-called “enterprise” level rather than within individual departments.
These glass boxes visually show how everything the AI algorithms have optimised and predicted, regardless of bias. XAI algorithms also figure out everything people need to see, so all relevant views are produced. And when algorithms find something they cannot solve or depict, they alert their operators accordingly so new and better programmes can be developed.
AI on its own presents challenges
Airlines today are figuring out how to use AI to personalise offers and proactively offer buyers products aligned with their personal willingness to pay, including ideas about what represents value for money for each individual person. But they need to watch out for the harm which opaque algorithms can cause when they lead to unintended results such as bias, racism , stereotyping or exploitation of vulnerable groups. When AI delivers contrary outcomes the whole airline is damaged as regulators act, brands are damaged and reputations lost.
Ethical AI has evolved to address these issues. So when an airline modernises its data across all departments and centralises it on a Customer Data Platform they will not run afoul of the law.
Ethical AI can also help customers build trust in their airline’s pricing. For example when frequent business travellers see they are being given a special deal on their holiday that is not available to others they will feel better about travelling the airline on business in the future, even if fares are more expensive than on competitors. Ethical AI will lead to improved commercial processes, robust customer engagements and faster time to market.
Complexity is endemic but tough to address
The more complex an algorithm becomes, the more difficult it is to understand how the underlying AI model has arrived at a decision. Unfortunately for XAI specialists in aviation, airlines and the algorithms they need are extremely complex.
For example, airline marketers need to trust that their Offer Management Systems (OMS) propose relevant flight, hotel, excursions and lifestyle combos at the individual buyer or passenger level. Their XAI wonks will need to understand how the OMS reached a certain decision and whether or not it hit the mark in terms of being close to the best the airline could have done at that time.
Measurement is tricky – XAI needs to define not only how many people have engaged and transacted, but also how many did not with ideas as to why. Hit and miss is not acceptable in a market where people can get captured by another airline’s loyalty programme (see article) and never return to the carrier making the first, badly targeted, offer.
Who should airlines work with to deliver XAI
XAI has three significant benefits for airlines. First it will help them increase the coverage and productivity of their offers, commercial models and decision-making capabilities.
Second it will increase the revenue and profitability of the commercial product due to better matching with buyer and passenger preferences. And third it will lead to higher accuracy, helping the airline make better and more profitable use of its capacity.
All the standard airline vendors claim to be AI experts. Considering that many of their core products could be obsolete airlines may be be unwise to rely on them. Besides, and in a future article, I will share why I believe the future of revenue optimisation is in proprietary systems for airlines that survive with differentiating service models.
Serious researchers in the field of XAI have been operating since about 2010. IBM (Cloud Pak for Data) was the first and InRule, Google Cloud Explainable AI, Rulex XAI and VirtuousAI followed. Airlines would do well to consider them in their XAI journey to help improve the deployment of AI across commercial functions.
One of the latest innovations to boost XAI is the so-called ‘data fabric’, an architecture that unifies and essentially ‘stiches’ data together. That, however, is for a future article.
ricardo DOT pilon AT millavia DOT com (author)
oliver AT ransonpricing DOT com (editor)