Data rather than AI will drive competition, as Formula 1 taps AWS

Formula 1 has just clinched a deal with Amazon Web Services (AWS) that raises interesting questions over the role of AI and machine learning for the drivers and constructor teams as well as viewers. Formula 1 is moving nearly all its IT and distribution infrastructure from on-premises data centers to AWS, while standardizing on the latter’s machine learning and data analytics services. This dovetails with Formula 1’s introduction in May 2018 of two OTT TV services signaling a move towards going direct to the viewer rather than through pay TV distributors like Sky, ESPN or RTL for example, which currently have rights in various countries.

We can now see the logic of the move emerging as Formula 1 seeks to exploit machine learning based analytics to engage more deeply with its global TV audience who can then relate more closely to the dynamics of the race. Such insights include predictions of the race outcome and comments on the performance of drivers on individual laps compared with their ability and prevailing conditions that can change considerably during the event.

Formula 1 is now busy training its data against Amazon SageMaker, a managed machine learning service aimed at helping developers and data scientists build and deploy suitable models. In this case, the data includes Formula 1’s archive of past races going back 65 years to the sport’s beginning in the early 1950s comprising race order, lap times, maximum speeds and some details of weather and track conditions. This is a foundation against which the same parameters gathered live from the event can be fitted and trained to yield insights and help make predictions.

Not surprisingly, much more data is available from sensors for collection now than early in the sport’s history. Cars are now fitted with up to 200 sensors collecting telemetry data on factors including vibration frequencies and amplitudes, temperatures and various pressures, as well as tire wear.

There is also data on the driver’s actions and behavior, including steering, acceleration, braking and verbal communications which could be subject to natural language processing.

There is then track data collected periodically on both environmental conditions such as wind speed and temperature distribution across the surface, as well as measurements of all drivers such as top speeds, pit stop times and lap times. This adds up to a lot of parameters that can be fed to the AWS deep learning model but raises the question of how this will overlap or conflict with AI-based systems being deployed by the constructors.

They are all racing to exploit AI to enhance their competitiveness both in engineering the cars and exploiting minute advantages during races that could lead to a gain in position. This has long been done by human experts but now they are deluged by more data than they are capable of processing accurately, and so the need for some automation is acute and long overdue.

Traditionally the main source of advantage for a given constructor has been engineering the cars and their engines for optimal efficiency within rules laid down by the Formula1 governing body, which are always changing. Increasingly data is becoming a second competitive front, with all the teams now engaging with it separately, even if a lot of it is available to everybody – including some of the sensor data from their own cars.

Teams are working with various AI service providers – Renault for example with Microsoft Azure Machine Learning, Microsoft Cloud and Azure Stream Analytics, rather than Amazon. This is synchronized with Renault’s own supercomputer for 3D virtual testing of car designs to reduce the need for expensive full-on wind tunnel tests. This is a very specific application but when it comes to exploiting vast amounts of telemetry and sensor data during races themselves it is unlikely any of the constructors will be able to outperform the computational capabilities of the big cloud players.

It will all come down to the data therefore and how it is applied rather than the underlying model, which is true for a lot of AI sectors whatever vendors will tell you. The implication is that constructors will become increasingly innovative over the data they collect and so gain an advantage by feeding their models with information their competitors do not have. This means going way beyond typical telemetry and sensor data.

One area with scope for competitive advantage is race simulation, which is related to what Renault is doing with Microsoft. Even after 65 years, only a small subset of all possible race conditions and outcomes have occurred and simulations can fill in the complete spatial background of all eventualities.

A huge simulated data set can be generated to test not just possible designs of cars but also to train machine learning algorithms to make more accurate or insightful predictions of race actions such as pit stops, tire changes, driver strategies and dealing with safety cars deployed temporarily to reduce risk of accidents while for example debris is cleared from the track.

Already, dangerous situations can be simulated to improve safety in design and operation but there is also scope to optimize for large numbers of non-exceptional conditions. It is possible to test a number of imaginary strategies to work out which is likely to be most effective for conditions anticipated just a day or two ahead of a race.

Perhaps most exciting is the prospect of gleaning data live during events by analyzing information readily available and quite legal, such as high-resolution video footage of a competitor’s car or recording of the verbal communications. At least one constructor is considering installing highly sensitive microphones in the car to capture information not just from inside but also of other cars in the vicinity.

This was inspired by observation of how some drivers can take advantage of knowledge that a competitor’s car has some problem that will restrict its performance until it has made a pit stop, or even for the rest of the race. Drivers can pick up clues during the race and in future the machine learning system can enhance this by becoming additional eyes, ears and brains.

The fact some constructors are even thinking about the potential of such data gives a hint of a new competitive landscape to come. One more contentious question though is the extent to which AI based features of autonomous driving will feed into Formula 1 racing. The idea of having some form of automated assistance in a sport that is supposed to test the human skill of drivers to the limit is anathema to many in the business.

But the reality is that in a sport that is already highly technical, the genie is out of the bottle and cannot be stuffed back inside now. Formula 1 has already restricted the application of electronic assistance but that is still playing an ever-growing role. The days when the best driver automatically won the championship are long gone and AI-driven autonomous assistance is going to add another layer between human skill and race outcome.