In the world of motorsport everyone wants to achieve the best result and be on the top step of the podium. The key to winning a race is choosing the correct setup of your vehicle. There are a lot of factors that can influence the performance of the vehicle. How much downforce do you put on a vehicle? What engine map should you use? What is the driving style of the driver?
As big motorsport enthusiasts, COMPUTD was excited to figure out if it was possible to find the optimal setup for a driver using telemetry data from free practice sessions in the motorsport. The main problem that is encountered is that this data is not publicly available as competitors might use it to get an advantage over the other teams. To solve this issue COMPUTD decided to focus on the Formula 1 category and use a Formula 1 simulator to gather data.
From the Formula 1 game, it is possible to extract your own telemetry data as well as from the AI (Opponents controlled by the computer). In the Formula 1 game, we can also tweak the setup of your car in much detail with settings like front and rear wing angle, the camber of your wheel, differential adjustment, suspension stiffness and many more options.
Gathering the data
The gathering of data was a process in which it was required to play the Formula 1 simulator against the AI with different setups on different tracks in different weather conditions. The game was also played on a force feedback wheel and pedals to simulate real-world telemetry.
The data that was saved contained all the different options of the setup and the telemetry of a practice session in the game. For each combination of setup, track and weather conditions, 10 laps of data and the respective lap times were saved.
Making a prediction
A machine learning model is used to determine what setup is optimal and gives the best lap times. However, the lap times are not the only evaluation metric. The setup should also be stable (in the data it can be observed if there is a lot of over- or understeer) and the tires should not burn the rubber (tire degradation is the main factor in determining the race strategy). To incorporate all of these factors a convolutional LSTM network is used to make a prediction for the ultimate setup.
The prediction network has good results and can determine a good race setup for the car. However, it predicts mostly setups close to the standard setup. The first explanation for this outcome is because the AI drives with standard setup and a lot of data is based on the standard setup. The second reason is that the evaluation of our setup is also based on tire degradation and how stable the setup is. The standard setup in the Formula 1 game is very stable and very good on the tires. Next to that the standard setup also gives decent speed.
The prediction network has also been tested when only lap times are used as an evaluation. From this, it can be concluded that it will predict the fastest setup but these setups are burning through the tires and are very unstable. In short: great for qualifying (where lap time is the most important factor) but not for the race (burning too much of the tire compromising a race strategy).