Formula E has grown in popularity as a sustainable sport that pioneers advancements in electric car technology. Its premise is not only that the cars are all electric but also that the 11 teams, each with two drivers, compete in identically configured, battery-powered electric race cars.
“How can we use the data to aid Formula E’s racing strategy?” – Vikas Behrani
Vikas Behrani, Vice President – Data Science at Genpact, spoke at the Deep Learning Devcon 2021, organized by The Association of Data Scientists. In his session, he discussed “Lap Estimate Optimizer: Transforming race-day strategy with AI” and gave insights into how a Formula E race is not only about driver ability and technique but also about data-driven strategy.
Characteristics of Formula E?
Behrani went into greater detail about the Formula E race’s characteristics. It is a racing series dedicated entirely to electric cars. Within each season, extensive performance data on racing dynamics, the driver, and the car itself from the previous seven seasons provide a great foundation for forecasting/simulation employing cutting-edge optimization and data science methods.
A wealth of available data ranging from past driver performances, lap times, standings in previous races, weather, and technical information about the car such as the battery, tyres, and engine, data scientists can forecast the number of laps a car can complete by quantifying behavioural characteristics such as the driver’s risk-taking appetite and other traits such as track information and weather that can affect a car’s performance. Additionally, Behrani discussed how this relates to the other industry – how similar models for racing strategy can be applied to banking, insurance, and other manufacturing sectors.
Forecasting using deep learning
Vikas stated during his discussion of the model that the objective of this exercise is to define the process for forecasting the number of laps a car would complete in 45 minutes during a future race using historical data. He then described the model for the Lap Estimate Optimizer. To forecast the number of laps completed at the end of each race, an ensemble model is developed using a combination of an intuitive mathematical model and an instinctual deep learning model.
There are numerous features such as lap number, previous lap time, fastest qualifying time, track length, and projected time. These characteristics will be fed into a neural network model used to forecast the lap time. We constructed and compared a total of 32 models.
Behrani went into detail regarding the steps involved in LEO.
Step-1: Collecting historical data on the quickest lap time.
Step-2: Collecting historical data on the fastest lap time of rank-1 drivers.
Step-3: Normalize the quickest lap time obtained in step 1 by subtracting it from the matching numbers in step 2.
Step-4: Using the distribution matrix from step 3, simulate data that follows the same distribution.
Step-5: In step 4, add the quickest lap time from qualifying and practise sessions.
Step-6: Add the values in the above matrix row by row until we reach 45 minutes.
Behrani later discussed the predictions for Santiago, Mexico, and Monaco and how the effort on the track translates into market impact. Finally, he went on to illustrate several use cases.
This exercise aims to determine how to use previous data to forecast how many laps an automobile would finish in 45 minutes. An intuitive mathematical model and an instinctual deep learning model are combined to anticipate the number of laps at the end of each race.