AI Bibliography

We discuss features contained in a machine learning software developed at MIT for professional car racing, to improve the predictions of track position changes within a race. We study pit crew performance and driver performance within selected races, and find that good combined performance for both correlates to better finish positions.

Secondly, we classify tracks based on tire wear and the ratio of 2 versus 4 tire change decisions for pit stops. We find that a driver’s performance in early stages of the race is similar to performance in later stages, suggesting that final race outcomes may be inferred from earlier stages of the race.

Thirdly, we look at how tire change decisions vary from track to track depending on tire wear, caution periods, and stages of the race to understand how teams adapt their tire change strategies as each race progresses. We propose heuristics based on these observations that may be used to improve the software.

Next, we test whether the construction of the machine learning dataset using similar and different track characteristics has a discernable impact on the predictive capability of the software. Our tests indicate that it may be preferable to aggregate different races together because there is no distinct difference in the results when compared to only selecting similar races.

Finally, we cover ideas about how new features could be implemented in the software, and touch on other factors affecting pit stop strategy in the quest for better predictive capability in the software.