SOSTA AI: Multi-Task Residual Learning for Lap-Level Pit Stop Decision Support in Formula One

Formula 1 being the pinnacle of motorsport, milliseconds can be the difference between making history or a disaster.Pitstops play a very crucial role at a time when teams are trying to gain every last tenth of a second until the checkered flag.In such a competitive environment taking a chance to introduce human error in the decision of making a pitstop is too much of a risk.Accurate decisions need to be made based on race conditions and data.We have built a system that considers a various race conditions like stint length, tire degradation et cetera and the historical data taken from data made public by Formula 1 itself and other trusted sources like the Ergast [5], Fast F1 API [6] and custom engineering it to optimize feature extraction.The novelty in the architecture is based on the cascaded multi-task approach that has been developed specifically for the objective of Formula 1 pit stop decision support, where the classification head makes the prediction regarding the next lap being a pit stop, and the output from the head has been complemented by the estimation from the regression head regarding the number of laps prior to the stop.We have achieved an AUC score of 0.92 and a best pit stop lap horizon MAE of 3.56 laps in test cases suggesting real world capability for accurate pit stop calls.