A new AI model designed to optimise race strategy in Formula One (F1) could help teams gain a competitive advantage.
The AI model, developed by a research team led by a computer scientist at King's College London, improved Lewis Hamilton's finishing position for the Mercedes-AMG PETRONAS F1 Team in simulations of the 2023 Bahrain Grand Prix.
The scientists believe the AI could provide F1 strategists with additional information to inform their decision-making during races - where quick decision making based on data can improve performance on track.
Dr Antonio Rago, a Lecturer in Computer Science at King's who began the research with a research team at Imperial College London before joining King's, said: "We discovered that AI models were able to not only replicate both strategy and tactics from the real world, but also that they outperformed existing race strategy optimisation techniques in many cases."
Race strategy is the art of making decisions that occur throughout the course of a race to gain an edge on rivals. This may include deciding when drivers should make pitstops and which tyres to use, in the hope of achieving the highest possible finishing positions.
Teams craft baseline strategies based on a variety of factors, such as their cars' pace, the track surface, tyre degradation and weather. These strategies are then used in simulations to calculate the statistical likelihood of what is likely to happen in a race over thousands of runs.
However, because such simulations are so complex, they can take a long time to run. These simulations also fail to consider the complex interactions between teams' strategies and tactics in the dynamic and unpredictable environment of the race, meaning they give limited accuracy.
Using a form of AI called reinforcement learning that uses trial and error to optimise the decisions it makes, the team trained the model using races across all tracks from recent seasons. These models were then supplemented by techniques from explainable AI, which provided the reasoning behind their decisions for F1 team strategists to review - encouraging them to have confidence in the models.
The models also exhibited "emergent tactics", which involves the AI model learning to deploy real-world race tactics without being directly trained to do so. For example, the team saw convincing evidence of the models performing "undercutting", where a car takes a pitstop earlier than the car ahead to take advantage of the pace offered by a newer set of tyres and overtake.
To test how well their newly developed methods worked, the research team took control of Lewis Hamilton's Mercedes car in the 2023 Bahrain Grand Prix in a simulation and compared, over the course of thousands of simulations, their own model, named race strategy reinforcement learning (RSRL), with existing techniques which do not use AI.
RSRL had an average finishing position of P5.33, whilst the best techniques without AI achieved P5.63, a significant advantage given the fixed pace of the car (the teams' cars had an expected finish of P5.50) and a large gain in a sport defined by tight margins.
It was also shown that training strategies can help fine tune the model to specific courses like Bahrain, while a more general approach could be applied to any circuit.
Incremental gains are crucial in a sport considered the pinnacle of motorsport, and the trustworthy decision support provided by our AI models could free up race strategists to focus on other factors, maximising the teams' chances of winning races. This is a very exciting time for the sport, and it may not be long before we see support from AI being crucial to winning championships.
Dr Antonio Rago
Dr Kedar Pandya, Executive Director of Engineering and Physical Sciences Research Council (EPSRC)'s Strategy Directorate, who supported the study, said: "This research highlights the growing potential of advanced AI to tackle complex, real-world decision-making challenges.
"By combining reinforcement learning, where systems learn through experience, with explainable AI that makes those decisions more transparent, Dr Rago and his colleagues have developed a powerful new approach to race strategy.
"Innovations like this, supported by EPSRC, are an important step towards wider adoption of AI in complex, multi-agent environments."
Read the paper, published in the Special Issue on Discovery Science of Machine Learning Journal: https://doi.org/10.1007/s10994-026-07081-3
The research was partially supported by the Engineering and Physical Sciences Research Council (EPSRC).
