This year’s winner of the 2025 INFORMS Edelman Prize was USA Cycling. This project generated well-deserved excitement within the Optimization and OR community.
In short, they built a Mixed Integer Program (MIP) to help the Women’s cycling team win gold in the Team Pursuit event in Paris in 2024. What makes this more impressive is that they were underdogs and needed to shave an improbable 8 seconds from their 2023 times.
Check out the video of their presentation. Here’s why I like it so much.
One, it’s a great story. It’s fun to hear about an underdog pulling off a big upset. It also helps that they use a MIP to do it.
Two, it’s a cool MIP application. The MIP was used to select the team that would work well together and to set the race strategy. For the race strategy, they modeled each cyclist’s individual characteristics, including power, energy, and recovery time. They also modeled the detailed aerodynamic impact based on position and developed strategies for switching positions. What was interesting, like what we’ve seen with deep learning algorithms playing Go and Chess, the MIP suggested a non-traditional strategy.
Three, the problem is non-linear, but was modeled as a MIP. I think MIPs get a bad reputation for not being able to handle non-linear constraints and variables. The dynamics of cycling are non-linear: aerodynamics, energy burn, and recovery time. However, they demonstrate that, with clever formulation, it can be modeled as a MIP.
Four, they have a nice way to present the MIP. See the video and judge for yourself. I picked up a few tips.
Five, you don’t need to have a big team to make a big impact. They built a complex MIP and convinced the coach and athlete to trust it with a tiny team.
Love this use case!!! Great application for optimization
This is really a very interesting use case Mike. The race strategy part is the one that caught my attention the most in terms on how the team and coaches were confident on the model recommendations and executed accordingly.
Wondering how they came up with which variables to consider (some are more obvious than others) and how they came up and fine-tuned the penalties in their model.
I'm sharing this article with my network...