Automated Generation of Training Programs for Swimmers - an interview

Here at svexa, we’re continually looking for ways to drive forward the science of sports performance. One of the areas we focus on is how to help coaches and recreational athletes to more easily build high quality, individualized training plans for their athletes or themselves. Recently grad students Rikard Eriksson & Johan Nicander worked with svexa’s guidance to develop a new system for exactly this. Their data model is now proven to successfully generate training programs for swimmers that closely match professional coaches.

Ahead of the full publication of their report, we talk with Rikard and Johan about this innovative work:

What is your background?

Our degree training was in Industrial engineering (Rikard) and Engineering physics (Johan). More recently we’ve both been working towards our M.Sc. in Data Science and AI. Outside of our studies we’re both recreational runners, so it was fun to study the sports field for our Masters thesis.

What was the impetus for this project, and how did you get the idea to want to work with svexa?

As runners, we naturally have an interest in the science of training. Previous courses where we worked with healthcare further ignited our interest in doing a project related to sports and exercise. We connected with Mikael and discussed different alternatives, the swimming project came up as an option that sounded interesting. We wanted to work with svexa because of the complementary knowledge. We know how to approach the problem from a data science perspective, svexa has a lot of knowledge related to training and physiology.

What’s the problem you are trying to solve, and could you describe roughly what the model does?

The challenge we wanted to tackle is that it's difficult and time-consuming to create good training plans that fulfill an athlete's individual needs. Each coach has their own style of training plans. We figured if we can learn how the best coaches in the world would plan training for their athletes, we can both reduce the workload of those coaches and help athletes that normally would not have access to high level coaches.

Based on that theory, our model takes in the weekly goals of an athlete and tries to provide a detailed plan that fulfills those goals in the style of a top-quality human coach.

What snags did you hit along the way?

Johan Nicander
Johan Nicander

The ‘solution space’ is very, very large and finding a good training plan is like finding a needle in a haystack. Many of the methods and algorithms we tried simply couldn't handle it or took way too long to do it. Finding an approach that could correctly balance all aspects of a good training plan and do it reasonably fast was one of the main issues.

More specifically, swim training is difficult to log automatically so we have had to work with the self reported data of athletes. This is an issue since humans tend to make mistakes, especially when doing something as tedious as reporting their training. This resulted in a lot of data cleaning to filter out the usable data.

The science of training is very complex and there is much that still is not known. More importantly there is stuff that we did not know when we started working on the project. It was absolutely crucial for us to get feedback from experts to be able to improve the model and incorporate systems that we would never have thought was needed.

What were the results?

We are excited to confirm that our model can indeed produce training plans according to specified requirements, with a structure that is similar to the coach. Top athletes and coaches we’re working with have confirmed that the plans produced are certainly good enough to be used by a wide range of athletes, tailored to the needs of each.

How will this new model be used going forward?

As a first step, our model will be integrated into svexa’s automated training plans system. We’re continuing to widen the capability of the engine, to produce plans for a range of sports. Over time, we’ll continue to refine our understanding of how the best coaches in the world plan their athlete training, and widen the toolset to offer these elite-quality plans to athletes at all levels.

Abstract from Thesis.

Optimal training planning is a combination of art and science, and a task that requires expert knowledge. This is a time-consuming task that is often exclusively available to top tier athletes. Many athletes outside the elite do not have access or cannot afford to hire a professional coach to help them create their training plans. In this study, we investigate if it is possible to use the historical training logs of elite swimmers to construct detailed weekly training plans similar to how a specific professional coach would have planned. We present a software system based on machine learning and genetic algorithms for generation of detailed weekly training plans based on desired volume, intensity, training frequency, and athlete characteristics. The system schedules training sessions from a library extracted from training plans written by a professional swimming coach. Results show that the proposed system is able to generate highly accurate training plans in terms of training load, types of sessions, and structure, compared to the human coach.