Incremental progress helps you find what works.

Why I'm belaboring exercise science.

Last week I had all of these posts about experimentation when n=1. Why have I been belaboring exercise training this week? Before diving deeper into the nuances of sports training, let me connect this thread to what I’ve been discussing so far.

Sports science highlights the issues of building an evidence base from randomized trials alone. The sports science community knows this better than anyone. Because sports science is underfunded, their trials tend to be small and, hence, always subject to the critique of being “underpowered.” But there is too much focus on power and not enough on validity and variability. 

What do experiments on novice athletes tell us about how to train professionals?  Are 4 reps better than 5 when training bicep curls with 20-year-old men? To improve a 5K time, is it better to train 100m or 120m intervals for college-aged women? And what does answering these questions with 20 people tell us more generally about training for soccer? You can’t build knowledge by doing millions of experiments like this. Similarly, we can’t only look at Olympic gold medal winners and generalize about general fitness. There has to be other thinking brought to the table. What is it? And how can it inform other areas of biomedical, social, and physical sciences?

Part of what intrigues me about physical fitness training is its inherently dynamic and incremental nature. Last week, I posed a riddle about language acquisition. The reason we believe that language lessons “cause” language acquisition is that spontaneous acquisition is impossible. But I also tried to emphasize that the incremental improvements in language capabilities were essential to our beliefs that we were progressing. These increments are foundational to how we teach language. 

I had written a first draft of that post using sports examples instead of language. But there are many more caveats with physical fitness. Because of the variability of human talent, some kids are naturally gifted runners who can get onto a varsity team without training. Some kids can just walk into a gym at 15 years old and bench press 250 pounds. It’s not as clean an example as trying to learn Japanese from scratch. But I’m interested in why even those with extreme talent get better with training. That kid with the crazy natural bench press will likely become a champion powerlifter if he trains for a decade. What exercise and language share is that no matter where a person starts, incremental progress accumulates into high proficiency.

Unlike language, fitness training has more quantifiable metrics of proficiency. This natural quantification lends itself to more natural measurement, forecasting, and mathematical modeling. I described Bannister and Calvert’s Fitness-Fatigue model yesterday. I mentioned off-handedly that a single control policy works for almost every setting of the parameters in the model. This jives with the conventional wisdom in fitness that the same basic training principles apply to every athlete, though you have to change how you train as you acquire expertise. 

From a more abstract point of view, this means the state of the system is more important than the parameters of the system. As someone who has spent too much time thinking about optimal control, I find this fascinating. From a research perspective, I have been working on the identification of dynamical systems for my entire research career. There are so many nuances to the dynamical system identification problem, and we’ve only recently been able to get a handle on how much data is needed to nail down a system model. But I was struck by conversations with Manfred Morari and Karl Astrom that there are many applications where you don’t need much of a model to design a robust controller. These fitness models are an amazing example of what Manfred and Karl alluded to. Humans are hopelessly complex systems, yet training them to improve on multiple axes can be principled and simple. Paula Gradu and I have been studying these models in depth for a while, and I’ll share some of our other findings in future blogs.

So bear with me on this topic for a bit. I’m going to dig in and try to keep it entertaining, and I will draw more connections to how I think this relates to other bio-medical topics as well.