Artificial Intelligence for run, cycle and triathlon training?
OK, the phrase Artificial Intelligence caused you to click this and you are sceptical, as was I. I imagined it was just another over-use of a popular buzzword but this time applied to run and bike training. So let’s clear that up first then I’ll give you the low down on AI Endurance‘s training plan if you are still interested.
The owner/developer is Markus Rummel, who was originally a physicist. His claim is that his AI Endurance platform is based on machine learning algorithms which he developed.
Specifically, what he is doing is connecting to your STRAVA account or your Garmin Connect account and interrogating your historical performances based on whether or not you are a runner, cyclist or triathlete and looking for patterns in the data.
Why it’s machine learning/AI, and why this approach is different
Every 24 hours, the model aims to correlate your best historical performances with your historical training inputs (pace, power, hr) to find a best-fit signature. This signature is then used to personalise relatively ‘normal’ plans that you would be familiar with ie you tell the model your training preferences and constraints. The resulting plans have some degree of adaptability both to your weekly schedule and changes in the signature.
Markus accepts that users rarely behave optimally and so has various fallback algorithms to deal with that.
So I hope you can see that this approach to planning our training is different. Normally our best performances are taken in isolation as a measure of capability at that time. Our best performances are used to create perfectly reasonable training zones and then somewhat arbitrarily a coach or plan might allocate us various amounts of time in those zones using previously used workout templates. At some point in the future, there might be a breakthrough point and then the normal models are reset. We have assumed that there are somewhat uniform responses from any given intensity-duration activity
Historically we have never really had a way of knowing if any of our training was a waste of time. Taking a silly, but illustrative example, if your body never-ever responds to runs over 1 hour, no-one would know that. Your coach would keep prescribing you with a mixture of speedwork, threshold and endurance runs but really, for you, all those extra-long endurance runs were a waste of time.
You can sign up for a free trial, further below, if this intrigues you. This is not a paid ad and I don’t benefit in any way from you signing up.
AI Endurance – Some Details
Markus claims to have a margin of error of less than 5% with power/GPS data from over 100 workouts.
The plans cover everything from a simple 5k to a more complex Half Ironman. Currently, there are over 50,000 variations of the plans and the best one is chosen that meets the athlete’s constraints. and each of those plan gives a continuous prediction of performance over the duration of the plan and a prediction of the end performance-state.
Please feel free to ask questions and comment below – be nice Markus is just a guy and not a large corporation.
AI Endurance – link to website
How it works
Reader-Powered Content
This content is not sponsored. It’s mostly me behind the labour of love which is this site and I appreciate everyone who follows, subscribes or Buys Me A Coffee ❤️ Alternatively please buy the reviewed product from my partners. Thank you! FTC: Affiliate Disclosure: Links pay commission. As an Amazon Associate, I earn from qualifying purchases.
I have always been curious in the idea of an AI/ML driven training plan. In theory there’s so much data collected by watches/foot pods/power meters/HRMs/etc… and so many people working out with them the curse of dimensionality should be somewhat avoidable, that an AI/ML driven plan should be viable. Certainly for the 90% of athletes who don’t need a trainer, most people aren’t that different. I’d like to pay $5-$10/month for an adaptive, semi-personalized training plan rather than $50+/month for a personal trainer.
However I’d like to see a little more rigor/proof that these models are doing anything more than just a standard workout plan. The “here’s proof that this works in comparison to a control sample” is rather low I’m finding, or at least some more details on the models. AI/ML is huge, are they using Neural Networks? SVMs? Unsupervised Learning? What’s the confidence interval on their predictions, etc…? And then I need more than just testimonials of a website. Ideally someone knowledgeable would do a side by side comparison of all of these and be able to say “This one works” or “This one is just snake oil”.
When Garmin introduced the Daily Suggested Workouts I hoped it would be something like this, but it doesn’t allow you to pick your goals or anything particularly detailed.
Hi driscoll42,
I’m Markus the Founder of AI Endurance.
Thanks for your comment! We are using neural networks and to estimate the prediction error we use cross-validation for time series as for example explained here:
https://robjhyndman.com/hyndsight/crossvalidation/
We are going to release some statistics on how (well) the algorithm works for predicting a user’s performance and what the predicted benefit would be compared to their baseline training on our blog https://aiendurance.com/blog over the coming weeks.
Thanks for sharing. This looks very like a very interesting approach.
I have uploaded my data and have chosen an Ironman plan.
I would like to share a couple of observations:
– My FTP was detected correctly and was identical to TrainerRoad’s ramp test.
– My current marathon time looks reasonable.
– My future marathon time also looks reasonable but maybe a little optimistic (10 min faster than my PB).
– Unfortunately, the site only allows events to be scheduled for the next 4 months which is unfortunate as most events will be more than 4 months out given that we are heading into winter. Also, it is not unreasonable to have training plan of more than 4 months for an Ironman.
– The plan does not include any swimming workouts.
– I would like to be able to set the days for my long rides/runs. I just will not be able to do a 3+ hour ride during the week.
– It would be great, if workouts got pushed to Garmin but realise that the site is small and just getting started.
In sum, this is super interesting and it is certainly worth looking into for others. This is much closer to what I hope Garmin’s Suggested Workouts would be, which generally advise rest for me. 😉
yes i think for IM/HIM a longer-term plan is normal (mine is about 7 months, although a proper build is only the last 4 months)
however MArkus will probably say that you could choose the maintenance plan right now.
yes you could choose the maintenance option for now up until you want to start your build
Hi Kuifje777,
Thanks for the feedback.
Looks interesting, think I’ll definitely take a closer look once it can talk to Garmin Connect as don’t use Strava.
now’s the time…links to Suunto and Garmin and training peaks!
interesting concept which i think has potential to be better than any reasonably priced standard coaching service which are generally not personalised enough.
however i am concerned that the assumption that doing what i have done before prior to top performances may be ignoring the number of things beyond training that influence performance for us amatuers eg work/life stress and hence sleep/recovery. Also, where i am now is not the same as where i was then so is doing the same again the best thing for me now? (i realise its probably a bit more sophisticated than that but that seems to be the general principle)
Hi Patrick,
Thanks for the feedback.
Regarding other influences on performance: in an ideal world one would combine all the data – not just training but sleep/recovery, stress, nutrition – into an input for a neural network. That’s the ultimate goal and would really drive the prediction error down I believe (combined with more accurate data) but there’s a lot of information in the training data already.
Regarding your other point, I agree taking the entire training history is not always optimal for the reasons you mentioned. We are planning to give the option to exclude certain periods of your historical data so you could for instance only use a most recent period with sufficient data and get a training plan based on that. This is also necessary to exclude time periods where bad data was taken such as mis-calibrated power meters or wonky GPS data inflating running paces.
looks interesting, I signed up for AI endurance’s free trial. I selected training for a HM just before Xmas and am a bit surprised that my recommended training doesn’t include any long runs. the longest training sessions are 1h – unusual for HM training I think
yep, will be interesting to see how you do
Hi Thomaek, thanks for the message. Running sessions are in multiples of 30 min right now (30 min, 60 min, 90 min, …) depending on overall volume. Will become more customizable in the future so you can pre-set wished duration of short and long sessions. Let us know how it goes!
I had a email from Markus to say this works with Garmin Connect now as well.
ty, added above