AI Training Systems

Hi,
Does anyone know much about AI (Artificial Intelligence) training systems and their value and ability to train up an athlete? I have recently started looking at Altabrio & AIEndurance and signed up to their free trials just to have a look at what their offering is. I have uploaded all my Garmin training data in to their systems but aside from predicting future race times their planned workouts seem very “hit and miss”. Thought I’d tap in to the worldwide Intervals brains trust out there to see if it is worth further exploring or not.

Not really much but I’ve read about it on Alan Couzens blog and on AI Endurance.
The principle goes like this:

  • You fill out some sort of survey with your goal(s), past training habits, available days/time and so fort
  • You are proposed a training plan taking the above in account
  • If you for some reason do not follow that plan, all future workouts are recalculated to get you to maximal Fitness/Form on event day.

That’s basically it: the AI part fills in different workouts and reiterates until it has found the sequence with the ‘best’ outcome on the day of the goal. It does that on every occasion. If today you decide to do something else then what is planned/proposed, the AI part goes back to work and adapts everything from tomorrow on, attempting to get you in the best possible condition on race day.
Some take HRV readings in account and adapt the daily training advice for your physical condition.

It’s what good coaches do based on feel/experience but in a mathematical way that takes advantage of machines to calculate just about each and every possible combination of load/time/TSS, type of workout (LIT, HIT, Recovery, Rest…).
How good these iterations are at this moment, is hard to say.

4 Likes

I’m using the TrainerRoad adaptive training which modifies upcoming training session based on how well you’ve completed past ones. It seems to take into account your rating of a session’s toughness when you’ve completed it, as well as your feedback when a session has been really a bit hard to complete - for example if I’ve had to pause a few times, I get a questionnaire about why.

It seems to work pretty well.

Robert

2 Likes

TrainerRoad’s “AI” is mostly in the workouts toughness ratings. Otherwise it follows a prescriptive plan where each workout is taken from a progression. The amount of work on each system is baked in the plan along with the exercises scattered around. If you complete the exercise X, then next time you’re ready for X+1. This is very sensible.

However I notice that by following the plan and progressing with my FTP I get kinda same workouts every cycle :slight_smile: So sometimes I ignore the “downgrading” of the workout levels after FTP raise to get something new.

Also, this winter I’m using Adaptive Training applied to the classic base and build plans instead of using the Plan Builder. This feels more meaningful.

P.S. Don’t forget Garmin’s workout suggestions! They are rudimentary, but are also looking pretty good.

3 Likes

I think the best application of AI nowadays would be to take a normal training plan and adapt it. There is research showing that following a structured plan generally beats unstructured training. And you can get even more by looking at your recovery metrics to decide whether to push on, stay at the same level or lie down and rest. (Kinda the same thing as TrainerRoad’s AT, but the “recovery metrics” is whether you are able to hit the power targets)

3 Likes

Thanks. As I am training/coaching myself for a 70.3 in 6 months it would be useful to use the AI as a tool to assist in the selection process of which workout to select at any given date/time depending on what the algorithm thinks is best for me. I have a structured 70.3 training plan so that should pretty much nail it for me (you’d think) but of course it is a “cookie cutter” plan and will not take in to account my “wellness”, recovery etc. If the AI is as good as I think it might be it might propose what workout is required and then I can select/modify that workout from the training plan. Thanks for your input.

Interesting. The new Garmin devices that plan weekly in relation to the future race and adapt to physiological values should also be taken into consideration.

1 Like

I’ve experimented with a couple of systems. Trainasone has been quite good for running only training. Humango has been ok when I want to have a mix of running and cycling training, but the workouts can seem overly complicated. I think with any of these you have to give them time for the AI to adapt to you and your training. However, in the end I always seem to revert back to setting my own schedules and adapting each day based on experience.

5 Likes

It’s the same workout, but with a higher FTP, correct?

So you should continue to get adaptations by completing it at X% of your new FTP.

AT 2.0 has been mentioned by a few TR staff already, and one if its aims is to include your outdoor/unstructured rides in as well to provide better suggestions. Not sure what else has changed, details are slim currently.

1 Like

Nothing wrong with stock training plans or plns that adjust to your progress//performance but you will never get the same sucesses as from a live walking talking hands on coach. Particularly if you have in person coaching. Wont belabor the obvious benefits. You can figure those out.

1 Like

Thanks, been coached for run, trails & triathlons since 2018. Time to self-coach for a change. It could be a success or disaster in June when I do my first 70.3. :grinning:

I’ve used Enduco for a couple of months.
Best part about it compared to trainerroad, it takes your unstructured and outdoor rides into account and adjusts future workouts.
It works with a polarized approach but unfortunately it didn’t explain much of the given structure.

It generated a plan based on your goals and calculated a weekly TSS and according to the developers it factors HR data aswell but it just seemed unclear why it adjusted certain workouts and in the end I never reached the targeted TSS score. At 15€ / month it was okay but nothing I couldn’t do without it.

I think the appeal of AI coaching is that a typical coach costs somewhere between 100-300€/month for a ambitious amateur and could go up even higher. For me that is not in my budget at least not this season as I’m just starting with structured training and are mostly working on consistency for which I don’t need a coach.
Will this year be greatly structured with the best possible outcome? Most likely not. Will I be prepared for next years season with some experience in structure and consistency? absolutely!
Will I have a budget for 100-300€ coach? I’ll ask my manager about that :smiley:

4 Likes

Hell just have fun. 70.3 great distance that you don’t have to surrender your life to. In way past did FL a couple of times and Vineman (still exist?) 2x. Wildflower too a couple of times-great event All self coached and BOP but who cares. I am high on in person coaching because of track cycling where a good coach’s expertise got me on the podium at Nats. Enough me-- Where you headed? and good luck/fun.

I did the trial on AIEndurance.
The concept is great, the training workout proposed sound good.
I did not continue on it because of AI needs correct data.
The analysis will work if you run on track, flat road.
If youd are like me who run mostly on hilly path then you AI is a bit lost.
If you do some other sport than running or cycling then the plan is useless.
I really enjoy the experience with AIEndurance, got quick answer from questions that I have from the coach. That’s great, but as I do not only run on my running plan, I can not use them as monitor.

1 Like

Thanks for the feedback.

Hi all,

The AI-coach company I co-founded in 2015 is called Athletica.ai. What we tried to do with the software is take the main concepts from a book called Science and Application of High-Intensity Interval Training (https://hiitscience.com/) and apply them to code. After connecting your wearable and plugging in your race date, we take your historic data and build your power/pace profile before leveraging Banister’s classic 1980 impulse-response training load model. We then adjust training to the model depending on what you do or don’t do. More recently, we’ve hooked up to ChatGPT and using prompt engineering analyze your session and comments (see session analysis screenshot). Overall, we are delivering a coach-like experience for a fraction of the cost. Moreover, our coach version leverages all the same principles for coaches freeing up their time to make deeper connections with their athletes.

Best,
Paul

2 Likes

Without real specificity regarding the event for which you’re training (i.e. starts like a crit for 7 miles, “settles” into a hard paceline on rolling terrain for 18, followed by a 4 mile climb at an average of 7%, etc.), combined with a way to analyze significant performance history, I don’t see how current tech can do much more than elevate your general fitness.

For example, is there an AI coaching platform that will allow a rider to:

• Upload a GPX file of a course
• Take the rider’s goals for that event into consideration
• And then provide a customized plan that includes a pacing strategy?

Discuss :slight_smile:

5 Likes

Garmin already does that. :wink:

1 Like

Hey Paul,

That sounds really good. Do you have any plans to incorporate training plans for specific types of riders? I race on a team in Zwift and would love to train specifically for sprinting because that’s the main way I think I’m able to help the team.

Best, Rune.

My interest this season is endurance and as a long time user of trainerroad I followed their polarisation plan.

After spending some time studying Polarisation methodology I concluded (others may argue) that TrainerRoad is not fully committed or adapted to the polarisation method.

Looking for an alternative I picked up a trial of JOIN. For me this was much more polarised focus. My numbers improved and I wasn’t left so drained after sessions plus it has an event focused option which suited some of my targets. (Fred Whitton & Dragon Devil).

Like other AI apps, it calculates and predicts FTP etc. and I found them pretty close. I recently undertook a lab Lactate test with the result for me was 133, JOIN had me at 135. As that figure can change daily I regard this as good as it will get.

JOIN isn’t perfect for me . I would like to see a data link to Hrv4training and a comprehensive training history but having said all that JOIN has been effective in helping me achieve my results.

I will keep my eye on others as this is a field that is going grow rapidly.

1 Like