Feature request: Could adaptive training provide the logic/rational for recommended adaptations?

There seems to be questions and interest from other users on why adaptations are recommended or not being recommended based on one’s completed workouts.
I too am in this boat as some times I think I understand them and sometimes I don’t.
It would be interesting to be told why the prescribed adaptations are being recommended.
For example, it could be something like:

  • completing the work out so TR will increase the PL level of similar workouts
  • based on the survey response TR thought you should rate the workout as easy but you indicated very hard so we will back off similar future workouts in order to get in line the workout level (e.g. achievable, productive, … ) with survey response.
  • and so on or whatever logic TR is using to adjust future workouts
1 Like

TR isn’t using logic / pre-defined heuristics for AT, but a machine learning model, so there isn’t a way to do what you are asking. It’s a limitation with ML models that you can’t create a human understandable reason for what it does.

6 Likes

I’m not sure the workout prescription process is completely under ML (?). Is that really the case? Would be interesting to hear the boundaries of the AI algorithms and use cases from TR…

I was assuming that ML is (only?) used for workout compliance after completing a workout (or for giving you a score (progression levels) with WLv2).

And that the next workout prescription was more of a deterministic logic alongside a ramp rate and/or the predefined plan structure (workout progressions) in case of compliance (with some setback on failure).

(Somewhere was a forum thread with graphs for ramp rate and progression logic iirc).

So in that case it would be possible to give some reasons for the adaptations but I assume it would mainly be sth like „progressive overload“ and „dialing back to bring you back on track“.

2 Likes

I took significant time off this fall or at best was inconsistent. Started a new plan this January. Completed an over/under workout and answer the question at the end as “Very Hard”. Then AI asked why and I said “Intensity of Training”. so what does Ai do but move future threshold workout up harder. I did send a question to TR and have not received a replay back yet. I find this confusing.

I will say AI is very good at predicting my FTP, where I had it predict and then also did a ramp test.

For this situation, I am guessing that one or more of the below may explain your situation:

  • you completed the workout so that is good and as such your next similar workout should be progressed in some way
  • I am not sure what “category” the work out was for you (e.g. achievable, productive, stretch, …), however, it if was a stretch workout, perhaps TR was shooting for you to answer it was an “all out effort” or “very hard” at least. However, you only said it was “very hard”, as such, it was OK to progress you?

The above are just some guesses.
I suspect you were hoping to not have that workout progressed as you think you will fail the next workout if it is “harder”?

Yes, i just think AI keeps progressing a little too fast from week to week. Moving up the difficulty significantly. In stead a .1 or .2 progression might make sense vs. .5 jump. Sometimes you need to prove you own a given level before getting pushed. I can definitely not accept the adaption but it just coms back all the more aggressive the next time.

Here is the history.
First week of low volume training. Gave me a sweet spot 4.0. I marked this as level 4 very hard. Seemed a bit much for the first week, but I do have a history of training. It moved future workouts up which confused me. But it did not move the 2nd week up. Gave me 3.6 sweet spot and I marked as level 4 again. This time it asked my why and I put “intensity”. I completed it but it was tough. Now it dropped future workouts (3 of them). Maybe AI knew at lessor level workout was coming and needed two of them (4.0 and 3.6) to make a true determination.

From the podcasts and other posts around this, marking one effort as very hard/all out or easy probably won’t cause AI to change your whole plan. Things happen in life, you didn’t sleep well, had a little cold, ect. that increased RPE for that one ride. If you did that ride again on a different day you might not rate it that high so AI doesn’t want to limit you for one response saying something was hard and slow progression too much. If, like you did, you mark efforts consistently at one end of the spectrum it will take those into account and ramp up or down the rate of progression.

1 Like

I also would like a more precise explanation to vanbc’s question on the data used to for adaption recommendations.
Copied from TR Website is the only info that I could find:
“Adaptive Training looks at your current capabilities and custom training plan to determine whether adaptations need to be made to your upcoming workouts.”
From "HOW TO USE ADAPTIVE TRAINING video; “adaptive training adjusts based on your performance.”
“Progression Levels are a complex calculation of many factors to include…
-Recent work out performance
-Training history
-Workout types
-Time between workouts”

The above clarifies somewhat, but is still too vague.

A more specific question i have is, “what data is used to analyze recent workout performance?” Is heart rate data used at all? If not, then it would be a pretty limited analysis as only options would be if the work out was completed or not and if breaks were taken. And the self reported ease of workout after the workout.

The other stuff (Recent work out performance, Training history, Workout types, Time between workouts) makes sense to me and I assume is similar to what xert has been doing.