It definitely doesn’t take sickness into account. A TR employee replied when I asked in another thread as I was curious after taking 10 days off and marking it sick if it knew the difference between my time off sick compared to time off relaxing.
I was sick for 3 days with a head cold back in July, missed one workout and AT changed my next workout to a breakthrough threshold workout the very next day after my flagged illness day!
I understand it has it’s limitations but I don’t see AT as anything more than a formula that simply adjusts future intensity based off a survey.
It’s missing a huge part of the picture ignoring illness and outdoor rides.
Don’t get me wrong, I’m not dissing AT. I think it has a bright future, but at this stage it’s got a long way to go yet.
As things stand, it can’t possibly do that, because it requires information it doesn’t have about what your personal goals are (beyond just ‘train for event x’). And then it would need a dataset across many multiple users to look at what goals they were aiming at, and what the outcomes were relevant to those goals. For that it would need inputs that it currently doesn’t have to capture those goals, and then it would need to build that dataset.
At the moment, its sole declared measure of success is ‘Did you go as fast as you could have in your priority event?’. Two users with the same target event, but with very different goals outside of that, look identical to it.
Has TR said they’re using the same algorithm for both? We already know that AI FTP detection does take into account outside rides and AT can’t recognize them or adapt from them, except from survey result. Seems like not the same.
TR is not using the same algorithm, because FTP and PLs are completely different kettle of fish.
As far as I understand AI FTP also uses data from outdoor rides. The TR team is currently testing PL scoring for outdoor rides. It is unclear when it will be released (as a public beta), but they are working on that, it is a very high priority.
I feel like there is a whole chunk of OP’s issue that isn’t being discussed.
Ignoring the relative practical sophistication of the models, he is claiming that one set of estimations of his work capacity is delivering a significantly suboptimal result, by comparison to the other.
The key question is: does the data set have a good understanding of the rate that PLs should decay without fitness.
A follow-on might be: if you were a bike training company that gave people gold stars for using your software and giving you money, do you have an undeclared interest in motivating people to give you money using that system and might that impact how you design a system? (Zwift does tonnes of this shit)
Probably not, but I also think pls decay too fast. YMMV.
Yours is the right question I think.
It would be tough for their ML to model every user’s PLs, for each of the seven PLs. And both progressions as well as decays.
That’s 14 trend lines for every user. Impossible without just throwing away training plans.
Good discussion, but ultimately we have to accept standard progressions and decays that best fit large populations.