I wasn’t trying to be condescending. Just pointing out that no one knows exactly what will happen between now and Feb 10th, so there’s no point worrying about the ride scheduled on Feb 10th when it’s highly likely to change, which is the amazing and awesome part of all this. There are lots of people explaining this and you’re choosing not to listen. You do you though. Good luck! I’m out.
4 Likes
Didn’t tr trial/have a feature where the future rides were just “threshold” “vo2 max” etc? Or was that just talked about as an idea?
Would make far more sense imo and lead to less confusion if that was implemented. If the plan is going to radically change workouts just use placeholders until the day of, then determine what the ai thinks is the appropriate workout and load that up.
2 Likes
I’m also seeing that my specialty block has me doing the highest PL (and highest Watts) for my first week, then lower PLs (and Watts) for the two weeks after that. The differences aren’t huge (e.g., 18W drop from this week to next on the same kind of anaerobic workout) but I’m having trouble reconciling this with the idea of “progressive overload”.
They have it in this update beyond the 28 day AI window. People still complain about that. They’d complain if they did that where you could only see your next workout and everything else was place holders.
I think the one thing I’ve realized from this update is you can not make everyone happy. Also people who enjoy the product don’t start new threads talking about it, we just hear about those that dislike it or are going to quit.
7 Likes
You’re not nuts, the system shouldn’t be changing its predictions unless something else changes. Now, one thing that can change is that they retrain their model on the newest, cleanest data and rerun it, and that causes a change to our calendars, which I think is fine.
Another thing that could be happening is that their plans are taking the uncertainty into account, which I think would be a bad design decision. So if you have a 40/60 chance of rating a workout hard/very-hard then the system considers it rated 0.4 * hard + 0.6 * very_hard . When you complete the workout and rank it 1.0 * very_hard then is actually a different number for the system, and it adjusts your FTP and/or workouts. I think this is counter-productive for showing a workout schedule, and if they are really monte-carlo-ing FTP estimations, they should show the distribution of predicted FTPs so it’s clear how much variance there is.
Edit: a third thing could be that heart rate data influences the future workouts. Hopefully their models aren’t too sensitive to noise in heart rate or power data though.
I haven’t thought about it as much as the TR engineers/researchers though, so I would like to hear more details about this stuff.
They are not yet doing monte-carlo estimations. Nate mentioned recently it was on the todo list. I’m sure they have a big todo list right now, unsure how soon that piece is coming.
2 Likes
Personally, I would not be burning too much mental energy contemplating workouts that are not in the current week (or not in the next few days.. or even not tomorrow or today)…
Being on a plan, you know what type of workout you’ll be doing on any given day (barring Fatigue Detection intervening), just not the specific workout that’ll be assigned on the day. The AI will choose the specific workout based on what it decides is appropriate, using all the model info it has available to it, which can change (if any model inputs change…) right up until the last moment…
At that point, pre-workout on the day, you can decide whether you trust the AI and are happy to do the workout it’s chosen, or you can choose an alternate instead, either by using the Workout Alternates feature or just choosing your own workout from the library. But there’s little point thinking ahead too much a week or two or three about specific workouts IMO as everything is potentially in flux. Take a glance but don’t obsess would be my suggestion.
My experience in the beta was that (in normal operation) workout selection by the AI is exceptionally good, and as a result I’ll perform much less manual workout selection under TR AI than I did with TR AT. Sure, if you’ve had awful sleep or wake up feeling ill - ie. you have relevant info the AI model doesn’t posses - you’ll be in a better position to make a good decision than the ML model, but absent examples such as that, the model does seem to be excellent at managing progression while accounting for fatigue.
In your case, with the model materially lowering your FTP and then assigning a low PL Threshold workout, it appears that it’s trying to feel out your ability at Threshold. In your shoes, I’d be looking to provide it good data by doing a Threshold w/o or two, with appropriate RPE feedback, and then see what it begins to serve up for you. If after providing it with good data, its suggestions look way off, then either call on Support (and/or override), until it truly gets fully on track. Once on track, it’s likely you’ll find its workout selection to generally be exceptionally good - much better than TR AT, and very commonly better than your own manual workout selections (unless ill, poor sleep etc) - as many people have commented upon, and it may take only a workout or two before it has things dialled in for you.
3 Likes
They aren’t doing a Monte Carlo simulation but they probably could add error ranges based on the likelihood of the expected path and number of workouts remaining. Wouldn’t need to run 1000 simulations for each person, they could estimate from their development work.
Here’s my thought process of what is going on with this…If you would have increased the difficulty level of your threshold workouts as the block progressed, the system is likely predicting that you will rate the workout “very hard” or potentially fail the workout, which is beyond the goal of the workout. Thus, AI “dialed down” the workout (as you mentioned above) to keep your RPE in the desired difficulty level (i.e. “hard”).
It learns when you rate a workout. Most of the workouts don’t have a distribution like 90% Hard 5% very hard 5% moderate. They are like 40/35/20/5.
So when you rate that workout as Hard, the next workouts probability may change. Before you were a rider who rated carillon as moderate but now you are someone with an ftp of 300 and rated Starlight -3 as hard and carillon as moderate. So maybe the probability of Cloudripper -2 switches slightly which changes what the workout after it is. It knows more.
But then why did it start off with a Threshold workout that it thinks (with >50% probability) I will rate as very hard?
I guess I expected the AI to use my already available historical threshold workout data to already “know” my ability at Threshold rather than to wait until I do a few more of them. But yeah, I’ll see what happens tomorrow. You can be sure I’m going to go through the gamut of ratings for that session to see how each one would affect future workouts.
I made the mistake of telling TrainerRoad about my gym workout today. It did not change my scheduled Threshold ride for tomorrow, but it did vastly downgrade some “endurance rides,” both the day after the Threshold ride and for next week’s endurance days. So much so, that I’d just rather not even do the session the day after the Threshold ride (or shorten it substantially) as I just don’t see the benefit of long junk mile rides, especially sitting inside on an indoor trainer.
I hear you. I suspect that the AI’s lowering of your FTP number by ~7% suggests that perhaps it doesn’t yet have a good bead on your Threshold abilities, hence it needs to triangulate things via some good Threshold-type data. Once it’s done that, I’d hope that it’d get things on track and dialled-in for you.
I was lucky, as my FTP number was identical under the old and new systems. Those people who saw significantly different numbers under the new system, whether that be higher or lower, have faced a few challenges adjusting to the new system (or rather, the new system adjusting to them…), but hopefully that’s just temporary. Good luck!
1 Like
Sharing an example from my own calendar where if all you look at is the workout level, you’d think your progression is going backwards, but if you look at the TSS/IF, things are escalating.
My workout today
My workout in 2 weeks
I’m not saying I think it will be this way for everyone’s calendar. Just an example.
Being honest, it doesn’t really matter to me because I know the workouts will change over time, but thought this was a good example where what appears to be a reversal is actually a gain.
5 Likes
We need a retro mode. It creates a training plan for the next 6 months and pins all of the workouts.
I’m so thankful we don’t have to use static plans any more that I cancelled my TP subscription. All those static plans I purchased are just sitting in reserve now.
What does the overall load look for the each week of the next 3 weeks?
From this past Monday until Feb 15 (so three weeks), the bike is pretty much flat week to week, but that’s not telling the whole story since I add runs and I know a lot of the bike workouts will change, plus my new AIFTPD will be Feb 6.
Just to keep the story flowing, I just completed the workout assigned today, Blackcap. All of the numbers and rating meet expectation, and the workout for Feb 15th is now this same exact one, Blackcap. Just showing why worrying about what’s on the calendar in 2 weeks doesn’t add much value for me. I’m sure it will change again.
1 Like
For people who don’t care that the schedule always changes, none of this matters. For others, and from my ML/software engineering/UX point of view, there is no reason that the schedule needs to always be changing and therefore useless. TR can create a schedule that, if the user chooses the most likely RPE rating and doesn’t show odd HR or Power behavior, remains the same before and after the workout is observed.
If TR can’t/won’t do this, then they should consider turning off predicted workouts. At best, people who care about this stuff notice that the system is often wrong in the workout it predicts and learn to ignore them anyway. At worst, people who care about this stuff start to wonder if there are underlying bugs in the TR system.
1 Like