Have you seen the new adaptive training of trainnerroad? Do you think it would be possible to have it also in Sufferlandria?
Very interesting and would definitely vote for this.
Not that much the TrainNow function but more than anything (1) the Plan Builder function they have and (2) the ability to adapt a training plan based on your performance in previous workouts.
That must be the future of training or the present !!!
To make this work wouldn’t you need to run the videos in level rather than Erg mode? If running in erg (which I think most people are most of the time), other than looking at heart rate data, what data would you have that allows you to adjust future sessions?
Assume you would need to run at least some sessions in level mode and so that the “adaptive training” functionality could then compare actual power with planned power for the session and hence adjust future targets up or down accordingly?
@JohnK It adapts the training plan and what workout you do next, not the specific workout you’re doing in real time.
Looks interesting, yeah. But interesting doesn‘t cut it - for me. Wake me up when they pull in „every-day data“ like steps, 24/7 hr, etc to make an informed decision.
For whom might be interested, Ray Maker posted an overview.
On another note, I sincerely believe (not know!) that every fitness company who offers some kind of training plans is at least looking into this whole „machine learning“ thing.
I do wonder if this might be something Suf are about to also announce. It has been a common request in the past to have alternative workouts to the ones in a plan, I wonder whether an adaptive plan, with some user choice, could be on the horizon?
@Tobias_Breme Yes, I appreciate that but I presume it adapts future workouts based on how “easy” or “hard” your find the existing workout. If you are running in ERG you either hit the targets or you grind to a halt. Assuming you do hit the targets how is the software going to know how much “effort” was needed to achieve that? Beyond looking at heart rate how would it know if you were flat out for the whole session or cruising along comfortably?
On the Trainerroad podcast they said there will be a one or two question survey at the end of each workout, which will give them your RPE for the session.
@JohnK TrainerRoad is also asking you how the session felt at the end to add a a subjective piece. So if you hit all the targets and said it was easy, I assume it will ramp up the intensity on the next workouts. TR did say for now they only increase the intensity score of the next workout selected. They aren’t increasing your FTP settings until you do another ramp test (for now). So if you were to repeat the same workout even after saying it was easy it’d still be at the same intensity, though TR would suggest a harder workout.
And that’s part of another another big difference between SUF and TR. TR will have a base workout and then multiple versions of the same one with variable numbers of intervals, like imagine Revolver +1, +2, +3, -1, -2, -3, etc. The Shovel +5 (…). Hard to see how SUF could do that with their video workouts. No Vid might be possible though.
Honestly with 4DP I feel like TSF is close to what TR is going to offer. TR is identifying 6 or 7 power levels and suggesting workouts based on your strengths and weaknesses (sound familiar?). I think if TSF could automatically detect 4DP increases based on HR and power data similar to how Zwift does it that’d be good. Especially if they could allow us to import and analyze external workouts. I could see AC and NM being pretty easy to analyze actually because 1 min and 5 second max pushes aren’t too uncommon when riding outside, at least for me. For example if my NM is 1000W and I have a 5 second effort of 1050W from an outside ride, SUF should be able to recognize that increase my NM. I think automatic FTP and MAP adjustments might be harder, but they seem to be doing based on power data and HR in the Half Monty. I wonder if similar principles could be applied to other rides.
I’ve been tempted to try TR a few times but really it just seems kind of boring. Their initial base build phase is nothing but Sweet Spot/Tempo rides for like 2 months without ever really pushing above threshold. I can’t speak to how effective that is but it’s just so bland. And don’t get me wrong. I can do 5 hour endurance rides and be perfectly happy. But I love the variety you get from day 1 with SUF.
Yes, I agree that automatic updates of NM and AC should (in theory) be pretty straightforward. On occaision I have run Open30 in level mode, done some sprints and one minutes efforts and then used that to manually change my 4DP numbers. Possibly a sub-optimal approach but if I mess up my own training I accept that it’s on me! There is “badass power records” in the passsport already so the app knows the numbers but I assume there is a #sportsscience reason they are not used to automatically update 4DP when you set a new record.
Like you I haven’t used TR and don’t really find it appealing for the reasons you say. I found Dylan Johnson’s recent video about it interesting: The Problem With TrainerRoad Training Plans - YouTube
big BUT at the end of this though is that I’m not a sports scientist so I really don’t know what I’m talking about! I have a plenty of faith in the SUF coaches who so far have helped me improve way more than I expected to!
It’s a good point. I think as you said (1) heart rate but also (2) I think you can correlate workout difficulty (seems TR developed a pretty sophisticated way of doing this based on how they impact different energy systems) to how likely someone is to complete it; specially based on previous completes / missed and your 4DP profile which measures exactly that.
Another way is to ask once you complete the workout how you felt and how ‘flat out’ you were to prevent excessive fatigue build.
It seem they clearly found a way.
Sorry for the last post, I answered from the phone and didn’t see the other posts
I did try TR twice (devoting quite a few months) and ad two main problems: (1) it is incredible dull indeed, even if you put your own media on the background, and probably more importantly (2) as Dylan Johnson points out, the training plans have an incredible amount of intensity which always led to me burning out after 1 / 1.5 months (and I’ve done a lot of intense plans in SUF and others).
Machine learning sounds like overkill here, more of a buzzword than anything else. (I say this as a software engineer who has taken a ML class, but who doesn’t work as a specialist in ML.)
However, I think there are simple heuristic-based improvements that could be very useful.
Here are some things the system already knows:
- Your heart rate and its relation to the targets.
- Whether you are successfully completing workouts at 100%.
- Your training load.
- Your age, which could be relevant to recovery needs.
Some additional info it could get easily gather:
- Ask you to rate how the workout felt on a scale, immediately at the end of a workout.
- As how fatigued you were on the day. I think this would also be best asked at the end, because I often can’t tell until the warmup is done.
If you’re consistently below the HR targets and you say workouts are easy, it could dial up the targets gradually. If you’re not able to complete workouts, it could dial down targets until you can. The four targets could even be adjusted separately, depending on the focus of the workouts and how successfully you did each type.
I would envision getting a notification upon completing a workout that this is going to happen for the next workout, and having an option to decline it. Also, the notification could say that it’s time to re-test. Perhaps the threshold for a re-test could be a bit bigger, and/or depend on the length of time since the previous test.
To add to your point, as someone who knows a bit about Machine Learning… for the purposes under discussion here, it really is just a refined way of doing statistical analysis.
The problem for me is that even with the known data, there is too much missing to build a model for ML.
There is too much missing to build a USEFUL model for ML. There are plenty of people out there building models for the sake of saying they have ML…
A few things …
(1) Predicting FTP: If I recall what was said during the TR announcement podcast, the ML-based FTP adjustments won’t happen until you’ve completed their ramp test (they added “progression levels” to help change the recommended workouts for you) at ~4 to 6 week training blocks in their plans. So no big difference in this regard from their current system, and not different from doing a Half Monty six weeks into your 12 week program. Also, this seems weaker than Xert’s algorithmic approach that updates your fitness signature after every workout, meaning your Xert-modelled FTP can change daily, in theory.
(2) Automatically updating your 4DP profile based upon bests from SUF workout history (badass power records) would be a bad idea as Full Frontal measures your power dimensions in relation to one another during the same test, in a set sequence, following a specific set of recovery periods, etc. To cherry-pick from your best power numbers without this context for your 4DP is just asking for trouble. I think the introduction of Half Monty (based on some modelling of The Sufferfest’s huge database of Full Frontal results) was a good move, providing way better bang for the buck than ML.
(3) “FTP is dead”: While machine learning seems to be a big piece of their approach, it seems to me that after years of investment in development and delaying sharing the value of their work with their users, they’ll end up in a similar place to The Sufferfest by creating a unique power profile that goes beyond just FTP for users to customize their workouts with, except ironically with even more (not necessarily better) power profile dimensions (5-6 “progression levels”).
There are plenty of bells and whistles and tech nerdery to get excited about in TR’s announcements for sure, but IMHO, the real value proposition for users doesn’t seem to have tipped the scale in TR’s favour … at all. I’m still pretty damn impressed by The Sufferfest, 4DP, FF and HM, and future opportunities to dial in my training while keeping me engaged through the process.
Thank you for making that explicit. Foolish me, I just assumed that when one said model it would automatically be associated with “useful”. Nonetheless, you are correct.
Hahaha, I wasn’t intending a rebuke of any sort! I’m glad a fellow nerd agrees with me.
For the non-nerds out there, there is a famous adage “All models are wrong, some models are useful.” All models are wrong - Wikipedia
Are we being a bit too fanatic here?
I love SUF as well but I think there are things to be learned from TRs new approach and I think it has little to do with FTP / 4DP and more with how plans evolve based on performance.
I don’t think we know enough to say their model is wrong or mot useful to be honest.