As some one who used to build super high fidelity models (defense sector), all of the current estimated FTP models suck. ALL. By suck, I mean that none of them are predictive, plus they just aren’t accurate. Take any model, and look at its predictions for your:
5 second power
30 second power
1 minute power
5 minute power
8 minute power
20 minute power
etc.
I can guarantee that any model you choose, will be lucky to estimate 2 of the above within 5%. Which means the model isn’t accurate and should be rethought. Further, if I’m doing sweet spot training, all models will show that my FTP is decreasing. Is this true? If not, again model fail. If I need to do specific all out efforts, then the model isn’t predictive, and isn’t better than looking at an actual power duration curve which has those same all out efforts.
I make the above point because when you start talking about “adaptive training”, you need to define ahead of time what are you adapting the training off of?
Additionally, there are some core assumptions that use total swags (e.g., CTL & ATL decay time periods). The current 42 & 7 days might be the correct population average, but that means for any specific athlete they are almost guaranteed to be wrong. So again, if you are using these as part of a predictive model for adaptive training, you are feeding your ML algorithm with garbage. I’d first like to see a predictive CTL & ATL decay rates. I have 7+ years of training data in TP (TR has the same), so why can’t either / both calculate these for values specifically for me?
The Firstbeat model seems a bit better, in that it doesn’t rely on max efforts. But jury is still out on that. I’ve got a similar background, been building models and supporting modeling tools for over 3 decades now (including working with neural networks and ML way way back). I’m playing around with WKO4/WKO5/Firstbeat right now, good times
I think the challenge with getting to the higher levels in your list is that you will need a lot more data sources than just what somebody is doing on the bike (even assuming you have all the data on what they’re doing on the bike - I can’t quite remember when TR started pulling in outdoor rides but presumably before that you only know what they were doing if it was a TR workout?). You’re going to need sleep, diet, life stress, body composition, etc.
E.g. If I look at my data over years, the times when I was strongest correlate at least as much, if not more, to how well sorted my life was outside training as they do to what training I was doing. Plus some training simply can’t be done unless you’ve nailed the recovery side of things. I’ve had periods where I was doing 7-800 TSS per week and thriving on it because I was getting a lot of sleep, eating well, not travelling, etc. And there have been times when I’ve attempted that kind of volume when I’ve a got of other stuff going on in my life and I’ve just dug myself into a hole. On the last podcast Chad said some people would likely be better off cutting back on training if it enables them to get 8 hours sleep and that rings very true with my experience. I don’t see how ML/AI is going to figure that out unless it has some measures of what else in doing in my life. The good news is that a lot of us these days are wearing sleep trackers, tracking nutrition, HRV, resting HR, using smart scales, etc. so a lot of the data is it there. I hope/assume that you guys are looking at ways to start sucking in those data sources as well, as they’re equally if not more important to how many watts somebody is pushing for how long each week.
On one hand I think there is more information about you in TR than you’d think. And on the other hand, maybe less is needed. If you don’t hit your training targets, the answer is nearly always to back off. It doesn’t matter that much why you can’t hit your targets, for future workouts you’d need to go easier. It’s a lot like the advice you get on this forum - workouts too hard, for a few workouts in a row? Lower your ftp. And, the other way round - if you’re often increase intensity or ace every workout - time for a new ftp test. Same if your outdoor rides come back with IFs over 1 - ftp is likely wrong. A more detailed analysis would look at the type of workout you’re struggling with, and draw conclusions about vo2max/ftp - you see that on here with ‘can’t do vo2max’ or ‘struggling with sweetspot’.
With regards to ‘TR knows more about you than you’d think’ - have you missed a couple of workouts in a row, but then returned to normal? Probably a work trip. Have you missed a couple of workouts, and then returned slightly weaker or changed your workouts down? You’ve been ill. Are you changing your training times a lot, miss and reschedule workouts, or frequently cut them short (despite hitting the intensity) - you’re busy and probably stressed. Or just the timing of your workouts can say a lot - very early morning vs during the day, weekends etc.
Yeah some basics like time of day v type of workout v compliance rate could be really enlightening for a lot of people, especially people who’re able to tally that against other metrics like sleep and diet/fueling.
But I’m thinking the real magic isn’t done pre-workout with AI/ML - it’s during the warmups or first interval. Really good software could reconfigure your next intervals by looking at your current form, and not simply reduce/increase power, it could totally remap the entire workout to something more suitable.
So this is my 2p. Take it with a pitch of salt, I merely DABBLE in data science.
So all these companies, Strava, TrainingPeaks and TrainerRoad, SufferFest (but mostly strava), now have VAST treasuretroves of data, start running it through tensorflow and I’m sure that it’s quite possible to create more accurate workouts. Imagine what you could do with all the data points say if you are garmin and you have even resting heart rate (so you can roughly see how people are dealing with recovery etc).
Imagine when you have access to be able to see how pro’s are training and even though they can’t see, you have a CRAZY amount of datapoints on each pro, each workout, if they have a smart watch even rest time too. Throwing AI at the problem could well be a way to then streamline and figure out what works best for who.
The future is wild and crazy. It’s exciting. I do wish sometimes I had access to such datasets as I’d breakout Hadoop and Tensorflow in a second. I bet even now it’s probably possible to predict which starters are going to be on their way to being pro if the data is there (I mean motivation willing)
Reminds me of the podcast with Steve Neal that’s been discussed recently, where you measure improvements by riding tempo intervals at a limit of 83%maxHR, and seeing if you can put out your target power for longer before you get to the HR limit.
If I did a couple of those a week, 1 longer endurance ride, 1 vo2 burst intervals session midweek, and a 1-hour crit on Saturday, I’d probably retain and maybe even gain a lot of fitness. But because crits don’t have max efforts longer than 30 seconds, the tempo rides are all sub-threshold, and the vo2 burst intervals are also sub-maximal individually, then a power duration model will struggle to capture my progress.
What I’ve always wanted, and I don’t know if this is part of what is planned, is some kind of intelligent system that will tell me what power to do Spanish Needle and other anaerobic workouts, as the ramp test definitely doesn’t get those ones right for me.
I think many are setting the bar a bit too high for the adaptive training. It will never be perfect for every athlete at all times. It will never know everything meaningful about your training and life in general. Neither will any coach be able to develop the absolute perfect training plan for their athlete. But surely adaptive training plans can be significantly better on average for most athletes than the current model in which you just choose a plan and set an FTP. That’s all that is needed for it to be worthwhile. If adaptive training plan would have 40 % chance making me 1 % faster, 20 % chance of making me 1% slower and 40% chance of no effect, I would take it.
I think adaptive training plans would be especially useful for people new in the sport but probably not so for people who already have a ton of knowledge. It could be just in the form of a virtual assistant that makes suggestions when it sees something that is not normal. E.g. “I noticed that you haven’t completed the last two workouts, shall I lower your FTP or give you more rest before the next workout?”
Xert is actually really good. The reason you don’t plan out a month of workouts is because it is adaptive! if you miss a workout, Xert adapts. Go on a smashfest rather than a recovery ride, xert adapts. Xert’s problem is that it is looking at this form a different angle. The norm that we are all used too is prescriptive training based on FTP -people think they can’t miss a workout or it all goes to pot. feeling tired…can’t take a day off. Feeling strong…oh the plan has me down for a recovery ride. Thing is, people aren’t static and neither should training plans be. Yes the Xert UI could do with development, but it is substance over style. I followed Xert’s adaptive training adviser for 12 weeks. made bigger gains then any other training plan…(I’ve been 2 years on TR) across all points of the PDC. Now TR will get you stronger, I’ve made good gains…but 3 years in it’s all getting a bit stale. Xert has taken me off that plateau. i just think because Xert is a new way of training, using unfamiliar terms, people aren’t comfortable in it or want to risk taking the plunge. In the not too distant future, training software, will be adaptive and just as importantly will focus on you as an athlete rather than prescribing intervals on the same FTP % for all athletes. Xert’s smart interval workouts (workouts that adjust in real time based on your effort) are incredible. their workout builder that allows you to build workouts not just on static FTP % but off your MMP and even by strain scores are really excellent.
Yes I get that and Xert have stressed it a number of times, but honestly it’s more to do with processing power and workload at Xert’s end.
There is no reason at all why an adaptive planner wouldn’t create a plan based on your current situation. You give it hours and days and it will fill it in based on your current situation. It would of course adapt the plan as you went along, but it wouldn’t alter significantly unless you significantly fell off track.
There is great comfort and enjoyment in looking ahead at your workouts. It makes you commit and it’s incredibly interesting to see the workout progression. Xert doesn’t offer that because it’s workout library is second rate.
Xert has made inroads here, you can spam the calendar now and get a plan build that way, but it is woeful looking. You’ll get the same workouts repeated on end. Forget the names now but it will literally repeat LSD 3 or more times a week depending on your settings. It’s too scientific and misses so much because of it. Worse, there is no explanation on the planner as to why it’s doing what it’s doing. The human element is definitely still needed.
This is how I see it as well. Adaptive training will be good for steering people new to training away from common mistakes, but for those with a few years training experience I can’t see it making too much of an impact.
Most of the experienced TR users on here aren’t blindly accepting the ramp test FTP and following plans to a t. They’re adjusting FTP, modifying plans, adapting workout intensity based on how they’re feeling etc… Best case scenario, the AI/ML is implemented so well that it automates what we’re already doing (I think it’d take years to get close to this point). Worst case, a ton of development work goes into this and we’re no better off than we are now.
It’s definitely something that’s worth researching, but I’d hope it’s more of a side R&D project rather than something that’s taking the main development focus.
I didn’t get on with xert’s smart workouts. That was about 2 years ago now, so they might have imrpoved. But back then, the workouts ended up being way to easy, especially as the recovery kept increasing if you didn’t drop power really low. I ended up getting off the trainer for a bit to make the recovery pass quicker.
Nah, I think it was because I did lots of outdoor riding, but didn’t have a power meter at that time (just power on my trainer), so xert didn’t get the full picture. I did like the different scientific approach to training, but in the end I wasn’t convinced it was working as well as they claimed, so stopped using it.
The workout library in xert is excellent actually. Again just doesn’t look great on initial view. The fact that the colour coding is based on fatigue rather than just zones so it shows how you will “feel” during the interval is actually better than anything out there at present. The reason you may get 3 LSD is the time that you are putting the plan together. Xert moves you through phases of training based on your selected athelete type. Or you may be in a tired state. You can actually adjust this. The human element is definitely still needed, but surely this is a good thing? Sure in the future we will have more accurate models that take into account sleep, stress, work etc etc but we aren’t there yet. I totally agree that xert has shortfalls. Again, you have to take a leap of faith. It is different to what is currently available and that will put people out of there comfort zones. It has taken me about 18 months before i fully trust it. i can only give my own honest opinion and for me whilst it may not look the best, it offers substance over style. I feel it knows me as an athlete, the workouts because they are based on my fitness signature rather than an arbitrary FTP figure are more beneficial. I’m not saying it is better or worse than any other offering…all have their limitations, I’m only saying what works for me.
Being the kind of person who knows on Sunday exactly what they’re having for dinner every single day until the following Saturday, I’m very much of this mind too.
We don’t like this approach. You’re assuming that people always give 100%. There’s lots of research that show that if people are given the option to go easier they will.
IE if you know if you just go a bit easier the workout gets easier lots and lots of people will be more likely to go easier.
.