This. It should be called ML at best. It’s more likely regression.
2 Likes
Here’s something to peak your interest. TrainingPeaks has had this capability for over three years now. People still complain about how bad it is IRL. For instance, your computed FTP(E) is 192 but your Total Time to Exhaustion is only 45 minutes. In reality, your FTP isn’t 192. There are issues with computed tests, such as Full Frontal, but for 99.5% of us, it is close enough to use. For the .5%, you would have to go to a place to have testing done, like the Wahoo Science Center in Boulder and spend thousands to be tested (any currency will do, USD, AUD, EUR, GBP). You would have accurate results. We aren’t professionals and we shouldn’t be that concerned if we are off by a few watts.
3 Likes
Careful now, you are going to start an argument about whether FTP is 1 hr power… 
AI FTP detection is a great marketing gimic for TrainerRoad user base who think everything can be reduced to data, but are relying on a single fitness metric FTP rather than 4DP? Their forum is full of people moaning that their VO2 max intervals are either way too easy or way to hard. On SYSTM with HM and FF FTP and MAP intervals are based on my personal physiology and are just right 
7 Likes
I can assure you that my MAP intervals are not too easy…
2 Likes
FWIW: My Garmin Edge will sometimes decide to spit out an FTP after a normal unstructured road ride. I have no idea how it works, but it usually comes within 5W of my FF FTP. I believe this is part of their Firstbeat Analytics, but I’ve never seen any detailed explanation of how Firstbeat works.
1 Like
Part of the problem is that there are many definitions of FTP, and it is being used beyond its original conception. It was defined as a relatively easy metric for cyclists to use to give a rough idea of where they stand.
Knowing my TTE at a given power level is very useful information. Where do you get your estimation that 4DP is useful for 99.5% of us? I have not found that it is very useful for estimating my performance in the real world. It is useful because it calibrates the workouts, and the workouts seem to improve my real world performance.
We really need to go back to cyclists developing a feel (RPE) for their riding, and not being driven by metrics.
4 Likes
99.5% is from Wahoo Science, not me. I, like you, are in the 1% of the 1% of being an oddball. I do FF and I can’t hit any of the numbers again. Some folks have complained they feel they are too easy. But I agree on “let’s return to RPE” as many have numbers wrapped to tight to what they are doing. The only two that I follow are Fatique and HRV. Both aren’t steering me in the wrong direction.
1 Like
I followed HRV for a while, but it did not tell me anything different from what my body was telling me. I do watch how long it takes for my HR to decline to normal after a workout. I do pay some attention to wake-up HR.
I would like to know where Wahoo Science came up with that number since they do not know what any of us does on a real world course. They do not know what state of fatigue or nutrition used on the various workouts and rides. Data obtained from elite athletes does not seem relevant to most of us.
Marketing gimmick. I can’t even imagine the bias on this considering who had to be in the training set. Can’t imagine it’ll generalize to my chubby, professional data nerd butt. I appreciate Wahoo sticking to testing protocols rather than going for things that sound cool in ads but don’t actually help.
2 Likes
I think it’s definitely possible to get a very good metric without a formal test, but it’s clearly only as good as the data and I feel like it needs one critical piece of input that is in FF but is not there in the data files themselves, and certainly not in other ride files. That is the effort level.
Imagine someone with data where they never tried harder than 70% of FTP for longer than 5 minutes. How could any intelligence ever figure out their FTP? At the very least any AI has to make some kind of assumption like that your best 15 minute effort really was the best you are capable of. If it at least knew your LTHR, but again, you get this from knowing something about effort, like in a ramp test where you must honestly go as hard as you can, and that’s an assumed piece of data. Once you know LTHR once, tracking relative changes in FTP over time is probably relative easy for an AI, or even a much simpler statistical algorithm. But it would seem that at some point, you need a data point where you know that the rider was at 100% effort, and that information just doesn’t exist in standard activity data.
That said, I don’t know why limited data should really cause serious overestimates, and if you’re pushing yourself, not just sitting at prescribed targets, it should eventually see what you can do. If an AI is overestimating FTP, in my opinion either someone gave it only sprint data, it’s poorly setup, or people are just interpreting the output of that particular metric in a way that’s not consistent with that particular metric.
1 Like
If you wanted AI or statistic to get an FTP metric compatible with FF you would just need to train the AI output, or correct the statistic, with/to FF-derived FTP’s for the same riders. A major difficulty though would be to what extent the activity data was itself influenced by prior FF results. If for example all activity data was in erg mode at exactly FF-prescribed levels, the algorithm may simply learn to estimate the previously prescribed FTP, and not the latest ability. The only way to even estimate latest ability would be through HR reduction. Again, this comes back to the need to assume that somewhere in the data, there must an all out unleashed effort, which anyway sounds a lot like doing an FTP test.
One of the issues I see with this isn’t even the accuracy. Let’s say you do all your workouts and your FTP keeps rising. Excellent. But you will have never done a maximum effort at that intensity to test yourself, to know your TTE, or how it feels. It doesn’t feel very tangible, but instead you keep saying my FTP is xxx watts because I’m told that’s what I should be able to do. But you’ve never done it. How do you know. And what happens then when you need to do that in a race/ride? Mentally you’re not prepare because you’ve never done it before, so the doubt can creep in. I actually feel the same to some extent about HM but am ok with that as it’s a mid way check in on things, but at the end of it all, you still need to do the FF to really know where you are.
3 Likes
full disclosure i have no experience with trainerroad’s AI FTP detection or whatever they call it but i have used various iterations of WKO. And i can confirm for you that, as long as you feed it enough of the right data, the outputs of the model will be a pretty good approximation of real world capabilities. I think one reason it works well is because it does not try to do too much. Like, it does not try to predict FTP from submaximal efforts or from single-best efforts in a workout (unless they happen to be spot on the right duration). These seem like pretty far logical leaps. Rather it assumes every effort is maximal, takes the best ones at each duration during a lookback period, and uses it to model a power curve. People get salty that it “understates” their FTP but really that’s a reflection of its strength, not weakness.
Couple of thins to note:
-
The real-world fidelity is only as good as the data you feed it. If you’re not periodically giving it max or near-max efforts at sufficient # of durations, you won’t get a useful output
-
that does not mean you are constantly needing to do maxes, just a sprinkling will do the trick when taken in conjunction with your workout interval data, which, is not max but generally “close enough” (think about The Trick)
-
Now of course, a gap in your model might mean a gap in performance (you’re bad at the duration) or an absence of data. but even the latter is actionable. E.g. someone (I think it was James) raised the example of TTE. TTE is notoriously squishy. So if you look at your model and it says TTE 35 mins, it should not be taken as gospel that your tte really is 35 mins. It could be that, OR it could be that you haven’t done sufficient longer continuous FTP and near FTP efforts for the model to know what your TTE really is. But the good news is, is that you don’t need ot know which because the right thing to do at that point is work on TTE (rephrased, either TTE is short or you haven’t worked it, in either cas, the answer is to work it!) Then you find out pretty quick what the real TTE is.
-
It doesn’t fully replace testing except in limited circumstances. 4DP and WKO have separate purposes. Even the coaches who designed WKO still use testing, except when WKO does replace testing, which is normally . . .
-
When you’re racing a lot and don’t have matches to burn on testing, then you can rely on WKO (and because you’re racing, you’ll likely have sufficient max efforts to give you a strong model). So we come full circle back to the exact use case the OP was looking for (which is fun)
So is modeled power data useful, absolutely yes. Should SUF build one, maybe, but not necessarily as it’s not critical for their training model and they are trying to keep things simple. OP, i’d suggest to keep using intervals.icu if it’s useful for you, otherwise consider giving WKO a shot. i think intervals is a little weaker (based on how it models) but it’s free and it’s probably close neough and if it works for you, that’s all that matters.
2 Likes
As I understand it (and I have played with WKO4 and 5 a little bit), it is very different from Trainer’s Road AI model, and Xert.
It is a descriptive model of where you are now. It makes no predictions. This podcast, with WKO5’s project leader goes into more detail:
1 Like
Another reason I like 4DP is this:
I created a custom MAP-focused workout in Zwift that I run in Zwift but record in SYSTM (long story that’s beside the point). Zwift bases everything on a standard percent of FTP. So, the first thing I had to do was to adjust the VO2 percentages up because they were too low for my MAP. That made it so the FTP sections hit my FTP and the VO2 sections actually hit my MAP numbers.
Well, I did a new FF and both my FTP and MAP went up. I ran my workout without thinking about it and found purely by accident than the FTP hit right, but the VO2 sections were too easy. Why? Because my MAP went up a higher percent than my FTP. So when I went back to do this custom workout all the FTP sections still perfectly hit my FTP, but the VO2 sections were now too low for my new MAP and hence too easy. So I had to go back in and re-adjust the VO2 sections up to meet my new MAP numbers. After that the workout felt right again.
And I had to do this to every custom workout I created (if I wanted to use them with my new numbers).
So, if TR or whoever can make an “AI” algorithm that can actually auto-detect my FTP, it still doesn’t really help that much. 4DP is one of the biggest reasons I stick with SYSTM (that and the whole Sufferfest thing).
I know that’s really a tangent as to whether the whole AI FTP detection actually works. But it does still help explain why Wahoo is not likely to switch over to some kind of algorithmic FTP detection. 4DP is more complex than that. And 4DP works a whole lot better, too. So if Wahoo/SYSTM ever does change to some kind mechanical detection, it’s going to need to do a lot more than auto-detect your FTP. And it will also likely include specific workout plans with specific workouts that will help test each 4DP value. But why go through all that work when the combination of HM and FF hit it pretty accurately as it is, and more accurately with better science and workout protocols than most?
Anyway… this message is my own opinion and was not paid for by Wahoo / SYSTM, etc.
1 Like
From a non-biased pov, because it is the tail wagging the dog syndrome that usually prevails, i.e. there are tons of athletes who do not want to stretch themselves with a physical test. Irony in its own right.
I reiterate what I said above, I seriously hope WahooX do not go down the AIFTPD rabbit hole.
2 Likes
True!
I honestly think the coaching team at Wahoo are far too knowledgeable for there to be much danger of that happening.
3 Likes
I have just come off of Xert (AI driven, no ‘FTP Test’ required!) & just starting Systm. Did a 4DP test & the results were within a couple of Watts. AI can work and saves having to prepare for and recover from more traditional FTP Tests. (PS, I’d forgotten how painful an indoor FTP test is!)
2 Likes
I think you are reading it wrong. Nothing is 100%.
I also use TR and have used the AI FTP detection for the last couple of months and it’s been dead on for me. Any bump in my FTP is fairly minor since I’m chasing marginal gains now but I’m able to complete my workouts as long as I’m fueled properly. It works.
2 Likes
Agreed. It is the same with everything, there are always a few people who never want to move on to better things. It happens in almost every sport, for example, bigger wheels in mountain biking. Innovation is the key to moving forward, and I think it would be a great feature to use in Systm. And if anyone wouldn’t want to use it, they could always just turn it off…
2 Likes