Xert is not AI. It is all based on a mathematical formula/algorithm modeling your specific power profile.
3 Likes
The strength of AI is finding and quantifying obscure and complex patterns and relationships in large datasets. Why try to detect something so crude as FTP? What weâd like to characterize is much more complex than that. AI could do better, identify more comprehensive ways of characterizing performance. (4DP is a tiny step in that direction).
1 Like
As noted elsewhere Xert does not use Machine Learning. Single cases can just be coincidence. Nobody publishes (SYSTM, Xert, etc.) their data sets, exact methods, etc., so no evaluation can be made of their general usefulness.
In my case, I stopped using Xert because I did not find it useful.
1 Like
You are assuming all these new approaches actually work. Nobody knows that.
I could also make the general statement that there are plenty of attempts at innovation that never work.
Thatâs fantastic for you. But one swallow does not make a summer.
I have yet to find a steady stream of TR users who share the same sentiments as you do. Other than AIFTPD not being accurate. Which in itself should be a red warning light.
I would not like to be training against power targets that are set 6-10% higher than they should be.
2 Likes
Me neither, but then there are some that like to seek out workouts that break them every time for whom a vastly over-estimated FTP might just be the solution they seek 
2 Likes
Yip there are those some that like to do that, but SYSTM has that covered by consciously and by design, allows one or some to alter the ride intensity in the settings menu of their chosen workout.
I donât want to be training unknowingly to inflated (inaccurate) numbers as a design to the SYSTM system. 
What gets me though, some use (train I assume) TR yet want these changes on SYSTM. Donât they have faith in the design of TR workouts? 
1 Like
Daily stream of not so positive remarks to AIFTPD.
How bad is a new feature when users are tentative to use it for fear of not completing their next workouts? Is that because the new feature overreads/is inaccurate/is not to be trusted and as such impacts the training progression of a user.
It is this user noise post an AI feature release which I fear will be unnecessary to the SYSTM ecosystem as it is just not well accepted.
The simple fact of using AIFTPD starts the notion of wanting things to be easier. An easy FTP test flows into easy workouts. Losing the impetus of needing to do physical FTP test starts to, IMO, create a bunch of bubble wrap users.
3 Likes
The real-world fidelity is only as good as the data you feed it. If youâre not periodically giving it max or near-max efforts at sufficient # of durations, you wonât get a useful output
The problem I have with this is the Sufferfest training model doesnât allow for sprinkling in max efforts here and there. It is based on prescribed efforts, most of which are not max, and all of which are pre-determined from your previous FTP test. Many people are now using erg mode too, so they will never go above that. So how can any algorithm learn what your new FTP is if all it ever sees is power that was pre-destined and fixed on your old FTP? Actually I think itâs possible to get it from trends in HR vs power to infer that recent activities were at reduced effort, but I donât know if any algorithm is presently that clever.
It used to be that several activities were FTP tests, and for normal training they were meant to be lowered by 5 or 10%, as the plans used to call for, but thatâs less true now and less consistent. In those days you could judge FTP pretty well just by trying to complete Thin Air or HHNF at 100%, or a 103% if you were feeling it was time to increase the pain. That was still putting in the effort to do the testing, but it was more fun.
1 Like
These are good points, though my personal counterpoint was that attempting 2015/2016 era Sufferfest training I always burned out quickly from too much intensity. As Neal & team say, âthe best training plan is the one you stick withâ which for me is the current incarnation.
Iâd also suggest that a number of the ProRides push you into âall outâ territory (though they are tricky to do sans-Erg mode) and RGT racing is another potential substitute so I see that as recognition from the team that theyâre not neglecting these efforts even if the focus has shifted for Sufferfest workouts. Iâm planning on trying a bit more RGT racing this winter and I hope other folks trickle over so thereâs more robust fields.
1 Like
Hi,
I donât think sufferfest is differentiated in this regard. There ARE opportunities to go max or near max (eg The Trick, Omnium) and you can always choose to push it on some workouts (thereâs even an article on the blog about it from a few years back). Second all training systems worth their salt have you doing prescribed efforts most of the time, not âall outâ. Finally, modeled FTP calcs donât need to fully replace testing, they just provide addl actionable data.
But what you say does drive home the original ping I made which was that, in my experience, best use case for modeled FTP calls was the racing season, bc you not only get plenty of maxes, but also have less time and energy to test.
Again though, I only have experience with modeled FTP / power curve based on your power data, not ML. ML like TR stuff may not have the same limitations, but Iâm sure it has its own limitations . . .
2 Likes
Yeah, Iâm not saying sufferfest is different in that regard. Iâm not sure how well the trick and omnium let you test FTP. In the end because of the repeated efforts, thereâs a big overall FTP component, sure. (and honestly the issue here is FTP, anyone can easily find out their 30s and 2~5 min power, even strava data on local hills gets that pretty good when you know you went for it, even without a power meter, because of the physics of hills). ML has the potential to take some information from heart rate reduction, but it still has to find a peg somewhere from some best efforts. If it sees HR reducing though, it at least knows (or could know) you can increase power enough to compensate that back up to what it had seen. If that was max effort or not in the first place is another matter.
Theres no suggestion that the Trick or Omnium give you an FTP measurement. The suggestion is that they contain 1min efforts in a fatigued state somewhat similar to that in FF, so if you are wanting to avoid FF and do just HM, say, than you can get a reasonable AC metric from one of those two workouts.
1 Like
Yes, the modeled FTP systems iâm familiar with need max or near max efforts at a few different durations to build the whole power curve and from it, derive FTP. Meaning, even including very short efforts that are not themselves reliant on FTP. Plus as leebo notes, they start to build out a model of your anaerobic capacity (FRC) etc.
However Still_D, itâs a good point that for it ti work best with Suf, youâd still need to either test long durations or occasionally push it in longer durations. i know suff doesnât have a ton of traditioanl FTP style intervals (like 3 x 20 or whatever) but there are some workouts that i expect would come close.
Iâd love to learn more about how the TR one works though. It sounds more like a true machine learning model vs. WKO, Golden Cheetah or intervals.icu.
i know suff doesnât have a ton of traditioanl FTP style intervals (like 3 x 20 or whatever) but there are some workouts that i expect would come close.
There are NoVid workouts that do have these and more! However, riding at FTP will NOT improve it but improve your endurance abilities (Hunter, et. al. have long articles on why this is true). You need MAP and AC workouts to improve both your FTP and VO2Max (they work hand in hand for some strange reason).
Oh man this has suddenly turned into a CONTROVERSIAL THREAD 
1 Like
I took a look at my power curve over different chunks of time; month, year, periods with more riding, or less riding. Combining indoor and outdoor riding. My data point of one shows a very tight correlation between ramp tests and my actual performance and a looser correlation with FF, but still well within usefulness boundaries. To me, this means that WahooX testing does a pretty good job of being close to your actual performance for that time period. I say pretty good as no model is perfect, and canât really account for âhow you are feelingâ during any particular chunk of time or on a given test.
I donât know TRs ML algorithm, but it seems reasonable that with enough data on your riding, their algorithm should provide something useful to guide training.
Sorry, the data scientist in me who models for food when not suffering escaped. Will try to do better in the future. I didnât even go into A/B testing.
1 Like
Actually, if the ML program has never seen the data, it cannot predict it. This is the reason that ML weather models do not predict 100 year storms â they have never âseenâ them.
1 Like
Heretic: Youâre right, itâs hard for a ML model to predict something with a small dataset. Itâs especially hard when that dataset has a lot of variability or is non-linear, which is characteristic of weather modeling. FTP prediction appears to be a linear model, with the data points having less variability, so the model needs less data to be trained. Does this mean a ML model with 50-60 rides over 30 min should be able to predict an FTP value? Sure. Does this mean that is your FTP at the moment? Thatâs a bit harder to say. However, it should be good enough to set some training thresholds. At the end of the day, I just like riding my bike.
1 Like
The variability with FTP prediction is that the data set is composed of many riders, each with their own peculiarities, and the responses changes with age, fatigue, fueling, weather conditions, etc.
2 Likes