I hope that TR is working on having AI determine your recommended workout rather than the current AT algorithm. Don’t get me wrong, I think that AT is a great innovation, and accurate in many cases, but my experience with the two features (especially when we used to be able detect FTP every day), AI FTP detection knows a lot more about whether I can handle a harder workout or whether my PLs need to be lowered.
To give a quick example, I recently took 4 weeks away from TR workouts (one week of a huge outdoor training camp-like block followed by three weeks of sitting around). When I returned, AI FTP had my fitness down by about 4% - ok, makes sense- but on top of that, AT reduced my PLs down to nearly the bottom of the ranges. My next event is less than four weeks after returning, and I don’t have time to wait for AT to slowly get me back. Instead, I’ve been selecting significantly harder workouts that what AT wants (but more in line with where FTP detections thinks my FTP is), ramping up manually, and successfully completing them as my fitness rebounds.
On the other side of things, after reading around on the forum, it appears that despite AT, users are still running into too much intensity and burnout because AT can’t recognize that they’re at or nearing a plateau. AT has to actually break you to realize you’ve plateaued, and then it shifts you a couple of weeks back. In my (admittedly limited) experience with AI FTP detection, it seems to know that you can’t always handle an increase of one full PL (about 2.4%). AI should be able to detect plateaus before AT and reduce the chance of burnout.
I am a very loyal TR user and advocate. Both features are good innovations, but AI FTP detection is at another level. Applying the AI resources to the selection of your next workout is the next logical step in my opinion.
Your suggestion comes too late: both are using the exact same technologies already and are designed to work hand-in-hand.
One way to think of PLs and your FTP is to make an analogy to games: characters often have a life bar and a power bar. Your PL corresponds to your life bar, your FTP to your power bar. When you detrain, you lower both, your life bar and your power bar. Hence, AT and AI FTP work hand-in-hand, and have been designed this way.
Really? That’s surprising. How could an AI model accurately determine my FTP and then put my anaerobic level at 1.3 and my sprint always at 1.0 when I could do at least a 6.0 on either? I’d be shocked if the PLs are AI determined rather than formulaic.
I think the only thing the AI influences is that a change in FTP determined by AI causes a formulaic shift in PL.
AI FTP and AT are based on the exact same mathematical techniques, and have been developed to work with one another. Both use the dataset provided by the millions of TR workouts to work out e. g. how quickly FTP decays after 4 weeks of inactivity on average given someone of your training status, gender, age, etc. What data is included in the model I don’t know, that’s likely a closely guarded secret.
Likewise, AT determines the length of your life bar. It determines the best ramp rates based on your previous performance and responses to post-workout surveys. And it determines how Progression Levels change when your FTP changes. It is important to note that AT is designed to err on the conservative side, so the drop might be too severe. (Ditto for AI FTP, by the way.) That makes sense on many levels. If you mark workouts as Easy often enough, it will ramp up your PLs quickly enough. You can also override AT in parts, but I wouldn’t recommend that unless you really know what you are doing.
The difference is that AI FTP takes non-TR rides into account but AT doesn’t. So if you’re doing a lot of unstructured (or structured but not TR) training and racing your AI FTP will still be accurate but your progression levels may well be much lower than what you’re actually capable of doing.
Do you have some source for this? They’re definitely designed to work together, but that doesn’t imply the PL adjustment is AI based. It really seems like a set of simple rules based on the survey answer and time.
PL/AT definitely seems formulaic, as it literally says “reduced PL due to 2 weeks of no workouts in this zone”. That’s not the output of a machine learning model…
Yes, just have a look at the page @ChefAcB linked to, you will then get to this explanation, emphasis mine:
AI (= artificial intelligence), Machine Learning, Big Data and a few other buzzwords are colloquially used synonymously to mean the same thing: a collection of methods to systematically extract and model relationships from a big amount of data. For non-experts machine learning and AI can be used synonymously.
ML-based methods generate their own algorithms by solving another optimization problem. That is different from most (if not all) other automated FTP estimators out there, which are likely based on an algorithm that has been programmed by a human. (Technically speaking, there are some purely statistical methods out there which are used as well, but I don’t want to overcomplicate the discussion.)
So AI FTP and AT are both based on the same mathematical methods and the same dataset. @patrickhill asked in the title of this thread whether AI FTP is “smarter”. That’s harder to answer. AT currently does not take outdoor rides into account, which is important information. (The reason is simple: scoring unstructured rides is a hard problem, but one that TR is currently working on as their #1 priority.) In that narrow sense you might say that AI FTP (which does take outdoor rides into account) is smarter than AT. But honestly, I wouldn’t frame it in terms of intelligence, I’d rather ask whether AI FTP works more reliably than AT or not.
Seems so, but only seems so. Yes, it is the output of a machine learning model.
In the same way that humans and chess programs can play chess, these days chess programs (including ML-based ones) beat any human any day. Still, the activity is the same, humans can play chess, too. And PLs are mathematically no different than a (ML-based or otherwise) chess program to score a position and quantify the advantage that black has over white or vice versa. Progression Levels do the exact same thing.
Let me explain OP’s complaint it in a simpler way then: it seems that the model that adjusts PL is significantly (over?) simplified compared to the one that calculates the FTP. We know that lack of outdoor rides is one part, but there’s likely others from the observable behavior.
All the blabber of “using the same techniques” is a bit besides the point. A linear regression is also machine learning, but saying that it’s arguably the “same technique” as a deep neural network won’t lead to very constructive conclusions or discussions. It’s possible the PL adjustments are based off of finding good values by mining the entire TR dataset, but the model those values belong to seems…not good enough and barely distinguishable - if at all - from a very simple set of static rules.
For all we know the ML system just filled in the x, y coefficients for “scale the PL progression by x if the user answered easy on the survey, scale it by y if they answered hard”.
It would also fit the given description.
The FTP calculation stuff seems more advanced, hence OP’s question. I don’t find it hard to understand without going off in the woods about arguing that it’s “the same AI system and technologies”.
My own view of the PLs when they go down in number it is just a function of time. I dont know the history for the drops but I do find the drops are always too severe for myself. My FTP has been in a similar range for the last 4 years. I find VO2 workouts easier then any other workout. A few times I have just jumped back to stretch workouts for three workouts in a row just to get them back to where they make sense for me. I really dont know how smart the PLs are when they drop due to lack of doing a workout.
I just listened to the podcast from last week. @1:12:40, Nate talks about how the AT is informed by ML and engineers’ analysis. They never discuss that it is specifically applied to your PL or workout recommendation. It’s clear to me that AT is formulaic. This is a big difference.
I think AT is about as good as you can get without specifically applying an AI/ML model to PL setting and workout recommendation, but in my experience, AI FTP detection is a more accurate tool. I look forward to the day where ML models are doing more specific work.
Does anybody else have the experience where when they start a thread on this forum, the first responses are trying to find flaws and very definitively telling you that you’re wrong, rather than discussing the the point of the post? Maybe it’s just a forum thing. I don’t participate in any outside of TR.
It’s a TrainerRoad forum. Many people’s first impulse is going to be to defend TR. It’s no different on any other forum. If you go to a Ford forum and say, “Fords aren’t good”, you’re going to get negative replies. I’m not trying to defend it, and i wish it didn’t happen, but t’s just human nature. Go on a Classic Rock page and say “Led Zeppelin sucks” and you’ll get the same.
It’s very common. That’s why Murphy’s law states that the best way to get the right answer on the internet is not by asking a question, but by posting the wrong answer.
Very true. The funny thing is that I’ve never met a bigger fan or advocate of TR than me, and I said so. It’s kind of like how you only tease people you like. One of my hopes with starting the thread is to get TR to look at possible improvement, although I doubt they’re not already working on it in some way.
Same! I just bumped two threads today to show Ivy and the team that I’ve encountered issues. I do that out of love for the product and team, not angst.
