Nope, you aren’t.
Except that TR already has that running internally, so we are not talking about vaporware and I wouldn’t call it “future state intent”. It has progressed far more than that, it is real code that is being tested against real rides. It might be very hard to get it to production-level quality (a failure rate of, say, just 1 % will lead to a huge number of rides being classified incorrectly) and to work out edge cases, I don’t know the current state of the project.
You are conflating structured training with fitness assessment, I think: I can assess my fitness on unstructured rides or races no problem (without ML). The ML algorithms that rate unstructured rides only do a fitness assessment.
That is different from intent, which may or may not fit into the way TR structures training — unstructured rides are your responsibility and not all “unstructured rides” are devoid of clear training goals. Can an unstructured ride be unproductive for your training? Sure. But unstructured rides are your responsibility. But you don’t need to know intent to draw useful conclusions for training. And they may have benefits that are not visible in power numbers, e. g. practicing group riding on the road, descending or pacing over a certain distance, they all do not fit into the “percentage of FTP” mold.
For example, you wrote:
In that case I just expect AT to conclude that the ride or race was well within your capabilities, i. e. an easy ride for you. If, say, your ride was an 8.0, then it will use that data point to change your future workouts.
It is interesting that you mention chronic training load and fatigue. I’d add things like TSS and IF as well. These are nothing but an earlier attempt to quantify efforts based on power numbers by taking weighted averages of your rides. The weighting is arbitrary in the sense that they assume your capabilities decrease at a certain rate (and decay type) and your freshness increases at a certain rate. I think TR’s ML models have had much more systematic testing than that. If you think about it, it is an ancestor of what TrainerRoad is doing now with new technology.
The reason why many feel more at home with CTL, etc. is that they have more experience with it. We “know” that on certain occasions an IF > 1 is possible (e. g. on a short crit race) while on others it isn’t. We “know” that not all TSS are created equal and yet have used TSS to structure our training (e. g. to ramp up training intensity and then lower it during rest weeks). That’s just us being used to an older system that attempts to do the exact same thing.
That’s precisely what ML is about: you don’t need to know the equation and solve for X as you put it, because the methods are statistical. (In fact, there was a recent preprint by physicist Max Tegmark who did exactly that, searching for new phenomena with ML when you don’t know the fundamental laws.) You look for predictive factors and generate a model based on them. That’s why ML is hard even though the methods are rock solid: you need to carefully create that model and be sure that your training data set is representative, which is super hard. Otherwise garbage in, garbage out. There is a reason why it took TR >3 years despite having the largest data set of structured training rides in the world or why it was so hard to use ML algorithms to reconstruct images from the merger of two black holes. (The reasons I give examples from physics are not accidental, I’m a mathematical physicist, i. e. I am at the intersection of applied mathematics and theoretical physics.)
You have no idea what they are running internally. You are assuming a mathematical model from very tenuous statements. I find it unlikely that based on the current algorithm used that outdoor free rides will fit the same model and adapt workouts applying the same weighted value for free rides as is applied to workouts for all the reasons I have already specified.
No, I am speaking directly to periodized training. You can enhance fitness with or without structure. Repeatability and recoverability are enhanced through a structured training model. It has nothing to do with just building FTP or power curves.
That’s not what I said. I said you need to know intent to determine the adaptations which is essentially determining which workouts are achievable, stretch, breakthrough or not-recommended. You absolutely need to know intent to put that kind of context around a workout.
This is short sighted. You don’t want it to conclude the ride or race was well within your capabilities or outside of it. You want it to know, If I hit an 8 in a race, and crumble into a million pieces and am garbage for 3 days, I sure as heck don’t want TR to adapt up my workouts. However, if I fail a workout at a 6.5 and give it feedback that it was all-out (or essentially didn’t select something like equipment failure to have it ignore the data point) I would want TR to effectively hold that data and “see what happens” on my next structured workout. If I fail again, it’s probably time to adjust down. I am clearly working through fatigue or some other concern. This applies logic to levels and adjustments. What you are suggesting they are doing is having TR try to turn unstructured work into some sort of workout on the back end. Your argument does precisely what you accused me of doing - conflating structured training with a fitness assessment. You can draw useful conclusions from unstructured work, the conclusions however, do not logically lead to an adjustment of levels.
Machine learning is not about turning a multivariate problem into a univariate solution. It is short-sighted to think that ML can use one dimension of training and extrapolate that to create a working model that flies in the face of decades of training data and science. AT as it stands does what it does very intelligently and in a way that makes sense from a training perspective - perhaps not the specialty of a mathematical physicist, your choice to include your credentials are at best an appeal to false authority, and in truth, poor form. Human physiology and training do not fit into a math model. The closest ML will be able to get to doing so is to include multiple variables i.e. strain, fitness, fatigue, FTP/Zone modeling, etc. Selecting the right workout against a plan based on adherence/attainment has immense value and TR has damn near nailed it in beta, a few small tweaks could really make this groundbreaking in the industry right out of the box. Follow the plan, the platform learns how competent you are in zones and can adjust intensity days around past performance to ensure that workouts are within the bounds of achievability. If a workout is a stretch or breakthrough it is labeled as such and the athlete can adjust preparation and expectations accordingly. I know unequivocally without even looking at the code that TR has not unlocked the secret sauce that coaches adept at reading power files down to the second in WKO+ have not. The machine learning can grab all the statistics in the world, it won’t know what to do with it unless a human being provides the boundaries (i.e. actual vs target and level evaluation frameworks) OR it is given multiple feedback variables.
Feel free to take the last word here, I have expressed my argument in detail and I understand yours. We are too far apart and continuing circular dialectic is not something I am interested in. As such, I won’t be returning to this conversation. Enjoy your day.
Meanwhile I can’t get my progression levels to move from 1.0. In the beta for over 10 days and doing mostly structured workouts in the great outdoors. Even substituted a TR workout to try and coax an update levels. Patiently waiting for custom workouts on Garmin to move the level needle. I’ve been working with neural networks and ML for almost 30 years, starting under a couple Stanford PhDs. My proof of concept is still at y=0. Not worth debating the abstract at this point, I’m “from Missouri.”
This isn’t exactly correct. I don’t follow a TR plan, and many of my workouts in TR are actually custom versions, often tweaked or slightly modified/extended versions of official TR workouts. And, I’m still using and progressing nicely through AT… It’s more manually intensive, deciding my own readiness along with the progression levels and productive/stretch/breakthrough/not-recommended tags based on my performance. But, you don’t need a TR plan or to only use TR workouts to see the results. Honestly, I think AT is pretty amazing… and I wasn’t expecting to use or really be impressed by it at all.
And, all of this being said… indoor vs outdoor; structured vs unstructured; success vs fail vs max effort vs max/min required effort (i.e. in a race) … AT and ML is based on a model. And if you aren’t going to feed it, it’s not going to give you meaningful data. So, with everything in training and everyone’s utilization and application of our available resources, YMMV.
The key thing in your use-case (which sounds like it’s working very well for you!) is that you’re the one doing the plan adaptation, ie. you’re manually selecting your future workouts based on their Workout Levels informed by your current Progression Levels, readiness, etc., rather than the machine doing this workout selection for you.
For the “standard” use-case, it remains the case currently that only workouts which are scheduled into Calendar via Plan Builder will be adapted automatically (ie. receive adaptation suggestions). I assume that’s what Cleanneon98 was referring to: the current requirement to use Plan Builder rather than being able to just drop training blocks (or individual w/o’s) into Calendar directly, if you want automatic workout adaptations to occur.
I have a coach who I’m really happy with so unlikely to use AT as things stand, but I do think it has a huge amount of potential and I’m really excited to see where things go. Truth is, none of us really know the limits of ML, and while there’s certainly a practical aspect to deciding whether it’s of use to you in it’s current iteration I think it’s pretty short-sighted to dismiss an emerging technology entirely because it’s not 100% perfect early on in the piece. Proclamations of what is and is not possible generally do not age well.
I also find it kind of funny that we’re holding up human intervention as the gold standard here when we are ultimately pretty flawed decision-makers If anything this forum is evidence that a lot of us are kind of dumb especially when it comes to our own training (myself included), and even professionals will readily admit they are far from perfect. ML already has significant strengths over human analysis in other applications particularly where large datasets are involved, and it’s a pretty amazing tool to have at your disposal.
TL;DR: The future is now, old man.
In the future when AT is widely available, I reckon it could become a tool like TP where coaches use fatigue, fitness and form to inform their view on the athlete’s fitness. Having progress levels gives them more fine grained information on the fitness of an athlete. Just like computers have gotten way better at chess than humans, I reckon they could become way better at choosing the right workouts after getting some input from coach and athlete. Simply put, humans tell AT what to do and AT will help them achieve it.
I would just like to point out the OP and question is about if you are bothered about Adaptive training in it’s current state/iteration
Saying the AI/ML is the future of training, is an opinion not a fact, it’s taken a long time for ML to get to the point where developers can use it freely, but what tends to happen in software is that something comes to market, doesn’t live up to the hype, and is replaced by something else that is more adapted to the problem
Saying the we should jump onboard TR because it’s the future is a little short sighted, who bought the first electric car, because it was the future, was it any good (Just as TR implementation might not be any good) and it’s very rarely the first version that gains market dominance, Apple didn’t invent the mobile phone, Tesla didn’t create the first electric car, it’s usually the second or third version that see’s the problems of the previous versions and delivers what is needed (bit like what I just said about ML)
There’s a lot of information in these very public forum about the private beta, that competition would find very interesting
I’m inclined to agree that we should temper our expectations for v1.0. You are right that ML/AI/Big Data is not a panacea and there have been cases where they had been e. g. trained improperly. Figuring out good, predictive models from a vast pool of data is hard. Nevertheless, I am enthusiastic about it, because I have some idea how hard it is and because I think I see the potential.
But I think there is reason for optimism akin to electric cars: you know that the recipe is on the precipice of working. You see prices falling. You can now buy good electric cars that are just good cars, period. I see the development in AI/ML/Big Data, and have been following on the periphery. The big difference to before is that thanks in part to Google and Apple spending time and money to develop APIs and accelerator hardware. When my best friend wrote his thesis on (in modern parlance) ML-driven automated software localization in the late 2000s, there were no such libraries available. Things started to change when I met some smart people in the ML/AI/Big Data community in 2014 or 2015 during a big three-month program at the institute where I had my fellowship. These tools became standard in specialized industry (e. g. automated processing of satellite imagery to estimate the area of rain forest and funnily enough baseball), but mostly required an expert in ML/Big Data/AI to be used properly. Among many experimentalists ML methods are being used to analyze data. One big breakthrough was the observation of a black hole merger where ML played an essential role in the data analysis.
Now these tools are accessible more regular people and experts in other fields. You are right that we don’t know whether AT will be successful. But I think it is the right bet to make and I see all the pieces falling into place. That’s why there are more and more features on our computers and smartphone that we use every day are driven by ML. Put another way, part of the problem, the tools, have become easier to use and much more powerful.
Does that mean TR will get it right on its second try? Nope. But we can still root for them.
To be honest, while I disagree with the @Sarah ’s premise and conclusion, I think she still has a point: the classification of outdoor rides is one of the hardest problems to solve, I think. There is no structure (doh!), lots of edge cases, some of which will undoubtedly only become apparent when it is released at scale. You have to deal with erroneous data, for example, very different types of rides (road bike group rides, mountain bike tours).
I wouldn’t worry too much about that to be honest for two reasons: the first one is TR’s dataset. That’s a huge advantage that few companies can match. Yes, Strava, Wahoo and Garmin might have similar numbers of logged workouts, perhaps more, but they aren’t structured. The second one is that revealing a class of methods is probably not all that useful to imitators. ML is an umbrella term for a shirt ton of very different techniques, some coming from classical statistics, others rely on neural nets, etc. Even when you narrow it down, I reckon the secret sauce is that you need to filter out the relevant variables that form the input. I doubt all that will be replicated very quickly.
Just to clarrify, not concerned, but I will just point out that the reason given for not giving figures about AT was that they don’t want to feed the competition information about interest and if it’s worth doing, the information can be gleamed from this forum, just look at the number of posts and fun bit of regex over a topic to get the number of users, all information gleamed from a public forum
I just find it fascinating that people can form such strong opinions about a product which doesn’t actually exist yet, based on some forum skimming. The beta can’t do a lot of things which I’m sure the final product will be able to, whenever it’s released. Best I can tell the folks who don’t want to use the adaptive portion of the plans can just schedule what they want and complete it on their own as before.
I tried this on Sunday and I don’t think it worked for me. I was able to do the survey and the ride shows as being associated with the workout on the calendar but there is no rating at the bottom of the ride on the calendar like my indoor rides (ex. "Productive +0.3). There was also no change in my progression levels after the ride.
You’re absolutely correct. So, it’s a specific use-case, and yes… does require me to work around the limitations of how I’m applying it for sure. But, i think the component that is so great… really for everyone, is that AT and specifically the Progression Levels – especially once outside (relatively unstructured) rides are taken into account and have a retroactive feature that allows for rider contextualization about what the goals/demands were – is that it’s going to keep ya’ honest… if you haven’t worked on a specific zone/energy system. It’s gonna show it in your PL list, and in my case, has motivated me to be more well-rounded in my training in certain scenarios, and more targeted in others. The truly “adaptive” feature isn’t really for me IMO… so I’m potentially missing out and/or utilizing the platform outside of its true design intent. But, it’s been great so far for what I’m using it for… and as such, would just encourage everyone to find if/how/why/under what circumstances AT can be beneficial for them, bc it seems like the use-cases and applications are pretty varied and there’s gotta be a way for AT to help anyone who wants to use it for themselves.
I’ve only been on AT for 2 weeks, but i can tell why it has made a difference to me -
I’m on a LV plan (plus outside rides to work, social rides and CX races) to build up to a State CX championship by end of august, but in 3 weeks I’m about to spend a week hiking 140km with my wife. I’ve just put the hiking in as a travel break and it has recalculated my plan, so i’m not going to be hitting that first workout after a heavy week off the bike and blowing up or putting myself in hole because the progression is out of step.
It’s winter here is Oz so I’m also expecting to pick up a cold over the next few months. In the past it’s been really hard to adjust a plan after that happens, but I’m not worried now. It’s taken the stress and decision making away which is great.
That’s why AT is still worth it to me.
I agree. To me, the key innovation here is Levels (Workout/Progression), providing us - and/or the machine - with conceptually simple but powerful insights we didn’t have previously, and all that flows from those insights in terms of tailoring training. I’m super impressed with the foundations TR has built with this.
That’s absolutely right
Athletes will be able to opt in or out depending on their preferences.
Be sure to try and match a workout that is somewhat close to you time and type of workout you did. It doesn’t show it on the ride but I believe the computer is processing for adaptations as you go. It took a week or two but mine updates as I go.