Workout Levels V2 update? [Unstructured Rides]

[WombleHunter]

@OreoCookie

@oldandfast

It’s a good think they are taking their time. It’s a hard problem to solve when you don’t have a model.

This. I think it would have been easy for TR to release something that sorta worked. I’d like them to ship only when they got it right.

I’m not sure I am totally onboard with this, or all the talk about data scientists, TR are not trying to fix string theory, in fact they created the answer, they created progression levels, so the answer can be what ever they want it to be, as it doesn’t have to fit in with anybody else’s “model”, they are having problems with the question, and have full control on both the question, and answer

So I would suggest that a lot of their problem lies with the fact that they don’t record workouts very well, an in a way that isn’t compatible with if the same workout was recorded on a external device, if you do a 6 x 3min V02 workout with 3 minute gaps in TR , and take a break in the middle of the V02 sessions, and then extend the rest periods, this does not affect your PL as you are given the score the workout intended, rather than what it achieved, but if you recorded the same workout on a Garmin would have 12 x 1.5 mins with 4 min breaks (for example) which would give you a different (and a data scientist would probably agree,) a more correct PL

Even if TR wanted to re populate previous PL’s they can’t do that as they don’t (seem to) record that information, which I’ve always found a little strange

My concern for TR is that it is neatly 2.5 years since Nate said they had a working version of outdoor workouts, and if it has taken this long to get it to beta, have TR built it on rocky foundations, although PL’s are great(ish) they aren’t really AI, and in that 2.5 years a lot have moved on, we have several apps from small companies doing actual AI training, Join, Breakaway e.t.c. TrainerDay are looking to beta test a ChatGPT based workout recommender, and are building a adaptive version of their (superior, if you buy into their training philosophy) plan builder, and Xert are promising “something market changing” (posted in this forum

And none of these companies seem to be looking at PL’s and thinking, hum … we gotta get us some of that action

*** Edit

All of the above is just speculation

[hpbieker]

The Xert app already has a few of the relevant components:

• fitness signature: threshold power, peak power and hight intensity power
• Maximum available power calculated every second
• Difficulty score that is calculated every second

The real time values can even be added to your head unit.

Would love to see something similar on TrainerRoad.

The current progression levels in TrainerRoad are also only working if you are able to execute workout exactly as prescribed. If you do an interval at a higher or lower intensity, the progression should reflect that, but it currently does not. You get the same score for a workout as long as you pass it.

[jz91]

@WombleHunter

My concern for TR is that it is neatly 2.5 years since Nate said they had a working version of outdoor workouts, and if it has taken this long to get it to beta, have TR built it on rocky foundations, although PL’s are great(ish) they aren’t really AI, and in that 2.5 years a lot have moved on, we have several apps from small companies doing actual AI training, Join, Breakaway e.t.c. TrainerDay are looking to beta test a ChatGPT based workout recommender, and are building a adaptive version of their (superior, if you buy into their training philosophy) plan builder, and Xert are promising “something market changing” (posted in this forum

And none of these companies seem to be looking at PL’s and thinking, hum … we gotta get us some of that action

This, 100% agree.

Is it fair to say then that TR has taken the wrong path? This is my fear to be honest.
I remain subscribed merely for controlling the trainer and because the calendar looks decent.

But in terms of training plans and flexibility it leaves a lot to be desired IMO. I’d definitely would have preferred TR to move in a similar direction as some of the names you have mentioned above.

Isn’t workout selection (aka PLs) something other apps (such as Join for instance) do already as part of a more comprehensive and flexible training solution?
I take that’s TR’s goal too, but given how long it is taking to implement a “simple” workout selector I fear it will take a very long time to get there.

[OreoCookie]

@WombleHunter

I’m not sure I am totally onboard with this, or all the talk about data scientists, TR are not trying to fix string theory, in fact they created the answer, they created progression levels, so the answer can be what ever they want it to be, as it doesn’t have to fit in with anybody else’s “model”, they are having problems with the question, and have full control on both the question, and answer

I don’t think your comparison works: Machine Learning/AI/Big Data is a collection of tools with which you can make other tools (algorithms). And the tool TR has built so far can only ingest structured workouts created in TR’s own workout creator.

@WombleHunter

So I would suggest that a lot of their problem lies with the fact that they don’t record workouts very well, an in a way that isn’t compatible with if the same workout was recorded on a external device, […]

I am in favor of recording more data, but I am not sure if this is the problem.

@WombleHunter

[…] and if it has taken this long to get it to beta, have TR built it on rocky foundations, although PL’s are great(ish) they aren’t really AI, and in that 2.5 years a lot have moved on, we have several apps from small companies doing actual AI training, Join, Breakaway e.t.c. TrainerDay are looking to beta test a ChatGPT based workout recommender,

I don’t think you are using AI correctly in this context, and everything you mentioned uses methods that is under the umbrella of machine learning/AI. ChatGPT is one particular type of machine learning model, a large language model.

A lot of companies are trying to integrate ML-based features into their products. It isn’t that hard to use some off-the-shelf algorithms. The hard thing is to train your model correctly, and this requires domain expertise (in endurance sports) and expertise in statistics and ML. Otherwise you produce an algorithm that reproduces your own biases or just garbage. You have to carefully select what you want to optimize for. E. g. if you only optimize for FTP, you’ll quickly get very skewed training plans that do not work as intended in the real world. I have no idea what TR does here, but that would be something they would want to keep very close to their chest.

Moreover, training ML algorithms requires a big dataset so that you have a statistically significant amount of athletes in all of your slices (e. g. gender, age bracket and choice of training plan). This is a huge hurdle for small players and TR’s dataset is basically the biggest in the business (when it comes to cycling). I would find it very hard to believe that someone just leapfrogs what TR has built given all of that. (TR has invested at least 5 years into this.)

Now I hope that TR gets some competition, but when I read that someone wants to beta test a large language model (such as ChatGPT) to recommend workouts, I just have to chuckle. The other day I saw someone ask ChatGPT for a recipe to make a steak. The text read like a recipe and technically you had a cooked steak at the end. It just wasn’t cooked very well.

My feeling is that a successful competitor will opt for a very different approach. FasCat’s OTSS and fatigue meter is very interesting, because it approaches the same problem from a very different direction and is based on very different data — data that TR (to my knowledge) does not collect.

[WombleHunter]

@OreoCookie

I don’t think your comparison works: Machine Learning/AI/Big Data is a collection of tools with which you can make other tools (algorithms). And the tool TR has built so far can only ingest structured workouts created in TR’s own workout creator.

I don’t think you have understood me correctly, I wasn’t talking about Machine Learning, I was talking about about everybody pronouncing that this is a difficult problem to solve (even though they have all the data, and the biggest collect of training data the world has ever seen), I was pointing out that this isn’t on the level of string theory as the solution doesn’t have to match a pre set answer, what a PL is, TR have definded that, they should be able to formulate the question to match the answer, the solution to the problem (ML or just straight code) is outside this

As for TR only being able to match structured workouts, it’s actually only match structured workouts that they have created

@OreoCookie

I am in favor of recording more data, but I am not sure if this is the problem.

Like I said, my post was just conjector, but the fact that OWO won’t match the same workout done in TR, (for the reasons I outlined), would be one of the reasons that I would guess that they are having problems (talking with my work buddies whilst writing software to do statistical analysis of huge datasets)

@OreoCookie

I don’t think you are using AI correctly in this context, and everything you mentioned uses methods that is under the umbrella of machine learning/AI. ChatGPT

I think you missed the point, I was pointing out

[WombleHunter]

@jz91

Isn’t workout selection (aka PLs) something other apps (such as Join for instance) do already as part of a more comprehensive and flexible training solution?

I agree, I had’t thought of it like that, I suppose they are doing something very similar, just not making a big deal about it

@jz91

Is it fair to say then that TR has taken the wrong path? This is my fear to be honest.

Oh absolutly, I have no information of TR long term idea fop adaptive training, so don’t really feel that I should be “wrong pathing” this, but usually if a feature takes this long to resolve (future not appliction) I would be questioned about what was going on, just interested for the outside

@jz91

I remain subscribed merely for controlling the trainer and because the calendar looks decent.

I like the calender, and is the main reason for “continued subscription” for now

[Pbase]

@OreoCookie

I completely agree with you that TR should be more disciplined in their communication or say honestly that “We don’t know when it will be released.”

I think they have done this many, many times.

@OreoCookie

Maybe scoring outdoor rides isn’t actually needed, but you might want to quantify fatigue incurred by an outdoor ride to adapt next week’s training. E. g. I would like AT to adapt my workouts next week taking into account that I did this long mountain bike ride in the Austrian alps last weekend.

100% agreed. They keep telling us people will be disappointed to find out how outdoor rides impact their PLs. If this is the case, let’s not focus on PLs then. Let’s focus on how “what we actually do” impacts the plan, and adapt the plan accordingly. This would be far, far more valuable. You could even make a simple statement that “if you want your outdoor ride to impact your PLs, then follow a structured TR outdoor workout, but know that this is not a perfect match and doing workouts indoors will always produce a more accurate result and response”.

@hpbieker

The Xert app already has a few of the relevant components:

• fitness signature: threshold power, peak power and hight intensity power
• Maximum available power calculated every second
• Difficulty score that is calculated every second

The real time values can even be added to your head unit.

Would love to see something similar on TrainerRoad.

The current progression levels in TrainerRoad are also only working if you are able to execute workout exactly as prescribed. If you do an interval at a higher or lower intensity, the progression should reflect that, but it currently does not. You get the same score for a workout as long as you pass it.

I could never get over some of what I saw as downsides to Xert, but the things you list were fantastic. I also loved having their data fields on my Garmin. I think TR could really benefit from implementing logic around “current difficulty” too. For example, we all know that Sprints are harder to do at the end of a long ride than they are at the beginning. A long interval at Threshold is harder to do after 2 hours of riding than at the beginning of a ride. These things need to be factored in.

@firemunki

There’s been plenty of other feature requests/ideas mentioned/stuff that could do with an update (having time limits for workouts when making plans, a modern calendar, better analysis, and a million other ideas).

These are the things I am waiting for. I have mentioned it before, but I was really disappointed to find out enough resources were focused on things like running that they had a big running launch when we were told long ago that the major focus was on outside rides and intelligent planning that allowed us to pick duration and type by day of week, add recovery weeks, etc. I would be more than happy to have our outside unstructured rides simply impact our plan from a fatigue standpoint, which other platforms have already incorporated.

[WombleHunter]

@Pbase

100% agreed. They keep telling us people will be disappointed to find out how outdoor rides impact their PLs. If this is the case, let’s not focus on PLs then. Let’s focus on how “what we actually do” impacts the plan, and adapt the plan accordingly. This would be far, far more valuable. You could even make a simple statement that “if you want your outdoor ride to impact your PLs, then follow a structured TR outdoor workout, but know that this is not a perfect match and doing workouts indoors will always produce a more accurate result and response”.

100%,

[Michael_D]

I hope V2 takes on board what you did on todays workout extra to or less than what was prescribed, then adjusts the next few workouts to compensate. Like a very clever version of what Join does and also Garmin suggested workouts do.

[Power13]

I’m just gonna say I couldn’t care less about PL’s….perhaps I don’t fully understand how they function within the TR AI, but it just seems like it is gamifying training.

Some find it very motivating (I like some of those aspects on Zwift), but for TR, I have no interest in it.

[Pbase]

I have to admit that I really like the gamification. I also really like that it tells me which rides are achievable, breakthrough, etc. Very helpful when picking a ride.

[Power13]

Please don’t take my comment as a criticism for those who do get benefit from it….it just isn’t something I personally care about in that aspect.

If it motivates some users (most, the majority, nearly all?), then mission accomplished.

[Pbase]

Yep. I totally get it. Mark me down as someone who doesn’t ever want to see banana peels in Zwift.

[Power13]

I mean they kinda essentially do that now with the anvil and the burrito, don’t they?

[OreoCookie]

@WombleHunter

I wasn’t talking about Machine Learning, I was talking about about everybody pronouncing that this is a difficult problem to solve (even though they have all the data, and the biggest collect of training data the world has ever seen), I was pointing out that this isn’t on the level of string theory as the solution doesn’t have to match a pre set answer, what a PL is, […]

I think you need to distinguish problems, which are conceptually difficult (e. g. string theory) and problems, which are practically difficult. Some problems are difficult, because the number of people qualified to work on them is tiny. Or that there is tons of competition and it is hard to attract those people, because Google et al. might offer them 2x or 3x the salary.

Just take one concrete example from physics: a few years ago, physicists observed a merger of two black holes. One key ingredient in this huge collaboration was a machine learning algorithm that reconstructed the complete image from a partial image. That part alone was hugely challenging even though it is easy to explain. One of the hardest parts was to make sure the ML algorithm isn’t biased and “wants to see” black hole mergers everywhere. So they fed it e. g. photos of Santa Claus and the like to make sure their algorithm wasn’t just reconstructing what the researchers wanted it to see.

Point is that even things which seem easy in theory can still be super hard in practice, and you shouldn’t underestimate the difficulty that goes into something like Adaptive Training. Friends who work in the field told me that many are simply lacking knowledge in statistics. (If you listened carefully to the introduction of AT by Nate, you can hear some key words from statistics.)

[WombleHunter]

I think you are missing the conversation for worrying about the words

Anyway, have a good one, as I no longer subscribe, I no longer have an opinion

[OreoCookie]

@Pbase

I could never get over some of what I saw as downsides to Xert, but the things you list were fantastic. I also loved having their data fields on my Garmin. I think TR could really benefit from implementing logic around “current difficulty” too.

This is clearly a weakness of TR’s apps: they do not seem well-integrated into the relevant platforms such as Garmin and iOS. You have no Apple Health integration, for example, so they have no access to sleep data for many of their athletes.

@Pbase

These are the things I am waiting for. I have mentioned it before, but I was really disappointed to find out enough resources were focused on things like running that they had a big running launch when we were told long ago that the major focus was on outside rides and intelligent planning that allowed us to pick duration and type by day of week, add recovery weeks, etc. I would be more than happy to have our outside unstructured rides simply impact our plan from a fatigue standpoint, which other platforms have already incorporated.

This one is very tricky, and I can see how TR arrived in the position where it is today: they had to allocate their limited resources to a few select projects, and scoring of outside workouts was likely a top priority for a while to the detriment of other features. It probably felt “almost done”, so they invested more time and person power, gave it one more concerted push, they got a bit closer, they pushed some more, … Sprinkle in some sunk cost fallacy and I think this could be how they got to where they are now.

Project management is tricky. On the plus side, now I see signs of TR choosing new features (e. g. the calendar), running, the launch of the ramp test free future. They have clearly turned their attention to other things.

[TooManyDogs]

@Pbase

I also really like that it tells me which rides are achievable, breakthrough, etc. Very helpful when picking a ride.

I like this aspect of it. I’m not chasing PLs but if I’m choosing a workout I like some guidance on what level of difficulty I can expect. If I’m not feeling great I’ll just choose “Achievable” feeling okay or really good I’ll choose “Productive” or higher.

[cyclhist]

@Power13

@OreoCookie

Maybe scoring outdoor rides isn’t actually needed

Except for many TR athletes, there is not much time spent inside during the season, so for AT to really function properly, it needs to take outdoor rides into account.

This is confusing. When PLs first came out I did a bunch of outdoor workouts on an old airport in Berlin and they definitely affected PLs. As far as I understand it, they do apply to outdoor workouts but only if they’re your TR workouts.

If you’re not doing TR workouts, I don’t see why that needs to affect progression levels. It’s a “nice to have” but back then they said that it doesn’t matter because the next TR workout you do will affect them again. If you’re not following a plan, then your PLs will deteriorate but then you’re not following a plan.

Did this behavior in AT change?

[Power13]

@cyclhist

As far as I understand it, they do apply to outdoor workouts but only if they’re your TR workouts.

Correct…if you do it as an outdoor workout, it counts towards PL. However, there is no evaluation of the actual data (as I understand it). It is simoly a “pass / fail” rating. If you say you did it, you get credit for it.

@cyclhist

If you’re not doing TR workouts, I don’t see why that needs to affect progression levels.

Because outdoor rides impact your fitness and ability to handle different workouts. If you exclude that data from the set, then the AI is getting bad / incomplete data upon which it is making recommendations / adjustments. So you end up with workout suggestions that aren’t appropriate.