🎉 🎉 🎉 Introducing Adaptive Training! 🎉 🎉 🎉

[D-9]

You can’t do that on the Android app.

[bobw]

A couple of thoughts/questions came to light.

Currently in the workout notes I record how much sleep I had the night before. So something like: 8hrs good or 6hrs interrupted. Obviously it’s freeform so not easy to parse all the workout notes and pull in that info as everyone will record it differently if at all. Could you add a couple of fields to the post workout questionnaire (both success and “fail”) for sleep duration and quality. They could just be drop downs. There’s likely to be many people who either don’t have a sleep monitor or don’t wish to use one, this could capture that information for those cases.

Periodisation - the plans as implemented at the moment, including Plan Builder, don’t really have an ebb and flow to them - they are pretty much geared towards always progressing whereas if you’ve done a big race/event or your target event if that’s different you need some downtime which depending on intensity and duration of the event could be a couple of weeks before gently starting back on progressing to the next race. When my main sport was climbing I’d basically take October/November off - I wasn’t doing nothing, any climbing I did was very easy but mostly I’d be out hiking getting my body used to the upcoming rigours of winter/ice climbing. It did mean that in 30 years of climbing I had very few injuries and zero burnout.

Finally does the ML avoid the equivalent of the social media echo chamber? I.e. you have expressed an interest in X therefore you might like X` and you’ll never be shown Y. I’m very much endurance biased but still need VO2max work, it’s a definite weak spot of mine, so I wouldn’t want the system to simply determine that I’m good at endurance so I’ll be recommended endurance, I can do that for myself!

[BCrossen]

Two things that I think address this based on all the communication and information they’ve shared:

1. Every workout has been reclassified in accordance with this new system of levels within energy systems. The relevance to FTP is taken into account, but is not the end all be all, particularly as the ML learns over time. Nate mentioned numerous features are considered when determining the difficulty of a workout, and those give it a prescribed level of difficulty (from 1 to 10) in the relevant zone. So what the ML feeds you is based on that new system, not based exclusively on %FTP (at least that seems to be the message they’ve stated).

2. Nate also mentioned that for people who are extreme outliers (which I would say you would be if your VO2 is 150% of FTP; note: I’m jealous as 120% is too much for me), the ML will likely have to adjust the levels to tailor to you, specifically. Here’s where Nate talks about the challenge that outliers pose for the ML: Announcing Adaptive Training: The Right Workout. Every Time. – Ask a Cycling Coach 298 - YouTube. I think he talked about outliers a couple times in the podcast, but the timestamps aren’t that narrow in scope.

[bazcurtis]

I have just caught up with podcast and can’t wait to try this. I have always appreciated that @Nate_Pearson knew Trainerroads value was their data and used correctly would be the basis of making it better. I heard a great quote the other day, data is king, but context is queen. Nate explain this well on the podcast.

@Nate_Pearson mentioned getting more data like sleep and steps. Couldn’t the new iOS app use Apple Health for this. I understand it wouldn’t cover everyone, but that data is held there and could be allowed to write to Trainerroad. It would handy to do weight that way too.

It is great to see @ambermalika starting on the machine learning model for training while pregnant . Congratulation @ambermalika

[SeanHurley]

The short answer is that if your FTP is set correctly, there will be a progressive workout for you! The ML is not only looking at % of FTP as the determinant of difficulty in a workout- things like interval structure, rest interval length, etc are super important. It’s extremely unlikely that your FTP and capabilities in a certain zone would be so dramatically decoupled that you’d be a 10 in one progression level and finding those workouts easy, while barely passing lower progression workouts in another zone. And in the few cases where there are holes in a certain progression within our workout library, we’re working to fill those in with new workouts.

[Nate_Pearson]

@gwhilts

@Nate_Pearson, I wish my TS and React skills were up to snuff, because this would be exciting stuff to work on, and y’all sound like a great place to work. Is your back-end and ML stuff in TS too?

It’s a mix of stuff. But I’m not really up to date on what they are using.

[Nate_Pearson]

@enki42

Too bad your optional is where I’m strong at. (C#)
What are your thoughts on part time workers?

No part time workers as it’s hard to make them part of a team when they are only working part time.

We were going to put a C# job up but we filled it internally first. We’ll have more C# jobs in the future though.

[Nate_Pearson]

@GregElwell

Sorry if this is a basic/daft question, but am I right to assume you would need to be on a TR plan in order for AT to work?
And if you write your own plan, or train outside, or whatever outside of a TR plan, then you could get the Progression Levels bar graph? Because that is a retrospective look at your performance/ability by zone.

If you’re not on a plan, then only thing you don’t get is the actual adaptation in the plan and FTP prediction. Everything else is there (Levels, progressions, etc).

[dendroica1]

I am very excited to try this all out, but I am one of the TrainerRoad users who just uses it in the winter and just rides for fun outside in summer.

@Nate_Pearson I was wondering if the next step with Plan Builder would be giving way more options as to what levels there are. I think it would be awesome if you could tell Plan Builder, here are my events, I want to ride X number of days and Y number of hours per week, and it gives you the best plan for your schedule

[Nate_Pearson]

@urnowsmoking

If the machine learning adaptive plan has the ability to take into account outside rides and rides on other platforms that would be mind blowing. It would mean that I have to get a power meter for my outside bike asap.

We understand that magnitude of being able categorize and score unstructured outside rides that then lets us feed that data into other ML systems.

There are so many possibilities for this and it would mean that you could get benefit from TR without ever being indoors or doing a structured workout.

[D-9]

With heart rate based tss, and ftp, it would be really tempting to do a workout that is not on the bike, but allow heart rate to put in a quasi FTP and TSS for the workout. For example, I could just start Petit, and then go on a run with a heart rate monitor. I know that the actual TSS and FTP would not be accurate, but it would be much closer than nothing. At first I was tempted to not even mention this, because I worry that workouts like this might screw up the machine learning, but somebody else is going to think about it. Maybe trainer road needs to put together a workout that is just “other sport,” so that folks who want to do this kind of thing can do it, but it won’t screw up the machine learning.

[Nate_Pearson]

@BCrossen

Pauses are accounted for in the ML, and a post workout survey allows you to tell it why the pause occurred. As has been mentioned several times in the thread, the ML is smart enough to know when those pauses are because the work was hard or because you got off to stretch or refill a bottle or something. And if it doesn’t know, it asks you why.

Yes and yes. It should be smart enough, but if it gets it wrong you can tell it and that will have human review and then we can use that as a case to train the model.

[pats]

Hi. I’m very excited to start using this training. I signed up to be added to the “list” on 2/25. Can you give me a guesstimate of when I will be able to start?

Ps any updates on when hr can be made invisible on the mobile app?

Thanks and great job on this new training feature. I think it’s gonna be a “game changer”!!

[Nate_Pearson]

@iamholland

A very exciting time, but execution is key here.

Such true words! We have a head start in my opinion over the other offerings I’ve seen and over our biggest competitors.

Now we have to move/learn/adapt as fast as we can.

[Nate_Pearson]

@denglish67

What is really exciting is that it should help to answer many of the ongoing questions related to the best structure for a training plan (periodized, pyramidal, traditional base vs sweet spot) and how much volume is needed for gains and what is a minimum effective dose. Most studies on these topics are necessarily based on a few self-selected participants in artificial scenarios. TR’s ML and huge data resource should be able to discern real-life results based on a wide-range of athletes. This will be a huge game-changer, not just for TR, but training in general.

Exactly. They are also usually college age men and for 4-6 weeks of training.

[Nate_Pearson]

@denglish67

One specific Q, and one area that I do not think has been mentioned so far, is how does your power curve feed into plan adaptations? And in particular, new power PRs. This is most relevant to data from outside rides, races etc. For example, during the course of my last two training blocks I have set a whole new set of PRs, but the majority of those have been when racing hard online. Of course, they are a result of the inputs (the training I have completed), but they too are a data point as to the power I can currently hold for a given duration - so will that also be taken into account in the progression?

It’s more than just your power PR, we have to look at repeatability too. That’s all the info I’m going to give on this question.

[Nate_Pearson]

@Dennis

Suppose I pass all my workouts but I’m actually over training. How will the system detect that? Same with under training I guess. In either case I might have a normal pass on most workouts. I’ve read anecdotal stories about riders doing the plan as prescribed, i.e., they’re able to complete the suggested workouts, but they’re spent at the end. Nothing left for the event stage.

If you pass all the workouts and have RPE responses that aren’t outside norms or trending in a bad way we wouldn’t know at the moment. We do follow patterns that should make it so you don’t overtrain, but we don’t know overall life stress so it’s possible.

We do have other things in the works for this. Subjective daily RPE, HRV, resting HR, and what ever else we come up with.

This is a really hard problem but we think we can get closer to perfect.

[Nate_Pearson]

@iamholland

Like a real coach, if you never tell the coach that you’re torched, you won’t get adjustment until physiological indicators start to crop up. Continuous failed workouts. High HR, lots of stops mid-interval, lower than normal cadence, etc.

Exactly. Those post workout surveys are going to be key.

[hubcyclist]

Dumb question, where do we find the ftp prediction for us beta users?

[ambermalika]

Yes - between ML and your survey response, we’ll be able to account for the workout accurately. The new workout filters (filtering by power zone and interval type/progression) can help you get a similarly-targeted workout at the right difficulty level for a shorter duration. Also coming soon is the option to do Harder/Easier/Longer/Shorter on any given scheduled workout … so hopefully we’ll make the need to fast-forward obsolete!