AI Training - Will it work?

Thru a little trial & error I’ve settled on Level = 2 in standard mode, and with that I can do zone2 work with seated sprints at basically the same power as outside sprints. That also works just fine for tempo, threshold, and short vo2/anaerobic intervals. If feeling strong or gearing feels off I’ll use Level = 3. My FTP is around 250W, my coach dials in short intervals at various power levels, and 5-sec sprint is about 4x FTP. Standard mode is better than resistance mode IMHO. Just start the workout in standard mode and treat it like outside. Hope that helps.

1 Like

Standard mode is the only mode! Honestly I do not remember when I used erg last time. For me standard lvl 2 is also perfect. And if I have a problem with exact power target I know that lvl 0 or 1 will do the job. And one more thing from my observation (anecdotal) that I recruit more muscles with standard than with lvl. Especially with threshold or some power bursts it feels a lot more realistic. Not to mention that holding power by feel is way more entertaining during these long Z2 rides and you have the freedom to do whatever you want… almost like riding a bike? :wink:

1 Like

I understand. But there is no surprise. Go to the Planner Tab and a calendar pops up. Then just press the “recommended workout” icon located on the date you choose.

But I don’t think it does.

The summary is…it depends on what kind of rider you are and what your individual power profile is. They have an example towards the middle of the article.

2 Likes

Not until it takes into account swimming, running and lifting. But they’re not doing that on any timescale.

George is correct; As far as the inner working of Xert is concerned, all power below Threshold is considered to be low stress and therefor sweet spot workouts get recommended as endurance even when they may not be the most appropriate.

The article is arguing that for some riders, low end sweet spot is actually below LTP, which may or may not be the case for the individual.

At this time Xert does not make any differentiation between workouts above and below LTP except that the difficulty rating is higher for the sweet spot work and won’t be recommended until you training load is high enough.

Mike

2 Likes

Agree that this is where proper AI/ML could get interesting. I have pretty good data on all of the things you list that could upload automatically. Think the key thing that is missing and would require more effort on my behalf than simply connecting accounts to each other would be accurate logging of diet. My power and HR are probably 95% reliable and the times they go wrong they normally go sufficiently wrong (dead battery or cut outs) that the data could likely be automatically excluded.

Definitely think there’s some value in augmented intelligence like spotting trends and offering recommendations and advice based on rules not true AI.

I think cycling training is a prime candidate for machine learning to be honest.

Loads of input options (workouts) and an easy to measure output (ftp) plus loads of available data.

I wonder what TR’s response will be when the machines think we should all be doing polarized though :wink:

1 Like

I think they are 10 years away. Lol.

The thing that has cracks me up about ‘AI’ is that its been just 10 years away, for the last 50 - 60 years.

1 Like

Its still new, but I thing practical quantum cryptography has been 10 years away for the last ~5 years.

I use something called TrainAsOne, it uses AI and picks up your workouts from Garmin / strava and then adjusts your next work out based on the workout or test that you have just done.

The App arrived a few weeks ago but normally its a desktop version. if you pay for the premium version it will send workouts to your watch, I just create the next workout when it gets emailed on my garmin and do it.

Its dependent on key workouts to measure progress and will set very random recovery workouts until you tick that type of workout off your list.

I have used it religiously for a block of training and noticed a massive difference when I stuck to it. and I am currently seeing good gains in my running whilst I am using it to replace the trainerroad run sessions in a triathlon plan.

The goal in both programs is the same date and a 2.5 hour half marathon.

The both set the same amount of running TSS during a week its just differently delivered.

One of the tests is a 6 minute assessment runs and I cocked it up on Saturday, the treadmill wouldn’t play ball and really didn’t want to do it in the cold. and its seen this and now putting in slower workouts.

If I delete that 6 minute test it simply schedules another test.

I think it would get me to a 40 minute 10km I don’t think it would get me to the 35minutes

1 Like

@AlphaDogCycling hit the nail on the head

The biggest challenge is which factors/training dimensions will have the biggest impact and which are just marginal gains. And once you think you know the dimensions you have to both be able to measure the raw values and weigh their impact in any algorithm you design.

Realistically at the macro level, the factors would probably be something like (1) overall training distribution; (2) nutrition; (3) recovery and (4) goals.

AFAIK, TR only really tracks overall training distribution. So there’s a few gaps in the measuring before getting to the weighing.

I could be wrong though - I’m not a sports science expert.

I guess that is what I meant, for some people the low end of sweet spot is “low” stress, but for some people it might be “high” stress. Maybe I still don’t really understand Xert. I love the idea of Xert and have been trying to understand it more.

I believe that they are looking at further division of the training stress allocation with the cut-off being at LTP.

At the moment XSS it is proportioned into 3 (Low, High and Peak) in varying degrees based on your power curve but all XSS is put in the Low bucket when you are riding below threshold.

XSS Work Allocations

Mike

Thanks for the clarification!

Although the TR charts only really show training distribution, would be interesting what other data they’re collecting . E.g. Garmin is tracking my sleep, resting HR, all day HR, recovery status (based on HRV), my weight via smart scales, etc. Garmin in turn also syncs with MyFitnessPal so if I was bothering to keep a food log that would be sitting in the Garmin database as well. I’ve authorised TR to sync with Garmin, and the TR blurb relating to that sync states that:

By connecting with TrainerRoad, you agree to share information from your Garmin Connect account to enhance your experience with TrainerRoad. This may include activities, location, heart rate and related metrics, calories burned and other health or personal data.

Have to assume therefore that TR has access to some or all of that other Garmin data beyond just ride data.

So of your 4 factors I reckon there’s a good chance TR already have a good data set for the first 3. And goals would be simple enough to add in. Or even just keep it simple and make raising 20 minute power (or ramp test power since it has more of an anaerobic component) the goal for everybody on the basis that it’s the rising tide that lifts all boats.

Plus I left out other sources of “training stress” that don’t get recorded / factories in.

For example for me, I signed up for the Get to Sesame Street 500 mile challenge, so I’m tracking miles I’m walking by walking my dogs, and so far this year I’m just shy (149 miles) of the 150 mile mark. This is purely the normal walks that my dogs get - I’m finally tracking because of the Get to Sesame Street Challenge. Without this challenge, I wouldn’t have any idea of how many miles I actually walk per month. My guesstimate was closer to 40 - 50 miles / month, instead of the almost 100 miles / month I actually walk.

I’ve never used it and no connection to it so just as a FYI…here is a training app for running, cycling and triathlon using machine learning:

Some interesting points from the FAQ

How well can AI Endurance predict performance?

We can predict your future performance at about 5% if you have at least 100 runs with GPS data or rides with power data.

How often do you train the AI?

With AI Endurance every user gets their own ‘digital twin’ - a machine learning model that represents all the possible ways you would respond to different training routines. The model is updated every 24 hours for every user taking into account each new activity you have uploaded via Garmin or Strava.

What about heart rate?

We use your heart rate data to gauge your performance. For instance, if you were able to run 5 min/km or at 200 W for an hour at a significantly lower heart rate than last week, the AI notices this as an improvement in your performance. However, generally heart rate is a much worse predictor for performance than power and GAP.

That’s what I love about the AI, ML, Deep Learning stuff lately. Tons of fancy buzzwords for task we can solve already.
If you give me 100 rides/runs with power I can certainly predict your performance with a calculator. No need for any AI wizardry.

4 Likes

5 Likes