I am looking forward to trying the adaptive training when it comes out.
One question that I have is how will the AI system assess how well I performed on a workout?
Will it simply be binary; based on whether I completed vs failed a particular workout/interval?
Will it look at HR data for particular workout/interval and compare it to HR on previous workouts/power numbers? (Which is how I currently look for small improvements by comparing max/average HR for a workout I have done multiple times).
I have been using Trainer road for 6 years. Will the AI be able to retrospectively look at personal data from previous years?
1 - Easy
2 - Moderate
3 - Hard
4 - Very Hard
5 - All Out
You don’t need to know what a workout should “feel” like. This is just how it felt for you. You might have an aerobic ride feel very hard because you’re thrashed, or you might have a vo2 ride feel moderate because you’re more fit than that level.
The idea is to not have anyone know what a ride should feel like relative to the workout. They just say on 1-5 how hard it was.
That would be great. I have been manually noting my Whoop recovery and strain alongside each TR workout for the last 18 months, so would be fantastic if TR could simply suck it all up automatically and use it to adjust plan, particularly in relation to fatigue/recovery.
I really really wish there was an option to add an endurance chunk at the end of workouts and not just a cooldown. Right now I have to finish a cooldown and then load up an endurance workout which usually has a small ramp. Kind of defeats the purpose a bit.
Garmin’s sleep tracking is WHACK. I have a garmin watch and it’s good for HR and GPS but not sleep. It’s a known issue. They’d be better off not even having sleep as a feature.
This may have been asked but will the ML adjust for time of day for the workout? For instance sometimes I do an “easy” morning workout just to wake up and get moving, like dans or petit, but because its in the morning and I typically train in the afternoon my R.P.E in the morning is almost always “moderate” or in super rare occasions “Hard”. Sometimes I have even turned the intensity down on lazy mountain in the morning but I sort of view these rides as separate from my training. Would the best way to make sure these don’t negatively effect future rides would be to just not accept the plan adaptions after these rides? Or will the ML learn to turn down my intensity in the morning but not in the afternoon or evening?
I don’t think we’ll need that much fine grain control. I could be wrong though.
But I want to stress this so hard…it’s not just “one” thing anymore. It’s hundreds of things and we might never understand the true relationship between everything.
And as we get more sophisticated and have access to more data we hope to make everything more accurate.
That would be sweet. We could do the first round that’s usually for bankers but make it just for TR users. And the more workouts you have in TR the more you could buy.
No plans for investment or IPO at this time though.
I’ve got a pretty big timeline goal in my head and I don’t want the external pressure to cut corners and get quarterly results.
Yes, we understand the difference between a pause during a 5 min rest interval to go to the bathroom and a pause in the middle of a 10 minute over-under set to make it easier.
This is one of the many reasons we used ML and not just do interval averages.
We had to classify thousands of rides that had different pause amounts to order for it to learn.
Exactly. As I’ve told you we want to “Make the World Faster”. I believe that we can use the same techniques/foundation that we’re using here in more sports and in health care/rehab.
But for now, we need to get this system flying and 100x than anything else you can get.
I hope you don’t break down the pass fail of a workout as just one data point. Talk about the rpe relates to this cause I was thinking two possibilities:
this feels hard so I will lower the resistance of the session
-i am struggling but will not touch tr
The first is obvious to see in the data. The second should be able to be picked up easily in someone using power match as the person’s power output will waver more then normal as they reach their limit.
@danredfern, Have you had a chance to watch the announcement on the AACC podcast? They went over a lot of (all of?) this there. To answer your questions in brief:
It is not binary. There is a scale of success/failure, as well as subjective input about the outcome (how you felt). For workouts that are cut short, there is also quick input to explain why (ran out of steam, ran out of time, equipment failure, …).
It does look at HR & power from previous workouts, along with 100 or so other features. The exact way it uses each of these data points is not always clear. Part of Machine Learning is that the system learns and adapts as it goes on which data points matter, how they matter, and adjusts its programming to continually improve its accuracy.
It will look at past data, but, if I recall correctly from the podcast, it won’t necessarily go back the entire six years.
(not whoop going by dcrainmaker, so will need a Garmin watch which gets the same basic info)
