Reading through the Dylan Johnson thread, I have seen AI training mentioned on several occasions. I think it’s a brilliant idea and I’d love to try it. If I’m not mistaken, I believe XERT already does this? Or something very similar. Full disclosure, I have tried XERT. I didn’t fully understand it and I don’t want to rubbish what might be a brilliant training platform.
I guess my question, or maybe provocation, is do you really see AI Training working for TR and maybe more importantly, the type of users TR currently attracts.
My honest answer. Not a chance.
Why? This forum alone is full of examples of why it won’t work. ‘WE’, the user(s), ultimately believe that we know better. Examples such as ‘The Ramp Test has given me the wrong FTP! I KNOW I can hold X watts for X amount of time. I did this only yesterday etc, etc…’ Then said user heads off to find another testing protocol that might give them a number they like the look of. What if the AI fixed that testing protocol? I’ve never done a 20 minute or Kolie Moore Test. I bet I’d suck! Could I, or anybody else, honestly set their next block of training to that number?
Other examples include just bumping your FTP by 5% because… because you completed Carparthian Peak +2 without your HR even coming close to where it should be for Threshold. Maybe you were on form that day. Congrats. What’s tomorrow or next week going to look like?
Would turbo trainers need to generalised? Smart Trainers only? All running the same software, auto updated.
If a computer algorithm was 100% in control of your training, could you get onboard with that? In this scenario, there is NO way to overrule the AI. You either follow the training, at the intensity it prescribes, or you find another platform.
Personally I think AI is useful to a format like TR for “optimizing” a product that is based on general application to a broad group of athletes. TR plans are a starting point/guidance. But the canned plans need to be individualized to the particular athlete. AI can never truly accomplish that because it can’t replace an athlete’s intuition regarding what their own body is telling them. I don’t think all the metrics and algorithms in the world can ever replace that. But I’m a skeptic by nature so open to reasonable arguments to the contrary!
I use both TR and XERT. I am a ‘time crunched, masters age cyclist’. I have used both platforms training plans and have found, for myself, that following TR’s LV plans whilst monitoring with XERT’s AI algorithm has given me the best results. There is a lot of rumbling about TR not having a masters plan, with merit. But I have found that the LV plans (2 high intensity sessions/week) + endurance Z2 and/or strength sessions is a great masters age protocol. I gained 30W+ last year doing so and hope to achieve in 2021 PR’s in FTP (TR lingo) and TP, LTP, PP best (XERT lingo).
Meaning we like to throw new, powerful technologies at everything. Then we realise it is not suitable for everything and dump the technology until it is used for things that it’s suitable for and not the things it’s not suitable for.
I personally don’t know much about AI. I think OPs point is fair, that I would propably fiddle with the plan anyway. I’m curious to see what the actual AI people here have to say about it
Could not an AI approach be better for this problem as it could take more information into account? It could take into account your power duration curve and know that my FTP higher than average relative to my MAP (~82% not ~75%) or vice versa for athletes that are anaerobically inclined? The downside is having to provide a series of maximal efforts over several durations, i.e. ‘feed the curve’.
I think if there was a way to give subjective feedback on RPE, how I feel throughout the day, whether I’m feeling recovered or not, maybe sleep data, maybe. I would want to evaluate over an extended period of time, but once the approach proved itself to my satisfaction I would most likely be okay with it.
Not trying to be cheeky, but what is optimal training? If expert, human coaches cannot give ‘optimal’ plans to their clients, is it reasonable to expect an AI approach to provide ‘optimal’ training?
As a computer science person, a pet peeve of mine revolves around ‘AI.’ Do we mean an expert system? Self-learning / neural network approach? Data mining and inferencing? The layman definition of AI tends to be quite different from CS theory and application.
I agree with the amount of second guessing users do. You also have the more is better crowd. If 85% is prescribed then 95% must be better. Or the I don’t want to waste my time doing easy endurance exercise - junk miles. Or those guys doing a plan plus chasing KOMs every other ride outside. No wonder they are burnt out. Or, these workouts aren’t hard enough, I should bump my FTP every week without having tested.
In all these cases people are second guessing the plans as designed.
Well, yes. I don’t agree with this point. The ramp test is influence by the percent of anaerobic and ones ability to suffer. I think a longer test is a truer test of maximum lactate steady state.
I’m not sure if TR needs AI. They could implement some simple adaptive features. Right now the adaptive features are in the notes and in the advice of the podcast but many don’t follow them.
The ultimate way to go might be to track a power duration curve and then give workouts based on the users’ actual performance.
What I envision as a perfect platform would be one that uses machine learning against a large set of data to set up a plan based on whatever multitude of existing factors/fitness it can gather. Then adjusts based on on-going factors, including the previous prescribed ride performance and other inputs (thinking stuff like HRV, etc., ala Whoop), all towards meeting whatever improvement goals you may have.
I.e., the next best thing to actually having a coach.
I had some more thoughts on this on my way home. i think for me the answer would be no. It seems to me that something AI based would inherently be quite the blackbox. I personally hate to do things without knowing why I should do them. Currently there is quite some detail written out what the purpose is of each workout/week /plan and so on. I don’t think I would be comfortable to just blindly follow, whatever the machine is telling me, even if it was more effective.
Im also a TR and Xert user and of the older trait (55).
The only issue I see with doing this is that xert sees SS workouts not as high intensity but instead as endurance work so in theory will potentially ask you to do high intensity work during the xert build and peak (athlete type permitting) even though you have just done a tough SS workout.
One of the biggest problems would be figuring out what the AI system should consider. That is, which of the following are signal, and which are noise:
Objective - What are you actually trying to optimize? 20 minute power? “FTP” [how do you define this?]? Repeatability?
HRV - plus how do you actually get this number? First thing in the morning, average of measurements through the day, etc.
Resting heart rate - plus how do you define this number?
RPE - I personally know I suck at this, plus the current RPE scales either 10 or 20 are too large to get repeatable data
Hours of sleep - do naps count?
Life stress - how would you measure day to day vs. I had one stressful day?
Weight? % Body fat?
Nutrition - is it calories? Macronutrient %? “Quality”? Combination?
Room temperature / amount of fan cooling
Muscle Oxygen - think a Moxy Monitor
To construct an AI Coaching system, you would have to figure out which of the above you are going to (and then how) / not going to include, and then have a dataset with your included parameters to use for the modeling / training / etc.
And then the $64,000 question: for the “average” person, is all of that really necessary / better than what can be done by a somewhat more advanced version of the current TR plan builder (e.g., allowing me to pick the work to rest ratio, allowing me to set max length of workouts during the week - e.g., I cannot do more than 1 hour rides, etc.)?
There are problems with TR that ML could likely help with fairly easily. For example, it could predict who over-, or under-tests with the ramp test, and then swap the test out in your training plan or adjust the FTP result to create a better match.
It could also probably find out what type of workout progression usually works best to achieve a specific goal (eg raise FTP or vo2max), depending on your past training history or the type of rider you are. It could likely create types of riders and identify them.
I think what TR is missing to make this work is a good suite of analytics to ‘power profile’ people though.
The main problem with all computer approaches is that the data to feed the model has to be good. The other issue is communication - did you stop the ramp test, because you couldn’t go on anymore, or because your child just burst into the room? How does the computer know?
How good are your Netflix and Spotify recommendations? Do you trust them enough to just open the app and hit play?
If not, do you expect a training app to be able to do the same with an infinitely more complex set of inputs and outputs?
Or does “working” mean making it easier to make decisions on a day to day basis?
ML and AI are great for helping humans simplify decision making in a constantly changing set of rules. They aren’t yet at the place where they can make intelligent decisions for us unless there is extremely simple and unchanging rule sets (poker, Starcraft, etc).
There are also a lot of things that don’t require expensive techniques to achieve. Workout tailoring based on power curves doesnt need ML or AI.
On the ramp test: TR is using a single % of max 1 minute power for everyone. It would be interesting if you could apply AIML to get a better individualized estimate of what the “right” % of max 1 minute power from the ramp test to use for me to set my FTP for training.
I think if you are interested in AI and coaching, Alan Couzens is worth watching as he actively develops his software. I think his conclusion that AI will support decisions not make them for a coach is the only model that will fully work. And here is some reading and his view:
I think it is partially a matter of the laymans’ definition of artificial intelligence (human-like general intelligence) is different from the domain experts definitions. I also am a little negative on AI as if you go back to the 60s you’ll see interviews where ‘AI’ is 10 years away. And in the 70s its 10 years away, and guess what? Its 2021 and we’re probably 10 years away. lol.
Lots of great ideas below. I think some basic things that could be achieved are:
More accurate FTP estimate from ramp / 20min tests or suggesting ‘optimal’ test protocol based on an individuals power-duration curve. Its obvious from mine that I have much better 20+min power relative to my anaerobic power and could adjust my x% of ramp test as FTP estimate.
WKO generates optimized intervals for supra-threshold intervals based on your PDC. I think its possible to look at the PDC and determine power levels and interval lengths based on that. For me, given my relatively weak anaerobic development, ideally a solution could look at my PDC and give me ~3min intervals at 116% instead of 120%. Or for people who have better anaerobic development give them intervals at 124% instead of 120%.
Adaptive plans? I think this is a very challenging problem from the definition side, because how do you know if someone is plateauing because they are approaching their potential vs plateauing because of sub-optimal training. But it would be interesting to see if adaptive plans could give me more SS/threshold if it figures out I respond well to that vs giving me VO2max work because I respond even better to that.
With all the data that is available I think there’s an opportunity on the data mining side via clustering and other approaches to identify high-level relationships between attributes / conditions and outcomes. Like, people who get to 5wkg tend to be light (duh), or people who get to 5wkg tend to ride 12+ hours/week. The risk is confusing correlation with causation, but there are probably some interesting trends that could be identified for future research.
I think an unstated assumption is that while developing software to make progress towards these problems, there is basically a human being defining what is ‘optimal’ or ‘right’ and the developers would be tweeking whatever approach they are using to try and get as close to that as possible. If one cannot define what is ‘right’ or ‘optimal’ its hard to adjust the approach to match it. That is kind of why I asked, ‘what is optimal?’. If coaches cannot define what that is, its hard to get there. A big part of solving the problem is articulating it and the solution or solutions.
I agree there are probably some interesting trends in the data (for instance, I’d love to see if greater compliance with their plans correlates with greatest improvements in FTP), but I wonder how reliable the TR data will be for anything other than inferring how best to implement TRs existing plans. The TR platform is designed to have people follow their plans and its data will reflect this. This data can be used to show relationships between people who have succeeded using their training plans but it can’t show that someone would respond better to a polarized approach, a different distribution of Z2 vs SS rides, or anything else that outside the prescribed TR plans.