Kettlebell A.I.

[ali]

A really wonderful read and insight from the world of AI.
A sophisticated algorithm based on machine learning coded for every training metric...hrv etc...recovery inputs both subjective and objective feedback matched to every known training variable with every known template known, or perhaps unknown, of classic work/rest/sets schemes.
Really good read......spoiler alert below the link with the result!

Lessons from my A.I. coach

What my A.I. coach has (re)taught me about High Performance Sport

alancouzens.com

"While it might be a little paradoxical that one of the most complex Machine Learning algorithms on the planet is ultimately pushing me back to the simple, time-honored, wisdom of some truly great Human coaches,"

and

only cares about what will give you the highest projected performance on race day & so it, smartly, says...

"OK, you know that set that you've been doing for the past 6 weeks? Add 10 minutes to it for the next 6 weeks"

Exciting? No.
Proven? Yes.
Optimal over the long term? Yes.

 

[ClaudeR]

Nice one! Should be interesting to see it evolve and if sometimes specialized approaches (or hopefully very unconventional things) appear

 

[ClaudeR]

By the way, it seems to be based on AlphaGo, which is notable for having taught itself to play chess (absolutely no input other than the basic rules) and won against every grand master. Time it took to learn that?

4 hours.

not a typo, it took all of 4 hours for it to evolve from zero to grand master, all by itself with no input whatsoever

scary and wonderful at the same time

 

[ali]

Yes, all that computer power, all the gazillions of possibilities and it's adaptive progressive overload. Checkmate. The simple basics, good inn'it?

 

[Van Der Merve]

I was surprised by what I saw: A simple reminder that large variations from the norm create exponentially larger fatigue & that a Milo-esque slow and very steady progressive overload remains a surprisingly powerful (if, admittedly, less exciting) strategy.

That's why "micro plates" work very well in lifting, despite of Pavel's insistence that large increases in training weights - S&S 2.0 - are superior for progress.

Not that anybody needed a sophisticated software to confirm this, of course.

 

[Anna C]

That's why "micro plates" work very well in lifting

Yep. Worked quite well for me. Can't say I love the microplates, but there were several lifting session where 127 lbs on the bench press was just right for a session to drive progress... Not 126. Not 128. Or 91 lb press, no more, no less. As for the lower body lifts, didn't ever microload those.

 

[Manuel Fortin]

I don't get how their AI training is done from the link in the original post. Maybe someone can enlighten me. I understand how neural networks work. I even spent a whole summer working on them in 1990, during one of the older waves of popularity for neural networks. However, to train an AI, you usually go as follows:
1 - take an input
2 - use the neural network (NN) to generate an output
3 - compare the output to what the NN gives you.
4 - adjust the parameters of the NN so that the observed output is closer to the desired output.
In other words, you need to "try" the results of the NN to adjust the parameters.

Alpha performed its training by playing against itself. That's fine for games, but how do you train the NN to maximize for example strength gains, things that are based in the physical world? You would normally have to try the program given by the AI on real humans, see how it goes, and adjust from there. This would take a very long time. If you use a computer model of the human body, you are simply training the AI to understand your model, which includes your preconceptions about how the human body works. The fact that it recovers what other people have proposed is not really surprising. You basically told the AI to understand the human body the way we understand it.

Maybe I'm wrong and someone can explain why this is a new development. To me, it seems like they are just riding on the popularity of AI.

 

[apa]

By the way, it seems to be based on AlphaGo, which is notable for having taught itself to play chess (absolutely no input other than the basic rules) and won against every grand master. Time it took to learn that?

4 hours.

not a typo, it took all of 4 hours for it to evolve from zero to grand master, all by itself with no input whatsoever

scary and wonderful at the same time

AlphaGo took a lot longer than that to beat Lee Sedol.

 

[ClaudeR]

If you use a computer model of the human body, you are simply training the AI to understand your model, which includes your preconceptions about how the human body works

that is what I suspect as well... although I think machine learning works differently now and a lot more intelligent rather than just calculations and then you refine the parameters (machine learning now does that by itself)

At some point you still have to give it the ground rules, which by default will contain all preconceptions
Would be interesting to know what the ground rules and initial parameters were andhow it evolves them

 

[ClaudeR]

AlphaGo took a lot longer than that to beat Lee Sedol.

True... limited by its opponent ? Go is also a lot more complex than chess

 

[Steve Freides]

That's why "micro plates" work very well in lifting, despite of Pavel's insistence that large increases in training weights - S&S 2.0 - are superior for progress.

Micro changes in weights is not what we teach here. Everything works for some people at least some of the time. StrongFirst doesn't try to teach every possible strategy.

Can't say I love the microplates, but there were several lifting session where 127 lbs on the bench press was just right for a session to drive progress... Not 126. Not 128. Or 91 lb press, no more, no less.

But there are other ways. Even though I only bench press around my bodyweight of 150 lbs, I don't use less than 5 lb plates as a rule. 135 x 5 or 145 for triples or 155 x singles is where I start if I haven't been doing the lift for a while. I haven't found a reason to try to bench press 140.

-S-

 

[Van Der Merve]

Micro changes in weights is not what we teach here. Everything works for some people at least some of the time. StrongFirst doesn't try to teach every possible strategy.

-S-

My post had nothing to do with StrongFirst methodology or teachings. Have you read the article?

 

[Alan Couzens]

I don't get how their AI training is done from the link in the original post. Maybe someone can enlighten me...

how do you train the NN to maximize for example strength gains, things that are based in the physical world?

In our model, the task of maximizing strength or fitness gains is accomplished via model-based Deep Reinforcement Learning.

We start by passing the model actual athlete data - a sequence of observations (workouts, sleep data etc.) and the corresponding result (in our case, aerobic fitness parameters such as the power generated at a given heart rate but it could just as easily be an athlete's 1RM). The model develops a 'map' of these action sequences and the corresponding performance *for a given athlete*. This is the supervised learning part of the process that you describe above - inputs such as chronic load, acute load, sleep, subjective wellness, heart rate variability are compared with the outputs of actual performance and session fatigue and a neural network model of each of these input-output relationships is constructed.

Once this model is built, we implement the reinforcement learning part of the process by running thousands of simulations on different workout sequences where we reward the model for picking the workout sequences that lead to the highest performance while staying within given fatigue bounds. This is the "maximizing gains" part of the process. At the end of this training, we are left with a list of optimal workout actions from any given fitness/fatigue state, i.e. from a given level of performance and fatigue we choose the actions that lead to the highest long term performance reward (in our case, the greatest gains in bike power output, run speed etc over the specified time period)

Of course, fitness and fatigue change dynamically over time, so the model recalculates any time new data is presented. In this sense, it adapts daily to the athlete.

Alan Couzens, M.Sc.(Sports Science)
CTO, HumanGo Inc.

Humango™ Training App

Humango™ is a digital training app designed by athletes to help you crush your goals through individualized and adaptive training plans.

humango.ai

 

[ali]

Thanks for taking time to answer Alan. Good to have the author's knowledge first hand!
Interesting to know, if your data is substantial enough for comparison, how different athletes perceive the AI data.
Do they fully trust it? ...is the question, really I suppose. Do they stick with the recommended programme or sort of?

 

[ClaudeR]

Yes thanks for your answer Alan, highly appreciated!
I didn’t realize you ran the output on actual athletes!
Did their results follow the prediction of the model?

 

[Alan Couzens]

Thanks for taking time to answer Alan. Good to have the author's knowledge first hand!
Interesting to know, if your data is substantial enough for comparison, how different athletes perceive the AI data.
Do they fully trust it? ...is the question, really I suppose. Do they stick with the recommended programme or sort of?

Thank you for the kind words and for posting, Ali. Very cool that our approach is finding its way to a different group of athletes!

One of my 'demands' in working with HumanGo to build this product was that we are up front in exposing our model error. So, we certainly are getting a good sense of how 'wrong' the model is in its predictions and how the amount of data influences this. As to comparison, at least comparison with coach predictions, that's a little harder as most seem reluctant to make, even medium term, performance predictions.

As for the question of whether athletes trust an 'A.I. coach' or, an 'A.I. assisted coach', I would say that the jury is still out but we're hopeful enough that we're putting a good amount of $ on the side that, given the right information, they will. We're in alpha testing at the moment, with a view to move to beta this quarter. We should have a much better sense of the answer to that question at that point. Fingers crossed.

Thanks again for the kind words and support,

Alan Couzens, M.Sc.(Sports Science)
CTO, HumanGo Inc.

Humango™ Training App

Humango™ is a digital training app designed by athletes to help you crush your goals through individualized and adaptive training plans.

humango.ai

 

[Alan Couzens]

Yes thanks for your answer Alan, highly appreciated!
I didn’t realize you ran the output on actual athletes!
Did their results follow the prediction of the model?

Thanks Claude!

Yes, I've been fortunate to collect thousands upon thousands of workout files from actual athletes the I've worked with over the past 20 years or so. Up until now, they were just taking up space on the hard drive. It's been nice to put them to use testing the model.

To this point, (assuming 2 years of athlete data) we've been able to get the model error down to an RMSE of ~3%. In our terms, an error between predicted model wattage and actual test wattage of ~10W. In our opinion, certainly low enough to give some useful insight on potential programming directions to take.

Thank you for the support!

Alan Couzens, M.Sc.(Sports Science)
CTO, HumanGo Inc.

Humango™ Training App

Humango™ is a digital training app designed by athletes to help you crush your goals through individualized and adaptive training plans.

humango.ai

 

[Tim Randolph]

Yes, I've been fortunate to collect thousands upon thousands of workout files from actual athletes the I've worked with over the past 20 years or so. Up until now, they were just taking up space on the hard drive. It's been nice to put them to use testing the model.

Alan Couzens, M.Sc.(Sports Science)
CTO, HumanGo Inc.

Humango™ Training App

Humango™ is a digital training app designed by athletes to help you crush your goals through individualized and adaptive training plans.

humango.ai

I don’t know how Strong First does it’s data analysis, but I have long been impressed that it collects data when experimenting with new protocols. There may be a synergy here.

 

[ClaudeR]

Thanks Claude!

Yes, I've been fortunate to collect thousands upon thousands of workout files from actual athletes the I've worked with over the past 20 years or so. Up until now, they were just taking up space on the hard drive. It's been nice to put them to use testing the model.

To this point, (assuming 2 years of athlete data) we've been able to get the model error down to an RMSE of ~3%. In our terms, an error between predicted model wattage and actual test wattage of ~10W. In our opinion, certainly low enough to give some useful insight on potential programming directions to take.

Thank you for the support!

Alan Couzens, M.Sc.(Sports Science)
CTO, HumanGo Inc.

Humango™ Training App

Humango™ is a digital training app designed by athletes to help you crush your goals through individualized and adaptive training plans.

humango.ai

Hi Alan,

thanks for taking the time to respond!
Your results sound promising, certainly something to keep an eye on!

brgds
Claude

 

[Dasho]

In our context, rather than focusing its efforts, as a human coach might, on entertaining you, or appearing smart, by talking in quasi-physiological terms, or coming up with a convincing story about how this particular one session is "specifically targeting the signalling pathway responsible for mitochondrial biogenesis"

Don't tell him about the new book :/