The hidden benefits of employing an A.I. 'assistant coach'
Alan Couzens, M.Sc. (Sports Science)
May 3rd, 2017
About a year ago, I wrote a post entitled “Why Coaches Should Learn To Code.” The post obviously struck a chord with quite a few coaches out there, becoming one of my most popular and spurring a good number of follow up inquiries from coaches all over the globe.
In that post, I outlined some of the benefits that I was finding as someone in a non computer related field (coaching) who decided to learn to code & apply some of these basic programming skills to my daily coaching life. Benefits like being able to quickly ‘query’ my database of athlete training files to ask any questions of the data that my curious brain could come up with! 😀
I also discussed another really useful application of coding in the streamlining of information delivery - in both directions. For example, I set up a simple script to send me notifications when certain limits are exceeded when an athlete uploads a file (e.g. if training intensity is more than 5% above or below plan), I set up a basic voice activated query that I can run each morning to check in on my athletes HRV and other metrics each morning to let me know if there are any 'red flag' athletes that I need to watch...
...&, going the other way, rather than having to manually look at an athlete’s current zones when putting workouts together and typing them into the boxes by hand (as I used to do), I let the machine do it for me – both the looking up and the typing. I still love watching my A.I. 'ghost' assistant coach going to work, pulling up Training Peaks, and typing the workouts I specify with the athlete’s individual details TP’s into boxes like a player piano....
Note: This is all PC typing, no AC typing :-) & it is not sped up! My A.I. assistant coach can type (& think through workout details for an individual athlete) much faster than the 'H.I.' (human level intelligence) head coach 😀)
Even this simple level of automation is a game changer in terms of efficiency & saving time. But, as time goes on, I’m starting to realize an additional, maybe even greater, benefit of bringing an A.I. ‘agent’ into the office to share the workload &, surprisingly enough, it falls on the emotional side of the coaching equation.
Here’s what I’ve discovered: Directing communication through a machine has another significant advantages in that...
(Click to tweet)
Simply, no matter what input you give it, a machine will only respond with objective, non-judgemental output &, in this day and age, there is something very appealing about that.
In the previous piece on coding, I showed a pic of my first computer – a gift on my 10th birthday, the mighty VZ300. One of the first programs that I typed into my beloved VZ300 was called Eliza – a BASIC adaptation of a famous computer program from the created in the M.I.T. artificial intelligence lab in the 1960’s that tries to mimic a Rogerian psychotherapist. A very, very early & simple application of Artificial Intelligence. Now, I’m no psychotherapist, but my hunch is that Eliza wasn’t especially good at her job. Her approach was a bit of a one trick pony – she would basically reflect what you said back to you in the form of a question..."
“I feel sad today, Eliza”
“I’m hearing that you’re feeling sad today, Alan. Why don’t you tell me more?”
But, the funny thing was, though clearly dumb as a post, and clearly a machine, she was very easy to ‘talk to’.
Looking back on it, I think I understand the reason for that:
A machine takes everything we say at face value & responds similarly.
Take your car’s navigation for instance – have you ever missed a turn that causes your navigation to go into “recalculating” mode? You’ll hear Gertie Garmin come on and say “recalculating”. Her tone doesn’t change if you miss 100 turns. No escalation of ‘pissed-offedness’ etc. Just, in a very stoic manner – OK, recalculating. How would the same scenario play out with a human navigator, maybe of the significant other variety? 😊 You miss a turn, forgivable mistake. You miss a second and you hear the map getting shuffled irritatingly. You miss a third and whether verbalized or not, you know that “I can’t believe you did it again, you Jackass” is in the air And there is something to this unquestionable objectivity.
Another e.g. – Have you ever received the dreaded “Insufficient Funds” message from an ATM? Not the best feeling admittedly, but compare it to the human alternative – you stand in line with your withdrawal slip for $20, walk up to the teller and she says to you “I’m sorry, Sir, it doesn’t look like you have enough money in your account to cover this $20 withdrawal” It carries a whole different emotional context to it, wouldn’t you say? More embarrassment. Maybe even a little more anger or desire to argue “No, that can’t be right…” Truth is, for all we complain about the lack of ‘human-ness” in the machine agents of today, there can be something quite appealing in pure, objective information transfer without emotional judgement or undertone. And I think that is what I discovered many years ago with my machine friend, Eliza.
So, how does this relate to coaching?
Well, I would argue that in coaching the emotional undertones in communication are especially strong. Athletes on the whole, don’t want to ‘disappoint’ their coach and may be especially sensitive to any communication perceived as critical. Sometimes to the point of not taking the communication onboard, but rather, in a knee jerk reaction, arguing against or dismissing it. Similarly, believe it or not, giving critical communication is not the most fun thing to do for a coach either. Giving the same critical information over and over (“you took a wrong turn again – recalculating :-)” can be quite wearing.
To this end, with a very simple application of A.I. you can create a basic proxy ‘domain expert’ system that offers frequent, simple, objective feedback to the athlete with no undertones or agitation.
In my team, we call him “Mini Al” and he currently operates on a couple of levels..
a) ’Morning check-in’ feedback
When an athlete completes their ‘check in form’ each morning – logging their sleep, soreness, mood, HRV etc (along with any more personal notes about how their current life status might be impacting the training), they receive an immediate response….
A response that will give them advice on whether they should alter the planned training for that day. This response is algorithmically defined by a machine by looking at how the metrics above compare to the athlete’s ‘normal’. For example, if HRV falls more than a given range from the mean, the machine may advise the athlete to omit the higher intensity portion of today’s training in favor of an easier session of aerobic base work. Additionally, I will receive an email from my machine assistant coach letting me know that the athlete’s HRV is low today. If the HRV has been low for a number of days, I will also receive an email letting me know that so I can intervene and change the programming if needed – moving a recovery week up in the rotation etc.
b) End of block feedback.
At the end of the training block (72hrs before the end of the block), my head will pop up on the athlete’s ‘block report’ with some summary thoughts on how the block went….
Again, the machine is helping me out dramatically, by sifting through the athlete’s (quantitative) data for the training block, comparing it against normal numbers for that athlete and returning (qualitative) insights that may be relevant to myself & the athlete
In both of these applications, I see a big advantage to having a machine start this conversation & to follow up as often as needed. This is not to say that it eliminates the role of the human coach in any way. Rather, in my view…
(Click to tweet)
It’s a long standing theme of effective teams: The whole is greater than the sum of it’s parts. Or you could look at it as the effectiveness of a ‘Good cop, bad cop’ strategy. Or, more eloquently from the book 'Smart Machines'....
Whichever way you look at it, the objective rational agent and the subjective human agent make for a powerful combo!
Is this all "A.I." is?
Not by a long shot! But it’s a start... “Expert” systems are the simplest form of A.I. They are obviously labor intensive to code as each ‘rule’ must be individually programmed…
If the athletes HRV is this, then say this….
Else, if the athlete’s HRV is one standard deviation lower, then say this instead…
The power of the A.I. agent starts to really kick in when the agent isn’t just parroting back your own feedback/your own ‘rules’, but rather is testing the validity of those rules & starting to deduce her own smarter, better, rules than you could possibly program into her even if you had time to do so!
While this may sound a little sci-fi – a machine learning to give better, smarter advice than a real life coach, when we break it down, it’s really not that hard to come up with a machine that does a better job of storing information and objectively deducing true, meaningful patterns than the (inherently non-objective) human mind. I looked at an example of using algorithms for relevant 'feature extraction' previously in the form of using ‘decision tree’ algorithms to predict injury, where the computer identifies the variables most important to the outcome and, furthermore, identifies the critical meaningful ‘cut off’ thresholds for each of those features.
If we're honest with ourselves, many of our own ‘rules’ in endurance sport coaching have pretty shaky evidentiary foundations. Often, foundations that are more based on folklore than fact. Rules that may have come from what was generally understood about the most salient features of one athlete or one squad. Take for example, Lydiard’s 100 mile week. While, maybe a decent generalization of appropriate training load in the base period for an elite endurance runner, it is clearly a ‘middle of the curve’ simplification. We know some middle distance and distance runners have been world class on 70mi/wk (e.g. Seb Coe). We also know others that have put in a huge number of 130mi weeks to reach a world class level (e.g Lasse Viren). Additionally, via simplification, it ignores the entirety of the Lydiard program. Was it the 100 mile weeks, was it the sequencing of the various phases, was it the emphasis on hill work, or was it how they all fit together? And (quantitatively) how important was each of these features in the mix?
A ‘squad coach’ may say, well in the absence of better information, doing what has worked for multiple athletes in a world class squad is a good start point. Tough to argue with that, but I would question - is there truly an absence of better information in this day and age? A coach with a bit more ‘processing power’ might say – let’s look at the specific features that define those athletes who do better on high v low volume. Let’s cluster these groups and look at what features separate them & then maybe when a new athlete comes on the scene, we can assess him/her for features, see what cluster they best fit into and start to move in that direction. While not terribly complex, these sorts of powerful computations clearly fall outside of the “do it trackside” variety & are far more efficiently performed by a machine.
Importantly, these clusters aren’t fixed or static. This isn’t a piece of research that the scientist does once and bases all ongoing conclusions off of. No, if the scientist is a computer, the ‘experiment’ is always running! These clusters and the understanding of the features that separate them will progress over time. As the data gets bigger, the AI coach gets smarter. Maybe at the start, our algorithm says “Athletes over 80kg don’t handle 100+ mi weeks very well” Then, maybe a year down the track, you get an email from your computer saying – “you know what Coach, on the recent data that’s coming in, it’s actually athletes over 75kg who don’t seem to be handling those big weeks”. Or maybe it says “Coach, I’m noticing that those athletes who are naturally faster over the longer distances right from the get-go, seem to also be the ones responding better to volume.” The point being, unlike human coaches, the A.I. coach isn't 'time poor'....
(Click to tweet)
And furthermore, the computer will always tell you “how it is” without fear of emotional consequence. It won’t pretend to be more certain than it is (and can even quantify as a percentage number, just how certain it is..
“If you continue on the current ramp rate, you’re going to crash and burn with 98% certainty” :-)
& it will continue to present the best information stoically, knowing that you are human and you may or may not always follow the most rational information.
For all of these reasons, I can’t think of a better ‘assistant coach’ to bring onto the team. As he gets smarter and smarter, I’ll be continuing to include him more and more as an essential component in the team’s success.
Don't miss a post! Sign up for my mailing list to get notified of all new content....