“Simply put, for the last 200 years, advisers have worked on the principle of information asymmetry, where they have better information than their clients. Today, we are at the point where machine intelligence is gaining information asymmetry over advisers, and that’s only going to get more acute and asymmetrical as time goes on. The only possible hope for human advisers is that they co-opt machine intelligence into their process.” – Brett King (Augmented)
For better or worse, we are living in a rapidly changing time. Only 20 years ago, as I moved from Australia to the U.S. and made my first trip to Disney World, I vividly remember the view into the future provided by the ride “Spaceship Earth.” As it spiraled upward through history, it culminated with a view into the living room of the distant future with a family engaged in various “futuristic” pursuits, ranging from kids video chatting face to face with their grandmother who lived across the country to these strange virtual reality headsets that enhanced the video gaming experience. Of course, they had to quickly change the ride only a few short years after my visit for one simple reason—time moves very fast and the future was already here!
Coaching hasn’t escaped these changes. In the world of endurance coaching, technology has facilitated the rise of “remote” coaching. Much, perhaps the bulk, of endurance coaching nowadays is done with the assistance of software platforms such as Training Peaks, which enable athletes to have access to quality coaches beyond their local area. Only 25 years ago, if an athlete wanted to work with a coach outside of their town, they would have had to move to that coach’s area or communicate their training back and forth via email. In the late ’90s, Joe Friel, Dirk Friel, and Gear Fisher saw a unique opportunity for a product that more easily facilitated this and, in doing so, remarkably changed the face of endurance coaching.
Current Applications of ‘AI’ in Sport
With the current influx of artificial intelligence (AI) into coaching, we are on the precipice of another change—a change that will be both another step into an unimagined future and, in some ways, a return to times past. So, what can we expect as artificial intelligence invades our world of coaching over the next decade? Let’s start by taking a quick look at where things stand currently.
1. Rules-Based Systems
Artificial intelligence is a broad concept with an equally broad range of definitions. Considering we have a hard time pinpointing exactly what human intelligence is, it should come as no surprise that we have an equally hard time defining the artificial variant. To some, a simple “rules-based” system where a human expert encodes part of their knowledge into a machine is sufficient to classify as artificial intelligence.Considering we have a hard time pinpointing exactly what human intelligence is, it shouldn’t be a surprise that we have an equally hard time defining the artificial variant, says @alan_couzens. Click To Tweet
If this definition is accepted, we are already there! There are a number of software packages on the market that attempt to do this for various sporting contexts. In triathlon this began with the “PC Coach” coaching software in the 1990s.
Your “PC Coach” asked a few “getting to know you” questions and then generated a basic training plan. Similar rules-based platforms exist to this very day. In fact, the majority of current commercial packages that use the term “AI” have little more than a bunch of “If-Then” statements under the hood. A simple pseudo-code example for the logic process of such a system might look like the following…
If weeks_to_event < 12:
- intensity = intensity * 1.1
In other words, as the event comes closer, crank up the intensity of the program by a given amount.The majority of current commercial packages that use the term “AI” have little more than a bunch of “If-Then” statements under the hood, says @alan_couzens. Click To Tweet
Clearly, covering all possible variations and decisions that a “real life” coach may encounter/consider over the course of a season is challenging, to say the least! What if the athlete is sick at week 12; would we still bump up the intensity? That logic would demand that we are: a) recording sickness/wellness information and b) covering all possibilities within the code. For example:
If weeks_to_event < 12 and sick==False:
- intensity = intensity * 1.1
elif weeks_to_event <12 and sick ==True:
- intensity = intensity * 0.9
What if the athlete is not sick but tired? We would have to write additional code covering each and every variant for each and every piece of data that we collect. This quickly becomes prohibitive, because with each new piece of data that must be factored in, the code grows exponentially.
This approach to AI was actually trialed in the field of medical diagnostics in the 1970s with a program called MYCIN. MYCIN was an artificial intelligence system that attempted to use encoded rules to identify which antibiotic should be prescribed on the basis of (a very long) series of Yes-No questions. While the accuracy of the completed engine was relatively good (comparable to human doctors) in this narrow domain, in the process of building MYCIN, the researchers identified a big problem with expert systems in general: “the knowledge acquisition bottleneck”—i.e., how do we transfer all of the knowledge of all possible variations that a human expert could possibly encounter into a long list of rules?
While the concept is easy enough to imagine, the actual process of covering all possible combinations of variables into sets of rules becomes impossible. In fact, this realization led to the first “AI winter” in the 1970s, when the Lighthill Report culled AI research in one fell swoop by rightly concluding that, due to the realities of “combinatorial explosion”—that is, stuff getting messy at the level of complexity found in real-world problems—expert systems were really only applicable to “toy problems.”
In the world of AI coaching, this reality of “combinatorial explosion” has led to expert systems falling under the “recreational athlete” category, with most serious athletes and coaches considering them significantly inferior to the abilities of a real-life coach to synthesize significantly more data and make better decisions on that basis.
2. Descriptive Analytics
On the more serious side of sports, the current definitions of “AI” and machine learning (ML) in sports have centered more around data acquisition and handling, says @alan_couzens. Click To Tweet
On the more serious side of sports, the current definitions of “AI” and machine learning (ML) in sports have centered more around data acquisition and handling. We are currently in the age of descriptive analytics; an age characterized by a plethora of different wearables/hardware coupled with “apps” that allow the coach/athlete to visualize the various channels of the hardware data stream over time. This ranges from daily health data such as heart rate variability, wellness questionnaires, and sleep trackers to the training data and game data recorded by heart rate monitors, accelerometers, and GPS units. Most of the current sports training software is focused on simply collating and presenting this information in one or more “dashboards” in a way that describes the current state of the athlete so that the coach can have as much information as possible in front of them prior to making a decision.
However, as the number of data streams grows with every new piece of hardware, making sense of all of these inputs in a way that truly informs decision-making is no small feat! In fact, some may legitimately argue that they add tasks to an already busy coach’s plate. Even with the best athlete monitoring systems (AMS) around, there is still a huge time investment to analyzing individual player responses on a daily basis and making individualized tweaks to the daily plan. To be completely frank, it is my position that this process is so labor-intensive that it is only done well at the highest levels of sport, in organizations that can afford to employ full-time data scientists.
A good analogy for the current state of affairs is to picture a Tesla driving down the road with all of its various data acquisition systems but no CPU to process them. You, the driver, are responsible for keeping tabs on the input from the eight cameras and various sensors on all sides of the vehicle and using that input stream to make better split-second decisions. Of course, a CPU takes in all of the data, weighs it appropriately, and combines it to automatically create an accurate model of the environment, resulting in a lot of stress taken off the driver. This process of using data to build representative models of athletes that coaches can utilize to predict the impact of different strategies represents the next phase in artificial intelligence in sport.
3. Predictive Analytics
Predictive analytics takes us one important step further in using the data to build models that can enable the coach to project ahead to what is likely to happen given a course of action. Currently, this is more the domain of research scientists who have used this approach to effectively predict things like injury risk and response from a given dose of training. For example, Jurgen Edelmann-Nusser et al.1 applied a neural network approach to modeling performance in elite swimmers based on training time in zone inputs and was able to achieve a prediction of within .05 seconds of the actual result! A similarly powerful example in the injury prediction sphere is a 2018 study by Rossi et al.2 that used a random forest model to predict injury to professional soccer players from training load and wellness data with an 87% accuracy!
By implementing this predictive modeling approach, we move away from the “driver keeping an eye on eight cameras” approach for the coach, where they have to independently monitor everything and determine “is the planned load appropriate for today?” by considering things like:
- How much high-intensity running is planned for today?
- Where is the chronic training load?
- Where is the acute training load?
- How well did the athlete sleep last night?
- What was the athlete’s HRV this morning?
- Has this athlete been injured recently?
Instead, we can feed all of the above inputs into one machine and it can spit out…
“Athletes chance of injury based on today’s inputs = 20%”
- “Athletes predicted fatigue based on today’s volume and intensity = 8/10”
- or, just as importantly,
- “Athletes predicted VO2 max improvement from today’s session = 0.03 ml/kg/min.”
The power of the decision-assisting insights provided as ML moves from research into application should be evident. Real-time assessment of the risk/reward of planned strategies for individual athletes will add a powerful tool to the coach’s toolbox as these systems make their way into production. However, there is a logical follow-up step that will be even more transformative and will truly be a “game-changer” when it comes to our job as coaches.
The Near Future: Prescriptive Analytics
While the move from description to prediction increases the real-world utility of data by a large margin, there is still a significant time investment running the various “what-if” scenarios. If I have my “fortune teller” app where I can plug in the plan for today’s training for an athlete and it can spit out predicted improvement along with fatigue and injury risk, and it tells me “That volume and intensity that you plugged in is a BAD plan,” the question remains—what’s a good plan? Given the data that I have, what would an optimal prescription be?
Consider, just for a second, what will happen when an algorithm comes along that is able to act as a data scientist for each individual athlete and communicate the insights into a simple actionable recommendation for the coach. Consider a machine that can digest (and remember) years and years of various input streams (training files, competition files, heart rate variability data, sleep data, wellness measures) for a given athlete and is then able to use this information to run thousands of simulations of different potential actions in its “brain” in a split second of what would happen if you chose a particular training session for that day and then is able to finally return a simple answer for “This is the optimal action for this athlete today.” This process represents the new world of prescriptive analytics.
Prescriptive analytics takes the data, identifies important features, creates a model of these important features to predict likely occurrences for different actions, and then runs the gamut of possible actions to decide the optimal prescriptive action to take on the basis of the model forecasts. In other words, it takes much of the “heavy lifting” that is limiting the widespread adoption of AI in sports out of the picture. Prescriptive analytics most closely adheres to my preferred definition of AI:
“The study of ‘intelligent agents’: any device that perceives its environment and takes actions that maximize its chance of successfully achieving its goals.”
- –Poole Mackworth Goebel, 1998
The move from prediction to prescription is not a technically easy one. Doing it effectively demands that we create an independent, intelligent agent that is able to perceive and learn from its environment and recommend actions that will maximize its chance of achieving its goals. However, if we accomplish it, prescriptive analytics moves us to a position where the data load on the coach is significantly reduced by creating an artificial assistant coach.If we accomplish it, prescriptive analytics moves us to a position where the data load on the coach is significantly reduced by creating an artificial assistant coach, says @alan_couzens. Click To Tweet
Importantly, especially in the beginning, the coach remains in complete control. We are merely providing an “assistant coach” that can serve an important advising role, offering the recommended range of prescription based on the data. We could think of our AI assistant similar to a golf caddy giving club advice. While useful, it’s ultimately up to the player to make the shot.
While the above technology of “artificial agents” might sound like the science fiction scene that greeted me at the top of Spaceship Earth all those years ago, I assure you, the technology to create these intelligent agents is already here.
In 2015, a computer program called AlphaGo rocked the world of AI when it beat the world champion in the challenging board game Go. More impressive than the victory, however, was the way in which it was achieved. Rather than the old methods of “hand programming” instructions to the computer about how experts play Go, the bulk of the knowledge accumulated by AlphaGo came from the process of optimization through learning that I outline above: Running simulation after simulation of possible actions and, from the data, learning what strategy leads to the highest reward.
In the world of AI, this process is called “reinforcement learning” (RL): The machine is given a goal—in the case of AlphaGo, to win the match; in the case of coaching, to achieve the highest fitness without exceeding a given fatigue or injury risk—and from there it’s left to the machine to determine the optimal strategy. It does this by rolling out hundreds of thousands of simulations, initially with random actions, but over time it learns and refines the actions to those that get the agent/athlete closer to its goal.
Because it trials random actions, the agent can show surprising levels of creativity in the solutions that it comes up with. In the case of AlphaGo, it created new strategies that were previously unthought of despite Go’s 2,000-year history. In our domain, it can prescribe sessions and sequences of sessions that may not fall within the dogma of any conventional heuristic or periodization scheme. Frankly, with its ability to “remember” huge amounts of historical data, coupled with its openness to possibility, it can come up with better solutions than any human could.
While clearly cutting-edge, the level of technology displayed by this decision-making agent is widely available to the point that it has been popping up more and more in commercial “real world” fields. Already companies like Google utilize the technology to optimize the energy usage of its server plants. RL is also being used in dynamically optimizing business recommender systems3, autonomous driving4, and traffic signal optimization5. Basically, in any field that involves optimizing individual decisions based on large amounts of data, RL is generally the superior approach.
You could argue that high-performance sports, a field where athletes can win or lose gold medals by hundredths of a second, represent the ultimate optimization task! Selecting the best possible action, day after day, over a long period of time, is what it’s all about! It follows that those who embrace the advantage of this new breed of artificial intelligence will have a decisive advantage over those who don’t.
My own work is currently focused on that end—bringing the power of deep reinforcement learning approaches to optimize the decision-making processes of coaches in sport. Through my work with HumanGo, we employ machine learning to the mass of athlete data out there in order to create RL agents that can learn to make better decisions than a human alone.
So, where is this all leading?
A Decade from Today
Picture, for a moment, just what a day in the life of a coach might look like 10 years from now…
You wake up in the morning to a message from your “assistant coach”:
“Good morning, Coach. I have received morning data from 19 of the 20 members of your squad this morning. I am still waiting on data from Maggie. Three athletes—John, Wendy, and Sue—were flagged as requiring a closer look this morning:
Wendy didn’t sleep well last night for the third night in a row. I have noticed a strong relationship between her sleep quality and her response to high-intensity training. Would you like to select a lower-intensity action item for Wendy for today?
John looks to be adapting really well to the training at the moment. All wellness metrics have been in the optimal range for the past four weeks. My calculations suggest that John may benefit from an increased intensity of 5% this block. Would you like to increase the planned intensity by 0, 5, or 10% for this block?
Sue has been consistently ranking her stress levels at higher than normal this week and has been showing some signs of maladaptation. I would recommend that we plan an unload week for Sue this week. Would you like me to go ahead and do that?
Thank you, Coach. Have a wonderful day.”
You then jump into your autonomous car to drive you to the pool, but instead of spending the time writing the training plan for today (after a quick review, it has already been sent to the swimmers’ waterproof iPads with lane assignments for the day), you jump on a private video call with Sue about why she’s been stressed out lately.
Everyone arrives at the pool, and the AI assistant coach has sent you a summary for each lane with key focus points for each athlete. Rather than spending the pre-swim time hurriedly writing on multiple whiteboards, you spend your time explaining these focus points to the athletes—demonstrating and connecting.
A few 100s into the warm-up, you get a notification that Jenny’s heart rate is unusually high for this set, and her stroke count is higher than average. You point your smart phone at Jenny to see if there is anything biomechanically amiss, and you get an instant diagnosis. Her range of motion on the right shoulder is 10 degrees lower than her norm. You couldn’t see this with your naked eye in the split-second period of max amplitude each stroke, but your machine friend can instantly go through it frame-by-frame and compare to her norm in a split second, without any high-end computer power, all within your smart phone’s browser.
You stop Jenny on the next length and ask her if her shoulder hurts. Reluctantly, she admits that it is a little sore, and so you update her program (in one click). This one click changes her program to emphasize kick sets for today, sends a note to her physio with a message and screen shot of her range of motion discrepancy, and updates her dose-response model with an injury occurrence for that day so that the ML algorithm can adjust and moderate the load to minimize the risk of future occurrences.
As you roll into the main set, you get another notification from your AI assistant coach that Brian’s heart rate is lower than intended for the prescribed set, and you may wish to increase the load. You stop Brian on the next rep and ask him how it’s feeling, and he says “easy,” so you instruct him to move up a lane to swim with the big boys. You check your app that has live heart rate data for each swimmer, and sure enough, that does the trick.
Delegating a large chunk of the planning to a machine might run counter to what you consider “coaches do best,” but you would be wrong. Simply put, machines run optimization calculations at levels that no human can. On the flip side, humans do human things (conversation, understanding and influencing fellow humans, seeing how things fit into “the big picture” that we call life) in a way that no machine can.As machines take over much of the number crunching, I predict that our obsession with the data side of coaching will diminish & there will be a return to the golden age of humanistic coaching. Click To Tweet
As we move out of the “Big Data” phase and into the “Big Understanding” phase, and the machines take over much of the “number crunching,” I predict that, paradoxically, our obsession with the data side of coaching will diminish and there will be a return to the golden age of humanistic coaching. After all, if we all have access to equally powerful data analysis and decision-making tools, and the difference between coaches is no longer one of “information asymmetry,” what will separate the best coaches from the rest? Perhaps the ability to positively influence, mentor, connect—to guide athletes along that perilous journey from the optimal plan “on paper” to the real-life manifestation of that plan.
As an old coach who has wound his way through his own Spaceship Earth—from the “good old days” of pen-and-paper, face-to-face coaching through the computer age of remote coaching and “data-driven” coaching, I’m more than ready for the next phase of machines and humans truly working together and using their respective strengths in the best possible way.
Since you’re here…
…we have a small favor to ask. More people are reading SimpliFaster than ever, and each week we bring you compelling content from coaches, sport scientists, and physiotherapists who are devoted to building better athletes. Please take a moment to share the articles on social media, engage the authors with questions and comments below, and link to articles when appropriate if you have a blog or participate on forums of related topics. — SF
1. Edelmann-Nusser, J. and Hohmann, A. “Modeling and prediction of competitive performance in swimming upon neural networks.” European Journal of Sports Science. 2002;2(2).
2. Rossi, A., Pappalardo, L., Cintia, P., Iaia, F.M., Fernàndez, J. and Medina, D. “Effective injury forecasting in soccer with GPS training data and machine learning.” PLOS One. 2018;13(7).
3. Zhao, et al., “ATRank: An Attention-Based User Behavior Modeling Framework for Recommendation.” Presented at the 32nd AAAI Conference on Artificial Intelligence (AAAI-18), 2018.
4. O’Kelly, M., Sinha, A., Namkoong, H., Duchi, J. and Tedrake, R. “Scalable End-to-End Autonomous Vehicle Testing via Rare-event Simulation.” Stanford University paper.
5. El-Tantawy, S., Abdulhai, B., and Abdelgawad, H. “Design of Reinforcement Learning Parameters for Seamless Application of Adaptive Traffic Signal Control.” Journal of Intelligent Transportation Systems. 2014;18(3):227-245.