Using Avatars to Optimize Human Performance

Alan Couzens, M.Sc.(Sports Science)

Jan 9, 2020

"Sometimes your whole life boils down to one insane move" - Jake Sully (Avatar)

The 2009 James Cameron film, Avatar, had a big impact on me. As the type of person whose link to 'the real world' is often tenuous at best, this should come as no surprise :-) The whole idea of moving through worlds, exploring and learning via an artificial representation of your self is an interesting concept. It's a concept that has obvious upsides! In the film, when Norm's avatar died, it was merely game over in the simulation for Norm. There were no ill effects to Norm in the 'real' world. This lack of consequences has obvious implications on how adventurous & bold we might be in our exploration. I'm sure we've all experienced a similar effect when playing video games: The consequences of our character 'losing a life' are far less than the consequences of us losing a life(!) & so we're far more likely to take risks and 'try stuff' that we wouldn't consider in real life like jumping from one building to another. It's in this 'trying stuff' that we often learn important strategies that help us win the game.

This idea of building a representative model to experiment on extends beyond the world of Hollywood to the world of science. For example, in the field of oil and gas exploration, digital avatars (geological models of well sites) are often used prior to actual drilling to help determine the optimal development plan for a particular site. From Wikipedia...

At HumanGo.ai, our development of Artificial Intelligence solutions for athletes and coaches in the field of sports science has been greatly benefited from our CEO, Eric Abecassis' previous experience in Oil and Gas modeling for the industry leader, Schlumberger. At HumanGo, we are employing a very similar approach of building a representative model of each athlete that we can experiment on with the courage that comes from knowing that if we screw up, it's just 'game over' in the simulation not 'athletic career over' or 'long term health over' as it would be if we made similar errors in real life.

As it turns out, in may ways, triathletes are quite similar to rocks! :-) In the case of oil and gas extraction, stresses are applied to the underlying geology as extractions are made. If improperly dosed, these stresses can be catastrophic! Furthermore, the energy within the well is a finite property & a quantitatively different property for each individual site. Therefore, the extraction processes selected & the 'aggressiveness' of the plan will be different depending on the dynamics of the individual avatar (model representation) of the site. Hopefully, it is clear from the above how this might relate to our domain. Replace 'reservoir' with athlete and the decision making process is quite similar.

By examining the individual relationship between load and state metrics (training stress, life stress, power, speed, heart rate, heart rate variability etc.) we are able to build an accurate dose-response model specific to each and every individual athlete. We call this individual model the athlete's 'Avatar'. This model tells us not only the expected performance improvement given a specific training dose but also how the training dose will affect the health & fatigue state of the individual athlete. This is important and something that is sadly lacking from most current performance models. It is all well and good to say that a CTL of 300 is predicted to produce a World Class performance but if the same dose will also kill the athlete in the process, the performance model is of little practical use :-)

So, just like the movie, once our Avatar, or 'digital twin' is built we can (safely) let it loose on some crazy simulations where we try each and every possible sequence of training actions under the Sun to determine which ones make the our digital twin better & which result in 'Game Over'. Again, importantly...

..with the creation of the Avatar we are able to try these things virtually in the simulation & save only the best (lowest risk, highest reward) actions for real life!

This whole process of Agent trial and error within a simulated environment is an important subcategory in Machine Learning termed Model-based Reinforcement Learning.

Model based Reinforcement Learning differs from many Machine Learning approaches in that, beyond pure learning, a key objective is developing understanding of the way the environment works and a consequent ability to plan prior to acting. This planning process is shown diagrammatically below...

Data is collected from training sessions conducted in the 'real world' and an 'avatar virtual world' that is specific to the athlete is built. In this world we try some initially very random, potentially crazy actions and observe the result - how beat up does the athlete get when, out of the gate we give him/her 6 hours of training every day for 2 weeks etc. :-) Over time, as we learn, our actions get a little less crazy as we select more of the optimal actions for our agent. By the end of the process, we are left with the best (highest reward, lowest risk) actions for our individual agent to take from any given state. We then take only those best actions in the real world.

As the avatar develops more and more understanding of the environment, they can 'look ahead' & imagine what the consequence of a certain series of actions will be, without having to actually perform the actions. This distinction may appear subtle but it's important. In the case of a robot learning from 'pure' reinforcement learning, it will have to drive itself off a cliff many times before it learns that a cliff face is a bad place. If, on the other hand, it has a (gravity complete) model of the world in its 'head', it can look ahead to the potential consequence of heading for the cliff before actually moving anywhere and can decide to head in the other direction. Or, in our world, our agent can think ahead and say "Keeping that training camp rolling for another week will almost certainly lay me low for several weeks after." In this sense the Avatar learns without requiring the athlete to actually experience 'real' overtraining. This is a very good thing!

This whole concept of planning before acting is central to good coaching and might almost fall in the "duh, that's what all coaches do" category. And you're right, of course, in the same way that you would be right in saying that good chess players simulate the consequences of moves in their head before actually moving the piece. The major difference between machines and humans is that machines do it better. Even in a very clearly defined rule based environment like Chess, machines have proven their clear dominance due, largely, to one simple fact - they have the processing power to 'look ahead' many more steps and search more of the 'decision tree' for the best possible action sequence. While a human may think a few moves ahead, the machine can think a few hundred moves ahead. Additionally, the model of the environment in Chess is very clear and easy to hold in the brain. In fact, it's the first thing that any aspiring player learns - the rules of the game. We know, for instance, where a Knight could potentially move. However, in athletic training, the player moves are individual and not so clear cut. In our world, one strong 'Knight' might have the ability to jump 6 squares ahead, while a weaker knight might only jump 2. In addition to thinking ahead many moves, the coach must also keep these individual 'piece' abilities in mind at all times (not to mention that the abilities change on a daily basis). This is challenging to say the least!

In my experience, the difficulty of holding all of these individual athlete models in one coach's head, often results in remarkably similar training plans being given to all athletes. The model becomes a generic one size fits all vs a model specific to the individual athlete. It's our belief at HumanGo that every athlete is an individual who deserves a training model and plan specific to them. The HumanGo team is currently working very hard towards that end and by Summer you can expect to be saying hello to your own personal Avatar - an individual digital representation of you, the athlete, that you can play with, beat up as needed :-), and, most importantly, learn from.

Train smart,

AC

  Tweet