Alpha Go and Soccer
If anyone hasn’t seen this documentary, I highly recommend it.
https://www.youtube.com/watch?v=WXuK6gekU1Y
Although I never played go, I was an experienced chess players and have built reinforcement learning algorithms to play abstract board games. I think there are a few key take aways from a movie like this and I hope to bring to life some of those key ideas as they pertain to soccer.
- Machines will change the strategy of soccer
- Actions in games are meant to be played with an understanding of how your opponent might react
- A machine needs to take input from relevant domain experts
Machines Will Alter Strategy
In the movie one of the most interesting scenes is this:
I think teams that use algorithms will start to make moves that seem bad at first to humans. If we are to strategically improve, we must do things that distinctly change the opinion in terms of what is a good strategy. It will at first seem like the teams that take on these risks are doing things that are inherently bad or wrong. This has happened in other sports (specifically basketball), when the Golden State Warriors started shooting more three point. The Warriors realized that three point shots had a higher expected value per possession then two point shots. Now every team in the NBA takes more three point shots.
Understanding your Opponent
One of the key thing that Alpha Go did was it predicted what sort of move an expert player was likely to take against it. This is actually interesting from a soccer perspective as well, because if the opposition left back has the ball at midfield, the play can progress in an infinite number of possibilities, but some are more likely than others. If you want to devise the best strategies to defend against the opposition you should understand that some moves are more likely than others. For example the opposition left back is much more likely to pass to a defensive midfielder than to a right winger on the opposite side of the pitch (especially in possession based teams). Understanding the game from your opponents viewpoint allows you to devise strategies to defend and attack against your opponent.
Machines need Human Input
One of the key breakthroughs in reinforcement learning was the idea of imitation learning. The simple idea is that in the infinite space of the world, machines work best when they can first learn what might be a best path from an expert. This sort of learning from an expert allows the machine to be calibrated at an already strong level instead of learning from the infinite space that exists in sport. It is also similar to the real learning that a human would undertake, to put it simply one of the ways Cristiano Ronaldo got better at soccer was watching some of the greats before him. He learned ideas of when and where he should shoot the ball and some of the tricks that he uses by observing other professionals. In statistical models it is important to really have that guidance of experienced soccer professionals to learn strong strategies.