12 July 2015

I don’t know how many people have experienced playground football (the roundball variant of the game that is). This is football where a cluster of kids chases around after a ball, fighting for posession. No one passes, it’s just a scrum. Apparently this was how the game used to be played back when it was invented. And when you think about that, you can understand how different games such as American Football, Aussie Rules, Rugby and Soccer all had a similar origin. They are separated by the way the ball is passed, handled or dribbled.

Anyway, in the playground game, it always seemed to me that there were only two or three players that ever got a touch on the ball. Often they were amoungst the most skilled dribblers, but somehow I think David Beckham would never have survived in that variant of the game.

Sometimes the field of research gets a little bit like playground football. A cluster of people chasing the ball. A lot of activity and shouts, and perhaps the occasional scrabbled goal, with multiple players claiming the last touch and very many different narratives about how the ball came to be near the goal in the first place.

Just back from ICML, and can’t help but feel that this is where the our field is at right now. This isn’t necessarily a problem, just an observation. There is a new shiny ball in the playground and lots of people want a kick at it.

So should you just dive in and try and have a kick yourself? Maybe, but then maybe not. It depends on how good your ball skills are.

First we need to do the analysis, and as ever in questions of strategic decision making, the right solution is to turn to the (interpreted) works of Bert Kappen. Or more generally the domain of stochastic optimal control. The lesson learnt, is that when the system is stochastic, rather than diving immediately in, it can be better to stand back and see how the game evolves.

Now, stochasticity here is a subjective phenomenon. The game doesn’t need to be inherently random for these ideas to come into play. The actual condition required is that the fluctuations in the state space of interest (i.e. the coordinates of the shiny ball) occur more rapidly than you are capable of responding to. Note that this is subjective. There are those in the middle of the scrum who have excellent ball control. This was developed by playing with the ball before everyone else noticed what fun it is. These people are ‘in the zone’. They can see where the ball is going and how to intercept it, they already know how to play, but it’s crowded and there are stray boots, and they are taking it on the shins.

The principles underpinning ‘Kappenball’ teach us patience. We expect that the dynamics of the game will evolve. In the long run, the right way of playing football is to position yourself intelligently and to wait for the ball to come to you. You’ll need to run up and down a bit, either to respond to how the play is evolving or to get out of the way of the scrum when it looks like it might flatten you. But if we assume ergodicity, the ball will come to you, the time frame is dependent on how cannily you are positioned.

So for people like me whose ball skills are either rusty or lacking, here is my advice: don’t dive in the scrum. Take up a position on the wings and start shouting encouragement. When you see a good opportunity, call for the pass. If you do get the ball, don’t dwell on it too long. Even if you never score, the connoisseurs of the game will still appreciate the spirit in which you are playing.

Neil Lawrence is a Professor of Machine Learning. From 1998 to 2000 he had an undistinguished 11-a-side football career as a central defender for St John’s College Cambridge’s graduate side. He scored two goals. Both from corners. He is fond of theoretical mathematical proofs that justify inactivity.

blog comments powered by Disqus