Hoang M. Le, Peter Carr, Yisong Yue, Patrick Lucey, 'Data-driven Ghosting using Deep Imitation Learning'
Why it's worth your time
'What if the computer could learn how teams defend, and tell you what an alternative could have looked like'. What's not to be excited about?
What it says
After working out the roles of players, the researchers use recurrent neural networks to train where players should be, in any given defensive situation, to have the best chance of stopping the opponent from scoring.
Given that spells of possession can be quite long, the 'ghosts' need to be able to correct their positioning if play develops unexpectedly. To quote the paper: "we want to train a model that can learn to recover from its own prediction mistakes so that the model can be robust over long sequences of decisions."
The model is first trained to learn how each of the ten outfield roles would behave on an average team, and then how multiple players would interact together.
At first, the ghosts are learning to be the average of how all teams, and players in those teams, defend. But the researchers then go on to show how the same process can be applied to learn specific team/player tendencies.
 You could rely on players being tagged as 'centre-back', but sometimes the tags are wrong and sometimes - like after a free-kick or after a substitution - players might find themselves filling roles they weren't initially tagged as.
Further note: As with anything, there are precursors. This work was inspired by/built on work done in the NBA by the Toronto Raptors (here's a Zach Lowe piece from 2013).
What's cool about it
Firstly, there's a 2:35 minute video produced by/with the research team available here. It's worth watching and shows some examples of the ghosts in action.
The idea of being able to see what could have been done gets to the heart of a lot of problems with trying to evaluate defending. Also, there's something quite funny about seeing a ghost in a position to intercept a pass, have the ball go right through them, and then have to change their positioning because the real-life player wasn't positioned as well as them.
Finally, assuming that the model is good and the process is sound, the possibilities of this excite me. I'm not sure to what extent the following is possible, but the idea that you could give this process a load of, say, Paolo Maldini tracking data, and be able to show youth players the Paolo Maldini ghost overlaid on their positioning, is very cool.
In a perfect world, could you, one day, identify several different players who you think represent the best practitioners of specific roles and then train ghosts on them? 'This is how a Virgil van Dijk-type would defend', 'this is how an N'Golo Kanté-type would defend', etc. You could do it to show forwards how to defend from the front, makeshift full-backs the moment their positioning went wrong, maybe even have to defend in a new formation. Those are big, big 'ifs'... but imagine...
'Research in Focus' is like SparkNotes for football analytics: summarising and analysing the best research out there. Get Goalside supporters get access to every post, with a rotating selection free to access for all. Follow this link for the list of all Research in Focus pieces.