A few weeks ago, I was invited to join my newsroom’s Fantasy Premier League. As an eager new American staffer at a British newspaper, of course I accepted the chance to join in assembling an imaginary team of real-life players and scoring points based on their actual statistical performance. One doesn’t want to be rude, after all.
That’s when the problems started, chief among them that I had no idea whom to select for my team. I enjoy football as an aesthetic experience but not so much as an analytical one; I’m no great student of formations, form, the transfer market, injury reports or management politics. So I searched for the shortcut that any clueless empiricist with a laptop would these days: machine learning.
On GitHub, I discovered the code developed by some real-deal programmers for a system called AIrsenal and stole it, er, implemented it on my machine. AIrsenal uses a “Bayesian approach”, estimates a “conditional distribution”, leverages “Monte Carlo integration” and uses more than a couple of Greek letters in its underpinning maths. Select, copy, paste, enter. A curtain of blocky green numbers geysered up my black terminal window, simulating fixtures yet to be played, goals yet to be scored, victories yet to be celebrated.
Thus artificial intelligence assembled my fantasy team; all I did was obey its instructions. This was deeply current and a bit weird, not just fantasy football but artificial fantasy football. Nevertheless, I was and remain thoroughly proud of “my” creation; it had geysered on to my screen, after all. In honour of some other influential newcomers to the (real) Premier League, my team is called American Billionaire (AB). I have no plans to disobey the computer’s recommendations in the months ahead.
The FPL rules, quickly: you get £100mn to assemble a fantasy squad of 15 real-life players. As they play in real-life games, you earn points from a menu based on their performance (eg, midfielder’s goal = 5, keeper’s clean sheet = 4). Each week, you can buy and sell players, whose prices fluctuate based on their popularity. Nearly nine million people are currently playing.
“Computers are useless,” Picasso once said. “They can only give you answers.” Pablo’s right enough, but the answers — my fantasy roster, the result of cold, mathematical optimisation — suggested the questions. Who are these human footballers the computer had divined in its fury of calculus? I had to know. Rather than divorce me from the humanity of the sport, the AI drew me closer.
When Newcastle’s Bruno Guimarães appeared in my squad, the output of an invisible algorithm, I learnt everything I could about the real man. He’s the son of a Brazilian taxi driver and wears his father’s cab number, 39, on his shirt. Chelsea’s Reece James, who is right-back for American Billionaire, dreamt of playing for the club (Chelsea, not AB) as a small child kicking a ball all night in the local park. And Manchester United’s Anthony Elanga, AB’s reserve midfielder, is fluent in French and Swedish, and aspires to learn Spanish and Portuguese.
There is a democratising power of technology. It has the ability to improve the quality of human knowledge and to accelerate its dissemination. This power plunges us faster and deeper into pursuits, where we can all mine the rich veins at their core. But all the while, I could imagine murmurs and objections rippling through the newsroom: he’s cheating!
I spoke with AIrsenal’s developers, Nick Barlow and Jack Roberts, former particle physicists and now researchers at the Alan Turing Institute in London. Like most good ideas, AIrsenal started, about four years ago, as a conversation on a train.
“The problem is not in the complexity in the game, it’s in getting the information in,” Barlow said. “Read newspapers and you know that this player is arguing with the manager or about to be sold to Barcelona. Our algorithm doesn’t know anything like this.”
Roberts added: “It’s much more difficult to represent the state of the FPL world than a chess or go board.”
Nevertheless, AIrsenal (opinions on pronunciation differ) has had its oracular moments. “We have our own move 37,” Barlow said, referencing the unthinkably beautiful go move by the AlphaGo AI in 2016. “It picked Joe Hart when he moved to Burnley.” They also eschewed Liverpool’s Mohamed Salah when he returned from the Africa Cup of Nations last year — an unpopular but wise move, it turned out. “Maybe it was AIrsenal’s lack of emotion,” Roberts said.
I also inquired about the ethics of my decision and was pleasantly relieved. It takes a certain level of expertise just to run the code, they said (select, copy, paste, enter) and my approach is just a different way of achieving what I could have gleaned by, say, reading gossip about David Moyes on Twitter. “The more, the merrier,” Barlow said. “I, for one, welcome our robot overlords.”
AIrsenal finished 976,423rd last season, solidly in the top 10 per cent. If they ever do win the top prize, the Alan Turing Institute’s ethics board has decreed that their creators give it to charity, and that’s just fine with them. I plan to do the same.
As I write, AB sit third from bottom, in the relegation zone.
Oliver Roeder is the FT’s US senior data journalist and author of “Seven Games: A Human History” (WW Norton)
Follow @FTMag on Twitter to find out about our latest stories first
