The Computer That Beat Two Million Humans at Fantasy Football
An AI program is already in the 99th percentile—now can it win?
Steven Cherry: Hi, this is Steven Cherry for IEEE Spectrum’s “Techwise Conversations.”
For decades, artificial intelligence was a field full of promise but little visible progress. For the past 15 years, though, AI researchers have been making steady—and quite visible—progress in computer vision, language translation, artificial speech, exoskeletons, and more.
The successes owe a lot to progress made playing games. The turning point, 15 years ago, was when IBM’s Deep Blue beat the human chess champion. More recently, when Deep Blue’s successor software beat the human “Jeopardy!” champion, it was a sign that the jobs of so-called knowledge workers weren’t beyond the reach of computer automation. That software is being turned toward medical diagnosis.
Now, researchers in the U.K. and Greece are hard at work trying to create the best player in the world at fantasy football, a game that attracts millions of people worldwide. They expect it to reap benefits for what they say are surprisingly similar tasks, such as how best to dispatch emergency responders.
In fantasy football—in this case, football in the European sense, meaning soccer—people construct teams out of existing football league players, 15 to a side, no more than 3 from any one team, within a budget. A season lasts 38 weeks, and only a limited number of changes are allowed without penalties.
A software program is already better than 99 percent of the 2.5 million people who play in the Fantasy Premier League, or FPL, which is based on the English Premier League. Now the question is, can it beat the remaining 2500 best competitors in the world?
My guest today is Sarvapali Ramchurn of the University of Southampton. He and his coauthors, Tim Matthews, also at Southampton, and Georgios Chalkiadakis, who is at the Technical University of Crete, are presenting a paper at the 12th [26th] annual conference of the Association for the Advancement of Artificial Intelligence, or AAAI, in Toronto. The paper’s title is “Competing with Humans at Fantasy Football: Team Formation in Large Partially Observable Domains.”
Sarvapali, welcome to the podcast.
Sarvapali Ramchurn: Thank you, Steven.
Steven Cherry: Maybe you could explain what is hard about dispatching emergency responders and how fantasy football is similar to it?
Sarvapali Ramchurn: So, when we looked at the problem of dispatching emergency responders in very dynamic and uncertain domains in a previous project at the University of Southampton called Aladdin, we found that forming teams of emergency responders is a challenging task because it requires you to forecast what’s going to happen in the environment and how the team is actually going to perform based on the composition of the team—so, based on the different abilities you have in the team and how they synergize to perform tasks better than possibly splitting them apart and distributing them over space. So team formation is a significantly challenging problem for existing emergency responders, and typically what they do is they train a lot, and they try and understand each other and try to understand what other capabilities they might need for different kinds of emergencies. But this is not entirely feasible unless you really analyze the statistics of how these emergency responders perform, both individually and as a team. So, that’s the kind of domain we’ve sort of been studying and trying to apply our techniques to in order to help them. We also look at forming teams of humans and robots, where the robots here could be unmanned ground vehicles or unmanned aerial vehicles, and these have very different capabilities and operate in very different environments.
Steven Cherry: Yeah, so you take what’s called a Bayesian approach. Maybe you could explain briefly what that means.
Sarvapali Ramchurn: A Bayesian approach basically means that you start with formulating a “prior” over what you believe the state of the world is and what’s going to happen with respect to certain variables in the environment, and then you keep updating those based on what observations you get. And basically you update the probabilities that certain things will happen in the environment. So in the case of emergency response, you’re trying to update what you think the emergency responders are going to do or how well they’re going to do, so you keep updating it after time, and that also takes into account various dependencies you might have between various variables in the environment. So the performance of an emergency responder in one place might be dependent on the UAVs that it has in that area, and then multiplying all these probabilities out gives you a probability of how well the team will perform.
Steven Cherry: So in the case of fantasy football, you don’t have observations, you have the actual performance of the actual player in an actual game.
Sarvapali Ramchurn: Yes, exactly. So the Bayesian approach also helps you look at what happens in a sequential team-formation problem, so when things keep happening and keep getting...and observations here are actual performances in the real world.
Steven Cherry: So, how does the software measure up right now? You say it’s in the top 1 percent of all players—does it actually play in the league?
Sarvapali Ramchurn: So, basically what we’ve done is take data from the previous season—or one season before that—and played out our automated fantasy player against the rules of the game. So we looked at the match outcomes for every single week; we played the team that was suggested according to the automated player and scored its performance based on the real scores of the fantasy football league. So, basically if we had entered for real, if we had actually played it in 2010, the performance would actually have been what we got, and we’re doing the same for the 2011–2012 league. So you can play things retrospectively in this case because it doesn’t really matter whether you’re playing it live—basically, if you get what I mean, it’s just based on the number of points that every player gets for a specific match, and therefore you can play everything retrospectively. So we’re planning to enter for real now, in the real league, and see whether we can actually show off how well the fantasy football player is playing against actual humans every week.
Steven Cherry: And you’re planning, I gather, to add some human intelligence—for example, information about last-minute injuries and other bits of information?
Sarvapali Ramchurn: Yes. So what we think is going to be the winning formula, is combining the human intelligence with the machine intelligence—basically, the machine being able to foresee what’s going to happen. So, using this Bayesian approach: being able to play out various futures and formulating probabilities of things happening in the future in a way that humans cannot do. And we saw an example of that today in Henry Kautz’s lecture at AAAI, where he was talking about how Kasparov and the machine play better than Kasparov on his own or the machine on its own, and then when you have a team of players playing with the machine, then they play better than Kasparov and the machine. So we think the same thing will apply to fantasy football, just because humans have got a better understanding of players. There are already some sources of statistics about player injuries and statistics that are more detailed than what we have used so far, so we are planning to exploit those in order to improve what the machine can do but also get input from humans to basically formulate a better strategy about how to go about playing fantasy football: picking out the players, for example, that are sort of growing in terms of performance or picking those players that are likely to perform well against teams that they can attack the defense better, for example.
Steven Cherry: So, I mentioned the IBM Watson software which beat the human champion in “Jeopardy!” is being turned toward medical diagnosis right now. Is there any particular worldly problem that you think your software could be directed toward?
Sarvapali Ramchurn: So, in terms of real-world problems, I mentioned the idea of coordinating humans and UAVs, for example, in disaster response. That’s one example of where we think sequential team formation is a significant challenge just because there are lots of uncertainties in the domains that these teams of humans and UAVs will operate in. And recently there was a piece by Nick Jennings on the BBC news talking about the Orchid Project, which is exactly about this domain where you might have human emergency responders sending out UAVs to take pictures of certain areas, but in the future you will have even more of these kinds of scenarios where humans and robots and even software will have to work together. So, we think this sort of teaming, this agile teaming of humans and agents and robots, will be more pervasive as we go into a future where robots become cheap, robots become more deployable and become used in more civil applications than how they have been used so far.
Steven Cherry: Fantasy football works at the level of individual players which are spread out among different teams. I’m just wondering: If you give the software just all and only the players on a single team and do that for two teams, you would have software that would try to predict who would win an actual football game. Have you thought about that?
Sarvapali Ramchurn: Yeah. So, we’ve been looking at analyzing the statistics to predict whether a certain team would win, and we do some of that already within the algorithm, but we don’t go too deep into that because there are already lots of research being done on predicting football match outcomes. We’re not trying to compete with that, but we’re trying to reuse some of that to play better fantasy football. So lots of these statistics are already being used by betting companies, for example, to set odds and by people who just like to research football, so there are lots of papers on predicting match outcomes. So what matters in our case is the sequential nature of the problem—so, when you need to pick a team that needs to perform well over a long period of time, not just for the next game. So, what could be probably looking at is looking at the best prediction for the winner of a league rather than for the winner of a match. So, looking at how many points a team might acquire over multiple matches given the formation of players it chooses to buy at the beginning of the league. So that’s the kind of thing we could do. So if you wanted to put a bet on the winner of a league, maybe our software could be used.
Steven Cherry: Or if you were the owner of a team and trying to figure out how to assemble the best players for next season. As we record this, you’re presenting your paper this week, so good luck with your session, and thanks so much for taking the time to talk with us.
Sarvapali Ramchurn: Oh, thank you very much for inviting us to talk about the paper. We’ve really enjoyed all the publicity, both for our projects and for AAAI.
Steven Cherry: We’ve been speaking with Sarvapali Ramchurn of the University of Southampton, who has coauthored a paper that chronicles the remarkable progress he and his colleagues have made in artificial intelligence by playing fantasy football. For IEEE Spectrum’s “Techwise Conversations,” I’m Steven Cherry.
Announcer: “Techwise Conversations” is sponsored by National Instruments.
NOTE: Transcripts are created for the convenience of our readers and listeners and may not perfectly match their associated interviews and narratives. The authoritative record of IEEE Spectrum’s audio programming is the audio version.