A Unified Theory of Basketball: Inside the machine

Photo by Liu Guanguan/CHINA NEWS SERVICE/VCG via Getty Images   Photo by Rachel Murray/Getty Images for Samsung
Photo by Liu Guanguan/CHINA NEWS SERVICE/VCG via Getty Images Photo by Rachel Murray/Getty Images for Samsung /

Some people will tell you that the Golden State Warriors are the future of basketball — a migration towards overlapping skill sets, a dozen similar 6-foot-8ish frames playing multiple positions, all switchy and, most importantly, twitchy from behind the 3-point line. This narrative hypothetical has been well-circulated — that we’re moving towards a version of basketball defined by a few unicorns and a surplus of redundantly versatile supporting players.

But what if the future of basketball doesn’t include players at all?

At the MIT Sloan Sports Analytics Conference at the end of February, researches Thomas Seidl and Aditya Cherukumudi presented a paper called Bhostgusters: Realtime Interactive Play Sketching With Synthesized NBA Defenses. The abbreviated summary is that they have developed an app which would allow an NBA coach to sketch a play on their tablet and then receive an animated estimation of how the defense would respond, taking into account the spatial tendencies of the specific defense as well as the players on the floor, defensive roles, the players’ foul trouble and potential fatigue based on minutes played, as well as the time and score. The program also returns an expected value for the offensive play the coach has drawn, given the projected defensive response.

You might have questions about the accuracy of the system’s defensive projections but the present efficacy of the system is much less important than the path it represents. The work of Seidl and Cerukumudi builds, at least thematically and creatively, on research that has been presented at the Sloan Conference over the past few years — Andy Miller and Luke Bornn on automatically identifying specific strategies and play sets from spatial tracking data; Avery McIntyre, Joel Brooks and John Guttag using machine learning to identify when a screen has been set in spatial tracking data; Alexander Franks, Andy Miller, Luke Bornn and Kirk Goldsberry on automatically estimating and assigning defensive responsibility to individual players in spatial tracking data; and Luke Bornn, Kirk Goldsberry, Dan Cervone and Alex D’Amour on tracking the expected point value over the lifespan of a possession to evaluate individual player decision making.

Read More Unified Theory of Basketball: In search of perfection

There are a few direct threads through all this work — spatial tracking data, machine learning, and automating the definition and categorization of different elements of movement on the court to allow data to be sliced in new and more meaningful ways. The work of Seidl and Cerukumudi represents something of a holy grail in that it has real-time applications. And even if the computerized intelligence in their system doesn’t yet measure up to the intuitive evaluations of experienced NBA head coaches, the fact that it would, someday soon, seems to be indisputable.

As I was sitting in the back of their presentation, I couldn’t help but drift towards future iterations of their work. If this system can project how each team would defend a certain scenario, could it then also design an optimal play to defeat each defense, including the very best in the league? Or, flip that around and design an optimized defensive strategy to stymie specific offensive sets? All backed with data and not just subjective human estimation? What if this artificial intelligence could design new offensive strategies that humans haven’t devised yet? And once we get to that point, what use is there for human coaches at all, at least in terms of designing on-court strategies?


Every fan will have their own favorite moment from this season. However, almost every fan and media member will be able to point to the worst moment of the season — five minutes into opening night, Gordon Hayward hitting the floor, both benches looking on in horror, a hushed arena and Kevin Harlan quietly repeating, “Hayward has broken his leg,” on the television broadcast.

#BanInjuries is a popular hashtag for a reason and, at some point, every NBA fan base will be wringing their collective hands and lamenting the frailty of the human body. NBA teams have invested innumerable resources in the pursuit of sustaining health for their players but physical chaos continues to rear it’s ugly head. If we really want an NBA without injuries, there is really only one possible solution…

…we need robot basketball players.

A recent report by Price Waterhouse Cooper, an accounting and consulting firm, found that as much as 38 percent of American jobs could be replaced by automation by the year 2030. This steady march of robotics and artificial intelligence is a discomforting existential threat and for many people the intellectual defense mechanism is trusting in the inherent human ingenuity and creativity that defines so many professions.

It’s hard to imagine a machine replacing athletes, especially in a sport that relies so heavily on physical and intellectual improvisation but a basketball league with no humans involved isn’t that far-fetched.

An artificial basketball league is probably not going to be a three-dimensional affair with plastic androids setting hard fouls and swishing jumpers. It’s likely to be an entirely digital entity, fans sitting down on their couch to watch simulated players playing a simulated game — an extension of what we already have with something like the NBA 2k video game series.

As a viewing experience, NBA 2k is already making inroads. They currently have over 550,000 YouTube subscribers for an account that heavily features clips of in-game play. Their Twitch account has over 100,000 followers. Seventeen NBA teams have signed on to run franchises in the inaugural NBA2k League. As strange as it might sound to basketball fans in the older age brackets, hundreds of thousands of people already willingly watch 48 minutes of artificial basketball.

The key difference between this, and the entirely artificial basketball environment I’m suggesting is that watching an NBA 2k video or live stream is embedded with humanity — actual human beings are on either side, controlling what happens on the court. This often includes trash talk and running commentary by the human players, which I’m told is a big part of the appeal of watching. However, that’s not always the case.

I asked Nekias Duncan, prolific basketball writer for FanRag Sports and NBA 2k afficianado, if he had ever watched an entirely simulated NBA 2k game (that is, no human players involved). He admitted, “Yeah, I did it a lot over the summer. It’s mostly to pass the time, but I also make a custom roster every year so I test them out that way. I’m…not sure I could stomach a season’s worth. And I say that as someone that plays/watches 12 minute quarters.”

When I asked what would keep him from watching an entire season of alternate reality he pointed to some of the heavily human details, saying, “The human emotion aspect especially. You don’t see techs (unless you have a mic and curse), no arguments or choppiness. The gameplay itself isn’t fluid enough to substitute real basketball. Altering tendencies helps some, but there’s no real read-and-react feel.”

In imagining this entirely simulated basketball environment, Duncan’s response highlights what would seem to be the two most obvious challenges — simulating the visual experience and the emotional and strategic complexities of actual humans playing basketball.

The visual component seems like the more viscerally disruptive challenge but, it honestly may be the lesser of the two. In Kofie Yeboah’s history of the NBA 2k franchise, senior producer for the series Rob Jones talked about how his team is constantly using new technology to improve the game play, telling Yeboah, “We get wind of new tech. We try to figure out how to maximize what we’re doing for that particular technology. We pushed 4K the moment it came out. We pushed VR last year. We’ll try something in every single way that we can.”

Graphics that are indistinguishable from reality are not just a pursuit of NBA 2k either. All sorts of video game companies, movie production studios and other entertainment entities are working on this problem simultaneously. The technology to make this viable has all sorts of application and thus the final piece of the puzzle could be assembling existing tech from different industries. Regardless, given how much the technology behind computer generated images and graphics have evolved over the last decade, it would take a leap of faith to assume that at some point in the near future, seamlessly replicating visual reality wouldn’t be possible.

As far as solving for complexity, I’ve often wondered (as a video game novice) how close something like NBA 2k is to reality and whether it could be used to simulate the on-court interaction of teammates to help evaluate potential trades or free agent acquisitions. LeBron James was way ahead of me.

To help James save face, let’s assume the inaccuracy here is a result of the game’s inability to project injury recovery or simply his own incredible skill as an NBA 2k player. But the fact that someone like LeBron would even imagine using NBA 2k as a remotely plausible projection of himself and a future teammate speaks to how much these games have advanced in mirroring the complexity of reality.

Understanding the incredible complexity of a real human professional basketball game is also being approached from multiple angles. Game companies like NBA 2k have been working on this for years, albeit to create a version of reality that is optimized in a certain way (maximizing the enjoyment of the human’s controlling the game).

Team analytic departments have been trying to map, analyze and evaluate the complexity of games for year as well, looking to gain competitive advantages. This process was well under way nearly four years ago when Zach Lowe wrote his seminal piece for Grantland on the Toronto Raptors and the “ghost defenders” their analytics staff was training on SportVU player tracking data.

"The Raptors’ analytics team wrote insanely complex code that turned all those X-Y coordinates from every second of every recorded game into playable video files. The code can recognize everything — when a pick-and-roll occurred, where it occurred, whether the pick actually hit a defender, and the position of all 10 players on the floor as the play unfolded. The team also factored in the individual skill set of every NBA player, so the program understands that Chris Paul is much more dangerous from midrange than Rajon Rondo, and that Roy Hibbert is taller than Al Horford.That last bit — the ability to recognize individual player skills — is crucial for the juiciest bit of what the Raptors have accomplished: those clear circles that sort of follow the Toronto players around and have the same jersey numbers. Those are ghost players, and they are doing what Toronto’s coaching staff and analytics team believe the players should have done on this play — and on every other Toronto play the cameras have recorded. The system has factored in Toronto’s actual scheme and the expected point value of every possession as play evolves."

What this system was working towards was recognition of the optimal strategy for different scenarios, which all sprung from a re-creation of the complexity of a game. It was one of the precursors to the Bhostgusters system of Seidl and Cerukumudi (and one of the reasons for the somewhat bizarre name). The proliferation of this technology and these techniques has been happening for years and the growth curve is exponential. Each new step forward clears the way for multiple leaps in different directions. We’re drawing closer to being able to accurately measure the complex elements of a basketball game. From there, it shouldn’t be hard to create it from scratch.

It may seem absurd now, but at some point there will be the technology to create a simulated basketball game the includes the inherent chaotic complexity of reality and is visually indistinguishable from actual humans actually playing the game. When that moment arrives, could it put human basketball players out of a job?

From a financial standpoint, the appeal of automation for a basketball league would be no different than any other industry — reducing cost. The technology to create, maintain and broadcast an entirely virtual league would certainly require a significant investment but it would pale in comparison to the cost of maintaining 30 teams-worth of human players and support staff, not to mention managing facilities and the entire apparatus of the league itself. The league would also simplify the business of battling declining in-person attendance for a product that is often best enjoyed remotely but still relies on the revenue streams from arenas to be viable.

The technology to make this happen will exist at some point and the financial incentives to try it will almost certainly still exist at whatever point the requisite technology coalesces. So why is a completely digital sports league so hard to imagine?

I spoke to several Step Back writers about this idea, all of whom professed to regularly play NBA 2k (I assumed the barrier to belief would be lower for them than someone for whom the existing digital alternatives to actual basketball are still so foreign.) To a person, they were skeptical and no one thought they could ever imagine themselves being invested enough in a completely digital basketball league to watch regularly. Their reasons varied but the best way to summarize it was that no one could imagine themselves rooting for algorithms.

Even if the game was impossible to distinguish from an actual NBA game, visually or in complexity, the viewer would still presumably know that they were not watching human beings, that they were watching a computer playing out scenarios against itself. Even with ELeague events, humans are still impacting and interacting with events, the digital world viewers are watching is just the medium for a human interaction.

These writers I spoke to may eventually be proven wrong. New technology has a way of seeming outlandishly superfluous when it’s introduced and slowly migrating towards indelibly essential. But if they’re right, it raises important questions about what it that we love so much about basketball. If a game can be digitally replicated, what would be missing that would keep you from watching? Obviously all the soup-throwing, Instagram trolling, bench-dancing, secret-tunnel exploring adds new dimension to what happens on the court.

Next: A Unified Theory of Basketball

So, the answer is emotion?

These algorithms may model strategy and the products of athletic performance incredibly accurately, but player emotions are something else entirely. Is DeAndre Jordan’s destruction of Brandon Knight as memorable without his post-dunk face, or Caron Butler walking away scratching his head? Is the Warriors-Cavaliers rivalry as compelling without Draymond losing his cool and kicking his feet? Would our fascination with LeBron and Kyrie have drifted away without moments like them battling to get the last head slap? Most fans will readily tell you that what they love about basketball is lot more than just basketball, but it could be that the scales are more tipped than we know.

In the end, the organic humanness of basketball culture, built by those in and outside the game,  may be what provides the most value added, and I guess we’re not quite ready to try and hand that over to simulations just yet.

A Unified Theory of Basketball is an irregular column exploring how aesthetics, technology, economics, strategy, narrative, and myth shape basketball.