Advancing Sports Analytics Through AI Research
deepmind.com
deepmind.com
There are still areas where analytics will continue to be important in sports and player assessment/team composition will always heavily rely on it. However the next steps for sports analytics in my mind are things that go on behind the scenes of the sport: personal training routines for players, managing wear and tear and predicting injuries, etc.
Also, the leagues themselves have lots of room to improve their products with analytics. Many complaints have been raised (which I agree with) about how boring modern basketball/baseball has become due to analytics causing teams to converge on strategies that maximize efficiency but are also quite boring. It isn't on the players/coaches to try something more exciting at the expense of winning. It's the league's job to make the product interesting. Analytics can be part of the solution, like determining that by pushing the 3-point line out several feet, the value of the shot is brought in line with other options that simply aren't viable right now.
Obviously the games are different, but perhaps there are some common themes or generalized structures that can help frame what questions to ask?
So, from what little I know of LoL, you’re simultaneously trying to attack the enemy base and protect your own. Presumably there are interesting second order things to study here - how can you get close to the base, how can you destroy it quicker than the opponent destroys yours etc. I assume if there are any epiphanies along the lines of “shoot threes” in basketball or “shoot closer to goal” in football, they’re probably well trodden.
Much of the game is presumably controlling the space between, trying to create overloads against opposition players, and not be overloaded yourself. The best teams in football are experts at using space, and at manipulating their opponent’s shape to create gaps or opportunities for ambushes. So an interesting thing to study would be teams tendencies to overextend - what sort of structures appear to offer them opportunities, but actually keep your players close enough together to pounce on them. The ghosting work mentioned in the article would be interesting here, as would work on pitch control:
https://www.researchgate.net/profile/William-Spearman/public...
I'd also imagine any data you can crunch from solo queue games is of limited use in competitive play.
The easily digestible stats like KDA or gold/min are relatively useless in this regard which is what I meant by them being low hanging fruit.
So I’m currently in a situation where I’m curious on how to best quantify or develop a series of metrics that can prepare teams for their opponents. The current system of watching vods, taking notes, and making slides is rather dull if I’m being honest...and really not paradigm shifting. If you’re interested I can show you videos/presentations/figures.
> Obviously the games are different, but perhaps there are some common themes or generalized structures that can help frame what questions to ask?
One of the most important things to understand is what the front office, coaches, and players want to know. Like I said, much of analytics these days (and not just in sports) is a solution searching for a problem. We build models based on what we can with the data, but the output of the models isn't actionable or something anyone is asking for. A better approach is to start from the questions that loom big in the eyes of those you are doing the analytics for. That sounds simple, but often doesn't happen, either because communication is poor, or neither party really understands what the other is trying to do. Also, often times a team doesn't even really know what questions they'd like answered, or at least not enough to verbalize it well.
One thing I've found consistently useful is focusing on measuring uncertainty. These days people are extremely excited about machine learning and "AI", and I work extensively with deep learning in my current position. But most times in analytics, especially sports analytics, simply getting the best prediction possible isn't actually useful. What is far more useful is measures of uncertainty through more traditional statistical methods, especially Bayesian statistics and Monte Carlo simulation. It's one thing to predict a player in the draft will be better than another. But a team usually knows the answer to that already anyways. Its far more useful to quantify the range of possible outcomes for a player. Is it practically a sure thing? Is it not a sure thing but a reasonable bet? Or is it really a toss-up? These are insights that teams often don't have but can find much more useful in their decision-making.
I generally don't focus on in-game strategy, since it is harder to give the team something they don't already know that they can reasonably act on. But meta-game insights could have a lot of potential. For traditional sports, this can be things like assessing player value, injury risk, draft capital, and contract valuation. Some of these may not translate well to LoL. I haven't played the game, but I've watched it a few times, and perhaps something could be done looking for insights regarding the hero pool. There are so many different heroes in the game that can act/oppose each other that quantifying those relationships somehow could be interesting.
I'd be super interested in that and seems like the more data we can collect from watches etc the more potential variables to model or train ML suggestions on?
lots of companies spending big money on this market too. Apple, pton, smart watches.
The documentary Breaking2 from nike about the marathon is really good, shows some advanced individual measurements that wouldn't be available to almost everyone (lactic acid blood draws while running..). in my sport climbing finding new ways to push lactic acid tolerance or better improving physiology through BFR or whatever training would be really valuable and maybe push the sport even further.
But basketball seems the opposite to me. Analytics has shown that except in a few superstar cases, back to the basket isolation mid range jumpers while the rest of the team stands around and does nothing (IMO the most boring play in basektball), is a terrible play call.
Analytics has shown that the pick and roll is a fantastic play in terms of points expectancy and I find it to be a really enjoyable play to watch because it often results in dunks and nifty passes or acrobatic layup attempts.
Analytics has shown that faster pace, shooting earlier in the shot clock, and getting out in transition more often to be hugely valuable strategies, all strategies that I think are more enjoyable to watch.
The perceived value of 3-pointer shooting has caused offenses to spread the floor creating more driving lines and making it harder for defenses to pack the paint which has created more offense which I think most people enjoy seeing.
I'd love to hear what you think analytics has done to make modern basketball more boring.
Analytics only uncovered these structural issues for teams to exploit. Personally, I think if the strikeout rate were tamed back to 1990s levels, the game would balance.
The optimal strategy for both sides is for pitchers to go for the K (increases K rate) and for hitters to try and hit the homerun (also increases the K rate).
Basketball is time-limited, and therefore the strategy of maximizing the number of shots within that time period as well as maximizing their value, make sense.
I'll note that I don't necessarily think one is better than the other, as each game can have a different approach. I also think that maximizing purely for excitement (from a rules perspective), can lead to gimmicks rather than genuine improvements to the sport.
Ultimately I guess I don't mind if some teams play like today's style of basketball, but I do mind if its the only viable strategy.
Arsenal bought an analytics company (more than one iirc), and did absolutely nothing with it. Man City are only just getting into the area. Liverpool have been doing it for a while (with fairly mixed success). There is almost no interest from management in applying analytics.
Additionally, sports are not equal. Baseball suits analytics. NBA is harder. And football is harder than the NBA (by a significant degree). All of this is measurable btw. One very simple point relating to this: there is no equilibrium strategy in football, it varies based on the opposition, teams that are bottom of the league can (under certain conditions) prevent a huge obstacle to the best team in the league.
Either way, it doesn't matter because almost no football teams are applying this knowledge (there are a few exceptions, some of the top teams are run by professional gamblers and they have been printing money from analytics for decades...so, in those cases, it is being applied).
The vast majority of teams that try to "buy the best players" don't succeed (usually there is only one or two players currently in the game who can win games by themselves, there is a substantially larger group of players who cost essentially the same and don't perform at all) because they pay too much/buy players randomly who don't fit their system/overestimate skill.
I would look more closely at what I said. I did not say that analytics hasn't been around for a long time (for some reason, the article you link fails to mention that Reep actually worked for clubs, and found some degree of success...but there is no Nash strategy, Bolton got to the Europa League playing long ball football...it is a fine strategy suggesting that it doesn't work on average means nothing because everything is conditional in football).
Btw, your model of salary cap=pay most for players is obviously flawed if you consider that the non-existence of a salary cap does not occur in a vacuum. Player salaries/transfer fees are a dynamic competition so the lack of salary cap means that most players are overpriced (because clubs are inherently overoptimistic...there are clubs who make money just by developing players and selling them to bigger clubs). The actual implication of salary caps (in combination with FFP) is that success is correlated to your ability to generate revenue. It isn't that the top clubs can pay more for players, it is that they can "lose" more overpaying than anyone else. As an example, Manchester United have rock-solid sponsorship revenue (iirc, they even have a "tractor partner"...they sponsors for literally everything) so they can overpay for almost every player, lose tons of money doing so, and end up doing okay. If a club lower down the table made deals as bad, they would get relegated. No salary caps just mean the rich clubs get richer.
No equilibrium strategy has been explained twice. If you don't know, I can't help you.
They aren't "farm teams". That only makes sense in the context of American sports (once again). Analytics is an emphasis at many of these clubs, it is how they win games (again, I am not sure if you understand that I am not talking about American sports...win probability is important because if you don't win games, you get relegated, and your club can stop existing in a few years...you need to win games).
All of what has been mentioned has been conceived of, thought about and discussed before, and some years ago. I've seen more advanced thinking in some models used to make bets on football matches.
As a Man City fan, I'll try and make this sound less snarky than it will come across, but it's going to be snarky: it's good to see Liverpool are finally - with Deep Mind's help - catching up to where many other teams in the league were more than half a decade ago.
Most players and coaches are not particularly interested in these models and the most useful applications seem to be related to scouting and the transfer market - "I need a replacement for player X, but 10 years younger and with a better left foot", is the sort of problem that data science is already helping with - and the clubs themselves prefer to put their money onto the field of play rather than expensive and experimental ML training, so the market for this sort of work is limited.
The challenges in this field are not really about data, algorithms or computational capability of the hardware.
The main issues are of applicability, cultural fit and economics.
Finding methods for addressing _those_ issues might be more useful to mainstream AI adoption than the data, algorithm and hardware problems many seem to look at first.
Your point about applicability is totally right though, see Arsenal with StatDNA and the arguably weak impact of Barcelona’s Innovation Hub. I’d argue that Leicester’s use of (at the time) very rudimentary stats, nevertheless well integrated into their process, was a big part of their sudden success.
[0]: https://www.liebertpub.com/doi/pdf/10.1089/big.2014.0018
Love that quote.
I agree that those sports have suffered in terms of watch-ability at the hands of maximizing competitive efficiency. But I don't really know why anyone would expect anything different. The point of the game from a team/player perspective is to win. Using analytics to determine that there are advantages to three-point shots or walks isn't any different than any other strategic layer of the game. The solution shouldn't be to rail against analytics (which is no solution at all), it should be for the leagues to use analytics of their own to control for some of these strategies, like a government is supposed to hedge against externalities in an market. For basketball, simply push the 3-point line back a few feet so its no longer so much more advantageous than a midrange jump shot. I'm not much of a baseball fan, but surely the league can do something, like increases the number of balls to walk to 5, that might bring some of the boring-but-efficient strategies into line with more exciting ones.
Though I suspect the author and many others would be against any substantial rule alterations given their abhorrence towards change, and would rather just pine about the good old days.
My personal interest is for soccer to allow more substitutions to let coaches tweak the team more as the game is played. Small change; maybe big results in scoring.
Usage of technology has ruined several aspects of traditional association football. The VAR for example, is ruining the serendipitous nature of football. Fans in general dislike the VAR because it removes this extra element of drama that the sport has.
I’m supportive of technology in sports though. I just think there should be a limit, otherwise they become a very boring and cold activity.
I’m not necessarily against the suggested approach here. Having analytics of how a game is going or went seems useful and a net positive. It’s just hard to tell if that could keep eroding those entertainment elements that make Football fun to watch.
Of course, but the reality is that analytics have proven that there are only really two shots worth taking in modern basketball: the 3 pointer and a shot near the rim. For anything in between the expected points for the shot is simply not worth it. Even an uncontested midrange shot is going to have similar or fewer expected points per shot than a contested 3-pointer or shot beneath the rim. And these two kinds of shots are spatially far from each other, so even if the opposing team knows this is your strategy, its still difficult to defend against. Teams and players that have tried to exploit the fact that modern basketball is fairly homogenous by taking relatively less defended midrange shots occasionally find streaks of success, but it hasn't proven sustainable, and certainly not enough to win a championship.
Basketball would probably benefit from either eliminating the 3 point shot, or, perhaps preferably, pushing the line back a few feet to even it out in terms of point expectations with midrange shots. Shots under the rim will still remain the most reliable point source, but teams won't have to divide as much attention on getting steam-rolled by three pointers. It will bring back scoring big men like a prime Shaq who have mostly disappeared from the game, still allow phenoms like Steph Curry to thrive (fewer players can match his ability to hit 3-pointers from further distance), and will make the midrange focused strategies more viable again.
The strategic trends always go in and out of fashion.
One thing that I see in Basketball that needs fixing though, is that the rules make the defender's job incredibly hard. Especially with the rules around fouls. They need to look into fixing that. I think a solution to that class of issues, will inadvertently fix the 3 pointer issues you cite too.
Yes, but by "inside the perimeter", you mean up-close to the basket. Zion is a great example. Take a look at his shot chart [0]. He doesn't even think about shooting more than a few feet away from the rim.
As I said, at this point, basketball has coalesced around offenses entirely focusing on shots at the rim, or beyond the 3-point line. Anything between there is a wasteland where teams don't even bother shooting anymore, despite Hall of Fame players like Michael Jordan and Karl Malone making those areas of the court a huge part of their game in the 90s. And this doesn't seem like much of a fad. Its a trend that has been becoming increasingly entrenched for nearly 2 decades at this point. The charts on this 538 article are all you need to see why this isn't simply a matter of what's fashionable, and is a simple mathematical reality under the current rules [1].
But I definitely agree with your last point. Basketball has numerous issues currently, of which unequal shot efficiency is only a part.
[0] https://www.statmuse.com/nba/ask/zion-williamson-shot-chart
[1] https://fivethirtyeight.com/features/how-mapping-shots-in-th...
the playoffs are all about effective shot selection, with the midrange becoming prominent because defenses tighten up arond the perimeter and the basket. lebron and anthony davis both explicitly featured their midrange game in last year's championship run. the clippers, incidentally, have the best midrange shooter (kawhi) and lost early last year because they relied too much on the 3-pointer. even the warriors of old used the midrange shot to keep defenses guessing enough to give their shooters space to shoot.
and that's exactly the issue with relying too heavily on analytics--humans aren't static, and pinning down the exact parameters where a model is valid is seductively elusive. mathematical models don't so much tell us about how the world is, but rather more about how the modeler sees the world. models aren't useless but they are limited (really knowing how is the hard part).
[1] - https://nbasavant.com/player.php?ddlYear=&ddlShotMade=&ddlTe...
In contrast, fans have been fine with the goal-line technology, which gives decisive (usually) and quick results.
The use of AI in this paper seems more relevant for setting team strategies and picking players rather than enforcing rulings.
For example, in basketball, it should be able to do goal-tending in near real time. Also 3s in the key.
The holy grail though is if it can help charge/block calls. Or we should consider changing the rule so that it can be consistently called, but that's a topic for a basketball forum.