While impressive, volleyball is one of the most formulaic sports when it comes to predicting what happens next. This is not a dig on volleyball (I'm a fan!), but it is how the sport works.
The first touch is always about receiving the ball and passing it cleanly to the setter. The second touch is about setting up the attacker (if possible). The third touch is a spike or other attempt to get the ball to the other side. The only significant decision points are usually who the setter chooses for the attack (among 2-4 players depending on formation), and on the defense side, where the blockers jump (3 players choosing spots to cover). Since the algorithm only predicts "actions", both of those don't matter (the setter is going to be "setting" regardless of who they choose, the blockers are always going to be "blocking" regardless of the spot they cover).
The "team strategy" part is in effect just a check of whether the algorithm knows what is happening on the court at all. Volleyball allows the aforementioned three touches, and what the team is doing for each touch is pretty much immutable (the only source of variation is if one of the "steps" fails).
This is of course still an achievement, but the choice of sport gives the researchers a leg up, since volleyball has significantly fewer "degrees of freedom" than other sports. In that context, ~80% accuracy at predicting an action 2 seconds out is less interesting, and it's unlikely that it's anywhere close to human performance for this particular sport.
Thinking the same thing! Volleyball is very structured, there are schemas and tactics basically for every event in the game. Additionally, at professional level the accuracy of the players is quite high so the variability should get lower e.g. the setter will set the ball very precisely and the spiker will attack from a specific position. What it might change is the timing (depending on how high the ball is), but usually the professional games are very fast so that should also simplify things.
Title is misleading. This is about classifying volleyball players and predicting their actions up to 2 seconds in advance. Which is really impressive. Humans do this automatically without even thinking about.
General artificial intelligence is a still long way away, but the slow progress and exploration is fascinating to watch.
I don't know volleyball well, but isn't it fairly deterministic within the window of 2 seconds? Like... your teammate just set the ball. What are you going to do? Isn't ALWAYS spike the ball?
The ball is moving fast enough that for at least some moves the previous person hasn't yet touched the ball 2 seconds before, you cannot block a spike if you are not in the air before the spiker touches the ball. The spiker of course will be watching you and change how the ball is spiked based on that (if you don't make any attempt to block at all the spiker will drive the ball hard into the ground, if you do it is either find the hole in your defense to spike through, or tip the ball over the blocker making the return a lot easier)
I only played at a college class level (I wasn't the best in my class, not a pro for sure) - yet I can do the above.
But then in this case the spiker is still spiking, the blocker is still jumping, no? There is no particular room for alternative states for competent players?
I don't know if this is progress towards AGI. All life and all intelligence is just taking input and matching it against previous inputs and/or some hard-wiring as a way to make a decision. So in that sense you could say that any improvement in pattern recognition is (slow) progress toward AGI.
I don't know if that is useful though, because I don't know if there is some progression from this to that.
In other words, using such broad criteria means you could say that improvement to any pattern recognition or any algorithm that takes an input and decides something to create an output is progress toward AGI, but it's clearly not.
So I'm not really sure if all this is any more than a souped-up "tree recognition neural network on a 16x16 pixel grid" that was the staple of "AI" neural net homework back in the 1980s, or even the old "expert system".
The paper doesn't mention any baseline, and so claims like "86% accuracy" are a bit useless: what if players do the same action 85% of the time? Then the model wouldn't be not predicting much. If on the other hand the most common action happens only 30% of the time, then that would be an incredibly strong result.
Bump-set-spike. That is how all great players play. There are exceptions for things like they messed up; or are exploiting surprise to do something wrong hoping it works out.
At amateur level it is much harder to predict as nobody follow thee proper order, and in turn this makes it really easy for someone just a little better than average win.
I'd like to see an attempt to predict the movement of an NFL defense based on the starting positions of both the offense and defense. Seems like a much harder problem than volleyball, yet is something that human experts can clearly do on the fly.
It’d certainly be interesting to see if 80% accuracy over a 2 second window was enough to exploit tactically. If so, you’ve probably just saved a few hours of a video analyst’s time. Interesting approach though. There’s been a lot of work on this over the years, including on (dare I say) more complex sports like soccer. This was from Sloan 5 years ago for example:
I'm pretty sure any skilled player can reliable judge what opposing players are going to in the next few seconds. Probably with closer to 95% accuracy.
Especially for volleyball, because the only open question are: “to which attacker the setter will pass”, “where's the attacker going to attack” and “will the defense be able to get the ball”.
For the “nine well-defined classes: spiking, blocking, set-
ting, running, digging, standing, falling, waiting, and jumping” used in the paper, the only things that are hard to predict is “jumping”, for the attackers, because it depends on who the setter will pass the ball to, and who's going to be “falling” because it depends where the attacker will shoot the ball and which defender is in range to catch it.
I've just tried to do it on a Youtube video, and it's actually a pretty fun game to play for a minute or two. (I think I'm above 85%, despite not having practiced volley outside of school's sport classes and never having watched it on TV, so I'm pretty convinced your 95% figure is well within reach of a skilled player).
The interesting thing is whether you can identify exploitable tendencies in opposition team play (again, faster than your human analysts would otherwise do). If so, that takes preparation with the whole team rather than split second decision making.
It’s worth noting this article doesn’t predict outcomes, but player roles.
It’s not “Team A will win” but “Player B will spike”. And even then, the prediction is only a couple seconds before the event (so you can’t, say, feed it an initial condition and make a bet that a player will get at least X spikes by game’s end).
Is there a betting pool for player roles, particularly in a fast paced game like volleyball?
The abstract talks about predicting 48 frames ahead. Is realtime sports betting even a thing? And if so is the liquidity and spread good enough to enter/exit say a $10,000 position that quickly?
> And if so is the liquidity and spread good enough to enter/exit say a $10,000 position that quickly?
In some sports, sure. Probably not in volleyball.
The threat in the field is that some people have a slightly faster access to info than others (3-5 secs for live versus broadcast, iirc).
If the result of an algo can push a betting line (e.g., predicting a tying point in a volleyball match), then automated bet placement software/script can make a prediction and get a bet in before the action is seen by bettors who are watching on broadcast.
Even without prediction software, this is a very real problem with real time sports betting right now. It seems like functional prediction software can make this knowledge gap larger.
Not exactly. So long as the house overall comes out ahead they love a consistent winner: people who lady luck smiles on tend to brag about how good they are not only bring in their friends, but also coming back to lose it all next time.
They can do this in a number of ways, some of them are because the game is setup so the house is the winner, others the house takes a fee off the top.
For the first we have games like blackjack where the house has a consistent winning strategy. They ban counting cards because this is a way someone can be a consistent winner instead of them. While there may be some skill in the game, no amount of skill will make you are winner, only luck.
For the second we have things like a sports betting where the they take a small commission but otherwise are not even playing the game. In these games they love consistent winners as such people tend to brag and bring in more people thinking they can do the same. If you are a great poker player it is worth paying the commission as the house provides security that the person you are betting against isn't cheating (marked deck, won't actually pay his bets...)
Betfair works like a market - you are not betting against the house, but rather against other bettors. Betfair takes a commission from each matched bet. They would probably allow it - More matched bets means more profit for them.
I'm not a lawyer, but out of curiosity I did a quick search through their T&Cs.
It seems to be a bit of a grey area.
Section 3 ("APIs") sort of allows it in a half-assed wording kind of way:
- *"Other than your permitted use of Betfair's endorsed API service (referred to above), the use of programmes or software designed to automatically place bets within certain parameters (i.e., "robot" players) is not permitted on any Games"*
Section 18 ("Bots") certainly seems to allow it:
- *"18.10 Some customers make use of programs designed to automatically place bets within certain parameters set by them (e.g., to back or lay at a certain price) ("bots"). These bots may be active in any or all Markets at any time and you should not assume that you can place bets on quiet Markets that will not be automatically matched simply because the Market otherwise appears quiet. Additionally, bot users should be aware that bots might be prone to exploitation by other customers. Users of these programs do so entirely at their own risk."*
Meanwhile, section 12 ("suspicious betting") seems to be very much against the idea:
- 12.1 For the purpose of this clause 12, "suspicious betting" refers to where we have reasonable grounds to believe that a Bet or a number of Bet have been placed in suspicious circumstances. Suspicious betting shall include, but not be limited to:
- 12.1.4 where we have reasonable grounds to suspect that a Bet or a string of connected Bets were placed robotically, by automated means, or otherwise than through the Account holder placing each Bet manually via their Account;
- 12.1.5 where we reasonably believe that you have used unfair external factors or influences connected with the event(s) the subject of any Bet(s);
- 12.2 If we discover or have reason to believe that you have participated in any of the activities described in clause 12.1 (which, as stated in clause 11.2, constitute "Prohibited Activities"), we shall have the rights set out in clause 13.
"Rights as set out in clause 13" being "Suspension or Termination of Your Account by Us".
That’s interesting. Do they publish the history of what it costs to place a bet?
There’s quite a few variables that go into the pricing: how much you think a team will win by and the payout if you win the bet. So there can be multiple tracks for “I think my team will win by 7 points” and there might be someone willing to pay 2 to one odds on that while another is 3 to 1.
Maybe betfair sets the rate? This number has to change over time though, can purchasers of a contract sell that contract later on if the odds or payout changes dramatically?
It would also be interesting to see where the “knee” is in those bets. For example winning by at least 7 or at least 8 points in American football should have about the same odds due to the point scheme that has a touchdown plus the extra point at 7 points. Course you could also try and go for two so maybe there is a price difference. Anyway it would be interesting to examine.
The first touch is always about receiving the ball and passing it cleanly to the setter. The second touch is about setting up the attacker (if possible). The third touch is a spike or other attempt to get the ball to the other side. The only significant decision points are usually who the setter chooses for the attack (among 2-4 players depending on formation), and on the defense side, where the blockers jump (3 players choosing spots to cover). Since the algorithm only predicts "actions", both of those don't matter (the setter is going to be "setting" regardless of who they choose, the blockers are always going to be "blocking" regardless of the spot they cover).
The "team strategy" part is in effect just a check of whether the algorithm knows what is happening on the court at all. Volleyball allows the aforementioned three touches, and what the team is doing for each touch is pretty much immutable (the only source of variation is if one of the "steps" fails).
This is of course still an achievement, but the choice of sport gives the researchers a leg up, since volleyball has significantly fewer "degrees of freedom" than other sports. In that context, ~80% accuracy at predicting an action 2 seconds out is less interesting, and it's unlikely that it's anywhere close to human performance for this particular sport.