Hi all, "Soccermetrics" has been an area of interest for a long time, so I'm happy to see there are others out there with a similar (slightly bent ) mind set. I've asked several people working outside our sport for their thoughts on soccer analysis. Perhaps nothing new, but I share their perspectives for your feedback (I thought the comment on neural networks was particularly interesting): 1. Baseball Analyst: "I think that continuous games such as soccer and basketball and hockey are much harder to analyze on a statistical basis than, for example, baseball. In baseball, the action generally proceeds in discrete steps, from one well-defined stage to another with stops between. One can define "states" and list all the possible transitions from one to another based on some primary activity--in baseball, the actions of the current batter. When there are not such identifiable discrete states, or a clear primary force in moving from one to another, matters are sticky. The essence would be to try to define the phenomena resulting in a scoring act, where those phenomena are, at least in principle, capable of being recorded as definite historical numbers. I don't know soccer, but if there are such data as ball possession, passes, and things of that sort, one might try to assemble some relation between them. If there are such data but no clear relation, one line of attack that is, so far as I know, still not much employed in sports analysis is what is called a "neural network", a complex kind of software that may be available in a package, like a word processor or spread sheet." 2. Basketball Analyst: "There's basically a few areas I think could be focused on -- one is the sense of touch/possession (how often a player touches the ball and what he does with it, how the other defense reacts to him -- closes instantly, double teams, etc), another is defending (one on one marking...spacing etc), and then more loose ball scenarios (winning a header, getting to a stray ball first)... With basketball there's such a gap currently in terms of a player's influence over a team's results, but in soccer since the major players play the whole game and the scoring is infrequent, it doesn't make as much sense to try and track the on/off stuff. Indeed the influence of one soccer player over a game must be less than the influence of a basketball star. Fundamentally though for soccer I would think that it comes down to what a guy does with his touches and how well he defends one on one. If you track these things you would have some interesting numbers like how often a guy's pass is intercepted, how often a guy's man can get by him on the dribble..."
Interesting thoughts--I think gaining the insight of basketball (and hockey) analysts would be quite helpful to our efforts. One thing I'm not too clear on is how a neural network would help. Did your friend have any more detail than this?
Before we talk about a collective consciouness I'd just like to welcome Siroco to the boards. Some good points, many of which we've discussed before but certainly warrent further thought. Welcome aboard.
Way beyond my expertise, but I assume such software would help identify patterns -- based on specific inputs (i.e. events/sequences potentially leading to scoring opportunities). From the software maker site www.calsci.com: "Neural networks are good at pattern recognition, generalization, and trend prediction. They are fast, tolerant of imperfect data, and do not need formulas or rules. Neural networks are trained by repeatedly presenting examples to the network. Each example includes both inputs (information you would use to make a decision) and outputs (the resulting decision, prediction, or response)...Your network tries to learn each of your examples in turn, calculating its output based on the inputs you provided..." Thought someone had some knowledge of this type of software or analysis.
Welcome, and thanks for passing those comments along ... from what I've seen, I think they do a pretty good job of reflecting what's been useful in each of these two sports. So far in soccer, we've seen some interesting patterns emerge from the limited data we have available, and I know that I've been surprised that some pieces of data turned out to be useful, while others did not. But really, the sticking point is that we don't have any systematic in-game tracking going on, beyond what MLSnet gives us. (For lack of further info, we should probably start archiving the match trackers.) Anyway, I'm afraid that until we get ahold of more empirical information, we can't get too far in investigating many of these ideas.
FWIW, I expect that ur_land knows a bit about neural nets ... the question is what kind of patterns we'd use them to look for, and whether they'd really be the best tool for the kind of tasks we're doing. Can't say I'm a big fan of them, myself.
I'm sure I don't have the experience with neural networks (or anything) that some people here do, but I've got a little. I think, way back when, I considered using them to evaluate lineups - which combinations had worked best with each other. However, I'm not at all confident that would work well, and there's not a whole lot of other stuff I can imagine that they would be useful for. Maybe it's just a lack of imagination on my part, I don't know. However, even if there are some uses, a central problem with neural networks is that, while they may be able to solve problems, it's very hard to interpret how they solved the problems. And, while the solutions are obviously interesting, I think we're far more concerned with the how here.
I remember reading some IEEE literature on using neural networks for soccer playing robots and software soccer simulation. It's an active field for artificial intelligence research because you have a set of simple rules that results in complex behavoir. As for it's application for soccer analysis: neural networks are typically used for pattern recognition and for systems that can learn or be trained. For example, credit card companies train neural networks to spot patterns of activity that are usually associated with stolen credit cards. Or Dept. of Homeland Security trains neural networks with profiles of real terrorists, then sees if the neural networks can spot terrorists on airline passenger lists. My question is what would be used to train a neural network for soccer analysis?
Maybe patterns of passes or possession or ball movement that lead to goals? Numerista is right--I do have a little familiarity with neural nets (at least as they are used for modeling human cognition and consciousness)--I lived for four years with a grad student doing neural networks examining vision and selective attention, and have taken a few classes in their application for neuroscience and psych. I was just brain dead when I posted, and couldn't think of applications for soccer. As we already talked about in this thread, I think the big problem is getting the data. Having both the uinput patterns and the results to train up the model would be a problem. Some of us need to start a company (like the elias sports bureau) for soccer, just so we can get this done (or get hired by an existing company.......any headhunters out there feel free to send me a PM!)
Matchanalyisis actrually did much of the same stuff that you are talking about Ur_land, led to soem pretty neat graphical analysis but I'm a little unsure as to exactly how practical it was. Of course with seemingly everything now they've revamped their site and it might be hard to get at them again but those guys definately had the same inclinations. Here's what they called it, Primary passing channels. PS. Thank you archive.org
Yeah--what we would want would be to take that data and then input it into a neural network. For example, if we had passing data, similar to what was on that webpage, we could use a pair of positions as an input (RB to RW), and then the output could be the pass completion % or a level of activation corresponding to number of completed passes. Then once it's trained up (on lotsof pairs or pass completions ), we could then give it novel inputs, (RB to F) and it would predict how good of a passign connection there would be between those positions. Not too useful in that example, but that's the idea. Actually, perhaps something better would be to get an idea of how well a new player would fit in. Let's say (for example) Il Bruce wants to try out Taylor Twellman and Eddie Johnson, up top, Ralston on the wing, and Bobby Convey in the middle. He can input into the neural network all of the passing completion % for Ralston and Twellman for Revs games, Ralston and Convey for nats games, and Convey and EJ for U-20 games. He can then use this data to predict how well Ralston will link with EJ, and how well Convey will link with Twellman. So inthis way, neural nets might be more useful for prediction than analysis.