Started By
Message
AI in football play calling
Posted on 8/21/24 at 9:10 pm
Posted on 8/21/24 at 9:10 pm
This seems to be an obvious future step. Wondering if anyone is doing any of it?
I think it would be a hugely expensive undertaking, even just collecting data points like “In their 3rd game Temple ran play X with personnel Y when the ball was on their 40 on 3rd and 4. Result was an incompletion to player 20.”
But with the money in football, someone is at least going to investigate its use.
I think it would be a hugely expensive undertaking, even just collecting data points like “In their 3rd game Temple ran play X with personnel Y when the ball was on their 40 on 3rd and 4. Result was an incompletion to player 20.”
But with the money in football, someone is at least going to investigate its use.
5
Posted on 8/21/24 at 9:17 pm to ChapelHillSooner
what could go wrong…..?
Posted on 8/22/24 at 8:11 am to ChapelHillSooner
quote:
AI in football play calling
Not sure if this was posted here originally or moved here... anyhoo...
Would be a fascinating exercise, but in my experience, AI isn't "great" yet, or at least consistent yet, at technical expertise. (Its primary niche right now is as a language model.)
And I think we've sort of been doing it for years with the predictive engines in video games. I even recall a rudimentary text-only football game that worked very similar to your description on the very earliest personal computers.
But to be fair, the moneyball-ization of sport is certainly real, and football coaches already do use complex metrics like you're describing to analyze and plan their concepts, so you're definitely on to something.
Posted on 8/23/24 at 8:25 am to ChapelHillSooner
quote:
I think it would be a hugely expensive undertaking, even just collecting data points like “In their 3rd game Temple ran play X with personnel Y when the ball was on their 40 on 3rd and 4. Result was an incompletion to player 20.”
This is done now without AI
Posted on 9/4/24 at 2:31 pm to ChapelHillSooner
I've looked at this thread and thought about replying multiple times.
I develop AI as a hobby, but AI is a broad term. I think most of the time people think it as only neural networks, and you look at things like Chat GPT and think it can do crazy things. But really that's just 1 small part of AI. It's not all neural networks. In many cases, neural networks aren't even the best solution(due to the resources required etc).
The biggest thing is they need a lot of data to train. You can't just feed AI a problem and expect an answer, it doesn't work that way. You have to train the AI by feeding it the questions and the answers. It then basically changes weights around in the network so that when the question hits the network it prints out the correct answer. And you need quite a bit of data to do this, and you need a variety so that it's able to generalize the data. So take something that recognizes cats and dogs, you start with a database of images that have cats and dogs labeled.
Chat GPT is able to have such success because it has tons of data available. It's basically an autocomplete machine, where it's always trying to guess the next word based on the previous words. It's obviously able to do this extremely well, even understanding context. This is what is possible with lots of computing power and tons of data.
So in order to do that for college football, you'd need to have a ton of data. You can't just feed it film or whatever, you have to also include the data for it to know what it's looking at. It's not going to be able to call a play if it doesn't know what a play is. So every play needs to have a value for that. And that means lots of human resources.
That's all supervised learning, unsupervised learning is based on rewards. But that's even more difficult. It would also require training in some kind of virtual environment because it needs to call games and see what happens, many games, hundreds and maybe thousands of years worth of games.
But like I said previously, neural networks are only a small part of AI. When I build AI, I don't use neural networks 99% of the time. I instead find logical ways to recognize a situation based on the variables and then find the appropriate reaction. So I don't need to feed tons of data to the AI for it to do what I want. I look at the current state and apply heuristics or whatever. I like to use a point type system often that is similar to a NN in that it adds weights in certain directions, but it's not a NN. A properly trained NN can probably do it better, but getting to such a point often requires extreme resources.
And then you might use neural networks for smaller specific jobs. The smaller the task, the less data you need etc.
That's basically how FSD has evolved with Tesla. It's not 100% neural network. It started out with the NN handling smaller jobs and the rest being coded by humans. Over time, they've been letting NN do more and more and it's improved as they've done that. It's like 1 step at a time.
And I think that's how such a thing would evolve for football. You can kind of look at people using general stats already. Maybe someone makes a program where you're able to put in the basic information and it spits out some situational stats etc. And maybe over time NN takes over parts of it. I'd probably start with things like identifying plays.
Things like that probably already exists. And then it's a matter of how good it will be. You'll still want a human looking over it for awhile, just like you still need a human supervisor for FSD. Because something weird may happen and the damn thing might punt on 1st and goal from the opponents 2 yard line and you need a human that says NOPE.
In the end, I think you'll have a program that can call plays better than the average person, but you'll still be looking at humans at the highest levels and for innovation for a long time.
I develop AI as a hobby, but AI is a broad term. I think most of the time people think it as only neural networks, and you look at things like Chat GPT and think it can do crazy things. But really that's just 1 small part of AI. It's not all neural networks. In many cases, neural networks aren't even the best solution(due to the resources required etc).
The biggest thing is they need a lot of data to train. You can't just feed AI a problem and expect an answer, it doesn't work that way. You have to train the AI by feeding it the questions and the answers. It then basically changes weights around in the network so that when the question hits the network it prints out the correct answer. And you need quite a bit of data to do this, and you need a variety so that it's able to generalize the data. So take something that recognizes cats and dogs, you start with a database of images that have cats and dogs labeled.
Chat GPT is able to have such success because it has tons of data available. It's basically an autocomplete machine, where it's always trying to guess the next word based on the previous words. It's obviously able to do this extremely well, even understanding context. This is what is possible with lots of computing power and tons of data.
So in order to do that for college football, you'd need to have a ton of data. You can't just feed it film or whatever, you have to also include the data for it to know what it's looking at. It's not going to be able to call a play if it doesn't know what a play is. So every play needs to have a value for that. And that means lots of human resources.
That's all supervised learning, unsupervised learning is based on rewards. But that's even more difficult. It would also require training in some kind of virtual environment because it needs to call games and see what happens, many games, hundreds and maybe thousands of years worth of games.
But like I said previously, neural networks are only a small part of AI. When I build AI, I don't use neural networks 99% of the time. I instead find logical ways to recognize a situation based on the variables and then find the appropriate reaction. So I don't need to feed tons of data to the AI for it to do what I want. I look at the current state and apply heuristics or whatever. I like to use a point type system often that is similar to a NN in that it adds weights in certain directions, but it's not a NN. A properly trained NN can probably do it better, but getting to such a point often requires extreme resources.
And then you might use neural networks for smaller specific jobs. The smaller the task, the less data you need etc.
That's basically how FSD has evolved with Tesla. It's not 100% neural network. It started out with the NN handling smaller jobs and the rest being coded by humans. Over time, they've been letting NN do more and more and it's improved as they've done that. It's like 1 step at a time.
And I think that's how such a thing would evolve for football. You can kind of look at people using general stats already. Maybe someone makes a program where you're able to put in the basic information and it spits out some situational stats etc. And maybe over time NN takes over parts of it. I'd probably start with things like identifying plays.
Things like that probably already exists. And then it's a matter of how good it will be. You'll still want a human looking over it for awhile, just like you still need a human supervisor for FSD. Because something weird may happen and the damn thing might punt on 1st and goal from the opponents 2 yard line and you need a human that says NOPE.
In the end, I think you'll have a program that can call plays better than the average person, but you'll still be looking at humans at the highest levels and for innovation for a long time.
This post was edited on 9/4/24 at 2:38 pm
Posted on 9/4/24 at 2:34 pm to ChapelHillSooner
The more you let robots do the more power they get. Stop letting them level up.
Posted on 9/4/24 at 8:10 pm to Lexag
quote:
The more you let robots do the more power they get. Stop letting them level up.
They are tools. Like any other tool, the problem is not the tool, but rather what some people might do with them.
What makes them scary(to me anyway) is they do not have a conscious and just do whatever they are trained to do.
Posted on 9/4/24 at 8:24 pm to 3down10
quote:
just do whatever they are trained to do.
the problem is AI is trained to train itself. Algorithms. Emotionally absent, problem solving.
Page 1 of 1
Popular
Back to top
Follow SECRant for SEC Football News