AI and Play-Calling

[Randy Carson]

Let's say that an AI system were to analyze EVERY single play from scrimmage run by some random team over a three-year period. Each play would be categorized for things such as down & distance, score, time on the clock, home vs. road game, etc.

Then during an actual game, a booth analyst would enter a few variables after each play ends to keep the AI system up to date on what the opposing team had done thus far.

Could an AI system review all the data, uncover patterns and determine that an OC tended to call one of three plays in a given situation? Then recommend the best defensive play to call?

 

[Enuratique]

Yes

 

[Blue&Gold1034]

Let's say that an AI system were to analyze EVERY single play from scrimmage run by some random team over a three-year period. Each play would be categorized for things such as down & distance, score, time on the clock, home vs. road game, etc.

Then during an actual game, a booth analyst would enter a few variables after each play ends to keep the AI system up to date on what the opposing team had done thus far.

Could an AI system review all the data, uncover patterns and determine that an OC tended to call one of three plays in a given situation? Then recommend the best defensive play to call?

Went down a quick rabbit hole concerning this and some engineers who specialize in pattern recognition have written papers on this very topic and posited software that can identify formations and personnel sets. The biggest problem will be dealing with certain variables such as specific players on the field, weather conditions, etc. I dont think AI will completely take over playcalling, but it will definitely help with the film review process for coaches leading up to game day.

 

[Enuratique]

I'm impressed with what CHATGPT can do. Given cleanly categorized data I bet it could pay dividends. And therein lies the rub.

 

[cpf2001]

You could probably make a good living playing both sides of this as a software vendor.

"Be able to tell your opponents tendencies and ask what influence other factors might have!"

"Be able to tell your own tendencies and keep the other team guessing by going off script!"

 

[slugboy]

People drastically underestimate what is required to build a system that would actually do a good job at that.

Here’s a good podcast dedicated to the topic https://twimlai.com/. Every episode is about the latest in commercial or research ML.

Signed,
A guy who did grad research in the topic and works with it today

 

[Lil G]

I think you would need AI to watch each every play in live time to be able to notice the things gametime coaches are paid to see, but even then there would be too many minor and potentially undetectable details if it were going off visuals and even audio.
Like if a safety on the previous play revealed that he is looking a certain direction too much, biting on certain route fakes or just got benched for another one that has different tendencies. Or let’s say a quarterback notices a LB thinks he is learning the signs/audible calls, and is positioning slightly different right before the snap when he hears it. Those might be rough examples but there has to be 20 other things you would be trying to tell the AI about inbetween plays that these teams are taking into consideration. Even without no-huddle and playing slow tempo, I don’t see how it would ever really be more than just complimentary ideas to what calls are being made on the field.
Just my 2¢ on game decisions.
I do think way teams prep the week before could be slightly altered if an AI was trained to watch film and pay close attention to player tendencies. But still, would be mostly complimentary study.

 

[stinger78]

People drastically underestimate what is required to build a system that would actually do a good job at that.

Here’s a good podcast dedicated to the topic https://twimlai.com/. Every episode is about the latest in commercial or research ML.

Signed,
A guy who did grad research in the topic and works with it today

From one who has done postgrad work in complex systems, they are extraordinarily difficult to predict with acceptable variance. Many say you cannot. There are too many feedback loops. Football is a complex system. Trends and tendencies is about the best one can do.

 

[MountainBuzzMan]

Let's say that an AI system were to analyze EVERY single play from scrimmage run by some random team over a three-year period. Each play would be categorized for things such as down & distance, score, time on the clock, home vs. road game, etc.

Then during an actual game, a booth analyst would enter a few variables after each play ends to keep the AI system up to date on what the opposing team had done thus far.

Could an AI system review all the data, uncover patterns and determine that an OC tended to call one of three plays in a given situation? Then recommend the best defensive play to call?

I was hoping this was not ever brought up on a forum as this is a huge yes. That and even down to player tendencies, matched up with our players and types of plays that are statistically more likely to work. You can do types of defensive calls on your opponent to determine the better O play calls.

I would even have a LLM running during the game to help with adjustments. Its going to happen. I was hoping we could get a few years of advantage until it is the nom

 

[MountainBuzzMan]

From one who has done postgrad work in complex systems, they are extraordinarily difficult to predict with acceptable variance. Many say you cannot. There are too many feedback loops. Football is a complex system. Trends and tendencies is about the best one can do.

Thats why you would break it down into a lot of smaller ones. One just for a teams play calling tendencies on O. Another for D. Another for key players ect.... As you accumulate the different models then you start running the output of all the other model into another one. But find the aspects of each game that could help and create a model just on that one aspect. Trying to model the whole thing, we are not there today. But it there is no reason why you cant take baby steps.

 

[Spalding Jacket]

I feel like the sayings “paralysis by analysis” applies here as well as “KISS.” Too many variables to keep track of, reminds me of Al Groh and his complex defenses.

 

[TampaBuzz]

Darn....I thought this thread was gonna be about Al Ciraldo and play calling.

 

[GTRambler]

Won’t work. Football will still be played by human beings.

 

[BuzzStone]

With the advancements that the AI and ML (Machine Learning) have made I see it as a massive edge for those teams that adopt it first. Using AWS or Azure or any number of platforms you can easily ingest a plethora of data and come up with some really good prediction models.