AI Playcalling

Submitted by Laveranues on

In light of recent news (this, among much else), how long until someone tries to implement AI playcalling? Some obvious difficulties in identifying and loading the pre-snap positioning / motion of the opposing teams (particularly on offense) and getting the play call in on-time (particularly on defense). But if that can be overcome - and it seems like it could, even with current tech - this seems like a legitimate, perhaps even inevitable, option.

Thoughts on feasibility? Legality? Impact on the game?

JHumich

October 19th, 2017 at 7:43 PM ^

kind of algorithm to spit out suggested plays.

But, there are so many things in the mannerisms and even the eyes of the players on the field, noticing a tell that is giving away the defense, etc....

Total replacement by AI would take a much greater level of sophistication than we are likely to have for a few years.

Laveranues

October 19th, 2017 at 8:02 PM ^

But, literally, only a few years. A lot of AI development is currently focused on image recognition, which could be used to catalogue the posture / positioning / mannerisms of players and take that into account. Obviously success is dependent on the quality of the algorithm and the machine learning that allows it to improve over time.

But that is the core competency of AI: it can evaluate every factor nigh instantaneously. In games that prioritize tactics over execution (Chess, Go), modern AI is quite literally unbeatable. It is hard to imagine this wouldn't lend at least some RPS advantage over a meatbag coach's decision making. (There is still execution to consider, though.)

Best case scenario: parity is improved as universities with an academic focus offset a talent disadvantage with superior machine learning and image-recognition algorithms.

4yearsofhoke

October 19th, 2017 at 11:52 PM ^

I think you could somewhat feed a list of plays coded by a number or something into a machine learning environment based on the opponents defense, down and distance, where you are on the field, and past performance etc....

However the amount of variables and seemingly the lack of a large enough sample size per play (because it’s be limited to one team) I think would make the algorithm worthless.

Mr Miggle

October 20th, 2017 at 1:25 AM ^

Which is not to say an AI program can't also be strong. Deep Blue had a tremendous hardware advantage over other engines and was still probably never the top engine, despite IBM investing a lot into its development and publicity.

I can't say as much about AlphaGo, but the amount of effort put into programming go engines is miniscule compared to chess. The AI approach was successful. Is it optimal for a finite game where computers have an overwhelming advantage in calculations? I'd guess not, but they demonstrated it could learn to beat humans and I suppose that was their point.

Note that it was quite different than what's talked about for playcalling. AlphaGo learned the game by playing against itself, not by observing humans and making assumptions about their body language, etc. 

go16blue

October 19th, 2017 at 7:35 PM ^

You could do that a couple of ways I guess. The most obvious one to me would be to build a model to predict the opposing team's playcall, and pick a play designed to beat it. But if you are using such a model and the other team knows it, they'd be able to intentionally trick it (because your play calling would be so rigid, you'd be much more vulnerable to the other team breaking tendency). You could also build a model to predict what plays will be most effective on down/distance based on previous data, but I'm just not sure that would give you good playcalls. So outside of the logistical challenges you said, my gut would say it's unlikely to give a strategic advantage. 

It should also be said that coaches do a lot of work to nail down the opposing team's tendencies. Having a model feed them the other team's likely playcall based on richer data than just their read from watching film seems like it could be a good idea, but isn't that different from what they already do. 

Laveranues

October 19th, 2017 at 7:42 PM ^

True, but eventually you could accumulate enough data to offset their attempts to exploit  tendencies. An algorithm could factor in what the opposition ran out of a similar formation many years ago (and all years since); also taking in to account all of the variations that were run against your offensive / defensive formation, probabilities of success, game time, down-and-distance, player skill, coaching tendencies, etc., to calculate the playcall / formation with the greatest likelihood of a positive outcome.

CommodityTrader

October 19th, 2017 at 7:55 PM ^

there is nothing about football that couldnt be accumulated into a data set and run through AlphaMaster or AlphaGoZero. AlphaMaster learned from testing human interactions and human play to become the Master. Zero learned Tabula Rasa... blank slate without human interaction or play and surpassed the Master level in 21 days if iterations. those who apply AI the fastest will have an advantage and they will learn how to improve it the fastest . to be fair... the team would be smart to engage the AI teams at thw school for project work. it would be hugely beneficial to learn human physical interactions.

treetown

October 20th, 2017 at 9:30 AM ^

"I don't care what the past records are, we need to go upgrade our Offensive Module to AlphaO 10 rev 2.1. It recently won the computer AI bowl - look 10,000 head to head games with a 76.55% win record for positive yardage is significant. Even if the license costs $40 million new dollars a season, it will be worth it. And will someone please update the tendencies in BlitzoMatic? The recent diary here shows we are trending towards predictability on 3rd down and not enough pseudo-randomness"

PrincetonBlue

October 19th, 2017 at 8:33 PM ^

I've thought a little bit about this problem, but it is, like others have said, whole orders of sophistication higher. It would be like playing Go, except the pieces on the other side change (the abilities of the other team's pieces depend on the team you are playing). AlphaGo works by using reinforcement learning, which requires tons of data (games played) to converge to the optimum state. Football games are not so easily played, and the additional degree of freedom in the pieces would make the required sample size to be so large that it would be ridiculous. Any play calling AI would probably have to work differently from AlphaGo.

Laveranues

October 19th, 2017 at 9:22 PM ^

But if you evaluate only the begin-state (pre-snap positioning & playcall & gamestate) and succes (yards gained), the aforementioned degree of freedom can be ignored. There's still no quantitative means of defining the ability of the pieces, but they could be estimated to improve the quality of the prediction.

PrincetonBlue

October 19th, 2017 at 9:24 PM ^

The reason AlphaGo is so strong is that all the relevant data that you need is right on a NxN board, which is not a ton of data to have to process.  On the other hand, what can you say are the relevant features that you need to consider before making a call?  Like you said, you have positioning & playcall.  Those two things are not quite as easy to see as pieces on a board.  AlphaGo looks at the state of the board, which is well-defined at all times, to make a decision.  On the other hand, AlphaFootball is going to have difficulty figuring out the "state" of the field.  Finally, it will also have to figure out opposing playcalls from the movements of the other team's players. 

It's all very complicated, not to mention other features you might consider, like opponent's playcalling tendencies, game film, and a whole host of other things that coaches use to inform their own decision making.

Not to mention, there's a tension here.  The more features you consider, the better you'll get at making decisions.  But with more features, you need more data.  I don't think many teams are up for letting a computer train on their practices for thousands of samples, and no team wants a useless bot that doesn't consider enough features.      

However, I might be wrong.  I don't claim to be an expert on the subject, since I haven't read the paper on AlphaGo yet, or any of the other DeepMind reinforcement learning papers.  Maybe I'll have a more informed opinion later.  Personally, I'm hoping that this is something currently doable, since I've been thinking about working on something like this for my senior thesis/junior independent work.  These are just all of the arguments that I imagine my professors making if I were to bring this up as a proposal.

Laveranues

October 19th, 2017 at 9:57 PM ^

Glad you mentioned it, since I was envisioning an NxN board for the pre-snap envaluation. It might be hard to assign coordinates to the O/D lines, especially with motion on the defensive side, but the rest of the formation could reasonably be assigned NxN coords. Not sure what size it would be, but any formation could be described with x,y coordinates of sufficient resolution.

5th and Long

October 19th, 2017 at 8:34 PM ^

The foundations for AI have already been laid in the NBA, and with other teams and leagues.  There are two companies that have camera systems that track the players in the game and analyze the information.  SportsVu is the company the NBA went with and installed them league wide in 2013.  They're system is tied to a statistic platform, however it can be used to track team tendencies and even player tendencies.  A google search will turnup a youtube video of the system in action.

Second Spectrum is the company that the NBA replaced SportsVu with starting this season.  They have camera tracking software for teams, that spit out data and analytics for teams, and the league.  But their marketed focus is on machine learning to help your team gain an edge...

 

 

 

UMfan21

October 19th, 2017 at 9:07 PM ^

is rather use AI to finally figure out what targeting is, and enforce it consistently.

Also, to correctly identify illegal men downfield, and intentional grounding.

ndekett

October 19th, 2017 at 9:30 PM ^

While there may be rules against it, the rules will eventually become unenforceable (counterpoint: amateurism). Computers will be infinitely better. I bet they never call a McDoom fade! Maybe no fades at all! Anyway, I love the post. Whether it happens or not, playcalling ought to be in the first 20% of human jobs replaced by AI since it just lends itself so well to Bayesian updating on the fly that humans could never do well.

treetown

October 19th, 2017 at 10:33 PM ^

Actually, play calling by programming in probablities based on the down and distance, time, and score has been used in some computer games simulating football. There was a pretty good game called PlayMaker Football back in the early Motorola CPU Mac days (back when 1 megabyte of memory was considered enormous) that did a pretty good job. It was later ported to the PC.

Each play was designed on a chalkboard mode and it was pretty sophisticated. Passing had multiple options of receivers, there were fake handoffs, pump fakes, shading of lineman on defense, zone, etc.

Esterhaus

October 19th, 2017 at 11:46 PM ^

AI system can adjust but will consistently revert to the highest probability play call. On-field players can almost always adjust faster than the system can adjust thus out-foxing and communicating it. Unless we wire on-field players to the AI, a smart football team of equal talent should win majority of the plays provided it can adjust rapidly and properly. Keep the comms to sideline signals and you can reliably defeat the ai ooda loop. Ai can be a helpful tool but it won't win against smart football teams.

bacon

October 20th, 2017 at 12:11 AM ^

Considering that Larry Page and Sergey Brin have connections to Michigan, they might be the ones who spearhead this idea. Google playcall could be an option.

laus102

October 20th, 2017 at 2:02 AM ^

I think that would be an incredibly difficult thing to use AI for. At this point, artificial intelligence is a misnomer. Even IBM's Watson is little more than a glorified dictionary.

Add in the fact that a lot of football is more than Xs and Os... there is so much more to a game then that, I'm not sure how useful AI would be.

Sam1863

October 20th, 2017 at 5:39 AM ^

Like several people, I saw this title and thought, "Who the hell is Al Playcalling? The new OC?"

No doubt he'll be followed by the new defensive coordinator, Greg Defensiveadjustmentsathalftime.

I know - quit drinking and go to bed.

Sambojangles

October 20th, 2017 at 9:13 AM ^

I think I the possibility exists on a spectrum, with ever-increasing resolution and effectiveness. It's possible today to write a program to suggest plays based on limited parameters: down/distance, game state, personell. Madden does it already.

I wonder if AI could be used for player evaluation. Every coach and pundit has their own biases when it comes to who should be recruited, given down by down playing time, and ultimately drafted. Perhaps a program could watch practice and game film to help show which player is better than another in certain situations, and inform player decisions on that level.