Beyond analytics?

[mjd1001]

Maybe anyone out there is interested in this topic, Its not something I think about all the time but I do think about it.

What is next after analytics...AI?  Or maybe AI makes analytics better/more useful?  And what exactly is AI we are talking about?  In some ways Artificial intelligence is overhyped, it is no where near doing some things we want. In other ways, it is so far advanced, doing things we haven't ever dreamt of.  In the middle, AI is advanced but it really isn't what we think of AI as...it is analyzing tons of data but in some ways its not quite teaching itself and evolving as much as we thought it would be.

Put that discussion aside, Can it give a team an edge in a sport?  Can AI analyze tactics in hockey, maybe play calling in football, and determine trends quicker than a whole room full of coaches does now?  Can it 'look for' trends analyzing ALL data (even data we don't look at) and learn which prospects are more likely to be stars?  Or maybe the particular type of 'development' a player needs to better reach their peak?

Think about this, trends that have worked in sports (the trap, the left wing lock in hockey, or different formations in football), at one point they weren't being used, someone thought of them, put them into play......can't things like that be 'thought' of quicker and discovered and tested 'quicker' by using complex programs/AI based software?

Maybe complex programs aren't there yet, and might be years away, but how much should each team be investing into this?  Should a team like the Sabres or Bills hold open interviews....interview software engineers who are involved in AI and basically say to them "We don't know much about AI in sports, but tell us what you think YOU can do to implement an AI program and tell us how it can help us?"

I saw a post in Twitter and it is showing the current level of software/AI...how much it actually is impacting Amazon right now, that is what got me thinking. Anyone have any thoughts?

Edited August 25 by mjd1001

[Taro T]

Nothing deeply thought out but do have a few thoughts on it.

1st.  AI or any search engine / data parser / whatever tool you're working with is only going to be as good at whatever task as whomever programmed it to be is and it will necessarily have the same biases as the programming team.  So, it could review data significantly quicker than a human (well, no Schlitz Sherlock) but it isn't necessarily going to see something in a different light - because it can't.

2nd.  As it does get used as a tool in new fields, it will necessarily get more useful as newer iterations come along.  No idea how long that would take as sports in general is a very big industry but it's also kind of a niche in that people making money off of owning professional teams are ridiculously few.  Now, could gamblers start using it?  The big ones almost certainly already are.  But, the gamblers aren't going to be suggesting line combinations nor forechecking strategies; they're going to look for trends in the end results and act accordingly.  Wouldn't even hazard a WAG as to how soon it could be affecting outcomes because of the front offices and coaching staffs making use of it. 

3rd.  AI doesn't actually "think" like we consider thinking so it's going to be way more useful in interpolating data and trends than extrapolating them.  And again, the team that's programming it would really need to know what they're looking for to try to take what it comes up with and take it "outside the box."

4th.  Expect an individual team would be better off talking to the major players in AI about how they could make use of what those companies have developed and outsourcing (with a ton of agreements protecting the propriety of what those companies come up with for the team) any AI "research" projects rather than trying to have a small team of their own developers trying to do something with it.  

[LTS]

AI will be extremely strong when given the task of working inside something structured, like writing code.  There is no end to the amount of sample code and best practices to draw from.

AI is not as good at concepts, but it is getting better.

However, to think of AI is as a data parser/search engine is definitely underestimating it.  While it's true that AI begins only as good as whomever programmed it, the good AI is reprogramming itself on a continual basis. The ability for AI to run an extreme amount of simulations with an inordinate amount of data points allows for it to hypothesize, test, and prove in rapid fashion.  Certainly much faster than any human.

If I were looking at sports applications I think an AI's given access to a depth of data could derive optimal strategies. It's a little mind boggling to think about how much data it could use to accomplish this. Imagine a world where athletes are covered in sensors that provide continual data points and then an AI that can consume that data.  It builds models that are predictive not only to each player but player types.  Once it models players it can then predict success rates based on comparing the data on player A versus Player B given the environmental variables of the moment. It could, in theory, analyze and entire defensive unit, its players, plays, etc. and build an optimal offensive strategy given the teams offensive players, plays, etc.

Certainly there's the fact that players are human and have to execute. Things like a foot slipping on turf at the snap could ruin the entire model, but given enough data an AI could even get to the point of predicting that based on the real time feedback from sensors. The player is exhibiting this level of exhaustion, the turf is demonstrating this level of wear, and so on.. then it predicts that the player's foot will slip.  If it's correct, it uses it in support of its model and if its incorrect it makes a tiny change to its model and then determines if that prediction is better.

It's hard to conceptualize that the entire world and everything that happens could be broken down into data and a predictable model created.  I firmly stand in the camp that there is no such thing as randomness. Any appearance of random is simply the result of a set of variables that at that point in time are incalculable. At some point the ability to collect all relevant data and process it into a calculation will eliminate what had appeared as "random" event.

[Taro T]

33 minutes ago, LTS said:

AI will be extremely strong when given the task of working inside something structured, like writing code.  There is no end to the amount of sample code and best practices to draw from.

AI is not as good at concepts, but it is getting better.

However, to think of AI is as a data parser/search engine is definitely underestimating it.  While it's true that AI begins only as good as whomever programmed it, the good AI is reprogramming itself on a continual basis. The ability for AI to run an extreme amount of simulations with an inordinate amount of data points allows for it to hypothesize, test, and prove in rapid fashion.  Certainly much faster than any human.

If I were looking at sports applications I think an AI's given access to a depth of data could derive optimal strategies. It's a little mind boggling to think about how much data it could use to accomplish this. Imagine a world where athletes are covered in sensors that provide continual data points and then an AI that can consume that data.  It builds models that are predictive not only to each player but player types.  Once it models players it can then predict success rates based on comparing the data on player A versus Player B given the environmental variables of the moment. It could, in theory, analyze and entire defensive unit, its players, plays, etc. and build an optimal offensive strategy given the teams offensive players, plays, etc.

Certainly there's the fact that players are human and have to execute. Things like a foot slipping on turf at the snap could ruin the entire model, but given enough data an AI could even get to the point of predicting that based on the real time feedback from sensors. The player is exhibiting this level of exhaustion, the turf is demonstrating this level of wear, and so on.. then it predicts that the player's foot will slip.  If it's correct, it uses it in support of its model and if its incorrect it makes a tiny change to its model and then determines if that prediction is better.

It's hard to conceptualize that the entire world and everything that happens could be broken down into data and a predictable model created.  I firmly stand in the camp that there is no such thing as randomness. Any appearance of random is simply the result of a set of variables that at that point in time are incalculable. At some point the ability to collect all relevant data and process it into a calculation will eliminate what had appeared as "random" event.

You're never going to get to a point where the world can become fully predictable because so much of the world (weather being a great example) is best described as strange attractors which are systems that in overly simplified terms essentially repeat but depending upon the starting point where you enter / start tracking the system you will wind up with wildly divergent outcomes.

And, no matter how hard one tries, really doubt AI will ever be able to account for the emotion of a situation.  AI will never (IMHO) be able to understand emotion as it will never be able to experience it.  And if it can't understand it, it will never be able to fully account for it.  And the emotion of a moment, and how humans deal with them, will make far more of a difference to an outcome than a random loss of footing might.  Again, MHO.

[mjd1001]

I guess my thinking is...We are where right now very close to what Taro T is saying, yet its possible to get to where LTS is saying.

My question is, how long between those? And at what point does some super-rich owner who may (or may not) know anything about AI programming decide he/she wants to develop a department to try to see what it can do for the team.

Maybe its not strictly "learning AI" but something like....analyzing film of your own players just like coaches do, but analyzing each player and doing so almost frame by frame, 1/10th of a second by 10th of a second, seeing how often they react one way vs react another way, spitting out trends on every player in the league in more detail than ever before..and going from there?  Seeing how you can exploit other players, or how your players can get better, not on a macro level, but on a tiny small micro level that coaches haven't broken down film before.

  Maybe almost as close as seeing a player who 'tips' which way he is going to pass the puck by the slightest positioning of his hands on his stick, or a small 'tell' that someone positions their feet slightly differently before a pass vs a shot. Something that coaches simply cannot analyze every single player on ever team, frame by frame of video...that a computer can do in a matter of minutes.

Edited August 25 by mjd1001

[R_Dudley]

I love the nerd discussion here.  With large language models(LLM) that AI uses along with the right data it is provided it can derive sentiment but not emotions or feelings. It's what they call training AI and is what is needed to have it hypothesize or generate the results for the models, approaches, or strategies you are using it for. A team would just need a good data scientist or two and several other staff(2-3) feeding the data points they would use to train the models they are working on.

I could see a team owner potentially going down this path and putting together a team of 4-6 staff to see what it can do. The challenge would be in establishing the baselines for getting the same data points being used across as many teams and player types as you could get assuming there is an agreed upon definition and collection for the data being used across teams.

Remarkably interesting technology discussion though and my first in the last 25 days since my retirement on 7/31 from my IT career of 43 years.    

[nfreeman]

It's only a matter of time until...

[French Collection]

6 hours ago, Taro T said:

You're never going to get to a point where the world can become fully predictable because so much of the world (weather being a great example) is best described as strange attractors which are systems that in overly simplified terms essentially repeat but depending upon the starting point where you enter / start tracking the system you will wind up with wildly divergent outcomes.

And, no matter how hard one tries, really doubt AI will ever be able to account for the emotion of a situation.  AI will never (IMHO) be able to understand emotion as it will never be able to experience it.  And if it can't understand it, it will never be able to fully account for it.  And the emotion of a moment, and how humans deal with them, will make far more of a difference to an outcome than a random loss of footing might.  Again, MHO.

The best sports example I can think of is golf. The elements can turn on a dime, especially in a place like Scotland. The field of play changes every day, pin positions, tee boxes, grass and turf conditions. Throw in the emotional response to great and horrible shots, being in contention, being near the cut line etc. and there is tons of uncertainty.

It was easier to predict in the Tiger era as he was so much better than his opponents. Even in that era he would still not win 2/3 of the tournaments he played in.

[LTS]

7 hours ago, Taro T said:

You're never going to get to a point where the world can become fully predictable because so much of the world (weather being a great example) is best described as strange attractors which are systems that in overly simplified terms essentially repeat but depending upon the starting point where you enter / start tracking the system you will wind up with wildly divergent outcomes.

And, no matter how hard one tries, really doubt AI will ever be able to account for the emotion of a situation.  AI will never (IMHO) be able to understand emotion as it will never be able to experience it.  And if it can't understand it, it will never be able to fully account for it.  And the emotion of a moment, and how humans deal with them, will make far more of a difference to an outcome than a random loss of footing might.  Again, MHO.

I disagree and I suppose it's a faith question because I certainly can't prove that we will. I think it's safe to say that emotion can already be predicted to a certain level and I think it's easy to mentally get to a point where understanding emotional response can be mapped to data points.  Again, we're talking about gathering an insanely large dataset in order to achieve this. Far beyond what is possible today.

4 hours ago, R_Dudley said:

I love the nerd discussion here.  With large language models(LLM) that AI uses along with the right data it is provided it can derive sentiment but not emotions or feelings. It's what they call training AI and is what is needed to have it hypothesize or generate the results for the models, approaches, or strategies you are using it for. A team would just need a good data scientist or two and several other staff(2-3) feeding the data points they would use to train the models they are working on.

I could see a team owner potentially going down this path and putting together a team of 4-6 staff to see what it can do. The challenge would be in establishing the baselines for getting the same data points being used across as many teams and player types as you could get assuming there is an agreed upon definition and collection for the data being used across teams.

Remarkably interesting technology discussion though and my first in the last 25 days since my retirement on 7/31 from my IT career of 43 years.    

AI training 100%. Imagine a body suit that is full of thousands of sensors that can read the movement, temperature, resistance, pressure, and more of the human body. Imagine a hockey stick filled with enough sensors to be able to measure positioning, motion, tension, etc. Now put that on hundreds of thousands of hockey players and draw the data models over decades.  Certainly it's not possible today, but its not hard to theorize what it would take, especially when engaging doctors, engineers, and other areas of specialty who are experts in their particular fields.

We're maxing out our current technology but quantum computing will take computational power to incredible new levels. This paves the way for the ability to collect and process exponentially more data than what is possible today.

[RochesterExpat]

10 hours ago, LTS said:

However, to think of AI is as a data parser/search engine is definitely underestimating it.

Most of the public AI products are glorified search engines with some rudimentary predictive analytics baked in. I can’t fault people for thinking this.

 

[Demoted]

AI is the future of everything. 

Edited Monday at 05:23 PM by Demoted

[PerreaultForever]

I've been hearing a lot of podcasts with "expert" guests (hockey people) saying there's a move away from analytics currently and/or a trend towards a more balanced approach of using analytics but also going back to the eye test and other factors. I don't know if that's true but maybe some analytic failures have made people second guess. Size is in vogue again as well. 

[mjd1001]

4 hours ago, PerreaultForever said:

I've been hearing a lot of podcasts with "expert" guests (hockey people) saying there's a move away from analytics currently and/or a trend towards a more balanced approach of using analytics but also going back to the eye test and other factors. I don't know if that's true but maybe some analytic failures have made people second guess. Size is in vogue again as well. 

Analytics looks just at the numbers for the most part. Very deep And specific numbers, but you're just analyzing the numbers. 

Is it possible artificial intelligence IS The move back to more of an eye test? The artificial intelligence could be what analyzes the videos, looks for trends, and merges that with the pure analytics?

[LTS]

11 hours ago, RochesterExpat said:

Most of the public AI products are glorified search engines with some rudimentary predictive analytics baked in. I can’t fault people for thinking this.

 

I think a lot of what people get from public AI stems from what they put into it. If you write a fairly generic prompt you get a fairly generic response. Of course writing a more detailed prompt requires one to have the capacity to structure the prompt in the first place. This means having a more in depth understanding of the information they are looking to get and a lot of people just want AI to answer questions for them.

It can of course do that and through iterative prompt writing its possible to eventually get to a more in depth response.

The search engine viewpoint is certainly solidified by Google's Gemini offering "AI" summaries in its search results. Those summaries generally just being recaps of the top 3 search engine results and then parsed to eliminate duplicate data.

 

[Big Guava]

On 8/25/2024 at 5:53 PM, LTS said:

AI will be extremely strong when given the task of working inside something structured, like writing code.  There is no end to the amount of sample code and best practices to draw from.

AI is not as good at concepts, but it is getting better.

However, to think of AI is as a data parser/search engine is definitely underestimating it.  While it's true that AI begins only as good as whomever programmed it, the good AI is reprogramming itself on a continual basis. The ability for AI to run an extreme amount of simulations with an inordinate amount of data points allows for it to hypothesize, test, and prove in rapid fashion.  Certainly much faster than any human.

If I were looking at sports applications I think an AI's given access to a depth of data could derive optimal strategies. It's a little mind boggling to think about how much data it could use to accomplish this. Imagine a world where athletes are covered in sensors that provide continual data points and then an AI that can consume that data.  It builds models that are predictive not only to each player but player types.  Once it models players it can then predict success rates based on comparing the data on player A versus Player B given the environmental variables of the moment. It could, in theory, analyze and entire defensive unit, its players, plays, etc. and build an optimal offensive strategy given the teams offensive players, plays, etc.

Certainly there's the fact that players are human and have to execute. Things like a foot slipping on turf at the snap could ruin the entire model, but given enough data an AI could even get to the point of predicting that based on the real time feedback from sensors. The player is exhibiting this level of exhaustion, the turf is demonstrating this level of wear, and so on.. then it predicts that the player's foot will slip.  If it's correct, it uses it in support of its model and if its incorrect it makes a tiny change to its model and then determines if that prediction is better.

It's hard to conceptualize that the entire world and everything that happens could be broken down into data and a predictable model created.  I firmly stand in the camp that there is no such thing as randomness. Any appearance of random is simply the result of a set of variables that at that point in time are incalculable. At some point the ability to collect all relevant data and process it into a calculation will eliminate what had appeared as "random" event.

AI is good at writing code at times. Mostly boilerplate code and unit tests, etc.

The more complex what needs to be coded, the less useful AI will be. In some instances it actually hurts the person using it because it constantly sends them down deep rabbit holes in the wrong direction or gives nonsense code that doesn't do anything like what you requested or to provide.

[RochesterExpat]

1 hour ago, LTS said:

I think a lot of what people get from public AI stems from what they put into it. If you write a fairly generic prompt you get a fairly generic response. Of course writing a more detailed prompt requires one to have the capacity to structure the prompt in the first place. This means having a more in depth understanding of the information they are looking to get and a lot of people just want AI to answer questions for them.

It can of course do that and through iterative prompt writing its possible to eventually get to a more in depth response.

The search engine viewpoint is certainly solidified by Google's Gemini offering "AI" summaries in its search results. Those summaries generally just being recaps of the top 3 search engine results and then parsed to eliminate duplicate data.

 

To your point, you are also able find information a lot easier using a search engine if you know how to craft a query. It's why the best members of the IT Department tend to be those who know how to write a better search query. For the overwhelming majority of users, "AI" is simply a search engine evolved. That's not a surprise either. It's what consumers want.

The majority of AI products the public has access to (and would be interested in anyway) are basically search aggregates that use NLP to formulate summaries of the top results. It's the NLP portion that people find impressive and that is why most people simply don't understand as since they believe all the model is doing is interpreting a written question, looking at the top search results, and generating a human-readable response. Which, funny enough, is pretty much exactly what it is doing. We know ChatGPT used Google's ranking algorithm to determine priority for neighboring relationships. It's why the answers are terrible because, as it turns out, if you use Quora as a source of truth, you've poisoned the well. Not to mention StackOverflow.

Even when we shift the conversation to generative AI, it's the NLP/NLG aspects that most people care about and are impressed by. In part, that's because modern public AI models are, frankly, awful at anything else. Once again, this is driven by the consumer. The niche case of the person using generative AI to to analyze proteins and develop new drugs is exceedingly niche. On the other hand, the person who wants help writing a thank you note or generating a background for a birthday party invitation is anything but niche.  Who do you tailor your model to?

I understand you are bullish on AI, but what you are discussing (thousands of sensors on an athlete's body, for example) is still very far away from being even considered for implementation. There's insufficient computing power to make it a cost effective venture. Especially when we're consuming resources so I can generate a video of a turtle doing the macarena. 

5 hours ago, mjd1001 said:

Is it possible artificial intelligence IS The move back to more of an eye test? The artificial intelligence could be what analyzes the videos, looks for trends, and merges that with the pure analytics?

It's probable that eventually it replaces the eye test. To answer your point, yes, AI can be used to analyze a video for any number of things. If we're considering how scouts evaluate a player with the "eye test" (namely positioning) this is a pretty trivial application and I would be surprised if teams aren't already employing it.

A simple understanding of a simple application for video replacing scouting would be to think of it this way. The team creates the ideal, "model" player for a position or role. This model would be used as a reference for on-ice positioning. You then compare the location of the player you are interested in evaluating on each video frame to determine how far the player is deviating from the model. This would allow you to score players. You can establish zones as well so each player is assigned a zone. This is basically what scouts already do during player evaluation. However, now I can process every single shift of every single CHL game for a prospect in less time than a scout can sit through a single game to scout a single player.

The simplest use case would likely be a goaltender. The hurdle here is it requires a video that shows the goalie's positioning before the play/puck arrives to him. A lot of the lower level leagues don't have the greatest camera work so it's doubtful you'd be able to seriously apply it, but this would almost certainly work at the NHL level. 

10 hours ago, PerreaultForever said:

I've been hearing a lot of podcasts with "expert" guests (hockey people) saying there's a move away from analytics currently and/or a trend towards a more balanced approach of using analytics but also going back to the eye test and other factors. I don't know if that's true but maybe some analytic failures have made people second guess. Size is in vogue again as well. 

I think part of this is due to the the league swinging too far into the analytics as teams were searching for any edge in order to win and data analytics being somewhat of a snake oil industry at times. The NHL is an incredibly small sample size group. Even the best analytics fail at times. When you start narrowing your data sets to search for even the slightest edge, your decreasing the sample size further, and the likelihood of an "analytic failure" increases. Look at Eric Comrie's numbers before he came to Buffalo. No "fancy stat" exists that can account for "what if this goalie plays more than 20 games?" and no "fancy stat" exists for "is this goalie able to handle the mental struggles that accompany an injury?"

It's quite possible--maybe even likely--that, given a 1,000,000 Eric Comries across 1,000,000 NHLs, the majority of them turn into top NHL goalies (as his stats the year before he arrived in Buffalo would suggest). The issue is when you have a single Eric in a single NHL and he gets injured and never recovers.

7 minutes ago, Big Guava said:

AI is good at writing code at times. Mostly boilerplate code and unit tests, etc.

The more complex what needs to be coded, the less useful AI will be. In some instances it actually hurts the person using it because it constantly sends them down deep rabbit holes in the wrong direction or gives nonsense code that doesn't do anything like what you requested or to provide.

"Has this coding problem or approach been thoroughly solved and repeated by a human?"

-> If Yes, AI is more likely to generate workable code.

-> If No, AI is more likely generate work for the programmer.

 The latter is honestly helpful on occasions where you aren't sure where to get started or how to approach the task. But when it's unhelpful, it's really unhelpful.

[PerreaultForever]

8 hours ago, mjd1001 said:

Analytics looks just at the numbers for the most part. Very deep And specific numbers, but you're just analyzing the numbers. 

Is it possible artificial intelligence IS The move back to more of an eye test? The artificial intelligence could be what analyzes the videos, looks for trends, and merges that with the pure analytics?

Can't see that working. AI is getting sophisticated but the old early days of programming adage still holds - "garbage in garbage out". Your AI is only going to work with the parameters you give it and if your weight certain aspects over others it will do that as well. 

In my view AI, or analytics in general for that matter, can help you sort through the massive amount of potential players out there, but after shortening your list you still want your version of Rick Dudley to go down to those rinks and say "that guy has it, that guy doesn't."

[DarthEbriate]

The ETA of an EEE AI scouting department is... 3 years?

[mjd1001]

3 hours ago, RochesterExpat said:

It's probable that eventually it replaces the eye test. To answer your point, yes, AI can be used to analyze a video for any number of things. If we're considering how scouts evaluate a player with the "eye test" (namely positioning) this is a pretty trivial application and I would be surprised if teams aren't already employing it.

A simple understanding of a simple application for video replacing scouting would be to think of it this way. The team creates the ideal, "model" player for a position or role. This model would be used as a reference for on-ice positioning. You then compare the location of the player you are interested in evaluating on each video frame to determine how far the player is deviating from the model. This would allow you to score players. You can establish zones as well so each player is assigned a zone. This is basically what scouts already do during player evaluation. However, now I can process every single shift of every single CHL game for a prospect in less time than a scout can sit through a single game to scout a single player.

The simplest use case would likely be a goaltender. The hurdle here is it requires a video that shows the goalie's positioning before the play/puck arrives to him. A lot of the lower level leagues don't have the greatest camera work so it's doubtful you'd be able to seriously apply it, but this would almost certainly work at the NHL level. 

 

Could computers/programming/AI 'reverse engineer' (in a way) what I highlighted above? The Computer/AI creates the model for the team by analyzing what is working.

Meaning, instead of you creating the ideal 'model' of what you want and then have the ai look for it...how about AI simply look at what is working (who are the best goaltenders) and analyze what they are doing that makes them successful?  Do they come out of the net a few inches more than other goalies?  Do they play a rush from 25% off the center of the net slightly different position than other goalies do?  Are there things ANY goaltender does when they do have a good game/good save that they don't even know they are doing that makes them successful?  The player themself might not know that...and the difference between success and failure might be something SO SMALL that the human eye doesn't recognize it.....but a computer/AI can analyze things frame by frame, hundreds of thousand of total frames of gameplay, and look for even the smallest things.   

It can then 'spit out' its conclusions, a human coach and look at it and see if its on to something that no one else saw (or had the time to see).  That can then be used as a coaching tool for your current players, or you now can have the SAME AI look for things in scouting video of free agents/prospects.

Edited August 26 by mjd1001

[RochesterExpat]

2 hours ago, mjd1001 said:

Could computers/programming/AI 'reverse engineer' (in a way) what I highlighted above? The Computer/AI creates the model for the team by analyzing what is working.

Meaning, instead of you creating the ideal 'model' of what you want and then have the ai look for it...how about AI simply look at what is working (who are the best goaltenders) and analyze what they are doing that makes them successful?  Do they come out of the net a few inches more than other goalies?  Do they play a rush from 25% off the center of the net slightly different position than other goalies do?  Are there things ANY goaltender does when they do have a good game/good save that they don't even know they are doing that makes them successful?  The player themself might not know that...and the difference between success and failure might be something SO SMALL that the human eye doesn't recognize it.....but a computer/AI can analyze things frame by frame, hundreds of thousand of total frames of gameplay, and look for even the smallest things.   

It can then 'spit out' its conclusions, a human coach and look at it and see if it’s on to something that no one else saw (or had the time to see).  That can then be used as a coaching tool for your current players, or you now can have the SAME AI look for things in scouting video of free agents/prospects.

In a sense that’s already how it would be done. You would train your model using the top tier goaltenders. This would be the “model” goalie used to evaluate others. As far as detecting subtle nuances to gameplay, a computer with sufficiently detailed video and a large enough data set is almost certainly going to reveal patterns of goalie behavior that we didn’t expect.

It’s unlikely the level of model training that you’re suggesting or trying to accomplish is technically feasible at present. But a rudimentary implementation is certainly doable.

[Jorcus]

I have a difficult time thinking of a use case for AI in hockey. I can think of many uses for AI outside of hockey but I can't see how you are going to use it to gain an advantage for on ice performance. For way too many years of my life I have handicapped horse races and have a good understanding of how advanced handicappers were developing AI type tools long ago to get an edge on the game. To some extent they were able to improve their chances of coming out ahead wagering but there were many things that were difficult for the system players to account for. One of the biggest issues was as much as you might know one horse is faster than the one standing next to it you really don't know how the horse is feeling that day. The best you can do is observe the paddock and the warmups but that is pretty much impossible to program and even when you think you know by observation you don't really know until they run the race. Like LTS pointed out a foot slip changes the dynamic of the outcome.  There are many other factors in a race that are difficult to account for that skew results and outcomes.

Hockey is like this in many ways. It's a very fast and reactionary sport. Players have to react to the conditions presented on the ice. They are given ideas of where they should be in relation to those conditions but a program can't do that for them. Maybe the program would want them to be some other place but they are not physically able to do that. It's like when a coach draws up a play during a time out late in the game. If that play does not work in the first few seconds there is no play anymore and it's back to reacting to the conditions on the ice.

  

[Big Guava]

On 8/26/2024 at 6:39 PM, RochesterExpat said:

To your point, you are also able find information a lot easier using a search engine if you know how to craft a query. It's why the best members of the IT Department tend to be those who know how to write a better search query. For the overwhelming majority of users, "AI" is simply a search engine evolved. That's not a surprise either. It's what consumers want.

The majority of AI products the public has access to (and would be interested in anyway) are basically search aggregates that use NLP to formulate summaries of the top results. It's the NLP portion that people find impressive and that is why most people simply don't understand as since they believe all the model is doing is interpreting a written question, looking at the top search results, and generating a human-readable response. Which, funny enough, is pretty much exactly what it is doing. We know ChatGPT used Google's ranking algorithm to determine priority for neighboring relationships. It's why the answers are terrible because, as it turns out, if you use Quora as a source of truth, you've poisoned the well. Not to mention StackOverflow.

Even when we shift the conversation to generative AI, it's the NLP/NLG aspects that most people care about and are impressed by. In part, that's because modern public AI models are, frankly, awful at anything else. Once again, this is driven by the consumer. The niche case of the person using generative AI to to analyze proteins and develop new drugs is exceedingly niche. On the other hand, the person who wants help writing a thank you note or generating a background for a birthday party invitation is anything but niche.  Who do you tailor your model to?

I understand you are bullish on AI, but what you are discussing (thousands of sensors on an athlete's body, for example) is still very far away from being even considered for implementation. There's insufficient computing power to make it a cost effective venture. Especially when we're consuming resources so I can generate a video of a turtle doing the macarena. 

It's probable that eventually it replaces the eye test. To answer your point, yes, AI can be used to analyze a video for any number of things. If we're considering how scouts evaluate a player with the "eye test" (namely positioning) this is a pretty trivial application and I would be surprised if teams aren't already employing it.

A simple understanding of a simple application for video replacing scouting would be to think of it this way. The team creates the ideal, "model" player for a position or role. This model would be used as a reference for on-ice positioning. You then compare the location of the player you are interested in evaluating on each video frame to determine how far the player is deviating from the model. This would allow you to score players. You can establish zones as well so each player is assigned a zone. This is basically what scouts already do during player evaluation. However, now I can process every single shift of every single CHL game for a prospect in less time than a scout can sit through a single game to scout a single player.

The simplest use case would likely be a goaltender. The hurdle here is it requires a video that shows the goalie's positioning before the play/puck arrives to him. A lot of the lower level leagues don't have the greatest camera work so it's doubtful you'd be able to seriously apply it, but this would almost certainly work at the NHL level. 

I think part of this is due to the the league swinging too far into the analytics as teams were searching for any edge in order to win and data analytics being somewhat of a snake oil industry at times. The NHL is an incredibly small sample size group. Even the best analytics fail at times. When you start narrowing your data sets to search for even the slightest edge, your decreasing the sample size further, and the likelihood of an "analytic failure" increases. Look at Eric Comrie's numbers before he came to Buffalo. No "fancy stat" exists that can account for "what if this goalie plays more than 20 games?" and no "fancy stat" exists for "is this goalie able to handle the mental struggles that accompany an injury?"

It's quite possible--maybe even likely--that, given a 1,000,000 Eric Comries across 1,000,000 NHLs, the majority of them turn into top NHL goalies (as his stats the year before he arrived in Buffalo would suggest). The issue is when you have a single Eric in a single NHL and he gets injured and never recovers.

"Has this coding problem or approach been thoroughly solved and repeated by a human?"

-> If Yes, AI is more likely to generate workable code.

-> If No, AI is more likely generate work for the programmer.

 The latter is honestly helpful on occasions where you aren't sure where to get started or how to approach the task. But when it's unhelpful, it's really unhelpful.

 

Yeah, we use GitHub Co-Pilot at work daily and honestly there are times I have been trying to get it to help for an hour and then go to stack overflow and I find what I need in like 5 minutes.

Always check Stackoverflow now before I ask to see if it's something I can find there first.