• Follow us on Twitter @buckeyeplanet and @bp_recruiting, like us on Facebook! Enjoy a post or article, recommend it to others! BP is only as strong as its community, and we only promote by word of mouth, so share away!
  • Consider registering! Fewer and higher quality ads, no emails you don't want, access to all the forums, download game torrents, private messages, polls, Sportsbook, etc. Even if you just want to lurk, there are a lot of good reasons to register!

Cleveland Browns (Finally drafting Buckeyes)

Jump to latest Follow Reply
Sort by date Sort by votes
AKAK said:
Is there a reason you're saying this? I'm not saying your not right, I just was wondering what the background is here.

(And if you say, "I walked by his office and he was playing solitaire on windows 3.1 - I will accept that)
Click to expand...

He got started in 2004 with Dodgers, using a different suite of stats to pick out players and this was revolutionary at the time.
But 15 years later, it's... underwhelming. I've never seen anything to indicate he's moved beyond that approach in both finding "new" statistics (ie espn' "Expected Gains" which is a respin of Basketball's Plus-Minus reflected in entire Offense / Defense / ST units) ... or in being focused on finding the right player more frequently.
One problem is that Baseball is a 1v1 sport most of the time. Pitcher vs Batter. It's an oversimplification, but it's one that's useful for statistics. Basketball becomes a bit more complex with 5v5... but Football jumps the shark completely. There's so much more going on in any given football play... the guys who don't have the ball are just as important as the one who does. DePodesta's approach seems utterly incapable of capturing the trenches, or the influence of OC vs DC mind games.
And I suspect this complexity is why a lot of us love football. It's the only sport I can think of where the Coaching Staff is a vital part of the team, a player themselves - though not on the field.
There's a lot of interesting things you could do, even with just vision processing that focuses on all22 footage, tracking the players, and having an intern add a few pieces of information (down and distance, yards gained, type of play - designed pass/run/broken)... and see what the computer finds as "meaningful" predictors. There's 256 games a season... go back 10 seasons, run the tape, and see what it picks out. Then start refining what the ML identified to pick out better features and rerun the same tests. You can run these tests 24hrs a day 365 days a year. Move to simulating 3 weeks into a season against a team with a new QB. The amount of cash, time, and good data available... everyone already has box scores and all22 footage... could make t a hotbed for Deep Learning and Neural Networks. Maybe I'm wrong, but I assume the guys in the booth are allowed to use a computer?
But even this raises some questions. Let's say your model predicts 80/20 run from a given formation and set with certain tells. And your DC lines up to stop the run, and they throw over your head with no Safety protection for a back-breaking 86 yard TD -- because the OC noted your reaction to his tendencies at halftime.
ML wouldn't pick up on things like Bud Foster running bear against us in 2014 either... if there's no sample data from somebody else doing it first, it's highly unlikely a model could predict that as an effective counter... and if you had a model that somehow did, it's unlikely to know the risk associated with doing it.
 
Upvote 0
kujirakira said:
He got started in 2004 with Dodgers, using a different suite of stats to pick out players and this was revolutionary at the time.
But 15 years later, it's... underwhelming. I've never seen anything to indicate he's moved beyond that approach in both finding "new" statistics (ie espn' "Expected Gains" which is a respin of Basketball's Plus-Minus reflected in entire Offense / Defense / ST units) ... or in being focused on finding the right player more frequently.
One problem is that Baseball is a 1v1 sport most of the time. Pitcher vs Batter. It's an oversimplification, but it's one that's useful for statistics. Basketball becomes a bit more complex with 5v5... but Football jumps the shark completely. There's so much more going on in any given football play... the guys who don't have the ball are just as important as the one who does. DePodesta's approach seems utterly incapable of capturing the trenches, or the influence of OC vs DC mind games.
And I suspect this complexity is why a lot of us love football. It's the only sport I can think of where the Coaching Staff is a vital part of the team, a player themselves - though not on the field.
There's a lot of interesting things you could do, even with just vision processing that focuses on all22 footage, tracking the players, and having an intern add a few pieces of information (down and distance, yards gained, type of play - designed pass/run/broken)... and see what the computer finds as "meaningful" predictors. There's 256 games a season... go back 10 seasons, run the tape, and see what it picks out. Then start refining what the ML identified to pick out better features and rerun the same tests. You can run these tests 24hrs a day 365 days a year. Move to simulating 3 weeks into a season against a team with a new QB. The amount of cash, time, and good data available... everyone already has box scores and all22 footage... could make t a hotbed for Deep Learning and Neural Networks. Maybe I'm wrong, but I assume the guys in the booth are allowed to use a computer?
But even this raises some questions. Let's say your model predicts 80/20 run from a given formation and set with certain tells. And your DC lines up to stop the run, and they throw over your head with no Safety protection for a back-breaking 86 yard TD -- because the OC noted your reaction to his tendencies at halftime.
ML wouldn't pick up on things like Bud Foster running bear against us in 2014 either... if there's no sample data from somebody else doing it first, it's highly unlikely a model could predict that as an effective counter... and if you had a model that somehow did, it's unlikely to know the risk associated with doing it.
Click to expand...
Analytics have a place in football as supplementary data. Leaning too heavily on them results in a roster of too many Seth Devalves, Corey Colemans and Cody Kesslers...which results in 1-31.
 
Upvote 0
NFBuck said:
Analytics have a place in football as supplementary data. Leaning too heavily on them results in a roster of too many Seth Devalves, Corey Colemans and Cody Kesslers...which results in 1-31.
Click to expand...

Any ML that leads you to draft Corey Coleman is flat out wrong.
And that's what im trying to get at with his being 20yrs in the past using 1v1 rudimentary baseball statistics vs the kind of stuff that lets me board ANA flights without ever being issued a boarding pass (not talking about phone either - they have a camera at the gate with facial recognition from passport and entry/exit customs.)

Im more interested in how ML could reveal offensive and defensive gameplans.
ML could get that 'scripted opening drive' effect every drive imo by analyzing the last few series against your history and their history.
But you're right that it will be an inferior tool in the hands of an incompetent staff.
It's not a replacement for having excellent play callers.
 
Upvote 0
kujirakira said:
He got started in 2004 with Dodgers, using a different suite of stats to pick out players and this was revolutionary at the time.
But 15 years later, it's... underwhelming. I've never seen anything to indicate he's moved beyond that approach in both finding "new" statistics (ie espn' "Expected Gains" which is a respin of Basketball's Plus-Minus reflected in entire Offense / Defense / ST units) ... or in being focused on finding the right player more frequently.
One problem is that Baseball is a 1v1 sport most of the time. Pitcher vs Batter. It's an oversimplification, but it's one that's useful for statistics. Basketball becomes a bit more complex with 5v5... but Football jumps the shark completely. There's so much more going on in any given football play... the guys who don't have the ball are just as important as the one who does. DePodesta's approach seems utterly incapable of capturing the trenches, or the influence of OC vs DC mind games.
And I suspect this complexity is why a lot of us love football. It's the only sport I can think of where the Coaching Staff is a vital part of the team, a player themselves - though not on the field.
There's a lot of interesting things you could do, even with just vision processing that focuses on all22 footage, tracking the players, and having an intern add a few pieces of information (down and distance, yards gained, type of play - designed pass/run/broken)... and see what the computer finds as "meaningful" predictors. There's 256 games a season... go back 10 seasons, run the tape, and see what it picks out. Then start refining what the ML identified to pick out better features and rerun the same tests. You can run these tests 24hrs a day 365 days a year. Move to simulating 3 weeks into a season against a team with a new QB. The amount of cash, time, and good data available... everyone already has box scores and all22 footage... could make t a hotbed for Deep Learning and Neural Networks. Maybe I'm wrong, but I assume the guys in the booth are allowed to use a computer?
But even this raises some questions. Let's say your model predicts 80/20 run from a given formation and set with certain tells. And your DC lines up to stop the run, and they throw over your head with no Safety protection for a back-breaking 86 yard TD -- because the OC noted your reaction to his tendencies at halftime.
ML wouldn't pick up on things like Bud Foster running bear against us in 2014 either... if there's no sample data from somebody else doing it first, it's highly unlikely a model could predict that as an effective counter... and if you had a model that somehow did, it's unlikely to know the risk associated with doing it.
Click to expand...

I think the true test would be to do all of ML learning work you noted and then let a computer come up with offensive and defensive game plans and call the plays against human coaches. It used to be impossible for a computer to beat a master in chess and now it's pretty regular. Given you 80/20 example it would be stupid for a coach to put all his eggs in the 80 basket in the second half. Sure you might lean toward the run, but the best coordinators are masters of breaking tendencies so I bet proper ML would take the game into account and drop that 80/20 knowing that it's highly likely that the OC is likely to break a tendency. Maybe display it as out of this formation they typically run the ball 80% of the time, but there is an 80% chance that they will break the tendency on this play.

One good thing about Stefanski is he wants the plays to look the same off the line. If everyone on an offense executes the defense has to play on their heals to make sure that it is a run or a pass before they commit. The Browns need to fix the mistakes and execute. That will be the biggest test for the new staff. 90% of NFL coaches are good at the X's and O's. It's the leading of men that has been an issue.
 
Upvote 0
exhawg said:
I think the true test would be to do all of ML learning work you noted and then let a computer come up with offensive and defensive game plans and call the plays against human coaches. It used to be impossible for a computer to beat a master in chess and now it's pretty regular. Given you 80/20 example it would be stupid for a coach to put all his eggs in the 80 basket in the second half. Sure you might lean toward the run, but the best coordinators are masters of breaking tendencies so I bet proper ML would take the game into account and drop that 80/20 knowing that it's highly likely that the OC is likely to break a tendency. Maybe display it as out of this formation they typically run the ball 80% of the time, but there is an 80% chance that they will break the tendency on this play.

One good thing about Stefanski is he wants the plays to look the same off the line. If everyone on an offense executes the defense has to play on their heals to make sure that it is a run or a pass before they commit. The Browns need to fix the mistakes and execute. That will be the biggest test for the new staff. 90% of NFL coaches are good at the X's and O's. It's the leading of men that has been an issue.
Click to expand...

The problem i see with that is Chess, Go, etc are all rigid rules. Go is fluid in playstyle and the follow-on possibilities are complex, but there's only 1 type of action.
There are rules in football, but their interpretation is subjective and random at best (Fiesta Bowl anyone?)
Could the computer win so predictably if its Rook had a fluidly changing skill level compared the opponent's Bishop?
Could the computer factor in difference between Rashan Gary and Joey Bosa?
How well can it capture Barkley being a shrinking violet in the 4th?
Remember those Chess algorithms were arrived at by simulating thousands of games over a few hours. That's not an approach that would work for football. There's about 2500 games in last 10 years. A lot more if you open up College tape, but how useful is Navy vs UCF to NFL ?

What id love to see is an AI that shocks OC and HC any time they start to call something cute and provide 3 Madden alternatives based on what its learned over last 10 years of games applied to the current game.
 
Upvote 0
You guys worry too much about the Browns when you should be computing how much beer you can drink during the games and how you're going to get the wife to pay for it, along with how many times she will bring you a fresh one .
 
Upvote 0
You must log in or register to reply here.
Back
Top