[BostonCard]

This was discussed in another thread, but it may not be so far out.

https://fivethirtyeight.com/features/the...for-poker/

The post discussed Texas Hold 'em, another situation when two opponents are working with imperfect information.  Carnegie Mellon developed a Poker Playing algorithm.  Turns out it beat the humans:

https://fivethirtyeight.com/features/a-c...-at-poker/

Here was an interesting passage...

There is one obvious advantage that computers hold over humans in this game: randomization. In certain situations, it’s optimal to play a “mixed strategy,” randomly choosing from a set of options. If you hold a very strong hand, for example, it may be optimal to mix between betting various healthy amounts, hoping to get paid off, and checking, hoping for a lucrative check-raise. This keeps the opponent guessing and dodges effective counter-strategies. Humans, however, even if they know something like that might be optimal, are no good at picking randomly. They also can’t effectively deploy a large menu of bet sizes without getting hopelessly lost and tangled in the thicket of their own best-laid plans.

As I have posited, it sometimes makes sense, in certain football situations to employ a similar "mixed strategy", for example on second and short, mixing between low risk options that have a high likelihood of getting you a first down and controlling the clock, and higher risk/higher reward strategies like play-action passing that give you a chance for scoring (but also increase the risk of either an incompletion or a turnover). 

BC

[BostonCard]

This was discussed in another thread, but it may not be so far out.

https://fivethirtyeight.com/features/the...for-poker/

The post discussed Texas Hold 'em, another situation when two opponents are working with imperfect information.  Carnegie Mellon developed a Poker Playing algorithm.  Turns out it beat the humans:

https://fivethirtyeight.com/features/a-c...-at-poker/

Here was an interesting passage...

There is one obvious advantage that computers hold over humans in this game: randomization. In certain situations, it’s optimal to play a “mixed strategy,” randomly choosing from a set of options. If you hold a very strong hand, for example, it may be optimal to mix between betting various healthy amounts, hoping to get paid off, and checking, hoping for a lucrative check-raise. This keeps the opponent guessing and dodges effective counter-strategies. Humans, however, even if they know something like that might be optimal, are no good at picking randomly. They also can’t effectively deploy a large menu of bet sizes without getting hopelessly lost and tangled in the thicket of their own best-laid plans.

As I have posited, it sometimes makes sense, in certain football situations to employ a similar "mixed strategy", for example on second and short, mixing between low risk options that have a high likelihood of getting you a first down and controlling the clock, and higher risk/higher reward strategies like play-action passing that give you a chance for scoring (but also increase the risk of either an incompletion or a turnover). 

BC

[TRD]

It is going to happen.  I don't see why Stanford wouldn't be the first to use big data and analytics for play calling.

[TRD]

It is going to happen.  I don't see why Stanford wouldn't be the first to use big data and analytics for play calling.

[CornFed]

Man, I don't know.  It seems to me that AI would put a real crimp in our second-guessing after games.  We'd only be able to get into questions of coding and program default language.  EEEEEwwwwww!!!

[CornFed]

Man, I don't know.  It seems to me that AI would put a real crimp in our second-guessing after games.  We'd only be able to get into questions of coding and program default language.  EEEEEwwwwww!!!

[Extra Point]

Man, I don't know.  It seems to me that AI would put a real crimp in our second-guessing after games.  We'd only be able to get into questions of coding and program default language.  EEEEEwwwwww!!!

Couldn't we just say that AI is stubborn!

[Extra Point]

Man, I don't know.  It seems to me that AI would put a real crimp in our second-guessing after games.  We'd only be able to get into questions of coding and program default language.  EEEEEwwwwww!!!

Couldn't we just say that AI is stubborn!

[GoodGrief]

It's going to feel really hollow to complain about the programming decisions instead of Shaw's conservativeness.

[GoodGrief]

It's going to feel really hollow to complain about the programming decisions instead of Shaw's conservativeness.

[fullmetal]

Just complain about the doofuses who wrote the algorithm.  Algorithms and AI machines are only as good as their creators.

[fullmetal]

Just complain about the doofuses who wrote the algorithm.  Algorithms and AI machines are only as good as their creators.

[BostonCard]

Just complain about the doofuses who wrote the algorithm.  Algorithms and AI machines are only as good as their creators.

If only we had a CS guy on this board...

My understanding is that for machine learning algorithms, the biggest source of bias is the data that is fed in.  For example, for a lot of the algorithms that recognize people in pictures, a lot of the training sets featured mostly Whites, which meant that the algorithms had trouble recognizing people of color.

BC

[BostonCard]

Just complain about the doofuses who wrote the algorithm.  Algorithms and AI machines are only as good as their creators.

If only we had a CS guy on this board...

My understanding is that for machine learning algorithms, the biggest source of bias is the data that is fed in.  For example, for a lot of the algorithms that recognize people in pictures, a lot of the training sets featured mostly Whites, which meant that the algorithms had trouble recognizing people of color.

BC

[murky]

[quote author=fullmetal link=topic=16522.msg187853#msg187853 date=1485979899]
Just complain about the doofuses who wrote the algorithm.  Algorithms and AI machines are only as good as their creators.

If only we had a CS guy on this board...

My understanding is that for machine learning algorithms, the biggest source of bias is the data that is fed in.  For example, for a lot of the algorithms that recognize people in pictures, a lot of the training sets featured mostly Whites, which meant that the algorithms had trouble recognizing people of color.

BC
[/quote]

Sound like how people's brains work.

[murky]

[quote author=fullmetal link=topic=16522.msg187853#msg187853 date=1485979899]
Just complain about the doofuses who wrote the algorithm.  Algorithms and AI machines are only as good as their creators.

If only we had a CS guy on this board...

My understanding is that for machine learning algorithms, the biggest source of bias is the data that is fed in.  For example, for a lot of the algorithms that recognize people in pictures, a lot of the training sets featured mostly Whites, which meant that the algorithms had trouble recognizing people of color.

BC
[/quote]

Sound like how people's brains work.

[Roberton3]

As I have posited, it sometimes makes sense, in certain football situations to employ a similar "mixed strategy", for example on second and short, mixing between low risk options that have a high likelihood of getting you a first down and controlling the clock, and higher risk/higher reward strategies like play-action passing that give you a chance for scoring (but also increase the risk of either an incompletion or a turnover). 

You're greatly understating your point.  Almost every play call in football is part of a mixed strategy.  The only exception is a situation where, even if the other team knows 100% for certain the exact play you're going to call, you'd still call that play.  Those situations are pretty rare.

[Roberton3]

As I have posited, it sometimes makes sense, in certain football situations to employ a similar "mixed strategy", for example on second and short, mixing between low risk options that have a high likelihood of getting you a first down and controlling the clock, and higher risk/higher reward strategies like play-action passing that give you a chance for scoring (but also increase the risk of either an incompletion or a turnover). 

You're greatly understating your point.  Almost every play call in football is part of a mixed strategy.  The only exception is a situation where, even if the other team knows 100% for certain the exact play you're going to call, you'd still call that play.  Those situations are pretty rare.

[fullmetal]

4th quarter, 2009.  Power, power, power  8)

[fullmetal]

4th quarter, 2009.  Power, power, power  8)

[gocard14]

[quote author=Boston Card link=topic=16522.msg187782#msg187782 date=1485906921]
As I have posited, it sometimes makes sense, in certain football situations to employ a similar "mixed strategy", for example on second and short, mixing between low risk options that have a high likelihood of getting you a first down and controlling the clock, and higher risk/higher reward strategies like play-action passing that give you a chance for scoring (but also increase the risk of either an incompletion or a turnover). 

You're greatly understating your point.  Almost every play call in football is part of a mixed strategy.  The only exception is a situation where, even if the other team knows 100% for certain the exact play you're going to call, you'd still call that play.  Those situations are pretty rare.
[/quote]

Of course, where a play calling AI would fall short is in its ability to assimilate information as the game went on. The hold-em AI is likely successful because in addition to randomizing strategies and being "un-readable," it can act on the probabilities for success that it knows based on the hand it has, the pot, and its analysis of the opposing players moves. The probability of converting a run up the middle on 3rd & 2 is highly variable on a game to game and situation to situation basis. A major portion of offensive design is creating favorable matchups for your players, and then exploiting said matchups.  The computer wouldn't know how they are responding to your personnel groupings, or smaller things like which O-lineman is consistently winning his matchup at the LOS.  The only way they could assimilate this data is either to have an army of people doing in-game data entry or hooking it up to an all-11 view of the game with incredibly powerful image processing.  While an AI playcaller with that knowledge would likely be superior, at that point it would sort of depersonalize football to acting out the commands of an all-knowing computer general, like the Star Trek episode "A taste of Armageddon."

A much more plausible use of computers in would come down to simple randomization. A coach could specify a pre-set number of plays to be considered in various scenarios and the computer could choose or suggest a play from within that subset.

As an aside, to me a computer "winning" at poker is a silly metric. By its very nature, poker involves reading the physiology and psychology of a human opponent. Playing against a computer that has no physiology that can be read is playing a different game.  This is also true for Jeopardy, as a computer can have a near infinite database of trivia facts, with the only limiting factor being the quality of its natural language processing algorithms. The fact that it is playing a "different game," so to speak, is precisely why I believe that human/AI interaction is going to be so important in the future. For example, in Radiology there are image processing algorithms that can detect abnormalities on a digital mammogram. I'm sure that in the future these technologies, if not already, will be more accurate then a human radiologist.  More importantly, however, is that these technologies tend to fail in different ways then a human radiologist would, and so the combination of the two can be better then either individually. 

[gocard14]

[quote author=Boston Card link=topic=16522.msg187782#msg187782 date=1485906921]
As I have posited, it sometimes makes sense, in certain football situations to employ a similar "mixed strategy", for example on second and short, mixing between low risk options that have a high likelihood of getting you a first down and controlling the clock, and higher risk/higher reward strategies like play-action passing that give you a chance for scoring (but also increase the risk of either an incompletion or a turnover). 

You're greatly understating your point.  Almost every play call in football is part of a mixed strategy.  The only exception is a situation where, even if the other team knows 100% for certain the exact play you're going to call, you'd still call that play.  Those situations are pretty rare.
[/quote]

Of course, where a play calling AI would fall short is in its ability to assimilate information as the game went on. The hold-em AI is likely successful because in addition to randomizing strategies and being "un-readable," it can act on the probabilities for success that it knows based on the hand it has, the pot, and its analysis of the opposing players moves. The probability of converting a run up the middle on 3rd & 2 is highly variable on a game to game and situation to situation basis. A major portion of offensive design is creating favorable matchups for your players, and then exploiting said matchups.  The computer wouldn't know how they are responding to your personnel groupings, or smaller things like which O-lineman is consistently winning his matchup at the LOS.  The only way they could assimilate this data is either to have an army of people doing in-game data entry or hooking it up to an all-11 view of the game with incredibly powerful image processing.  While an AI playcaller with that knowledge would likely be superior, at that point it would sort of depersonalize football to acting out the commands of an all-knowing computer general, like the Star Trek episode "A taste of Armageddon."

A much more plausible use of computers in would come down to simple randomization. A coach could specify a pre-set number of plays to be considered in various scenarios and the computer could choose or suggest a play from within that subset.

As an aside, to me a computer "winning" at poker is a silly metric. By its very nature, poker involves reading the physiology and psychology of a human opponent. Playing against a computer that has no physiology that can be read is playing a different game.  This is also true for Jeopardy, as a computer can have a near infinite database of trivia facts, with the only limiting factor being the quality of its natural language processing algorithms. The fact that it is playing a "different game," so to speak, is precisely why I believe that human/AI interaction is going to be so important in the future. For example, in Radiology there are image processing algorithms that can detect abnormalities on a digital mammogram. I'm sure that in the future these technologies, if not already, will be more accurate then a human radiologist.  More importantly, however, is that these technologies tend to fail in different ways then a human radiologist would, and so the combination of the two can be better then either individually. 

[BostonCard]

I believe the image processing needed to detect alignment from an all-22 view may be in my pocket; recognizing faces is harder than recognizing where players are.

And a computer would be able to learn as it went along.  Third and two gives you one yard?  That can get incorporated into the next decision on third and two.

BC

[BostonCard]

I believe the image processing needed to detect alignment from an all-22 view may be in my pocket; recognizing faces is harder than recognizing where players are.

And a computer would be able to learn as it went along.  Third and two gives you one yard?  That can get incorporated into the next decision on third and two.

BC

[Hurlburt88]

If we fast forward several years later see debate about AI as "performance enhancing".  I would love be interested to see how that Plays out

[Hurlburt88]

If we fast forward several years later see debate about AI as "performance enhancing".  I would love be interested to see how that Plays out