Move 12 discussion: Everyone vs stockfish

Originally Posted by Brian-E

I don't feel qualified to comment on your remarks about "surprise" or the very philosophical discussion about prederminedness in the rest of your post, and maybe someone else will take that up instead, but I can't let the above quoted sentence pass without sticking my oar in again.

All practical experience of playing against programs has led to the opposite conclusion to what you write above. If there is one area where the best human players still stand a chance against computer programs, it is the technical type of position where calculation of precise move sequences is largely irrelevant. And, very roughly and with plenty of exceptions, the more pieces that are exchanged the more "technical" the play becomes and the less the program will be helped by its ability to examine millions of positions.

First, I do not intentionally philosophize and second I respect your opinion. You are experienced and I am learning - can you provide exceptions to the rule as stated within your second paragraph if you are aware of any? If none exist, then I can take this as proof that what you state is correct.

Some considerations I apply are from the following games: blackjack, when played well requires card tracking usually relative to some point count basis; bridge/poker/gin require card location skills as well as an ability to read intent;GO is like playing with ripples on a
still pond and Snooker/Pool requires an ability to place and land bank shots and collisions
by thinking ahead and wiping up the table appropriately. When the above are played in a
hostile environment and out of one's `comfort zone` you play your best game every time.

[Brian-E]

Well, I did say there were "plenty of exceptions".

However: it is time for me to get off my horse, I think.

You call me "experienced", but that chess experience does not extend to playing against chess programs in recent decades.

I reacted to your previous posting based on my experience of playing chess programs from 30 years ago and more. At that time all chess programs were hopeless at all endgames and technical middlegames. A bit of reading about strategy against chess computers now has shown me that this is no longer true.

I now rather agree with your opinion that, very generally, reduced material on the board helps the machine benefit more due to its increased search depth. The positional understanding which the programs have in being able to evaluate final nodes is so vastly better than it was a few decades ago that what I stated in my previous post can be called nonsense.

Your comparisons with other games are interesting. I guess chess is more complex than the card games you mention, but less complex than Go. I imagine a game like Blackjack, while obviously not a complete information game which chess and Go are, is simple enough for a machine to be programmed perfectly so that its winning chances are maximised. Perhaps poker is similar, though the characteristically human bluffing element in that game might present problems in the programming, I don't know. Bridge is a bit weird because you play in partnerships and, at the bidding stage at least, the information exchange is a bit subjective and ill-defined. That's my impression anyway.

[gd_barnes]

Even today, chess programs are still better at tactics and not quite as good at long term strategy. Getting down to fewer pieces so that long term strategy is required should benefit us. Of course we'd be at a disadvantage in a "solved" 6-piece end game but anything likely with ~10 pieces on the board should be easier for us to compete with a machine than with ~20 pieces on the board. Of course that's a very general statement.

[VBCurtis]

Originally Posted by Brian-E

Your comparisons with other games are interesting. I guess chess is more complex than the card games you mention, but less complex than Go. I imagine a game like Blackjack, while obviously not a complete information game which chess and Go are, is simple enough for a machine to be programmed perfectly so that its winning chances are maximised. Perhaps poker is similar, though the characteristically human bluffing element in that game might present problems in the programming, I don't know.

I play mid-stakes poker as a hobby, didn't think mersenneforum would ever overlap poker! Poker software is as good as solved for one variety (so far): 1 on 1 fixed-bet ("limit") holdem. If bet-size can be varied (no-limit, TV poker), the search space of possible actions is too large to be yet solved even approximately- though the best no-limit computers are the equal of the best no-limit professionals when played 1-on-1.

Computers are not nearly expert at regular poker, because it is not a 2-player game. The multiple-villain aspect is too much complexity (and too little game-theory theory) for a computer, even if betting sizes are fixed. Skilled humans do not fear a 'bot' player in a multi-way poker game.

Bluffing isn't a human 'thing', really- there are optimal bluff percentages that can be calculated for almost any situation, and a player who bluffs less than optimally is less profitable than one who does it correctly; so a computer may calculate as if the villain bluffs optimally, and should win more if the villain does not.

[Brian-E]

Originally Posted by VBCurtis

I play mid-stakes poker as a hobby, didn't think mersenneforum would ever overlap poker! Poker software is as good as solved for one variety (so far): 1 on 1 fixed-bet ("limit") holdem. If bet-size can be varied (no-limit, TV poker), the search space of possible actions is too large to be yet solved even approximately- though the best no-limit computers are the equal of the best no-limit professionals when played 1-on-1.

Computers are not nearly expert at regular poker, because it is not a 2-player game. The multiple-villain aspect is too much complexity (and too little game-theory theory) for a computer, even if betting sizes are fixed. Skilled humans do not fear a 'bot' player in a multi-way poker game.

Bluffing isn't a human 'thing', really- there are optimal bluff percentages that can be calculated for almost any situation, and a player who bluffs less than optimally is less profitable than one who does it correctly; so a computer may calculate as if the villain bluffs optimally, and should win more if the villain does not.

Thanks for the insights from someone who knows the ins and outs of poker. It's interesting that the multi-opponent nature of the regular versions of poker make the game still impossible, currently, for programs to play well, and also that bluffing can be coded optimally.

I've never played poker. Did once read a fascinating book by David Levy (a chess International Master and games programming expert) on programming games, which included a chapter on poker. I think he concentrated on what he called "stud poker". It was written back in the 1980s, I think, but I assume the fundamental principles of games playing algorithms haven't changed since then even though their implementation has been fine-tuned and the hardware available has improved out of all recognition in that time.

Brian, here is another connection you may not have been aware of regarding poker:
https://en.wikipedia.org/wiki/Herbert_Yardley

Reasoning about games is a fun way to do hard work because it takes imagination and creativity. Abstracting those aspects that apply to complete/incomplete information and disinformation as well as P vs NP categorization can be useful within the appropriate context. Chess is a great game that Mr. Barca would have appreciated.

[Brian-E]

Originally Posted by jwaltos

Brian, here is another connection you may not have been aware of regarding poker:
https://en.wikipedia.org/wiki/Herbert_Yardley

Reasoning about games is a fun way to do hard work because it takes imagination and creativity. Abstracting those aspects that apply to complete/incomplete information and disinformation as well as P vs NP categorization can be useful within the appropriate context. Chess is a great game that Mr. Barca would have appreciated.

Thanks. Not only was I not aware of this connection between cryptoanalysis and poker theory, I'm ashamed to say hadn't even heard of Herbert Yardley. I have now.

[LaurV]

Same here! Nice reading about that man. Thanks.

[VBCurtis]

I have his 1957 poker book, had no idea he had a math background. Indeed, thanks!