Chess vs. Poker vs. Computer

[diddle]

[ QUOTE ]
The experts would simply play to make the computer think it was playing a fish and then exploit it.

Fish play bad all the time, not that "simple" to fake inexpensively. The computer can treat any borderline players as experts and use game theory, while cleaning out the fish with special modules.

[/ QUOTE ]

Incredible. Someone in this thread who is not a retard on the subject.

[Knockwurst]

[ QUOTE ]
My statement assumes both chess and poker computers are using the strategy that has the best chance against experts.

In the poker case that means it will try too many bluffs against a sucker.

In the chess case it means that it will assume the bad player will find the best move against it.

Stuff like that.

If you want to postulate a computer that deviates from optimum strategy when it sees bad play you run into the problem of counter strategies designed to make that occur.

A counter counter strategy would be nightmarish to program, I think. Practically, the best one could hope for would be a program that plays optimally. One that would not crush bad players to the degree an expert human would.

[/ QUOTE ]

I think an expert chess program and poker program would have many more differences than similarities.

The chess program goes through simultaneous calculations of different analytical trees, giving an evaluation at a certain depth of the tree based on checkmate, material gain or positional considerations with a weight assigned to each.

It is true the program will make the best move based on an evaluation of the position after what it thinks the best response to its move is. However, it is also simultaneously calculating the best way to exploit less than the best response made by an opponent. And there is every reason to believe that a computer is at least equal to a human in efficiently exploiting a bad move.

Now you and others seem to suggest that the human would have the better ability to play incorrectly in order to more efficiently exploit the weaker player. (The example you used was a bonus for winning in under 45 moves. A bad example by the way because the average chess game is about 40-45 moves and a computer could beat a weaker player in less moves.)

But this is exactly how chess and poker are different. As you and Ed Miller pointed out in NLHTP, poker and in particular No Limit is a battle of the mistakes. The better player should play less than optimally in order to exploit or take advantage of the weaker player's bigger and more costly mistakes. (Sorry if this is only a very rough characterization of your theory, I don't have the book in front of me). In chess, it isn't necessary to play suboptimally in order to exploit your opponent's weak moves. And playing suboptimally will not more efficiently exploit those moves.

A perfect example is Paul Morphy who has some of the most spectacular minatures where he beat opponents in under 25 moves. But examining most of those games, you will see that he played optimal developing moves for the most part and only once his opponent failed to make the optimal response, he would counter with the optimal way to exploit those mistakes. There is no reason to believe that a computer cannot do the same.

In short, neither the human chess player nor the computer need deviate from optimal strategy to most efficiently exploit the weaker move or player. Whereas I believe, as you have pointed out, it is necessary in poker to play suboptimally in order to more efficiently exploit the weaker player in the game.

[djames]

[ QUOTE ]
My statement assumes both chess and poker computers are using the strategy that has the best chance against experts.

...

In the chess case it means that it will assume the bad player will find the best move against it.


[/ QUOTE ]

This is not how chess programs work. They choose the best move at every decision point contemplating "all" possible future moves of their opponent. These decision trees are weighted, and the highest score wins. Then the next decision point comes and a new tree is constructed. The opponent need not choose the "best" move for the computer's previous choice to still have been the best. Perhaps there would have been a better choice against the particular move the opponent will choose, but that's not the point. The computer picks the best possible move at each time. If the opponent then blunders, the position presented at the next decision point may alter the computer's planned action (based on its prior move) since the blunder opened an even more exploitable line.

Chess computers can *crush* opponents, weak & GM alike, way better than the top GMs can. The top rated chess engine rates over 200 Elo points more than the highest rated human. Fritz is over 3,000 while the highest rated human was near 2,850 for a short time.

Perhaps you don't intend to imply that a player rated a lowly 1,000 will be tougher to beat by Fritz than a player rated 2,500 .. but that's what you've implied.

[StregaChess]

[ QUOTE ]

If you want to postulate a computer that deviates from optimum strategy when it sees bad play you run into the problem of counter strategies designed to make that occur.

A counter counter strategy would be nightmarish to program, I think. Practically, the best one could hope for would be a program that plays optimally. One that would not crush bad players to the degree an expert human would.

[/ QUOTE ]

Total absurd bull-feathers. There is a misunderstanding here of how computers play chess and what their accomplishments have been. It’s not just a brute force issue. Mr. Sklansky and others here may have forgotten more about poker than I shall ever hope to learn, however I am also sensing a bit of false pride in that my “skill set” can’t be accomplished by a computer because of “X”. Do not forget that the computer does not play chess, it does not play poker, it only implements a game plan that a human being put into motion in writing the code and rules of logic. I’ve read comments like “heads up” yes.. but in Ring games and tourney’s no….. I don’t think so, I would suggest the opposite may be true.
Poker is a game of incomplete information, however let’s create a scenario where a top player sits down for a 2 year 24/7 session continued session, the end result is that a program could be developed to own that player or any player. The fallacy in this logic is to think of poker as only a game of incomplete information, view it instead as a game where bits and pieces of information are constantly leaking out over time. The longer the playing session the more information that’s gathered. A great player can change what he’s going to do on the next hand but he can’t change what he’s done in the past 300. And those 300 hands will yield information that can be used to optimize decisions going forward. Fuzzy Logic could be used as part of the basis to develop these actions and in the end, if enough time and resources are feed into it, the top poker players would be crushed…..

[djames]

I thought it was common knowledge that the only reason computers can't dominate multi-handed poker games is due to computing power. The search trees in heads up limit hold 'em are still gigantic, however programs exist that crush most players. The UC Alberta site has such programs and have posted results against a player whom they state is "world class." If you read their research, they state that computing power is the bottleneck which prohibits multi-game world class poker programs. Nothing more.

[marv]

[ QUOTE ]
I thought it was common knowledge that the only reason computers can't dominate multi-handed poker games is due to computing power. The search trees in heads up limit hold 'em are still gigantic, however programs exist that crush most players. The UC Alberta site has such programs and have posted results against a player whom they state is "world class." If you read their research, they state that computing power is the bottleneck which prohibits multi-game world class poker programs. Nothing more.

[/ QUOTE ]

This is simply untrue.

The state of the art techinques for HU don't apply to multiplayer games since multiplayer equilibirium finding can't be expressed as a linear programme. You can find near-optimal strategies for 2 and 3 player limit using other solution techniques (which are slower than solving an LP) but not for 4 player and no-way-in-hell for 6 players or more.

People vastly underestimate how hard it is to write a bot which can construct a model of its opponents play accurately enough to exploit it (except for really poor opponents). Saying "its easy, just use pokertracker ... fuzzy logic ... blah blah" have cleary never tried, nor are familiar with the published work on the subject. Even mediocre humans are currently far better at figuring out how to exploit their opponent than the state of the art modelling bots.

A good game theory bot built using current techniques played against a state-of-the-art opponent modeller will crush the opponent-modeller, even over hundreds of thousands of hands. The UofA learnt this the hard way after their bot Poki-X (a combination game-theory/opponent-modeller bot) was defeated in a very public competition by another bot, and then by a well known pro, so in the recent AAAI pokerbot competition they just mixed two game-theory bots rather than even try to exploit its opponent. And they won, convincingly.

Unlike chess/backgammon, game theory (which is just minimax for complete information games) is not the full answer to how to write a pokerbot. The reason? Even a perfect game-theory pokerbot, though unbeatable in the long term will make very little against a reasonable competant player. Against anyone even close to expert, the bot will have essentially no edge whatsoever. This is the difference between chess and poker.

More mythbusting: HU NL Holdem is not actually that much harder than HU limit. Most HU NL games don't involve may different raise amounts, and usually have far fewer raises than HU limit. The tree is bigger, and you have to solve a different tree for each depth, but it's not hard to write game-theory NL bots which are actually quite sound (far better than the UofA's NL bots), but they are definitely not as good as their limit counterparts.

Marv

[FortunaMaximus]

[ QUOTE ]
I thought it was common knowledge that the only reason computers can't dominate multi-handed poker games is due to computing power. The search trees in heads up limit hold 'em are still gigantic, however programs exist that crush most players. The UC Alberta site has such programs and have posted results against a player whom they state is "world class." If you read their research, they state that computing power is the bottleneck which prohibits multi-game world class poker programs. Nothing more.

[/ QUOTE ]

Nope. That's fallacious. They simply can't come up with the proper algorithms or logic system for it to work accurately over the long-term. Thus far, they've been unable to introduce the right mix of unpredictability.

[George Rice]

[ QUOTE ]
Fish play bad all the time, not that "simple" to fake inexpensively. The computer can treat any borderline players as experts and use game theory, while cleaning out the fish with special modules.

[/ QUOTE ]

What I don't understand is how the computer will know it's playing a fish, borderline player or expert. Are you talking about programming a bot on an online site? Or are you talking about a stand-alone machine that the players will know they're playing? If the latter, I don't think the computer has a chance against the expert, not in the near future anyway. And the expert will win more off the bad players.

And not just for David's reasons. You would need a world class player who can also disect what he does to help program the machine. Even then, there are so many variables I doubt it can be done. For one thing, a computer can't read tells. So an expert will win more off the fish than the computer.

[djames]

[ QUOTE ]
A good game theory bot built using current techniques played against a state-of-the-art opponent modeller will crush the opponent-modeller, even over hundreds of thousands of hands. The UofA learnt this the hard way after their bot Poki-X (a combination game-theory/opponent-modeller bot) was defeated in a very public competition by another bot, and then by a well known pro, so in the recent AAAI pokerbot competition they just mixed two game-theory bots rather than even try to exploit its opponent. And they won, convincingly.

[/ QUOTE ]

Not sure what you're (mis)reading, but according to this report, UofA won at AAAI '06, they didn't lose. (Edit: Ah, I reread to see you reference Poki, which is a model that's a few years old now.)

Both your post, and the following post are claiming facts without evidence. I cited my evidence. Please cite yours to refute what I've said. Thanks!

UofA main poker page.

AAAI 2006 results page, showing UofA's Hyperborean CRUSHING the opponents.

UofA and AAAI on youtube.

[marv]

[ QUOTE ]
[ QUOTE ]
A good game theory bot built using current techniques played against a state-of-the-art opponent modeller will crush the opponent-modeller, even over hundreds of thousands of hands. The UofA learnt this the hard way after their bot Poki-X (a combination game-theory/opponent-modeller bot) was defeated in a very public competition by another bot, and then by a well known pro, so in the recent AAAI pokerbot competition they just mixed two game-theory bots rather than even try to exploit its opponent. And they won, convincingly.

[/ QUOTE ]

Not sure what you're (mis)reading, but according to this report, UofA won at AAAI '06, they didn't lose. (Edit: Ah, I reread to see you reference Poki, which is a model that's a few years old now.)

Both your post, and the following post are claiming facts without evidence. I cited my evidence. Please cite yours to refute what I've said. Thanks!

UofA main poker page.

AAAI 2006 results page, showing UofA's Hyperborean CRUSHING the opponents.

UofA and AAAI on youtube.

[/ QUOTE ]

[I'm *very* familiar with the UofA's work, and I'm not misreading anything - I think you may just have misread my post (I agree Hyperborean won). Search for marv at the PA forums if you're under the impression I'm just some random unkown guy.]

Poki-X was their state of the art last summer. It actually has nothing to do with their 'Poki' engine which is quite old (as you said). The name Poki-X was chosen for continuity/marketing reasons. This not the same as Hyperborean (the bot which won the AAAI competition this summer), which doesn't have an opponent modelling engine at all, just a rule to choose between the two game-theory strategies (search the PA forums for Darse Billings - he's the UofA's main brain on this subject for more info.)

Marv