Poker-playing computer almost beats human pros

From here:

As the article explains, poker is a harder game for a computer to play than chess because so much depends on who you’re playing against. When the computer does win, it’s going to be big news.

I don’t think a computer could ever consistantly beat the best human players. Be on par with them and win a good amount, sure. I could even concieve a computer winning a World Series bracelet. There’s just too much uncertainty and luck involved, and so many numbers like pot odds and starting-hand percentages and the like have been ultra-anylized already. They’re best to stick with limit games.

I’m not familiar with Ali Eslami but I can name two dozen players off the top of my head who are better than Phil Laak. Phil is one of the most erratic players in the world, in terms of doing strange things and making odd moves, but I don’t believe anyone would rank him as “the best.”

But wouldn’t that play style be “the best” if you were playing against a purely logical computer program?

But, hell, what do I know, I often lose against hand held games!

In a semi-related story (about an admittedly simpler game, but still an impressive feat) from last week, programmers have “solved” the game of checkers, creating a program that is physically impossible for a human to beat. Non-believers can even play and fail miserably at the game online! :slight_smile:

Chess is a much deeper game than checkers. So far, computers have ‘solved’ all chess endings with 5 pieces and are currently working on 6 piece positions. This is done by creating endgame tablebases.
Once computers are able to handle all possible 32 piece positions, they can play perfect chess. Might take a while, though - in the meantime computer programs being able to look about 10 moves ahead is enough to beat top grandmasters.

(If anyone would like, I can give references for any of the above…)

Oh, I understand and don’t doubt you. I just thought it was an interesting, recent, and somewhat relevant story that I’d throw out there.

That is cool, last I heard checkers was very close, but not solved yet.

Tey mention Connect Four and Awari, wonder what is between those and checkers, and is there anything between checkers and chess.

I also wonder where Shogi places. Go is probaby the most computationaly (tho theoretically solvable) popular game.

Brian

I believe that this might be the same team of researchers. I do know that the checkers people were going to move onto to poker and were about to play a test against real players.

Looks like you’re correct. Both articles mention that the developers are from the University of Alberta.

More detailed article on Chinook:
http://www.spectrum.ieee.org/jul07/5379

Wikipedia article on solved games:
http://en.wikipedia.org/wiki/Solved_board_games

(Wiki classifies Checkers as “weakly solved”)

Brian

I’m good friends with one of the researchers on the U of A team. We used to play poker together when we were both making a living at it. He’s a great player, and a very smart guy. His group was working on poker ten years ago. It’s an extremely hard problem, because it’s a game of incomplete information. Unlike Chess, where all the information the computer needs is stored in the board positions, Poker has bluffing and unseen cards. The computer has to have ‘judgement’. It has to look at past betting patterns and make guesses as to what kind of hand the opponent might have. This is a completely different class of problem than that experienced by chess computers.

It’s also a potentially much more useful problem to solve. The applications of this type of analysis to military strategy, business, and any other decision-making under uncertain conditions are tremendous.

Plus, you can use your program to make money a poker if it gets good enough. That’s a good litmus test for the program, and the U of A group is trying to achieve that now.

Seems this stuff is spewed about a lot. What are there now, like 300 players that are ranked as the best in the world?

Maybe not the best, but the guy is pretty damned close as far as I can see. He’s able to win consistently despite his erratic behavior, and from what I can tell after watching him for many hours is that he is extremely talented when it comes to putting opponents on a hand.

And he is a far better ambassador for poker than the vast majority of bracelet winners as of late.

Yes, and this is an important distinction - they didn’t produce a complete tablebase of all possible positions. There are positions where the program doesn’t know the best move.

Similarly, in theory it might be possible to prove that Chess is a win for white with 1.e4. A computer could then play white perfectly by always opening with this move, and have no idea what to do in other positions with a white pawn on e2.

Part of what I meant by “strange things” was his table antics, which would have no effect on a computer.

Plus he stands as a shining beacon of hope for erratic poker nerds everywhere that one day they too might land an incredibly hot movie star girlfriend.

I don’t mean to suggest that Laak’s not a terrific player. Just not one who I would call “the best in the world.”

It’s not only what kind of hand the opponent may have (i.e., strictly playing the odds). AFAIK, this is almost always done with “user models”; poker is a good test-bed for this because there’re a small number of player types (tight, aggressive, chaotic). Based on the player’s actions, the computer assigns a (hopefully flexible) type classification that can be adjusted as more is learned about a particular opponent (i.e., more games are played).

I thought about pursuing this a while back as a side project (I’m a grad student in AI myself). Not enough time, though. Most intriguing to me was being able to have the computers collude in real-time throught their own socket connections (separate from the game server). Taking it a step further, how to identify when players (possibly computers) were doing exactly that so that the game was kept fair. One could become very wealthy doing so, although I was thinking about it before on-line gambling became a legal issue here in the states.

The link in the OP doesn’t say what kind of poker was being played. As I understand it, the two forms of poker that are easiest for computers are limit holdem and turbo no limit holdem sit-n-gos. For sit-n-gos, a model of the structure known as the Independent Chip Model comes damn close to being a Game-theoretically optimal solution. If the structure of the contest was similar to a sit-n-go, which I suspect it probably was, this result is less impressive than it may appear at first. Additionally their sample size is ridiculously small to even base speculation on. The fact is, in a normal ten man turbo tournament, the best player will probably only win about 15-20% of the time, if that. (The numbers are just my estimates, but my point is that short-term variance is the dominant factor in any individual contest.)

When a computer can consistently beat, over thousands and thousands of hands, great human poker players, I’ll be really impressed. And I hope it never happens, cause I don’t want to have to get a real job. :slight_smile:

And fwiw, there have been numerous bots online playing mid-limit cash game Limit Holdem for several years that have been making money. No time to dig up a cite now, but the twoplustwo.com forums have some good information on known bots, if you’re interested.