Game Theorists Crack Poker | Hacker News Reader

Game Theorists Crack Poker

scientificamerican.com

45 pointsjonbaer29 comments

https://news.ycombinator.com/item?id=8858280

(no relation to submission.)

I prefer this title (Game Theorists Crack Poker), article (sciam), and our thread here, and wouldn't have clicked the other one if I had seen it. (and didn't read it now that I did open it:). the ieee one is one block of text without subsection titles, and starts with the 'All your poker chips may soon belong to the computers' which made me want to close it.)

I'm missing something:

    "But Bowling and his colleagues have demonstrated that their algorithm always wins in the long run."

So how does it fare when its opponent is the same algorithm ?

Yet another article[1] stated that if playing it's self the bot with the better hands/chance 'won' in the long run.

[1] http://motherboard.vice.com/en_au/read/this-robot-is-the-bes...

In another article I read they said it would "never do worse than even", and I would assume that the even case is when playing itself (or we have a logical contradiction). A lot of game theory winning strategies work that way; when you play another perfect strategy you draw.

Losses equal to rake percentage

- We are talking about HULHE only. If the game were no limit and 9 players the numbers presented in the article regarding possible states grow, by a lot. AFAIK there is no bot that could break even against a table of decent players playing NL. "Crack Poker" seems an overstatement to me.

- The articles states the bot always wins in the long run but most correctly this bot can break even at worst scenario. I mean, if you face it against itself they can't both win in the long run.

> AFAIK there is no bot that could break even against a table of decent players playing NL.

While it's true that NL ring is far from being solved in the same mathematical sense the HULHE has been, there exist very good bots that can certainly beat a table of decent players, and perhaps even world-class players. There were a number of scandals with ring NL bots on UltimateBet and other online sites, with huge samples showing solid win rates at midstakes.

There is also pokersnowie:

http://www.pokersnowie.com/

and here is a huge thread about it from when it first arrived on the scene: http://www.deucescracked.com/forums/131-Poker-Theory/topics/...

This is Oskari Tammelin of Buzz tracker fame (audio software): http://www.jeskola.net

Links to some counterfactual regret minimization code is on his site here: http://www.jeskola.net/cfr/

I'm disappointed, was hoping for more -- this is "cracked" for heads up limit only, and even then I have my doubts. But would like to know more or play against it or see some code or something...

You shouldn't have your doubts. The pros stopped playing even inferior bots a long time ago in HULHE. We're talking 5 years ago.

I have very little doubt that this bot cannot be beat.

Sounds like they've developed a very good bot, but I don't believe it's possible to ever completely "solve" an imperfect information game.

Solved here just means there is no profitable strategy that can be played against it, like playing rock, paper, scissors by choosing randomly. It doesn't mean it maximally exploits other strategies, like always playing paper against someone who always plays rock.

Kuhn Poker [1] is a game of imperfect information, which was solved by its creator in 1950. The only fundamental difference between Kuhn Poker and LHE, the game discussed in the article, is that regular LHE has much more cards involved, making the game more complex.

Your statement is technically correct because of the use of the word "completely", but wrong in a practical sense. The algorithm used to solve these games is iterative, so the solution only ever approaches a complete solution, but this is only meaningful if you want to argue that a 0.0000001BB difference in win-rate per hand is really meaningful.

[1] http://en.wikipedia.org/wiki/Kuhn_poker

It is proven that for heads-up poker there is a Nash equilibrium, so it is indeed possible.

Do you have a good resource I can read about that?

HULHE has been known to have very strong bots[1] for many years, mainly thanks to the University of Albert and the research they have done.

There is the Computer Poker Competition that is held every year too[2].

Everyone I know has been staying away from HULHE for awhile, preferring either HUPLO or HUNLHE instead.

1: https://en.wikipedia.org/wiki/Polaris_(poker_bot) 2: www.computerpokercompetition.org

HULHE SNG's especially.

Short stack NLHE too.

http://poker.srv.ualberta.ca/

I'm very curious to read the paper if anyone can find it posted online.

https://pdf.yt/d/qv-O9AwQuV1Kjb04

The article mentions randomness, incomplete information, and how often to bluff, but I don't see anything about tells. Unlike humans, computers are excellent at hiding their bluffs and their excitement at holding a royal strait. Hiding tells doesn't matter in checkers or chess, but it's essential in poker.

Tells, in the sense of ticks and such, are mainly ignored by good poker players. The board situation is where the info is at. A good player won't go thinking "its sweety! he could bluff", but instead "his range based on previous hands on this position is X so I estimate a chance Y for him to bluff".

Computers do have a great advantage by ignoring the whole psycho layer and focus on data only.

That's not true. I played and studied poker for a many years, and I will say for certain that physical cues is a huge part of live games, and definitely not ignored by good players.

It's not ignored, but it's not a 'huge part' of the game.

It's one of thousands of factors that weigh in, and in my mind, one of the least important factors.

I've played several million hands of poker fwiw.

The usefulness of physical tells has been wildly exaggerated to make poker in TV/movies more interesting. They can be useful, but betting patterns are far more important.

Not really. If the goal is to maximize profit then tells allow you to make (game theory) suboptimal plays to maximally exploit your opponent. Something the bot would never do. A real player could thus win more money off players with tells, or weaker players in general.

The bot just plays an unexploitable strategy where at best you break even against it, given you never make a mistake.

In online/computer poker, there are a lot less tells. In fact the only tells are bet sizing (which is static in limit hold 'em, which is the main reason it is easier to solve), and the timing of bets/checks/folds.

Tells in online poker amount to bet sizing and bet timing tells. Both of which good humans find easy to hide.

Reply on news.ycombinator.com