von Neumann’s poker-playing machine

by on May 6, 2006 at 7:26 am in Science | Permalink

Here is Tim Harford’s article on poker and game theory, from today’s Financial Times.  Excerpt:

…he admits that it is only a matter of time before anyone will be able
to download a free poker robot that will outplay the world champion. At
that point, people may not care to risk money online against
unidentified opponents.

Alex Ambroz May 6, 2006 at 10:12 am

This actually already exists. I had a friend in college who would play 4 screens at once against mediocre players – made a good bit of money at it too. We got to talking about this exact issue and he stated that it was already in existance and was pretty simple actually (don’t know much about poker, but I assume it’s just a probability engine tied to an executable). The major websites put a lot of effort into detecting whether or not programs are actually sitting behind avatars.

I think what’s even more fascinating about it is that eventually, you might get to a point where we’ll just have the programs playing against each other.

Gordon Mohr May 6, 2006 at 5:04 pm

I suspect there are already robots that are good enough to earn a positive return in online play. Though online card rooms claim to fight robots, their incentive is chiefly to appear vigilant against robots — not to actually be vigilant.

A good robot is no worse for card room revenues than a good human player, and could actually help revenues, by raising total betting volume. There’s a reason why casinos and card rooms use shills to keep games going. A robot who plays at a monetary break-even level (exclusive of table fees) could be a perfect shill.

I would even worry about room operators selling bulk hand data to robot operators. Even if anonymized, so that robots don’t have perfect records of other exact players’ tendencies, this data could give robot operators unparalleled insight into what strategies, on average, generate positive returns given different actual populations of players.

I suspect robot operators would pay a lot for for this data; the countervailing incentives for room operators not to sell would be (1) the risk their reputation could be sullied by if such collusion with one class of players was discovered; and (2) the long-term risk that players would become discouraged from wagering by always being beaten by better-informed opponents. Yet the willingness of people to play games with guaranteed negative returns suggests (2), at least, would be far fram fatal to card rooms’ businesses.

If Levitt’s out-in-the-open Pokernomics has any hope of generating useful results from mass data, secretive syndicates have probably already managed the same.

John Forsberg May 6, 2006 at 8:13 pm

“Tyler quoted Harford paraphrasing Darse Billings saying that one day (and presumably thereafter) in online poker the best robot will beat the world champion. No one is claiming that robots will be supreme face to face.”

Uhm, yes there are in fact loads of people claiming that. Maybe you mean something along the lines of “the best exploiter of poor human opponents in a live setting will always be human”. And maybe you’re right about that.

The claim “the best computers will be able to beat the best humans in 1-on-1 games regardless of setting better than any human” certainly has been made, and in my opinion pretty much is a lock. Tnis is because tells based on physical appearances are of very little value against expert opponents who simply don’t get very excited about pulling a huge bluff or having a lock on the pot. Been there, done that and thrown away the tacky t-shirt.

Possibly there is some validity to the claim “the best performer against an expert human during very long sessions in a live setting will be human” due to the fact that some humans might be better exploiters of exhausted expert players. But I’d would think that the computers complete lack of exhaustion for any time period would well compensate for not being able to extract every last bit of value from an exhausted expert.

As for the existence of programs playing online with a positive expectation, they already are (and yes, I do know of specific examples). And it’s not that much of a challenge due to the vast masses of incredibly poor players playing online. I’m pretty sure I could, with a rubber hose and a sufficient supply of bananas, teach a chimp to beat the low limit online games. The quality of play is simply appalling. Of course, beating the higher stakes games is a lot more difficult due to the generally stronger opponents.

As for hand history data, they already are publicly available in that any observer can pull up any game and watch. I download around 5mb of data a day using a fairly unsophisticated program and I’m sure someone who wanted to fairly easily could access far larger volumes than that with a bit of ingenuity.

There’s also the *huge* issue of collusive strategies which, while cheating, are very hard to detect. I believe that programs utilizing those kind of techniques could beat the very best players today fairly easily if working in a 2-against-1 setting. I’d expect those sort of bots to become a lot more common, and they would be unbeatable unless countered with the same kind of techniques.

Obviously a two-player optimal strategy does not break down once it’s know. It’s kind of the point of an optimal strategy.

The point about the appearance of stopping cheating being more important that actually stopping cheating is right on the money though. Card room operators usually claim that they can’t disclose their procedures due to security reasons, but I know of quite a few cases of incredibly blatant cheating that hasn’t been detected. Customers, cheaters included, are incredibly valueble (I paid something like $10k in fees last month and $1k is very common) to the game operators and they’re very reluctant to lose any of them.

For the record, I *am* a professional poker player. Though I don’t play one on TV, so don’t expect to know who I am.

John Forsberg May 7, 2006 at 2:03 am

“The nature of the private information in poker and game theory precludes all 100% answers to every situation.”

No it does not.

A trivial example is when checked to on the last street of betting when you hold a hand that cannot be beat and is last to act. You bet 100% of the time, no exceptions. Those sort of situations are not that uncommon really. Especially not in a 2-player game.

Further, there are lots of solutions to poker games that include very little or no mixed strategies. Instead you mask your hand strength by playing different hands the same way. Eg betting the top X% and bottom Y% of your hands at the last round of betting when checked to.

I even think there might be some proof that there needs to be *no* mixing in any 2 player poker game, though I’m not at all sure about that. If there is, it has to do with solving linear equations. At the least it’s possible to confine mixing to a few key hands.

Here is a link to a fairly simple, hold’em-like game solved by Alex Selby for those interested: http://www.archduke.demon.co.uk/simplex/index.html

If you take a look, you’ll notice that one of the solutions he presents only uses a mixed strategy for 42/1326 hands.

BillWallace May 7, 2006 at 4:37 am

There must be some situations with mixing…. Not all.
Obviously some situations are 100%.

I’m almost certain the optimal heads up limit solution would include mixing.

linda October 9, 2006 at 9:45 am

Comments on this entry are closed.

Previous post:

Next post: