incomplete information - so what?

[runway model]

There seems to be a widespread argument that IN THEORY and even in faceless internet play, a poker computer cannot play as well as the best humans because poker is a game of 'incomplete information'. True but irrelevant.

So long as the computer has access to ALL the same information as the human (however complete or incomplete that information is), then the computer can be programmed in theory to play just at least as well and almost certainly better that the best humans.

Of course, the computer would need to keep huge database records of opponent characteristics and tells including opponent response times. In fact everything that a human can observe or process on the internet already. Potentially, powerful computers can use Poker Tracker type info far more efficiently than humans.

Yes, in practice it's a huge task. But rest assured in years to come, the best computers will be the best players. Better at exploiting weaker players, better at beating the strongest players, better h/u, better in multiway games, better at making the most +EV decisions against random combinations of strong and weak opponents. Not for a few years maybe, but in time you will believe...

[AaronBrown]

I'm not sure where you came across the "incomplete information" nonsense, but I wouldn't pay any attention. Computers have no problem with incomplete information.

I agree with you that in heads-up limit play, computers should be able to beat the best humans. It will take some work, but I expect it to happen. As algorithms get better and hardware gets cheaper, I think it's a matter of time.

No-limit is a much harder problem, in fact it becomes hard to define "beat". You can solve that by setting a horizon, like play until one player is broke, or play 1,000 hands. But that's moving in the direction of a limit. It will be interesting to see how well computers can play no-limit, and how broadly they can define no-limit.

I don't expect to see a computer program that can play well at a table of humans anytime soon. If someone did that, it would be a major breakthrough in machine intelligence.

In an important sense, I think that means no one has any idea how to build a poker machine that can compete with good humans. It's true I expect someone to figure out how to beat heads-up limit poker, but only because it exploits aspects of the game that are not essential to the poker. We didn't say computers beat the best human chessplayers until they could do it under tournament-like conditations, the way important chess is played. The fact that a computer could win at speed chess was a relatively minor accomplishment. People play speed chess and heads-up limit poker, but to be considered a human-beating poker player, a machine has to come to a no-limit table with low blinds and no timeclock.

In terms of tells, if you allow the computer the use of machine sensing devices, it should have no trouble beating humans now. That's already being done in the lab. It's possible that humans could learn to beat the machines, but at the moment sensing of voice timbre, micromovements, pupil dialation and other indicators removes the incomplete information from the game.

[thechainsaw]

[ QUOTE ]

I don't expect to see a computer program that can play well at a table of humans anytime soon. If someone did that, it would be a major breakthrough in machine intelligence.

In an important sense, I think that means no one has any idea how to build a poker machine that can compete with good humans.


[/ QUOTE ]

I am not sure that this follows. There is a difference between knowing how to do something and actually doing it. You can solve chess in about 5 minutes, and program it in less than a day, but current hardware would be too slow to run it in a tournament setting, so the work invovles speed optimizations and cutting corners, but people know HOW to do it.

In poker, we just dont know how to do it, or at least we are unsure. I think that's the important distinction with the idea of incomplete information. Addionally, just because no one has done it so far, does not mean that no one knows how to do it. To say no one has "any idea", based this observation, I think is a non sequitur.

The same computer vs human argument always is posed in bridge, and the argument that no one has built anything anywhere decent yet always comes up. But imho, if you were a programmer good enough to solve bridge, you probably wouldnt be writing bridge software.

[Copernicus]

I agree with AB, incomplete information has nothing to do with it. Humans and machines have the same constraint of II, so that can't possibly be the barrier to a winning poker bot.

In no-limit I would liken the problem to "graininess". A programming team can build a data base of all of the relevant information, and program the game at the 0th and 1st levels without encountering too much "graininess", and can probably become a decent player.

The problem as I see it is that the architecture of the program still needs a way to sort through all of the information and make a decision. The human brain is quite adept at recognizing patterns...those things that lead to "intuition". But just as a photograph that is focused on too closely loses context amidst the detail, and becomes an incoherent mass of grains, pixels or whatever you want to call it, poker thinking at the 2d level and beyond can get to bound up in the data.

Until there are machines built to learn the way the brain learns...multiple synapse/neuron trees linked with each new memory, I don't see a machine being able to balance the "big picture" and the data well enough to beat a full table of NL players.

Its more of a SMP topic, but obviously I don't think the Turning test has much validity.

[David Steele]

[ QUOTE ]

I'm not sure where you came across the "incomplete information" nonsense

[/ QUOTE ]
I see it too in almost any thread on this subject.

[ QUOTE ]
In terms of tells, if you allow the computer the use of machine sensing devices, it should have no trouble beating humans now.

[/ QUOTE ]

Just to clarify, so the fair test would be some internet style poker game where neither type has tells?

Machine vision, audio analysis etc at one time was considered to be AI, until it worked.

Its much harder to solve a problem with a computer and require that it be solved very close to the way a human solves the problem. It works both ways, it would be crazy to require the human to play chess and analyze a million positions and not use much pattern rec.

D.