SoylentNews is people
SoylentNews
AI's have beaten the best human players in chess, go, and now poker.
In a landmark achievement for artificial intelligence, a poker bot developed by researchers in Canada and the Czech Republic has defeated several professional players in one-on-one games of no-limit Texas hold'em poker.
Perhaps most interestingly, the academics behind the work say their program overcame its human opponents by using an approximation approach that they compare to "gut feeling."
"If correct, this is indeed a significant advance in game-playing AI," says Michael Wellman, a professor at the University of Michigan who specializes in game theory and AI. "First, it achieves a major milestone (beating poker professionals) in a game of prominent interest. Second, it brings together several novel ideas, which together support an exciting approach for imperfect-information games."
Source: Poker Is the Latest Game to Fold Against Artificial Intelligence
Is there anything at which AI's won't soon be able to beat humans?
(Score: 5, Insightful) by ledow on Friday January 13 2017, @10:58AM
Poker has fixed odds. They're easy to calculate, card-count and whatever else.
The problem is not the card game, but the betting. Betting has much more possibilities (if I have $1000, no limits, etc. that's nearly 1000 possible combinations, which turns into a different number next walk round the table).
But, again, if you can work the odds out, you can come up with a betting strategy. Your opponents can only do the same, and it's the interaction between what they each bet and what you bet that determines who wins the overall game 50 hands or whatever later.
But it's still a game you can build a graph of and find an optimal route through, it's just more complex than knowing there are 3 Aces left in the pack.
Humans can't calculate those kinds of betting odds - even professionals. That's why professionals will tell you that they read their opponent and so on in preference. That's something they can do that computers can't.
However, given enough rounds, enough processing power, large enough game tree branches, it all comes back to simple game theory and - in essence - graph theory. It's just that we have the capability to handle that now.
This isn't anywhere near "gut feeling" or AI or any such nonsense.
The only real surprise in AI in the last 20+ years was Google Go machine which DOES NOT try to iterate every possibility (like Deep Blue effectively did, which is why Deep Blue isn't THAT impressive). It can't. The game tree for Go is unbelievably huge, far, far huger than anything chess could ever approach. That Google's AlphaGo (or whatever it was called) can beat good humans at Go, that's a true leap in capability that's not just approaching the brute-force requirements.
Any variation of poker, though? Though the game tree is bigger than you might think, it's nowhere close to Go still, even with extremely fine betting quanta.
(Score: 2) by The Mighty Buzzard on Friday January 13 2017, @01:19PM
Me, I'm still dubious about it consistently beating professional poker players for the simple reason that they lie. Regularly and with great skill. You'd need to build a play style profile on your current opponent(s) to even be able to guess at when they're bluffing and if you started to and started winning, they'd change.
My rights don't end where your fear begins.
(Score: 0) by Anonymous Coward on Friday January 13 2017, @02:46PM
Lying is only effective if you don't get caught and professionals aren't trained to lie to computers. The computer just has to call the bluff enough times to balance out some of its benefits and its superior odds calculations will finish the rest.
(Score: 2) by The Mighty Buzzard on Friday January 13 2017, @02:52PM
And the human just has to adjust how much he bluffs to make that strategy the losing one.
My rights don't end where your fear begins.
(Score: 2) by bob_super on Friday January 13 2017, @06:52PM
Bluffing only gets you to a certain point against something which knows exactly the odds that your cards are better.
Playing completely against the usual raise logic will either augment your losses or lower your gains. They may get a few big ones, but the machine apparently grinds them out over the long term.
(Score: 1, Insightful) by Anonymous Coward on Friday January 13 2017, @01:38PM
Card counting is for Blackjack, not Poker. Also was the AI up against anyone who could bluff worth a damn?
(Score: 0) by Anonymous Coward on Friday January 13 2017, @01:54PM
I'm not sure this dude has ever actually played poker.
(Score: 1, Touché) by Anonymous Coward on Friday January 13 2017, @09:04PM
The cards in your hand and the cards up on the table change the odds of any given hand your opponent can make.
Example: You are holding two kings and there is one on the table. Now, what are the odds that your opponent will make a hand with three kings?
(Score: 3, Insightful) by FatPhil on Friday January 13 2017, @03:32PM
Journals were rejecting papers on neural nets, as they were "academically" analysed to death, and even if new results came in with slight improvements over domain-specific expert systems they weren't "advancing the art". It's only when outsiders (OK, they were technically academics, but they weren't playing the academic game, they were bought ivory towers by corporations/funds, and told to just ponder and play) stormed in and started winning every AI competition that existed (image recognition, drug discovery, speech recognition, you name it), that academia realised that neural nets were, again, third time lucky, the next big thing. As did corporations, and basically everything exploded, and error rates plummeted quicker than you could publish your own new world's lowest results. Finally, some time in 2016, deep learning neural nets, trained on *enormous* amounts of data, and making wide use of GPUs, finally basically became better at recognising everything than humans.
Paired with learning in the inputs -> virtual-understanding direction was the virtual-understanding -> outputs direction too, and suddenly they became able to not just do translation better than ever before, but even voice recognition + translation + voice sysntesis in real time. Shrink that down to a small circuit that will fit in or behind the ear, and you've basically got a babelfish. There's no reason why the input and output need to be even in the same domain - image-input -> virtual-understanding -> english-output is perfectly possible, and currently (mid 2016) gives descriptions of images that humans think are perfectly adequate, and are even preferred to ones given by humans 25% of the time. Expect that figure to rocket, the system had only been training for a few weeks when that measurement was taken.
The AI's got no tell, the human's got no advantage. Not only can the robot learn to play Nash-optimal poker, but it can also learn to exploit any weakness it may find in a human who's not Nash-optimal. To be honest, this is a relatively small victory, the only people who are shocked are people who have wildly inaccurate views on how competent humans are.
Youtube has a bunch of stuff by people like Geoffrey Hinton and Jeremy Howard on the subject of Deep Learning, most of them are both informative and entertaining. A few (such as some UCLA lectures) are a bit technical and not for someone already quite familiar with neural nets (and similar constructs).
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2) by goodie on Friday January 13 2017, @06:49PM
Your comment made me think that it would be interesting to have an AI learn about people's physical clues when they play poker. Could be interesting to determine how to play the next round etc.