CMU Researchers Reveal How Their AI Beat The World's Top Poker Players (triblive.com)

2017 began with an AI named "Libratus" defeating four of the world's best poker players. Now the AI's creators reveal how exactly they did it. An anonymous reader quotes the Pittsburgh Tribune-Review: First, the AI made the game easier to understand. There are 10**161 potential outcomes in the game of poker -- that's a one followed by 161 zeros, potential outcomes in a game of poker. Libratus grouped similar hands, like a King-high flush and a Queen-high flush, and similar bet sizes to cut down that number. Libratus then created a detailed strategy for how it would play the early rounds of the game and a less-refined strategy for the final rounds. As the game nears the end, Libratus refined the second strategy based on how the game had gone.

A third strategy was at work as well. In real-time, Libratus created another model based on how its play stacked up against the play of the humans. If the humans did something unexpected to Libratus, the AI accounted for it and built it into the strategy. Instead of trying to exploit weaknesses in the play of the human, Libratus focused on improving its play.
The AI was created by a computer science professor at Carnegie Mellon University and his Ph.D. student, who argue in a new paper that "The techniques that we developed are largely domain independent and can thus be applied to other strategic imperfect-information interactions, including non-recreational applications."

"Due to the ubiquity of hidden information in real-world strategic interactions, we believe the paradigm introduced in Libratus will be critical to the future growth and widespread application of AI."

Comments

Well that's the end of online poker.

[Mal-2]

It was one thing when bots could beat up on donkeys, but when even the best human players can't win it means only bots will be left standing. That doesn't mean humans are totally out of the loop, someone still has to be standing by to talk to the admin when questioned about their human status -- for now. That too will probably fall before long.

The micro-stakes tables will probably remain largely human because there's very little to lose (or gain) down there, but for high-stakes games this signals a rapidly approaching end.

Not comparable

[nospam007]

It's way harder if you know you may end sleeping in your car for a couple of years if you lose.

It cheated ...

[140Mandak262Jamuna]

By counting cards. The bouncers will have it thrown out in a minute.

So?

[fluffernutter]

It kind of defeats the point of AI if you preload it with all kinds of statistics a human wouldn't have access to while playing a game. The point of AI is to make a computer think and learn like a human, not to prove that a computer can beat a human. We already know computers are better at calculations than humans.

Re:So?

[hey!]

The point of AI is to make a computer think and learn like a human, not to prove that a computer can beat a human.

Well, that's one possible point of AI. Other possible goals would be to generate better results, or good-enough results but much more cheaply. Yes, there have always been people who have hoped to shed light on human intelligence by creating machine intelligence, but by in large they haven't had as much to show for their efforts as people trying to improve on people in some way (performance or cost).

Re:So?

[mentil]

You're kidding, right? Read any 'how to win at blackjack/poker/gambling' book and, right after the rules, it will give all of the precise odds for all of the means of winning. Amateur gamblers are expected to memorize ALL of these odds, this is Gambling 101. IIRC, it's allowed to bring a 'cheat sheet' of all of these odds into a casino, as well. In this case, humans know a bunch of statistical information regarding the game. That said, this AI did improve its strategy over time, which could be called 'intelligence'. It doesn't have to precisely mimic human learning in order to be 'intelligent', they weren't calling it 'Artificial Human Intelligence' AFAIK.

Re:So?

[Kjella]

I feel it's been more the opposite, people claiming machines can never do this or that until they have human intelligence. And since we don't really know what that is or how to build it, we can't have it. And then we have AI - but since so many people seem to get their panties in a bunch about that term I'll just say adaptive algorithms - that do it better anyway. And despite it's encroaching on more and more human jobs we're clinging to our own unique abilities as important.

If I'd like to eat a hamburger, is there any part of its creation I don't think can be done by a robot if I extrapolate? I think a robot can cook it. I think a robot can get the ingredients to cook it. I think robots can grow the crops and vegetables, grind the flour, bake the bread, raise the cattle and so on. I think a robot housekeeper can wash my clothes, iron my shirts, do the dishes, vacuum the floors, dust the shelves and so on. The building I live in, I think could be built by robots. The excavation, foundation work, carpentry, plumbing, wiring, HVAC, painting, floors and tiles and doors and whatnot.

Pretty much every practical aspect of my life I think could be potentially created by or done by robots. And that possibly includes a sex bot to replicate the physical experience of having sex, it should certainly beat jerking off. So what does that leave for humans? Creativity, deep thought, comprehension, philosophy and emotional connection. Which sounds like pretty big outs, until you start looking at them in more detail. For example, creativity are often variations on a theme and just because it's unique to me doesn't mean it's totally unique. For example in a lifetime I'll eat 80*365 = ~30k dinners. There's probably way more than 30k recipes out there and there's no harm in eating a great meal twice. I don't actually need a chef to come up with new meals for me to have a new meal.

Deep thought maybe, to the degree that you don't have algorithms smashing you on speed or parallelism like playing chess or Go. The computer does a thousand simulations refining its solution in the time you do one. Comprehension yes, like what do these business specs actually mean. But we're working on digital assistants that maybe one day will evolve into rapid prototyping that'll do that kind of bouncing back and forth until people get what they want. Philosophy yes, though you can probably fake it pretty well with a textbook since nobody has real answers. Emotional connection yes, if so many weren't happy with faux emotion. Like a sex bot should in theory fake it equally well as an escort/prostitute.

Overall, it's getting pretty hard to see a niche that is like guaranteed forever "human" and where we're not just tweaking some kind of meta-meta-meta-parameters of what an algorithm is doing. Sure it's a long way from where we are now to there but the principal hurdles have turned out not to be so principled after all, mostly just practical. There's tons of people working on automating all these little niche tasks and put together they are automating pretty much all of society. Turns out what humans do is taking one extremely unique and flexible brain and using it to do pretty many not that complex, limited tasks that can be automated. Basically, our mind is overkill and "AI" is sufficient.

Only win in an ultra simplified tournament

[Blue23]

I followed this as it was happening. This is NOT about bots being able to beat human players. It's about bots being able to beat human players in the simplest possible space that doesn't mimic 99% of actual poker play.

It was only heads-up with 1 human a time, not vs. a table. After every round the money was reset so it never had to play from low amount of chips, or have to try to bully with it's chip advantage. The amount of chips vs. the big blind was a very large stack in the first place even before it reset every hand, so the blinds were statistically little more than noise in the amount that was going back and forth.

Don't get me wrong, this is really interesting and great strides. But this is far from a bot being able to play at a full table and having to deal with a few bad hands taking it out of the place where it's betting is suited for. (If you have less of a stack, you have less of an upside so draw hands aren't worth as much.) Or to have someone with a larger stack push it beyond it's acceptable betting and make it fold because it can.