Yet another game where AI beats human

[tj86430]

Libratus won against (or should I say crushed) professional poker players:

http://www.pokerlistings.com/libratus-p ... -end-42839
https://www.bloomberg.com/news/articles ... play-poker
http://www.theverge.com/2017/1/31/14451 ... tournament
https://www.engadget.com/2017/01/31/lib ... an-rivals/
https://www.wired.com/2017/01/mystery-a ... ers-poker/

[Kirby]

Andrew Ng, a famous AI researcher (and co-founder of Coursera), seemed significantly impressed with the feat, noting that it was a significant milestone for AI.

Othello/Checkers/Chess/Go were theoretically solvable with minimax tree search and sheer computation; but poker, which requires bluffing, needs sophisticated modeling of your opponents and new algorithmic principles.

As reluctant I've been to be enthusiastic about computer go, I personally feel that Ng's statement plays down the significance of AlphaGo - I would consider their implementation to be sophisticated, and it somewhat downplay's AlphaGo's success (i.e. it took a lot more than a "minimax tree search and sheer computation" to create a successful go-playing program).

Nonetheless, I agree with Ng's assessment that the betting aspects of poker, such as bluffing, introduce a level of complexity that required sophisticated opponent modeling, as he suggested.

I was just a little turned off by the sweeping statement that appeared to downplay the complexity of go.

[Uberdude]

Well, Go is 'theoretically' solvable in such ways if you don't mind waiting for the universe to die out several gazillion times per move, so not such a great theory in practice. As you say AlphaGo needed some rather clever algorithms, as will this poker AI to play an imperfect information game. Maybe it's relevant that Andrew Ng co-founded Google Brain, which got superseded/subsumed/overtaken by DeepMind so there could be a bit of sour grapes re AlphaGo. He's now in charge of deep learning at Baidu, I wonder if they are making a Go AI as many other Chinese tech companies are (e.g. Tencent, and who made Xing Tian?).

[pookpooi]

As there are many imperfect information games, some of them might be harder to surpass human (like StarCraft, which DeepMind is aiming right now) and some of them might be easier (like Scrabble, surpassed since 10 years ago). I don't know where Poker sit between this, but the AI Libratus is consuming even more than AlphaGo. With with more than 15 million core hours of computation in training, and throughout the competition, Libratus recruited the raw power of approximately 600 of Bridges’ 846 compute nodes. Bridges total speed is 1.35 petaflops, about 7,250 times as fast as a high-end laptop and its memory is 274 Terabytes, about 17,500 as much as you’d get in that laptop.

[thegrrr]

I wonder, maybe it will kill online poker sooner or later.

[tj86430]

I wonder, maybe it will kill online poker sooner or later.

Probably at some point, yes. It will take some time, first the players will move to more complex games (Omaha, 5 card omaha, hi/lo games etc) but eventually it is likely that online poker will either die or at least suffer badly. Also Libratus used really massive amount of computing power in this match, but Moore's law and other developments will take care of that eventually.

Some people also say, that it is possible to recognize poker bots (and it is to some extent), but I believe the bots will become smarter in this regard, too.

BTW, FWIW, I think this achievement will have much much more effect in our whole society than any of the other AI achievements so far. This is the first time a complex game of imperfect information is mastered by AI at this level, and it will have huge implications in commerce etc.

It is also worth looking at other achievements of prof Sandholm. They are very interesting indeed (of course I may have a slight nationality bias here)

[pookpooi]

Talking about other field application, can you guess from this abstract?

"....In exploratory studies, we demonstrate that MCTS with neural network policies outperforms the traditionally used best-first search with hand-coded heuristics."

No, it's not another deep learning go AI paper, it's "Towards "AlphaChem": Chemical Synthesis Planning with Tree Search and Deep Neural Network Policies"

[dfan]

I wonder, maybe it will kill online poker sooner or later.

Online poker in the US was already kind of killed off by the UIGEA, but even without that, bots are already better than the average person at medium levels.