Blog
Poker is about to be tackled by AI
It does come as a surprise to me that it seems that only one year after Go, poker is going to be the next game to be conquered by AI. Poker is way more complicated than Go, because of the unlimited nested levels of bluff (bluff, bluffing that you bluff, etc.), it not being a perfect information game, it being played in groups of various sizes with players of varying strengths, it having no limits (that is, in no-limit versions of the game), it having psychological elements (exploiting tilt), etc. Of course, if we see poker as a combination of (1) calculating chances, (2) opponent dependent strategising, and (3) psychological trickery, we could still say that the computer’s unsurprising superiority in tasks (1) and (2) compensates for its lack of capacities related to (3). In that way we can still deny that the computer lacks social intelligence of the kind that poker players use, even though it is able to play a very decent game of poker.
This must be a blow to online poker. But it is not to poker in general. It is quite clear that it will not be much fun playing poker against a machine.
“REINS” is a research project funded by an ERC
Check out the video as well!
https://www.theguardian.com/technology/2017/jan/30/libratus-poker-artificial-intelligence-professional-human-players-competition?CMP=share_btn_fb
Some further research means that we need to curb our enthusiasm a bit more. It turns out that the A.I. is not playing the 4 professional humans at the same time. Instead the humans take turns playing the A.I. heads-up. Heads-up poker is a totally different game from ring game Texas Hold em. Heads-up play is much more mathematical and it is especially important to randomize things like pre-flop raise/fold ratio, bet sizes etc. Randomizing is, of course, something Humans are notoriously bad at. And we would also expect computers to be better at calculating odds. Hence, in retrospect the achievement by the engineers of this poker bot is, while still impressive, not as monumental as it sounded at first.
To somewhat curb the enthusiasm. Of the pro’s playing this both the top one (Jason Les) is rated 598th in the GPI (global poker index) The others are all ranked lower than 10.000th. So it is not like the computer is really beating the best of the best humans. Furthermore, they are playing 120,000 hands in total, and are now about halfway. This is still a relatively small sample size. It will take more hands to determine if the bot is really consistently winning. Also, when more hands are played the pro’s might be able to figure out how quickly the bot is able to adjust to tactics changes etc. And they might be aple to exploit that. And finally, these pros all know each other and know how to play each other, the bot is the new player on the table so it can profit from that. In poker a new human player often splashes onto the scene winning big in his/her first year only to lose it all again in the years thereafter because the other pros have adjusted. So while the achievements of this bot and the other achievements of deep learning systems are certainly impressive. It is much harder to say when a bot is consistently better than humans at poker, then, say, better at Chess or Go.