Feb 04, 2017
Libratus Crushes Poker Pros by $1.8m - 5 Big Implications
By RTR Dennis
For the second time in three years, researchers from Carnegie Mellon University trotted out an artificial intelligence program designed to play heads-up no-limit Texas hold'em.
The program, Libratus, faced off against four successful poker pros – Dong Kim, Jason Les, Jimmy Chou, and Daniel McAulay – in a battle for poker supremacy.
Many felt that the second edition of the Brains vs. Artificial Intelligence contest would again go to the humans because they beat Claudico by $732,713 in 2015.
However, the evolved Libratus won convincingly, beating the four humans by a combined $1,766,250 over 120,000 hands. Only Kim managed to play respectable against the program, losing $85,649 over 30,000 hands. Everybody else lost between $277,657 and $880,087.
Obviously this is a monumental step in poker AI history, given that a computer program beat four skilled pros at a game of imperfect information. But just how big is this? Let's look at three implications below.
1. HU No-Limit Hold'em Is/Will Be a Conquered Game
Libratus hasn't proven it can beat elite high-stakes cash pros. Furthermore, 120k hands isn't the infinite sample size to end all. But given the convincing victory over four pros who are highly regarded, we have to assume that HU no-limit Texas hold'em will become a conquered game – if it's not already.
Other games that AI programs have beaten include limit poker, chess, checkers, Go, Jeopardy!, and Othello. Texas hold'em is more complex than chess, checkers, and Go due to its imperfect information. And that's what makes this amazing because beating HU NL hold'em was considered the gold standard for AI technology.
Many will argue that a larger sample size needs to be played. But Libratus shows that AIs have advanced greatly in the last few years, so much so that they're capable of beating top pros.
2. Libratus Will Only Get Better
Unlike Cepheus, a University of Alberta bot that plays perfect limit hold'em, Libratus does not represent perfection in NL hold'em strategy. This means that some top pros could possibly beat Libratus over a reasonable sample size.
But we can also conclude that above-average NL hold'em players have no chance against this super AI. Bots have been beating low-stakes online poker for the past several years. And Libratus shows that bots are no longer just micro-stakes dwellers – they can beat any solid player.
This is comparable to Deep Blue, the chess-playing AI that defeated champion Garry Kasparov in 1997. But even in the early 1980s, chess bots had the skills to beat mid-range grand masters and masters. The improvement by Deep Blue shows that with time, Libratus will eventually beat any poker player without question.
3. This Win Has Broader Implications for AIs
While the Brains vs. Artificial Intelligence challenge was certainly fun, beating Texas hold'em isn't the end game for Libratus and AI technology as a whole.
The thing about AI programs designed to beat a game is that they're quite dumb when trying to manage the complexities of broader thoughts. So the goal for Carnegie Mellon researcher Tuomas Sandholm and his team is to improve Libratus' artificial general intelligence (AIG).
Rather than having a narrow focus, like beating poker, a high-powered AIG could move from diagnosing medical conditions to maintaining perfect cyber security. This means programs like Libratus could one day think like humans – only they'd be better at mastering and remembering tiny details of in-depth subjects.
4. Online Poker Is in Even More Trouble
As mentioned in point #2, some low-stakes games have/do feature poker bots. Online poker rooms have done their best to monitor the situation and ban AI accounts. But Libratus shows that poker sites now have to worry about superior bots infesting NL hold'em at any stakes.
The good news is that reputable poker rooms are still banning AIs. Furthermore, it's not like Carnegie Mellon is renting Libratus out to amateurs. But one day, we'll have unbeatable bots that can also evade security measures set forth by online poker rooms.
When discussing if AIs have ruined online poker, Kim said “Not in the near future, but we should be worried. I'm no rocket scientist but I assume that anything with computers grows exponentially. The end is near. It was a good run.”
This certainly doesn't meant that internet poker is in immediate danger of going extinct. But security measures will have to continue improving and perhaps evolve into something altogether.
5. Live Poker Will Largely be Unchanged
While online poker faces more challenges than ever, live poker will mostly be the same game. Sure, some players use sophisticated bots to improve their own skills away from the tables. But it's not like people can use supercomputers at a live poker table to beat the game.
Relating poker to chess one more time, the latter is still popular among live players. However, you're not going to see chess players betting money online due to the perfect AIs.
We can expect the same scenario in poker within 10 years, where few people play online for real money, but they still show up to live tournaments and cash games.