can you imagine phil hellmuth playing against libratus? "it called with ace-ten honey!" "this computer is the worst poker player I've ever seen! really it probably won't last another hour!"
I wonder how the study could go if for the first day you tell the players they are playing against an anonymous pro who is and in another room? I feel there can be a placebo effect with players being conscious of the AI. This was a very interesting report, thanks for posting!
Poker is still a game of luck. All the AI is doing is trying to approximate the Nash equilibrium strategy as closely as possible. It doesn't win by being "good" it wins by making less mistakes than humans
Nullpointer but if it wins it is the better player. The better player makes fewer mistakes. It (the AI) doesn't get tired, frazzled, pissed, euphoric or impulsive.
I forsee a future where a bearded Poker-playing AI with chicken leggs, is posting Instagram pictures filled with hired models, guns and luxurious mansions. And people will like the AI for that.
They are not the best HU players in the world tho.. Which kinda sucks, would be great to see the comp vs the best of the best. They're very good pros.. but best HU they are not.
tfw this is actually fake news too... The test is flawed, they are playing mirrored hands and are equity chopping all in scenarios which fundamentally changes the game and the way mirrored hands should work. Dong Kim explains this in his interview with Doug Polk
I am really interested in seeing how Libertus adapt his play when facing itself. I can reason that it would reach several stationary states all of which are only metastable. The issue is how can it closely approach the optimal strategy while the other Liberatus is doing the same. Also it is amazing how easily Liberatus drifts far away from GTO to exploit the opponent! Great job!
1. Those players are not "absolute best". Far from it 2. The player hadn't had the usual incentive to play to win. 3. The players chose to play as a marathon (being tired), to get the same money in less days 4. It's a statistical tie (insider joke)
They are very good players but not the "best" in the world. Also, the computer can not predict what cards are coming out next. Nor would it know what you mucked. Also, the computer has a bit of an advantage because it can instantly calculate odds that would be too complicated for a human, especially since humans get tired. I would like to see it play against a mix of regular players and much higher level players. Also, the computer is playing unfairly. Human plays aren't bringing their charts, computers, calculators, or consulting outside sources. Level the playing field and let the humans use any resources they want which would cut down of fatigues and human error.
@@jburch5752 I think they were using some tools to analyze what was working and not working but yes you are right computers can do calculations in second. I'm reading the poker blueprint and the formula to find ev is too long to do each hand. That's why we have to just estimate and review later unlike computers
Yep, AI can make very weird moves that you rarely see high level players use, it's just because they can go through every scenario that pro level players don't consider due to an undiscovered way of thinking. Chess, poker, whatever. Any player that played with an AI can notice a difference.
It's tue that old style AI's (hard-coded logic with future-move analysis) often have really weird moves. But modern Neural-network AI's do learn like humans (from other humans, through experience, just at a really high rate) and have much more human-like moves. In fact, AlphaZero, the version of the Alpha AI that leans without being fed human game replays, has rediscovered many human strategies by itself. AI's in much more complex real-time games (like DOTA and StarCraft) do also play a bit like humans (although they feel oddly smart and skilled), and show human-like behaviour like faking, baiting, smart exploiting of game mechanics,... But when they feel inhuman, it is mostly because they have much more experience, they have seen and played millions of games.
The most interesting bit is that, if the network was trained against a computer algorithm it would learn the loopholes embedded within its syntax. When trained against a human it is learning our flaws and mannerisms. I say we probably shouldn't strive for a general A.I unless we wish to be fully displaced.
5:20: How did he arrive at that figure? Don't get me wrong, the number of possible situations that can occur in a full ring game are incalculable, but the amount of possible permutations in a HU game is roughly 5.75 quadrillion.
Have a computer on the side playing online poker and each hand you play just enter the exact cards you have and the flops then let the Ai tell you what to do. Mark witch position is checking or raising and let the computer give you the next move. (Example, fold because that that that or 3 bet because i see weakness, you will win millions) I’m programming my own poker AI assistant for online games
I would like to see if there would be any variance if the cards were real and the info of the hand was inputted into the Libratus instead of playing the game in the computer that has Libratus on it.
I'm wondering how this AI would do in 6 handed sit&go's for example. I'll bet that it won't be long before most poker sites will use a version of this AI to consolidate and probably even increase their margins. And I'm wondering how long it will before an AI will win the WSOP using 'speech play'?!! Like a boss!!
6max sitngoes can be beaten by bots pretty easily. However, a bot beating a 6max cash game table over 100k hands would be extremely hard for a bot if all other 5 players were extremly talented (such as the ones in this video). It would be even harder to beat if it was a 9max cash table.
The theoretical problem with a 6 handed cash table is there are exponentially more situations that the bot could come across. Logically it would mean that they need a lot more computing power to pull it off than they do even now!
As for the best strategy for improving odds, considering the ai has better statistical data, strategy, and is overall steps ahead of the player, I think the best odds you have are going all in on premium/high valued probability hands, as the players started to do. The ai will still beat you at valueing it's hand, but by making it all in every time, it removes the river, flop, and turn betting strategies, and puts more into preflop odds, as well as limit the ammount of hands the game is played, limiting how much the statistics effects the outcome. Although, the ai would win over 50%, I presume this is the closest you could get it to even win/loss. Simply do your best to remove all strategy from the game, and put your chips in on luck.
This is a monumentous task involving having to vary your strategy over time and even changing the rate at which you change strategies over time, whilst also having each style be concrete enough in order to not leave too much up to chance. All the while, the AI's model has a discrete memory. Since it's playing against multiple people and learning from them whilst still keeping individual strategies for individual players, they are playing against a machine with the solid foundations of many which can focus that knowledge in order to wreack havoc. Like a prized warrior faced by the collective efforts of a small army, terrifying and also very instructive and awe inspiring!
last year their bot got crushed in almost the same way it did the humans this year but the company declared it a statistical tie in an official report. so gratz on a statistical tie.
So many people commenting that the ai is cheating because the "computer" knows the cards ... are people really that dense they think 5 pro poker players and multiple leading scientists didn't think about that. I'm pretty sure the ai didn't know the cards or his winrate would have been 95 %
I'm wondering if it took 20 days of poker play for Libratus to solve the game? Can it fair as well in a normal poker time frame like 3-8 hours of gameplay?
Libratus is likely the best bot at this point, but you can check out the Anual Computer Poker Competition, where many good bots play against each other. Even an earlier bot from the makers of Libratus and other universities as well.
Questions:- 1) Did the blinds increase over time or were blinds the same through out the competition? Could the computer account for the Blinds increasing? 2) Could the AI in its present form compete at a 10 person table tournament? 3) What other games left does google AI want to compete in?
1. The blinds stayed the same throughout. Also each hand started with both players having a 20k stack. So each hand both players were equally deep. 2. This is just a guess, but I wouldn't be surprised if it would do reasonably well. The team behind Libratus has said that in their estimates Libratus could crush 6-max in 2 years. I think this is a careful estimate, and believe it can be done sooner if they put their effort on doing it. 3. Your guess is as good as mine here.
I understand that they're probably more interested in the practical applications of this AI, but I would LOVE to have Libratus play itself for thousands of hours until it converges (or comes close) to the Nash Equilbirum for HUNL. It would really improve the level of human players, in much the same way that chess engines have vastly improved the play of chess professionals.
So what would happen if you pitted this algorithm against itself? Would it basically never converge and just have the strategies oscillating around equilibrium?
i am just saying that the raw computing power that is necessary to solve HUNL is just not there yet. As the video says, there are like 10^160 scenarios and therefore a huge number of game trees each with different possible bet sizings etc.
Finding the Nash equilibrium for Texas Holdem is like finding the perfect strategy (God's hand) for chess or Go; it's impossible with a computer smaller than the observable universe. But I imagine it should be possible to go far beyond the best human player, albeit the result is less apparent since victory is still dependent on chance.
10^160 poker combinations. "More combinations than there are atoms in the universe." This is true, but it's also a gross understatement. If you created a copy of the universe for every atom in our universe, *then* counted the number of atoms in all of them combined you'd be far closer.
Does it learn from data specific to each player or that specific aggregate group and their various learn exploits? How would this scale if it was just a station game at a casino with no hand large hand history for whoever sat at the machine? Is it optimising for poker in general? or against the specific player and their data?
Not really, people can't afford to make those kinds of bots...atleast poker players can't and wealthy businessman wouldn't bother as they play for fun..also NL texas holdem is only 1 form of poker, there are far more complex game like PLO(Pot limit omaha) which is theoritically impossible to solve.
You live in a naive reality if you believe people with the financial ability to employ this technology in order to cheat ordinary folks out of money " wouldn't bother" . Go watch some video on how technology is used to provide fractional edge to traders in the stock market. The rich will pursue any, ANY, method to extract money from people. I can't emphasize the word " any " quite enough.
torgo4ever Think about it for a second, how many multi millionaires play poker ? Prob less than couple hundred, then they have to have the motive to cheat, which maybe let's give a optimistic guess around a dozen, then they gotta find someone to make such kind of technology and lastly even somehow couple people manage to do it they gotta worry about not getting banned or variations in formats being bought in which drowns millions of dollars which doesn't sound like a great businesses plan, see the flaw in your logic ? If they really wanna cheat they will find more reliable and cheaper scams.
torgo4ever And lastly this would only be at the highest stakes which only few dozen even play on weekends....99.95 poker population doesn't get effected by this...
The way to beat Libratus is by not giving it enough hands or information to figure you out. You have to beat it quick and be almost random. Machines can't make sense out of utter randomness.
Yeah it is learning the players and what the players do, not what all players do. The same way completely new players can defeat seasoned players simply because they don't understand what they have or what the other person is pretending to have. Most professional players play the math/odds game. They bet certain amounts based on their cards to get payouts proportionately. The supercomputer outmaths them.
These players and future players whether pro or amateur should worry about AI. This is like 60 million minds vs your own mind. But AI should be limited and controlled by the Government. We can't let companies filter out all of the humans from their jobs. President Biden should come up with new laws that limit AI being in all jobs. AI is good but if you apply it to teaching in schools then you are going to fire all teachers. You apply that to Nursing then you replace all of the human nurses at the hospitals here in the USA. This technology is dangerous. This is nothing but "revenge" of the nerds and revenge of the ugly people kicking everyone out of the workforce. This is a digital Oligarchy even though people right now don't see it. They will see it once "the cop pulls them over". Either AI gets limited or you will be visiting Goodwill, getting Food stamps, and visiting food banks when it comes to assistance. May 4, 2021 [Tuesday].
9:16 If I may add a slightly more grim approach. (First of all, I want to say that I think tool development is one of the best things to help human kind overall) But, the argument that he uses about having AI helping negotiations, by that point in time (and maybe, unfortunetly, it willl always happen), the "boss" will always have a better tool than the employee, specially for big corporations. Maybe progress will hault and become accesible, but we often have to think about the implications of our technological improvements.
Was it adjusting on its own when they were finding hoes to exploit or were they actually going in and tweaking the programming manually? I think the AI team were adjusting it's play themselves and not the AI doing it all by itself.
It was adjusting on its own, from what I've gathered this bot runs on a neural network, seeing others experiences, and even having my own with Neural networks, they can definitely adjust its strategy.
I see that there is more and more videos on the Internet about how AI defeats people in various games. Let him beat me in washing the floor or in winning a discussion with a pissed off wife then we will talk. What is this poker game in which you do not have physical contact with an opponent? Is it still poker? Without this most important thing? I don't speak english very well. Sorry about that.
AI can't wash the floor or win a discussion better than humans, but it's going to get there slowly but surely! With poker, there's live poker and online poker. Live poker is where you do have physical contact with an opponent, but pretty much all professional poker players play lots of online poker as well.
The comment at the end that AI is always helping mankind is what so many scientists/futurists are worried about. Programming AI with that kind of amorphous, subjective criteria is beyond difficult. Awesome story here, regardless.
These are not the best players in the world. They don't have much information about their success on the internet. They're playing 1500 hands a day which is an enormous amount, with 10 second shot clocks. Under those conditions, an OK professional player isn't going to be playing anywhere near the best. These players admitted to playing terrible strategies just hoping to find a weakness in the bot, rather than playing their own best strategies. The robot team is overstating their accomplishment here. I'd like to see this bot challenge the actual known best players in the world.
It would be intereesting to see a total rookie being told what to do by the program playing against a pro. I wonder if the pros could read the rookies.
lol, he's giving a lecture there where the tables are empty, literally no players are playing around 7 am, I'm probably one of the weirdos playing lmao
Actually, heads up is harder than a full table. You have the luxury of waiting for cards on a full table. However, heads up, you are blinded in every hand. You are forced to outplay your opponent without good cards.
Faster game does not mean a harder game. More players means more variables, both known and unknown. There are other videos about this AI that explains that it doesn't do well against more than one opponent ....yet.
More players the more variables. For the computer playing heads-up is far easier. Example, pocket aces heads up is far stronger than it would be in a 10 handed game.
No shit? A machine that is designed to calculate odds can beat a human which can make mistakes, want to push the limits, and experience fatigue? No fucking shit? Next you are going to tell me the sky is blue, and grass is green.
