Slashdot Mirror
An AI Is Finally Trouncing The World's Best Poker Players (cmu.edu)
Halfway through the "Brains vs. AI" poker competition, an AI named Libratus is trouncing its human opponents, who are four of the world's top professional players.
One of the pros, Jimmy Chou, said he and his colleagues initially underestimated Libratus, but have come to regard it as one tough player. "The bot gets better and better every day," Chou said. "It's like a tougher version of us"... Chou said he and the other pros have shared notes and tips each day, looking for weaknesses they can each exploit. "The first couple of days, we had high hopes," Chou said. "But every time we find a weakness, it learns from us and the weakness disappears the next day."
By Saturday, the AI had amassed a lead of $693,531 after 56,732 hands in the 120,000-hand match (which is being livestreamed by the Rivers Casino on Twitch). "I'm feeling good," said Tuomas Sandholm, the computer science professor at Carnegie Mellon who co-created the AI. "The algorithms are performing great. They're better at solving strategy ahead of time, better at driving strategy during play and better at improving strategy on the fly."
By Saturday, the AI had amassed a lead of $693,531 after 56,732 hands in the 120,000-hand match (which is being livestreamed by the Rivers Casino on Twitch). "I'm feeling good," said Tuomas Sandholm, the computer science professor at Carnegie Mellon who co-created the AI. "The algorithms are performing great. They're better at solving strategy ahead of time, better at driving strategy during play and better at improving strategy on the fly."
Cue the goal post shifting.
Those who do not learn from commit history are doomed to regress it.
Without the psychological aspect of staring your opponents in the face it's just a calculation of odds. Takes most of what makes poker poker out of the equation.
Thank God. All the poker playing jobs are safe for now.
Unlike with games like Chess (best moves can be precisely calculated) and Backgammon (simple probabilities), Poker requires adapting to human behavior, indeed varying your play depending on what you learn about your opponent. The techniques are going to be applicable to a wide range of situations. For instance, I will go so far as to claim that we will shortly be wise to use an AI to advise us on investment decisions. (In the past, the computer has been used for speed, and reacting to subtle market signals, but not so much for long term investment planning.)
The next challenge is going to be independent learning. I believe human experts still supervise the learning process of all the best AIs. Once the need for the human adviser goes away, AIs are literally going to be everywhere. Your phone AI will recognize and react to your current mental state, as well as help you overcome everyday problems. The AI in your fridge could become a huge help in keeping you compliant with your diet plans.
Should it fold if the opponents bet outweighs the probability?
If the AI uses probabilities, it itself becomes predictable and therefore trivial to beat.
The game is setup to make you lose if you only play good hands, so there is no playing safe in poker.
Probabilities in poker are nearly meaningless if you play against even half-decent amateurs.
Slashdot social media options: AIM, ICQ, Yahoo, Jabber and Mobile Text. Why no MySpace?
Isn't card counting such an effective way to win at poker that casinos ban it? And shouldn't a poker AI count cards pretty much by default? So no wonder it's effective.
Heads-up (2 player) Texas Holdem is not the most commonly played version of poker.
Most people play Texas Holdem in groups of 6 or 9 players. Working out an optimal strategy to beat multiple opponents is a LOT harder than beating a single player. We may have a dominant heads-up poker AI soon, but I would expect it to take several more years for a dominant multi-player to be created.
Uh, several more years? Allow me to quote one of the poker players:
"...every time we find a weakness, it learns from us and the weakness disappears the next day."
Let's not underestimate the power of learning at damn near an exponential rate. I expect an AI multi-player tournament next year to crush human opponents. How quickly do you think an AI supercomputer could process every single hand of play that a professional poker player has ever made in their life to analyze and exploit every weakness to be able to predict behavioral patterns with great accuracy? Lather, rinse and repeat for the top dozen poker players in the world. How to get AI to beat humans in a game of finite limits and statistical values is not exactly a mystery.
The largest mistake mankind could make is underestimating the speed at which AI will prove it can do a lot of things better, faster, and more accurately than any human could ever do. Underestimating that speed will greatly reduce our ability to properly prepare for a world of unemployable humans.
tl;dr Poker isn't dead, yet.
AI beating humans at a game is merely a beta test. The real application will feed unending greed, which will never die.
If it was real AI it would be self-aware. And the computer would be wearing mirrored sunglasses.
I would use it to earn money.
What if...wha...oh. What if Trump's entire campaign is driven by an AI?
"Now dispute crowd sizes. Celebrities and news media will double down on their loud Hollywood mouths. It matters little in the short run but builds background distrust of them. This will be used in 4 years."
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
Uh, several more years?
Yes, years. Libratus uses 16 Terabytes of memory for just a 2 player game. The size of the game tree increases by at least a factor of 1000 when moving up to just 3 players. That's significantly more memory. Also the computations themselves take much longer when there's more than 2 players as something called "card removal" comes into effect.
Not so long ago, people assumed that a world class Go playing computer would also take years to create, and all of of a sudden there was AlphaGo beating them.
The solution is to combine probability calculations with unpredictable behavior.
Real intelligence is also just computation.
So, will Libratus play against DeepStack (from the University of Alberta etc.), which also claims to be able to beat professional level humans...?
DeepStack: Expert-Level Artificial Intelligence in No-Limit Poker
DeepStack becomes the first computer program to beat professional poker players in heads-up no-limit Texas hold'em
https://arxiv.org/abs/1701.017...
No, we did.
Faster! Faster! Faster would be better!
Re:
"...every time we find a weakness, it learns from us and the weakness disappears the next day."
Let's not underestimate the power of learning at damn near an exponential rate.
It does not look like AI learning at exponential time. It looks like nightly patches to a program to remove discovered exploits. Let us wait until "the weakness disappears immediately without any human intervention".
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra
Just don't hook it up to any missile command system.
Yes, years. Libratus uses 16 Terabytes of memory for just a 2 player game. The size of the game tree increases by at least a factor of 1000 when moving up to just 3 players.
This is why better algorithms are almost always a bigger factor than increased computing power when solving these problems. They won't solve playing against 3+ players with more RAM, they will solve it with better algorithms. By some cases algorithmic improvements can be 43 times more important than computing power improvements.
Considering this AI is already dealing with unknown information, I doubt the size of the "game tree" increases by the factor you cited as you add more players.
-- All that is necessary for the triumph of evil is that good men do nothing. -- Edmund Burke
“Hillary Clinton I think is a terrific woman,” he told Greta Van Susteren. “I am biased because I have known her for years. I live in New York. She lives in New York. I really like her and her husband both a lot"
https://www.twitch.tv/libratus...
It's in the damn article if you read it.
since most stock "advisors" are actually salesmen who get commissioned for pushing certain stocks (and IIRC they'll even tell you that, albeit with the weaselistly words possible).
What worries me is that this is another case of increased efficiency in our economy. Inefficiency is a huge part of what keeps it all going. Now, it's certainly true that it's ridiculous to pay people to break windows to employ window makers; but I'm not convinced we're going to have anything for those window makers to do if tomorrow the number of broken windows drops 90%. We either need a solution to keep the money and resources flowing or it'll collect at the top like it always has.
Hi! I make Firefox Plug-ins. Check 'em out @ https://addons.mozilla.org/en-US/firefox/addon/youtube-mp3-podcaster/
I've never used an online gambling site, but doesn't the existence of this AI kill off the fairness of these sites?
If a user is running this in his or her basement, wouldn't it pay more to just babysit the AI, acting on all the human-check capchas the sites deploy, and just doing what the AI decides?
This makes online poker effectively gold farming?
Why aren't you encrypting your e-mail?
It's exactly what they are doing, so no, they haven't overlooked it.
Self awareness is only a small part of intelligence. Our brains do a lot of processing that's not self aware.
I think it's fantastic that we're making progress with 'smart pattern matching'. However, I still get disappointed when I read the hardware specs required to do so. Imagine for a moment: 5 humans competing in a highly complex game with several hundred inputs to each player; spatial, acoustic, thermal, temporal etc. The complex task of facial recognition of multiple players, and how that relates to our 'operating system'. Hiding your own emotions, doing the best you can with statistics, the sounds, the environment, etc. The total thermal dissipated power of the brains at the table of 5 players is on the order of 100W? The short term memory capacity on the order of kilobits? To me, this isn't even a 'game' in the sense that there's any chance for competition. No human can (as in gets the opportunity to) analyze the complete history of an individuals games. No human can spend more than ~40W of TDP at any given time, and the memory limit, while not a well defined quantity is certainly much much less than the memory available to the cluster. For this to be a fair game, they should plug the entire system into a 2,000 Calorie battery, and run it till end of each day, with a memory limited to the span of interaction with a group of players. Else it's just an exercise in how many teslas you can afford to throw at a concurrent problem.
Where does it say anything about human intervention? The thing is learning while it sleeps.
Chris Mesterharm
Libratus has a poker tell. His CPU fan speeds up whenever he gets a good hand.
Have gnu, will travel.
This is not true/strong AI,
You keep using that word. I do not think it means what you think it means.
AI has always mean imitating intelligent behavior through clever algorithms. (Almost) no one researching AI is looking for machine consciousness - why would you want that? They're trying to solve real-world problems with engineering solutions. We want a self-driving car, not a self-aware car.
Socialism: a lie told by totalitarians and believed by fools.
If it was real AI it would be self-aware.
No, that just not what "AI" means, any more than "sentient" means self-aware. Science Fiction keeps abusing those terms, but they have mainstream meanings. AI is clever algorithms that imitate intelligent behavior. Which means it could still be wearing mirrored sunglasses.
Socialism: a lie told by totalitarians and believed by fools.
Heads-Up Hold'Em is a completely different game than standard 9 Player Hold'Em. It's not just that there's 8 people to analyze, you also have to analyze the relationships between every player, which is constantly changing with the flow of the game. Player A crushed B who's on tilt and will be extra aggressive, or maybe B's on-tilt the other way scared and easy to push out of hands, or maybe player B will just shut down and not play anything at all. Then long-time friend C sits down which perks B up, and their style completely changes again. Meanwhile Player A is flush with chips and extra loose with the range of hands their playing if they aren't disciplined or cocky and over aggressive, or disciplined and staying balanced.
With every development I have to adjust my play, and so would an AI. Hats of for the AI that can solve all that, but it's long way from the sterile heads-up match they're doing now.
Also most professional players don't post their hand histories, in the way Go players do. The newer pro's are more open to that, but nobody's going to offer up their lifetime hand history to opponents.
hey shut down a program of the foundation
Yea, the program that was using foundation money to buy Hillary the Oval Office.
Aaaand fail. I did write "strong/true AI" and hence you are the one that does not understand the "AI effect". Incidentally, the AI effect proves my point, because there are perfectly good terms for what often is called (non-strong/non-true) AI these days and hence there is zero need to call it AI. Pattern recognition, statistical classification, automation, etc. all far better terms than the entirely misleading unqualified "AI".
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
An unsubstantiated claim at best. At least actual Science claims no such thing.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
These are defined terms. They are not _my_ terms. Look them up before spouting complete nonsense.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Those types of games computers have been beating for years. Given irrational non-statistically valid play simply playing a boring statistical game works really well. Now of course as the M ratio goes up the game gets harder for both computers and humans but for computers faster. Given how easy an M ratio of 50-200 was solved and how small the bots were, solving 1000 or so when players are being irrational shouldn't be hard.
A static system fails to good play. Take the most common situation, computer is in the big blind (i.e. you had a forced blind bet before you were dealt). 1/2 the hands dealt will be less than average. Folding a raise with a substandard hand subsidizes your opponent almost always raising regardless of his cards. So you can't fold. Calling a raise with a substandard hand subsidizes your opponent better hands, he raises when he is good and mucks when he isn't. So you can't call. Reraising with substandard hands makes the whole situation even worse.
There is no static solution against a dynamic opponent in partial information games.
What in your opinion is the difference? How can I tell something that looks intelligent from something that is intelligent?
Of course it does. More and more we are learning the mechanism by which simple brains work. They are reproducible in software.
Let's not underestimate the power of learning at damn near an exponential rate.
Except that you pulled that exponential rate out of your ass. AlphaGo which has by far the best record in self improvement, learns at a linear rate.
There is zero evidence that the brain is magic, even for low values of magic such as "quantum." Actual "Science" does regard the brain as performing computations, massively in parallel.
Ironically, modern neuroscience is beginning to find evidence that your self awareness might consist in large part of your conscious simply being told of decisions that your subconscious has already made. Disturbing thought, isn't it?
Bro, do you even dictionary? It does have a defined meaning and it's not machine consciousness. For fucks sake, that's science fiction and only science fiction. It's not the common meaning of the term. It's not what "AI researchers" research. It's fanwank. Get over it.
1 : a branch of computer science dealing with the simulation of intelligent behavior in computers
2 : the capability of a machine to imitate intelligent human behavior
Socialism: a lie told by totalitarians and believed by fools.
No it does not. Read maybe some actual scientific results? The current scientific of how this works is "we do not know". That is for brains where actual intelligence can be observed. A fruit-fly, for example, cannot be called "intelligent".
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
No, it does not. Actual science at this time says "we have no clue how this works". There is zero evidence either way and that makes the question open. Or have you forgotten that actual intelligence gets observed nowhere else? That alone makes a default to the physicalist explanation exceptionally non-scientific. Or maybe you think consciousness and intelligence are emergent properties of complexity? If so, that would be "magic" right there, because the whole cannot be more than the sum of its parts in physics.
The actual scientific fact at this time is that the question is completely open.
Incidentally, you are wrong about "quantum". The human brain is awash with quantum-effects. They happen all the time in the synapses, and there are about 1,000 trillion of those in a human brain backed very densely. Nobody knows what even tiny deviation in the probability distributions could do.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Heads-up (2 player) Texas Holdem
Came here to say this.
Also the real test is not against a 1 to 9 pro players. The real test is against a mix of really good players and a few who have no idea what they re doing.
Well, both "strong AI" and "true AI" are keywords in Wikipedia. It is defined (simplified) as the ability of a machine to perform "general intelligent action". There is no consciousness requirement, but a "generality" requirement. And that makes all the difference. Strong/true AI is AI that is not specialized for one tiny problem, but can solve general problems.
Whether actual intelligence (whether natural or artificial) is possible without consciousness is an open question and besides the point for the current discussion.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
and bring it to the casino?
Not so long ago, people assumed that a world class Go playing computer would also take years to create, and all of of a sudden there was AlphaGo beating them.
People in the nineties assumed that a world class Go playing computer would take years to create. They were correct.
I'm a minority race. Save your vitriol for white people.
and yet, even with all that unfathomable and incalculable complexity, humans manage to play and win the game against other humans
An AI doesn't have to analyse the game situation perfectly, it just has to do it better than humans. and that's a hell of a lot easier than perfection.
I, for one, will be glad to see the macho bullshit associated with the idiotic practice of poker disappear - there's no pseudo-testosterone in being thrashed by a computer, or in being - at best - a second-rate poker player.
You have never done any game development, it's obvious.
The step from single-player game to multiplayer game is not a simple upgrade, it's a complete shift in everything. It requires a completely different approach, not a refined version of the same approach.
In any non-trivial multiplayer game, the interactions between all the players matter, and the complexity of those is subject to combinatorial explosion. Poker being a relatively low-interaction game will not make this as bad as some others, but beating one person and beating a table of people is not the same system with a little more cycles, it quite possibly requires a different approach altogether.
It will be interesting to see the jump happen, but it is a jump, not a step.
AI beating humans at a game is merely a beta test. The real application will feed unending greed, which will never die.
Greed is a game.
Assorted stuff I do sometimes: Lemuria.org
That's not what exponential growth means. Exponential growth means that the variable is within the exponent; not that the function contains an exponent. Exponential and linear growth are two completely different things.
woosh.
I'm probably at a 1300 ELO in chess, which means I can probably manage to not completely embarrass myself against the chess club president at Podunk High School, but reliably beat anyone who hasn't given the game some relatively serious study. I've read a dozen or so books about openings and endgames, and I keep some chess engines kicking around the smartphone and computer, but I've never had any serious interest in mastery nor any real hope of it. That said, one of the first things one notices about computer analysis of chess openings is that most paths get pruned *very* quickly. With expert play, it is very easy to turn a positional advantage into a material advantage, and this is true for human experts as well. The distribution of chess openings used in master-level play is extremely similar to the computer ranking of those openings, with the exceptions being the more obviously ridiculous things like 1. a4 ... 2. h4 that no computer would ever consider playing. It's also widely said that chess is the game where the winner is the person who made the second-to-last mistake. So while the theoretical number of possible moves in chess is large, in practice the number of viable positions is much smaller, and as long as you're sufficiently clever when evaluating positions and pruning branches then your average desktop computer will be able to completely evaluate all interesting positions up to 20+ ply (20 half-moves, so 10 moves from each player) in advance of a given position in just a few minutes, even without opening tables or endgame tables. The opening is an even worse example. The opening in chess is not just less complex than go, it's actually completely solved, and it's not inconceivable that at some point in the next century or two that the game could be solved completely. So in terms of how the game is actually played, this means that both chess engines and human players will have an "opening book" and not even start evaluating positions for the first 10-20 moves.
Chess games are as sharply decisive as they are because, I believe, most of the pieces can affect most other pieces, and because captures remove pieces from the board, which tends to further increase any advantage. It's fairly easy to evaluate at any given position, and one can prune branches extremely quickly. I don't personally have any idea how to evaluate a Go position; even on small boards I lose without knowing why. However, if we are to have a single definition of complexity, it must be the mathematical one. Go has more potential positions and more decisions, and whether or not it's easier or harder for computers or humans to actually play it is more of a tangential issue. I'm not sure on quite what basis you can dismiss that definition, but even aside from that, perhaps you can provide any example of a way in which Chess is more complex than Go, because your specific example proved rather the opposite point.
Those who advocate genocide deserve every protection afforded by law, and none afforded by common human decency.
But that's a thing of science fiction. No one is seriously working towards that - why would we want it? It's neither what academic nor industry AI researchers are doing.
Socialism: a lie told by totalitarians and believed by fools.
you can still have lots of rounding error like the USA, and you can't even directly vote for the leader, even though most treat their local vote as such.
Ironically, Russia does this part right. While their elections may be invalid for other reasons, their voting system makes more sense. You get a country wide horse race for the President, no rounding, and a separate riding system to ensure local ministers represent you.
Actually, people _are_ working towards that, but there has been very little progress, and hence it does not get reported often. On the other hand, general artificial intelligence even far below what a human moron can do would be extremely helpful. For example, the robotics people would be hugely interested and some other fields too.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
All four heads look like they were photoshopped on. :D
In reality, this is how the US election works for the most part.
When voting, it may say Donald J. Trump and Hillary R Clinton, but you are really voting for electors which vote for the president. It is a system meant to balance the needs of high population areas and low population areas so that it isn't a tyranny of the majority, but for some reason there are many who seem to think tyranny sounds pretty good.
APK likes to ask for responses to the same things over and over. Maybe he just likes the responses?
The US lost because there were no good candidates. I mostly blame the terrible primaries for that though.
APK likes to ask for responses to the same things over and over. Maybe he just likes the responses?
"It is a system meant to balance the needs of high population areas and low population areas so that it isn't a tyranny..." Not quite. Its a system designed to work with two limitations. One was the technology of the time. The second was the uncooperative states; they were offered more voting power than otherwise entitled in order to entice them along. The founders of this system were not fans of it themselves, but offering some states increased voting power was the only way to get those states to commit. Our founding fathers didn't like this system, but it was as close to a democracy as could be created given the players. Personally,I think every vote should be counted, and the electoral college scrapped. It is outdated.