Not reading through all of this, but I'd like this question answered:

Did Libratus REALLY beat humans OR

Did it have unlimited time for decisions?

Did it use a **** ton of bet sizings that humans cannot use?

That's completely unfair and changing the game in a way that is to the complete advantage of the computer. I mean obviously any endeavor computers are better at if we give them unlimited time where human brains will eventually fail and have them randomize things to confuse people, but what the real accomplishment is is beating humans within human constraints and until that happens, i.e:

Libratus is given 15 seconds to act and a certain amount of time bank. Libratus uses 6-8 sizes. I don't really see it as a "victory."

And it won the best paper award at NIPS!
"Safe and Nested Subgame Solving for Imperfect-Information Games"
https://nips.cc/Conferences/2017/Awards

But hmm, it seems like unsafe "subgame solving" was fine all along:
"Nevertheless, in practice unsafe solving achieves strong performance and exhibits low exploitability, particularly in large games."
https://arxiv.org/pdf/1805.08195.pdf

So, errr, the major "breakthrough" was that it was Nested!

"The key breakthrough that led to superhuman performance was nested solving, in which the agent repeatedly calculates a finer-grained strategy in real time (for just a portion of the full game) as play proceeds down the game tree"

The results of the Modicum agent described in the new paper seem very impressive, but that is a separate topic from Libratus' success. It is also based on a very similar idea to what I proposed last year and strangely my paper was not cited http://nebula.wsimg.com/d32f48e088b2...&alloworigin=1
(I admit the results in my paper are terrible, but not much more that was possible with the time and resources available, and that also doesn't detract from the general approach.)

Yes


No


The game is called no-limit for a reason.

I think they should make a rematch with isildur playing, it would be sick lol

fyp

As a human you want to be able to use any size you want, which btw there are no bet sizings a human can't use. Using weird sizing abstractions was really the only way the humans had a chance of exploiting the bot.

So when Doug said it would tank for 5-10 minutes on certain rivers, he was wrong? What tournament or current poker room allows such behaviors?

How many people really play it no limit style? Nobody, there are probably 6-8 bet sizings you need to get down to be good and beyond that maybe 12 you really want to master.

You can program computers to do all sorts of things that confuse human beings and make them impossible to track, bet sizings is one of those as if Doug is correct and it used 27 different bet sizes, there is very little EV advantage it gained apart from adding an extra layer of confusion for humans or if it did gain EV, it was long term EV from slightly improved bets.

The more I read about this challenge, the more I realize how flawed it was. Especially the no hud thing, the bet sizings and time make it very questionable, but no hud....lol that's not even close to beating humans if that was the case.

A fairer challenge would've been:

1.Agreed upon time limits

2.Agreed upon bet sizings

3.A HUD for humans

Our minds can only handle so much data, we might decide between 1/4, 1/3 and 2/5, not between 2/17, 3/19, 4/23, 5/24, 6/27 etc.

This really only benefits a perfect algorithm and adds a non competitive layer of confusion that has very little to do with poker and more to do with the fact that humans would need a team to begin to understand the bet sizings and why they're done v a computer which naturally has that memory and randomizer space.

The humans were also able to take 5-10 minutes, so the time limits were fair.

That's just handicapping the computer for no good reason. It would be like someone playing hu vs Isildur in 2010 and crying about him having 3x pot overbets in his arsenal because "it's not fair that he gets to use bet-sizes that I'm not accustomed to".

If you allow a hud then it's not a computer vs human match, it's a computer vs human+computer match.

So you're saying a computer has an advantage over humans because it can use more refined strategies? That was basically the point of the whole experiment.

I think in general you are missing the point of this test. It wasn't to decide who is the champion if you pit a computer vs a human in regular poker settings. It was to demonstrate that computers are getting closer and closer to finding a GTO solution to poker, and when they employ that strategy then humans get crushed because we play so far from GTO (even the best of us). The time limits are immaterial and are only constrained by cpu power. Give it a few years and we'll have machines that will take milliseconds to make the decisions that took 15 minutes in this match.

Just a quick addendum to Wolfram's excellent post.

7OAD, this challenge occurred well over a year ago. This thread was created so people interested in the challenge could discuss it while it was on-going (and the inevitable post-mortems). The discussion naturally died down a long time ago since the challenge ENDED over a year ago.

If you are interested in the details of the challenge, I suggest skimming this thread. In addition I believe that "popular" as well as "academic" articles have been written on this challenge that you can probably find if you are interested.

Humans won't take 5-10 minutes though, their time has a value on it, they can't afford to spend a year playing a bot because they want to play each situation perfectly, there is a good enough element in this that the computer is not previewed to. A computer has nothing to do but that.

Bad comparison, there is a huge difference between going from playing with 6-8 sizings, to 12-14 vs up to 27 which is what the computer had. The problem isn't facing a few different sizes, it's when it begins to use 2/9 instead of 1/5 and at times 4/38 because it can and you do not have enough info to understand why. This acts only to the advantage of the computer, but it doesn't prove anything apart from confusing humans is really easy and our brains are not as competent as computers.

Haha now we go from "oh big deal it's using many sizings, it's called NO LIMIT hold em" to "woah, you can't just use a HUD, that would be unfair!" How is it fair that a computer has unlimited memory and can use all sorts of NON poker advantages like sizings and unlimited time, but a human can't use basic software?

This test doesn't really prove that, it more so proves:

Under perfect circumstances when we tilt the favor towards computers, they can beat humans.

In the same way if you gave computers from 1940 a day to run calculations, they probably could've beaten a chess grand master back then too.

Obviously computers will be able to beat humans eventually, hell they are now with the script kiddies in EE, but the real question is that if they're playing poker and not "we use unlimited time, have unfair sizings humans aren't used to and then take away software that allows them to compete."

If this was presented in the light it happened, it would've been an interesting step, but instead it's actually shrouded in half truths that it takes a more investigative look to fully understand what happened and it's not even close to how it's presented as humans bumbling their way through real poker with real poker constraints and using regular poker equipment (HUDS).

In actuality, the real lesson to take away is that knowing Moore's law, computers will likely soon be able to beat humans within the realistic match terms as they can currently beat them by tilting things in their favor (slightly to some).

2p2 is such a weird place. A post as clueless as this bump brought out someone as expert as Sam G...

I loved following the Brains vs AI challenge and am always optimistic when this thread gets bumped that there will be updates on the poker AI landscape.

If anyone were able to give a really clear and accurate explanation of what set Libratus apart from other AIs, that would probably help the confusion. Popular articles are written by people who are so non-technical that their rewording of what they half-understood from experts is often simply false whereas the academic articles are so dense and jargon-infested that I certainly really struggle with them, even despite being at least somewhat technically familiar with the topic.

Will do that in a bit, but my main point is that this was presented as "Computers beat humans" and that's all there is to it, like in Chess where it's 1 on 1, the time constraints are the same and it's just software v humans.

However, poker being a far more complex game, that isn't the entire truth. The humans were under unfair constraints when you factor what computers actually are, so I would hardly call this some massive win for bots.

If they want to prove me wrong, just do it again:
1.Humans get a HUD
2.Bet sizings are constrained
3.There is a time limit to action

If the computers are better, they would do well to make their lives harder and see what needs to happen in order for humans to currently beat them, just like chess sims have levels of difficulty that are 1200-1500-1800-2100-2400+ and beyond GM.

7OAD, it is my understanding that the point of the poker AI was to demonstrate the effectiveness of depth-limited solving in imperfect-information games by building a master-level heads-up no-limit Texas hold’em poker using only a 4-core CPU and 16 GB of memory. It wasn't to prove computers are better than humans at poker, or to develop a chess-like program for humans to play poker against. All of your concerns have nothing to do with this challenge and shouldn't be polluting this thread.

This is laughably wrong by many orders of magnitude. The fastest supercomputer in 1993 computed 125 GFLOPS and the fastest in 2016 computed 93 PFLOPS--in other words, the 2016 supercomputer was ~1 million times faster. Computers in the 1940s were still using vacuum tubes and weren't even in the same universe of performance as a modern computer.

Why stop at a hud? Why not just give the human their own copy of libratus to consult on the moves? Wouldn't that be even more fair?

Out of all your points this one is the silliest. Deciding the correct bet size is a part of the game. An expert player will use better and more varied bet sizes than a beginner. That's part of realizing the edge for the expert. If the computer is able to use more bet sizes than a human then that is proving that the computer is smarter at that aspect of poker. All your talk about it being "unfair" or "confusing" is just gibberish.

But fine, lets limit the bet sizes. That experiment was already run with a lhe bot called cepheus (much earlier than libratus). It crushes humans.

Fine, next time lets use shot clocks at a more reasonable limit than 15 minutes. Trust me, it won't make a difference.

This test was exactly like the Deep Blue match vs Kasparov. It was to demonstrate that a computer is able to beat a human. The funny thing is it turns out it isn't just able to, it demolished all of its opponents beyond any reasonable doubt.

That fact seems to have a weird effect on some people, who get stuck in cognitive dissonance and denial. Tribalism is a hell of a thing.

i agree with Wolfram, giving humans a hud, independent of its definition, is sort of denying the purpose of the experiment, which was brains vs AI. limiting bet sizes works against team humans, not in favor of.

it was def a blunder not to set a time limit on action. there are two sides to this; the first is that if they had agreed to smth reasonable - say 20sec per decision point and an additional 5min time bank that auto-refills every 100 hands - team AI either would've faced a bigger challenge in preparation of the experiment or they would have had to use nested subgame solving less often or use less precise results, leading to libby playing a worse strategy overall. the other side is that team humans expected the days to be severely shorter than what they turned out to be, which caused fatigue issues. in addition, the waiting time in b/w hands was often so long that it disrupted the flow of the game. imagine you open pre and get called, flop goes check check, and then your opponent goes in a 1min tank... and checks. it seems likely that a more reasonable time limit would have led to team human playing better overall. it's a pity that they didn't foresee that team AI were likely to add nested subgame solving compared to the first challenge vs claudico, as it was kind of obvious given the emerging literature at that time, and that public solvers who were available back then used subgame solving techniques as well.

The point was that if you want to give unlimited constraints like Libratus had, the computer's edge will become even bigger.

Rightttt giving the humans software ALL humans use=giving them software no humans use. If we want to prove AI is smarter, let's just create super robots that genocide humanity, oh wait looking at things at black and white and taking things to their logical extreme isn't an argument? Well neither is suggesting that using a HUD is the same as using a super computer.

Most websites generally have 3-4 sizes, the best players are using more, is it by a magnitude of 7x+, from what I've seen, no.

I don't think it's gibberish, you just obviously misunderstand it in the same way you desire to strawman my position of using a HUD. IF using unlimited sizes was the perfect solution, why did it stop with 27? Why didn't it go to 200, why not 2000? It's almost like once you understand the EV will likely be similar, you can add a layer of confusion that human minds cannot solve without a ready answer and use this cheapo tactic to add an extra advantage to the computer.

Ok computers are already better at limit, great, why was this done then? We could merely extrapolate the data at that point, better at limit=better at no limit. Oh wait, there was an attempt to prove something here and only people who have a limited view of what poker is think that thing was really proven.

If it won't make a difference, then add it, simple.

Lmfao, not even close, Deep Blue was computer v man within the chess constraints everyone faces, not within some constraints that are outside the boundaries of normal chess.

1.Getting better humans would've been nice

2.Using a HUD which is how humans are used to playing would've been fairer

3.Not allowing a machine who's only function is to do this thing unlimited time v humans who don't want to play the rest of their lives so that they can beat a machine and likely earn what their hourly is regularly v the bot

4.Agreeing on sizes or giving the computer sizes and hand history to the humans before would've also made it fairer

There's no tribalism, I literally said that computers will obviously be better no matter the constraints within a matter of time (although reading what I said isn't your strength), I am merely saying that isn't what happened here. This was the equivalent of giving Kasparov an hour and Deep Blue 10 hours, it was so obviously tilted towards the computer any non anti human tribalist like yourself (see, two people can throw out baseless accusations) could see it.

Your whole idea is that the computer is clearly better

Ok great, let's see how much better it is when we start equalizing the playing field. If it's SO much better, there should be NO problem. Hell, people like you can call it giving the humans an advantage (as if there's no way to EVER advantage the computer) and we'll still get a better show than this was.

Chess already proved that computers are smarter than humans, this experiment is dropping a bowling ball and a bag of feathers in 2018, it's been done.

Giving a HUD allows humans to make adjustments like the computer can based on human strategy, denying them the HUD just means it's computer v brain, as I said, been done. We already definitively know computers are better than our brains.

The programmers and people who ran this were smart, they knew what they were doing, the poker players had nothing to lose regardless.

But hey, if the computer is so much better than the humans, let's stack things in our favor and see how it does by having the best humans play with a HUD, giving them computer hand histories, having time limits and constraining bet sizes.

My bad for discussing the rules of a challenge as if they had no impact, in which case if they didn't, they should have no problem giving us humans a chance to win by stacking things in our favor.

Sorry for polluting this dead thread, I shouldn't be interrupting high quality discussions on if more rake is better, what meme Doug Polk posted and the other gossip threads.

I never said using huds was the same as using a supercomputer. I was pointing out that once you allow software assisted play, you are no longer measuring man vs machine, you're measuring machine assisted man vs machine.

More illogical gibberish. The point wasn't to prove AI is smarter than people. The point was to show that AI plays poker better than people.

I didn't suggest that.

Every site allows you to bet any amount from zero to the size of your stack. Because that's the rules of the game.

Your position isn't that complicated. And you don't know what strawman means.

I can't believe that this is the hill you choose to die on. I have no idea what sizes are optimal, neither do you. The computer is trying to figure that out. Why would you artificially limit its strategic ability?

To see if computers are better at No-limit. You know, the game where you can bet any amount you want.

You're the one that wants to limit the available bet sizes because your view of NL is too limited.

Wrong. The humans got the same amount of time as libratus did. And they were allowed to use the same bet sizes.

Wrong. Computers are better at chess than humans. They are not generally smarter (yet).

Wrong again. Libratus doesn't adjust based on the human's strategy.

You swear up and down that you are not being tribalistic, but then you keep making emotional statements like this throughout your post. You keep making up convoluted reasons for why this was unfair instead of viewing it as what it was: a demonstration that expert players are playing far from GTO, and that we are very close to solving for GTO in poker. Which will have great implications for machine learning in other fields.

It wasn't a dick measuring contest.

Stop replying to the brainlet troll, you explained the flaws in his logic, he is never ever going to listen. I mean he is arguing that having fixed bet sizes would advantage humans ffs.

Yeah, this was fun, but I'm done now

In the next Human vs Computer chess match, the computer shouldn't be allowed to move its knights, because that's unfair to the humans and I'm not used to moving my knights.

Congratulations to everybody involved for helping to destroy the game of poker and the livelihoods of those who depend on it!

Question: are AI developers less, equally, or more responsible for the impending demise of poker than Black Friday was?

less, but i'm gonna have to ask you to show your work on an "impending demise"

The world doesn't owe you a living. Either you adapt or you die.

You are a fool if you think black friday will be a bigger contributor to the death of online poker than the technological advancement will be. After it's dead, and we look for the cause, black friday won't even be considered. That was nothing. In fact, we won't even have to look for the cause, as it will be blatently obvious to everyone who doesn't have their eyes completely shut. Solvers are and will be the death of online poker. Still pointless to blame the people that happened to be the software developers. Someone would have done it regardless.

Many players realize this already, and many of the sites do too. Pokerstars are ruining their reputation, in pursuit of short term gains. It's not a stupid decision, because there is no long term.

(the rest of the post is somewhat of a derail)
Some sites do not realize however. When pressed on the issue on JoeIngram's podcast, Phil Galfond admitted that bots (and I'm not talking about ****ty bots that you can beat) are the main threat to online poker, and claimed to be taking it extremely seriously. He was invited on the podcast to discuss certain decisions that RunItOncePoker had made (such as banning huds, and having anonymous tables), that he outlined in a blog post. I read the full blog post and bots were mentioned once: They did not want to allow HUDs partly because they fear that recreational players would see HUDs and confuse them with bots.

Maybe I'm being unfair to Phil Galfond by arguing that he doesn't take the issue of illegal software or bots seriously, based on not mentioning the issue in his posts. Maybe he doesn't want to talk about the issue publicly, not revealing his strategies to deal with it. After all, this was in essence Trumps foreign policy, and that seems to be going well so far.

The real nail in the coffin is the fact RunItOncePoker are going for anonymous tables. This is heaven for someone running a bot or using illegal software. The number one mechanism that we have for catching people is player reporting. You need to be good at the game to be able to spot someone who is a potential cheater. Random staff members inspecting tables won't be able to. If the sites doesn't ban a suspected account, at least the regs can avoid him. With anonymous tables all of this is gone. You can't report someone because you don't know who they are, and you can't avoid them for the same reason. A good example is the partypoker heads up lobby. The anonymous tables are filled with bots, but in the normal tables you at least have a chance at a fair game. The pokerstars games are the cleanest, because everyone knows the screenname of every reg, and who they are, and their history on the site. If you play on obscure sites, you are much more likely to get cheated than most of you probably realize. I think that MTTs are reasonably safe as of yet, but all other NL formats are at risk. Of course it's harder to have a 6max bot than a heads up bot, and harder at higher stacksizes than at lower ones. But don't be fooled, 6max bots are very possible. And even if you think it's not possible, people can still have assistance in heads up pots easily. I'm not qualified to speak about bots in for example PLO, but I'm confident in what I have said. The future of online poker is short and bleak.