Libratus Poker AI beats humans for $1.76mln - how was the computer made? - Page 2

On February 03 2017 01:08 Stroggoz wrote: hmm having got my math degree last year all i understand is about 1/3 of that paragraph. how this mirroring thing works is this: if dong gets dealt cards against computer, jason gets dealt the cards that the computer got dealt against dong. So in other words if dong runs good, jason will be on the other end of that run good. Stop saying jason is a fish lol he is prob worst on the team but by a very slim margin and certainly one of the best hunl players in the world. Dong ran hot, Jason ran bad. Their combined team results were almost the same as the other team.

Not quite though, anytime they went AI they got their EV back. The only thing that was mirror'ed was the opposing hands, so there were coolers in some situations and reverse situations.

Pretty telling that Les was the only losing player against Claudico (-80k) when humans beat it for overall 800k, including his loss.

On February 03 2017 05:02 Baalim wrote: The no all-in variance and mirror probably ironically increased variance since lets say you get allin vs libratus KK v AA, you lose 80% stack, but if your mirror buddy slowplay AA on a xxxxK board you lose your whole stack.

This is true. Equity chops decrease the effectiveness of mirroring for basically the reasons you stated. They still reduce overall variance together more than either alone but it isnt that much, like 20% reduction in the stdev

On February 03 2017 06:27 NewbSaibot wrote: Would a pot limit version of holdem be good or bad for the bot?

Bot would have higher nl winrate however it should be able to produce a strong PL strategy more easily5

The bot only tanked the turn all other actions were instant. It had its pre and flop strategy completely mappes out in a database essential. It had a tool for recalculating its strategy on the turn so it would tank then act instantly on river again. The pros all said turn was the bot's strongest street (and i think the hardest street in general personally)

Can we also toss jason les out of future competitions? Get wcgrider and jungleman. I dont even know who donger kim is. What ever happened to those two dang brothers that use to play nosebleeds i would rather have them in, people that have been playing since the manifestiation of online poker boom

On February 03 2017 06:27 NewbSaibot wrote: Would a pot limit version of holdem be good or bad for the bot?

it would be EXTREMELY good, limiting the sizing limits the complexity of the code, it doesnt have to react to overbets

On February 03 2017 23:42 JohnnyBologna wrote: Can we also toss jason les out of future competitions? Get wcgrider and jungleman. I dont even know who donger kim is. What ever happened to those two dang brothers that use to play nosebleeds i would rather have them in, people that have been playing since the manifestiation of online poker boom

rofl.... all these players would destroy the Dang brothers come on this isnt 2008.


The key to defeating in most cases AI is to find a chink in the armor, not to play straight up, I think the most effective way is to try exploit him in very unhortodox ways, like betting 10x pot, betting 0.1pot, open raising to 20bb and shit like that, trying to find holes in the adaptability, the previous match with the other AI made this clear.

In general AI works like that watch the Starcraft AI and see how easy they are defeated when they are exploited and how stupid would be to play them straight up.

i wonder why old pros who were considered high level now suck vs the new guys? is it more complacency than anything? the game is not physical unlike mma, boxing which youth can change the dynamics, while poker is more cerebral like chess which older people dont get their asses kicked by younger players. or just the sheer scale of smart people unlike the smaller pools before?

libratus uses a meta algorithm which identifies ways its being exploited and adjusts, basically making it impossible to find just a few flaws in the bot and exploiting them.

On February 04 2017 05:29 Stroggoz wrote: libratus uses a meta algorithm which identifies ways its being exploited and adjusts, basically making it impossible to find just a few flaws in the bot and exploiting them.

So did Claudico, all advanced A.I. tries to adapt and its just not a script it reads input too, but how good they do it is the thing.

Dong Kim had by far the best results against both bots, he seems to understand better whats going on with the AI

On February 04 2017 04:17 whamm! wrote: i wonder why old pros who were considered high level now suck vs the new guys? is it more complacency than anything? the game is not physical unlike mma, boxing which youth can change the dynamics, while poker is more cerebral like chess which older people dont get their asses kicked by younger players. or just the sheer scale of smart people unlike the smaller pools before?

because they quit online poker, the Dang brothers quit poker to become entrepreneurs, Dwan went to live (and obv his skill quickly degrades), EmpireMaker, Hallingool, caracas etc all of them lost interest and stopped studing the game, but there are some who have lasted a looooong time at the top like Phil Galfond or Daniel Cates

Wcg, jungle, galfond, ivey would be my picks

Neilly, Spitfiree, Doc[Lem0n] and Floofy would be my picks.

No ai can even begin to comprehend what will be coming its way.
The common denominator here is persistence. These 4 would outwork the computer.

Def not ivey or galfond. All 4 of rhoae guys were significantly better. Fortheswarm was considered top3 for a long time.

It would be jungle, basjkorven (he plays most like the ai), OTB, (maybe sauce, wcgrider or kabyDF if they put significant time into brushing up there hunl games. Mummiposs and linuslove would be good options. Have no problem with dong and swarm playing again.

Prob is bot smashed at 10bb/100. Dont think any reg is that far above the group that played at all. Maybe uf you clone jungle 3 times

Last think id like to see is jungle vs the bot straight up. He ia an asbergers savant with a mind that can think like the ai and find those cracks more than any other. Id call him a prodigy. I wouldnt say that about all of the last 4 who playd. Libratus is likely unbeatable though and only getting better.

Much like alphago who beat go 10 years aheas of schedule. A far more complex came than chess. Was thought of as the holy grail of ai. It beat pro seedol 5/4 (the game the human one was amazing) mod 2016. Alphago has further improved and popped up on korean go servers mid january 2017. It was already much improved. It 3-0d the consensus #1 in the world and amasses a 60-0 record. More go possible moves than atoms in the universe. Far larger than hlholdem or plo.

On February 04 2017 06:45 Baalim wrote: Show nested quote + On February 04 2017 05:29 Stroggoz wrote: libratus uses a meta algorithm which identifies ways its being exploited and adjusts, basically making it impossible to find just a few flaws in the bot and exploiting them. So did Claudico, all advanced A.I. tries to adapt and its just not a script it reads input too, but how good they do it is the thing. Dong Kim had by far the best results against both bots, he seems to understand better whats going on with the AI

Actually Dong and Les were the 2 humans who did the worst against the bot last time (and were the 2 to repeat this challenge). Dong was up 70k, Les was down 80k. Bjorn li and our boy Doug Polk were up a combined 750k vs Claudico.

On February 04 2017 22:25 Endo wrote: Show nested quote + On February 04 2017 06:45 Baalim wrote:   On February 04 2017 05:29 Stroggoz wrote: libratus uses a meta algorithm which identifies ways its being exploited and adjusts, basically making it impossible to find just a few flaws in the bot and exploiting them. So did Claudico, all advanced A.I. tries to adapt and its just not a script it reads input too, but how good they do it is the thing. Dong Kim had by far the best results against both bots, he seems to understand better whats going on with the AI Actually Dong and Les were the 2 humans who did the worst against the bot last time (and were the 2 to repeat this challenge). Dong was up 70k, Les was down 80k. Bjorn li and our boy Doug Polk were up a combined 750k vs Claudico.

is there a link for the results vs Claudico ?

https://en.wikipedia.org/wiki/Claudico

les' results are minorly skewed vs claudico due to him blasting off stacks the last few hours trying to finish up

On February 05 2017 02:18 PoorUser wrote: les' results are minorly skewed vs claudico due to him blasting off stacks the last few hours trying to finish up

was the stack blasting the difference between a statistical tie?