^^ Not much info at all on how the bot works, no open source. At the end of the last competition, both Sam and Noam came into the thread and gave out a bunch of info, hopefully Noam will do the same this time. Wouldn't expect the code to be published

I did.

I did.

I'm not really sure what it is that you are trying to say I have misinterpreted.

Do you think that this:
"Libratus will continuously sharpen its strategy during the Brains vs. AI competition, performing computations with the PSC’s Bridges computer each night while the pros get some shuteye."
...means it must be developing exploitative adjustments (i.e. looking for frequencies or ranges that are not theoretically sound)?
They have run trillions of trials up to now and will be gradually be accumulating more data to better design further trials. Manually adjusting the bot's strategy to exploit leaks in the players' strategies, however, doesn't really make sense in the context of what they are trying to achieve.

I guess it could be playing against a version of itself that emulates the humans' strategies - that would make some sense and would result in adjustments that focus on the strengths and weaknesses of the humans' play. But if it is just using the hands played as an input to see which actions it regrets, it's way too small of a sample to make a significant difference.

I'm not saying you are wrong with anything; in fact I completely agree with the above and pretty much the rest of your posts. However, when you read that, in the first competition, the researchers "did make several modifications throughout the competition that were intended to plug some leaks, and possibly counter-exploit some of the humans' exploitative tendencies", and, in addition, used "different versions against different humans" based on which version they thought was more likely to beat the respective human, then why would you conclude that the same isn't going on this time?


And in case it's not clear by now, CMU have their very own perspective on what "the context of what they are trying to achieve" really is

Indeed. Unfortunately, humans have a basic design fault that means they can't evolve quite as quickly as machine learners.

If this competition had a shot clock I believe results would look a lot different. Machine is taking minutes on decisions humans are making in seconds.

Okay - I think we understand each other now. I would be really keen for Noam to clarify some of those points, for sure.

It is probably GTO for the players to throw the match and let the computer think it has figured out a great strategy

Sacrifice some money to delay the rise of bots for a bit longer


Then they can book a bunch of action against themselves on Pokershares


win/win

The prof said if AI wins this then 'no human on the planet' can beat the bot and they would move onto different things AKA they would never be able / interested in raising money to pay good players a 2nd time to risk losing their AI>player title. aka they gotta beat this if they would ever want another shot probably.

It wouldn't really make sense for them to simply be running more GTO neural net trials. If they have already ran trillions, what would a few thousand more do? It seems more likely to me that they are making exploitative adjustments, which is dangerous for them given they don't know exactly why the bot is doing what it is doing.

Eh do they start at different times on purpose so you can't check hands without a hassle ? Nobody here who wants to track some interesting hands ?

All 4 start playing at the same time, but the streams for the two playing in the upstairs room are delayed + no twitch chat so that viewers cannot relay information to their mirrored partners about holdings. ( jason + dong are mirrored, dan + jimmy :: jason will play hands with the bot and then dong will play with what the bot had vs what jason had)

Good to know the matchups, thanks. One stream is missing though.

Someone made the observation on one of the live stream chats that the bot was originally trained on those trillions of hands playing itself. So now that they are playing a different opponent (set of opponents), there is potential to get "better" versus those specific opponents by using those hands to "re-train" using the current match hands.

I think that is what the press release may be referring to what the bot does while the human players are sleeping. But who knows?

And, if this is true, it brings up an interesting question. Can the bot utilize the humans' hole card information on hands that do not go to showdown? The bot (programmers) know all hole cards since the match is paired. Not sure if that would be kosher.

Anyway, I am not even convinced that any of this even makes sense.

On stream the bot developer Noam said the bot was seeking a GTO solution and not looking to exploit humans because he said when you exploit you open yourself up to being exploited. But I do question after trillions of hands what are they tweaking overnight?

This is cool:
https://arxiv.org/pdf/1701.01724.pdf

If it can analyze the human's strategy based on hole card info of hands that do not showdown, then it is solving some game other than NLHE. Also, the stacks resetting to 100BB makes this a game variant other than NLHE.

Reminds me of the Piosolver vs. Doug Polk argument where the creator of pio challenged Doug to a match where only certain bet sizes could be used and Doug responded with something like, "why would I agree to play some game that is not No Limit Hold'em?"

Do you have a source for this? Sounds interesting.

Their description of their technology is helpful and quite interesting but their methodology for the contest itself is super sketchy imo. They only pay the top three finishers by chip count which encourages gambling and giving up if you aren't up early. They only played 3K hands per person and the total prize pool was only 8.5K CAD.

Then the bot was super slow so of their 33 players few finished and some quit after < 20 hands. I'm also pretty sure only one of two of their players actually plays heads up at all seriously. If you look at the names of the people who played you probably won't recognize more than 2-3 of them.

There was an article about it here where they interviewed one of the human contestants: https://calvinayre.com/2017/01/12/po...-holdem-clash/

Relevant quotes:

------------------------------------------

“I played few hands, and gave up,” Said Luca Moschitta.”The software they used was that slow that it was making me very tilted and I realised I was playing poorly.”

Fintan Gavin faired a little better competing in 1,555 hands.

“I felt privileged to be offered the opportunity and found it to be a good overall experience with surprisingly decent software with almost zero glitches,” said Gavin. “The biggest challenge for me was completing the 3,000 hands within the time allotted.”

So how did this affect the outcome for Gavin?

“I crushed the bot in the first half of the match. But then in an attempt to speed up, I gambled a lot & lost all my later sessions.” ...

“I would be willing to place a large wager vs. the bot over any amount of hands.” Said Gavin."

------------------------------------------------------------

TLDR I am skeptical that Deepstack is on par with Liberatus or top human HU players. Their technology sounds cool and it may well be, but their match up vs the humans seems extremely poorly designed.

I'm not sure that makes sense. If their training sample is trillions of hands and they are using normal Bayesian updating, a few days worth of trials with the competition hands shouldn't have much of an effect on the overall strategy of the bot. They would have to be weighting the competition hands very heavily, which is dangerous.

FWIW that was me (GTORB creator not Pio) and I wasn't serious about the challenge. I thought that he was claiming that GTO solvers don't do what they say they do (it seemed that way to me at least) as he was speaking very negatively about Pio on reddit while trying to market his own product as an alternative. He said Pio was "basically terrible" without much detail as to why and I thought he meant he didn't believe the solution output was correct within the given bet sizing abstraction.

I said, "if you don't believe that GTO solver algorithms work then surely you would take this challenge" and he replied as the poster above stated. I meant it as a rhetorical device just to point out that if anyone really thinks that GTO solver's results are mathematically inaccurate as presented that they should then be willing to take on such a challenge. I did not mean it as a serious challenge that any sane person would agree to.

I then asked him if he'd play 20bb cap vs a GTO bot if I made one (which is a real game that people play for money) and he never responded. Thread is here: http://forumserver.twoplustwo.com/16...64/index3.html

This is a quote from his post on reddit and then another user brought it up on 2p2 which sparked the discussion:

"I think PioSolver is basically terrible. I don't know what the high stakes six max players are using, and I'm not claiming to be one of the top players in those games. But what I can say is that the players in HU that use this type of software to play, end up playing horrific post flop."

This was in a thread where he was marketing his own preflop software that is not GTO based.

http://forumserver.twoplustwo.com/sh...3&postcount=74

Is there a place where all the counterintuitive bot hands are being complied?

Jason 226 A7cc, libby 678 T7hh, Jason call.
Flop(1356) 653shh
Libby 339, call
Tu(2034) As
Libby 2847, call
Riv(7728) Ad
Libby jam 16136, call

Thought the bot was supposed to understand removal

blocked A7, not sure what more you want from it :P

Surprised to learn Doug said that about pio. I remember Jungleman saying he thinks it's funny that Doug doesn't like pio because a bunch of his proteges use fixed bet sizes. Jungle also said he used pio to improve his own game.
Apologies for derail.