Poker-AI.org

So, it looks like HUL has been solved (by neopokerbot)?
http://www.computerpokercompetition.org ... l=&start=1

And, congrats to Marv for regaining his bankroll throne.

Neo Poker Bot beat Hyperborean (IRO) by 11.2 +/- 6.2 millibets in this year's run off competition. Last year's Hyperborean (or which ever strategy the U of A discussed in their papers), was exploitable by 62 millibets.

I'll set up the server again if you want to play again Marv.

the slumbot NL 's paper :

http://www.aaai.org/ocs/index.php/WS/AA ... /view/7044

So what? It means that Best Response vs Hyperborean's strategy will exploit it 62mb/h. But accurate nash equlibrium strategy can potentially beat it by 11.2mb/h.
Maybe someone remember how much space in terabytes full FLHU strategy takes? If we assume keeping data in doubles, and without compressing?

I'm guessing there's some degree of exploitability in NPB's play, if only due to compressing/truncating the game for the competition. By that +11 millibets, I'm guestimating how close it is to a real EQ.

Hello all. I was interested in seeing some of the NL hands between Slumbot and Hyperborean.iro so I converted about ~100k of their hands into PokerStars hand history format so Holdem Manager 2 can import them. It's setup so the holecards are visible while watching in the hand replayer. If anyone else is curious, I just put it on the first file sharing site that Google returned: http://www.filedropper.com/hyperboreanirovsslumbotnl

I couldn't help but notice that Ben Sulsky submitted a bot to the NL competition too. He's better known online as Sauce123 and is one of the best big bet cash game players... Cool to see him experimenting.

Edit: the previous file hosting site I used had a pretty short expiration time... Here's a new link: http://www.fileswap.com/dl/Q7X0GOP31V/

Edit: new new link: http://www.fileswap.com/dl/7n7nKXVfYd/

TY !!!!



Sauce123
I think that it's the only botter of the competition who is better than his bot

it's so close

^
what's up with the showdown/non-showdown lines? is slumbot in the same position every hand or how can they be so asymmetrical?

no there is no filter.
they play a différent style
superborean seems more aggressive, he 3b 15%, 2.52 postflop AF whereas slumbot 3b 10% & 1.86 postflop AF

Confirmed. I imported a few more hands than you.



Also it seems like slumbot limps a lot more preflop.

I played the neopoker bot for a few hundred hands and it plays terrible for HU NL. The first thing I noticed is that it folds to a 3 bet about 95% of the time. It also does not have a 3 bet flatting range so its either fold or all in. So OOP to beat it you just 3 bet every hand and fold to the all-in. It raises about 90% of buttons so the 3 bet is highly effective against it. I raised 90% of buttons as well and Its flatting range is typical..broadways..suited aces etc so its highly predictable. It 3 bets top 10% or so and wider if its short of course. A 4 bet will just result in an allin just like OOP. I saw no adjustment whatsoever.

Post flop it plays way too tight and its easy to range it. It does not handle a check raise or a double/triple barrel that well. When it does 3 bet or raise its 2 pair or better unless its short. I have yet to see a bluff raise. It also does not seem to notice differences in bet sizing and what that could mean. Like it was very weak against a triple deke type of play.

Most bots do not handle aggression correctly and this one is no exception as its easy to run over. The flaws with this bot showed up after just a few hands of play.

Update: I played a few hundred more it did start calling a few of my 3 bets but certainly not enough to overcome my 100% 3 bet advantage. I also saw it bluff raise/stack off with bottom pair which was rather bold based on the board texture

I was thinking about this as I work on board texture. I think they used statistics to classify "same" hands according to metrics that allowed for real-game precision.

I can reduce the flop board from about 22k combinations to about 2.2k by using the stats method I've mentioned elsewhere on this forum. Of course, my method probably isn't precise to the real game, the point is even without clustering, just removing duplicates, it's a 90% reduction.

Ty, Sir. What did you use for HH conversion to PS format?

Regarding Sauce, it is his dad that is the programmer behind the project.

I noticed that hyperborean_iro never flats premium hands from the BB (QQ+, AK) facing an opening raise (it always 3-bets) which sounds logical. I am wondering if approximate NE for shorter stacks (say 25bb) still never flats those hands. Anyone that has solved these short stack games find the same strategy?

Thanks.

Great discussion, unfortunately it died down.

I tried to use nonpareil download links but none of those 2 seems to be working anymore.
Could anyone be so kind and reupload or pn me with instructions?

Thanks guys and keep up to good work!

Would you mind sharing your code for creating importable PokerStars-HH?