Poker-AI.org

Re: Leaks in plain site

2013-12-26T05:29:01+00:00

I found notes I had written down from Sauce about foldinq frequencies:

Quote:

Folding frequencies for nlhe seem to hover around alpha % in most situations. Alpha= (betsize in pots)/1+(betsize in pots). In mid stakes games, mean turn barrel size tends to be around 3/4th pot, which means you should be folding around (.75/1.75), or around 42 percent of the time.

Statistics: Posted by shalako — Thu Dec 26, 2013 5:29 am

Re: Leaks in plain site

2013-12-19T05:54:50+00:00

maikon wrote:

PT4 imports play money games on Stars, so you can have decent HU opponents play your bot there.

Also, what makes you believe folding 33% on turn is correct?

Yeah I always test my bots in play money. Its not very realistic as a whole but I am able to test it against many different different player types and betsizes which allows me to program for just about every scenario imaginable.

33% is near correct because if I am folding to much more then that then the villain is always getting value from his bluffs on a potsized bet. I asked that question to a bunch of GTO guys and I got the same response from all of them that 33% was almost near optimal. I got it right and didn't even know it. I thought it was over calling the turn which is why I asked the question.

Statistics: Posted by shalako — Thu Dec 19, 2013 5:54 am

Re: Leaks in plain site

2013-12-18T22:37:51+00:00

PT4 imports play money games on Stars, so you can have decent HU opponents play your bot there.

Also, what makes you believe folding 33% on turn is correct?

Statistics: Posted by maikon — Wed Dec 18, 2013 10:37 pm

Re: Leaks in plain site

2013-12-15T02:51:35+00:00

Having it play a pro (or even myself) in an environment like you say would be the best way for sure to debug a HU bot. Play a few thousand hands against it aggressively and then run the stats thru PT or that site you mentioned to check for issues.

I have resorted to putting my bot into certain situations over and over to generate stats. Like the last one I ran was too see how often it was folding the turn to a double barrel after it called the flop. I found that it was ok and only folding about 33% of the time which was just about right.

Statistics: Posted by shalako — Sun Dec 15, 2013 2:51 am

Re: Leaks in plain site

2013-12-15T00:16:54+00:00

You definitely need to get a decent HU player to play against it a good sample of hands in an environment where HUD stats are available. He will easily point out a lot of leaks that go beyond what you have listed. Unlike 6max, HU huds contain many more stats due to the wealth of screen space. Also, since it is rare that you get regular action, it is easy to prepare for a match in advance away from the tables. Alternatively, you can run a big sample of hands through sth like pokerarithmetic.com in order to find sequences that can be exploited.

Statistics: Posted by maikon — Sun Dec 15, 2013 12:16 am

Leaks in plain site

2013-12-10T18:00:34+00:00

Lately I have been working on ensuring my HUNL bot does not have any obvious leaks and that its adjusting quickly before losing any significant money. I found a thread somewhere in which a guy found a bot net and found their weaknesses very quickly and made a pile of money. After he figured out the bots ranges and problems he quickly crushed them as the bots never adapted. I was happy to see he listed several of the bots weaknesses so I that I could make sure mine didnt do the same:

1. Folds to high percentage of 3 bets preflop. The problem was so severe this guy 3 betted 90% of hands.
2. Hi Flop C bet % in position. The bot would fold to 65% of check raises.
3. Gives up on turn in position when weak or high turn fold %. The most common post flop problem.
4. Very passive on the turn with Top Pair OOP. It would pot control making the hand transparent.

Problems 1-3 are pretty bad. I noticed the Neo bot had similar 3B issues. Anyway..this kind of stresses the importance of running sims to see what the bots stats are in various situations to ensure its not folding too much. And not only that it needs to identify situations quickly when the villain is trying to exploit it (especially preflop)

Statistics: Posted by shalako — Tue Dec 10, 2013 6:00 pm