Poker-AI.org
http://poker-ai.org/phpbb/

Preflop branch selection rule
http://poker-ai.org/phpbb/viewtopic.php?f=24&t=2709		 Page 1 of 1
Author:  jcfifo [ Mon Feb 17, 2014 8:58 pm ]
Post subject:  Preflop branch selection rule
I'm not a good player of poker so I decided to make a tree search poker bot with the min possible number of rules encoded in it.

Postflop game is decent, it's funny to see my bot check-raise or bluff taking into consideration I have never told it how to do it. And it basically needs only one rule: select branch with max expected ROI.

But in preflop that doesn't works because full tree can't be simulated, so it only simulates preflop tree and calculates pot equity comparing each player hand potential. It doesn't take into consideration postflop game so the bot becomes too much agressive trying to maximize preflop winnings with the "max expected ROI" rule.

So

- In postflop the objective is to maximize winnings and minimize risks -> "max ROI rule"

The question is:

- In preflop, what is the objective? And, what would be a simple rule that would codify this objective?
Author:  spears [ Mon Feb 17, 2014 9:28 pm ]
Post subject:  Re: Preflop branch selection rule
It's still max ev. So if that isn't working you are not taking into account all relevant factors including your appearance in future hands. Are you taking into account villain's view of you? Are you using a mixed strategy to hide both weakness and strength optimally? Do you assume that villain will use the same strategy on the current hand that he has used in the past? Have you thought about game theory?
Author:  shalako [ Mon Feb 17, 2014 10:40 pm ]
Post subject:  Re: Preflop branch selection rule
max ev preflop could depend on a bunch of factors but the biggest one is the villains folding frequency. I am not sure if we are talking about HU or not but if so then you max out ev with fold equity if you do not have hand equity. So you raise 100% of the time on the button until he adapts and its no longer profitable. You also 3 bet OOP a high percentage of the time until he adapts.

The flip side of this is your bot must also recognize when the villain is using fold equity against you and adapt accordingly by increasing your calling or 3B frequency.
Author:  jcfifo [ Wed Feb 19, 2014 11:34 am ]
Post subject:  Re: Preflop branch selection rule
Thank you, your answers made me think in a different way and I quickly found the problem: hero was allowed to do one more raise per street than villains -poker tracker 2 actions per street database limit makes me a lot of troubles :-( -. Now, preflop statistics make more sense.

spears wrote:
It's still max ev. So if that isn't working you are not taking into account all relevant factors including your appearance in future hands. Are you taking into account villain's view of you? Are you using a mixed strategy to hide both weakness and strength optimally? Do you assume that villain will use the same strategy on the current hand that he has used in the past? Have you thought about game theory?

I haven't worked game theory, I use a simple "brute force" algorythm focused on a fast, heavily table based, code. For the moment only works with datamined decision trees with overall statistics, nothing about hero image or villain changing strategies. I expect it to beat PA and microlimits, then start tweaking with game theory

shalako wrote:
max ev preflop could depend on a bunch of factors but the biggest one is the villains folding frequency. I am not sure if we are talking about HU or not but if so then you max out ev with fold equity if you do not have hand equity. So you raise 100% of the time on the button until he adapts and its no longer profitable. You also 3 bet OOP a high percentage of the time until he adapts.

The flip side of this is your bot must also recognize when the villain is using fold equity against you and adapt accordingly by increasing your calling or 3B frequency.

It's NL multiplayer poker, bot estimates villain action probabilities and card ranges with static datamined decision trees. Dynamic in the future. ;-)
Author:  ibot [ Wed Feb 19, 2014 12:24 pm ]
Post subject:  Re: Preflop branch selection rule
Slightly off-topic, but what kind of accuracy are you getting with your decision trees? Is it predicting F - X/C - B/R, or different?
Author:  shalako [ Wed Feb 19, 2014 11:25 pm ]
Post subject:  Re: Preflop branch selection rule
Quote:
It's NL multiplayer poker, bot estimates villain action probabilities and card ranges with static datamined decision trees. Dynamic in the future.


So by static do you mean its using datamined PT player averages and not adjusting to the players currently in the pot or ?

Also..be careful with using Max EV/ROI to always make a decision. It sounds weird I know but you will find the bot will not be balanced which makes it very very predictable. The biggest hole to watch out for is your bots folding frequencies. You do not want the villain to always get value on his bluffs. This is why calling with many -EV hands are necessary such as backdoor draws, pair and overcard, two overcards, etc.
Author:  jcfifo [ Thu Feb 20, 2014 12:36 pm ]
Post subject:  Re: Preflop branch selection rule
ibot wrote:
Slightly off-topic, but what kind of accuracy are you getting with your decision trees? Is it predicting F - X/C - B/R, or different?


Four decision trees (A is allin)

- preflop XBA 88.2% (big blind)
- preflop FCRA 82.6%
- postflop XBA 73.3%
- postflop FCRA 78.6%

These figures could be about 2% better, but I sacrified it for speed and uniformity.
Also could be much better if it not were for poker tracker database 2 actions per street limit :(

Also 4 regression trees to estimate (rough) raise amount

BUT, in Poker Academy hand history you know opponent cards, so these figures are not reliable to compare with real poker ;-)
Author:  jcfifo [ Thu Feb 20, 2014 1:10 pm ]
Post subject:  Re: Preflop branch selection rule
shalako wrote:
Quote:
It's NL multiplayer poker, bot estimates villain action probabilities and card ranges with static datamined decision trees. Dynamic in the future.


So by static do you mean its using datamined PT player averages and not adjusting to the players currently in the pot or ?

I mean I datamine PT action history offline, the bot knows villain averages but can't react if he changes his strategy

shalako wrote:
Also..be careful with using Max EV/ROI to always make a decision. It sounds weird I know but you will find the bot will not be balanced which makes it very very predictable. The biggest hole to watch out for is your bots folding frequencies. You do not want the villain to always get value on his bluffs. This is why calling with many -EV hands are necessary such as backdoor draws, pair and overcard, two overcards, etc.

Thank you for your advice, I agree with you, I have to solve this weakness. It could be a coincidence, but I have noticed that, when I update something in the bot that changes its statistics, it becomes winner and after a few hundred hands it looses again. It is a good question: how many -EV hands to play, when and how much risk can be taken
Author:  shalako [ Thu Feb 20, 2014 7:12 pm ]
Post subject:  Re: Preflop branch selection rule
Quote:
Thank you for your advice, I agree with you, I have to solve this weakness. It could be a coincidence, but I have noticed that, when I update something in the bot that changes its statistics, it becomes winner and after a few hundred hands it looses again. It is a good question: how many -EV hands to play, when and how much risk can be taken


Well the problem with always playing +EV is that most of the time you flop nothing or some kind of marginal hand or draw so the bot is going to play way too tight and its going to get run over. Marginal holdings and bluffing is where the money is as the good hands play themselves.

Since your playing in multiplayer games you might want to start with a well balanced preflop isolation strategy which will make playing post flop much easier. The other alternative is a small ball implied odds strategy which involves lots of set mining and flopping a disguised hand vs big cards/pairs. An example of that is calling 75s vs an early position raise. Your gonna want to see the flop with this strategy about 25% of the time so your gonna be raising first in with a wide range of hands and calling with the intention of flopping two pair/draw or better. Any top pair or worse type hands your gonna want to fold multiway.

I recommend aggression vs small ball which is rather passive but both have their merits. One thing I like about small ball is that it made me realize why 75s is better then ATs when calling a raise. I know both my cards are likely good and if I get the right flop I may get his entire stack. So technically this is -EV long term but the implied odds are so good that I believe its +EV depending on the situation. The other thing about this type of preflop strategy is that they are easy to get away from post flop. If you do decide to call ATs you call only with the intention of flopping the nut draw. If you do flop top pair then you want to try to play a small pot as and try to get to showdown cheaply otherwise its better to just fold (especially multiway).

Now..the flipside of small ball strategy with ATs. That would be prime hand to actually 3B due to its Ace blocker, especially for a raise from early which consists of many Ax combos. 75s might be a good hand to add into your 3B bluff range with this type of strategy too.

So many different ways to play a hand. That is what is great about the game of poker.
Page 1 of 1 All times are UTC
Powered by phpBB® Forum Software © phpBB Group
http://www.phpbb.com/