Ideas on automating table selection

[bziegler2]

Table selection is probably the second biggest factor in determining a player's win rate (BB/100). I'm assuming skill is the biggest factor. Improving and automating your table selection could dramatically improve your win rate.

How do most people select their tables now? Looking at the overlay stats and thinking "hey there are X # of fish at this table"?

I think it would be more efficient to have it automated. One way would be to score each player with a "fish factor score", we could use the results from Predicting BB/100 given Player's Stats to do that part. The next tricky part would be to calculate some single Table Score based on the "fish score" (not sure how this would work).

Using historical data I think we could run some queries against past data to see if the theory (i.e. the table score function) really works. You would look for a correlation between the Table Score, and the player's win rate at the tables. This would require some iteration and expert knowledge to improve the "magic formula" used to calculate the Table Score, but eventually you'd have a formula that could calculate with high probability your chances of winning more at a given table.

If anyone is interested in helping me pursue this let me know.

-Ben

[jukofyork]

[ QUOTE ]
Table selection is probably the second biggest factor in determining a player's win rate (BB/100). I'm assuming skill is the biggest factor. Improving and automating your table selection could dramatically improve your win rate.

How do most people select their tables now? Looking at the overlay stats and thinking "hey there are X # of fish at this table"?

I think it would be more efficient to have it automated. One way would be to score each player with a "fish factor score", we could use the results from Predicting BB/100 given Player's Stats to do that part. The next tricky part would be to calculate some single Table Score based on the "fish score" (not sure how this would work).

Using historical data I think we could run some queries against past data to see if the theory (i.e. the table score function) really works. You would look for a correlation between the Table Score, and the player's win rate at the tables. This would require some iteration and expert knowledge to improve the "magic formula" used to calculate the Table Score, but eventually you'd have a formula that could calculate with high probability your chances of winning more at a given table.

If anyone is interested in helping me pursue this let me know.

-Ben

[/ QUOTE ]
I'm interested and have just been working on a similar idea too. I've so far a fair prediction of BB/100 for each player based on a LWS of just 4 features (VPIP, PFR, PFA & WtSD), but somehow I need to better handle the samples with low amounts of data.

The samples with low amounts of data have alot of noise (especially PFA, which can be almost +/- inf) which causes problems. The problem is that these players you haven't seen before, tend to be the biggest fish of all (ie: NumHands correlates with BB/100) and I want to be able to quickly classify a table which has a few players with very few hands played (assuming they are just average players [average with few samples too] to start off with, when you have no data on them yet).

I can see that much better, even positionally aware (fish to the left, rocks to the right, etc), formulas could be made, but I think it's gonna take a huge amount of data for that.

One extension of my idea of predicting BB/100 against any player, would be to try to classify the player types first (mainly to try to find TAGs and LAG/TAGs or other profitable styles), then create an energy flow matrix of what each player type is winning/losing to other player styles (I'm guessing TAGs love Calling-Stations more and LAG/TAGs love Weak-Tights more). Then using this you could try to come up with a better LWS which predicts the BB/100 a TAG or LAG/TAG (depending on your style) could make from a table.

Juk [img]/images/graemlins/smile.gif[/img]

[bziegler2]

Juk,

What is LWS? [img]/images/graemlins/smile.gif[/img]

[ QUOTE ]

and I want to be able to quickly classify a table which has a few players with very few hands played


[/ QUOTE ]

That will be tough. It might be a bit easier if you have done LOTS of data mining, and could be confident that someone with a low number of hands just hasn't played before, because otherwise you'd have data on them.

[ QUOTE ]

I can see that much better, even positionally aware (fish to the left, rocks to the right, etc), formulas could be made, but I think it's gonna take a huge amount of data for that.


[/ QUOTE ]

Yeah, that's why I'm asking for help. [img]/images/graemlins/smile.gif[/img]

Your energy flow matrix sounds like a neat idea. But how would that actually work, i.e. combine it into one number to use as a Table Score.

[Mempho]

I've come up with some spreadsheets that give a raw score on how fishy a player is. Looseness and passivity create a score...the more loose and passive a player is, the higher the score. It's a little more complex than that, but you get the point. I suppose one could score this somehow on a per table basis. Even better would be doing it on a per seat basis. This could be incredible and such a program would be worth lots of money.

[WolfElder]

isnt this similar to that what the program poker table scan does?

[Mempho]

[ QUOTE ]
isnt this similar to that what the program poker table scan does?

[/ QUOTE ]

I don't know of this poker table. Can you describe it?

[jmillerdls]

wouldn't your own personal style need to be involved somewhere? I'd think a TAG would like different tables than LAGS would.

[jukofyork]

[ QUOTE ]
What is LWS?
Quote:
--------------------------------------------------------------------------------


and I want to be able to quickly classify a table which has a few players with very few hands played



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



That will be tough. It might be a bit easier if you have done LOTS of data mining, and could be confident that someone with a low number of hands just hasn't played before, because otherwise you'd have data on them.

[/ QUOTE ]
Sorry, LWS = Linear Weighted Sum, so atm I have:

BB/100' = A*VPIP + B*PFR + C*PFA + D*WtSD

All attempts to train up non-linear stuff so far haven't worked well, mainly because of the PFA problems when using low amounts of data (my data is 6max, and PFA is highly correlated to winning).

(Only spent about 2 hours on this so far, but results were looking ok [apart from the PFA outliers causing huge winrates]... not had time to use PCA or clustering to preprocess the inputs yet).

My other problem is if I just choose the players I have alot of data on (say 1k+ hands), then these tend to be the winners anyway (which I would rather not have at my 6max table), so I have to try to make as best use as I can from the low data samples. I think possibly even adding in another feature like E*log(NumHands) may work, as like I mentioned before the more hands I see a player play, the better they are in general.

[ QUOTE ]
Your energy flow matrix sounds like a neat idea. But how would that actually work, i.e. combine it into one number to use as a Table Score.

[/ QUOTE ]
The only interesting row would really just be your own playing style (assuming TAG or LAG/TAG), and then I was just going to sum the elements to try to estimate the amount your playing style would likely win (basically just an extension of my current generalized player type idea).

I'm not sure if any of this is possible though (may need a huge amount of data), but I think it's an interesting idea to look at.

Juk [img]/images/graemlins/smile.gif[/img]

[bziegler2]

Thanks for the replies.

My problem is I like to write poker software, but I just don't play that much. I only have about 5000 hands played. Something like this is going to need about 50k hands on 1 player, and it would be good to have multiple players with different styles to see if the formulas are consistent.

Does anyone want to donate some data to get this effort off the ground?

[bziegler2]

I forgot to mention, I think I have an easy way to figure out some good weighting numbers. I could use the "Goal Seek" function in Excel to change the weights until there is a maximum correlation between the Table Score and the player's win rate at the table.