Holy Grail of Poker
10-19-2010
, 08:21 PM
10-19-2010
, 08:48 PM
10-19-2010
, 09:58 PM
Quote:
I re-imported as you said and over another 10 occasions it thought flop bet was better 9 times. On some occasions this difference was very large.
It "accepted" my turn raise 3 times out of 10 but each time it still had the EV of a call as slightly higher.
This raises 2 questions:-
1) How far out does a deviation have to be before it decides an action is "wrong"
2) Does it consider bet sizing as a "tell". Clearly the program is saying my turn raise is wrong (and it may well be, especially seeing the probable width of his range
) but part of my "read" was his small bet which although it is sometimes a trap is much more often a weak hand. Does the AI consider that?
I prefer not to post a HH on here to protect the innocent. However, if I could send it to you directly for investigation and you undertook not to use/release it for any other prupose, I am fine with that.
Cheers
T
It "accepted" my turn raise 3 times out of 10 but each time it still had the EV of a call as slightly higher.
This raises 2 questions:-
1) How far out does a deviation have to be before it decides an action is "wrong"
2) Does it consider bet sizing as a "tell". Clearly the program is saying my turn raise is wrong (and it may well be, especially seeing the probable width of his range
) but part of my "read" was his small bet which although it is sometimes a trap is much more often a weak hand. Does the AI consider that?I prefer not to post a HH on here to protect the innocent. However, if I could send it to you directly for investigation and you undertook not to use/release it for any other prupose, I am fine with that.
Cheers
T
2) Yes, it makes bet size reads, and this is in the process of being enhanced. However, off the top of my head I'm not sure it's making the same read as you. PM me the hand history -- and I won't share it. But I can analyze it, and I can see what our AI is thinking about that bet size in that situation.
Thank you for the terrific questions / advice.
10-19-2010
, 10:24 PM
10-20-2010
, 01:55 AM
Quote:
When playing against simulated game against the AI, we have three import options: 1 all hands, 2 no hands, and 3 only when playing with advice hidden. I wanted to play using Hero while I import all hands (or only when playing with advice hidden), and using Hero2 while I import no hands. I think I figured out how to do it, but it was a bit confusing (I had to right-mouseclick when I was playing at the table and create a new hero that way).
The goal is to be able to play using one of them while I look at the advice, and to play using the other one while I do not look at the advice. I will look at the advice when I play using Hero2; but I will not look at it when I play using Hero and import at the same time. When I play using Hero, and do not look at the advice, and import, that gives HGP the ability to adjust to my play. But if I start looking at the advice while playing and importing with Hero, then HGP will not adjust correctly to my play, so when I want to look at the advice I switch over to Hero2.
As far as the GUI is concerned, here is how I think it would work well:
That drop-down box that shows Hero playing is disabled, but it would be better if it was enabled and allowed the user to choose Hero, Hero2, Hero3, etc. (There also needs to be a screen/window for creating those other heros too.)
By the way, here is how I ended up creating (and switching to) Hero2, but I do not think it is good to have to do it like this:
The goal is to be able to play using one of them while I look at the advice, and to play using the other one while I do not look at the advice. I will look at the advice when I play using Hero2; but I will not look at it when I play using Hero and import at the same time. When I play using Hero, and do not look at the advice, and import, that gives HGP the ability to adjust to my play. But if I start looking at the advice while playing and importing with Hero, then HGP will not adjust correctly to my play, so when I want to look at the advice I switch over to Hero2.
As far as the GUI is concerned, here is how I think it would work well:
That drop-down box that shows Hero playing is disabled, but it would be better if it was enabled and allowed the user to choose Hero, Hero2, Hero3, etc. (There also needs to be a screen/window for creating those other heros too.)
By the way, here is how I ended up creating (and switching to) Hero2, but I do not think it is good to have to do it like this:
Thanks!
10-20-2010
, 04:32 AM
I do not know how you can get round this potential problem but I throw it out there to see what you and others think.
You say that the program is intended as a training tool as if we have a coach sitting looking over our shoulder. However, due to processing time pressure it runs the simulation a relatively small number of times compare to infinity That can create, as it did in the case of the J9 hand I highlighted, a situation where it gives advice that HGP itself thought was wrong the next 9 times it tried.
Because it looked "odd" I questioned it and you explained the method of how to get the prog to re-import to give it another chance to consider the hand.
However, I am a player with many years of experience. What if I was a relatively new player? I would have no experience base on which to "question" HGP and would now be left thinking that "slow playing" top and bottom pair on a 2 flush 3 str8 board is the correct play.
(I can just about see an argument - but still disagree - for not betting on the basis that my equity is not brilliant when called by worse and awful when called/raised by better. If that equity balance was such that I had a very small edge when he did not fold, I may indeed make more money on the hand by inducing bluffs or hero calls later but I would not call this a slow play; it is simply a pot control type line looking to maximise EV by inducing because I make litttle "profit" from a value bet)
Would I notice, as a novice, that in a few thousand hands time, HGP then advised betting when a similar situation re-appeared?
The problem of different assumptions will always remain but the randomness caused by small sample runs could be confusing to relative beginners.
Cheers
T
You say that the program is intended as a training tool as if we have a coach sitting looking over our shoulder. However, due to processing time pressure it runs the simulation a relatively small number of times compare to infinity
That can create, as it did in the case of the J9 hand I highlighted, a situation where it gives advice that HGP itself thought was wrong the next 9 times it tried.Because it looked "odd" I questioned it and you explained the method of how to get the prog to re-import to give it another chance to consider the hand.
However, I am a player with many years of experience. What if I was a relatively new player? I would have no experience base on which to "question" HGP and would now be left thinking that "slow playing" top and bottom pair on a 2 flush 3 str8 board is the correct play.
(I can just about see an argument - but still disagree - for not betting on the basis that my equity is not brilliant when called by worse and awful when called/raised by better. If that equity balance was such that I had a very small edge when he did not fold, I may indeed make more money on the hand by inducing bluffs or hero calls later but I would not call this a slow play; it is simply a pot control type line looking to maximise EV by inducing because I make litttle "profit" from a value bet)
Would I notice, as a novice, that in a few thousand hands time, HGP then advised betting when a similar situation re-appeared?
The problem of different assumptions will always remain but the randomness caused by small sample runs could be confusing to relative beginners.
Cheers
T
10-20-2010
, 05:26 AM
PS After I wrote my previous post HGP came up with another odd result. I had misplayed it preflop (so true!) but on the flop it suggested raising. My thought was WTF? (or simlar). I re-ran it and suddenly it said call was "correct" by 25% of the pot. A couple of other re-runs gave similar big +EV results for just calling. Then it said I should raise again - back to call etc. etc.
What I would like to be able to do is to get longer runs for POST play analysis without needing to do this manually.
I understand this AI has a lot of processing to do, so, if it is supposed to be an immediate post hand critic, it has to reduce the number of trials to a relatively small number.
However, in the calm of the following day when a more experienced player may be interested in whether his game has a few leaks he should be thinking about, the output is sometimes incorrect based on the small sample run. Not only does it throw up "mistakes/deviations" that are not present over longer runs but the short run may also fail to find problems that would exist over a longer one so we never even know it was there.
It would be useful to be able to ask HGP to re-run ALL of our hands that are filtered for analysis but to allow it to run many more cycles (choice made by the user depending on the time available to THEM) so that its results are based more on a long term basis and are more reliable.
Alternatively if time was a little more pressing for the user, we could ask HGP NOT to re-run hands that were so "certain" over the initial smaller run performed on import that a re-run would be most unlikely to change its opinion.
These choices of a) number of cycles b) probability of initial run being correct would allow users to only see results that we were pretty sure were likely to be useful rather than have a look and say WTF? and then need to re-run the hand several times manually to see what an "average" result looks like.
I cannot imagine implementing this would be difficult for post play analysis; but what do I know?
T
) but on the flop it suggested raising. My thought was WTF? (or simlar). I re-ran it and suddenly it said call was "correct" by 25% of the pot. A couple of other re-runs gave similar big +EV results for just calling. Then it said I should raise again - back to call etc. etc.What I would like to be able to do is to get longer runs for POST play analysis without needing to do this manually.
I understand this AI has a lot of processing to do, so, if it is supposed to be an immediate post hand critic, it has to reduce the number of trials to a relatively small number.
However, in the calm of the following day when a more experienced player may be interested in whether his game has a few leaks he should be thinking about, the output is sometimes incorrect based on the small sample run. Not only does it throw up "mistakes/deviations" that are not present over longer runs but the short run may also fail to find problems that would exist over a longer one so we never even know it was there.
It would be useful to be able to ask HGP to re-run ALL of our hands that are filtered for analysis but to allow it to run many more cycles (choice made by the user depending on the time available to THEM) so that its results are based more on a long term basis and are more reliable.
Alternatively if time was a little more pressing for the user, we could ask HGP NOT to re-run hands that were so "certain" over the initial smaller run performed on import that a re-run would be most unlikely to change its opinion.
These choices of a) number of cycles b) probability of initial run being correct would allow users to only see results that we were pretty sure were likely to be useful rather than have a look and say WTF?
and then need to re-run the hand several times manually to see what an "average" result looks like.I cannot imagine implementing this would be difficult for post play analysis; but what do I know?
T
10-20-2010
, 06:05 AM
Quote:
I do not know how you can get round this potential problem but I throw it out there to see what you and others think.
You say that the program is intended as a training tool as if we have a coach sitting looking over our shoulder. However, due to processing time pressure it runs the simulation a relatively small number of times compare to infinity That can create, as it did in the case of the J9 hand I highlighted, a situation where it gives advice that HGP itself thought was wrong the next 9 times it tried.
Because it looked "odd" I questioned it and you explained the method of how to get the prog to re-import to give it another chance to consider the hand.
However, I am a player with many years of experience. What if I was a relatively new player? I would have no experience base on which to "question" HGP and would now be left thinking that "slow playing" top and bottom pair on a 2 flush 3 str8 board is the correct play.
(I can just about see an argument - but still disagree - for not betting on the basis that my equity is not brilliant when called by worse and awful when called/raised by better. If that equity balance was such that I had a very small edge when he did not fold, I may indeed make more money on the hand by inducing bluffs or hero calls later but I would not call this a slow play; it is simply a pot control type line looking to maximise EV by inducing because I make litttle "profit" from a value bet)
Would I notice, as a novice, that in a few thousand hands time, HGP then advised betting when a similar situation re-appeared?
The problem of different assumptions will always remain but the randomness caused by small sample runs could be confusing to relative beginners.
Cheers
T
You say that the program is intended as a training tool as if we have a coach sitting looking over our shoulder. However, due to processing time pressure it runs the simulation a relatively small number of times compare to infinity
That can create, as it did in the case of the J9 hand I highlighted, a situation where it gives advice that HGP itself thought was wrong the next 9 times it tried.Because it looked "odd" I questioned it and you explained the method of how to get the prog to re-import to give it another chance to consider the hand.
However, I am a player with many years of experience. What if I was a relatively new player? I would have no experience base on which to "question" HGP and would now be left thinking that "slow playing" top and bottom pair on a 2 flush 3 str8 board is the correct play.
(I can just about see an argument - but still disagree - for not betting on the basis that my equity is not brilliant when called by worse and awful when called/raised by better. If that equity balance was such that I had a very small edge when he did not fold, I may indeed make more money on the hand by inducing bluffs or hero calls later but I would not call this a slow play; it is simply a pot control type line looking to maximise EV by inducing because I make litttle "profit" from a value bet)
Would I notice, as a novice, that in a few thousand hands time, HGP then advised betting when a similar situation re-appeared?
The problem of different assumptions will always remain but the randomness caused by small sample runs could be confusing to relative beginners.
Cheers
T
10-20-2010
, 06:07 AM
Quote:
PS After I wrote my previous post HGP came up with another odd result. I had misplayed it preflop (so true!) but on the flop it suggested raising. My thought was WTF? (or simlar). I re-ran it and suddenly it said call was "correct" by 25% of the pot. A couple of other re-runs gave similar big +EV results for just calling. Then it said I should raise again - back to call etc. etc.
What I would like to be able to do is to get longer runs for POST play analysis without needing to do this manually.
I understand this AI has a lot of processing to do, so, if it is supposed to be an immediate post hand critic, it has to reduce the number of trials to a relatively small number.
However, in the calm of the following day when a more experienced player may be interested in whether his game has a few leaks he should be thinking about, the output is sometimes incorrect based on the small sample run. Not only does it throw up "mistakes/deviations" that are not present over longer runs but the short run may also fail to find problems that would exist over a longer one so we never even know it was there.
It would be useful to be able to ask HGP to re-run ALL of our hands that are filtered for analysis but to allow it to run many more cycles (choice made by the user depending on the time available to THEM) so that its results are based more on a long term basis and are more reliable.
Alternatively if time was a little more pressing for the user, we could ask HGP NOT to re-run hands that were so "certain" over the initial smaller run performed on import that a re-run would be most unlikely to change its opinion.
These choices of a) number of cycles b) probability of initial run being correct would allow users to only see results that we were pretty sure were likely to be useful rather than have a look and say WTF? and then need to re-run the hand several times manually to see what an "average" result looks like.
I cannot imagine implementing this would be difficult for post play analysis; but what do I know?
T
) but on the flop it suggested raising. My thought was WTF? (or simlar). I re-ran it and suddenly it said call was "correct" by 25% of the pot. A couple of other re-runs gave similar big +EV results for just calling. Then it said I should raise again - back to call etc. etc.What I would like to be able to do is to get longer runs for POST play analysis without needing to do this manually.
I understand this AI has a lot of processing to do, so, if it is supposed to be an immediate post hand critic, it has to reduce the number of trials to a relatively small number.
However, in the calm of the following day when a more experienced player may be interested in whether his game has a few leaks he should be thinking about, the output is sometimes incorrect based on the small sample run. Not only does it throw up "mistakes/deviations" that are not present over longer runs but the short run may also fail to find problems that would exist over a longer one so we never even know it was there.
It would be useful to be able to ask HGP to re-run ALL of our hands that are filtered for analysis but to allow it to run many more cycles (choice made by the user depending on the time available to THEM) so that its results are based more on a long term basis and are more reliable.
Alternatively if time was a little more pressing for the user, we could ask HGP NOT to re-run hands that were so "certain" over the initial smaller run performed on import that a re-run would be most unlikely to change its opinion.
These choices of a) number of cycles b) probability of initial run being correct would allow users to only see results that we were pretty sure were likely to be useful rather than have a look and say WTF?
and then need to re-run the hand several times manually to see what an "average" result looks like.I cannot imagine implementing this would be difficult for post play analysis; but what do I know?
T
10-20-2010
, 06:39 AM
Quote:
I do not know how you can get round this potential problem but I throw it out there to see what you and others think.
You say that the program is intended as a training tool as if we have a coach sitting looking over our shoulder. However, due to processing time pressure it runs the simulation a relatively small number of times compare to infinity That can create, as it did in the case of the J9 hand I highlighted, a situation where it gives advice that HGP itself thought was wrong the next 9 times it tried.
Because it looked "odd" I questioned it and you explained the method of how to get the prog to re-import to give it another chance to consider the hand.
However, I am a player with many years of experience. What if I was a relatively new player? I would have no experience base on which to "question" HGP and would now be left thinking that "slow playing" top and bottom pair on a 2 flush 3 str8 board is the correct play.
(I can just about see an argument - but still disagree - for not betting on the basis that my equity is not brilliant when called by worse and awful when called/raised by better. If that equity balance was such that I had a very small edge when he did not fold, I may indeed make more money on the hand by inducing bluffs or hero calls later but I would not call this a slow play; it is simply a pot control type line looking to maximise EV by inducing because I make litttle "profit" from a value bet)
Would I notice, as a novice, that in a few thousand hands time, HGP then advised betting when a similar situation re-appeared?
The problem of different assumptions will always remain but the randomness caused by small sample runs could be confusing to relative beginners.
Cheers
T
You say that the program is intended as a training tool as if we have a coach sitting looking over our shoulder. However, due to processing time pressure it runs the simulation a relatively small number of times compare to infinity
That can create, as it did in the case of the J9 hand I highlighted, a situation where it gives advice that HGP itself thought was wrong the next 9 times it tried.Because it looked "odd" I questioned it and you explained the method of how to get the prog to re-import to give it another chance to consider the hand.
However, I am a player with many years of experience. What if I was a relatively new player? I would have no experience base on which to "question" HGP and would now be left thinking that "slow playing" top and bottom pair on a 2 flush 3 str8 board is the correct play.
(I can just about see an argument - but still disagree - for not betting on the basis that my equity is not brilliant when called by worse and awful when called/raised by better. If that equity balance was such that I had a very small edge when he did not fold, I may indeed make more money on the hand by inducing bluffs or hero calls later but I would not call this a slow play; it is simply a pot control type line looking to maximise EV by inducing because I make litttle "profit" from a value bet)
Would I notice, as a novice, that in a few thousand hands time, HGP then advised betting when a similar situation re-appeared?
The problem of different assumptions will always remain but the randomness caused by small sample runs could be confusing to relative beginners.
Cheers
T
I think we'll add the ability to change # of simulations to reduce this issue, and have the default be longer/more simulations.
Thank you for your insight.
10-20-2010
, 10:08 AM