OK. Latest version of JBMCCFR posted, but I've moved the files. You can now find them at
Notifications
Clear all
I commented on
and now Barry Carter writes
http://www.pokerstrategy.com/news/world-of-poker/How-to-understand-extreme-variance_87242/
And speaking of
PokerStars Tournament #984180120, No Limit Hold'em
Buy-In: $2.85/$0.15 USD
3 players
Total Prize Pool: $30.00 USD
Tournament started 2014/10/01 22:15:28 WET [2014/10/01 17:15:28 ET]
Dear jbpatzer,
You finished the tournament in 1st place. A USD 30.00 award has been credited to your Real Money account.
Congratulations!
Thank you for participating.
Spin and Go. The most perfect recreational iPad poker game ever invented. When those numbers spin and I get to see what we're playing for, I get all excited! Maybe I should get out more.... 
Oooooo! That's good. 
Petition update
Richard Richardson just posted an update on the petition you signed, Jenny Williams (Commissioner and Chief Executive) Gambling Commission: Let UK poker players keep automated rebuys and top-ups.
Gambling Commission OKs Rebuys
Oct 02, 2014 — The Gambling Commission have issued a clarification in their FAQ blog for the new licence regime....
http://licensingadvertisingact.blogspot.co.uk/2014/10/42-can-licence-holders-offer-auto-rebuy.html
Originally posted by jbpatzer
PokerStars Tournament #984180120, No Limit Hold'em
Buy-In: $2.85/$0.15 USD
3 players
Total Prize Pool: $30.00 USD
Tournament started 2014/10/01 22:15:28 WET [2014/10/01 17:15:28 ET]Dear jbpatzer,
You finished the tournament in 1st place. A USD 30.00 award has been credited to your Real Money account.
Congratulations!
Thank you for participating.
Spin and Go. The most perfect recreational iPad poker game ever invented. When those numbers spin and I get to see what we're playing for, I get all excited! Maybe I should get out more....
Well done!
iPoker has Twister, which are pretty much the same thing. They are ideal for the "Bingo" players who like to take chances. So much so that I've dubbed them "Spingo".
You are welcome to use the term as often as you wish without attribution or royalty. 
The buggers are fun though. ♠_biggrin:
I think that the thing with the November Nine is that it is so noticeable.
Hands like the one where Quad Aces is beaten by a Royal flush have lower probability, but are buried in the noise of everyday life. In the first live tournament I ever played, the same thing happened at the table behind me. They had to pause the clock while everyone had a look.
People don't recognize that the improbable happens all the time. It must. If it is possible, eventually it will happen.
I am convinced that every winter, at least once, there are two snowflakes the same.
Cheers,
--VS
I've had several interesting thoughts since last night. BTW, this is all about my code which, even though only me and Jules seem to be interested in it, is going to be the most awesme piece of poker software ever written (my unbiassed opinion), so if you want some lolz, go read a young person's blog.
1) I don't know why it hadn't occurred to me before, but it would clearly be fiddly, but straightforward in principle, to set the code up to start from the turn or river, thereby reproducing all of the functionality of GTORangeBuilder, and more, but without the sexy graphics. I didn't really set out to do this, but these are easier problems than the one I'm solving. Running it from the river would produce pretty much instant results, and from the turn, would probably take minutes at most. I could envisage this possibly being a commercial product, but tbh I doubt whether I cba to put the effort in to trying to make money out of it, because I suspect the market remains limited. It seems to me more likely that I'll eventually release a gui that allows you to produce data files after inputting a flop, turn or river board and two player ranges. The fiddly bit is inputting the ranges, but I'm sure I can sort that out.
2) The fact that you can't put coloured text or suit symbols into the text box of a MATLAB gui is seriously tilting me. If anyone can make anything of this, earn a gold star, and then explain it to me please. I've asked my friend Doug Hull, who writes poker books and works as a MATLAB expert for the MathWorks, if he can help.
3) I worked out a structure for my data that stores it in what I suspect is the optimal way, with no wasted space. I'll code it up this weekend and see how much space I save. This should give me the way in to using larger game trees (more bet sizes).
4) I read a very interesting paper by Sandholm last night. I didn't follow the details, but basically it seems to be a framework that allows you to optimize the bet sizes during an MCCFR run. The paper's a bit impenetrable (surprise, surprise) but the idea is out there. If, with a range of bet sizes, the CFR solution starts to converge to just one or two, this could potentially allow the game tree to be pruned, and the bet sizes adjusted to maximise the EV of the strategy. This then really would be a GTO solution of postflop NLHE. I have no idea whether I can get that working, but I'll bear it in mind.
5) One question that seems to me not to have an obvious answer is whether a GTO strategy would involve a single, optimal bet size for a given hand, or more than one bet size, i.e. a mixed strategy for bet sizing. It's not obvious from the solutions I've computed so far. For example, if the solution at some decision node is to check 50% and bet three quarters pot 50%, does than mean that the real GTO solution would involve always betting 37.5% pot, or some other approach? Once I have more bet sizes, it should be clear whether the solution homes in on some particular bet size, or sometimes bets high and sometimes low. If the former is true, Sandholm's approach could work; if the latter, it's not so clear. I have some ideas of my own too.......
6) Janda's example hands 6-10 involve fractional ranges, i.e. calling on the button with just one combo of AA and three of AKo, iirc. I can include this, but it's a right pain in the arse for little reward imo. I will, however, do it, because I want to solve those ten hands, but I'll be pissed off for the whole time that I'm coding that bit.
Mrsjb's off to sit a counselling exam in Sheffield this afternoon, so it's me, the kids and the ironing today. Ho hum....
Originally posted by jbpatzer
I've had several interesting thoughts since last night. BTW, this is all about my code which, even though only me and Jules seem to be interested in it, is going to be the most awesme piece of poker software ever written (my unbiassed opinion), so if you want some lolz, go read a young person's blog.1) I don't know why it hadn't occurred to me before, but it would clearly be fiddly, but straightforward in principle, to set the code up to start from the turn or river, thereby reproducing all of the functionality of GTORangeBuilder, and more, but without the sexy graphics. I didn't really set out to do this, but these are easier problems than the one I'm solving. Running it from the river would produce pretty much instant results, and from the turn, would probably take minutes at most. I could envisage this possibly being a commercial product, but tbh I doubt whether I cba to put the effort in to trying to make money out of it, because I suspect the market remains limited. It seems to me more likely that I'll eventually release a gui that allows you to produce data files after inputting a flop, turn or river board and two player ranges. The fiddly bit is inputting the ranges, but I'm sure I can sort that out.
2) The fact that you can't put coloured text or suit symbols into the text box of a MATLAB gui is seriously tilting me. If anyone can make anything of this, earn a gold star, and then explain it to me please. I've asked my friend Doug Hull, who writes poker books and works as a MATLAB expert for the MathWorks, if he can help.
3) I worked out a structure for my data that stores it in what I suspect is the optimal way, with no wasted space. I'll code it up this weekend and see how much space I save. This should give me the way in to using larger game trees (more bet sizes).
4) I read a very interesting paper by Sandholm last night. I didn't follow the details, but basically it seems to be a framework that allows you to optimize the bet sizes during an MCCFR run. The paper's a bit impenetrable (surprise, surprise) but the idea is out there. If, with a range of bet sizes, the CFR solution starts to converge to just one or two, this could potentially allow the game tree to be pruned, and the bet sizes adjusted to maximise the EV of the strategy. This then really would be a GTO solution of postflop NLHE. I have no idea whether I can get that working, but I'll bear it in mind.
5) One question that seems to me not to have an obvious answer is whether a GTO strategy would involve a single, optimal bet size for a given hand, or more than one bet size, i.e. a mixed strategy for bet sizing. It's not obvious from the solutions I've computed so far. For example, if the solution at some decision node is to check 50% and bet three quarters pot 50%, does than mean that the real GTO solution would involve always betting 37.5% pot, or some other approach? Once I have more bet sizes, it should be clear whether the solution homes in on some particular bet size, or sometimes bets high and sometimes low. If the former is true, Sandholm's approach could work; if the latter, it's not so clear. I have some ideas of my own too.......
6) Janda's example hands 6-10 involve fractional ranges, i.e. calling on the button with just one combo of AA and three of AKo, iirc. I can include this, but it's a right pain in the arse for little reward imo. I will, however, do it, because I want to solve those ten hands, but I'll be pissed off for the whole time that I'm coding that bit.
Mrsjb's off to sit a counselling exam in Sheffield this afternoon, so it's me, the kids and the ironing today. Ho hum....
Long wall of boring text too long to read..... are the zimmers still running 
Originally posted by roopopper
Originally posted by jbpatzer
I've had several interesting thoughts since last night. BTW, this is all about my code which, even though only me and Jules seem to be interested in it, is going to be the most awesme piece of poker software ever written (my unbiassed opinion), so if you want some lolz, go read a young person's blog.1) I don't know why it hadn't occurred to me before, but it would clearly be fiddly, but straightforward in principle, to set the code up to start from the turn or river, thereby reproducing all of the functionality of GTORangeBuilder, and more, but without the sexy graphics. I didn't really set out to do this, but these are easier problems than the one I'm solving. Running it from the river would produce pretty much instant results, and from the turn, would probably take minutes at most. I could envisage this possibly being a commercial product, but tbh I doubt whether I cba to put the effort in to trying to make money out of it, because I suspect the market remains limited. It seems to me more likely that I'll eventually release a gui that allows you to produce data files after inputting a flop, turn or river board and two player ranges. The fiddly bit is inputting the ranges, but I'm sure I can sort that out.
2) The fact that you can't put coloured text or suit symbols into the text box of a MATLAB gui is seriously tilting me. If anyone can make anything of this, earn a gold star, and then explain it to me please. I've asked my friend Doug Hull, who writes poker books and works as a MATLAB expert for the MathWorks, if he can help.
3) I worked out a structure for my data that stores it in what I suspect is the optimal way, with no wasted space. I'll code it up this weekend and see how much space I save. This should give me the way in to using larger game trees (more bet sizes).
4) I read a very interesting paper by Sandholm last night. I didn't follow the details, but basically it seems to be a framework that allows you to optimize the bet sizes during an MCCFR run. The paper's a bit impenetrable (surprise, surprise) but the idea is out there. If, with a range of bet sizes, the CFR solution starts to converge to just one or two, this could potentially allow the game tree to be pruned, and the bet sizes adjusted to maximise the EV of the strategy. This then really would be a GTO solution of postflop NLHE. I have no idea whether I can get that working, but I'll bear it in mind.
5) One question that seems to me not to have an obvious answer is whether a GTO strategy would involve a single, optimal bet size for a given hand, or more than one bet size, i.e. a mixed strategy for bet sizing. It's not obvious from the solutions I've computed so far. For example, if the solution at some decision node is to check 50% and bet three quarters pot 50%, does than mean that the real GTO solution would involve always betting 37.5% pot, or some other approach? Once I have more bet sizes, it should be clear whether the solution homes in on some particular bet size, or sometimes bets high and sometimes low. If the former is true, Sandholm's approach could work; if the latter, it's not so clear. I have some ideas of my own too.......
6) Janda's example hands 6-10 involve fractional ranges, i.e. calling on the button with just one combo of AA and three of AKo, iirc. I can include this, but it's a right pain in the arse for little reward imo. I will, however, do it, because I want to solve those ten hands, but I'll be pissed off for the whole time that I'm coding that bit.
Mrsjb's off to sit a counselling exam in Sheffield this afternoon, so it's me, the kids and the ironing today. Ho hum....
Long wall of boring text too long to read..... are the zimmers still running
There is no hope for you. No, no zimmers. Want to start it up again?
Originally posted by jbpatzer
Originally posted by roopopper
Long wall of boring text too long to read..... are the zimmers still running
I don't think "running" is the right word for zimmers. "Shuffling" is about all I can manage these days....
I'm game for a game though.
Shuffle up and deal!
Nice to hear from you again, roo
--VS
Originally posted by jbpatzer
Originally posted by roopopper
Originally posted by jbpatzer
I've had several interesting thoughts since last night. BTW, this is all about my code which, even though only me and Jules seem to be interested in it, is going to be the most awesme piece of poker software ever written (my unbiassed opinion), so if you want some lolz, go read a young person's blog.1) I don't know why it hadn't occurred to me before, but it would clearly be fiddly, but straightforward in principle, to set the code up to start from the turn or river, thereby reproducing all of the functionality of GTORangeBuilder, and more, but without the sexy graphics. I didn't really set out to do this, but these are easier problems than the one I'm solving. Running it from the river would produce pretty much instant results, and from the turn, would probably take minutes at most. I could envisage this possibly being a commercial product, but tbh I doubt whether I cba to put the effort in to trying to make money out of it, because I suspect the market remains limited. It seems to me more likely that I'll eventually release a gui that allows you to produce data files after inputting a flop, turn or river board and two player ranges. The fiddly bit is inputting the ranges, but I'm sure I can sort that out.
2) The fact that you can't put coloured text or suit symbols into the text box of a MATLAB gui is seriously tilting me. If anyone can make anything of this, earn a gold star, and then explain it to me please. I've asked my friend Doug Hull, who writes poker books and works as a MATLAB expert for the MathWorks, if he can help.
3) I worked out a structure for my data that stores it in what I suspect is the optimal way, with no wasted space. I'll code it up this weekend and see how much space I save. This should give me the way in to using larger game trees (more bet sizes).
4) I read a very interesting paper by Sandholm last night. I didn't follow the details, but basically it seems to be a framework that allows you to optimize the bet sizes during an MCCFR run. The paper's a bit impenetrable (surprise, surprise) but the idea is out there. If, with a range of bet sizes, the CFR solution starts to converge to just one or two, this could potentially allow the game tree to be pruned, and the bet sizes adjusted to maximise the EV of the strategy. This then really would be a GTO solution of postflop NLHE. I have no idea whether I can get that working, but I'll bear it in mind.
5) One question that seems to me not to have an obvious answer is whether a GTO strategy would involve a single, optimal bet size for a given hand, or more than one bet size, i.e. a mixed strategy for bet sizing. It's not obvious from the solutions I've computed so far. For example, if the solution at some decision node is to check 50% and bet three quarters pot 50%, does than mean that the real GTO solution would involve always betting 37.5% pot, or some other approach? Once I have more bet sizes, it should be clear whether the solution homes in on some particular bet size, or sometimes bets high and sometimes low. If the former is true, Sandholm's approach could work; if the latter, it's not so clear. I have some ideas of my own too.......
6) Janda's example hands 6-10 involve fractional ranges, i.e. calling on the button with just one combo of AA and three of AKo, iirc. I can include this, but it's a right pain in the arse for little reward imo. I will, however, do it, because I want to solve those ten hands, but I'll be pissed off for the whole time that I'm coding that bit.
Mrsjb's off to sit a counselling exam in Sheffield this afternoon, so it's me, the kids and the ironing today. Ho hum....
Long wall of boring text too long to read..... are the zimmers still running
yes yes let's have a Zimmers reunion
Originally posted by VorpalF2F
Originally posted by jbpatzer
Originally posted by roopopper
Long wall of boring text too long to read..... are the zimmers still running
I don't think "running" is the right word for zimmers. "Shuffling" is about all I can manage these days....
I'm game for a game though.
Shuffle up and deal!
Nice to hear from you again, roo
--VS
Well that's three of us in, I had better head over to Tim's Blog next 
Originally posted by roopopper
Originally posted by VorpalF2F
Originally posted by jbpatzer
Originally posted by roopopper
Long wall of boring text too long to read..... are the zimmers still running
I don't think "running" is the right word for zimmers. "Shuffling" is about all I can manage these days....
I'm game for a game though.
Shuffle up and deal!
Nice to hear from you again, roo
--VSWell that's three of us in, I had better head over to Tim's Blog next
Originally posted by Tim64
Hello mateI'm in especially if we can spin n go for random prize of $.10 - $1 ♦
Next stop Gadget's blog ![:]](https://www.pokerstrategy.com/_forum-runtime/wp-content/uploads/wpforo/emoticons/pokerstrategy-legacy/basics/pleased.gif ":]")
Originally posted by jbpatzer
a list
I had a bit of a holiday. Where the second half has been nursing the side effects from 'too much party'... back to it now.
1) GTORB is getting there with turn and river play. Last I spoke to him, I think he said he had some solutions down to around ~30min. I'm not sure how detailed his bet sizes are and stuff so I suppose it's not a great comparison. But it looks like his latest blog post (which has a pretty cool brainteaser) he's getting closer to releasing something publicly.
As for a commercial product. As is, yeah... As a tool, it's obviously very, very niche. But even so, you could probably make some pocket change out of it, if the effort of commercialising doesn't sap the fun out of it.
Regardless, you're thinking like an English professor. This is a stepping stone.
One day you have a little CFRM program, the next you are selling DeepMind to Google for $650m.
The day after...
5&6) I'd think that very, very often mixed strategies will win. So I'd lean towards some high bets and some low bets. We can see this when we look at some solved river situations right? I can't see why it would be different when adding more streets/opportunities.
But from a practical point of view I'm not sure it really matters too much. I'd guess, they will usually be pretty damn close in EV. Perhaps you could do both for a couple of hands and compare the EV's?
Close enough is good enough for me, when I'm usually going to be picking an exploitable line anyway, whilst staying as close to My_Best_GTO_Attempt as necessary.
But I suppose if there are big differences, it would be nice to see how these ranges are constructed.
I haven't gone over the J64 and 886 hands yet, but will later today.
Originally posted by jules97
Originally posted by jbpatzer
a listI had a bit of a holiday. Where the second half has been nursing the side effects from 'too much party'... back to it now.
1) GTORB is getting there with turn and river play. Last I spoke to him, I think he said he had some solutions down to around ~30min. I'm not sure how detailed his bet sizes are and stuff so I suppose it's not a great comparison. But it looks like his latest blog post (which has a pretty cool brainteaser) he's getting closer to releasing something publicly.
As for a commercial product. As is, yeah... As a tool, it's obviously very, very niche. But even so, you could probably make some pocket change out of it, if the effort of commercialising doesn't sap the fun out of it.
Regardless, you're thinking like an English professor. This is a stepping stone.
One day you have a little CFRM program, the next you are selling DeepMind to Google for $650m.
The day after...5&6) I'd think that very, very often mixed strategies will win. So I'd lean towards some high bets and some low bets. We can see this when we look at some solved river situations right? I can't see why it would be different when adding more streets/opportunities.
But from a practical point of view I'm not sure it really matters too much. I'd guess, they will usually be pretty damn close in EV. Perhaps you could do both for a couple of hands and compare the EV's?
Close enough is good enough for me, when I'm usually going to be picking an exploitable line anyway, whilst staying as close to My_Best_GTO_Attempt as necessary.
But I suppose if there are big differences, it would be nice to see how these ranges are constructed.I haven't gone over the J64 and 886 hands yet, but will later today.
Glad to see you're back as you seem to be the only person who's interested.
Once I've squeezed every ounce of performance out of my code (i.e. parallelized it and got it running on a server with shitloads of RAM and cores), let's choose a situation, I'll compute the crap out of it and we can write something together about it, just to show what it's possible to learn.
Originally posted by jbpatzer
Glad to see you're back as you seem to be the only person who's interested.Once I've squeezed every ounce of performance out of my code (i.e. parallelized it and got it running on a server with shitloads of RAM and cores), let's choose a situation, I'll compute the crap out of it and we can write something together about it, just to show what it's possible to learn.
Cool. I don't know why more aren't interested tbh... their loss ♠_biggrin:
Flop J:c6:h4:d 75% single raised pot
The preflop ranges, tend to focus a little more on low connected hands than higher equity hands with blockers, like suited aces and even KJo in MP's opening range. But they're pretty good. About the same width.
Flop
Janda Cbets 39% as MP, has no x/r range
As BTN, no raising range, calls 67%.
JBR cbets 30% as MP, x/r 17%!
As BTN, calls 58% and raises 5%!
First thing I noticed that seems odd is, how as MP, JBR handles K:dQ:d (and K:hQ:h) compared to K:c:Q:c. cbetting ♦ 96% and the other 56%. My initial thought was there isn't much difference. Which then moved on to, well one has a Royal Flush draw... so still not much difference. So I'm not sure what's going on there. They're both pretty similar in EV, so I guess it doesn't really matter. Maybe a royal flush draw matters? Perhaps my strategy of never folding a royal flush draw has merit...
JBR obviously loves it's mixed strategies. Where practically for humans it's more common to balance ranges with whole groups of combos. For example. JBR here will cbet sometimes with a hand like K:dT:d and check sometimes for uses later on.
Where most human players, will choose to do something like always cbet K:dT:d and have something else in their checking range, perhaps say A:d:Q:d. The human approach has the advantage of being easier, but the disadvantage of a more transparent range.
Whilst it's impossible for humans to balance individual combos as detailed as a machine does. It is definitely possible (with the help of some kind of pseudo random number generator, i.e. a clock/suits on flop or whatever) to play more balanced, mixed strategies than currently commonly used.
This is something that you mentioned JB in your post earlier, point 5 I think. And I said that I think mixed strategies will be MORE GTO, but probably not too important. However, thinking about it more, mixed strategies certainly have far greater information hiding. And for that reason alone, in a practical sense (maybe not against a GTO bot?) where humans can exploit transparent ranges, mixed strategies should come out on top.
The way TT is handled kind of demonstrates this.
Janda cbets TT no the flop. For I'm guessing similar reasons JBR cbets sometimes 77-TT for a similar number of combos. This makes sense to me, but JBR has a less transparent range, so I think it's doing a better job.
JBR cbets it's OESD's 75s at as high percentage as any cbet, which have about as much EV, 6bb, as a weak top pair.
Only one gutshot, the nut one 78s has 3.3bb EV, better than AK at 2.2bb.
Now, about JBR's flop x/r range as MP.
This is a spot where very, very many humans players don't have a x/r range, Janda didn't. Equities even, equity distribution even & a very dry flop.
So, what's JBR raising?
By the way, that (63% of combos) stuff is awesome.
JBR raises ~60% of set combos it gets there with, in a fairly even spread, weighted towards the better ones. Which is about 36% in total. So lets say equivalent of one set combo.
It also does a handful (about 20% of total combos) of an even mix of QQ-AA. And about bout 25% of total AJ combos. So in human terms, we could put that as approximately the equivalent of all AA combos.
For bluffs, it's picking mostly the worst 6x & 5x combos. So mid set blockers, which looks pretty good. And then it throws in a very mixed range of stuff like gutshots, broadways/top_pair_blockers+bdfd's (i.e Q:dQ:d).
So I think the range looks fine. The question is, why or why not have a x/r range? I think it's a bit of a fad to not have a x/r range on dry flops. I'm not really sure why. Probably because peoples checking ranges are terribly broken...
I think there we have 2 pretty important things to take away and I only got as far as the flop and MP's x/r. 
oh, I just realized in the above post where I talk about mixed actions. And previously you were talking about mixed bet sizes... my bad, not related... nothing to see, keep moving
Originally posted by jules97
oh, I just realized in the above post where I talk about mixed actions. And previously you were talking about mixed bet sizes... my bad, not related...
Great post. Keep it up. I should be able to release the 0.5 pot, 0.75 pot, pot version of Janda hand 1 soon. That will be interesting.
Originally posted by gadget51
Well well if it isn't the roo, hello there! I suppose I've been volunteered then have I? ♦In other news, wtf is wrong with zimmers shuff-el-in anyway! ♠_love:
Be well and run good and wear slippy shoes.
looks like the old gang is coming together 
Should we try and recruit some fresh oldies too ?
Btw does the crazy Argentinian still lurk around these forums?
Originally posted by roopopper
Should we try and recruit some fresh oldies too ?Btw does the crazy Argentinian still lurk around these forums?
The more oldies the better.
Crazy Argentinian is rock climbing in Yosemite Park. She's got worse since you last spoke to her.