Same

the goat miss that guy bring him and raisedbyjews back


Sent from my iPhone using Tapatalk

Same problem

hello from the past

I heard you were dead.

What is this nonsense? Newff, Spamz, and Maru showing up in the same week?

C-c-c-combo post. Where's Zach?

Big plans to make a new cartel!

You're a legend spamz! You were my go to guy for hand analysis when I was last playing HUSNG's like 6 years ago. What's up man? Do you still play/coach?

Nah not at all. Went back to uni good 3 years ago and now working on my masters degree.

I'm back too.

Hi,

I posted an ad in the 'Classified Listings' section about my GTO project I have been working on for the last few years but since most people interested in this kind of thing are gathered here I thought to give you a 'heads-up' on its existence. Also since that forum doesn't allow replies, if you have any questions regarding my project please post here and I will answer (within reason of course!)

P.S.: Moderators I hope this does not cross any of your policies, if it does please delete this post but I sincerely hope it's ok to post about this here. Maybe it will generate a nice discussion about GTO solutions if nothing else.

This is the thread he's talking about:

http://forumserver.twoplustwo.com/19...oject-1643245/

As far as capabilities, what you developed is nothing new. Both Pio and SPF can calculate all flops, given enough RAM. I don't understand how yours stands out in any way. Is it faster? Can it reach higher accuracies? Does it require less RAM? What's so great about it?

As far as I looked at both SPF and piosolver preflop trees that are sold were generated on a limited set of flops (276 for spf, can't remember for piocloud but I think it was less than 200). I used those sets myself at some point but didn't get very good results. My code uses all 1833 unique flops, all turns and all rivers. Memory consumption is high too, on a machine with 512 GB RAM I can solve trees of max 400-450 nodes at the moment. As for speed it will solve a gametree in 8-12 days (depending on hand samples I set it to use when solving). I would appreciate if anyone has any link on how much pio or spf will solve a preflop game tree.

It's true that commercially available solutions don't use all flops. Instead they use weighted subsets that have been shown to be good estimates of all possible flops combined. However, anyone can calculate with all flops themselves with Pio, for example, if they have the edge version. All they have to do is rent a big enough server. So, again, I don't see what's unique about your software. It would have to be faster or more efficient or something like that.

This is what I was saying, I tried this in my project and compared to my results when using all flops they didn't look so good. So I would say that my project does produce better results as I bucket hands and not reduce flops. This does transform the game tree into a digraph (as hands on same cluster on the flop do usually get on different clusters on turn) hence the difficulty of transforming the algorithm to work with this. In the end I did all with four passes (one for each street) and this does work.

Have you personally used piosolver to generate preflop solutions or do you have a link for anyone using it on all flops (not a subset) ? Because my code can do exactly that here and now, as this was the goal from day one. I have a huge respect for piosolver, GTORB and the rest of the solvers but I do think that mine can be used right here right now to generate a complete headsup gto strategy given a game tree.

SPF as I see generated/sells only preflop solutions, piocloud sells complete gametrees for piosolver but they are solved on a restricted set of flops and focus on preflop solution. My solver generates and stores complete gametree. I would be extremely interested in links related to this, in order to compare other work if it exists.

And browsing the gametree will be the next step, so far I focused on generating the data, now I need a mean to view it. I have a flop visualizer that given a flop shows the strategy but I want to make it for the whole game tree.

Hope that answers your questions!

Unfortunately, I lack the expertise to understand this.

I have used Pio to generate preflop solutions. Pio can also save full game trees. But that seems very impractical because that's 100s of GB. When it calculates preflop solutions it calculates solutions for all streets. It's just that people usually only save strategies up to the flop because otherwise files would be humongous. But you can save the full tree if you want to.

Pio used calculations on all flops to create their subsets. That's what they used to measure how good subsets were. You can read about it here: http://piosolver.myshopify.com/blogs...the-whole-game

Ok but have you used all flops or subsets ?
Also how easy is to see postflop strategy for various flops ? If you want to see strategy for a flop not included in the subset what would you do ? (that is if you used a subset)
That is the main difference, what I do is generate the complete solution, pre and postflop and keep it for study. Still working on the visualization part as mentioned.

Thanks, I have read that a while ago.
My point is that while for starters taking a set of flops from them is better than taking the same set of random flops - nothing beats working on all flops.

I haven't used all flops because it would require a very expensive server.

Viewing postflop strategies works great. Try it yourself with the free version.

For flops not in the subset, you can simply use the preflop ending ranges (limped pot, MR pot, etc.) from the preflop calculation using subsets and do a flop calculation, which doesn't take long. It might even take less time than loading a 500GB tree that has all flops So I think the general consensus is that you don't really need whole precomputed trees.

No one will disagree with you there, but like I keep pointing out, Pio and SPF can already do this. It's just that people generally don't choose to for the reasons above.

If you don't use all flops when computing a strategy then the pf ranges you use for a new flop might not be so good approximations of the real ones and some errors will propagate IMO.

I used to do something like this in two passes
Generate first an approximation of preflop solutions in one pass then for all flops generate complete solution. It was more computationally expensive and I found out the solutions were not that good.

I am not in competition with any of the other solvers.
What my project does is generate the complete solution for a game tree in one pass using all flops/turns/rivers which I believe results in higher accuracy of the solution. Also considering that it does this in 8-12 days for a game tree I think it is pretty fast considering the number of samples but that is in the eye of the beholder. If anyone is interested in backing up this project good, if not also good.

Don't want to fight with you but I hope you realize the difference between theoretically possible and actually done. I asked for some reference to someone who already did this, because as you add more flops the memory requirements increase and the scale of all increases. I know because I had my share of revelations when I went from 144 flops to 900 to all 1833.

Maybe post some examples of how your results differ with all flops vs the current commercial methods.

I tested my code with subsets from piosolver vs my code with all flops.
Subsets usually tend to favour some hands, for example call openshoves w/ 54s or hands with 6 in them or something like that because no matter the subset chosen it will have some cards w/ different distribution than normal one. Best subset I tested was just around 300 flops, worst results I got with those around 100 flops.

I haven't generated results with commercial programs due to their costs, I prefer to write everything from scratch, but I did test my solver via free piosolver on some smaller examples and found similar results.

Still generating my game trees as we speak, if anyone would be kind to provide some preflop results obtained via commercial programs I would be more than happy to provide a preflop analysis vs my results - as soon as current batch is ready which would take at least 1 week from now.

To compare solutions, preflop and postflop configurations (bet sizes, etc.) would need to be mimicked exactly.

Even in that case since we are dealing with mixed strategies it's hard to evaluate by eyeballing a solution. Some things might pop out as anomalies like bad calling openshove range in a wrong solution or if a certain card (5/4 etc) gets called more than higher cards.

One test would be to play one strategy against the other but that's no guarantee as strategy A can beat strategy B which can beat strategy C which can beat strategy A.

Best would be to run best response on each strategy and see how exploitable is each. But need full game trees for both and one must make sure the best response code is 100% bug free.

Just my two cents on the subject ...

Pio automatically calculates exploitability for each calculation. So yes of course you would to compare that. What I'm pointing out is that it's meaningless to compare it without first matching configurations exactly. And you're asking for example solutions calculated with Pio without sharing your own configurations, so I don't see how they would be useful at all.

Partially I agree that it's impossible to automatically compare two strategies, especially if one uses a limited amount of flops and the other doesn't.

Still what I had in mind was for a given stack depth let's say 20bb deep which I am generating now - I provide my preflop solution and someone provides piosolver or other solver solution for the same depth and people here decide.