I actually wrote something about real fe previously,

I calc your real fe of your example, using the ev to reverse calc the real fe,
For 8s5x it's 0.503,
For 8x5s it's 0.5012,
For other it's 0.49535,
but unfornately I can't see this info from your ui.


We still fill up worse hands before filling up better hands for the same type of hands,
for different type of hands, we need to use ev comparison.

You said earlier indifference principle is useless, but unfornately indifference principle for
mixed hands in equilibrium still holds true. Otherwise I can construct a higher ev line
and then by definition your's solution are not GTO(I have sent your construction formula before).

I had asked how to identify spots "0% if solving down" at previous street,
So by your answer I suppose your imply meaning is we won't be able
to unless we have supercomputer?



happen to see this post when I'm writing my post, but the downlink in that post doesn't work?

This feature (to see how many combos there are in opponent's hand assuming we have a specific hand) is going to be available in the near future.

By "indifference principle" I meant specifically what was used in Mathematic of Poker to solve toy number games. You assume there is an indifference point and you try to find it. This is useless for real poker because:
-some hands are not mixed and you can't know which ones are mixed without solving it
-the order of hands are not as well defined as in number games; if you play with number you know that 1 is worse than 5 and 5 is better than 3. In real poker some hands might be better bluffs (even though they have lower equity), some might be better checks etc.

You won't be able to get to exact equilibrium.
Even University of Alberta when they published they Cepheus result (solving limit holdem HU) only calculated the solution to 0.1bb/100 and they used a lot of core years of computing power and could afford to use an algorithm that converges better and uses more RAM.
There will be an alternative Pio algorithm in the near future which converges to 0.01% (or better) on almost all trees in reasonable time (that still can be many hours) and it will use way more RAM but it's not available right now. The one we are shipping now has many nice properties:

-converges fast to good ballpark numbers, overall frequencies rarely change much once it's 1% exploitability or so
-uses little RAM
-doesn't fluctuate much (for most trees)

but it is not as good at nailing last 0.1% of exploitability. Good thing is that it doesn't matter at all for any kind of practical purposes. If you really want to solve a given river more precisely just copy the config from there, make a new river tree (there is a feature that does that or you automatically) and solve that.

Hey my dedicated server just got shut down and I would like to deactivate my licence there so that I can activate it somewhere else. Unfortunately I do not have access anymore.
Licence email is [email protected]

Kind regards,

How can I download one more time if I deleted without previous deactivation?

Hey guys, question.

Recently, when I've been using PIOsolver, whenever I change trees (e.g. click check/ bet 15 etc, or with hotness chart, select a new turn card etc.), I have to click the strategy button again, before it shows me the new strategy. It used to auto switch such that I did not need to re-click the strategy tab again for it to update.

Is there a setting to change this back?

since the 0% accuracy setting is impossible and impractical for us to use,
I want to ask a question, would pio stick to use wrong hands to bluff due to his
history computation? Which I mean, let's say, a previous iteration, your program
calcs bluff with K-high has ev -0.005, and then bluff with A-high has ev +0.008,
so it choose to bluff with A-high instead of bluff with K-high, after several iterations,
it still stick to bluff with A-high when it reaches its "equilibrium" since it's still within the "small accuracy", but in true equilibrium we may bluff with K-high instead of A-high.

Also you said earlier toy game is wrong and useless, why is that? The toy game has two assumptions:
1.there's no worse hands than bluff
2.you play a mixed strategy of bluff
so if there's worse hands than bluff, you simply remove it, the remains defending
hands still satisfy (1-a) argument for the bluff.
(Even the hand equal to bluff should be removed)

It's the best to not post any private information publicly.
Please follow the instructions here:
https://licences.piofiles.com/

and if that doesn't work please email [email protected]

There is a link to the installer in the link above.

Yes, there is a checkbox in the browser tab on the right side. This one:
https://gyazo.com/cefa61fc0afa1892a6ca49159199f7a9

(we recommend checking square size proportional to weight as well)

Probably not although in the big tree there are bound to be some hands which are not in correct ballpark even at 0.15% frequency. My tests with algorithm converging to 0% though show that it's very rare and usually the perfect equilibrium is about adjusting frequencies slightly, not changing which combos bluff unless it's very very close in EV to begin with.

Yes but the way it works is that it will slowly go in the direction of K-high. Again, you can solve a specific reason to 0% (almost) and see.

Both are incorrect in Holdem. For example in the river example I provided above 8s5h bluffs 100% of the time and the bluff is +EV.
Other hands never bluff and bluffing with them would be -EV. Even though some of those never bluffing hands have lower equity than some of the bluffing hands.
1-a doesn't work in poker. It's a decent approximation in some specific lines (for example bet/call, bet/call, bet/call usually the 2nd and 3rd call are close to 1-a frequency) but it doesn't work at all for other lines. The reason for it is that EV from the whole line matter not the EV at the last decision. You can have profitable bluffs at the river with hands that shouldn't be there in the first place as calling on previous streets gives up too much EV.
Usual argument for 1-a is: "if you fold more they can bluff any two cards and profit". This doesn't work in multistreet games because even though they could profit with any two cards on the river it would cost too much to get there in the first place so we can happily fold a lot. If you run aggregation reports for example you will see that folding frequency varies according to turn/river cards as well as betting line. Sometimes they are as high as 90% on some runouts and as low as 50% on others.
Runouts matter as well. Just because you are folding 60% on a specific reason doesn't mean the opponent can profitably call the turn because you might be folding only 40% on most rivers and they don't know which card is going to land on the table.

As an example I solved a very simple single raised pot HU example and looked at 3barrel line (IP is the one barelling). The board is Js 7s 2c 5s, calling frequencies on various river cards:
https://gyazo.com/692574eb8487d8751acdee2715ca55de

You can see that folding frequency is as high as 62.5% on Ks and as low as 44% on many other cards. This can't be explained by 1-a reasoning.
This is also the simplest possible spot (3barrel from IP and calls from OOP). There are situations where those discrepances are bigger.

Still question about accuracy, in your graph in your range explorer,
now we know that a lot of this hand won't exist in equilibrium,
but it just clutters the graph, would it possible filter out these hands?
Like say, don't display hands with freq <5%?

This can be explained by 1-a reasoning.

I've constructed a very simple example,
using board 2c 2d 2h 3s 2s,
OOP range, AA,KK,QQ,88 and 0.5 66
IP range,JJ,TT,99,77

starting pot, 100, let OOP bet 50





you can also tried bet 33,still holds

so that's a two step process,
first, you find the bluff hands which takes a mixed stragety,
which means its ev_check=ev_bet,
then you remove the hands worse than it, as you can see, 77, doesn't count,
since it will lose to 88 no matter we bet or check, so we don't take him into consideration.
After removing 77, the remaining hands satisfy (1-a) constraint.
Also you can see that OOP 66 is bluffing 100%, since it's strictly ev_bet>ev_check,
it takes a pure stragety.


I also adds QQ to IP's range, as you can see, still satifised 1-a.
a=0.46*3/4=0.34

Also I want to ask a question about stack size,
since we're impossible to run infinity stack size sims,
Let's say I run a sim for stack 100,
how much ev loss would be if I play the hand same as stack 100 when we're deep, say stack 200?
the function of ev should be smooth right? Which I mean, it's similar to play a hand when stack is 100 vs when stack is 101,
also it's similar to play a hand when stack is 101 vs when stack is 102, so we can give some reasoning about this right?
but when the gap becomes bigger, I'm not sure how much ev loss would that be

And then you constructed a toy game without card removal to show it :-)
I mean, I am not really interested in discussing it anymore. My posts contain enough information for you understand this subject.

It's not well defined what "play the hand same as stack 100" means as there are either more options or different options with bigger stacks.
I don't know either. You have the solver to run your own experiments though.

sorry if this has been posted already but what are the main benefits of piosolver edge over pro?

Still question about accuracy, in your graph in your range explorer,
now we know that a lot of this hand won't exist in equilibrium,
but it just clutters the graph, would it possible filter out these hands?
Like say, don't display hands with freq <5%?
You can see many hands like T6s really clutter with the graph.

I have no way to verify with card removal effects, since we're unable to see real fe right now, right? I can only predict even with card removal effects, the 1-a still holds true

I am running simulations over 1755 flops using Flop: IP single bet size [no donk]; Turn: OOP 2xbet sizes [no donk] & IP 4xbet sizes; River: OOP 2xbet sizes [2xdonk] & IP 4xbet sizes. I am solving to 0.50%.

I've finished roughly 145 flops, all of which each took 50 minutes or less (at least the ones I've viewed, which is many of them). My current simulation has been running for nearly three hours and keeps fluctuating between ~0.80% to ~1.10%. I know this is part of the process, but in general what would cause this particular flop analysis to do this when the others did not? (I'm not sure which flop it is...)

Also, the times between each cycle are only ~80 seconds, compared to a couple of minutes in the other simulations.

Sometimes specific flops are difficult for the solver. Can you send us the tree config and the board you have trouble on to [email protected]? I will take a look if something can be done to improve things.

This suggests that it's either same rank or same suit flop. Sometimes on those solver has problems converging. Again, please send us the config and I will see what can be done. If it doesn't solve you can kill the script, edit it manually to remove that one flop (or change accuracy on it) and then finish all the rest.
Meanwhile I will test if changing some parameters makes convergence better.

Hello, have you seen my question?
Is it possible to filter out clutter hands in the graph?

It's not possible at the moment.

I sent the offending tree: It was 9h9d9c. As an experiment, I let it run for 18 hours and it didn't solve (which I expected).

When this occurs, if I would examine several "0.80%" solutions (the closest it got), how close would their strategies be to each other?

Hi again,

I'm sorry, I don't completely understand that first part. Actually, I didn't get the 2nd part about weight either, but it doesn't matter. All of this is a bit overwhelming and probably a little too advanced for me.

Am I in the ballpark here? The PIO subset is created to get as good estimations of EV/EQ on different flops? And the Tipton subset for example is created to contain as many of the different characteristics of the different types of flops? Is it not possible to do both? Or what's the difference?

I've tried reading different posts on forums on picking subsets and also the official PIO blog post, but this basically only made me more confused.

Either way - thanks again for a great product and support.

/Hall

Let me try again then:

1)There are 22100 different flops in Holdem

2)We notice that some of them are strategically the same. For example As9s2s is the same as Ah9h2h if we just calculate results for As9s2s and then change spades to hearts in all hands then we get results for Ah9h2h

3)We can now group flops so strategically the same ones are in the same group

4)It turns out some of those groups have 4 flops (As9s2s, Ah9h2h, Ad9d2d, Ac9c2c), some has 12 flops (AsKsQd and all combinations of AK being the same suit and Q being another suit) an some has 24 flops (rainbow unpaired boards).

5)If we now list all the groups there will be 1755 of them. We choose one flop for every group and assign appropriate weight to them: if a flop comes from a group of 4 flops it gets weight of 4, if it comes from a group with 12 flops it gets weight of 12 and if it comes from a group of 24 flops it gets assigned a weight of 24. This way you get 1755 flops with weight which represent all the 22100 flops

6)If you now calculate results on those 1755 flops and then calculate weighted average on them using the weight from point 5) you will get the same result as calculating 22100 results and taking simple average

7)The problem is that 1755 is still a lot of flops, maybe we could sacrifice some precision and only take a fraction of those and still get a result close enough for practical purposes. The task at hand is to find if we can have a subset of those 1755 flops with some weights that gives results close to calculating on all flops. Again, we only care about averages here: we want to know for example what's EV of QQ is when we face a 3-bet or how often we 4-bet from the button if BB 3bets us etc.

8)Intuitive approach to the task described in 7) is to find flops which has certain characteristics (A-high, K-high, Q-high, flush draw, straight draw, connectedness etc.) and try choose them in such a way that resulting subset has characteristics similar to what occurs in all flops. The hope is that a subset constructed this way will yield good results when calculating averages results on them (good = similar to calculating on 1755 flops)

9)This approach was tried by Will Tipton and then improved by some of our users to include more flop characteristics. It works somewhat OK but not great.

10)It turns out that algorithmic approach that doesn't care about flop characteristics but just search for the best subsets in not so clever way (the description of this algorithm is a bit too complicated for this post but it comes down to trying a lot of subsets and then improving them gradually) gives much better results. In fact the results are so much better that even 25 flop subset beats 100-something Tipton's one and once you allow for around 90+ flops in a subset it gives almost perfect results.

11)We don't have intuitive explanation for the fact in 10). We know it's true because we (with help of some of our early days users) calculated a lot of solutions on all 1755 flops and then compared averages on all flops to what would happen if we used various subsets. Some of those benchmarks are published in the old blog post you mentioned.

Now commenting on your post:

It's created to give averages (EV/EQ/strategies) close enough to what would happen if you calculated them on all flops.

Tipton subset was created with the same goal in mind. It just doesn't work that well.

It is but it turns out that caring about flop characteristics is detrimental to the task of approximating overall EV/EQ well.

How many cores can Pio use? (Eyeing AMD Ryzen 3990x when released with 64 cores)

You gave me two improvements for my tree building

1. use a cap of 4 so 4th bet is rounded to all-in.
2. Build the tree, then press ctrl+b to get arbitrary solver command, type:
change_step 0.6

Questions:
a. What does the second step do?
b. After making these adjustments, can I finish running all my flops with this new configuration and then run reports over the entire set of flops (including those flops I ran with the old config)?

Right now up to 64 threads due to limitations of the compiler we are using. On a 64 core CPU you would need to disable hyperthreading to get max performance.

It tells the solver to start solving using smaller steps. This usually means worse convergence in the first few steps but better overall in case of more difficult trees
You can't as they would have a different structure so if you want to finish that for a report don't change the tree structure (don't introduce the cap for that batch of trees).

How do I add the "change_step 0.6" to a script (...containing 200 flops)?

You will sadly need to add that line after every build_tree command.
Maybe you can do that automatically in a decent text editor.
We will work on improving the defaults in the future. It seems our users like solving much bigger trees than we expected (often 4-5 or even more bet sizes) and back then when I worked on default solver settings I was mainly testing it on much smaller trees.

I've installed new hardware on my PC and now I need to reactivate my license. The license I used originally does not work at the moment because it says it is still active. What do I do to deactivate the license so I can use it again with my upgraded PC?

Hi,

About memory for ThreadRipper 2950x build.
Do you know - how much memory speed and latency affects sims solving time?

For example, I’m choosing between this options:
G.Skill
F4-2400C16Q-64GFX Flare X DDR4-2400MHz CL16-16-16-39 1.20V 64GB (4x16GB)
F4-2933C14Q-64GFX Flare X DDR4-2933MHz CL14-14-14-34 1.35V 64GB (4x16GB)
F4-3200C14Q-32GTZ Trident Z DDR4-3200MHz CL14-14-14-34 1.35V 32GB (4x8GB)

Would be difference somewhat near meaningful?

Thanks in advance.