Poker-AI.org

Seems to be a difference of general opinion on whether or not following solver outputs 1 to 1 would be able to beat the rake (without rakeback) in real money online 6max cash games. The popular opinion on poker forums seems to be that the "GTO" bot would crush the games at a high winrate. However, many of the GTO bot/RTA makers as well as some voices in the botting community seem to disclaim that following the solver solutions 1 to 1 will not be a profitable strategy when considering the rake - and that deviating from the outputs at obvious spots would be necessary to profit.

Does anybody have experience or data with this debate? And/or an opinion on it either way?

Thanks

Who do you think that understand the game better poker pros or botters ?

As you already know the poker community is clear on position that gto would crash. If you watch some game play where a pro is playing 500nl or higher you will see that he is basically trying to mimic gto as best as he can. In other words the best pros out there are just trying to play as close to gto as possible. At a high level this is what makes only sense.

You need to actually test it otherwise I dont think that is possible to estimate what win rate it can have. Rake back and every other bonus should be considered as part of your winnings. So if you paying 10 bb/100 rake in a site and you are getting 5bb/100 back through bonuses and rake back then your win rate needs to be bigger than 5 bb/100 to be winning. Now I can not show any data but only speculate. People make huge mistakes all the time and I refuse to believe that a shit reg that is playing on such high rake game is winning, where a good gto model with 3-4 sizes on every street would lose. The tricky part here is to have a decent model. You need to understand the game so you can judge where you can cut corners and where you need more precision. Every model is as good as the parameters you provide.
Having said that obviously there are some spots where your win rate could be improved by deviating from gto but I dont thing this is necessity to make the gto winning.

To answer the question, following solvers like Pio probably isn't going to beat the best players because you have to make a lot of assumptions about what villain holds and what the possible actions are.
Edit: Last time Iooked PioSolver was a GTO solver for an idealised problem that assumes you know what villain holds and assumes a limited number of bet sizes. So it is GTO solver for a game we aren't playing.

But bots that solve the entire problem will beat the very best players. Pluribus beats high ranking pros at a rate of about 5bb/100 at 6 max NL holdem.
https://ai.facebook.com/blog/pluribus-f ... yer-poker/
It doesn't adapt its strategy to its opponents, so even greater win rates are possible
Edit ... even more so if its competition aren't pro players.


Software developers using machine learning techniques don't have to understand the game, though it helps a bit. Machines now beat all humans at backgammon, chess, go, poker and 99.8% of all humans at starcraft. All these AIs were developed by humans who would be absolutely crushed by their own creation.
Edit Developers are not writing programs that incorporate expert knowledge of game playing: rather they are developing programs which learn how to win through self play.

OP was asking a specific question. There is a difference of opinions and I can confirm this by myself. I have seen someone that is selling rta services claiming gto alone can not win etc. while providing some pure solver solutions with only 1 bet size per street. Of course such a bad model is not close enough to real gto solution in order to be profitable.

GTO does not have any assumptions what so ever. The only parameters that make the strategy the way it is are preflop ranges, rake, stack depth, and post flop bet sizes. Gto does not care if the game is tournament cash or whatever but again only take into consideration the previously listed parameters. Saying that your opponent need to have a specific hand or a specific range in one spot in order gto to work is a pure misunderstanding of gto theory. Poker community understand the theory much better compared to developers and this is exactly what I meant.

In a rake less game there is absolutely nothing a player can do in order to beat a good gto model by definition. With rake things are more complicated and there is no way to estimate if gto will win or lose although I really doubt it will lose even on the highest rake game.

Why are telling us this?

Why are you telling us this?

Who is saying this?

What is your evidence for this statement? Why is it even relevant?

Read the thread from the beginning and you will understand why I am telling this.

The reason why there is a difference in opinions is a misunderstanding of gto and you just confirmed this by telling that your gto need assumptions in order to work. I am not looking on arguing or anything but thinking that you need your opponent to have a specific holding or range on any street is a common mistake. If your opponent arrives on the river with a different range compared to gto that means he previously deviated. Any deviation is an immediate lose and since this is a zero sum game the ev goes to the opponent.
He may have no bluffs on the river and you would think that gto is losing but you are missing the most important point. Poker is played on 3 street and while gto would lose more money with bluff catchers on the river gto will gain ev from earlier streets with other parts of the range and it is guaranteed that the EV gained will always be more than EV lost.
This is not just an opinion but a pure fact and any decent player will agree with this. Maybe you disagree in which case I would like to hear what you have to say rather than simply saying that gto need assumptions to work.

I read the thread from the beginning before I wrote what I did. I didn't understand why you were telling me so I asked the question. Now please answer.

Where did I write that?

Looks like that to me.


Who is saying they think that?


Where did I write that?

Ok then I must misunderstood you. I apologize for that.

OK. I will alter my answer a bit to try to make it clearer,

Greatly appreciate the insight guys.

My question was more towards putting a solver like Piosolver (with sufficient bet size availability) into a 6max game. Seems agreeable that it would beat even tough games pre-rake, just not sure if it would win at a clip high enough to beat raked games. Can you really sustain a pre-rake 10+ BB/100 win rate without ever adjusting to imbalances of your opponents? Pluribus is a bit of a different discussion but even it had only a ~5BB/100 winrate before rake.

The vast majority of postflop Pio solutions are mixes. Many spots where there is an "incorrect" option to take will only cost villain a small amount of EV if chosen. Not sure if there are enough -EV decisions (according to Pio) being made to make up for the rake, unless villains are punting off stacks with massive -EV decisions. This is true for poker in general, but adjusting to these large imbalances and obvious fish seems to be where most of your EV as a player comes from.

Was hoping somebody had some data/experience with a pure solver based bot, I'm sure there are some out there by now. But any educated insight is welcome and appreciated.

Thanks again

With Pio, I really doubt it, though obviously it depends on the skill of the opposition. The only evidence I have is how much better Pluribus / Libratus were against the best AIs that preceded them. In theory at least, those predecessors should have been much better than Pio because they weren't approximating ranges and bet sizes (much)

But that is against pro players. Against amateurs it's going to be higher, by the margin that pros live on.

"Pluribus beats high ranking pros at a rate of about 5bb/100 at 6 max NL holdem."

If Pluribus wins with 5bb/100 it doesn't mean that live solvers would win as well. Pluribus is as well just trying to play GTO as much as possible, it doesn't have "blueprint" (gto) solution for each spot so it does abstraction and subgame solving to overcome this issue. I would say, it depends how many solutions live solver has to cover the entire game tree, will depend how much winrate it has.

Pluribus / Libratus use many bet sizes even preflop while live solvers use 1-2 bet sizes preflop. I think this is the main difference, and where live solvers loose big portion of their EV and would be heavily exploitable if opponent would use larger bet size than standard one. Considering that most people who use RTA play mid / high stake, there they rarely face unconventional bet sizes, as nearly all regulars use standard bet sizes, try to play like solvers. Only recretationals use non-standard bet sizes but human players don't need solver to beat them.

Besides this, Pluribus can play solid multiway, while live solvers offer no solution for multiway spots, or some rare preflop spots like e.g. limp/3bet. This is why people who use live solver need to be solid players so that they don't burn money multiway without solver hints.

Another important aspect is that solid human players who use live solver can play max exploit strategy against weak players which would yield much higher winrate than "GTO" strategy that Pluribus plays or someone who strictly would follow RTA hints.

This is why estimating winrate for RTA is very difficult, and depends how well human player can perform in spots without solution.

"because you have to make a lot of assumptions about what villain holds and what the possible actions are."

GTO doesn't require assumptions about what villain holds or possible actions. GTO is supposed to be unexploitable against any strategy or actions, assuming it has solution for any given spot. The only way to exploit
GTO bot or RTA would be if such bot or RTA is not having enough solutions to cover big parts of the game tree, and human knows exactly where such spots are and exploits it. E.g. if RTA / GTO bot, has only RFI size 2bb or 3bb,
and human knows it, it could exploit it by raising e.g. 10bb, as GTO bots / RTA usually round the actual bet size to the nearest possible one for preflop.
Pluribus / Libratus were trained with many different bet sizes even preflop, and humans were able to exploit it during the match using big bet sizes, but not for long enough as team of Libratus would run CFR for exactly those bet sizes that were missing, and tomorrow it would be able to play GTO for even such big bet sizes (i find it unfair to fine tune the bot between the matches exactly for the missing bet sizes).

The way RTA overcomes the issue of facing actions / bet sizes that it doesn't have solution is that human player takes over and plays its own strategy in such spots.

Yes, that's my point.


It's true that it doesn't have blueprints for all spots, but the live subgame solving is supposed to make up for that deficiency.

I didn't say it did. If you are using a GTO solver like Pio in 6 max games you have to make a lot of assumptions about what villain holds and what the possible actions, because you are using a GTO solver for only part of the game, and the GTO solution isn't for the game you are solving.

You are right that gto would need to round the actual facing bet size to match your gto model. But given the fact that most are using similar bet sizes like 33%, 66% and 100% do you think that it will matter if you treat an 75% like an 66%. Of course there will be spots where you need to hard code your bot because the action villain took can not be put into any category. For example villain goes all in a spot where gto never does.

As for raising 10bb preflop to counter gto I dont think that this is practical. First of all you would lose a lot of EV by doing so. what ranges are you going to use and how you are going to play post flop. Since you are raising so big you need to have a very tight range and the range your opponent will call such a raise will also be very tight. If you decide to use 3bb open ranges by raising 10bb you will just end up burning your money. You will win many small pots preflop and then lose huge pots post flop because you will face a very strong range and will often be dominated. People would also notice after some time and start 3 betting you light. I would start raising to 2-2.5x and now you find yourself putted 10bb raise with a shity range facing a 3bet. Furthermore it is easy to counter such a big raise. I would just code and treat such a raise like an all in and just call with top of my range.

In a short there is nothing a human can do against a good gto model with 3 bet sizes and some additional coding to counter situations where villain takes a zero frequency line. I agree though that most of the winning are coming from fish and while gto would still win against them it will not win as much as an experienced poker player but even here we can hard code our strategy and instead of using gto use this strategy against fish.

I know at least 1 legit person that is using a gto bot and it is working for him.

Pre-rake 10 bb/100 seems way to high. Just take a look at high stakes pros win rate to see what is possible. There are some players that have over 10bb/100 but those are exceptions and usually the highest win rates are achieved on the softest games like ignition.
A pure solver based bot is not possible because you will still face situations where opponents take a zero frequency line. That means your solutions does not have an answer to the current situation. Take an example where opponent donks while your gto solution does not have any donks in that spot. You need to hard code those situations.

At midtskaes, rake takes 6-7 BB/100 winrate from you. So to beat the games at a reasonable clip it requires winning ~10BB/100 before rake. The win rates you see posted from people are including rake. I do believe the consensus is that a solver could beat a tough game without rake, nobody just seems to know what realistic winrate would be; and if it even would be enough to beat the rake.

I do understand that there will be situations that a solver does not cover, such as the donk bets you mention and limped/multiway pots, etc. And that these will account for part of the win rate of the bot.

I agree that estimating RTA is going to be very difficult for the reasons you mentioned. I am trying to gauge a reasonable winrate for a bot that strictly follows solver plays 1 to 1, and is coded to play a basic solid strategy in spots that solvers don't cover. Or, the winrate of an RTA player that never deviates from the solver outputs, and plays basic/solid in spots not covered.

Opinions seem to be all over the place depending on the source.

Yes subgame solving is what made it win vs humans, difference between losing and winning. They made experiments without subgame solving, it was much worse. Pluribus and Libratus are basically brute force bots that thanks to powerful GPUs pre-calculated and saved huge amounts of blueprint strategies and then use subgame solving on turn and river where it is fast enough to calculate it while playing.
I don't find it impressive as humans learn games by very few samples and learn how to generalize, while libratus and pluribus iterate over almost whole abstracted game tree and write solutions on disk, it is difficult to even see elements of AI / ML in this. I find deepstack more interesting as it at least uses DNN to predict CF values, instead of brute force run all blue print strategies.