Poker-AI.org :: View topic - AlphaPoker realtime GTO assistance?

mlatinjo might be closer to the truth than he expects.

Speaking from experience: online gto is hard. Very hard.
Suppose you are on the flop, hu 100bb, single-raised pot. Quite common spot. Let's consider cfr+ as our solver.
UofA did 1k iterations on the flop in their DeepStack, Sandholm & Brown did 1k-1.5k iterations on the turn. Let's not push it and suppose that our solver gives acceptable convergence in 200 iterations - which is already a gross assumption.
Average reaction time in hu is 5 seconds, so we are expecting an average iteration time = 5000 ms / 200 iterations = 25 ms per iteration.
And that's where the big problems come. Remember that cfr+ is a vector-vector algorithm, so it must went through all nodes, all turns, all rivers. An average tree in our situation is 120-150 nodes. Solver must process single node with all boards in less than 1ms, and it's just plain impossible with current computation power.

So, their solution must be abstracted to unknown degree (knowledge brings fear, eh?

) or they are using a precomputed solution at least on the flop. Precomputed solutions are easily exploitable.

And even more, they are providing eq solutions in chinese rooms which is a very specific case: super deep stacks (300 bb+) and various antes up 1 bb, making eq computations even more complex.

All in all, I think they might provide _some_ eq solutions, but their quality is questionable.

Real time gto solvers are using precomputed gto solutions which are stored on the servers. It requires powerfull servers with a lot of storage. That is why it can provide gto solution in short time making it applicable real time. But off course they can't solve all spots, there are so many, and they can't solve multiway pots.

mlatinjo might be closer to the truth than he expects.

So, their solution must be abstracted to unknown degree (knowledge brings fear, eh?

) or they are using a precomputed solution at least on the flop. Precomputed solutions are easily exploitable.

And even more, they are providing eq solutions in chinese rooms which is a very specific case: super deep stacks (300 bb+) and various antes up 1 bb, making eq computations even more complex.

All in all, I think they might provide _some_ eq solutions, but their quality is questionable.

In regards to putting villain on a ‘range’, we provide a solid “default” game-plan, meaning we assume the villain is playing near optimal.

Hi Oleksandr nice to meet you and thanks for the nice explanations.

I thought that the intention of the tool is to play as a bot or that any user should follow 100% of instructions. But if the intention is that it displays solid ranges, and opponent
should use own skills to use that advice to make better strategic decision (even if making exploitative) in that case it makes sense for me.

What is not clear to me is why don't you use it as a bot if you already have to use stealth? It shouldn't be hard to improve it to play also exploitative vs weak opponents.
Also, why would you sell this tool to some guys if it is going to print money? why not keep edge for yourself? Also why only to limited number of guys (i actually think that it is kind of commercial only limited)? Is your tool able to read table information, read previous actions of villain's all automatically, or user has to do it manually?

Quote:

In regards to putting villain on a ‘range’, we provide a solid “default” game-plan, meaning we assume the villain is playing near optimal.

Does it mean that you define villain range using some set of predefind rules, or you use gto solution for villain range assumption?

You also wrote that it can handle multiple bet sizes, raise sizes. How many exactly? Is it just handling few bet sizes or any possible bet size and raise size? same question for stack sizes. For me that is the crucial question for real time "approximated gto solver"

Author:	Boreku [ Thu Apr 05, 2018 2:34 pm ]
Post subject:	AlphaPoker realtime GTO assistance?
New to forums, Not sure if in correct place to post this (please move if needed mods) has anyone used this software? a friend of mine suggested to take a look, looks like a great idea and support seem super friendly but also looks new to the market so im curious if anyone can give feedback before i try it out? https://gyazo.com/b3c6c79fd3f7318786c36ddd77c0e28c https://gyazo.com/d377853996ab9980994909d2beaff04c https://gyazo.com/c06bf1fe907153d591f25f2c5c845159 they are related to this site http://www.alphapoker.org

Author:	mlatinjo [ Mon Apr 09, 2018 9:15 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
And i guess you are the owner of alphapoker making here marketing

Author:	HontoNiBaka [ Tue Apr 10, 2018 1:53 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Well, it depends how close they are to GTO, hard to say how they abstarcted the game etc.

Author:	sn0w [ Tue Apr 10, 2018 2:51 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
mlatinjo might be closer to the truth than he expects. Speaking from experience: online gto is hard. Very hard. Suppose you are on the flop, hu 100bb, single-raised pot. Quite common spot. Let's consider cfr+ as our solver. UofA did 1k iterations on the flop in their DeepStack, Sandholm & Brown did 1k-1.5k iterations on the turn. Let's not push it and suppose that our solver gives acceptable convergence in 200 iterations - which is already a gross assumption. Average reaction time in hu is 5 seconds, so we are expecting an average iteration time = 5000 ms / 200 iterations = 25 ms per iteration. And that's where the big problems come. Remember that cfr+ is a vector-vector algorithm, so it must went through all nodes, all turns, all rivers. An average tree in our situation is 120-150 nodes. Solver must process single node with all boards in less than 1ms, and it's just plain impossible with current computation power. So, their solution must be abstracted to unknown degree (knowledge brings fear, eh? ) or they are using a precomputed solution at least on the flop. Precomputed solutions are easily exploitable. And even more, they are providing eq solutions in chinese rooms which is a very specific case: super deep stacks (300 bb+) and various antes up 1 bb, making eq computations even more complex. All in all, I think they might provide _some_ eq solutions, but their quality is questionable.

Author:	mlatinjo [ Tue Apr 10, 2018 4:41 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
In gto solvers, you define villain range and software calculates you approximated gto ranges. It is approximated because you selected limited number of bet sizes, raise sizes, and you defined villain range yourself. It is surely a very solid range and it is very good way of getting an idea how to construct solid unexploitable range. It is aslo approximated because villain range on the current street is defined. The huge difference between offline gto solver and real time gto solver is that real time gto solver has to define villain range itself (likely using what it thinks is preflop gto range then for each action what it thinks is gto range using limited bet sizes and raise sizes). But vs real opponent, which is not using gto range that a real time solver assumes, it is going to make huge mistakes in estimating value range. As an simple example, if villain who is very tight has on the river only nut hands, and gto solver assumes that villain has what it thinks is gto range, then estimated value hands are going to be totally wrong. That way also bluff frequency is totally wrong because value range is estimated totally wrong. Real gto ranges should be unexploitable vs any kind of villain range and not just vs 1 defined villain range and vs few bet sizes / raise sizes. As another example, lets say we are on the river, pot is 100bb both players have 200bb left. Gto solver might only solve spots with e.g. 33% bet size, 50% bet size, 75% bet size, 100% bet size and 200% bet size. But in this case villain bets 1bb. How is solver going to approach that spot? In offline solver you don't have this problem because villain can use only bet sizes that you define so gto ranges are calculated only for those bet sizes. Another big issue is balance of villain and own balance. Gto solver has to solve hands only for limited number of balances. E.g. for 20bb, for 40bb, 60bb, 100bb, 200bb. But what if villain has 68bb and hero has 58bb? Off course in all those spots, the only thing solver can do is approximate even harder, meaning that e.g. for 68bb balance it will provide you gto solution for the closes balance it has solved, in this example 60bb. And that way, stack to pot ratio is different, meaning that ranges are going to have again big mistakes.

Poker-AI.org http://poker-ai.org/phpbb/

AlphaPoker realtime GTO assistance? http://poker-ai.org/phpbb/viewtopic.php?f=26&t=3125	Page 1 of 2

Author:	mlatinjo [ Tue Apr 10, 2018 4:50 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
sn0w wrote: mlatinjo might be closer to the truth than he expects. Speaking from experience: online gto is hard. Very hard. Suppose you are on the flop, hu 100bb, single-raised pot. Quite common spot. Let's consider cfr+ as our solver. UofA did 1k iterations on the flop in their DeepStack, Sandholm & Brown did 1k-1.5k iterations on the turn. Let's not push it and suppose that our solver gives acceptable convergence in 200 iterations - which is already a gross assumption. Average reaction time in hu is 5 seconds, so we are expecting an average iteration time = 5000 ms / 200 iterations = 25 ms per iteration. And that's where the big problems come. Remember that cfr+ is a vector-vector algorithm, so it must went through all nodes, all turns, all rivers. An average tree in our situation is 120-150 nodes. Solver must process single node with all boards in less than 1ms, and it's just plain impossible with current computation power. So, their solution must be abstracted to unknown degree (knowledge brings fear, eh? ) or they are using a precomputed solution at least on the flop. Precomputed solutions are easily exploitable. And even more, they are providing eq solutions in chinese rooms which is a very specific case: super deep stacks (300 bb+) and various antes up 1 bb, making eq computations even more complex. All in all, I think they might provide _some_ eq solutions, but their quality is questionable. Real time gto solvers are using precomputed gto solutions which are stored on the servers. It requires powerfull servers with a lot of storage. That is why it can provide gto solution in short time making it applicable real time. But off course they can't solve all spots, there are so many, and they can't solve multiway pots.

Author:	sn0w [ Tue Apr 10, 2018 6:09 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
mlatinjo wrote: Real time gto solvers are using precomputed gto solutions which are stored on the servers. It requires powerfull servers with a lot of storage. That is why it can provide gto solution in short time making it applicable real time. But off course they can't solve all spots, there are so many, and they can't solve multiway pots. Ah, and that's the question: to which part "realtime" corresponds to? I was talking about computing eq in realtime from scratch, like DeepStack did. If you think that they are providing precomputed preflop+flops, then this solution (at least in hu) will be just straight bad. A year or two ago those precomputed flops were fine. But in today's games humans at nl200+ are exceptionally good at finding leaks in static action abstraction. And static solution will be even worse in 6max, especially in chinese rooms. As far as I know, both Simple Postflop and PioSolver quit their helpers based on static precomputed models. SP didn't evolve and develop their helper and it becomes outdated (info is from friend of mine who was their client). Piotr went to omaha. And why is that? Because there is no easy money in gto helpers. Despite their neurobullshit, DeepStack did one thing right - they were using continuous resolving. Evs are much more robust than ranges. Even bad evs are better than not-so-bad ranges when confronting with rapidly changing stacks and bet sizes. But it doesn't really matter if they are providing precomputed or realtime gto; gto alone won't crush chinese games where 50% of the time you are playing with effective stacks 300-700 bb in hu and where exists helpers specifically tailored for teamplay in 6max.

Author:	HontoNiBaka [ Tue Apr 10, 2018 6:16 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Do you think there is a good way to build a bot or advisor for Chinese sites? I have been thinking about it for some time now.

Author:	mlatinjo [ Tue Apr 10, 2018 10:34 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
even if gto helpes would be really increasing profit in HU games only, today it is so tough to find HU games anyway. Even at stars you see just several regulars waiting in HU room for fish to join. And if you join a table with regular, he will anyway not play against you, he will go to sitout after few hands and will write in chat "what a fuck is wrong with you". And if you play vs fish because you got lucky that fish joined your table (or weak reg), why would anyone want to use gto ranges vs weak opponent if max exploit strategy is going to print money. Regarding real time helpers i think the only help would be to show what hand to play preflop (using gto charts) for the games that is hard to memorize. E.g. in omaha or tournaments or spin & go it helps a lot. Or postflop to display how good opponent range and own range hits board by showing % of draws, strong hands, top pairs etc. And with that info players could increase proift likely 1-2 max BB/100, which is still not enough to risk using software that is against T&C.

Author:	Alpha_Poker [ Fri Apr 13, 2018 9:47 am ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Hey Guys, My name is Oleksandr; I am the co-founder of AlphaPoker. We had this discussion linked to us by a team member, and I would like to try to answer some of the points raised and clear up misconceptions on how and what we provide. I would like to add that this brochure was meant only for pre-release and was not ready for public viewing. However, your discussion has been productive and insightful; I think there are a few key takeaways which we will improve on, especially around our messaging. I will answer your questions keeping in mind that I need to be mindful of our IP. In the brochure we avoid promoting “GTO ranges” we were careful to provide approximations. We have 3 layers to our AI engine; this allows us to attack multiple pots, different stack sizes and raise bets. We have taken a blended approach to real-time solves and pre-solved solutions, this is what allows us to return results for all spots within seconds. Its fast, effective and fit for purpose. On the rare occasion, it does yield unusual recommendations, but these are sub 1%, and our team are actively making updates to our engine. In regards to putting villain on a ‘range’, we provide a solid “default” game-plan, meaning we assume the villain is playing near optimal. With expected values of each decision provided, you can use that information with any other “reads or exploitative data you have gathered” to deviate from the Alpha Poker recommendations. This is going to evolve rapidly, and we have started working on player-specific recommendations. Stealth has been one of the more manageable challenges to overcome, we are confident that unless specifically targeted, sites will not be able to detect Alpha Poker running. We also have caps on our player pool to reduce the risk of exposure. The goal is to increase profit by - Providing our members with decisions/lines that play a stronger more defined strategy. - The ability to play longer sessions also becomes more comfortable. - Other opportunity costs that professional poker players have to give up to stay ahead. sn0w wrote: mlatinjo might be closer to the truth than he expects. Speaking from experience: online gto is hard. Very hard. Suppose you are on the flop, hu 100bb, single-raised pot. Quite common spot. Let's consider cfr+ as our solver. UofA did 1k iterations on the flop in their DeepStack, Sandholm & Brown did 1k-1.5k iterations on the turn. Let's not push it and suppose that our solver gives acceptable convergence in 200 iterations - which is already a gross assumption. Average reaction time in hu is 5 seconds, so we are expecting an average iteration time = 5000 ms / 200 iterations = 25 ms per iteration. And that's where the big problems come. Remember that cfr+ is a vector-vector algorithm, so it must went through all nodes, all turns, all rivers. An average tree in our situation is 120-150 nodes. Solver must process single node with all boards in less than 1ms, and it's just plain impossible with current computation power. You are spot on! Currently, it is impossible to provide real-time `GTO’ solutions in sub 10 seconds. If someone is promoting this, you should think twice. We did not approach this in the same way as research institutes, as our objective for Alpha Poker is not aligned, we have no intentions of becoming a bot. sn0w wrote: So, their solution must be abstracted to unknown degree (knowledge brings fear, eh? ) or they are using a precomputed solution at least on the flop. Precomputed solutions are easily exploitable. As I mentioned before we use a blended strategy, but whether we provide precomputed solutions or use other methods, the opponent does not know of this. If I provide you OTB_Redbaron’s strategy on paper, of course, you can pick it apart and find ways to make money, until he realises you have this paper and then the meta begins. sn0w wrote: And even more, they are providing eq solutions in chinese rooms which is a very specific case: super deep stacks (300 bb+) and various antes up 1 bb, making eq computations even more complex. All in all, I think they might provide _some_ eq solutions, but their quality is questionable. Again you must consider what we are providing – an Assistance program. If you are one of the few players in the world with an eidetic memory and top notch exploitative game, then Alpha Poker’s value is questionable. However, we know our target demographic love to know their strategic options when facing tough decisions. mlatinjo wrote: Real time gto solvers are using precomputed gto solutions which are stored on the servers. It requires powerfull servers with a lot of storage. That is why it can provide gto solution in short time making it applicable real time. But off course they can't solve all spots, there are so many, and they can't solve multiway pots. You are assuming all services take this approach. (if any exist for nlhe cash games - please share) We appreciate all the comments, feedback and thought-provoking discussion. It is great to see a community of poker enthusiasts, professionals and hackers come together and push the boundaries. If you would like to ask any further questions, feel free to email me (My email is in the brochure). Oleksander)

Author:	mlatinjo [ Fri Apr 13, 2018 10:16 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Hi Oleksandr nice to meet you and thanks for the nice explanations. I thought that the intention of the tool is to play as a bot or that any user should follow 100% of instructions. But if the intention is that it displays solid ranges, and opponent should use own skills to use that advice to make better strategic decision (even if making exploitative) in that case it makes sense for me. What is not clear to me is why don't you use it as a bot if you already have to use stealth? It shouldn't be hard to improve it to play also exploitative vs weak opponents. Also, why would you sell this tool to some guys if it is going to print money? why not keep edge for yourself? Also why only to limited number of guys (i actually think that it is kind of commercial only limited)? Is your tool able to read table information, read previous actions of villain's all automatically, or user has to do it manually? Quote: In regards to putting villain on a ‘range’, we provide a solid “default” game-plan, meaning we assume the villain is playing near optimal. Does it mean that you define villain range using some set of predefind rules, or you use gto solution for villain range assumption? You also wrote that it can handle multiple bet sizes, raise sizes. How many exactly? Is it just handling few bet sizes or any possible bet size and raise size? same question for stack sizes. For me that is the crucial question for real time "approximated gto solver"

Author:	Alpha_Poker [ Mon Apr 16, 2018 2:04 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Hi Mlatinjo, Thanks for your response and glad that I could clear up our approach. mlatinjo wrote: Hi Oleksandr nice to meet you and thanks for the nice explanations. I thought that the intention of the tool is to play as a bot or that any user should follow 100% of instructions. But if the intention is that it displays solid ranges, and opponent should use own skills to use that advice to make better strategic decision (even if making exploitative) in that case it makes sense for me. The objective of our software is not to ruin the game We love. Poker is evolving, more bots are appearing, collusion is prevalent online and there are closed groups that have built their own software which aren’t publically available. mlatinjo wrote: What is not clear to me is why don't you use it as a bot if you already have to use stealth? It shouldn't be hard to improve it to play also exploitative vs weak opponents. Also, why would you sell this tool to some guys if it is going to print money? why not keep edge for yourself? Also why only to limited number of guys (i actually think that it is kind of commercial only limited)? Is your tool able to read table information, read previous actions of villain's all automatically, or user has to do it manually? Our team are technologists at heart. We enjoy the research and data science element of poker. Alpha Poker allows us to do something we are passionate about and support others. In saying that, one of our team members is a professional poker player for many years and provides great feedback by using and testing our software. mlatinjo wrote: Quote: In regards to putting villain on a ‘range’, we provide a solid “default” game-plan, meaning we assume the villain is playing near optimal. Does it mean that you define villain range using some set of predefind rules, or you use gto solution for villain range assumption? You also wrote that it can handle multiple bet sizes, raise sizes. How many exactly? Is it just handling few bet sizes or any possible bet size and raise size? same question for stack sizes. For me that is the crucial question for real time "approximated gto solver" Alpha Poker does automatically read all table action in real-time and presents the spot within seconds to the user. We do the latter and approximate ranges as best as currently possible. We have built Alpha Poker to handle, solve and return a response for all scenarios. Our blended strategy returns very accurate approximations.

Author:	HontoNiBaka [ Mon Apr 16, 2018 4:22 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Do you use OCR to get the table state? The way most emulators lagg, I don't think it will be very stable.

Author:	mlatinjo [ Tue Apr 17, 2018 11:01 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Do you have solid stealth? Considering that your tool is against T&C of most famous poker sites, how do you hide your tool? If somebody playing mid or high stakes uses your tool it is a high risk of getting banned and risk losing a lot of money in account.

Page 1 of 2	All times are UTC
Powered by phpBB® Forum Software © phpBB Group http://www.phpbb.com/

Author:	HontoNiBaka [ Wed Apr 18, 2018 10:41 am ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Those Chinese sites don't have strong security, also as long as the software does not run inside the emulator, I don't see how it could be detected.

Author:	mlatinjo [ Sun Apr 22, 2018 9:54 am ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Ok so as i understood now, alpha poker can be just used on Chinese sites like pokermaster? Those sites are also very rigged and full of collusion, due to low security.

Author:	HontoNiBaka [ Sun Apr 22, 2018 7:22 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
There is a lot of collusion, but I don't think it's rigged. I kmow a few winners on PM and they claim to be gaming labs certifief.

Author:	mlatinjo [ Fri Apr 27, 2018 10:20 am ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Well having plenty of collusion, 5 accounts at the same table sharing hole cards is making it very hard to win. You say it is not rigged, but if site allows them to play like that how is it not rigged? Or what is your definition of rigged?

Author:	HontoNiBaka [ Fri Apr 27, 2018 4:12 pm ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
Rigged for me would mean, if the site itself manipulates something, like using an unfair random number generator or super user accounts. I had many people claiming, that the site will create set over set situations and so on more often, because it generates more rake, but I never observed that. Collusion is just missing security, it's not the site's fault directly, I mean you can bot on most sites, but I still would not call them rigged.

Author:	mlatinjo [ Sun Apr 29, 2018 1:41 am ]
Post subject:	Re: AlphaPoker realtime GTO assistance?
.