[article] How does Artificial Intelligence learn to bluff at poker?
I have been a recreative player since 2008. I've spent so much much time on this forum. I was a big fan of High Stakes Poker, the legendary heads-up battles between Isildur1 and Tom Dwan, and of course, Phil Galfond’s challenge.
I’m also an artificial intelligence engineer. During a break in my career, I decided to write an in-depth article breaking down the ReBeL algorithm (2020) in a way that’s easy to understand. ReBeL was developed by Noam Brow, the same researcher behind Libratus—the AI that, as you may know, defeated for the first time, a group of professional poker players in 2017. Here it is : https://medium.com/@sergi.nakache/rebel-...
I would like to make sure that the content of the article is accessible to any poker lover so I would love to have your feedbacks.
introdue humour to the AI. Make bluffs charming.... remember that they wont always work and that if you try to dictate number seeds into variants, the bluffs will become predictable. Now have a fun time teaching AI ROCKEGG
Nice article, Serge!
I particularly enjoyed your boxing analogy:
A consensus has been emerging in recent years among professionals: poker should be approached similarly to boxing. By default, one should maintain a high guard (unexploitable strategy) and launch precise attacks whenever a weakness in the opponent is detected (exploitative strategy).
Your breakdown of CFR felt a bit cumbersome, though, and could be simplified. I had the sense you were copy-pasting sections from academic papers rather than trying to translate it to layman. I say this as someone who's written about how CFR solvers work and understands the difficulty of such a task. Other technical sections faced similar challenges. But I still enjoyed it.
That said, your article is impressively detailed, and anyone developing a solver would be fortunate to find it. High-quality explanations on this topic are rare.
Much of the latest solver technology, including GTO Wizard AI, is built upon ideas Noam Brown pioneered with ReBeL.
Nice article, Serge!I particularly enjoyed your boxing analogy:Your breakdown of CFR felt a bit cumbersome, though, and could be simplified. I had the sense you were copy-pasting sections from academic papers rather than trying to translate it to layman. I say this as someone who's written about how CFR solvers work and understands the difficulty of such a task. Other technical
Hey, tombos21, thanks for your feedbacks, I appreciate. Indeed, the CFR part is not very fluid. I'm gonna re-write it. I'm a bit curious, you said " I say this as someone who's written about how CFR solvers work and understands the difficulty of such a task". What have you written?
Hey, tombos21, thanks for your feedbacks, I appreciate. Indeed, the CFR part is not very fluid. I'm gonna re-write it. I'm a bit curious, you said " I say this as someone who's written about how CFR solvers work and understands the difficulty of such a task". What have you written?
I wrote this article for GTO Wizard:
It's written in a way that's easier for poker players to digest.
Rereading my old work, I also forgot to mention in that article that it's the average strategy that converges, not the iterative strategy. I guess I should update that lol.
I wrote this article for GTO Wizard:
It's written in a way that's easier for poker players to digest.
Rereading my old work, I also forgot to mention in that article that it's the average strategy that converges, not the iterative strategy. I guess I should update that lol.
Wow! Very well formated! Do you mind if I take inspiration on it?
I tried to send you a private message in order to avoid flooding this thread but I have not found out how to!
Ahh thank you! Feel free to DM me or post here, either is fine
Wow! Very well formated! Do you mind if I take inspiration on it?
I tried to send you a private message in order to avoid flooding this thread but I have not found out how to!
@Serge, you'll need more posts/activity before you are able to utilize this site's internal messaging system (a week or two of activity is likely sufficient).
Good poker AI should use some randomization factor, in case someone decodes the pattern it uses.
introdue humour to the AI. Make bluffs charming.... remember that they wont always work and that if you try to dictate number seeds into variants, the bluffs will become predictable. Now have a fun time teaching AI ROCKEGG
Beyond the probable troll, I do agree about what should be an entertaining Poker IA. It should take slightly suboptimal decision in order to entertain the public and the players.
My dream is to play the NL poker variant I have designed (Triple Slot Poker) against a smart AI.
😀
Beyond the probable troll, I do agree about what should be an entertaining Poker IA. It should take slightly suboptimal decision in order to entertain the public and the players.
This is a really interesting point. I think all grinders forget, to our detriment, that this is foremost an entertainment industry. Losing conscious appreciation of that key fact we hurt our own profit margins dramatically.
Solver/AI based, cleverly targeted 'hijinks' would be a fantastic turnaround in poker theory. In fact, I've seen examples of it already with some of these new 'jam all-in' pre strategies that seem to be emerging from the data. People need to be having fun at the end of the day.
I think there's a significant difference between good "cash game AI" and "tournament AI". In MTT against bots, you are 1 against hundreds of players and try to win. In a HU game, you're 1 vs 1, so the AI should have a different profile (script).
I think there's a significant difference between good "cash game AI" and "tournament AI". In MTT against bots, you are 1 against hundreds of players and try to win. In a HU game, you're 1 vs 1, so the AI should have a different profile (script).
Noam Brown spoker about this topic here : https://www.youtube.com/watch?v=ceCg90Q9...
😀This is a really interesting point. I think all grinders forget, to our detriment, that this is foremost an entertainment industry. Losing conscious appreciation of that key fact we hurt our own profit margins dramatically. Solver/AI based, cleverly targeted 'hijinks' would be a fantastic turnaround in poker theory. In fact, I've seen examples of it already with some of these
What do you imagine exactly? Online poker looks dead to me, I can't see any solutions
It's a new NL game, but it's only 2 weeks old, so it hasn't been playtested yet. I'm currently searching for playtesters. I know it's great because I designed it, so my analysis of it is kind of deep and precise. Don't worry about all the skeptic posts, people just got used to a situation that someone comes up with a new poker variant idea.
You can see the rules here (posts: #1 and #26):
https://forumserver.twoplustwo.com/29/ne...
But I think it would be difficult to make good AI bots in this game, because of the cards interaction.
What do you imagine exactly? Online poker looks dead to me, I can't see any solutions
CO opens 5x and you're in the BB with QQ. According to some new AI model the most profitable play against humans is to cold jam, because the psychology of the fish is to think we're punting or always on AK and lol call much lower than they should.
Stuff like that. Ludicrous exploit lines that only emerge from large data sets and AI to piece it all together and calculate the risk. Moar action -> moar fun -> more action etc
But online poker can't die anyway. Poker strategy being ostensibly infinite. The bottle neck is intelligence + comprehension, not gto. I don’t know why people claim this. To claim this everyone would need to know all the strategy that hasn’t even been mined yet. All the exploits to every player type for every spot. It's not physically possible. I'd hear a case that this is the still the beginning of poker strategy actually.
Plus, in the future, the general public will gravitate to simpler games with rich lore and strategy content when the nuclear fallout curfews kick in and all the Fortnite server farms melt down.
Can't wait. $£€¥$£
It's a new NL game, but it's only 2 weeks old, so it hasn't been playtested yet. I'm currently searching for playtesters. I know it's great because I designed it, so my analysis of it is kind of deep and precise. Don't worry about all the skeptic posts, people just got used to a situation that someone comes up with a new poker variant idea. You can see the rules here (posts: #1
I have read it all. I pretty understood the format but it seems it's way to noisy to develop any stratégies. Community cards is evolving too much for me. It's like playing craps.
CO opens 5x and you're in the BB with QQ. According to some new AI model the most profitable play against humans is to cold jam, because the psychology of the fish is to think we're punting or always on AK and lol call much lower than they should. Stuff like that. Ludicrous exploit lines that only emerge from large data sets and AI to piece it all together and calculate th
In think I do agree that you can build more sophisticated poker bot but it's still the same problem at the end. Since it's pretty easy to play according to the Nash Equilibrium with bots, you can't make winners and losers at this game anymore. Which means nobody wants to play it.
I have read it all. I pretty understood the format but it seems it's way to noisy to develop any stratégies. Community cards is evolving too much for me. It's like playing craps.
Cool, thanks for reading. Reading opponent's range is just much more complex and difficult. Only 60% of the board is "noisy". You can also bluff with cards, not only with chips - this increases the complexity.
What is the best now available game to buy (or free) that has a very good poker AI you can play against ? (and which platform)
What is the best now available game to buy (or free) that has a very good poker AI you can play against ? (and which platform)
GTOwizard I would imagine. I think that's the current gold standard.
This chap is building something that people are playing around with that's more MDA based. But I think it's still mainly in dev mode.
How does Artificial Intelligence learn to bluff at poker?
As an answer to the main question, I would say that good poker AI doesn't necessarily need to learn in real time how to bluff. You (the programmer) can teach it how to bluff and do it in a tricky way, so that it's not easy to discover a solution.
If you program AI which is learning during the game, you can as well program it to be able to secretly see your hole cards occasionally. Would this be cheating ? Well, if it's beatable, it's fine with me - I hate AI anyway 😀 That's why I love to play against it.
