An AI has just beaten “elite” poker pros at 6-handed No Limit Hold’em poker

An AI has just beaten “elite” poker pros at 6-handed No Limit Hold’em poker
My Comment: Although utterly unsuitable for a casual reader (including me, unfortunately), this magazine article explains how an Artificial Intelligence called Pluribus was able to consistently out-play a variety of top professional poker players in a six-handed format. The pros involved are names you will likely recognize. Enjoy–if you can. The machines are coming for us.

7 Likes

A very interesting read for sure. This has definitely been a long time coming, especially as this is not the first time that creating a 6-Max poker AI via Monte Carlo counterfactual regret minimization has been attempted. One example which comes to mind is PokerSnowie, which is a bot you can actually get preflop ranges from for free and play against with a subscription. However, that bot was quite limited as it only had a few bet sizes and played quite exploitative and suboptimal in some situations because it learned by exploiting the large online player pool of real people it was playing against. This meant that PokerSnowie won a lot against the online player pool overall but was beatable by some elite level players.

From what I can tell, Pluribus is quite a big step up from that but is still not perfect. The speed at which Pluribus plays is relatively impressive, and being able to run on 128GB of RAM is also pretty good in comparison to some of the supercomputers required to run some of the other poker bots. As well as this, Pluribus is capable of betting at ‘between 1 and 14’ considered sizes at each point. It also learned by playing a huge number of hands against itself rather than via playing human players. However, one important thing to note with this bot and also PokerSnowie is that Pluribus does not converge to a nash equilibrium like tools such as PioSolver are able to for heads up poker. I think it’s long been known that trying to find a nash equilibrium for 3+ player games isn’t really possible with current technology because of how big the game trees are. This means that despite what the articles are saying I don’t think we can conclusively say with this that Pluribus is miles ahead of the top pros. The sample size is quite small and the winrate (given in mbb/hand for some ungodly reason) of 4.8bb/100 is not actually that high. With this said it may still be possible for the absolute best human 6max cash game players such as OtB_Redbaron to still beat this bot. Not for much longer though; I’m sure some other bot is sure to come along before too long.

It’s also important to say that I think we probably have at least a few more years before AI is able to beat top humans at MTT and STT formats. This is due to the essentially infinite variety of ICM situations which can arise (meaning that it must be a lot harder to find optimal strategies for every situation).

In conclusion, I think it’s quite impressive that poker has gone this long without being beaten by bots and I actually don’t think that we’re quite there yet from what I’ve read here. It’s bound to happen eventually though- ultimately there is nothing that a human can do that a robot could not theoretically do better! Just gotta sit back and enjoy the ride I guess.

2 Likes

I guess it’s a big deal to some,as for me i’m here for the laughs lol

1 Like

Keep in mind that this bot is new and nobody has really had the time to study its game. Since it’s virtually impossible for it to play perfectly, there must be holes to exploit, especially in multiway pots. It will be interesting to see how fast it can adapt as it goes.

Back when I was learning NLHE, I played against a bot called Vexbot. This was an early adaptive AI from the Ai lab at the university of Alberta. It played to not lose, and gave me fits until I found its weaknesses. Once I found one, it would adapt and plug its leaks.

That was like 15 years ago, I shudder to think about how good the newer ones are!

1 Like

I think the key to beating this new AI would involve players adopting a swarm mentality and playing looser than they normally would. Since the AI will be strongest in heads up situations, it benefits everyone to force it into more multi-way pots, where it will be far less likely to find that optimal strategy.

Also, keeping pot sizes to the “paradoxical region” as described in Morton’s theorem should maximize the player’s EV, even when the AI does make optimal or nearly optimal decisions.

Basically, I’m suggesting a kind of implicit collusion not unlike that seen when a player is allin and there is no side pot. It’s in everyone’s best interest to eliminate that player, so people tend to check it down unless they have a very strong hand. It’s in everyone’s best interest to eliminate the AI, so it makes sense to work together to accomplish this.

That’s good thinking! Bravo!

1 Like