"The program considered 24 trillion simulated poker hands per second for two months".

But could it beat Negreanu if it had two weeks?

can you explain please sir

1. In the specific case of drawing to the nuts with one card to come is it possible that the computer would fold getting a smidgeon above correct odds (as the casino bot sometimes does)? If you cant say no with absolute certainty, why not add the instruction not to?

2 Do you actually know the precise strategy that beats the computer or do you just know that it exists? If you know it, what can you say about strategies that are in between your strategy and that strategy?

please tell me that you are not doing NLHE next.

See post 46

not enough computing power for that yet.

They do have a HU NLHE project. I would describe their current AI as quite competent but non-expert. The most interesting thing about it is that it's play is pretty different from human norms as it's generally quite passive. This isn't to say that it's right to do so, but it's neat that it's going in a different direction than humans so far, unlike the LHE AIs which for the most part do what humans do better.

whoops, my bad. ty sir

The other major piece of context is that the game is such that it's practically guaranteed that the opponent will bet the flop after a pre-flop 3-bet and raising most flops does little to define one's hand range, giving one the opportunity to keep the pot small with weak hands like the high card portion of one's range at little cost to the strong portion of one's range. Communicating adequately with non-players is ultimately almost impossible, which isn't to say that misrepresentations don't bother me, too.

Can we play vs the bot on the website? I assume yes, however the play button on the site brings me to a white screen. Probably just overload right now, but would really love to try it out.

Website's a bit fried right now, but Cepheus should be available to play once it cools down a bit.

why the name Cepheus? to keep up with the space theme?

I guess we can't say it doesn't ever do this with absolute certainty, and there's always a bit of noise because it's essentially solved, not perfectly solved. It's certainly possible to check scenarios, one at a time: given a specific hand, board cards and betting, I can ask what Cepheus does (as could you or anyone else - it will always be a bit slow, on purpose - but the strategy query tool is up and running again after the big media hit.)

Why not use a hard coded rule? We've gotten (lightly) burned trying to do this in the past. At one point, back in the days of PSOpti, we added some hard-coded rules to force correct play in a few circumstances where the bot would otherwise do strange things. This code sat there, turned on, effecting everything we did for the next 8 or so years. Eventually, we realised it was still on, tested things with it off, and discovered the hard-coded rules weren't quite correct: once we finally got good enough, they made things just a little bit worse.

So it's safer just to let the algorithm do its own thing. We can verify after the fact that anything strange it's doing is either i) correct, despite expectations, or ii) not worth worrying about.

Now that we've computed the strategy, could we possibly do something like this to clean up a few messy small numbers? Maybe. But why? Good poker players would already know because it was using existing human knowledge, and science wouldn't care...


We know one strategy that "beats" it, in the sense of getting as much money as possible: itself. That's one property of a GTO solution. For two player games, each player's strategy has to be a best response to the opponent's strategy. It's just that they're special best responses. Ignoring the "essentially" part of "essentially solved", there is no way to play against Cepheus that gets more money than Cepheus itself gets. *

If you were emphasising "precise" strategy in the sense of the exact strategy that wins 1/1000 big blinds a hand against our strategy, we don't have that. We computed it, one bit at a time, in order to find out how close to GTO we are, but we throw it away as we go to save space (and time.)


* Breaking even against a GTO doesn't mean you have perfect play: think "always rock" against the 1/3,1/3,1/3 equilibrium in rock-paper-scissors. Similarly, bad play doesn't necessarily mean you will lose against GTO. You'll only lose if you make a mistake -- something NO GTO solution would ever do. That's easy in rock-paper-scissors, but poker is a bit harder and it's hard to make absolutely no mistakes.

Yup. The previous bot in the Man vs Machine matches was called Polaris. When Mike Bowling read that Cepheus was going to slowly replace Polaris as the North star, we had to stick with the space theme and use Cepheus.

Do you have any estimate what WR Cepheus would expect against a good or great human HUHU player?

That would be crazy and is really possible the way tech is advancing. Might really be rip to online poker

Why hack? Wasn't it Open Source?

The completed strategy for HU LHE doesn't really mean anything for other formats (other than as an indicator of how this sort of tech is progressing). Cepheus is literally a massive spreadsheet of probabilities for HU LHE and nothing else.

I'm not sure you realize that the only situation that concerned me was folding a draw to the nuts getting correct odds with one to come. How hard could it be to add the specific instruction to make sure its not doing that? (Even if the only reason was to prevent detractors from jumping on this simple clearcut error that even layman would understand if it were to ever occur.)

ok you can beat me over a large sample but what if i hit and run you for and you never see me again

My understanding is that adding extraneous code always carries some risk of complication. While there may be some hypothetical possibility that Cepheus could fold the nuts (due to the massive strategy table that would need to play infinity hands to converge on perfect) if that probability is assumed to be epsilon then it's not worth tinkering with the whole AI to try and fix it.

For sure - if there is now a bot that can beat HU limit holdem, it's a complete certainty that someone will be able to steal it, make it undetectable by poker sites, and then destroy all of online poker with its unbeatable play of the most popular version of Texas Holdem.

This is witchcraft and the creaters should be burned at the stake

For this alone, I'm a fan.

No money in Holdem imo, time to play Omaha. Or tournies...