Cepheus - an unbeatable poker program finally a reality?

That's a bold claim to make:scientists at the University of Alberta believe they have finally achieved a goal of creating a poker playing program that is unbeatable.

The program is named Cepheus and the bot plays Limit Hold'em. By using a statistical analysis and pre-determined strategies for every possible move any player can make, Cepheus could play a heads-up match against any other player and, at worst, come out as a break-even, which means his play has to be perfect.
“We’re not saying that it’s guaranteed to win money on every single hand,” says Michael Bowling, a computer scientist at the University of Alberta and a co-author of the study. “What we’re saying is that, in the long run, if you looked at all the hands that could happen and you averaged all of those, then the computer can’t be losing, at a losing rate — it has to be either breaking even or winning.”

“We had this training phase where the program started off playing uniform random against itself, meaning that (Cepheus) had no idea what it was doing other than following the rules of the game,” explains Michael Johanson, another computer scientist also at the University of Alberta and Bowling’s co-author of the study. The program supposedly started to update the strategy as the database became bigger and bigger based on the fact if the decision was +EV or -EV.

If you are worried about Cepheus making it into live play, don't bother - scientists reported that it took nearly 70 days for the program to gather enough information to reach the perfect strategy, and that happened with use of 200 computers with 32GB RAM and 24 CPUs each.

Also, they stated that there is still some of improvement, but it's so marginal it's not really worth it: “Even if the program spent a lifetime in training, it is so close to perfect now that it wouldn’t have much value other than academic novelty,” Bowling stated.

On another note, Cepheus was able to demonstrate academically and mathematically something that has always been a basis for profitable poker play. During the study, Cepheus demonstrated that the dealer has the advantage over its opponent, but the advantage is miniscule. “We can now prove that the dealer has an advantage of what we call ’88 millablinds’ per game,” Johanson said. “That’s .088 of a big blind per game,” Johanson reported.

You can see the full strategy and even play against Cepheus under this link: Cepheus Poker Project.

i may be paranoid but i'd prefer not to play against him not to give them data ;P

On January 14 2015 19:14 MadeInPolanD wrote: i may be paranoid but i'd prefer not to play against him not to give them data ;P

If you are worried about Cepheus making it into live play, don't bother - scientists reported that it took nearly 70 days for the program to gather enough information to reach the perfect strategy, and that happened with use of 200 computers with 32GB RAM and 24 CPUs each.

yeah i don't think they would even bother collecting your data :D

but now that the 70 days on crazy computers has already been spent solving it, it now CAN affect live play. If you have the 12 terabyte solution you can play optimally in a reasonable amount of time.

for example you can go to the alberta website right now AFAIK and look up the answer to any spot in HU Limit Holdem.

is it a foregone conclusion that rake structure doesn't affect optimal play... ?

I don't believe them, I think they are wrong. What they are saying does not make sense.

Poker is a game of adjustment, there is no perfect. Range and strategy can always be adjusted to beat another strategy and range set. So how can there be a perfect?


And if they made it so that their bot analyzes hand statistics and adjusts, then that fact itself can be abused to outplay it once you know how it works.

Perfect is not possible in these types of games.

lol

yea well its LIMIT holdem

im curious about the bots that play NL

On January 15 2015 00:21 asdf2000 wrote: Poker is a game of adjustment, there is no perfect. Range and strategy can always be adjusted to beat another strategy and range set. So how can there be a perfect?

Generally its the poker version of randomly playing 1/3 rock, 1/3 paper, 1/3 scissors, called the Nash equilibrium. This doesn't necessarily mean being unbeatable by itself since certain positions are better than others, but when each player gets an equal opportunity with those positions it does.

And if they made it so that their bot analyzes hand statistics and adjusts, then that fact itself can be abused to outplay it once you know how it works.

This however, is correct. It tilts me a bit when use language like "unbeatable", "solved", and "perfect" when the very method they're using to find their strategy is to find the best strategy that beats it. The strategy that has the highest EV that they have found (as in least negative EV) against the best possible counterstrategy is what they consider the "perfect" strategy.

Basically its a huge a complicated math problem, but the "solution" is being found in this case by trial and error instead of actually solving the problem. But you can't really blame them because actually solving the problem is something they know how to do but requires a lot more processing power than trial and error.

i think this is pretty cool they made it, i think its terrible they posted it all over the internet.

just what we need is a bunch of fish programmers sitting around all day perfecting bots..

in my mind, this promotes that bots can just play 24/7 all day while u go la de da around and come back to see ur profits with lhe tables running, im sure some of the bots out there are very very advanced already..

oh well, we may all have to turn into robots to counteract. i call laser eyes.

Is anybody else surprised that the advantage of having position is worth less than 1/10th of a big blind? I think that's by far the most interesting part of the article.

^ it's hard to believe that...over the millions of hands i've played my button stats isn't a 1/10 of a bb edge....

On January 15 2015 11:57 HungarianGOD wrote: Is anybody else surprised that the advantage of having position is worth less than 1/10th of a big blind? I think that's by far the most interesting part of the article.

per game means per hand not per 100.

Its still a little surprising, but I think the main reason we consider position so powerful is since normal players (even good ones) fall into certain patterns that give away information. The bot in question won't be falling into such patterns.

Also there's not much folding in limit preflop, so its not like you have a huge advantage from posting less.

the only unbeatable program vs humans is pokerstars.exe

Not surprising that position its not very relevant to a GTO bot, it is important for us flawed humans playing against exploitivly against other flawed humans

position is only relevant for determinating ranges, everybody feels negative emotions when playing oop coz his range is weaker, also its hard for a human to incorporate balanced donkbetting strategy

On January 15 2015 10:19 NMcNasty wrote: This however, is correct. It tilts me a bit when use language like "unbeatable", "solved", and "perfect" when the very method they're using to find their strategy is to find the best strategy that beats it. The strategy that has the highest EV that they have found (as in least negative EV) against the best possible counterstrategy is what they consider the "perfect" strategy. Basically its a huge a complicated math problem, but the "solution" is being found in this case by trial and error instead of actually solving the problem. But you can't really blame them because actually solving the problem is something they know how to do but requires a lot more processing power than trial and error.

Even in case there were formulas for finding equillibrium, i guess we could only bruteforce our way to near-perfect strategies by trial and error method.

Oh sorry, there is a way to find out mathematically correct formula for LHE equillibrium, but they are not solvable atm because of huge numbers

On January 15 2015 14:23 YoMeR wrote: ^ it's hard to believe that...over the millions of hands i've played my button stats isn't a 1/10 of a bb edge....

well, it's for Limit. (i'm assuming your millions of hands are from NL)

Is it me or are they " undergoing temporary maintenance. Please check back soon! " permanently

On January 15 2015 20:17 Baalim wrote: Not surprising that position its not very relevant to a GTO bot, it is important for us flawed humans playing against exploitivly against other flawed humans

position more important in NL imo? since limit is comparable to super shallow stack NL, and being deeper favors the IP