So glad that these guys could potentially destroy thousands of online poker careers for the sake of curiosity and academia. Such heroes.

This bot is nothing special, just combination of the solutions from "Simple Preflop Holdem" and "Simple Postflop" + good conditions for bot to play.

as far as i'm aware off, this is not true since the bot started from absolute scratch and just got better by playing against versions of itself. don't think there any preprogrammed ranges involved whatsoever tbh

No, it has baseline strategy , it uses self play but its based off of strategy

hmm, must have missed that part then. could you perhaps provide a source for that statement?

I thought the idea of the CFR algorithm was this (source: https://www.quora.com/What-is-an-int...minimization):

It doesn’t matter if a human isn’t calling with many worse hands. A human likely should be calling with many worse hands. If the bot doesn’t get value with its QJ very often there it means the human is over-folding and is going to get owned by the bot’s bluffs (assuming the bot is in fact balancing correctly)

Let’s say the human players only entered multi-way pots with crazy strong ranges and were giving up EV by not defending their Bb enough multi-way. Then the bot wouldn’t do well in multi-way pots but would still be using a good multi-way strategy. I’m sure there are better and more complex examples, but that’s one example.

In Simple Preflop Holdem preprogrammed ranges are also not involved. You just specify tree, stacks and start solving equilibrium from the scratch. The process of the solving could be called playing against itself - it is true. After each iteration the solution gets better, more accurate and less exploitable.

The name of the algorithm in the paper is the same as in the solver.

So, my statement is true.

Obviously the objective is to be applied to other fields, but one always should be cautious about AI applied to the real world. It's chaos out there, many times without clear predictable data. Remember LTCM and other guys, thinking that they can predict behaviour of market participants.

Curious also the first associations of this AI research is with the american military and with facebook, always good to see science put to good use.

In the solver you could choose variations of Monte Carlo counterfactual regret minimization algorithm:

cool, thanks for the information. however saying 'you are not impressed' because they use a well-known algorithm is quite denigratory. no need to reinvent the wheel either

Lots of players check better than QJ on the turn in those spots these days.

That AJ bluff probably worked some years ago but today wouldn't against most players.

"Depth-limited search in imperfect-information games
The blueprint strategy for the entire game is necessarily coarse-grained owing to the size and complexity of no-limit Texas hold’em. Pluribus only plays according to this blueprint strategy in the first betting round (of four), where the number of decision points is small enough that the blueprint strategy can afford to not use information abstraction and have a lot of actions in the action abstraction. After the first round (and even in the first round if an opponent chooses a bet size that is sufficiently different from the sizes in the blueprint action abstraction) Pluribus instead conducts real-time search to determine a better, finer-grained strategy for the current situation it is in. For opponent bets on the first round that are slightly off the tree, Pluribus rounds the bet to a nearby on-tree size (using the pseudoharmonic mapping (39)) and proceeds to play according to the blueprint as if the opponent had used the latter bet size."

and it assigns playing styles.

"Pluribus instead uses a modified form of an approach that we recently designed—previously only for two-player zero-sum games (41)—in which the searcher explicitly considers that any or all players may shift to different strategies beyond the leaf nodes of a subgame. Specifically, rather than assuming all players play according to a single fixed strategy beyond the leaf nodes (which results in the leaf nodes having a single fixed value) we instead assume that each player may choose between k different strategies, specialized to each player, to play for the remainder of the game when a leaf node is reached. In the experiments in this paper, k=4. One of the four continuation strategies we use in the experiments is the precomputed blueprint strategy, another is a modified form of the blueprint strategy in which the strategy is biased toward folding, another is the blueprint strategy biased toward calling, and the final option is the blueprint strategy biased toward raising. This technique results in the searcher finding a strategy that is more balanced because choosing an unbalanced strategy (e.g., always playing Rock in Rock-Paper-Scissors) would be punished by an opponent shifting to one of the other continuation strategies (e.g., always playing Paper)."

so this isnt nash equilibrium ( this would change constantly ) . i think.

inb4 google makes some kind of alphazero bot and crushes everyone with it

It's 20 seconds per hand, not per action

It is trying to approximate nash, it doesn't assign playing styles

There will always be live poker.

I like you as a politics personality/poster but DONNY YOU ARE OUT OF YOUR ELEMENT in this particular spot

deppressing as **** that you're inclined to point that out

It's fairly standard to overbet turn with QJ+ there for most regs in that spot, so i like the x/r. I don't think most regs are certain enough about other people's ranges to make the x/r this big though, but perhaps that kind of x/r size will become more and more common, i like it. Piosolver is going with mixed strategy there on that river, i think it mostly just leads as a huge overbet on river when it gets to this spot, but it does the big x/r as well.

So is there any access to more hands than just the 3? kinda unfair that these guys get to learn multiway strat from the bot while the rest of us don't.

Washington Post has an article about this:
https://www.washingtonpost.com/scien...=.a62550f00c44

An aspect of all this that I don't think has been mentioned yet ITT:

"One intriguing, or perhaps disturbing, element to the story is that the bot achieves these results without paying attention to the personalities, habits and strategies of its opponents. The bot does not concern itself with human psychology. It doesn’t know whom it is playing or try to calculate what the mental state of the opponent might be.

That’s in contrast to what’s happening this week in Las Vegas at the World Series of Poker. A television viewer will notice that the players spend a lot of time scrutinizing one another, trying to figure out who is bluffing and who is not — looking for the “tell.”

What Pluribus suggests is that humans may overrate the psychological part of the game. Getting the math and probabilities right seems to be all that’s necessary to be a champion.

It doesn’t matter who’s twitching and scratching and blinking at the table."

bolded is surprising to no one other than maybe Double A and Charlie Carrel

I would also like to know more about bot-detection, but surely poker sites are heavily incentivised to build AI detecting AI or they will inevitably lose their action too?

A large percentage of my strategy has been informed by machine-play (off-table). From my perspective 'they' have been and always will be better than me so the only question to worry about is whether online sites can maintain the integrity of the games, and I believe they can - or at least will likely do so - because they ultimately have more to lose than the players.

Stars have made/will make an effort but I sincerely doubt anyone else will come up with something remotely effective to detect bots

thats like saying: both the pluribus and GTO solvers both run on PCs, therefor they are essentially the same and pluribus is nothing special, when in fact they use vastly different methods. while solvers try to find nash equlibria in 2 player situations, to my understanding it isn't even clear that there is such a thing as a nash equilibrium in multiway pots, unless obviously all players "agree" to find such a thing and cooperate in doing so. so no, pluribus is not just a "better" version of a solver.

"The core of Pluribus’s strategy was computed via self play, in which the AI plays against copies of itself, without any data of human or prior AI play used as input. The AI starts from scratch by playing randomly, and gradually improves as it determines which actions, and which probability distribution over those actions, lead to better outcomes against earlier versions of its strategy. (...)
Pluribus’s self play produces a strategy for the entire game offline, which we refer to as the blueprint strategy. Then during actual play against opponents, Pluribus improves upon the blueprint strategy by searching for a better strategy in real time for the situations it finds itself in during the game. In subsections below, we discuss both of those phases in detail, but first we discuss abstraction, forms of which are used in both phases to make them scalable."
https://science.sciencemag.org/conte...cience.aay2400

to me some of the plays it made indicate the accuracy of the solutions it came up with was prob fairly low so it was exploitable, just less than any human