News: Google AI called Deepmind with AlphaZero self-learning algorithm is the most powerful and highly rated entity to have ever played the game of chess, AINEC.

View: Google Deepmind, after only 12 hours of learning Holdem poker, would beat Libratus (the strongest poker engine) for >10bb/100.

Gossip: So why has this not already happened? Does Google simply not care enough about poker? Is there a potential weakness for Deepmind when the game features incomplete information (highly doubtful IMO)?

Or, does Google not desire to put an entire generation of Computer Scientists out of a job, given that poker is the most popular way for those scientists to study AI and earn a paycheck?

google has already implemented deepmind poker bots and is crushing all internet poker sites as we speak

I still trust that Google is socially responsible.

Agree can’t see why a game of incomplete info would be a problem for it

I honestly think that if deepmind could be configured to learn poker it'd be a pretty long stretch to imagine it beating Libratus for over 10bb/100

We need Deep Mind to stream on twitch so us mere mortals can learn the ways of the robot.

Hours is not a good metric for training time.

Neural networks currently struggle in games with incomplete info.

It would need to give us information that is more than we already get from books, GTO softwares and what we can buy. Did we get to see ten chess games or something like that? So, some hand histories perhaps, some percentage perhaps?

I believe that it is much easier to determine the max that your computer can be beaten for than to come close to GTO. Alberta did that although I am sure its a lot simpler to do that for limit vs no limit. Anyway would not it be quite possible that Libratus could not be beaten for 10bb with the perfect counterstrategy let alone with perfect GTO which would do worse.

Yeah in limit there are no abstractions so you can just calculate the ev of the best possible counterstrategy. It isn't possible measure that for something like Libratus because you have the option to not play in the abstraction. I'd guess it's exploitable to some extent in some bizarre lines. One of the key strengths of Libratus though was that it could train overnight to find weaknesses in the abstraction that the players were exploiting.

Everything Libratus can do Deepmind would do better, easy victory.

Generally the top AI's "retire" once they achieve their task. More accurately they are then moved into real practice. For instance, IBM Watson hasn't played Jeopardy since it's big win but has been doing real work in medical records for instance. My understanding is that AlphaZero and AlphaGo are now retired from their competitive careers and are back to being secret, proprietary working programs.

There's nothing really in it for Google to play poker. It's a big deal to be the first program to beat a human in a given activity (see Watson, AlphaGo, Deep Blue, Libratus), not so big a deal to be second or tenth. It's surprising actually then how much press AlphaZero got, although AlphaGo got way more.

Interesting. Thank you. I am sure you already know everything I am saying below, this is just my discussion of the matter:


My interest in the AlphaZero algorithm is that it specifically beat its own predecessor, AlphaGo *because* it was not specifically trained to play the game of Go. It is only given the rules of the game, and learns the game from scratch.

This leads Deepmind AlphaZero to play games such as chess and go in a way that is more familiar to humans, but at a super-human level, that is best described as a highly evolved alien being might play.

So, DMAZ is not simply crunching numbers to evaluate a position. It has a PLAN for every position.

How well would this translate to poker?

That is debatable, since poker *strategy* may in fact be more about raw evaluation of odds, and merely limited by the ability of the human or computer to efficiently evaluate the portion of gametree for expectation.

So, for me to say that the DMAZ will beat Libratus for a meaningful margin, is to say that DMAZ will actually be playing poker, with a plan for every combo, right from the start. This will find weaknesses in the ability of Libratus to survey the tree, with powerful CFR. Further, the weaknesses will be numerous, yet tiny, and beyond the precision of Libratus to detect and correct.

I do not understand what your wrote.

I would be amazed if Deepmind haven't tried poker. The founder of Deepmind, Demis Hassabis, is not a stranger to the WSOP.

http://pokerdb.thehendonmob.com/player.php?a=r&n=42073

The relative easiness of computing counter strategies comes from the fact that you need only consider pure strategies and that the EV of each individual hand can be maximized in isolation vs. villain's strategy. Even for no-limit not much abstraction would be needed. I think something like requiring bets/raises be a multiple of 0.5bb would make the computation feasible.

That statement is highly questionable. AlphaZero played a match against stockfish, which is the strongest chess program existing, but stockfish was severly handicapped during the match. they never played a match with stockfish at full strength, so we dont really have evidence as to what the result would be.

Who wins OTB or Deep Mind?

Amusingly, the latest competitive AI from DeepMind is called AlphaFold, but it has nothing to do with poker. It's 'playing' a much more complicated game of predicting the 3D structures of proteins.
"It’s never been about cracking Go or Atari, it’s about developing algorithms for problems exactly like protein folding," Hassabis said. The number of possible protein structures is around a googol cubed, or 1 followed by 300 zeroes. It's a much more complex 'game' than poker.
Source: https://www.theguardian.com/science/...es-of-proteins

Deep Reinforcement Learning from Self-Play in Imperfect-Information Games

David Silver
Juk

I suppose I am jealous of the chess community, since Chess24 has post game analysis of the world championship match with input from DeepMind (free to view, highly recommended).

I wish us poker players could get to hear how DMAZ would play a certain hand, or a certain range of hands.

My statement has just recently been confirmed. Deepmind has released the results of a few thousand more matches vs Stockfish in different configurations.

There is no doubt, Deepmind AlphaZero is the strongest chess playing entity ever.

So, with regards to poker, again I am puzzled that Deepmind is not entered into the competition for poker AI’s.

Clearly self-learning is the future of AI. However, with poker, increasingly efficient CFR has been the key.

Can DMAZ teach itself enough CFR to win?

I think so.

It seems so but it is a private lab test and some people are still skeptical.

http://science.sciencemag.org/content/362/6419/1118

no real incentive for them to play poker, but it would be interesting to get a take on how good A0's appraoch would work for probabilistic imperfect information games.

Anyone remember when Facebook made a couple AI accounts/profiles and they created/were communicating via their own language and it could not be interpreted and they shut the entire thing down? That was creepy.