Alpha Zero in 4pc

There's been a lot of buzz over Google's Alpha Zero soundly defeating the Stockfish 8.0 chess engine. 

A brief recap: Alpha Zero, using a novel form of machine learning, "studied" the game of chess for 4 hours (it was told nothing but the basic rules, and spent 4 hours playing games against itself).  After those 4 hours, it played 100 games against the Stockfish 8.0 engine (the winner of the 2016 chess-engine championship).  And Alpha Zero CRUSHED Stockfish.  It won 28 of the games (25 as white, 3 as black) and drew the rest.   While there's a lot of debate raging over some of the nuances of this match-up (Stockfish didn't have its opening book, it wasn't the latest version, it wasn't run on optimal hardware, etc.), the key takeaway is that in a very short period of time, Alpha Zero was able to teach itself chess at a level at least on par with the best chess engines ever built.  

The waking thought I had this morning, and that's been bouncing around in my head since: what if Alpha Zero were trained to play 4PC? The basic development principle should be the same: just give it the rules of 4pc and 4 hours to play itself (technically, play 3 different versions of itself). 

How would it stack up against human players?  In a game between two solid "Leaderboard" human players and two Alpha-Zero opponents (with game anonymized to avoid any transparent "Fight the machines!" ruses), would the humans have any chance at finishing first? What about with 3 humans vs 1 machine?

What would its game play look like?  Obviously the engine would employ superhuman tactical capabilities (if final two players were human vs Alpha Zero, with roughly equivalent material and points... yeah, game over).  But what about meta-strategy?  Would the highly skilled Alpha Zero engine use game-theory to its advantage?  Would it learn to cooperate with opposite player (e.g. to coordinate attacks against adjacent players, or perhaps to rescue the opposite player from danger for no immediate gain)?  

It's utterly pointless speculation, of course.  I expect Google tackled Chess *not* to create a masterful chess engine but simply as a marketing stunt that made a point about its AI's current capabilities.  Even that point merited just 4 hours or so of computational time.  I would not be surprised if - to the utter frustration of the chess world - Google never again (officially) entered the foray of Chess.  And I can think of no motivation for Google to dedicate any additional resources (at least, not any time soon) to teach Alpha Zero to play more obscure games (especial games without clearly-established standardize rules like 4PC).  Nevertheless, I thought it was a fun thing to contemplate, and figured it might be interesting to hear the thoughts of others on this forum.

Better yet, what if it were trained to eradicate humans.  Time to stop being so impressed.

Yes, obviously computers will exterminate us in another 10-20 years.  Kind of a ho-hum topic.  But how would they do with chess.com's 4PC today?  That's not so obvious.  

4 Player Chess is now not even a baby, it is an embryo. Any chess engine can impress us at this stage. Even top players, including the current Top 5 guys, still blunder their queens in Teams Mode. If the engine won't blunder like that he'll be Top 1.

well we would have to give the machines information on what humans are so wikipedia may be the reason the the human race gets exterminated

The moment a publicly available engine comes out that can play 4 player the game is ruined.

Nahh, we won't be exterminated any time soon. Computers won't kill us, for the same reason that other vicious predators, like cats, are not killing us. There is a secret code shared by all cats, that goes like so: keel a humin and you has food for one day, let him stay alive and you has food forever. 

@ giusbold In 2016 Alpha go ( the ancestor of alpha zero) beat the best human at go....  people thought engines could never be thaught to play go...  Recently, Alpha zero go beat alpha go 100 -0....   I repeat: 100 to 0!

alpha zero would learn FFA too.

But Trinhlan could still beat it! 

@kevinkirkpat   Clearly it would.   It would play millions of games against itself and gain a superhuman insight about how to play 4PC and utterly destroy the best 4PC players,   wether in team mode or in FFA mode.

There already are computers playing poker and it's a big problem for sites like pokerstars. Bots as they call them, certainly can apply game theory principles.

but through cheating we could still easily beat it.......... ......

totally agreed, as you can see by my profile picture i really like cat memes, someone should make that a meme

Computers don’t have their own thoughts like humans do, the just do what their code tells them to do.

They will exterminate us thoughtlessly.

Then exterminate them before they exterminate us

Wait. Isn’t this supposed to be about AlphaZero in 4pc?

It will win in Teams, but lose in FFA. Because in FFA humans play 3 vs 1 against the strongest player.

It would "learn" optimal attack timings that we don't perceive, it would "learn" odd defenses based on attacks that the other players were setting up on one another, it would probably trade unevenly and unpredictably. I'd be very interested in seeing the types of opening principles it valued, whether it preferred points vs position, how it valued promotions, where it placed the king, etc. Odds are we are playing 4pc drastically wrong

@pjfoster1313 nahh, FFA is not a deterministic game to have objectively good or bad ways of playing. Any strategy can be beaten, if it makes your opponents annoyed enough to gang up on you. And you can't deduct what annoys your opponents solely by examining the chessboard, you must also examine (read) their minds. Alpha Zero would have a serious disadvantage against human opponents, because it couldn't understand what is written in the chat (if was not allowed to /stop-chat at first move of course).