x
Chess - Play & Learn

Chess.com

FREE - In Google Play

FREE - in Win Phone Store

VIEW

AlphaGoZero masters Go entirely on its own?

  • #1

    This sounds amazing. Is anyone trying to do the same for chess?

    "Starting with zero knowledge of Go strategy and no training by humans, the new iteration of the program, called AlphaGoZero, needed just three days to invent advanced strategies undiscovered by human players in the multi-millennia history of the game."

    "After three days of training and 4.9 million training games, the researchers matched AlphaGoZero against the earlier champion-beating version of the program. AlphaGoZero won 100 games to zero."

    https://www.quantamagazine.org/artificial-intelligence-learns-to-learn-entirely-on-its-own-20171018/

  • #2

    This indeed sounds amazing.I wonder why they haven't done it with chess yet.

  • #3
  • #4

    Machine Learning approaches are pretty common now in AI. It's been done with chess, but as far as I know still loses to brute force approaches. 

    Machine Learning is better at some things though, like playing old Atari games

  • #5
    hairhorn wrote:

    Machine Learning approaches are pretty common now in AI. It's been done with chess, but as far as I know still loses to brute force approaches. 

    Machine Learning is better at some things though, like playing old Atari games.  

    It would probably goes into chess if they support huge chess programming fund similar to alpha go fund. Current top 3 chess engine programmers has no support. Stockfish programmers are volunteers. Komodo is run by two programmers and Houdnini by single programmer and they get support/rewards from selling those commercial engines to small chess addict community .

  • #6
    hairhorn wrote:

    Machine Learning approaches are pretty common now in AI. It's been done with chess, but as far as I know still loses to brute force approaches. 

    Machine Learning is better at some things though, like playing old Atari games.  

    Nobody has tried to train a machine learning AI enough to beat a brute force chess engines. Giraffe was abandoned to create AlphaGo. If they did, the machine learning engine would eventually beat the brute force calculator.

  • #7
    But since brute force has been wildly successful in chess (and much less so in go), there's no real push to do it, outside of the odd student project.
  • #8

    The interesting thing is that the machine plays with zero knowledge of Go strategy and no training by humans. So it has no opening book, essentially, among other things. And yet it is so good. How good are the top "brute force" chess engines if you disable their opening book completely?

  • #9
    keju wrote:

    Is anyone trying to do the same for chess? 

    Some chess engines are improved (continuously) by a process which can be decribed as "self learning"

    Stockfish has played more than 500,000,000 games (engine vs engine) and the results are used for incremental tweaking to make the engine better. (This is about 100x the games that were used to make the Go engine). A webpage to see the running scorecard is here:
    http://tests.stockfishchess.org/users

    I think as a whole, chess engines have been developed closer to "perfection" than Go engines.

    Unfortunatelly there's no way to prove it because chess engines can't play Go, and vice-versa.meh.png

  • #10
    vickalan wrote:
    keju wrote:

    Is anyone trying to do the same for chess? 

    Some chess engines are improved (continuously) by a process which can be decribed as "self learning"

    Stockfish has played more than 500,000,000 games (engine vs engine) and the results are used for incremental tweaking to make the engine better. (This is about 100x the games that were used to make the Go engine). A webpage to see the running scorecard is here:
    http://tests.stockfishchess.org/users

    I think as a whole, chess engines have been developed closer to "perfection" than Go engines.

    Unfortunatelly there's no way to prove it because chess engines can't play Go, and vice-versa.

    What you describe is very different.Tweaking the program or the parameters is actually improvement of the programmers not of the engine itself.We are talking for an engine that can learn from it's mistakes and self improve it's parameters.

       The engines in chess are indeed close to perfection but the dream of an engine that can "think" still stands.

    Maybe the more interesting would be creating engines with personality(if that is even possible).Would an engine be able to immitate   Tal's styles if it learned chess only by Tal's games?Imagine a world championship with engines that immitate players of all eras(Morphy , Anderssen, Steinitz , Alekhine . Tal ,Fischer, Kasparov , etc.That would be really interesting. I know it has been tried by some programs(chessmaster) but failed miserably(they all play very similar , you never feel that you play a different "person").

  • #11

    I agree that the programming style (between a computer learning to play Go, and a chess engine being tweaked to be better in chess) are probably very different.

    There is probably no more opportunity to let a computer "learn" to play chess on its own, because they already know how. (Although it can be done for experimental purposes).

    I believe both these tasks require lots of computer resources (consider the number of game played).

    I don't think there is any purpose for a chess engine to learn from any grandmasters anymore. Chess engines are now completely unbeatable and they have their own style: they win games.

    Learning the "style" of human players could be interesting, but it's academic.

    I think programmers are currently interested in a higher objective - in engine vs. engine, what is the best engine?

    Human play is obsolete when the goal of "perfection" is the objective.happy.png

Top

Online Now