AlphaZero defeats best programs at Chess, Shogi and Go Chess forum

7 replies. Last post: 2017-12-08

Reply to this topic Return to forum

AlphaZero defeats best programs at Chess, Shogi and Go
  • Carroll at 2017-12-07

    https://arxiv.org/pdf/1712.01815.pdf

    It seems to have an elo rating 100 points superior to Stockfish 2016.

    It looks only at 80K moves a second with MCTS compared to 70 000K positions for Stockfish.

  • Carroll at 2017-12-07

    I’m not sure why AlphaZero achieves better results than AlphaGo Zero on Go.

    What is the improvement or the side effect if anyone knows that can explain why a general purpose neural net performs better than one which was designed specifically for Go ?

    The only difference I saw was that instead of producing 0-1 results for win/loss on Go, for chess where draws are possible it has to give a linear result from -1 to 1.

    Also from the games I saw it does not seek to exchange material much, could its performance compared with Stockfish be linked with a different way to handle contempt factor?

  • Tasmanian Devil at 2017-12-08

    And when will it take on all the other games offered on Little Golem??

  • David Milne at 2017-12-08

    It took huge resources for Big Blue to reduce the world chess champion to tears. Now software on a home pc can outclass any professional chess player.

    I think that within ten years anybody will easily buy stock software that allows them to develop an unbeatable opponent in any abstract game that they choose, ... including all the games on LG.

  • lazyplayer at 2017-12-08

    Caroll, but what hardware they used?

    There are some rumors these games weren’t played on “fair” hardware.

    P.S: Also keep in mind GPU are like 10x faster than CPUs if you can run your code there properly.


  • lazyplayer at 2017-12-08

    Nonetheless it’s an extremely interesting thing, because it shows a better evaluation function is so damn important even in chess.

    Basically it shows us that searching long and deep like stockfish does is not enough to discover all the “tricks” available in chess to create a favorable position...

  • Carroll at 2017-12-08

    I don’t know, the hard part is training where they use massive Google power.

    To play they just use a TPU which is an expensive specialized fast 64 cores GPU.

    I’ve read that Neural Net + MCTS sort of averages on a subtree the evaluation error while classical alpha-beta goes deep in the tree but gets the evaluation error to the top of the tree (like the horizon effect).

Return to forum

Reply to this topic




Include game board: [game;id:123456] or [game;id:123456;move:20] or [game;id:123456;move:20;title:some text]