TOP Comments
  • Thinker point
    This is illustrative as to why supercomputers have been in a kind of arms races for the past couple of decades. The realization has dawned that the first player with something like this that's capable of gaining a definitive edge in in world markets will probably control it all before the next player can even enter the game. It's difficult to see how such a force could ever topple, as its power and reach would go well beyond the traditional "hydraulic empires" of old.
  • Resident Futurist point
    I am unconvinced that it is better than Stockfish. They used Stockfish 8 rather than the latest development version which is 40+ Elo stronger. It looks like they are using the equivalent of a learning openings book in Alpha Zero while Stockfish has no built-in openings book learning or otherwise. And Stockfish is designed to work with endgame tables, that judging by the 10 games they released, they neglected to provide to Stockfish.

    They probably should have used Brainfish with the Cranium book it is designed to work with. It is not a learning book, but that is the best that is out there. The endgame tables that should be included are 3-man to 6-man and are available for download or they can generate them using a program that can be if they want the 6-man tables sooner. 7-man has been generated and can be accessed over the web but is useless to engines because the access time is far too slow. The main use of 7-man currently is in postal games (in tournaments where it is permitted) and game analysis.

    As you can see on this page, the latest Brainfish with the Cranium book is at 3486 Elo relative to Stockfish 8's 3396 Elo. www.sp-cc.de That shows the difference between having a good openings book and just being given an openings position to play from.

    Could AlphaZero become much stronger with more training? Sure. They just have not proven that. Also, both engines should be provided equivalent hardware. There has been some concern that Alpha Zero is using far more hardware resources than Stockfish in their tests. You certainly can't say your engine is better if the tests do not have the same hardware for each player. Any strong engine given substantially more time or processing is going to win against another.

    I wonder if the Shogi engine had a book. It would not be fair to it either, if it didn't.

    I think their claims are premature. Also, just releasing 10 cherry-picked games does not tell us much. If you say there were such and such games, you need to provide a pgn of those games so we can see that it is true and Stockfish is not playing crippled in some way. They talk about Stockfish 8 being the TCEC champion but it had those endgame tables in TCEC. So one can't really claim AlphaZero is playing against the same thing.

    This is not sour grapes. I do think they can likely achieve what they are claiming they achieved. I just think they are a bit short of that at least at the time of their publication. And the games they provided are interesting. AlphaZero does some great stuff with its rook humans and other programs are not doing...unless the teem just liked games where it used it rook more. We need to see more games. 10 games is just far too few.