Random Posts

Tuesday, June 21, 2016

Is the Giraffe Dead?

     Back in September I reported on Giraffe, a deep-learning machine that learned to play chess like humans do. Also, see the article in MIT's Technology Review. 
     The bad news is the project has been discontinued by the author; he explains why on Talk Chess. You can also read an additional article HERE.  The good news is, it is now open source and it will interesting to see if it will ever reach the point where it can challenge the commercial engines like Stockfish and Komodo. 
     Giraffe's CCLR rating is only 2414 placing it way down at 167th place. In obtaining that rating it has not played any of the top engines. One disadvantage Giraffe has it that only uses one core.
     Yesterday I played a match against a version of Giraffe that I download of the CCLR website. It was played at 4-minutes per game using my Expanded Fritz 12 Opening Book and the results were Stockfish scored a decisive +4 -0 =1. 
     Playing over the games I noticed that it played very well then it would make a positional error. On a couple of occasions it seemed like the engine "didn't know what to do" and so made a meaningless King move. And, unlike the leading engines, it committed an occasional tactical error. Also, on more than one occasion, it's evaluation was nowhere near Stockfish's. Maybe this is the result of having learned by "studying" human games!? 
     Thinking that perhaps 4-minute games were not long enough I also did a test game at 10 minutes and it's presented below. Things are pretty even until Giraffe starts making some small positional bad calls in the ending. 
     The claim is that Giraffe taught itself to play at the IM level.  Compared to the super-GM level other engines play at that's pretty good.  But, does it really play at the IM level?  Judging by this game, it looks like it, but not being an IM, I can't say for sure.

No comments:

Post a Comment