Turing Chess
At the HTM Workshop there was a lightning talk by David Doshay on Computer Go which is another application I thought of as soon as I read On Intelligence.
During the break after his talk a bunch of us were talking and he basically said that a lot of researchers were moving to Go because Chess was a solved problem (perhaps I should have pursued it more back when I was interested in 2001).
I asked if he knew of anyone who, instead of switching from Chess, was going back to make Computer Chess more human-like rather than simply better. He wasn't aware of any such work.
It seems to me that an interesting pursuit would be a sort of Turing Test of Chess where the goal is not to beat the human but to trick other humans reading a transcript of the game as to which was the human player.
(Yes, I read Blondie24: Playing at the Edge of AI a few years back)
Comments (5)
James Tauber on July 8, 2008:
To me a computer that is indistinguishable in playing style from a human at a particular level is a far more interesting insight into intelligence than a computer that can beat anyone but is playing in a completely different way (algorithmically or behaviourally) from a human.
A Turing Chess Test might therefore not even go far enough. I'd love to see algorithms that were biologically plausible as well as human-like in their behaviour.
Carl Patterson on July 8, 2008:
We know it has a saddle point, (we know go has a saddle point), but unlike say, tic-tac-toe, we have no idea what that saddle point looks like, nor do we know what optimal play is (unless I've truly missed something recently - a possibility).
Also, I enjoy the mix of topics on your site.
James Tauber on July 8, 2008:
Glad you enjoy the mix of topics. I often worry it drives people away.
Nicolai Czempin on July 16, 2008:
When Computer Chess started, beating humans was such a ridiculous-seeming goal (like beating humans at soccer seems to be now) that it was hard enough for many to get involved. It is this problem that is solved, but the way in which it is done ("brute-force") doesn't satisfy everyone.
One could stay in computer chess and try to find other metrics than playing strength if the goal is more in the idea of a chess Turing Test. Or one could change the "underlying" and move to Go or Poker, where for now the brute-force approach doesn't seem fruitful (although one of the Deep Blue people has already claimed that he'd find it feasible even for Go).
Any progress in Computer Chess certainly doesn't seem to happen in the academic world any more (although Robert Hyatt, behind the legendary Cray Blitz, is still somewhat competitive with his Crafty engine). The currently strongest engine, Rybka, was conceived by a private individual in his spare time (he has gone commercial by now), and he hasn't shared his "secrets".
I myself have coded an engine in my spare time, with the specific goal of a playing style that is "more human" (than the other engines at a similar playing strength). It is not competitive with the strong engine, but it is stronger than a large majority of human chess players. It doesn't use any revolutionary techniques, just a little more emphasis in a style of play that I considered to be human-like.
IMHO it would be possible by changing a few parameters to change any of the strong engines into one that feels more like a human player. It will also become weaker that way, and most of the coders seem to be most interested in getting the strongest engine out.
My engine "Eden" is written in Java (that alone makes it mostly non-competitive) and I have taken care that each new version plays stronger than the previous one, but in a similar way to the way human players progress.
I have received several comments that people find its playing style more "natural" than if they took a strong engine and made it artificially weaker by e. g. reducing its time per move to half a second or its search depth to one ply.
Add a Comment
Last Modified: July 7, 2008
Author: jtauber
Bill Mill on July 7, 2008: