I don't know how I missed this, but the 5.5-0.5 defeat of Grandmaster Michael Adams earlier this summer by the chess computer Hydra was a major event for chess, and probably for the field of artificial intelligence as well. It likely signals the end of the whole "man vs. machine" phenomenon in chess, because after this it's hard to see, given this result, how even the best players will be able to compete anymore.
What impresses me most about this latest computer is its ability to "prune" irrelevant variations before they are analyzed; whereas Deep Blue analyzded every single possibility, no matter how unlikely or disadvantageous, Hydra dismisses unlikely paths before that analysis gets started, which helps speed up the process and allow more in-depth investigation where it's more likely to pay off. This moves chess computing in the direction of pattern recognition, so that it more closely resembles a human player's thought process (except, of course, in terms of the available computing power).
Just to put some numbers out there... Deep Blue averaged 12 ply (that's 6 moves for white and 6 moves for black or vice versa) in analysis. In a rook ending against Adams, Hydra averaged 26 ply and at one point followed a variation for 55 ply in its calculations. The fact that it won this ending in convincing fashion is really disturbing. In general computers have performed horribly in the endgame where strategic long-term thinking is key, but it seems that Hydra has breached some analysis horizon where such long-term thinking has finally become a computational possibility. It's not really surprising to see Adams get crushed tactically, but to see him outplayed from equal or even slightly worse positions (in the one draw that he managed Adams had what every human commentator thought was a nice opening advantage which slowly deteriorated until Hydra played a positionally "unthinkable" move that turned out to be tactically justified. In the end Adams had to sacrafice an exchange just to hold a draw) is kind of horrifying. Even worse was an opening novelty that the computer found in a position that has been topical with the world's best players over the last two years. The computer not only found and played a move that no human had been considering, it used this "idea" to beat a Super-GM. Undoubtedly, humans will begin to follow suit in this opening variation...
Interesting that the first example (and maybe the second as well) goes back to the whole brute force model of analysis again. I assume the reason Hydra was able to go 55 ply was that many of the positions had been ruled out by its pruning faculty?
I guess I'm not as impressed by all this when I wrote the post before. It doesn't seem like it would take a very sophisticated program to know when to throw out a variation; all you really need is a second scoring system that goes on concurrently to decide whether a particular position holds promise -- and of course those variables are input by humans, just like all the rest.
Post a comment