Author: Mig Greengard
Date: 21:00:55 04/26/01
Go up one level in this thread
I certainly can't guess at what these recent additions to Junior are, but from looking at the 11 games there are several things that strike me as a master player with a lot of experience playing against programs. (I know, I know, 11 is an infinitesimally small sample, nothing less than 5,000 games at 2 hours per move will do...) It seems that Junior has taken some of the good of human play (positional evaluation) but avoided the bad (tactical weakness). At the simplest level of description it has played very much like how GMs usually lose to Fritz! Giving up pawns, playing for piece activity, opening diagonals, and avoiding piece exchanges are all very human things that computers usually do only when concrete calculations show a return of the all-mighty material. GMs get dangerous, even winning, positions against comps this way, then watch as the comps play perfect defense and win with the extra material. I believe CS Tal had similar problems playing against the fastest programs, despite playing aggressive chess that would be lethal against most any human. I obviously don't know precisely why this new DJ's eval shows almost no dip when it's down pawns in these positions, even when the pawn is a "clean" one and there is no forced draw. I suppose there are two approaches, one, more general, giving a high value to piece activity and giving pawns a low value, maybe even less than 0.6. The other would be deep, narrow searches that kick in when the previous activity value is a good one. The latter would be a step toward eliminating a huge weakness comps have against GMs: understanding the transition from middlegame to endgame. DJ's fearlessly entering these complicated (without forced and/or TB draws in sight) endgames down material and drawing easily (or even winning!) is very impressive in this regard. It's love of bishops is also clear, at least in these games. This is another human characteristic (that might be a problem for a comp against a human, because humans can more easily limit the scope of a bishop). Though not a practical example, near the end of game ten, Fritz with its useless extra pawn gave a reasonable eval of -.38 for itself. Junior with its bishop vs. knight saw it +.15 despite being down a pawn! In game 11, 27...Nb3 almost looks like DJ did not see far enough ahead to the loss of the isolated b-pawn after 28.Bxb3. But look at its eval throughout. It barely moves even when the pawn is taken and even gives itself a tiny plus, which it fully deserves. After 38...Ba7, black dominates the board with beautiful piece placement and it's no surprise DJ easily drew this (and with such pieces, it would probably even win against a strong human). All I can say is that its positional evaluations are excellent. The game eight loss was interesting as well, but this was a typical comp-comp win out the book despite the position being equal. Black's plan is much clearer than white's in that position and it would take long-term strategic play to play it well and understand the weaknesses. Instead, DJ panicked when it saw it might lose the b-pawn to Be7 and gave up the a-file, which was positional suicide. Playing for e4 is thematic in this line, but it needed to be played earlier. Basically though, a comp with white after move 16 is going to be in trouble. I don't see why it didn't play Ra1 early on and Ra2 was a complete waste of time, of course. On April 26, 2001 at 19:18:14, Fernando Villegas wrote: >Hi Amir: >Yeah, DJ is speaking and how it does! Let me guess: the thing on the top of J6 >is some kind of attacking algorythm. Kind of positional module or whatever >capable of smelling oportunities along very long shots. Or perhaps some kind of >tactical extension that goes very very very far. Well, anyway there is something >new there besides the sheer processing power of the "deep" thing. Be lucky. It >is neccesary to beat the blond beast. My grand fathers would love it.. >Shalom >Fernando
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.