Author: Robert Hyatt
Date: 08:23:35 09/19/05
Go up one level in this thread
On September 19, 2005 at 10:06:16, K. Burcham wrote: > > >Pretty remarkable, and it shows that they were extremely strong compared to >everyone else during that period. >DB was just "a lot faster, and a lot smarter" than deep thought. It was (and >would still be) competitive... > >Robert Hyatt > >Robert, you have always had "faith" in Deep Blue playing in a tournament against >todays programs. What do you base this on? Mostly just a gut feeling? Is there a >game that you were impressed with Deep Blue knowledge? Maybe just the fact that >Deep Blue held its own against Kasparov? >I read your point that you thought Deep Blue was strong for several years, but >its competitors may not do so well against todays programs. I base it on several things. (1) their search speed was (and still is) remarkable. Deep Thought was able to find several of the Nolot solutions, while DB was 100X faster (I never saw any published results for DB for such tests, but they did publish speed numbers of course). (2) their eval was not pitiful. For example, we played in a game where I was noticing that Crafty was ahead, but in an opposite bishop ending. (not against deep thought). Hsu was sitting at the table with us and said "no, this is winnable. We reached this kind of position a couple of times and [some GM whose name I forget] helped us get this right. Later he showed me some DT analysis on the same position and explained why it was choosing certain moves, with a couple of strong players (like Valvo, Berliner, etc) looking on. All agreed it knew what it was doing. We didn't have such knowledge, although we won the game due to a blunder by our opponent. I saw examples of other "things" that they did that the rest of us were unable to do because of computational cost. They didn't have any computational cost to speak of, they just added in another parallel computation in the hardware and it was free. (3) I played them multiple games. I watched them in every ACM event they played in. I watched them in exhibition matches at places like SuperComputing, where they were playing GMs in 4-6 game exhibitions at the IBM exhibit booth. In getting to see them play, you began to realize "this thing is not just fast but stupid, it has chess knowledge that is letting it compete at a level that was unknown (for computers) in the early 1990's. As I said, I am one of the _very_ few that enter into these discussions here that actually _saw_ the thing play chess, that saw the output on the monitor, that occasionally saw the output of their system vs the output of mine when we were playing, etc... The very first versions played _UGLY_ chess. But by the time 1990 or so rolled around, they were beginning to play remarkable chess moves positionally, not just tactically. > >Maybe you are saying that with improvements between 1997 and 2005, Deep Blue >would be very strong today. Are you saying that Deep Blue, exactly the way it >was in 1997 would be competitive today, with its 1997 search depth and 1997 >knowledge? Yes. Based on the fact that deep thought produced a 2650+ Elo performance rating over 25 consecutive 40/2hr games against nothing but human Grandmasters. Do you believe every program of today could do that? I don't. Some, on new AMD hardware might pull it off, but not many. So yes, I believe they would be highly competitive, unless you really believe that to be competitive you have to have a FIDE-approximate rating of 2750 or beyond. Note that the 2650+ was produced by deep thought, not deep blue 2, which was about 100X faster than DT, and also had even more chess knowledge than deep thought or deep blue 1. > >I am a fan of Deep Blue, its hardware and what they accomplished. I have spent >days trying to find a line or move that todays programs will not play. I cannot >find this, in fact we know that Deep Blue could not see 44.Kh2 wins in game 2. >Instead Deep Blue played the draw move 44.Kf1. The knowledge and depth was not >there to avoid this move. Todays programs also will not play this with winning >eval. Some will play, but like you said once, not for right reason. For every good move played by DB, or for every bad move, you can find programs that will play the same good moves, or avoid the same bad moves. But the question is, what does that prove? 99% of the time, they play the good move even when it is for a seriously wrong reason, or they avoid a bad move for the same reason. > > [D] R7/1r3kp1/1qQb1p1p/1p1PpP2/1Pp1B3/2P4P/6P1/6K1 w - - 0 44 > > >2006 Unlimited World Open >Fruit 2.2 >Zappa 2.0 >Deep Fritz 9 >Crafty 21.4 >Deep Blue 3.1x >Shredder 10 >Hiarcs 10.2 >Deep Junior 10 > >kburcham I wouldn't begin to try to predict such things. Rather than saying that (a) they would be crushed or (b) they would crush, I tend to think that 1997 DB would be _very_ competitive today. At the very least. And had they continued development, they would crush everyone. Look at the problems computers have with Hydra, yet it is nowhere near as fast as DB would be today using ASICs rather than FPGAs...
This page took 0.01 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.