Author: Robert Hyatt
Date: 08:07:48 07/18/00
Go up one level in this thread
On July 17, 2000 at 22:18:18, stuart taylor wrote: >On July 17, 2000 at 20:14:39, Robert Hyatt wrote: > >>On July 17, 2000 at 19:10:26, stuart taylor wrote: >> >>>On July 17, 2000 at 11:23:52, Robert Hyatt wrote: >>> >>>>On July 17, 2000 at 09:44:57, stuart taylor wrote: >>>> >>>>>On July 17, 2000 at 09:32:18, Robert Hyatt wrote: >>>>> >>>>>>On July 17, 2000 at 08:05:57, blass uri wrote: >>>>>> >>>>>>>On July 17, 2000 at 07:22:41, Graham Laight wrote: >>>>>>> >>>>>>>>I'm afraid I still feel that Junior could have come out ahead (instead of >>>>>>>>level)in this tournament by beating Bareev and Khalifman - and possibly by not >>>>>>>>losing with such apparent ease to Kramnik. Continuing the game against Anand >>>>>>>>might possibly have gained an extra half point as well. >>>>>>>> >>>>>>>>I think that Amir has an aspiration to make his program demonstably better than >>>>>>>>Deep Blue (this certainly comes across in his interviews published on the >>>>>>>>Chessbase Website coverage of Dortmund (www.chessbase.com) before the Kramnik >>>>>>>>game). If so, as a (hopefully!) impartial member of the viewing public, I'm >>>>>>>>afraid to say that I've yet to be convinced. >>>>>>>> >>>>>>>>As evidence, I point firstly to the games against Bareev and Khalifman. On both >>>>>>>>occasions when Deep Blue '97 gained an advantage over Gary Kasparov (who's a >>>>>>>>better player than anyone at Dortmund was), it parlayed that advantage into >>>>>>>>victory - whilst Deep Junior twice failed conspicuously to "slam in the lamb". >>>>>>>> >>>>>>>>I would also point to the game against Khalifman. Here we see Deep Junior lose >>>>>>>>to a combination of blocked centre and king attack - classic anti computer >>>>>>>>methods which have both been well known for a long time. They work because, in >>>>>>>>this case, nothing short of truly massive search depth is going to help you to >>>>>>>>make the correct moves. >>>>>>>> >>>>>>>>However, for both king attack and blocked centre, Deep Blue '97 demonstrated >>>>>>>>that it's evaluation knowledge was able to adequately handle the challenge. >>>>>>> >>>>>>> >>>>>>>I guess that the evaluation of Deep Junior could do better if Deep Junior could >>>>>>>search the same number of nodes. >>>>>>> >>>>>>>I believe that Deep Junior is better than Deeper blue if you assume 200,000,000 >>>>>>>nodes per second for deep Junior. >>>>>>> >>>>>>>Uri >>>>>> >>>>>> >>>>>>I believe pigs can fly. But only if you increase the density of the atmosphere >>>>>>by a factor of 10,000 or so. >>>>>> >>>>>>DB has two almost insurmountable advantages: (1) it is faster than anything is >>>>>>going to be for a _long_ time; (2) using special-purpose hardware they did >>>>>>everything in the eval that was suggested by GM players, because they could do >>>>>>so with no speed penalty. DJ and every other PC program has _many_ >>>>>>"concessions" in the evaluation due to speed considerations. DJ's king safety >>>>>>would fail if it was 1,000 times faster... because there are some things that >>>>>>speed won't help until we reach the point where the computer can see 30-50 plies >>>>>>into the future. You either understand the Stonewall (and its kin) or you get >>>>>>beat by it, regardless of how deep you can see. I don't claim to have solved >>>>>>this either, but I don't see Crafty losing Stonewall games on ICC today, where >>>>>>3 years ago it was getting killed by this attack, and my defense was to hack the >>>>>>book repeatedly. It will certainly lose one every now and then as my randomness >>>>>>(on ICC) will occasionally cause it to play a stonewall as black. But book >>>>>>learning closes that hole, and once out of book, it doesn't have great >>>>>>difficulty avoiding the problem pretty well. >>>>>> >>>>>>There are a couple of ICC "regulars" that are a problem for computers, >>>>>>cptnbluebear is one, and insight is another. cptnbluebear doesn't play crafty >>>>>>much any more because other programs are easier to 'stonewall'. Insight still >>>>>>plays a lot, but he _rarely_ wins. He seems to primarily play for draws, which >>>>>>are easier to do, but still very difficult to pull off. >>>>>> >>>>>>I've done this with special eval code, not with speed... and I have a long way >>>>>>to go myself... >>>>> >>>>>To Dr. Hyatt, So how far do you beleive it is possible to go without tremendous >>>>>speed? If software was maximised the most possible, could 1ghz. ever overtake >>>>>D.B.? or maybe 2 ghz? What is the potential that still hasn't been realised? >>>>>S.Taylor >>>> >>>> >>>>this is an old theoretical question. A similar one: what is the maximum >>>>bandwidth over a single piece of copper? Answer? 1 / signal-to-noise-ratio. >>>>If you get SN to 0.00, the bandwidth is infinite. But that is quite hard to >>>>do of course. :) >>>> >>>>same thing for chess engines. In theory, today's hardware ought to be fast >>>>enough. But the programming is hundreds of years behind what evolution has done >>>>to our "personal biological computer system" we all carry around. It will catch >>>>up at some point of course. >>>> >>>>As far as overtaking DB, that is another matter. Whatever commodity micro- >>>>processors can do, DB (or a new successor) can do 1,000 times faster, easily. >>>>So the hardware we use won't _ever_ be as good as the special purpose hardware >>>>that can be designed/built to handle a specific thing like chess. >>>> >>>>The current DB is going to be untouchable for at least another 5 years, maybe >>>>closer to 10. By then Hsu _could_ do something that would again be untouchable >>>>for another 5-10 years. The special-purpose vs general-purpose issue won't go >>>>away, ever, most likely. >>> >>>Atleast you seem to be conceding that PC software MIGHT overtake the latest >>>version of DB, due to better computing. I mean, it might start catching up a >>>little bit with our "personal biological computerized system", enough, even >>>before another 5 years. >>>S.Taylor >> >> >>I have _always_ conceded that micros will catch 1997 deep blue in 5-10 years, >>based only on raw hardware improvements. I think I have said that dozens of >>times here. Of course, a new DB-3 chip would spread that gap back to a factor >>of 1,000-2,000 times faster again... assuming anyone was interested in building >>the thing... > >I'm sorry I don't read everything. >Hardware? x over a thousand? >But you were conceeding that software can do it-theoretically. >S.Taylor Not exactly. I said "Deep Blue is approximately 1,000 times faster than today's PC's, assuming 200Knps average for todays programs, and given DB's 1B nps peak and 200M nps average in 1997. Based on that, considering the fact that hardware speeds are doubling about every 18 months, take log2(1,000) (roughly 10) and multiply by the doubling interval (1.5 years) and you get 15 years for the hardware to catch up to deep blue. Nothing to do with software. Just pure hardware advances...
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.