Author: Robert Hyatt
Date: 17:58:21 10/17/99
Go up one level in this thread
On October 17, 1999 at 19:32:48, blass uri wrote: > >1)I understood from odell hall that lev albert said that deep blue and not deep >thought is 2500 elo(I believe deeper blue is clearly better). > IF he thinks DB is 2500, then DT must be lower, correct? Because DT searched 2-4M nodes per second, and was replaced by DB1 and then DB2. Now if DT was way under 2500, what is the probability that it could play 25 consecutive games and produce a performance rating of 2650? What is the probability that it could take a very low initial rating (due to many hardware/software problems early in its life-cycle) and _still_ produce a real rating of 2551 (USCF)? I'd say the odds that DB was only 2500 is approximately zero. Because DT was obviously stronger than that, and DB1 was 100x faster + a much better evaluation in hardware (DB2 was even better in the eval department). >2)The difference against humans is not the same as the difference against >computers so you cannot learn from the rating of DT or DB and the results >against micros about the rating of micros aghainst humans. See above. _ALL_ of those games were 40/2hr games against nothing but humans... USCF didn't allow computer vs computer games in their rating system. > >Uri >> >>either he is wrong, or the micros are 2100. I don't believe micros are 2100 >>players... I think they are 300-350 points higher. >> >>IE if you pick the right set of 25 consecutive games from deep thought, you get >>a performance rating of about 2650. If you take _all_ games it played, >>including early deep thought versions with bugs, you get 2551. All of these >>are USCF ratings of course. Best guess is to subtract 50-60 to convert to FIDE. >>Either is impressive. > >I do not know if deep thought has better results against humans in the last >games that it played. >You are probably right that the early deep thought had more bugs but humans did >not know how to play against deep thought before 1990 when they knew better >later. > However, the humans did best when DT first came out. The last 25 games it played were simply monstrous (TPR=2650). >I know they won a tournament in 1989 with 6.5 out of 8 and with performance >above 2600 when they did later only 2.5 out of 7 in a tournament (performance >2410). > >I read that they had better hardware and could see 10,000,000 positions per >second the second tournament when in the first tournament of 1989 they could see >less than 1/10 of it. According to Hsu DT could never search 10M nodes per second. That was probably the 'theoretical max" while reality was about 1/3 of that or so... > >They did better result later when they lost 2.5:1.5 agains bent lersan and won >the danish team 3:1 in a tournament time control. > >Uri As I said, they had their share of technical problems. Cray Blitz lost games for the same reason. Once one (out of 16) cpus forgot how to multiply by -1 and get the right answer (sign). I used that to change sides. It simply lost instantly. Very much like the Rebel game where the hardware was flakey. It happens. The games 'count' as far as USCF is concerned.
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.