Author: Vincent Diepeveen
Date: 05:48:01 07/04/05
Go up one level in this thread
On July 04, 2005 at 08:26:30, Vincent Diepeveen wrote: >On July 04, 2005 at 08:17:40, Thomas Lagershausen wrote: > >>>>37.Be4 !! after 1 Min 44sec on a Athlon 3800+ (512 MB Hash) >>> >>>As proven by Seirawan 37.Qb6 is 1-0 >>>Be4 is not so clear >> >>Why is 27.Qb6 a win? >> >>Where can i find the analysis of GM Seirawan? >> >>All this is new for me. >> >>TL > >This is not new, i posted this already 100 times. > >See analysis of Seirawan ICCA journal june 1997. > >Kasparov has drawn attention away from his poor play by doing statements like >this. That Kasparov played real ugly bad that match, don't tell me that this is >news for you, is it? The reason why Qb6 is winning is a simple one. After Qb6 and a few moves later you exchange queens you have a won endgame of opposite bishops. Very easily won. However for computer knowing about opposite bishops, and surely in 1997 this was primitively implemented, it will give a penalty and not take the opposite bishop endgame. This apart from fact that you must realize Deep Blue just searched 12 ply here. I realize how hard this is to realize for you, but without hashtables and without nullmove 12 ply is quite deep. You guys are comparing to a real ugly bad program from 1997, which for 1997 standards searched deep in the opening (10 or 11 ply), but which of course real ugly tiny depth even for those days in endgame. Here is exact output of Deep Blue for this move 37: hash guess Pa6b5p,Guessing axb5 8(4) #[Qb6](30)[Qb6](30) 30^ T=1 qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Rc8b8 ra2a6 Qe8d8 pd5d6 Bc7b6 8(6) #[Qb6](53)#################################### 53 T=4 qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Rc8b8 ra2a6 Qe8d8 pd5d6 Bc7b6 9(6) #[Qb6](53)#################################### 53 T=9 qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Rc8a8 kg1h2 Ra8b8 pg2g3 Qe8f8 10(6) #[Qb6](55)#################################### 55 T=33 qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Qe8f8 kg1h1 Bc7d6 ra2a6 Rc8d8 ra6a7 11(6)<ch> 'ab' #[Qb6](32)########[Be4](37)############################ 37 T=182 bc2e4 Rc8b8 pg2g3 Qe8d8 ra2a6 Ra8a6r ra1a6R Bd6c7 ra6f6P 12(6)[TIMEOUT] 37 T=199 bc2e4 Rc8b8 pg2g3 Qe8d8 kg1g2 Ra8a2r ra1a2R Bd6c7 qf2a7 Bc7b6 qa7a6 Qd8d7 --------------------------------------- --> 37. Be4 <-- 3/37:56 --------------------------------------- So Deep Blue sees initially Qb6 with a score of 32 which is exactly 0.5 pawn. Then later on it plays Be4 with a score of 37/64 = 0.58 pawn You see th eproblem of deep blue? It was a 2200 rated program with near to no evaluation. It didn't even know like most todays programs that having a queen there is great. At a depth of 12 ply you hardly even see all the tactical lines. I can't test this position for you at diep now, as diep is busy doing a long test for the world champs at my machine. However, you will see how quickly todays software skips the 12th ply. Most people here are just completely away from this planet. In 1997 chessprograms play so ugly bad, when seen from professional level, that Kasparov could do anything and still win the match with eyes closed. Todays software is so strong and agressive, as you can see from scoring of deep blue this is not the case for 1997, deep blue was a piece square table program with some gnuchess mobility elements in its eval, which for 1997 was impressive, but sure not close to todays standards. If you would play the full blown deep blue from those days with the book of those days; deep blue had an impressive 4000 bookmoves hand given in by grandmaster, versus todays chesssoftware have a 2 million bookmoves hand given in tournament book by openingsbook creators, who will win? If you put Diep at a quad, you will really win with a 100% score against deep blue from those days. Not a single draw will be there. It starts with that you play a better opening, then you outplay based upon knowledge it in the middlegame, then if deep blue is so lucky to reach the endgame, you outsearch it a ply or 10, which for sure will lose for the passive deep blue. Todays agressive tuned, deep searching software, is really no compare to a chesscomputer of those days. By 1997 the deep blue project was completely outdated. It was obvious its branching factor + overhead was not good enough to survive in the modern time. Hydra has been programmed in a programming language for the FPGA. That is portable to new technologies and new cards. Deep Blue was made in the logic blocks which is non portable. So it would run forever at those 25Mhz chips in 0.6 micron technology (even in 1997 that was outdated technology). Deep Blue couldn't get tested well, because only once each half year or so they could print cpu's. Imagine that you can only do 2 modifications a year in a program nowadays. See the problem of tuning? They printed the deep blue chips 2 weeks before the match. During the match, deep blue crashed many times. See todays software crash during games? I expect to see ZERO crashes of the top programs in world champs 2005. Kasparov really served it well by playing so bad. The completely outdated project could get stopped, put a nice box around it, and serve it out. Kasparov first minutes after the match was very polite towards IBM team, i saw this live at CNN. Then of course his management told him: "you idiot, you should have won the match, you will NOT get a rematch". Then kasparov realized what he had done and furiously reacted again live at CNN. Kasparov slammed around him with all kind of stupid statements. As if the poor level of the match, deep blue could have cheated. Well let me ask you, a 2200 rated match, if deep blue cheated, it sure was some poor player from 1800 or so who did the cheating, for sure not a GM, as the total level of the games was so so ugly bad. 5 to 6 bad moves per game. You cannot compare it to today. Programs get huge search depths now and in combination with better optimized chessknowledge and algorithms and agressive play, there is a world of difference between todays software and a chesscomputer from the far past. Hydra gets tested every day, of course it is difficult to make any chessknowledge improvements in Hydra. Far more difficult than software, but at least it is POSSIBLE to do. In deep blue this was near to impossible to do. We are talking about this: 0x9239349219 s0sdlk09 294929 010934091 0943090 09209 3091093409 19034091209430912209 0932090909 09019409 a b0d09 090 c 0d0909 e0d09 e090de Now from that you can remove the bug and 6 months from now you can see whether your bugremoval worked. Do you see how hard it was for Hsu to make a chip play chess *anyway*? He could just test on paper and then wait for 6 months or so until the next ASIC processors were printed in order to test it. For Hydra it goes easier already. It is programmed in a low level C a look like language (verilog). Still very hard to program, but if he makes a change then 24 hours later he can test that change. Software improvements go so so much faster nowadays. Vincent
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.