Author: Vincent Diepeveen
Date: 04:12:16 08/20/02
Go up one level in this thread
On August 18, 2002 at 21:48:08, Robert Hyatt wrote: >On August 18, 2002 at 21:15:47, Vincent Diepeveen wrote: > >>On August 18, 2002 at 15:08:49, Gian-Carlo Pascutto wrote: >> >>>On August 18, 2002 at 09:06:02, Jorge Pichard wrote: >>> >>>> Kasparov proved that he can defeat programs at fast time controls when he >>>>defeated Deep Thought in a game/90 two games match in 1989. This program was >>>>weaker than Deep Junior is today, as it searched well over 2,000,000 NPS, but >>>>didn't have as much chess knowledge as Deep Junior. He also defeated Deep Blue >>>>in 1996. This program is obviously much faster than Deep Junior is today, but in >>>>my opinion Deep Junior still has more chess knowledge than Deep Blue had back in >>>>1996. >>>> >>>>PS: It is hard to compare Deep Blue of 1997 vs Deep Junior of today, but in my >>>>opinion Deep Junior Chess Knowledge could make up for the difference of Deep >>>>Blue super calculating power of 1997. >>> >>>How do you know all this? >>> >>>How do you know Deep Junior has more chess knowledge? >>> >>>I mean, we don't know what Deep Blue evaluated exactly (save a few things >>>that are published). We know *nothing* about what Deep Junior evaluates >>>exactly. >> >>yes we know that. Look at the paper it describes about 40 patterns and if >>you multiply that with arrays of 64 (that's how it goes in hardware) >>and add to it piece square tables it is exactly what theydid. of course how >>well defined the patterns are is a different case. >> >>We can see that at the rude play very easily. Look at game 1 from 97, >>where it played manoeuvres like qa5 bc7 qc5 which even seirawan comments >>correctly in his 1997 analysis. Gnuchess accuracy it is. Very rude and >>primitive, but for a program with a leaf evaluation (even though some >>tuning by preprocessor took place) with several tens of patterns (and >>as a result of that several thousands of adjustable parameters) that >>means it was searching deeper than any program with that amount of >>knowledge in evaluation in 1997. I for sure had more in 97 (though i >>used arrays less back in 97 than i do now as i'm not hardware but >>software and L2 caches were performing bad in general back then until >>pentium pro which took a few years to adjust to) so had others, but we >>all shared that at a 200Mhz pentiumpro we searched 8-9 ply, NOT 11-12. >> > >I'm not going to comment on the rest of your nonsensical statements, but >the above is clearly wrong and that is provable. I played in Jakarta on >a pentium pro 200. And _My program_ searched 11-12 plies. I have the logs >to prove it. And anyone that wants to download the crafty (jakarta) version >can find the same thing... > >So if you are going to make statements, at _least_ verify that there is some >basis of truth to them first. _you_ might not have been able to hit 11-12 >plies on a P6/200, but I did... And others did as well. - no checks in qsearch - forward pruning last ply which also hurted nullmove incredible - no mating extensions (not solving win at chess 141 even soon which in 1997 was a 9 ply trick for me). Fritz3 also got 11 ply but it was also with a lot of dubiousy. In fact some hit 15 ply as well with major forward pruning back then. But that is not a fair compare. We must compare programs that searched in the same way Bob. So not forward pruning, at most nullmove. and strong in the leaves. Crafty even today is very weak there. > > > > >>So then 11-12 wins simply. 8-9 with evaluations from that period lost >>simply. period. >> >>Evaluations 8-9 nowadays are a different case (extending way more nowadays >>too than in 1997 too). The evaluation of DIEP is a top grandmaster relatively >>seen when compared to 1997 where it knew shit from endgames for example. >> >>>As a consequence, you can't possibly support any of the claims or >>>suggestions you make. >> >>Yes we can. The biggest evidence is the games played. Statistical evidence >>on how programs play moves is the best. The major problem is that you need >>to invest time if your chess level is not so high to see it and even a high >>rated chessplayer who knows nothing from how chessprograms evaluate will >>completely fail here (though Seirawan came pretty far but as he was paid >>by IBM he described it in a positive way, leaving conclusions to the reader). >> >>We have seen some marvellous conclusions already by Uri here based upon >>logfiles from the IBM computer. From evaluation viewpoint we >>see for example from the mainlines that it gives a big bonus for a bishop >>attacking its own queen. We also see it only cares for how many squares the >>queen can go to, not caring for patterns there. Very basic things which >>were at the time very normal in gnuchess type programs. >> >>We also see that it knows really nothing from good/bad bishops (not >>surprising, only 1 program had in 97 this thing and it was mine). It >>simply didn't care for the center at all. This is amazing nowadays >>comercial programs *only* care for the center. >> >>Also its knowledge on passers was very primitif. We see for example that >>it doesn't see difference even between covered passers and very good >>blocked passers. Regrettably that didn't happen a lot on the board. >> >>The most amazing thing by far is its huge penalties and bonuses for a few >>king safety things. that of course led to big patzer play which is nice >>and nowadays very normal. These penalties/bonuses are in complete >>contrast to pawn structure aroudn the king. In many games we see >>major mistakes here. game 1, but if i remember well game 4 where >>deep blue castles long and then plays horrible king moves and b4 b5. >> >>From the many king moves in the game and in the logfiles we see clearly that >>it had a very primitif 'opponent pieces to my king' distance feature. >> >>I remember how DIEP back in 95,96 wanted also always ka1 because that would >>mean the king is further away from the pieces. A very basic mistake we >>still see in some engines. It is a non-preprocessor mistake obviously. >> >>but it doesn't take away that the pattern is a very primitive heuristic. >> >>nearly all kind of bad moves are explained by simple bugs in evaluation. >>100% the exact bugs gnuchess also has. >> >>the comparision with gnuchess is not fair, but for evaluation it sure is. >> >>We see that the 'new gnuchess', sorry to call it like that, zarkovx, >>is the program which when getting 10-11 ply is playing from all chess software >>nearly exactly every move which deep blue also played. don't use the >>dos-zarkov, but i mean the 4.5xx versions of zarkovx where John hardly >>nullmoves the last few plies (they take some time to get 10-12 ply, >>horrible branching factor). It makes the same weird moves, same mistakes, >>same strong moves. It is a perfect match for how deep blue played. >> >>A person who can't play chess at all and whose program is exactly making >>the mistakes a beginner makes when making a chess evaluation. >> >>Best regards, >>Vincent >> >>>All of this talk about Deep Blue this and Deep Blue that is just pure >>>bullshit. Maybe Fritz 7 would kick its ass. Maybe Fritz 7 would get >>>its ass kicked. Maybe they're about as strong. I dont care either way >>>since Deep Blue doesn't exist anymore and it certainly doesn't look as >>>if it's ever going to play again. So why care about it? Why keep making >>>totally unfounded speculations? What's the frigging point? This kind >>>of discussion comes up about once in every 2 months and there has NEVER >>>EVER come anything insightful out. Instead, a lot of people are making >>>claims or saying things that they can never ever support, or even are >>>demonstrably wrong. >>> >>>Mention the words 'Deep' and 'Blue' to anyone who works in computer >>>chess, and all sanity suddently grinds to a halt. >>> >>>-- >>>GCP
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.