Author: Robert Hyatt
Date: 10:25:47 09/10/01
Go up one level in this thread
On September 09, 2001 at 08:27:50, Amir Ban wrote: >On September 08, 2001 at 23:24:41, Robert Hyatt wrote: > >>On September 08, 2001 at 14:07:22, Gian-Carlo Pascutto wrote: >> >>> >>>The question was a post where DB saw something the current >>>programs can't see. This is one. >>> >>>-- >>>GCP >> >> >>I wish I could give more information about "the position". All I recall is >>that Bert and I were sitting at the table, with Murray and Hsu on the other >>side. Hsu left to go to the restroom. after a couple of minutes into a >>search (I don't recall whether it was a ponder search or right after we had >>made a move) murray noticed that the program had failed high. He commented >>"it is picking on your bishop and the score just jumped." When I asked him >>"how much" he said "+2". Since we were not seeing any problem, Bert and I >>were assuming they had some sort of quirk in the hardware or software. But >>our score steadily dropped until several moves later the bottom dropped out >>and we were at -2 or so. >> >>I remember that the problem started on the move of the c-pawn, as discussed >>a year ago or so. And at that point their search failed high. I don't recall >>the depth. I do remember that in their PV (the part displayed by the software) >>they indicated which moves were singular-extended, and most of them were in >>that particular position. >> > >27...c5 doesn't win material, and it's easy to see this by the way the game >continued. > >White's 34.Nb2 is an unforced blunder. > >White can play better earlier, like 32.Bg5, and it's black advantage but not >more. > >All this has been pointed out before. What we are asked to believe here is not >only that DT saw a (non-existent) combination no one else can find, but that so >did Cray Blitz. > >Amir > A couple of things. 1. I obviously can't be certain about what DT saw. I personally saw their score. I personally saw their PV although as I mentioned before, their PVs are always short since the hardware doesn't back them up and the hash table is not always a reliable source for obtaining a PV. IE I know what I saw during the game. 2. I obviously can't be certain that their +3 eval (really more like +2.4 when you factor out their pawn value) was due to something forced vs something that was due to a bug in their search or hardware. I wouldn't know how to prove that even if I had their hardware in my office. I saw Cray Blitz play what appeared to be an ugly move against Nuchess in the 1984 ACM tournament, only to discover that it gave NuChess the opportunity to make a fatal error that would lose the game, which it promptly did. We could never reproduce that move again. Was it a bug? Was it a hashing artifact? Was it cosmic rays? No way to prove this, period. So proving whether their fail high was a bug or real is not going to be very easy. 3. Cray Blitz did _not_ see the combination. It blindly played on, with its eval slowly dropping, until it finally saw it was in trouble. If you want to say they won because we played poorly, perhaps that is right. Perhaps what you call a blunder is really a reasonable move _if_ your program can see deeply enough to realize it is already in trouble at that point. Again, this is difficult to prove, and probably not worth the effort either, since those that don't want to believe DT/DB was very strong are not going to believe it no matter what evidence is shown, and vice-versa. I personally prefer to rely on logic. And logic tells me that a machine that can search 200X faster than mine is going to be far stronger than mine. They proved this over a long period of time via WCCC and ACM tournaments. They then proved it against Kasparov and other GMs at the various exhibition matches they played. If you really think your programming skills are enough to offset their 200X speed advantage, knowing Murray, Hsu, and the rest, then that is just an opinion you will have to keep. I don't believe your (or anybody else's here) skills are better than mine, and I don't believe mine are better than the DB team's skills. I also don't believe that mine are better than anybody else's, nor do I believe there is much difference in programming skills among most of the chess-program authors. Which leads me to believe that in that light, a speed advantage of 200x is simply a killer. With Cray Blitz, people used to say "the program is not very smart, it is just 10x faster than everybody else, which is why it wins." If 10x was good enough back then, 100x is _certainly_ good enough to do the same today. (and many are now learning that CB wasn't just fast, it probably did some things that other programs were not doing at all, both in evaluation and in search. the "just fast" excuse is just that, an excuse. And a very feeble one at that.) Back to the DT/CB game. I've seen _nothing_ that proves that the pawn push doesn't win material. I've seen nothing that proves that it does, either, other than their output. I don't buy any of the mucking-around analysis either. It may well be that a highly non-obvious move is needed at some key point to pull this off, and humans tend to miss those moves and conclude that it doesn't work when it actually does. DT saw something. I personally saw the output, as did Bert Gower, Murray Campbell and Hsu. Perhaps Mike Valvo even saw it, as he was the TD that year. Whether it was a reliable piece of analysis, or the result of a bug, is impossible to tell. Of course, you are free to conclude anything you like. Including stuff like "Deep Junior is superior to DB" and so forth. Impossible to prove. Improbable to believe. But a statement, of sorts, nontheless. I doubt you would have made such a statement between 1987 and 1995 however, because it would be very likely that you would have had to back that up with a game, which would not be so easy as it is today with no DB available to contradict such a statement. > >>I had seen them do this to others, so it wasn't particularly remarkable at the >>time and I didn't take any careful notes. I did make a note on the log file, >>but that paper log has long since been lost somewhere or other.
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.