Author: Robert Hyatt
Date: 07:17:03 11/16/98
Go up one level in this thread
On November 16, 1998 at 09:55:46, Amir Ban wrote: >On November 16, 1998 at 02:13:01, Ed Schröder wrote: > > >>>Yes, you got it right. >> >>>The PV for Deep-Blue on all iterations except the last starts 36.Qb6 Qe7 >>>37.axb5 Rab8 38.Qxa6 e4 39.Bxe4 Qe5. In the last iteration, there's no PV. >> >>Thanks for posting the main variation. I clearly remember the end position >>again. Based on this impressive main variation I can come to no other >>conclusion other then that Deep Blue must have a speculative king safety. >> > >I'm not ruling this out as an hypothesis, but please tell me: How does a >speculative king safety agree with moves such as g5 in the 1st game, b4 in the >4th game, and a generally solid style throughout the match ? > > two choices: (1) in that game, I believe that even Kasparov said that g5 was the only real possibility to consider there. At least one other GM said this during the match as well. I don't recall exactly why they thought this was necessary, but some did. (2) it is not difficult to have a king safety evaluator that will suggest g5 there. I've done it *many* times although I generally consider it a "bug" in such positions and try to tune it out. But in doing so I would *only* be looking at the king-safety issue and the gash in my kingside pawns that g5 produced. But if there are other deep issues here, and some suggested they were, then g5 isn't so mysterious. We ought to run some deep searches to see how different programs evaluate g5 and whether they would play it, or how close it is on score to the move they would prefer instead (if they won't play it.) >>The other explanation, a bug, sounds not fair to the Deep Blue team. >> > >Actually bugs, or general malfunction during this part of the game, is quite >high in my list of possible explanations. There are a lot of strange things >going on. Here's one of them (the PV for iteration 10): > >36. Qb6 Rab8 37. axb5 Rab8 38. Qxa6 e4 39. Bxe4 Qe5 40. Bf3 Rd8 41. Qa7 Qxc3 42. >Bh5 > >Question: how to explain the appearance of the move 41.Qa7 in the PV ? > >Amir I believe they have answered this already. They have problems in getting a PV back from the "hardware" as I understood it a good while back. IE we use an array to back up the PV when we back up a score. They can't do this, because the hardware doesn't work like that, and there are multiple chess processors for each SP processor, making this impossible to handle since the processors can't communicate with each other. As a result, I believe that they use the hash table to construct the PV after the search, since the hardware can't push the PV back to the SP processors. And doing this, you can definitely get bizarre PV's. I played with mtd(f) a year or so ago, and tried to do this to construct a PV since *every* move fails high or low and my PV code was not going to *ever* return a PV. And when I did this, I got some "interesting" stuff back, which was one of the reasons I gave up on the idea. I believe others have tried doing this (PV reconstruction by probing the hash table) and found the occasional "oddball" move on the end... I "extend" my PV when I get a hash hit by going to the end of the PV, and probing to get the best move, then making that (if legal) and probing to get the next move, and so forth. And I have, on rare occasions, gotten something odd. If you check my "ponder" code you will notice that I check the 2nd PV move for legality, because it might have come from the hash table, once every now and then it would be illegal. ANd when I "made" it it tore hell out of my data structures and resulted in a lost game (flag) on ICC. If it can happen to me, I'd assume it can happen to them... probably more frequently to them if they have to use the transposition table to form *every* PV. And at their search speed, anomalies would be even more likely...
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.