Author: Amir Ban
Date: 11:36:05 10/21/97
Go up one level in this thread
On October 21, 1997 at 13:23:29, Bruce Moreland wrote: >On October 21, 1997 at 11:55:20, Amir Ban wrote: > >>The PV given by DB up to ply 10 starting with 36.Qb6 Qe7 37.axb5 Rab8 >>38.Qxa6 e4 39.Bxe4 Qe5 evaluated +0.74: >[snip] >>1. The DB choice was based on evaluation. > >It is possible, and I would think extremely likely, that there is more >to this PV than what you see. Eight moves is not a lot to return from a >10-ply search, so it appears to be truncated, and given their preference >for expensive extensions, possibly extremely truncated. > The full line given is Qb6 Qe7 axb5 Rab8 Qxa6 e4 Bxe4 Qe5 Bf3 Rd8 Qa7 Qxc3 Bh5 +0.74. If you follow this line to the letter, you will probably get an even lower eval. The reason is the move Qa7? which simply gives away the position. I can't justify it at any ply depth. Even if you assume there are 4 hidden extra plies here, you are pretty close to the horizon so what's going on ? When I initially looked at the printouts, it was obvious that some of the PVs are just nonsense. I already mentioned the nonsense PV given for 37.Be4. People told me that there are some TT techniques that cause that, so I didn't press the issue (not that this is really a satisfactory explanation here. I already speculated on machine/program malfunction). The line up to Qe5 appears on several PVs and seems to be "in earnest", so I tried to see what others see from that point. >On the other hand, it would surprise me a little if a micro couldn't see >as much after giving it 8 plies of "solution" for free. > >I let mine search for an hour or so, and still +1.27 (from white's point >of view). > >>2. That evaluation was "non-standard" for our field. >>3. The analysis does not prove cheating, but is no great help in >>disproving it either. The status quo remains. > >I thought that this was going to be the primary evidence for cheating. >If this evidence doesn't work, what other evidence is there? > The analysis cannot ever prove cheating if you are allowed to hypothesize any evaluation function you want, since in that case any PV and any eval can be justified. Actually this analysis was an attempt to disprove cheating by finding between us a line that DB probably relied on. In that case we could perhaps dismiss the fact that the PVs it printed were actually different and its other obscure behavior in timing and PVs. We didn't find it. The origin of this investigation is Kasparov's belief that no computer could prefer axb5 over Qb6. The investigation also discovered things that are incidental to this central issue, such as the timing behavior (which in itself is serious enough to need an explanation). Remember Tan's evasive reply was "We are flattered that Garry thinks DB is not like other computers". OK. He was right. None of us do this, or can even come up with the answer months after the event. We give up, Dr. Tan ! Tell us the answer. >I mean, eventually someone is going to say that Karpov was making all of >the moves, and the whole match is something out of "Capricorn One", and >there will be no way at all to "disprove" that, will there? > Sure there is. If there is a good answer, the DB team know it. For example: "In this position we evaluate king safety at -2.5. Hey, maybe we are wrong, but that's what we do." "The axb5 PV surfaced when it did because of this-and-that mechanism". Amir >bruce Amir
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.