Author: Robert Hyatt
Date: 07:31:57 08/30/01
Go up one level in this thread
On August 30, 2001 at 05:28:14, Gian-Carlo Pascutto wrote: >On August 29, 2001 at 23:10:56, Robert Hyatt wrote: > >>The "weights" don't always reveal what they stand for. I've had the DT >>code for at least 2 years. It isn't new. But I can give you the weights I >>use in Crafty and they don't tell you _what_ I am actually doing with them, >>only the numbers I am using. Their bishop of opposite color + pawn ending >>evaluation was _very_ good. Once Hsu explained it to me at an ACM event. >>You won't find that explanation of how it works in the stuff you reference. >> >>Which is a shame, actually. There's a lot in the thing that we won't ever >>know in great detail. > >The code includes the full eval itself. You can check it out and see >how they did kingsafety, bad bishops, passed and blocked pawns, etc... > >The tuner has that code because it is useless without it. You can't >tune an eval if you haven't got any. When I looked at the code it was horribly messy. I could not figure out everything they did without someone to ask questions to... The version of this code I had was pretty old, and the comments were few and far between. I haven't looked at what is on Tim's site to see if Andrew cleaned it up and made it more readable. If he did, that is a good thing. > >What I don't see is the endgame stuff you talk about. I see two possible >explanations: > >a) they thought it was so great that it shouldnt leak out and carefully >removed all references from it from the tuner > >b) they simply didnt _have_ it yet at the 1988 US Open. Perhaps it was >added afterwards in DT, DT2 or DB, and you are confused about when they >talked about it to you or implemented it. I am _certain_ it wasn't there in 1988. Hsu mentioned the Opposite Bishops stuff around 1993-1994. We were playing in an ACM event and we had reached such a position (I don't recall which side we had). Both thought it was a draw, but Hsu pointed out that they had had a long conversation with a GM the previous year and they had specific knowledge in DT that knew how to win this ending. He explained why this particular case was winning and another GM there said "that is correct..." I think the code Tim has goes back far further than that... > >Make your pick. > >I think what the code shows is that in the 1988 US Open, Deep Thought >did not have great sophisticated evaluation. An ok one yes, but it's >certainly been surpassed by the micros in the meantime. > >Which doesn't mean anything about the evaluation of DT after 1988 or >of its succesors, but I find it awkward to be making much fuss about >DT's supposed evaluation if you can _look_ at it and see what they >did and did not do. You really can't do that. You have a snapshot that is probably 13 years old. It changed a lot over the years, as they have said. I doubt you will find any 1988 programs that did what they did, for example... And don't forget, it _was_ good enough for that 2655 rating over a bunch of 40/moves in 2hrs games... > >-- >GCP
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.