Author: Don Dailey
Date: 14:37:25 12/25/97
Go up one level in this thread
>Your reply to my small comment has material to start 3-4 extra long >threads :) I hope so! >> Your program will look great in a few tactical >>positions where your extension really helps you nail it but in most >>cases all you will be left with is a general slowdown. >> > >A pawn push doesn't need to be good to be extended. Yes, and I this is a point often missed by people. The idea is that the search "learns" something about the position. >In most endgames, advancing passed pawns is not an important issue, it's >the only issue. I don't think of this as a slowdown, because the program >still does the same number of nodes/sec, doesn't it. If you would rather >look at 12 plies of aimless king and rook maneuvers rather than >concentrate on how to stop that dangerous pawn, then I would love to >play against you. As a completely random and AdHoc test, I opened up an endgame book called Chess Endings for the practical player by Ludek Pachman and picked out the very first position my eyes fell upon. I decided I would repeat the experiment until I happened upon a position with at least one passed pawn. The first position I looked at had a passed pawn, strengthening your argument that passed pawns are important (and which I do not contest.) 8/5kpp/8/1R6/1P4PP/5K2/8/1r6 w - - It was position 131 on page 113 of this book. The book line given is 1.Rb7+ Kf6 2.Rb6+ Ke5 3.Rb5+ Ke6 4.h5! h6 5.Ke4 Rg1 6.Rb6+ Kf7 7.Kf5 Rf1+ 8.Ke5 Rg1 9.Rb7+ Kg8 10.Kf4 Rb1 11.b5 Rf1+ 12.Kg3 Rb1 13.b6 Rb4 14.Kf3 Kh7 15.Ke3! After 21 ply, the first move of a passed pawn was made. This was preceded by 20 ply's of "king and rook manuevering." Now would aggressive passed pawn extensions help a computer understand this position? It's very possible, since the real reason for the maneuvering was definitely centered around the passed pawn. But the main point that this illustration should make obvious is that there is nothing aimless (or uncommon) about this kind of maneuvering in computer chess. >>A case in point is check extensions, most programs extend all >>checking moves, and it's certainly not because they are usually >>good moves because they are only occasionally good. But checks >>are usually just annoying delaying moves. They are important >>to extend even if they are completely irrelevant because they >>steal 2 ply from your tactical awareness. So even bad checks, >>or unimportant checks help return meaningful information. >> > >Yes. For a program to look at passed pawn advances only in consideration >of whether promotion takes place is very much like ignoring checks >because after all we only care about mate. >>If you do this well your program can remain quite selective without >>being stupid with regard to passed pawns. >> > >I don't know if this is directed at me. I'm not a null-mover, but I know >what you mean. I did find that the "regular" search techniques don't >cover pawn advance lines very well, so had to find something that does. >I don't particularly like the passed-pawn extension, there's something >so ad hoc about it. When I find something better, I'll adopt it in two >seconds. My comment about being stupid with regard to passed pawns was not directed at anyone. I was refering to the computer being stupid in the sense that it misses passed pawn opportunities because of prunning. >You may want to consider my recent experience in Paris. In only two >games did I play against a faster program, Ferret and Fritz, and both >were decided by pawn pushing, both offensively and defensively. I'm not >even nearly the world's greatest pawn pusher. Probably Chess Genius, to >quote another wildly extended program, deserves the title of Pawn-Pusher >Supreme. These are good wins. I know little about Ferret but do know it's highly spoken of. But I don't doubt for an instance that any extension algorithm will win games. My only question is have you every lost a game because of them? I'll bet you have, and I'll also bet that you probably don't know it. It's always very obvious when an extension helps the program but a loss is NEVER attributed to the extension. You don't even know how the earlier part of the game would have developed without the extension. You might also have won these games without the extensions. At any rate, I'm definitely interested in your extension ideas. I want to implement something more aggressive with Cilkchess and will get ideas from this discussion. I have to admit I am skeptical about being really aggressive with them. But that doesn't mean I'm not intrigued by the possiblity ... -- Don
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.