Author: martin fierz
Date: 18:40:53 09/08/02
Go up one level in this thread
On September 08, 2002 at 17:43:24, Gerd Isenberg wrote: >On September 08, 2002 at 15:41:50, martin fierz wrote: > >>On September 07, 2002 at 19:46:49, Peter McKenzie wrote: >> >>>I've been having a break from computer chess post WCCC, but have now started >>>going over some of Warp's games. First up is Warp's worst game, the loss vs >>>Brutus. In this game, Warp showed a total lack of understanding of its >>>centralised king in the early middlegame and lost without a fight. >>> >>>I present here the position after move 21 in the game. White has grabbed a pawn >>>thinking this position is OK, but in reality the white king is hopelessly stuck >>>in the centre. Also, white is passive, black is active and has a safe king, >>>therefore white is totally winning here: >>> >>>[D]r1r3k1/1p3ppp/b5q1/p7/4n3/P3PNB1/1P3PPP/1Q1RK2R b K - 0 1 >>> >>>I am curious what the static evaluation of various programs is here. The >>>version of warp used in Maastricht gives 0.238 in favour of white. Ideally the >>>static evaluation should favour black here I think. >>> >>>cheers, >>>Peter >> >>the only important thing to recognize is that >>Ba6 prevents castling && lots of material (queens!) on board. > >most important? if you like, yes. to me as a chess player it is the first thing i see when i look at the board, and the only relevant feature in this position. of course, you can say that the black king is safe, which is naturally also important, but gets blocked out of my mind when i look at this :-) >>i think you can easily give this kind of feature a value of 1 pawn. >>however, you have to be kind of certain that it is here to stay. which >>means that you should make sure that there is no move like c2-c4 before >>applying this kind of rule. gerd gave a nice description of what his eval >>is doing here, but i think part of it is no good: his count of "There are four >>black pieces controlling squares with distance two from king." is probably no >>useful measure. e.g. place the black bishop on b6 instead of a6, and you still >>have that feature, but now it is black who has to hope for a draw. > >The order of the statements was rather random and doesn't imply any importance >or weights. The most important is the (long time) lack of doing a castle or a >indirect castle via Kf2,Re1,Kg1. That in conjunction with the rather missing >ability to move a pawn or light piece to the diagonal. > >"There are four black pieces controlling squares with distance two from king" is >a light weight, but is simply a heuristic that may improve the probability of >"correctness". sure, i understand that. but it's a very general heuristic and i'd be surprised if it's really useful. i generally find that making eval terms more specific helps my checkers program a lot. it would be interesting if you gave the weights of these eval terms too. >>i know what you are worried about when not giving the "Ba6-prevents-castling & >>queens on board" feature a higher eval. it might go away again, if white could >>play something like Be2 or Ne2 and then castle. i have often thought about >>similar things in my checkers program. in the end, i usually put high eval terms >>for this kind of feature in anyway, because i think that if white can free >>himself from this type of bind, the search will find it. however, if he cannot, >>then it is very important that the eval tells you that he is in trouble. of >>course i can search much deeper in checkers, which helps me believe that the >>search will find it... >> > >yes, but sometimes it's a decisive leaf-position :-) of course... i know the feeling :-) but the Ba6 preventing castling (or more commonly Ba3) is a very common theme specially in open games, where white often sacs an exchange for it. i'd rather have my engine play a few more spectacular games and also lose a game or two because of this than not know about it. of course, i don't have a chess engine so it's easy to say that :-) aloha martin > >cheers, >Gerd > >>aloha >> martin
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.