Author: Robert Hyatt
Date: 17:56:10 07/09/99
Go up one level in this thread
On July 09, 1999 at 18:40:54, KarinsDad wrote: >On July 09, 1999 at 16:01:08, Robert Hyatt wrote: > >[snip] >> >>Real positions don't matter here... it is "what can the search produce from >>the root to the tips?" that you have to deal with. _lots_ of strange things >>happen in a full-width search. While > 8 queens are not likely, I recently >>discovered a bug in my eval that would break if one side had > 4 queens, and >>they were all on the side of the board away from the kings. And this was in >>a _real_ game position that wasn't made up... > >A search will never be able to find a 9 queen position from a real position. It >doesn't matter how far it searches. > >From a created position, a 9 queen position could be found in a search. > >Having 4 queens is light years away from having 9. > >To get to 9 queens in a real game, one side must not capture any pawns or queens >of the opposing side while at the same time, it must give up at least 4 pawns of >it's own. This is forced (basically, read on). Not only that, but the most >connected pawns that the side that is giving up the minimum of 4 pawns can have >is 2 (for all intents and purposes unless the "winning side" captures a lot of >pawns and pieces with it's pawns to shift the promotion locations). It is real >hard (i.e. impossible) to protect your king when the opposing side is a computer >program and it is starting to acquire more and more queens and the most pawns >that you can have in front of your king is 2 and those pawns cannot be a/b pawns >or g/h pawns. > >Also, out of the 8 additional queens, 2 must be promoted on the a or b file, 2 >must be promoted on the c or d file, 2 must be promoted on the e or f file, and >2 must be promoted on the g or h file. So, it will be impossible for the king to >hide on any of those files. Now, granted, this assumes that the 8 pawns did not >take a bunch of pawns/pieces and move over more than 1 file each. But if that is >done, then there are even fewer pawns/pieces to protect the king that is trying >to hide from this increasing number of enemy queens. > >It's exponential. The more queens one side gets, the fewer pawns/pieces the >enemy has and the higher the probability of mate. The more enemy pieces allowed >to be captured in order to allow the enemy king to have a pawn fortress around >him, the fewer pieces the enemy has to protect. > >And, the more positive the score, the more likely the computer is to trade off >pieces in order to take advantage of that positive score. > >It is extremely unlikely (probably impossible) that you could play against a >program (no matter how you played) and the program (assuming a decent program) >could even get 7 queens (i.e. 6 promotion files out of 8) in it's search before >finding better mating lines. 9 queens is definitely impossible (assuming a >program of at least a decent search capability). > >KarinsDad :) None of that goes with "alpha/beta minimax depth-first". I _must_ search to tip positions before looking at any score... And I'm not willing to bet that searching 30+ plies deeply along selected lines will not find some way to make this happen. Remember minimax doesn't search "logical" lines. It searches +all+ lines. The position I reached with 5 white queens was done instantly in a blitz game. At the starting position both sides had _one_ queen. By the time I did an instant 8 ply search + extensions, etc, and got to the first call to Evaluate() I had 5 white queens, which produced a score way > +infinity and blew me up. I agree that it is very unlikely. But full-width searches often act like "help-mate" searches... one side does everything right, the other side does everything wrong, leading to very bizarre positions. I'd at least cover the case so that it wouldn't blow up. Either by allowing 64 queens as I do now, or by making sure no promotion is allowed if it takes me over my max limit.
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.