Author: Robert Hyatt
Date: 15:04:06 08/14/05
Go up one level in this thread
On August 14, 2005 at 16:08:35, Brian Richardson wrote: >On August 14, 2005 at 15:33:00, Robert Hyatt wrote: > >>On August 14, 2005 at 14:57:50, Brian Richardson wrote: >> >>>Just wondering if others see this behavior or not. >>> >>>In the position below, Tinker prefers Qc7 after a bit. >>> >>>[D]5rk/3b2p/1q1p1p1p/2p1p/1pP1P1P/1P1PQP/r2N2KR/7R/ b 0 1 >>> >>>However, in the "flipped" position (horizontally around a rank axis) >>>it does not like Qc2 (I think the flip is correct) >>> >>>[D]7r/R2n2kr/1p1pqp/1Pp1p1p/2P1P/1Q1P1P1P/3B2P/5RK/ w 0 1 >>> >>>When I ran this test with Crafty, and flipping, the moves are also a bit >>>different, as are the scores for each full ply. >>> >>>So, I'm wondering if this is an asymmetrical evaluation problem, or search >>>instability, or no big deal. >>> >>>Thanks >> >> >>Here's the thing for crafty. >> >>It is 100% symmetrical if you just flip the board along either axis. I test >>this all the time. But if you do a search, it might not be, because my move >>generator generates moves in a specific order, and if you flip the board, the >>move order will change (think about FirstOne() applied to a bitboard that is >>rotated, you get a different first or last bit set, etc... >> >>a 1-ply search should be perfect unless there are two moves with the same score, >>and flipping the board causes the generator to produce a different order for the >>two moves in questio... >> >>My "rule of thumb" is _everything_ is a big deal, until properly explained... > >Thanks for the reply. I can understand different search trees based on >different move generation order and such, but shouldn't the final score for each >full ply be exactly the same either way? > >Perhaps you could see if Crafty does this for you, or not, in which case I'm >just missing something. > >Thanks again. To test in crafty for yourself, set up a position, type "score". The type "flip" or "flop" which rotates the board along the x or y axis. flip also changes white piece to black and vice versa. flop does not, but it interchanges a and h files, etc... Note that if you search the same basic tree, but change the ordering, the hash table can further alter the tree and even the final score. Fine 70 is a great example where minor changes produce major differences at a given depth.
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.