Author: Uri Blass
Date: 21:17:54 12/22/00
Go up one level in this thread
On December 22, 2000 at 22:40:27, Robert Hyatt wrote: >On December 22, 2000 at 15:52:19, Ernst A. Heinz wrote: > >>>On December 22, 2000 at 14:45:03, Ernst A. Heinz wrote: >>> >>>>>Did you look at what happens if you play a4? Your kingside gets totally >>>>>shredded, starting with Qxf7+... >>>> >>>>Yes, true -- but then Black's King is not standing there! >>>> >>>>It rather seems quite safe on d8 while White's King gets >>>>into mighty trouble on d1 after the sequence a4 - Qxf7+ - >>>>Kd8 - Qxg7: >>>> >>>>[d]2rk3r/6Q1/4p3/1pqpP3/p3b1PP/1B6/PPP2R1R/2K5 b - - 0 1 >>>> >>>>"DarkThought" quickly scores this as substantially positive >>>>for Black, locking onto Qe3+ as the best move almost instantly. >>>> >>>>=Ernst= >>> >>>I have no doubt that black might actually be winning here. But that wasn't >>>the point. You said you picked a4 at depth=10. For that to happen, you have >>>to ignore king safety and sit in the middle of the board, with a queen at f7, >>>a rook on the open file, the king rook hanging, the king can't move to connect >>>the rooks, etc. >>> >>>IE at 10 ply it is all judgement, not "truth" as the search probably can't see >>>the final outcome. >> >>"DarkThought" gets a fail-low on 0-0 in iteration #10 with the >>score dropping from +1.1 in iteration #9 to just +0.39. Then, >>it locks onto a4 and shows the correct PV for it already in >>iteration #10 (with the white King being driven to d1 where it >>is similarly exposed as the black King on d8). >> >>So, why should that mean to ignore king safety or all the other >>evaluation stuff you are talking about above? >> >>To me, it simply looks like our search picks up the crucial >>threats earlier (maybe, by means of more appropriate extensions >>in this particular case) -- no more, no less. >> >>=Ernst= > > >Note that I am not saying your eval is bad, your search is bad, nor am I saying >mine is right. I am simply saying that in _this_ position, your program is >probably not seeing the "truth" any more than any of the other programs that >have been tested on this position see it. It appears, to me, that it becomes >a matter of how the evaluation is tuned. IE you fail low on castling at depth >10. Crafty likes Rc7 until depth=12 or so where it fails _high_ on castling. I can add that Crafty can see nothing wrong with 0-0 even at depth 15 after hours. I also cannot see something wrong after 0-0 with other programs. a4 seems to be even better than 0-0 but I do not see a reason that 0-0 should fail low. Uri
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.