Author: Dave Gomboc
Date: 03:49:49 06/24/98
Go up one level in this thread
On June 24, 1998 at 00:07:42, Bruce Moreland wrote: > >On June 23, 1998 at 18:25:49, Amir Ban wrote: > >>The critical line is: >> >>1.Rxe6 fxe6 2.Nxe6 Qa5 3.Bxd7+ Kxd7! 4.Nc5+ Ke8 5.Qd7+ Kf7 6.Nxb7 Rc7 7.Nd6+ Kg7 >>8.Ne8+ Rxe8 9.Qxe8 Nf6! >> >>The rook sac is not at all speculative. It's just wrong. Your eval should be a >>significant plus for white at at least two points in it, but this doesn't hold >>because black has these counter-resources. Black has only one defense, but it >>works. The hard part for black is seeing 3...Kxd7. 3...Kf7 is a dead loss. > >All times on a Pentium Pro 200. > >What happens when I run this from the root is I get Bh6 in under a second, at >about -0.8, holding for about 3.5 minutes, until in ply 10 Rxe6 fails high, and >the re-search fails low. > >I just ran this one for five minutes, so I don't know what happens later. > >If I force Rxe6, and search on the resulting position (I didn't think to do this >earlier), the program wants to play fxe6 (about +1) for about a minute, then >something bad happens, it goes down to -0.6, and a few seconds finds Rc7, which >is about zero, and which it wants to play until a little before six minutes. > >Then it fails high on fxe6 again in ply 10, with a score of +1.7, and it goes up >from there. > >>>My own analysis shows a nasty check or two, then black is simply up material. >>> >> >>No, there's more to it as you can see in the line. I don't understand how your >>program can dismiss this line without passing through intermediate depths where >>it would think it works. > >I think I lucked out a little bit. > >bruce You can call it luck, but I'm not sure it is luck. I remember during the prototype Rebel-Crafty NPS game (not match, as it turned out :) that there was some discussion over whether solving something positionally was really "solving it", because if some weights were tinkered with the software might no longer solve the position correctly. There are a lot of positions that I "solve" at the board, not because I see all the tactics in the position, but because I have a reasonable idea of what is likely to be the better alternative, based upon some chess criteria that isn't specific to the position being analyzed. For instance, it could be that a move does a better job than others at preventing defenders from returning to the area of attack. It is clear that a tactical analysis of the entire situation is going to guarantee the best assessment possible, but if software has the right rules of thumb (and that includes when the rule is applicable!) I think that the software will do much better than the equivalent software without such small bonuses and penalties. Both versions will be able to calculate the knockout blow, but I suspect that the latter version will be finding that blow for the opponent player more often than for itself, if you get my meaning. The situation with the actual problem being discussed above is a little different, but my thinking is along the same lines. Something in your search/eval causes your program to like Bh6 more than Re6, and it has to do with what positions you reach when searching the subtrees, and how you assess them. Other programs don't share Ferret's preference. Your program is making the correct decision here, and it could be just luck, but it probably isn't. Dave Gomboc
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.