Author: Dann Corbit
Date: 10:58:48 03/07/06
Go up one level in this thread
On March 07, 2006 at 04:39:52, Uri Blass wrote: >On March 07, 2006 at 03:47:06, Dann Corbit wrote: > >>On March 07, 2006 at 03:02:17, Uri Blass wrote: >> >>>On March 07, 2006 at 00:46:27, Dann Corbit wrote: >>> >>>>On March 07, 2006 at 00:41:55, Dann Corbit wrote: >>>> >>>>>On March 07, 2006 at 00:34:48, Stuart Cracraft wrote: >>>>> >>>>>>On March 07, 2006 at 00:31:45, Dann Corbit wrote: >>>>>> >>>>>>>On March 07, 2006 at 00:27:43, Stuart Cracraft wrote: >>>>>>>[snip] >>>>>>>>Very interesting indeed. A clever test. >>>>>>>> >>>>>>>>If one's results do not rotate approximately as described >>>>>>>>for the four positions and you say the evaluation is an >>>>>>>>issue, what kinds of evaluation issues have you seen that >>>>>>>>could explain it?!? >>>>>>> >>>>>>>The most common thing that I see is something that is good for white being >>>>>>>counted as positive for black also on the evaluation. Often, when we are >>>>>>>writing the eval, we are thinking from the perspective of white. And so if we >>>>>>>are not very careful, we may invert the sign of some evaluation component and >>>>>>>count something that is good for white as something that is good for black (or >>>>>>>vice versa, though the reverse is seen less often for some reason). >>>>>>> >>>>>>>There are, of course, many other possible causes besides that. >>>>>> >>>>>>A good point. I try to avoid that by always doing things from the >>>>>>side on move, almost always. There are a few in there however with >>>>>>respect to white and black specifically, but they are then folded >>>>>>together with the stm variable and stm^1 which translate to white/black >>>>>>or black/white depending on who's on move. I could try this: rerun >>>>>>your rotation test with successively less in the evaluation table >>>>>>until nothing but material and see what happens. >>>>> >>>>>Right. If you have divided off the eval components, you could binary search >>>>>until you find the problem component. >>>>> >>>>>Now, we do not know for sure that it is an eval sign problem. However, the fact >>>>>that the records are similar in pairs makes it very suspicious. >>>> >>>>I guess that when you have gotten your eval symmetrical, you will miss less than >>>>ten problems on WAC. >>> >>>I think that you are wrong here. >>>Stuart may have evaluation bugs but his main problem is the search. >> >>I think it likely that it is both. >> >>Given: >>5rk1/2p4p/2p4r/3P4/4p1b1/1Q2NqPp/PP3P1K/R4R2 b - - bm Qg2+; id "-rotXTDa.1"; >>1kr5/p4p2/r4p2/4P3/1b1p4/pPqN2Q1/K1P3PP/2R4R b - - bm Qb2+; id "-rotXTDg.1"; >>r4r2/pp3p1k/1q2nQpP/4P1B1/3p4/2P4R/2P4P/5RK1 w - - bm Qg7+; id "-rotXTDc.8"; >>2r4r/k1p3pp/PpQn2q1/1B1P4/4p3/R4P2/P4P2/1KR5 w - - bm Qb7+; id "-rotXTDe.8"; >> >>When I changed to material only eval, here is the result: >> >>st 5 >>ts >>position file? [wac.epd] rot.epd >># of test positions to test? 4 >>maxtime = 500 >>Interrupt current ply and return move at timeout >>Testsuite: rot.epd 4 positions >>*** Problem Solution(s): Qg2+ (bm) >>[D] 5rk1/2p4p/2p4r/3P4/4p1b1/1Q2NqPp/PP3P1K/R4R2 b - - bm Qg2+ >>*** Problem Solution(s): Qg2+ (bm) >>-- ** -- ** -- BR BK ** >>** -- BP -- ** -- ** BP >>-- ** BP ** -- ** -- BR >>** -- ** WP ** -- ** -- >>-- ** -- ** BP ** BB ** >>** WQ ** -- WN BQ WP BP >>WP WP -- ** -- WP -- WK >>WR -- ** -- ** WR ** -- >>mv 1 stage 0, black to move, computer plays black >>hash=62305c813f5fad4 >>pawnhash=3da7edf6c1ba87ea >>0 0 0 0 0 0 >>Alpha=-400 Beta=400 Maxdepth=9999999 MaxTime=500 xboard=1 >>Ply Score Time Nodes PV >>1. 40 3 12 c6d5 e3d5 >>1. 900 5 74 f3g2 e3g2 > >Qg2+ is a sacrifice so it is not logical so if the computer choose it at depth 1 >then it means that there is a serious bug. It may find the result by quiesce(). I think you are probably right about search problems also. However, having errant terms in his eval, and having search stability problems, and yet still solving 3/4 of WAC, I suspect he will make rapid progress once he irons out some simple details.
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.