Author: Komputer Korner
Date: 13:22:44 02/01/99
Go up one level in this thread
On February 01, 1999 at 15:19:20, Dann Corbit wrote: >A while back, someone posted Hiarcs 7 epd analysis that showed acn counts and >also acd as 'dep'(plies forward). > >How do you set up Hiarcs 7 to produce this wonderful analysis. Inquiring minds >want to know. > >I want to know also. The following will help. -- Komputer Korner hyatt@crafty.cis.uab.edu (Robert Hyatt) wrote: > > Richard A. Fowell (fowell@netcom.com) wrote: > : <snip> > > : > > : >RAF> nodes/ AMD K6 Pentium P5 AMD/P5 > : >RAF> second 200 Mhz 133 Mhz *133/200 > : >RAF> ========================================= > : >RAF> average 21434 8983 1.58 > : > > : >Hm, AFAIK Hiarcs doesn't show the nps, so how do you get these results? > : > > : > > : >Harald Faber > > : When you analyze/process an EPD test file in HIARCS 6.0, > : it will insert > : "acs" (analysis time, seconds) and > : "acn" (analysis count, nodes ) in the processed EPD file. > > : ** HIARCS tip *** > : In order for HIARCS to compute the "acs" and "acn" tags, it appears > : that the input file must includ "bm" (best move) tags. > > : It also seems that HIARCS halts analysis as soon as the PV begins with > : the "bm" key move. > : fowell@netcom.com (Richard A. Fowell) > > I would hope this is wrong. Me too, but I doubt it. See below. > Otherwise the hiarcs results are invalid, True. > because many programs will, on occasion, like a "key" move for a ply or > two then change to something else. Seen it many times. > > I have this as an option in crafty, in the test command, as follows: > > test <testfile> <early-exit-count> > > I typically use an early-exit-count of "2" which says that the solution > is correct when two consecutive search iterations stick with the same move > that is in the solution set. For results I report here, I omit this, > which disables the early exit. Many positions will fool a program that > quits when it finds the key move, and this will sometimes greatly inflate > the test result score, falsely... > Well, I agree it is unfortunate, but I think it is true. Perhaps some other HIARCS owners can confirm/deny. My experiences here went like this: 1) Set up 33-problem test set at 270 seconds/problem - noticed HIARCS 6.0 PC programs often took *way* less than 270 seconds, and that these cases corresponded to PVs that began with the "bm" move. 2) (wanting 270 seconds of run time) Deleted bm tags - found that this made the "acs" "acn" output disappear too - couldn't compute nodes/second. 3) (still wanting 270 seconds of run time) Changed moves in "bm" tags to the the absolute worst moves I could find in each position, figuring that HIARCS would never think that these were best, hence would run 270 seconds on everything. 4) Found that this worked in all of my positions but this one: (all machines now take 270 seconds an all problems except this one: 8/1p3pp1/7p/5P1P/2k3P1/8/2K2P2/8 w - - bm Kb1; id "F.01"; "Kb1" was the worst move I could think of here. The actual solution for this problem is "f6". When given "bm f6;" HIARCS 6.0 PC finds this solution at ply 9: ce -190; pv f6 gxf6 f4 Kd4 Kb3 Ke4 g5 ; dep 9; acs 15; acn 187965; But when told the false "Kb1" as best move, HIARCS 6.0 PC exits at ply 7: ce -146; pv Kb1 Kd4 g5 f6 gxf6 gxf6 Kc1 ; dep 7; acs 2; acn 27054; I don't know if it has your "2 ply" check or not, but it definitely gets the wrong answer here. I agree that this could easily lead it to be scored too highly on problem tests. ====================== My current workaround: ====================== Run each test twice: once with "bm" tags, and once without. For problems where it finds the "bm", double check the non-bm output to make sure it has the same solution at the end. If not, score that problem as failed. This still won't catch case's where it finds/loses/finds the solution, but in my experience those are at least rare. (Since the last version of HIARCS I had wouldn't let me compute either solution times or nodes/second, I view the situation as an improvement, if not ideal.) The HIARCS folks often lurk here, but I will forward this message to them. fowell@netcom.com (Richard A. Fowell) # Visit the Macintosh Chess FAQ Homepage at: # # http://www.geocities.com/SiliconValley/Pines/6827/ #
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.