Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Debugging hash tables & fingerprinting versions

Author: Charles Roberson

Date: 09:26:32 03/13/04

Go up one level in this thread



  In a 15 0 game, you should be able to search better than 6 ply without
  transposition tables. I suggest dropping transposition tables and
  figuring out the problem.

  Your problem could be the timing algorithm !!

  I suggest testing your program out on a small benchmark test.
  I used the same six positions as crafty to use as a comparison.
  It is simple to setup a benchmark procedure assuming you have the
  ability to parse fen strings. Such a benchmark proves excellent for quick
  tests of changes/fixes.

  One of the nice things about the procedure is I run the benchmark to
  fixed depths (not time). This isolates the timing code from the search.
  It prints out the chosen moves for each position and stats for the total
  six positions (nonquiesce nodes, quiesce nodes, time, hash table hits,
  pawn hash hits....).

   With this information, you will be able to zero in closer on the real
   problem.

   The following is sample output from noonianchess after a benchmark run.
   Just for reference I've added the fics info for the same version.

(2/27/2000)
NC 2.8 perfomance numbers on bench with fixed depths for comparisons.
     (All tests are on 300Mhz Pentium II with 67 Mhz motherboard)
Depth  Time(s)     Nodes     LeafNodes    TotNodes     T/6(s)     TN/6   EBF
  3
  4      1.00      36,944        74,988      111,932    0.167     18,655  --
  5      3.00     105,756       205,085      310,841    0.5       51,807  2.78
  6      8.00     461,411       770,918    1,232,329    1.33     205,388  3.96
  7     32.00   1,862,531     2,942,007    4,804,538    5.33     800,756  3.90
  8    207.00  11,250,076    18,566,478   29,816,554   34.5    4,969,426  6.21
  9    751.00  39,864,355    74,131,319  113,995,674  125.167 18,999,279  3.90
 10   2794.00 164,822,825   247,551,761  412,374,586  465.67  68,729,098  3.62
 11  11944.00 675,064,407 1,062,372,517 1737,436,924 1990.67 289,572,821  4.21

  Time = time in secs to complete all six benchmarks to the specified ply.
  Nodes= Nodes hit in search procedure not counting the quiesce nodes
  LeafNodes = Nodes hit in the quiesce search
  T/6  = Time / 6 = avg time per position
  TN/6 = Total Nodes / 6
  EBF = effective branch factor based on time = time(ply)/(time(ply-1))


                   Lightning      Blitz     Standard
    FICS rating      1825          1818       1893
    Games             103           448         26
    Peak             1825          1914       1910
    Date             2/29/2000    4/5/1999    2/29/2000


   For comparison, I modified crafty to produce the same info.
   (FYI: I have noticed that this amount of data from the benchmark creates
      a nice finger print to identify a version of the code).

(2/27/2000)
  Crafty perfomance numbers on bench with fixed depths for comparisons.
     (All tests are on 300Mhz Pentium II)
 6xDepth    Time(s)      Nodes    MultNodes    Time/6    Nodes/6       EBF
    1                      336                                56.0     ---
    2                    2,584                               430.67    7.69
    3        0.24       11,695                   0.04      1,949.16    4.53
    4        0.56       35,541    3.04           0.09      5,923.5     3.04
    5        2.32      174,971    4.92           0.39     29,161.8     4.92
    6        6.35      504,658    2.88           1.06     84,109.67    2.88
    7       12.23    1,017,208    2.02           2.04    169,534.67    2.02
    8       27.94    2,384,657    2.34           4.66    397,442.83    2.34
    9       78.73    6,866,830    2.88          13.12  1,144,471.7     2.88
   10      249.37   21,415,181    3.12          41.56  3,569,196.8     3.12





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.