Computer Chess Club Archives


Search

Terms

Messages

Subject: Comparison of chess engine EGTB probing

Author: Mike Hood

Date: 18:00:53 02/08/03


Today I set up a position to compare the EGTB probing efficiency of several
Chessbase engines. My PC is a Pentium III 733 MHz and I have all the 3/4/5-piece
EGTBs. For the tests I gave the engines 128 MB hash tables. The position is:

[D] 8/6k1/6P1/p4K2/8/8/4P1P1/8 w - - 0 1

The optimal win is Mate in 25 with minor variations in the moves, for instance

1. Ke5 a4 2. Kd4 a3 3. Kc3 a2 4. Kb2 a1Q+ 5. Kxa1
(Here it enters the KPP-KP EGTB)
5...Kxg6 6. Kb2 Kf5 7. Kb3 Kf4 8. Kc4 Ke4 9. g4 Ke5 10. Kc5 Ke6
11. Kd4 Kf6 12. e4 Kg5 13. e5 Kg6 14. e6 Kf6 15. Kd5 Ke7
16. g5 Ke8 17. g6 Kf8 18. Ke5 Ke8 19. g7 Ke7 20. Kd5 Kf6
21. g8Q Kf5 22. e7 Kf4 23. e8Q Kf3 24. Qe4+ Kf2 25. Qgg2# 1-0

The purpose of the test was not only to find the best first move (Ke5 or Ke4),
but to recognize that this move is Mate in 25. This should be an easy task for
any engine with EGTB probing, but almost impossible for an engine without
probing. For instance, Fritz 5.32 finds the correct line, but after 30 minutes
only gives a score of +5.84

I ran the test with all the engines that use EGTB probing, and I stopped the
test either (1) when the correct result was found or (2) after 5 minutes. Of the
21 engines I tested only 11 found the correct result.

Correct Results:

Engine           Score  Time  Depth    Nodes    tb=  Line
Shredder 6        #25      0  11/22     17kN   1864  1.Ke4 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1B+ 5.Kxa1 Kxg6
BamBam            #25      1  14       760kN  58165  1.Ke4 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1
InmiChess 3.06    #25      1  13        19kN   1145  1.Ke4 a4 2.Kd3 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1
Fritz 7           #25      1  13/26    714kN    484  1.Ke5 a4 2.Kd4 a3
Crafty 19.01      #25      2  10/12     18kN    335  1.Ke5 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1
Hiarcs 7.32       #25      2  11/25    154kN    487  1.Ke5 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1
Shredder 7        #25      2  12/12     27kN   2311  1.Ke5 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1B+ 5.Kxa1 Kxg6
Fritz 8           #25      3  13/27   1258kN   2744  1.Ke4 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1 Kxg6 6.Kb2 Kf5
Goliath Light 1.5 #25      3  12/21    722kN   3078  1.Ke4 a4 2.Kd3 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1
Gromit 3.1        #25      3  11/52     61kN   1404  1.Ke4 Kxg6
Hiarcs 8          #25      5  11/25    601kN   2232  1.Ke4 a4 2.Kd3

All the engines that find the correct answer need less than 5 seconds, even on
my slow computer. Although I can't vouch for the scientific accuracy, it seems
that Shredder 6, Fritz 7 and Hiarcs 7.32 are all faster than their newer
versions.

Incorrect Results:

Engine           Score  Time  Depth    Nodes    tb=  Line
Junior 7          #23     12  21      1788kN  13585  1.Ke4 Kxg6
SOS               #23     54  17/33  11995kN  29437  1.Ke4 Kxg6
Chess Tiger 14    #24      3  13       767kN   1101  1.Ke4 Kxg6
AnMon 5.07        #26      5   9       136kN   1440  1.Ke4 a4
Exchess 4.02      #24      4  11       786kN unknown 1.Ke4 a4 2.Kd3 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1
Patzer 3.11       #24      1  10        54kN   1544  1.Ke5 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 a1R 5.Kxa1
The Crazy Bishop  #29   3:19  17     86975kN unknown 1.Ke4 a4 2.Kd3 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1 Kf6 6.g7 Ke5 7.g8Q Kd6 8.Qe6+
Ikarus 0.18       #30   1:40  16/30  18302kN 105079  1.Ke5 a4 2.Kd4 a3 3.Kc3 a2
4.Kb2 Kf6 5.g7 a1R 6.Kxa1 Ke5 7.g8Q
Comet B55         #75   1:26  16/31  26455kN  25065  1.Ke4 Kf8 2.g7+ Kf7 3.Kd4
Nimzo 8          3.96   1:20  20/40  52331kN   2210  1.Ke4 a4 2.Kd3 a3 3.Kc3 a2
4.Kb2 a1Q+ 5.Kxa1 Kg8 6.Kb2

Junior, Chess Tiger and SOS all make the same error of expecting 1...Kxg6 (Mate
in 23) instead of the best reply 1...a4 (Mate in 25). Chess Tiger makes the
additional error of scoring this line as Mate in 24.

AnMon, Exchess and Patzer find the correct line, but the score is wrong.

The Crazy Bishop finds the correct line as far as the tablebase entry point
5.Kxa1, but chooses an inferior path afterwards.

Ikarus and Comet fail to find the correct line.

Nimzo 8 is the biggest failure. Even though it probes the endgame tablebases, it
doesn't seem to use the probe results correctly, so it works even worse than the
engines with no probing.

I deliberately confined myself to testing Chessbase engines. I wanted to leave
the other forum members to test their favourite Winboard and UCI engines
themselves :)



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.