Author: Dieter Buerssner
Date: 05:41:25 09/25/01
I made a small, not very scientific experiment, to get a feeling
for the significance of hash collisions.
For this, I tested 22 positions of WAC (the ones, that I would call the
most difficult ones). I tested those for 2 million nodes. Yace will use
HTs in quiescence search, so the table gets filled very fast (on average
almost 80% of the searched nodes get stored). I used only a small HT
for this test - 1e6 bytes, which is about 80000 entries - to make sure,
that the table really gets overloaded. I made 4 runs of the test positions:
with a 32/16/12 and 8-bit hash-values (the hash index is calculated
independently of these). Here are the results. I hope, it is more or
less self explaining. The number of hash errors given, is just the
ones that were detected by checking for the legality of the move. This is
only done, when the hash info would give a cutoff, or an adjustment of
the search window. When the move is not legal, no cutoff/adjustement of search
bound is allowed.
TEST test.ci by Yace 0.99.56 32-bit hash search_nodes 2000000 Tue Sep 25
14:14:58 2001
table_add=30392714, table_store=30392714, table_probe=38930034, 29.22% found
hash_errors 0, egtb_probe=0, egtb_found=0
22 tested, 19 found, 3 not found, mates 4, time 118.71
test nodes 17047943 win nodes 11047915 mate nodes 2601918
win time 78.65 mate time 17.41 av depth 8.045 (nm 8.222) maxdepth 32 tu 744
TEST test.ci by Yace 0.99.56 16-bit hash search_nodes 2000000 Tue Sep 25
14:21:48 2001
table_add=30433572, table_store=30433572, table_probe=38892887, 28.97% found
hash_errors 22, egtb_probe=0, egtb_found=0
22 tested, 19 found, 3 not found, mates 4, time 116.31
test nodes 16872328 win nodes 10872318 mate nodes 2669471
win time 76.86 mate time 17.11 av depth 8.045 (nm 8.222) maxdepth 32 tu 748
TEST test.ci by Yace 0.99.56 12-bit hash search_nodes 2000000 Tue Sep 25
14:29:01 2001
table_add=30559112, table_store=30559112, table_probe=38955354, 28.70% found
hash_errors 441, egtb_probe=0, egtb_found=0
22 tested, 19 found, 3 not found, mates 4, time 122.58
test nodes 17320105 win nodes 11320088 mate nodes 2614227
win time 82.12 mate time 17.71 av depth 8.045 (nm 8.222) maxdepth 34 tu 777
TEST test.ci by Yace 0.99.56 8-bit hash search_nodes 2000000 Tue Sep 25
14:36:32 2001
table_add=32512573, table_store=32512573, table_probe=40699069, 26.80% found
hash_errors 7128, egtb_probe=0, egtb_found=0
22 tested, 17 found, 5 not found, mates 3, time 167.31
test nodes 23403260 win nodes 13403242 mate nodes 2620719
win time 99.20 mate time 17.77 av depth 7.864 (nm 7.947) maxdepth 32 tu 734
Only in this last run, the search got very confused at times. One example:
161407 1.228 0.39 5t+ 1.Nfg5 fxg5 2.Nf6 Ba3 3.bxa3 Rb7 4.axb5 axb5
{-80}
167981 1.300 0.93 5++ 1.Nfg5 bxa4 2.Nxh7 Rxd4 3.Nxf8+ Kg8 {330}
170673 1.333 1.93 5++ 1.Nfg5 Bc5 2.dxc5 Rd4 {260}
172307 1.361 6.93 5++ 1.Nfg5 Rd5 2.Nxh7 Nd7 3.Nxf8+ Kg8 4.Nxd7 Bxd7H
5.axb5H {570}
179777 1.405 Mat04 5t 1.Nfg5 fxg5 2.Nf6 Bxf6 3.Be4 Kg8 4.Qxh7#H {-690}
182542 1.428 Mat04 5. 1.Nfg5 fxg5 2.Nf6 Bxf6 3.Be4 Kg8 4.Qxh7# {-690}
183483 1.433 Mat04 6t 1.Nfg5 fxg5 2.Nf6 Bxf6 3.Be4 Kg8 4.Qxh7# {-690}
201820 1.493 Mat03 6t+ 1.Nxf6 Bxf6 2.Ng5 Bxg5 3.Qxf8+ {-190}
203865 1.505 Mat04 6t- 1.Nxf6 Bxf6 2.Ng5 Bxg5 3.Qxf8+ Qxf8 4.Bxc6
{-1111}
203865 1.505 Mat04 6t 1.Nfg5 fxg5 2.Nf6 Bxf6 3.Be4 Kg8 4.Qxh7# {-690}
227459 1.595 Mat04 6. 1.Nfg5 fxg5 2.Nf6 Bxf6 3.Be4 Kg8 4.Qxh7# {-690}
239645 1.648 14.24 7-- 1.Nfg5 fxg5 2.Qxe6
263834 1.745 13.24 7-- 1.Nfg5 fxg5 2.Qxf8+ Bxf8 3.Nxg5
268473 1.770 8.24 7-- 1.Nfg5 fxg5 2.Qxf8+ Bxf8 3.Nxg5
1152818 9.489 -2.45 7t 1.Nfg5 fxg5 2.axb5 Rg8 3.bxa6 Nxa6 4.e3 Rg6
5.Qh5 Kg8 {HT} {-350}
1309369 10.801 -2.44 7t+ 1.Nc5 bxa4
1724535 13.784 0.43 7t 1.Nc5 Bxc5 2.dxc5 Rd5 3.axb5 cxb5 4.b4 Nc6
5.Rfd1 {-80}
1978437 15.616 0.44 7t+ 1.axb5 cxb5 2.Nfg5 Rd5 3.Nxh7 Bb7 4.Nxf8+ Kg8
{490}
2000006 15.750 0.44 7t 1.axb5 cxb5 2.Nfg5 Rd5 3.Nxh7 Bb7 4.Nxf8+ Kg8
{490}
2000006 15.758 0.44 7u. 1.axb5 cxb5 2.Nfg5 Rd5 3.Nxh7 Bb7 4.Nxf8+ Kg8
{490}
Regards,
Dieter
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.