As a notorious SF-tester (private) i noticed that since about 2 months the newest SF-dev.-versions have problems with the famous "Kasparov-Topalov"-game from 1999.
In the position below SF (Timestamp: 1390645772 from january 25.) "realized" quite quickly that after Re7 white is at least not worse
[D] b2r3r/k4p1p/p2q1np1/Np1P4/3p1Q2/P4PPB/1PP4P/1K2R3 w - - 0 2
bm Re7
Newest SF-versions getting the clue much slower (b4-move)
It seems to me that this began with the dev-version from 26.jan 2014 and is getting slightly worse until today.
Also the following position was "solved" about twice as fast by the version of jan 25. mentioned above compared to nearly all versions that followed up since then:
[D] 1rbq1r1k/pp2n1bp/3p1p2/2p1pPP1/P3B2Q/2NP3N/1P5P/n1B2RK1 b - - 0 16
bm Be6
To me these problems seem to be serious honestly and i hope this can help the further developement a little tiny bit
Deep respect for the SF-Team, i love you guys
tested on winXP32 on a core2duo e6600 with th=2 and 256mb hash
regards from germany
possible SF regression (testposition) ?
Moderators: hgm, Rebel, chrisw
-
- Posts: 416
- Joined: Thu Aug 02, 2012 7:48 pm
- Location: Germany
possible SF regression (testposition) ?
Wahrheiten sind Illusionen von denen wir aber vergessen haben dass sie welche sind.
-
- Posts: 6442
- Joined: Tue Jan 09, 2007 12:31 am
- Location: PA USA
- Full name: Louis Zulli
Re: possible SF regression (testposition) ?
Based on the dates, what you observe is likely related to changes in null-move pruning.
-
- Posts: 416
- Joined: Thu Aug 02, 2012 7:48 pm
- Location: Germany
Re: possible SF regression (testposition) ?
Thanks for your reply Louis
Yes, thats right...there are much more positions (mainly midgame-positions) which are handled much better by SF-versions before 26. january !
Yes, thats right...there are much more positions (mainly midgame-positions) which are handled much better by SF-versions before 26. january !
Wahrheiten sind Illusionen von denen wir aber vergessen haben dass sie welche sind.
-
- Posts: 6442
- Joined: Tue Jan 09, 2007 12:31 am
- Location: PA USA
- Full name: Louis Zulli
Re: possible SF regression (testposition) ?
Here's what the latest SF does, using just 1 thread:Spliffjiffer wrote:As a notorious SF-tester (private) i noticed that since about 2 months the newest SF-dev.-versions have problems with the famous "Kasparov-Topalov"-game from 1999.
In the position below SF (Timestamp: 1390645772 from january 25.) "realized" quite quickly that after Re7 white is at least not worse
[D] b2r3r/k4p1p/p2q1np1/Np1P4/3p1Q2/P4PPB/1PP4P/1K2R3 w - - 0 2
bm Re7
Newest SF-versions getting the clue much slower (b4-move)
Code: Select all
Searching: b2r3r/k4p1p/p2q1np1/Np1P4/3p1Q2/P4PPB/1PP4P/1K2R3 w - - 0 2
infinite: 1 ponder: 0 time: 0 increment: 0 moves to go: 0
1 -1.92 00:00 94 Qxd4+ Kb8
2 -1.82 00:00 361 Qxd4+ Kb8 c4 bxc4 Nxc4
3 -1.96 00:00 754 Qxd4+ Kb8 c4 Rhe8
4 -2.31 00:00 1513 Qxd4+ Kb8 c4 Rhe8 Rc1
5 -1.99 00:00 2604 Qxd4+ Kb8 c4 Rhe8 Rc1 Nd7 cxb5 axb5
6 -1.99 00:00 3266 Qxd4+ Kb8 c4 Rhe8 Rc1 Nd7 cxb5 axb5
7 -1.99 00:00 4708 Qxd4+ Kb8 c4 Rhe8 Rc1 Nd7 cxb5 axb5
8 -1.47 00:00 12201 Qxd4+ Qb6 Re7+ Nd7 Rxd7+ Rxd7 Qxh8 Rxd5 Nb3 Rd1+ Ka2
Bxf3 Qxh7
9 -0.74 00:00 22916 Qxd4+ Qb6 Re7+ Nd7 Qc3 f5 Nc6+ Bxc6 dxc6 b4 axb4
Qg1+ Ka2
10 -1.59 00:00 68852 Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ b3 f5 Rxh7 Kb6
Rxh8 Rxh8 Qxh8 Kxa5 Qh4 Bxf3
11 -2.21 00:00 141984 Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ b3 Kb6 Rxd7
Qc5 Qxc5+ Kxc5 Nb7+ Bxb7 Rxb7 Rd2 b4+ Kc6 Rxf7 Rxc2+
Kb3 Rxh2
12 -2.03 00:00 155626 Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ b3 Kb6 Rxd7
Qc5 Qxc5+ Kxc5 Nb7+ Bxb7 Rxb7 Rd2 b4+ Kc6 Rxf7 Rxc2+
Kb3 Rxh2 Bd7+
13 -2.04 00:00 205695 Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ b3 Kb6 Rxd7
Qc5 Qxc5+ Kxc5 Nb7+ Bxb7 Rxb7 Rd2 b4+ Kc6 Rxf7 Rxc2+
Kb3 Rxh2 Bd7+ Kd6 Bg4
14 -2.19 00:00 354074 Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ Nb3 f5 Qc7+
Ka8 Rxd7 Rxd7 Qxd7 Qd1 Qc7 Qxf3 Qb6 Bxb3+ cxb3 Qb7
Qf6 Rb8
15 -2.25 00:00 476682 Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ Nb3 f5 Qc7+
Ka8 Rxd7 Rxd7 Qxd7 Qd1 Qc7 Qxf3 Qb6 Bxb3+ cxb3 Qb7
Qf6 Rc8 g4 Rc6 Qe5
16 -2.53 00:00 1103K Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ b3 Kb6 Rxd7
Qc5 Rb7+ Bxb7 Qxc5+ Kxc5 Nxb7+ Kc6 Nxd8+ Rxd8 Kb2
Rd2 Bc8 a5 h4
17 -2.73 00:01 1672K Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ Nb3 f5 Qc7+
Ka8 Rxd7 Rxd7 Qxd7 Qd1 Qc7 Qxf3 Qd6 Bb7 Qe7 Qd5 Bf1
Qh1 Bd3 Qxh2 Nc5 Qh1
18 -2.34 00:01 3200K Qxd4+ Qb6 Re7+ Nd7 Qc3 Qg1+ Ka2 Bxd5+ b3 Kb6 Rxd7
Qc5 Nc4+ bxc4 Qf6+ Kb5 a4+ Ka5 Rxd8 cxb3+ cxb3 Qc2+
Ka1 Qd1+ Kb2 Qxb3+ Kc1 Qc4+ Kd2 Rxd8 Qxd8+ Kxa4
19 +0.00 00:02 3969K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ b4 axb4+ Qxb4 Qc7+ Ka4 b3+
Ka3 Qe5 Qc5 Bf1 Rc8 Qa1+ Kb4 Qa4+ Kc3 Qa1+ Kb4
20 +0.00 00:02 4531K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6 Rd7
Rxd7 Bxd7 Qg1+ Ka2 Qc5 dxc6 Qc4+ Kb1 Qf1+ Ka2 Qc4+
21 -1.33 00:04 8794K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 f5 c7 Rc8 Re6+ Kxc7 Rxa6 Rhe8 c3 Kb7
Rd6 Re2 f4 Rxh2 Bf1 Rh1 Rd1 Kb6 Kc2 Rh2+ Kb3 Re8
22 -1.38 00:05 10969K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 f5 c7 Rc8 c4 Rxc7 Rxc7 Kxc7 cxb5
axb5 Bf1 Kc6 Kc2 Kc5 f4 b4 Bg2 bxa3 bxa3 Re8 Kd3
Rd8+ Ke3 Rb8
23 -1.30 00:06 13330K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 f5 c7 Rc8 c4 Rxc7 Rxc7 Kxc7 cxb5
axb5 Kc2 Re8 Kd2 Kb6 Bf1 Kc5 Bd3 b4 a4 Rd8 Ke3 Ra8
Bb5
24 -1.32 00:06 14233K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 f5 c7 Rc8 c4 Rxc7 Rxc7 Kxc7 cxb5
axb5 Kc2 Re8 Kd2 Rd8+ Kc2 Kb6 Bf1 Kc5 Bd3 Rd4 Kc3
b4+ axb4+ Rxb4 b3 Rb7 g4 fxg4 fxg4
25 -1.53 00:12 27345K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 f5 c7 Rc8 c3 Rxc7 Re6+ Rc6 Re2 Rd8
Kc2 Rcd6 f4 R8d7 Bf1 Rd1 Bh3 R7d6 Bg2 R1d3 Kc1 a5
Kc2 a4 Bf1 h6
26 -1.58 00:18 40528K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 f5 c7 Rc8 c3 Rxc7 Re6+ Rc6 Re2 Rd8
Kc2 Rcd6 f4 R8d7 Bf1 Rd1 Bh3 a5 Bg2 R1d6 Bf1 a4 Bg2
27 -1.78 00:34 76217K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 Rd1+ Ka2 f5 c7 Rc8 Rxh7 Rxc7 Rh6 Rc6
c3 Rd2 g4 Rxc3 Rxg6+ Kc5 Rg5 Rxf3 Rxf5+ Rxf5 gxf5
Rxh2 Bg4 Rf2 Kb3 Kd4
28 -1.68 00:41 90818K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 Rd1+ Ka2 f5 g4 fxg4 Bxg4 Kxc6 h4 Rd2
Kb3 h5 Be6 Rf2 Rg7 Rxf3+ c3 Rh6 Rf7 Rxf7 Bxf7 Kc5
Kc2 a5
29 -1.99 01:09 152874K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 Rd1+ Ka2 f5 g4 fxg4 Bxg4 Kxc6 f4 Rd2
Bf3+ Kd6 Ra7 Rxh2 Rxa6+ Ke7 Be4 Rf8 Bd3 Rxf4 Kb3 g5
Bxh7 g4 Bd3 g3 Rg6 g2 Bxb5
30 -2.02 01:30 198938K Re7+ Kb6 Qxd4+ Kxa5 Qc3+ Kb6 Qd4+ Qc5 Qxf6+ Bc6
Qxc6+ Qxc6 dxc6 Rd1+ Ka2 f5 g4 fxg4 Bxg4 Kxc6 Kb3
Rd2 Re6+ Rd6 Re7 h5 Bh3 Rf8 Bg2 Kb6 Re5 Rff6 Rg5 h4
h3 a5 a4 bxa4+ Kxa4 Rf5 Rxf5 gxf5
31 +0.00 01:54 250839K Re7+ Kb6 Qxd4+ Kxa5 b4+ Ka4 Qc3 Qxd5 Ra7 Bb7 Rxb7
Rhe8 Rb6 Ra8 Rxf6 Qc4 Qxc4 bxc4 Kb2 Re7 f4 Kb5 Rd6
f5 Bg2 Rc8 Rd5+ Kb6 Rd6+ Kb5
32 +0.00 02:01 265004K Re7+ Kb6 Qxd4+ Kxa5 b4+ Ka4 Qc3 Qxd5 Ra7 Bb7 Rxb7
Rhe8 Rb6 Ra8 Rxf6 Qc4 Qxc4 bxc4 Kb2 Re7 f4 Kb5 Rd6
f5 Bg2 Rc8 Rd5+ Kb6 Rd6+ Kb5
33 +0.00 02:05 272920K Re7+ Kb6 Qxd4+ Kxa5 b4+ Ka4 Qc3 Qxd5 Ra7 Bb7 Rxb7
Rhe8 Rb6 Ra8 Rxf6 Qc4 Qxc4 bxc4 Kb2 Re7 f4 Kb5 Rd6
f5 Bg2 Rc8 Rd5+ Kb6 Rd6+ Kb5
34 +0.00 02:22 310397K Re7+ Kb6 Qxd4+ Kxa5 b4+ Ka4 Qc3 Qxd5 Ra7 Bb7 Rxb7
Rhe8 Rb6 Ra8 Rxf6 Qc4 Qxc4 bxc4 Kb2 Re7 f4 Kb5 Rd6
f5 Bg2 Rc8 Rd5+ Kb6 Rd6+ Kb5
35 +0.39 02:43 353041K Re7+ Kb6 Qxd4+ Kxa5 b4+ Ka4 Qc3 Qxd5 Ra7 Bb7 Rxb7
Rhe8 Rb6 Ra8 Rc6 Re2 Rc5 Qd1+ Kb2 Rxc2+ Qxc2+ Qxc2+
Rxc2 Re8 Bf1 Nd5 Rc6 Re6 Rxe6 fxe6 Bh3 Ne3 Bxe6 Nd1+
Kc2 Nf2 Kc3 Kxa3 Bg8 h6 Bf7 Nd1+ Kd2 Nb2
-
- Posts: 416
- Joined: Thu Aug 02, 2012 7:48 pm
- Location: Germany
Re: possible SF regression (testposition) ?
yep....unfortunately i dont know how to post engine outputs but eg the version from 25. january by Joona is significantly faster here...
Wahrheiten sind Illusionen von denen wir aber vergessen haben dass sie welche sind.
-
- Posts: 4567
- Joined: Sun Mar 12, 2006 2:40 am
- Full name:
Re: possible SF regression (testposition) ?
It is a bit of an open door for anyone with some experience with computerchess, but just to mention it, it is a bit of a tradeoff; you do more nullmove pruning (in this case, the nullmove searches are made shorter so they take less time so you can search deeper) so get a better depth overall. The result of the change Variable null-move value based reduction is about 2.6 elo. But in some positions you will miss things. The depth of the verification search is also reduced so it will be slightly worse detecting any Zugzwangs.
Maybe a list of positions like this if it can be shown to be related to this change, could be helpful. Maybe some unrelated search improvement would help, but the weakness involving Zugzwang will not go away unless nullmove is totally disabled.
P.S.: engine output in most GUIs can be simply copied I think and then posted/pasted here in Talkchess. That should not be a problem?
Eelco
Maybe a list of positions like this if it can be shown to be related to this change, could be helpful. Maybe some unrelated search improvement would help, but the weakness involving Zugzwang will not go away unless nullmove is totally disabled.
P.S.: engine output in most GUIs can be simply copied I think and then posted/pasted here in Talkchess. That should not be a problem?
Eelco
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
-
- Posts: 6442
- Joined: Tue Jan 09, 2007 12:31 am
- Location: PA USA
- Full name: Louis Zulli
Re: possible SF regression (testposition) ?
The very latest Stockfish (signature=7451319, Threads=16, Hash=1024) still finds this fairly quickly:Spliffjiffer wrote:
Also the following position was "solved" about twice as fast by the version of jan 25. mentioned above compared to nearly all versions that followed up since then:
[D] 1rbq1r1k/pp2n1bp/3p1p2/2p1pPP1/P3B2Q/2NP3N/1P5P/n1B2RK1 b - - 0 16
bm Be6
Code: Select all
info depth 29 seldepth 55 score cp 117 nodes 307014343 nps 14255866 time 21536 multipv 1 pv c8e6 f5e6 f6f5 g5g6 e7g6 h4d8 b8d8 e4f5 a1b3 c1g5 d8e8 c3b5 f8f5 f1f5 e8e6 f5f7 e5e4 d3e4 h7h6 f7b7 b3d4 g5d8 a7a6 b5c3 d4f3 g1g2 g6h4 d8h4 f3h4 g2g3 g7c3 b2c3 e6e4 h3f4 h8g8 a4a5 h4f5 g3f3 e4e3 f3g4