24 Doublecheck 3.0 blitz games

[Mike S.]

It's just an "informal" low tech test:

3m+2s, Atom N455/1,66 Ghz
64 MB hash tables each
ponder off, 1 thread, 32-bit
4 short opening variations
Arena 2.0.1, Windows 7

Code: Select all

   Engine             Score        Do
1: Doublecheck 3.0    8.5/24 ········ 
2: Pro Deo 1.74 smart 6.0/8  011=11=1 
3: SlowChess 2.1      5.0/8  1=0=1110 
4: Cuckoochess 1.12   4.5/8  111000=1

35.4% --> -107 Elo relative to the opponent average (whatever that is, depending on the rating lists each).

Nevertheless, this could be a future top engine as well. They all had to start somewhere. While watching a few games live, Doublecheck once rushed to 127 plies of depth almost instantly, in a pawn endgame!

[lucasart]

It's just an "informal" low tech test:

3m+2s, Atom N455/1,66 Ghz
64 MB hash tables each
ponder off, 1 thread, 32-bit
4 short opening variations
Arena 2.0.1, Windows 7

Code: Select all

   Engine             Score        Do
1: Doublecheck 3.0    8.5/24 ········ 
2: Pro Deo 1.74 smart 6.0/8  011=11=1 
3: SlowChess 2.1      5.0/8  1=0=1110 
4: Cuckoochess 1.12   4.5/8  111000=1

35.4% --> -107 Elo relative to the opponent average (whatever that is, depending on the rating lists each).

Nevertheless, this could be a future top engine as well. They all had to start somewhere. While watching a few games live, Doublecheck once rushed to 127 plies of depth almost instantly, in a pawn endgame!

Thanks for your interest!

Do you have the position where DC reached 127 plies ? It's more likely to be a bug

[Mike S.]

Do you have the position where DC reached 127 plies ?

20: Doublecheck 3.0 - Pro Deo 1.74 smart, Double3-Blitztest 2012
[d]8/8/p1p1k1p1/P1p3P1/2P1K3/8/8/8 w - - 0 52
The games were in Arena, but this was reproducable in the Fritz GUI too:

Analysis by DoubleCheck 3.0 JA:

52.Kf4
= (0.01) Depth: 127 00:00:01 7kN
= (0.01) Depth: 127 00:00:01 7kN

If you are planning to enhance the endgame knowledge, here is something to fix:

12: SlowChess 2.1 - Doublecheck 3.0, Double3-Blitztest 2012
[d]2B5/8/1b3k2/1P4pp/8/5K2/8/8 b - - 0 94

Analysis by DoubleCheck 3.0 JA:

94...g4+ 95.Ke4 Kg5 96.Kd5 h4 97.Kc6 Bg1 98.Kd5 h3 99.Ke5
-+ (-1.97) Depth: 11 00:00:01 58kN
-+ (-2.29) Depth: 23 00:00:38 11461kN

Does not expect 95.Bxg4! =

The game showed that Doublecheck knows that K-KB is a draw - thus not capturing the last white pawn - but does not know that KP-KB is a draw too (with exceptions) as evals remained about +2 in that 4-piece ending.

After pondering about this, I am guessing that it's not very easy to implement that because there are KB-KP positions where the pawn cannot be stopped. I understand that therefore, 0.00 most probably should not simply be assigned just based on the material. Or can the search always "overrule" this if required? I'm not a programmer.

[d]8/B7/8/8/4k3/3p4/8/7K w - - 0 1

[tpetzke]

When you assign a Draw Score to KBKP then you should have a recognizer that is able to spot the exceptions (where the pawn cannot be stopped) otherwise you should only make the score more drawish but let the search continue.

Probably you should start to make the score more drawish already in the originating position because of the opposite colored bishops.

Thomas...

[lucasart]

Do you have the position where DC reached 127 plies ?

20: Doublecheck 3.0 - Pro Deo 1.74 smart, Double3-Blitztest 2012
[d]8/8/p1p1k1p1/P1p3P1/2P1K3/8/8/8 w - - 0 52
The games were in Arena, but this was reproducable in the Fritz GUI too:

Analysis by DoubleCheck 3.0 JA:

52.Kf4
= (0.01) Depth: 127 00:00:01 7kN
= (0.01) Depth: 127 00:00:01 7kN

Reproduced it too. DC is correct, and finds a forced repetition, after which depth increases fast (due to hash table) => OK

If you are planning to enhance the endgame knowledge, here is something to fix:

12: SlowChess 2.1 - Doublecheck 3.0, Double3-Blitztest 2012
[d]2B5/8/1b3k2/1P4pp/8/5K2/8/8 b - - 0 94

Analysis by DoubleCheck 3.0 JA:

94...g4+ 95.Ke4 Kg5 96.Kd5 h4 97.Kc6 Bg1 98.Kd5 h3 99.Ke5
-+ (-1.97) Depth: 11 00:00:01 58kN
-+ (-2.29) Depth: 23 00:00:38 11461kN

Does not expect 95.Bxg4! =

The game showed that Doublecheck knows that K-KB is a draw - thus not capturing the last white pawn - but does not know that KP-KB is a draw too (with exceptions) as evals remained about +2 in that 4-piece ending.

Confirmed: DC has no knowledge of KBKP, but it knows KBK is draw (according to chess rule, not even "knowledge" actually).

After pondering about this, I am guessing that it's not very easy to implement that because there are KB-KP positions where the pawn cannot be stopped. I understand that therefore, 0.00 most probably should not simply be assigned just based on the material. Or can the search always "overrule" this if required? I'm not a programmer.

Yes, it's possible to detect whether a KBKP is a draw in the eval. So long as the Bishop can at least sacrify itself against the pawn we know it's a draw (B can't mate alone, and opponent cant either since the B can sacrifice itself against the P).
What is more interesting is the KBPK, where one has both a B and a P but still cannot mate (due to stalemate problem, when bishop is of the wrong color and lone king can keep opposition).