Rybka, Fritz 10: some possible bugs

[ml]

Rybka evaluates the following drawn position as a win, scoring about +6.0 (Rybka WinFinder scores +0.10; most other programs immediately see the draw without the 4-piece tablebase). I wonder if this was done to eliminate excessive code and use endgame tablebases instead.

[d]6k1/8/7P/7K/8/3B4/8/8 w





Fritz 10 thinks the following is a draw after 1.Rb1+ cxb1=q 2.Qxc3+ kxc3 (Fritz 5.32 sees the right move 1. ...cxb1=n). Has this been corrected in Fritz 11?

[d]8/8/3q4/7p/6pP/2r3P1/1kp3PK/2R1Q3 w

[Mike S.]

It's not a bug, it is missing knowledge. Some endgame knowledge has (unfortunately!) been removed between Rybka versions 2.2 and 2.3.2a.

See the output from Rybka 2.2 (without any tablebases!!):

Analysis by Rybka 2.2 32-bit:

1.h7+ Kh8 2.Bf1 Kxh7 3.Bd3+ Kg7 4.Kg5 Kf8 5.Kf6 Kg8 6.Bc4+ Kh7 7.Kg5 Kh8
= (0.09) Depth: 2 00:00:00
= (0.09) Depth: 37 00:00:10 1700kN

It has been announced that (all or most of the?) endgame knowledge which has been removed after version 2.2, will be included again with version 3.0. I do not know any more details about it. But it is clear that the blind bishop knowledge NEEDS to be included into a 3000+ Elo engine, otherwise it's a joke.

[ml]

So the endgame code was removed on purpose, as I suspected. The only reason for removing endgame knowledge would be to improve the endgame in other areas, or improve the overall strength of the engine, e.g. by relying on tablebases. According to the developer of the engine, version 2.3.2a plays a better endgame than previous versions. But I agree that the endgame knowledge should be included in the engine itself where possible.

[ozziejoe]

You could have it both ways. You could have an optional paramter that, if checked, gives you the endgame knowledge (including bishop underpromotion etc), but perhaps weakens the engine in the opening and middle game.. Most of us use Rybka for analysis. If rybka is making this kind of mistake at the endgame, then it could be making a similar mistake earlier on (e.g., when it calculates a particular middle game position as good because it can see trading into a particular endgame that it thinks is good)

best
J