Positional Test Position

[Joerg Oster]

I don't know what others are doing, but I quit testing locally at all.
Running local tests is simply not reliable, unless you play a lot of games. And Fishtest was created to do exactly this for you.

[Leto]

I agree, it is a difficult test position that can distinguish stockfish from komodo currently. IMHO more tasty would have been to skip the forced move sequence and start from here
[d]r1b3k1/ppp2pp1/3p3p/8/2P4N/P5P1/1P2qPB1/R1B2RK1 w - - 1 18

White will loose a pawn soon and then the formal material balance is even.
Stockfish believes the position is balanced, komodo believes white is to be preferred and shows a convincing pv, leading to this position:

[d]r6k/p1p2Bp1/1p5p/1bp1N3/q7/P3B1P1/3R1P2/R5K1 b - - 1 26
Suddenly both engines recommend 26..g5?(?) as if that would be completely normal.
They then follow up with 27.f4! and black is busted soon. However there must be better attempts to defend, e.g. 26..Be8!? and no mate attack is there.

So we can safely conclude that unbalanced positions are difficult.

Yes 26...Be8 seems stronger but I think the position is still lost for black. For example 26...Be8 27.Bd5 Rb8 28.f4 Qa5 29.Rad1 Qxa3 30.Rd3 Qa6 31.Be4 Qc8 32.Kf2 c4 33.Rd4 c3 34.g4 a6 35.g5 Qh3 36.Rh1 Qe6 37.Rdd1
and I doubt black can survive this

[d]1r2b2k/2p3p1/pp2q2p/4N1P1/4BP2/2p1B3/5K2/3R3R b - - 0 37

If 28...Kh7 then 29.Bf2 Bb5 30.Re1 Kh8 31.Re3 and Stockfish now sees black is in trouble.

[BeyondCritics]

Ok, thank you.
I assume now, it is ok to test if you lunch the fishtest worker sufficiently often at the same time.
Another question: What do you do if parameters are sought?
I know there is some specification about spsa and clop available, but how would you actually do it?
Oliver

[Joerg Oster]

Ok, thank you.
I assume now, it is ok to test if you lunch the fishtest worker sufficiently often at the same time.
Another question: What do you do if parameters are sought?
I know there is some specification about spsa and clop available, but how would you actually do it?
Oliver

That would be great if you would join. Running the worker is not necessarily a precondition for queuing some tests.

I cannot say much about SPSA tuning, as I never used it.
Actually I am running some CLOP sessions locally. Unfortunately, CLOP doesn't work in the framework.

For further questions I recommend the Fishcooking forum: https://groups.google.com/forum/?fromgr ... ishcooking

[BeyondCritics]

Yes agreed, black is totally dominated by the white pieces after 26..Be8.
My point was, that 26...g5 seemingly was not recognized by stockfish as a weakening -- you may conclude this in multi pv mode on lower depth -- and i didnt wanted to make the post more complicated.

[d]r1b3k1/ppp2pp1/3p3p/8/2P4N/P5P1/1P2qPB1/R1B2RK1 w - - 1 18
In this position i believe white has clearly the better end. The question is how big the advantage is and how to realize it.
A mere mortal might have preferred something like 18.b4 to keep a smooth pawn structure intact, but the engines are programmed for activity and therefore 18.b3 Qc2 19.Be3!? Qxb2 and now i believe that bringing back the knight with 20.Nf3! is really good:
[d]r1b3k1/ppp2pp1/3p3p/8/2P5/Pq2BNP1/5PB1/R4RK1 b - - 1 20
Seemingly black is in trouble. I think it is established that after 20...Qa4 21.Rfd1!? Bd7 22.c5! white has clearly the advantage or more and after 22...dxc5 23.Ne5 Bb5 24.Bd5 he is very likely already winning.

In any case it is a tough position, testing evaluation features in an unexpected way.

[BeyondCritics]

Thank you for your assistance. My next questions are directed to the fishtest forum. Maybe some other developers here are interested too, so i think my questions were not so way off.
Oliver

[Eelco de Groot]

I agree, it is a difficult test position that can distinguish stockfish from komodo currently. IMHO more tasty would have been to skip the forced move sequence and start from here
[d]r1b3k1/ppp2pp1/3p3p/8/2P4N/P5P1/1P2qPB1/R1B2RK1 w - - 1 18

White will loose a pawn soon and then the formal material balance is even.
Stockfish believes the position is balanced, komodo believes white is to be preferred and shows a convincing pv, leading to this position:

[d]r6k/p1p2Bp1/1p5p/1bp1N3/q7/P3B1P1/3R1P2/R5K1 b - - 1 26
Suddenly both engines recommend 26..g5?(?) as if that would be completely normal.
They then follow up with 27.f4! and black is busted soon. However there must be better attempts to defend, e.g. 26..Be8!? and no mate attack is there.

So we can safely conclude that unbalanced positions are difficult.

Another modified Stockfish at depth 40, 26... g5 is still there but 26... Be8 has just suffered a Fail Low throwing it out of the best four, and out of nowhere a new best move appears, 26... a6 This is all without any change to King safety, but you would think the depth can compensate it a bit.


[d]r6k/p1p2Bp1/1p5p/1bp1N3/q7/P3B1P1/3R1P2/R5K1 b - -

Engine: Sf20140810_022 MOD MP (Q6700, 32 bit 4 threads, 512 MB)
by Tord Romstad, Marco Costalba and Joona

40 513:17 -0.79 1...a6 2.Bg6 Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 7.Bf4 Re6
8.Rxe8+ Rxe8 9.Bxe8 Kf8 10.Bg6 (97.998.983.449) 3182

40 513:17 -0.84 1...g5 2.f4 gxf4 3.gxf4 Rf8 4.Kf2 Be8
5.Rd8 Kh7 6.Bc4 Qc2+ 7.Kg3 Qc3
8.Bd3+ Kg7 9.Rg1 Kf6 10.Kh2 h5 11.a4 Rh8
12.Rc8 Qb4 13.Rxe8 Rxe8 14.Rg6+ Ke7 (97.998.983.449) 3182

40 513:17 -0.95 1...Kh7 2.Bg6+ Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 7.Bxe8 Kh7
8.a4 Qc3 9.Bb5 Kg6 10.Kg2 Qb4 11.Bd2 Qb3
12.Bf4 Re7 13.R8d7 Rxd7 14.Rxd7 a6 (97.998.983.449) 3182

40 513:17 -1.00 1...Qe4 2.Bf4 Qf5 3.Bg6 Qf6 4.Rad1 Kg8
5.Bc2 Rc8 6.Bb3+ Kh7 7.a4 Be8 8.Nd7 Qe7
9.Rd3 c4 10.Bxc4 Bh5 11.Rc1 Bg6
12.Rdd1 Bf5 13.Ne5 c5 14.Nf7 Bg4 (97.998.983.449) 3182

[BeyondCritics]

Hi Eelco,

yes this _is_ an interesting position, very suitable for deep computer analysis.
Exploring interesting(!) positions with super deep analysis was often very exciting for me. Take for example the question: What the heck is 1...a6!? good for??. Looking somewhat closer at the lines given, it can be said that it restricts the white bishop somewhere deep within critical variations. Take e.g. this position after 1...a6 2.Bg6 Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 and now not 7.Bxe8 Kh7 8.a4
[d]3RB3/2p3pk/pp5p/q1p1r3/P7/4B1P1/5P2/3R2K1 b - - 0 33
because of 8.. Qc3 9.Bc6 Rxe3! with only slight winning chances for white according to stockfish.
On the other hand after 1...Kh7 2.Bg6+ Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 7.Bxe8 Kh7 8.a4 Qc3 9.Bb5!
[d]3R4/p1p3pk/1p5p/1Bp1r3/P7/2q1B1P1/5P2/3R2K1 b - - 2 34
9...Rxe3?! win easilly for white.
Now the "dual" question is: Why does the winning plan employed after 1..a6 does not work after 1..Kh7? The problem seems to be the weakening of the black pawn chain. E.g. 1...Kh7 2.Bg6+ Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 and now not 7.Bf4 Re6 8.Rxe8 Rxe8 9.Bxe8 because of 9..g5 10.Bxc7 c4 with enough counterplay for a draw:
[d]4B1k1/p1B5/1p5p/q5p1/2p5/P5P1/5P2/3R2K1 w - - 0 36
On the other hand after 1...a6 2.Bg6 Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 7.Bf4 Re6
8.Rxe8+ Rxe8 9.Bxe8 not 9..g5 10.Bxc7 c4
[d]4B1k1/2B5/pp5p/q5p1/2p5/P5P1/5P2/3R2K1 w - - 0 36
because of 11.Rd7! winning this time.
The placement of the most innocent looking piece on the board makes a decisive difference!

[Leto]

Hi Oliver, after 1...a6 2.Bg6 Kg8 3.Bc2 Qa5 4.Rad1 Re8
5.Bg6 Rxe5 6.Rd8+ Be8 7.Bf4 Re6
8.Rxe8+ Rxe8 9.Bxe8 here's how white wins. Very instructive endgame technique:

1... a6 2. Bg6 Kg8 3. Bc2 Qa5 4. Rad1 Re8 5. Bg6 Rxe5 6. Rd8+ Be8 7.
Bf4 Re6 8. Rxe8+ Rxe8 9. Bxe8 Kf8 (9... g5 10. Bxc7 c4 11. Rd7) 10. Bc6 Qxa3
11. Bxc7 Kf7 (11... b5 12. Rd7 Qc3 13. Kg2 c4) 12. Bxb6 a5 13. Kg2 a4 14. Bc7
Qb4 15. Rd7+ Ke6 16. Rd6+ Kf7 17. Bd5+ Ke7 18. Ra6 Qb5 (18... c4 19. Bd6+ Qxd6)
19. Ra7 c4 20. Bf3 g6 21. Ba5+ Ke8 (21... Ke6 22. Bc3 Kd6 23. Bd2 h5 24. Ra5
Qb3 25. Ra6+ Ke5 26. Bh6 Kd4 27. Rxg6 Qc2 28. Bg7+ Kc5 29. Rh6 Kb4 30. Rxh5 Qg6
31. Bf8+ Kb3 32. Rb5+ Kc3 33. Rb7 Kd3 34. Bg7 c3 35. Rd7+ Kc2 36. Rc7 Kb3 37.
Bxc3) 22. Be4 Qc5 23. Ra8+ Kd7 24. Bc3 Qa3 25. Ra7+ Kc8 26. Bd2 g5 27. Ba5 Qc5
28. Ra8+ Kd7 29. Bc3 Qb5 30. Ra7+ Kd6 31. Rb7 Qa6 32. Rh7 Qb5 33. Rxh6+ Ke7 34.
Bc6 Qf5 35. Bxa4 Qe4+ 36. Kh2 Qf5 37. Bf6+ Kf8 38. Bd4 Ke7 39. Be3 g4 40. Rc6
Qh7+ 41. Bh6 Qh8 42. Bc2 c3 43. Bf5 Kf7 44. Kg1 Qe5 45. Bxg4 Qe1+ 46. Kg2 Qe4+
47. Bf3

position after 26...g5:
[d]2k5/R7/7p/6p1/p1p1B3/q5P1/3B1PK1/8 w - - 0 27

Here the goal for white is to break the h pawn's defense by forcing the black king into the same rank. Position after 32.Rh7:

[d]8/7R/q2k3p/6p1/p1p1B3/2B3P1/5PK1/8 b - - 0 32

After the h pawn falls black's defense crumbles as the queen can not protect her three isolated pawns.

[BeyondCritics]

Yes that makes sense. But i have one objection. Instead of 35.Bc6 my copy of stockfish thinks 35.Bg6! is stronger:
[d]5k2/2p3p1/pp4Bp/q1p5/5B2/P5P1/5P2/3R2K1 b - - 2 35
Now black should run according to stockfish and then another pawn drops: 35..Ke7 36.Bxc7 Qxa3 37.Bxb6 c4 38.Bd4 and the rest is similar.
In any case its instructive that white always wins.