Further weaknesses

[Uri Blass]

In the first line stockfish suggests 52.Qb1 and not 52.Qb3 but
at the end of your line 55.Qb1 wins

[D]8/8/4pk2/1P5p/3p4/P2Qb1PK/5r1P/8 w - - 0 55

stockfish give it evaluation of more than +9 pawns in a few minutes

After 52.Qb3 stockfish suggest 52...Rd2 53.Qb1 d3 54.b5 with evaluation of more than 3.5 pawns for white so I guess 52.Qb3 is winning for white.

In the second line after 51.a4 h5 52.Qxd6
white is winning by more than 4 pawns.

[D]8/8/3Qpk2/7p/PP1p4/4b1PK/5r1P/8 b - - 0 52

If black replies 52...Re2 53.b5 Bg1 54.g4 black has no mate and white wins

[D]8/8/3Qpk2/1P5p/P2p2P1/7K/4r2P/6b1 b - - 0 54

for example
54...Rxh2+ 55.Kg3 h4+ 56.Kf4 Rf2+ 57.Ke4 Re2+ 58.Kd3 Re3+ 59.Kc4

[D]8/8/3Qpk2/1P6/P1Kp2Pp/4r3/8/6b1 b - - 7 59

None of your posted positions is very clear for me, I am not a machine to compute all lines after all. I think everywhere the balance could swing both ways.

OK, so your main line would be 52.Qb1. I would suggest the following continuation: 52...e5 53.b5 d3 54.Qd3 d4
55.Qb1 Rd2 56.b6 d3 57.b7 Ba7 58.b8Q Bb8 59.Qb8 Kf5

[d]1Q6/8/7p/4pk2/8/P2p2PK/3r3P/8 w - - 0 60
Who is winning now?

In the diagram white is winning

60.Qb4 with the idea Qg4+ and Qf3+ to win the pawn d3
black cannot prevent it.

60.Qb4 Rf2 61.Qf8+ and 62.Qxf2
60.Qb4 Rd1 61.Qg4+ and 62.Qxd1

OK, complicated.
Humans need more time in open positions.

I now try 49...Rc5
The position starts getting interesting for me, I will win this.

[d]8/8/3ppk1p/2r5/3pQ3/P3b1PK/1P5P/8 w - - 0 50

OK, after 49...Rc5, I see the following line: 50.b4 Rc3 51.b5 Rb3 52.Qf3 Ke7 53.Qb7 Kf6 54.b6 d5 55. Qc7 d3 56.b7 Ba7

[d]8/bPQ5/4pk1p/3p4/8/Pr1p2PK/7P/8 w - - 0 57

Any suggestions to improve on that line?

The diagram is better for white if white plays 57.Qh7 but earlier the computer prefers 52.a4 and not 52.Qf3

[D]8/8/3ppk1p/1P6/3pQ3/Pr2b1PK/7P/8 w - - 1 52

52.a4 Rb2 53.Qd3 Ke5 54.Qa3 Rb1 55.a5 Rxb5 56,a6 Rb8 57.a7 Ra8
58.Qa6

[D]r7/P7/Q2pp2p/4k3/3p4/4b1PK/7P/8 b - - 2 58

is the main line at depth 34
white wins here
for example 58...d3 59.Qa1+ Kf5 60.Qh1(with the threats Qxa8 and Qf3+ )

[Lyudmil Tsvetkov]

You use too much processing power, Uri!
Indeed, white might have advantage in the above line.

It is difficult to analyse just on your own, especially when you are not in good shape.

Last try, 51...Ra3 instead of 51...Rb3

[d]8/8/3ppk1p/1P6/3pQ3/r3b1PK/7P/8 w - - 0 52

[Uri Blass]

You use too much processing power, Uri!
Indeed, white might have advantage in the above line.

It is difficult to analyse just on your own, especially when you are not in good shape.

Last try, 51...Ra3 instead of 51...Rb3

[d]8/8/3ppk1p/1P6/3pQ3/r3b1PK/7P/8 w - - 0 52

Does not help
The queen likes to make forks

52.b6 Rb3 53.Qa8 when black cannot capture the pawn because of a fork.

[Lyudmil Tsvetkov]

You use too much processing power, Uri!
Indeed, white might have advantage in the above line.

It is difficult to analyse just on your own, especially when you are not in good shape.

Last try, 51...Ra3 instead of 51...Rb3

[d]8/8/3ppk1p/1P6/3pQ3/r3b1PK/7P/8 w - - 0 52

OK, white seems to win in all lines after 52.Qb1.

[d]8/8/3ppk1p/1P6/3p4/r3b1PK/7P/1Q6 b - - 0 52

There are always some nasty queen checks.

In the future I will just stick to playing closed positions against engines.
But that was my general impression: when engines have more time, like in this case, it is close to impossible to beat them in the endgame, as it is just simple for them to compute and they do not make mistakes. You can beat them in a complicated middlegame, where processing power is insufficient, and especially in closed positions. Still, I wanted to excercise myself a bit.

Looking forward to the next challenge.

[Lyudmil Tsvetkov]

I like it when the framework is busy.

Just to mention, if anyone attempts to add another file to storming pawns, my vision is that the values for the 4th file should be somewhere twice lower as for other files. I.e., with black king on g8, if f file scores 30cps, e file would score 15cps, if f5 scores 30cps, e5 would score 15cps, if f6 scores 40cps, e6 would score 20cps.

Same for pawns on the 4th rank, much lower than 5th rank.

If I were the SF team, I would dismantle all storm code and give it a fresh try.

It is really nice to see STF busy.

[mcostalba]

If I were the SF team, I would dismantle all storm code and give it a fresh try.

Hi Lyudmil,

one part of the evaluation that is very ELO sensible is the evaluation of passed pawns. In SF this has been quite extensively tuned, but IMO there is still something to grab there.

According to your deep chess knowledge, what are the factors to consider here?

Below you can find the link to passed evaluation in SF, I know you are not a programmer, but that part is heavily commented, so you can read in the comments the rules we have applied.

https://github.com/mcostalba/Stockfish/ ... e.cpp#L582

[Uri Blass]

I like it when the framework is busy.

Just to mention, if anyone attempts to add another file to storming pawns, my vision is that the values for the 4th file should be somewhere twice lower as for other files. I.e., with black king on g8, if f file scores 30cps, e file would score 15cps, if f5 scores 30cps, e5 would score 15cps, if f6 scores 40cps, e6 would score 20cps.

Same for pawns on the 4th rank, much lower than 5th rank.

If I were the SF team, I would dismantle all storm code and give it a fresh try.

It is really nice to see STF busy.

It seems that Lucas has a different opinion and even when I give no patches he does not like to see the framework busy because he likes to complain about wasting resources

From his post in the stockfish forum:
" Enough resources have been wasted on tuning that array. That's all I'm saying. Let's just find the CHEAPEST way out of this situation and stop draining framework resources on that."

[mcostalba]

From his post in the stockfish forum:

Uri, please, this thread is interesting, don't spam it.

[Lyudmil Tsvetkov]

If I were the SF team, I would dismantle all storm code and give it a fresh try.

Hi Lyudmil,

one part of the evaluation that is very ELO sensible is the evaluation of passed pawns. In SF this has been quite extensively tuned, but IMO there is still something to grab there.

According to your deep chess knowledge, what are the factors to consider here?

Below you can find the link to passed evaluation in SF, I know you are not a programmer, but that part is heavily commented, so you can read in the comments the rules we have applied.

https://github.com/mcostalba/Stockfish/ ... e.cpp#L582

Hi Marco.

Very kind of you to ask my opinion on the matter, I am flattered.

You are probably kidding about my deep chess knowledge, but never mind.

I will take my time reading through SF passer specifications, it might take a while. My first problem arises when I open Sf code with a notepad and it is so difficult to read there, no text structure, nothing, and I do not know how to make it more readable. I will do my best to read it by tomorrow.

You already pushed 2 passed pawn patches that scored positively, so maybe indeed something could be done about passers. My personal feeling, however, when I look at SF games, is that SF plays very well with its passers. I have not observed really any problems in that area.

You know, when I look at 100 games, only in 25 of them I observe something going wrong that could be corrected in the eval. If search and eval are split 50/50 in their elo weight, I can say nothing about search, but concerning eval, for me only 50% of eval problems are traceable on the surface. The other half have to do with better tuning, you see SF loses a game, you know it is because of imperfect eval, and not a search problem, but you simply can not say what it is about. It is so small, so invisible. In order to correct those problems, you possibly have to tune at least 5 or 10 parameters simultaneously, but I know this is impossible for you.

So that basically I am posting only issues that I have observed as a repeated behaviour, a pattern of failure in a sufficiently large number of games. When there is a pattern, chances are that correction might succeed. I guess it is the same about search: in half of the cases you can spot the problem, in the other half the problem is so invisible that you need to tune at least a couple of parameters simultaneously to correct it.

[mcostalba]

You are probably kidding about my deep chess knowledge, but never mind.

I am not kidding, you are way stronger than me (but this is easy) and I think also to most people here.

I will take my time reading through SF passer specifications, it might take a while. My first problem arises when I open Sf code with a notepad and it is so difficult to read there

Just open the link I have posted above, don't use notepad nor any other editor, just follow the link.