18 days from SF4 release and about ~30+ ELO gain!

[zullil]

What is a "dangerous" move?

Miguel

Code: Select all

      dangerous =   givesCheck
                 || pos.is_passed_pawn_push(move)
                 || type_of(move) == CASTLE;

[JuLieN]

Code: Select all

|| type_of(move) == CASTLE;

That's why I never castle either. Too dangerous.

[Don]

Code: Select all

|| type_of(move) == CASTLE;

That's why I never castle either. Too dangerous.

Good policy. It's just suicide to get that king trapped over there in the corner.

Promotions are dangerous too. If you lose a pawn, it's just a pawn. But if promote to a queen and it gets captured you have lost a queen and do not even have a pawn to show for it!

[michiguel]

What is a "dangerous" move?

Miguel

Code: Select all

      dangerous =   givesCheck
                 || pos.is_passed_pawn_push(move)
                 || type_of(move) == CASTLE;

Got it. They are dangerous for the opponent. I thought it could have been leaving pieces in prise or something like that. Now Uri's message has a complete different meaning for me

Miguel

[gladius]

It seems to me that most of the progress that stockfish make is thanks to search improvements.

I feel that stockfish's search is still weak part of it otherwise stockfish's developement version could solve the following position fast

[D]6k1/1r4p1/2p2Pb1/2p3R1/p1Ppq3/Q7/PP3R2/6K1 b - - 2 98

I can accept being twice slower than most top programs but stockfish developement version seems to me more than 10 times slower than most top programs after the stockfish team decided to reduce dangerous moves because unfortunately it gave some elo improvement.

Note that my last patch that passed both stage I and stage II correct the problem but marco is going to test a simpler patch that helps to find Rb3 faster than the developement version but still in the order of 10 times slower relative to my patch(and it is not that my patch could help stockfish to find Rb3 faster than the best top programs so I do not feel that I extend too much).

Note that my intuition tells me that if we want stockfish to be strong against weak opponents then we need to have less positions when stockfish is significantly weaker than them in tactics.

Maybe my intuition is wrong but my intuition tells me that the patch of not reducing dangerous moves has to be counter productive
if we test against opponents that are 200 elo weaker that is important for rating lists.

All pruning is making trade-offs in the search tree. If we extend dangerous moves in some useless section of the tree, we may not get to use those nodes for searching a valuable part of the tree.

Besides, intuition is not how patches get committed . Your patch looks very promising , but if the LMR dangerous moves patch had not been committed, what is the likelihood this area would be getting explored? In the end, it should be a big strength/tactics win!

[Don]

It seems to me that most of the progress that stockfish make is thanks to search improvements.

I feel that stockfish's search is still weak part of it otherwise stockfish's developement version could solve the following position fast

[D]6k1/1r4p1/2p2Pb1/2p3R1/p1Ppq3/Q7/PP3R2/6K1 b - - 2 98

I can accept being twice slower than most top programs but stockfish developement version seems to me more than 10 times slower than most top programs after the stockfish team decided to reduce dangerous moves because unfortunately it gave some elo improvement.

Note that my last patch that passed both stage I and stage II correct the problem but marco is going to test a simpler patch that helps to find Rb3 faster than the developement version but still in the order of 10 times slower relative to my patch(and it is not that my patch could help stockfish to find Rb3 faster than the best top programs so I do not feel that I extend too much).

Note that my intuition tells me that if we want stockfish to be strong against weak opponents then we need to have less positions when stockfish is significantly weaker than them in tactics.

Maybe my intuition is wrong but my intuition tells me that the patch of not reducing dangerous moves has to be counter productive
if we test against opponents that are 200 elo weaker that is important for rating lists.

All pruning is making trade-offs in the search tree. If we extend dangerous moves in some useless section of the tree, we may not get to use those nodes for searching a valuable part of the tree.

Besides, intuition is not how patches get committed . Your patch looks very promising , but if the LMR dangerous moves patch had not been committed, what is the likelihood this area would be getting explored? In the end, it should be a big strength/tactics win!

The term "tactics" is ill-defined anyway and almost useless except as something we as humans can wrap our brains around. It really means finding a checkmate or the win of piece AFTER someone has made a stupid move. It has nothing to do with the most important part of chess, the part where using finesse and skill you create a position that makes it difficult for your opponent. In other words that is the hard part - the tactics is the easy part that comes only after you already have a won game.

In general tactics is what impresses weaker players the most, but is not what really separates the strong players from the weak ones. That's why it's never been a huge interest of ours to make Komodo into a dumb tactical monster. It's already a tactical monster by human standards and the far more interesting part is getting to play chess as if it understands what it is doing. That's where all the focus should be.

People have made the claim that for analysis "tactics" is most important. I don't really see that either. I use computers basically to suggest moves and to check my own lines of play and track down lines. If you try a move that is a blunder ANY program will quickly tell you that. In general if the move is "losing bad" a tactical program is not superior to a positional one of the same strength. Komodo is much stronger than some more tactical programs and it cannot be that way because it cannot take advantage of bad moves.

It might turn out that some other program will see the win of a piece sooner -but if you make a bad move it will show up very quickly as a score drop - which (in human terms) means that Komodo "senses that something is wrong."

Once in a while a more tactical program will find the right move where a less tactical one won't. So what? The program that is strongest is going to find the right more more often and that is what you want for analysis.

[gladius]

The term "tactics" is ill-defined anyway and almost useless except as something we as humans can wrap our brains around. It really means finding a checkmate or the win of piece AFTER someone has made a stupid move. It has nothing to do with the most important part of chess, the part where using finesse and skill you create a position that makes it difficult for your opponent. In other words that is the hard part - the tactics is the easy part that comes only after you already have a won game.

In general tactics is what impresses weaker players the most, but is not what really separates the strong players from the weak ones. That's why it's never been a huge interest of ours to make Komodo into a dumb tactical monster. It's already a tactical monster by human standards and the far more interesting part is getting to play chess as if it understands what it is doing. That's where all the focus should be.

People have made the claim that for analysis "tactics" is most important. I don't really see that either. I use computers basically to suggest moves and to check my own lines of play and track down lines. If you try a move that is a blunder ANY program will quickly tell you that. In general if the move is "losing bad" a tactical program is not superior to a positional one of the same strength. Komodo is much stronger than some more tactical programs and it cannot be that way because it cannot take advantage of bad moves.

It might turn out that some other program will see the win of a piece sooner -but if you make a bad move it will show up very quickly as a score drop - which (in human terms) means that Komodo "senses that something is wrong."

Once in a while a more tactical program will find the right move where a less tactical one won't. So what? The program that is strongest is going to find the right more more often and that is what you want for analysis.

I totally agree! Strong engines play strong moves throughout the game, slowly increasing their advantage. But it's natural for humans to want to find the killer shot. So the 999 other positions that the "tactically weaker" performs better on are put to the side.

[Henk]

One can define a chess program as pure tactical if it only counts material.

[mwyoung]

It seems to me that most of the progress that stockfish make is thanks to search improvements.

I feel that stockfish's search is still weak part of it otherwise stockfish's developement version could solve the following position fast

[D]6k1/1r4p1/2p2Pb1/2p3R1/p1Ppq3/Q7/PP3R2/6K1 b - - 2 98

I can accept being twice slower than most top programs but stockfish developement version seems to me more than 10 times slower than most top programs after the stockfish team decided to reduce dangerous moves because unfortunately it gave some elo improvement.

Note that my last patch that passed both stage I and stage II correct the problem but marco is going to test a simpler patch that helps to find Rb3 faster than the developement version but still in the order of 10 times slower relative to my patch(and it is not that my patch could help stockfish to find Rb3 faster than the best top programs so I do not feel that I extend too much).

Note that my intuition tells me that if we want stockfish to be strong against weak opponents then we need to have less positions when stockfish is significantly weaker than them in tactics.

Maybe my intuition is wrong but my intuition tells me that the patch of not reducing dangerous moves has to be counter productive
if we test against opponents that are 200 elo weaker that is important for rating lists.

Uri,

I am no longer finding this to be true as I have said before in the newest release versions of Stockfish. I see no big searching issues with Stockfish in my testing.

To see if I could expose this weakness in search.

(slow tactics vs other programs weaker and stronger).

I grabbed the latest compile of stockfish and I did a quick test.

I played two 1/0 bullet matches to see what stockfish could do. I never test at this speed. So I had no idea what would happen, but thought this would expose such a weakness if it exist the fastest.

Here are my results. I don't know if this proves anything, but I see no big tactical search issues in regards to seeing tactics quickly in this bullet test.

I would expect to see something.

Code: Select all

Blitz, Blitz 1m  0

Stockfish 100913 64 SSE4.2   - Rybka 3                        16.0 - 4.0    +13/=6/-1    80.00%
Stockfish 100913 64 SSE4.2   - Houdini 3 Pro x64              10.5 - 9.5    +6/=9/-5    52.50%

[Mithu]

All pruning is making trade-offs in the search tree. If we extend dangerous moves in some useless section of the tree, we may not get to use those nodes for searching a valuable part of the tree.

Besides, intuition is not how patches get committed . Your patch looks very promising , but if the LMR dangerous moves patch had not been committed, what is the likelihood this area would be getting explored? In the end, it should be a big strength/tactics win!

The term "tactics" is ill-defined anyway and almost useless except as something we as humans can wrap our brains around. It really means finding a checkmate or the win of piece AFTER someone has made a stupid move. It has nothing to do with the most important part of chess, the part where using finesse and skill you create a position that makes it difficult for your opponent. In other words that is the hard part - the tactics is the easy part that comes only after you already have a won game.

In general tactics is what impresses weaker players the most, but is not what really separates the strong players from the weak ones. That's why it's never been a huge interest of ours to make Komodo into a dumb tactical monster. It's already a tactical monster by human standards and the far more interesting part is getting to play chess as if it understands what it is doing. That's where all the focus should be.

People have made the claim that for analysis "tactics" is most important. I don't really see that either. I use computers basically to suggest moves and to check my own lines of play and track down lines. If you try a move that is a blunder ANY program will quickly tell you that. In general if the move is "losing bad" a tactical program is not superior to a positional one of the same strength. Komodo is much stronger than some more tactical programs and it cannot be that way because it cannot take advantage of bad moves.

It might turn out that some other program will see the win of a piece sooner -but if you make a bad move it will show up very quickly as a score drop - which (in human terms) means that Komodo "senses that something is wrong."

Once in a while a more tactical program will find the right move where a less tactical one won't. So what? The program that is strongest is going to find the right more more often and that is what you want for analysis.

My emphasised words are to be engraved in gold! I consider this an excellent deduction of unknown factors! I am enjoying this thread a lot.