Stockfish Singular Extension, does it make sense?

[Mincho Georgiev]

Well I have no extensions in non-pv nodes now in the strict sense of the word. I got read of the check extension in the previous version, but i do not reduce it either. Things I used to extend before have now become don't reduce. This really gave me some elo and was somewhat surprising after years of 100 elo from check extension. Maybe I will try to not reduce singular moves instead of extending but it seems like a lot of effort just to avoid reduction but you never know..
EDIT: nevermind , I don't reduce first move anyway.

Daniel, I remember that some time ago, 2-3 years maybe, you've made a test with different versions of Scorpio with diff. extensions at winboard forum. Sorry if I'm wrong, but if I'm not, could you post the results if you keep them? Thanks!

[Daniel Shawul]

I vaguely remember back in 2004 I threw out mate threat and passed pawn extensions by this idea of not reducing rather than extending. As regards the tests I wouldn't trust me simply because 100 games were a lot to me at that time Others have done more extensive tests and have more or less come to the conclusion that they are not needed except maybe at PV nodes where you want a longer PV and possibly surprise the opponent..
For instance not reducing mate threats is very costly especially if you try null move all the time (i.e not only for nodes where eval() >= beta), which I feel like doing the regular extension of adding depth. There was also another of those null move threat extensions ( Botvinnik markov) that I tried in the past and got rid off the same way.

[Don]

Moral is that this is not something you can just cut and paste and have it work, without lots of testing, tuning and hard work.

If this is really the case then I think we were extremely lucky with Stockfish. First implementation (and window size) we tried gave around +40 elo.

We of course tried to fine tune things, but were never able to improve.

We got something similar in Doch or Komodo, when I first implement it the program immediately gained a LOT of ELO.

It did slow down the search fairly significantly, but there was no question about whether it was worth it in time control games.

[bob]

Moral is that this is not something you can just cut and paste and have it work, without lots of testing, tuning and hard work.

If this is really the case then I think we were extremely lucky with Stockfish. First implementation (and window size) we tried gave around +40 elo.

We of course tried to fine tune things, but were never able to improve.

We got something similar in Doch or Komodo, when I first implement it the program immediately gained a LOT of ELO.

It did slow down the search fairly significantly, but there was no question about whether it was worth it in time control games.

It is on my test schedule once this LMR/noLMR test ends (thread is on openchess) where I am playing 10m+10s games to measure whether LMR is better at deeper depths. My testing says "no change" with Crafty with longer time controls. Now says the same for Stockfish 1.8 as well... Turning LMR off in Stockfish 1.8 (I edited the source to be safe) drops Elo by about 60 after a couple of thousand long games (about 50 after 30,000 faster games so very close). Crafty lost much more, for reasons not yet understood., Old version ( year+ ago) lost about 40 if I removed LMR but left null-move in. New version loses more than triple that. Old version was -120 for LMR + NM removal, new version is -250.

[Don]

Moral is that this is not something you can just cut and paste and have it work, without lots of testing, tuning and hard work.

If this is really the case then I think we were extremely lucky with Stockfish. First implementation (and window size) we tried gave around +40 elo.

We of course tried to fine tune things, but were never able to improve.

We got something similar in Doch or Komodo, when I first implement it the program immediately gained a LOT of ELO.

It did slow down the search fairly significantly, but there was no question about whether it was worth it in time control games.

It is on my test schedule once this LMR/noLMR test ends (thread is on openchess) where I am playing 10m+10s games to measure whether LMR is better at deeper depths. My testing says "no change" with Crafty with longer time controls. Now says the same for Stockfish 1.8 as well... Turning LMR off in Stockfish 1.8 (I edited the source to be safe) drops Elo by about 60 after a couple of thousand long games (about 50 after 30,000 faster games so very close). Crafty lost much more, for reasons not yet understood., Old version ( year+ ago) lost about 40 if I removed LMR but left null-move in. New version loses more than triple that. Old version was -120 for LMR + NM removal, new version is -250.

ELO is roughly 100 ELO in Komodo for LMR. I can believe it is a little less in Stockish because I think they are more selectivity (than Komodo) in other ways - especially near frontier nodes.

Don

[bob]

Moral is that this is not something you can just cut and paste and have it work, without lots of testing, tuning and hard work.

If this is really the case then I think we were extremely lucky with Stockfish. First implementation (and window size) we tried gave around +40 elo.

We of course tried to fine tune things, but were never able to improve.

We got something similar in Doch or Komodo, when I first implement it the program immediately gained a LOT of ELO.

It did slow down the search fairly significantly, but there was no question about whether it was worth it in time control games.

It is on my test schedule once this LMR/noLMR test ends (thread is on openchess) where I am playing 10m+10s games to measure whether LMR is better at deeper depths. My testing says "no change" with Crafty with longer time controls. Now says the same for Stockfish 1.8 as well... Turning LMR off in Stockfish 1.8 (I edited the source to be safe) drops Elo by about 60 after a couple of thousand long games (about 50 after 30,000 faster games so very close). Crafty lost much more, for reasons not yet understood., Old version ( year+ ago) lost about 40 if I removed LMR but left null-move in. New version loses more than triple that. Old version was -120 for LMR + NM removal, new version is -250.

ELO is roughly 100 ELO in Komodo for LMR. I can believe it is a little less in Stockish because I think they are more selectivity (than Komodo) in other ways - especially near frontier nodes.

Don

For Crafty (current version) it is right at +200 if there is no null-move. Or null-move with no LMR is about +150. Add both and it is a little less than 300 total. This was results ran with the current 23.3 code as part of a discussion on OpenChess...