singular extension

[xr_a_y]

I see that stockfish or xiphos use singular extension only once in a line. They don't use "recursive" singular extension. Can someone explain why please, I mean not by the elo argument.

Thanks

[hgm]

Is there anything else than the Elo argument that drives the development of Stockfish?

[elcabesa]

I see that stockfish or xiphos use singular extension only once in a line. They don't use "recursive" singular extension. Can someone explain why please, I mean not by the elo argument.

Thanks

from my understanding singular extension can be found more than one time in a line. Stockfish will not singular extend a move after another extended one. But it can happen that it will extend more than one time for a line. Usually it's not extended 2 times in a row to avoid search size ecplosion

[xr_a_y]

Is there anything else than the Elo argument that drives the development of Stockfish?

eheh ! at least there is something else that drives Weini development ! => learn about chess programming

[elcabesa]

My explanation is not probably totally correct... but I still believe that more than one singular extension will be possible in a line.

[xr_a_y]

My explanation is not probably totally correct... but I still believe that more than one singular extension will be possible in a line.

I must say I think I tried both and do not see a great elo gain... As stockfish comments "+60elo" inside the source code in front of singular extension, I was hopping for something better.

[elcabesa]

Some time ago stockfish calculated elo value of each technology used. If you use all the algorithm together they will have less elo gain. I don't remember how they calculated the elo gain of each algorithm.
1)disable everything and enable a single algorithm a time
2)enable everything and disable a single algorithm a time

[xr_a_y]

Due to the easily editable configuration of Weini, I used to do this also.
I'll run a big test soon to display some results.

[Cardoso]

I must say I think I tried both and do not see a great elo gain... As stockfish comments "+60elo" inside the source code in front of singular extension, I was hopping for something better.

It would be very interesting making a new test using SF9 on singular extensions playing 20k+ games.
Perhaps one test using just one core, and a second test using the maximum number of cores available (preferable on a system with more than 32 cores).
If you have big hardware (many cores) maybe it makes some sense doing singular extensions but even so I don't think they will give much ELO.
Singular Extensions is a bit of a myth, the Deep Blue team used them and talked about them, so it became a "must" for every chess engine programmer.
But the Deep Blue team didn't test SE using 20k+ games at long TC!
I don't have a chess engine myself, my engine is a spanish checkers engine and my experience is that if you test SE in sharp/tactical positions you get longer pvs and the engine zooms into the "right score" slightly faster, so you think it must be a good thing.
But when you make a few thousand test games things won't look so rosy. The problem is not the excluded search, but the extension it self. I made some tests just doing the excluded search and not making the extension and the increase in nodecount was negligible. But when you trigger the extension (usually one ply) the trees get larger, much larger. So you pay a price in nodecount. And in equal/quite positions SE is not helping, at least that is my experience.
The thing about extensions in general is that they increase the nodecount, maybe we should think that near the tips how much of those extensions will make it into a new pv. I know SF only does SE only if "depth>8*ONE_PLY" but I'm talking now about extensions in general and not specifically about SE.

[corres]

For getting an effective singular extension at first you should optimize the parameters of your engines.
In particular the values of the figures, the position dependent bonus/penalty of figures, the bonuses of free pawns and the parameters of material imbalance are very important.
Without optimization the singular extension technique is rather ineffective.