STS [1-10] Stockfish 1.8 Tactical

[Swaminathan]

This is the settings from Lucenathelucid



http://sites.google.com/site/strategict ... st-results

1000 Positions
10 seconds per position
Hardware: Q6600, 32 bits, 2 GB RAM, 2.4 GHZ. Arena 2.01 GUI.

[LucenaTheLucid]

Thanks Swami...

I was expecting a much bigger gain.

I also tested these scores using WAC and they scored 288/300 compared to 284/300 for the default settings. Also according to these 1/1 blitz games I am playing:

Stockfishtest19 2010

Deep Rybka 4 w32 - Stockfish 1.8 JA 125.5 - 98.5 +79/=93/-52 56.03%
Deep Rybka 4 w32 - Stockfish 1.8 JA TACTICAL 114.5 - 108.5 +70/=89/-64 51.35%

These settings are promising for a small ELO gain. Thanks again Swami!

[Hagen]

it would be interesting to see if Stockfish's results could be tweaked even more by altering the settings. I'm almost certain a change in the setting could make it even stronger. I'm guessing these results were using only the default listings in Stockfish right out of the box?

[royb]

Thanks Swami...

I was expecting a much bigger gain.

I also tested these scores using WAC and they scored 288/300 compared to 284/300 for the default settings. Also according to these 1/1 blitz games I am playing:

Stockfishtest19 2010

Deep Rybka 4 w32 - Stockfish 1.8 JA 125.5 - 98.5 +79/=93/-52 56.03%
Deep Rybka 4 w32 - Stockfish 1.8 JA TACTICAL 114.5 - 108.5 +70/=89/-64 51.35%

These settings are promising for a small ELO gain. Thanks again Swami!

Lucid,

Are these the same "spoon" settings you outlined back when Stockfish-1.7.1 was the current release of Stockfish? Or are these different settings from the 1.7.1 "spoon" settings?

Thanks.

[LucenaTheLucid]

Different settings. I will not release them quite yet. I want to make sure its stronger.

[LucenaTheLucid]

Just some more updates

Silent But Deadly

Stockfish 1.8 Default-----123/134
Stockfish 1.8 TACTICAL --124/134

New Win At Chess

Stockfish 1.8 Default----- 284/300
Stockfish 1.8 TACTICAL-- 288/300

Arasan test

Stockfish 1.8 Default-----026/225
Stockfish 1.8 TACTICAL--035/225

Eigenmann Endgame Test

Stockfish 1.8 Default-----028/100
Stockfish 1.8 TACTICAL--029/100

Head to Head vs Rybka

Deep Rybka 4 w32 - Stockfish 1.8 JA-----------------186.5--154.5--+114/=145/-82---54.69%
Deep Rybka 4 w32 - Stockfish 1.8 JA TACTICAL----173.0--168.0--+102/=142/-97---50.73%

[Hagen]

Oh my god. You gotta release those tweaked settings for Stockfish 1.8 Heh. To think I was about to take the plunge and actually buy Rybka 4. Now I think I'll pass. We may not have to wait for Stockfish 2.0 to beat Rybka 4. Based on your new settings...Stockfish 1.9 (when it comes out) may be the one to beat Rybka 4. It's a wonder I'm not seeing these numbers on the Rybka forum.

[LucenaTheLucid]

Settings are:

Check Extensions (pv nodes) - 0
Check Extensions (non pv nodes) - 0
Single Evasion Extension (pv node) - 0
Single Evasion Extension (non pv node) - 0

Everything else default. Also these settings seem to improve the endgame phase:

Pawn Endgame Extension (pv node) - 0
Pawn Endgame Extension (non pv node) - 0

I'm still not sure whether they are statistically better than the default or not.

[LucenaTheLucid]

Last update:

Another reason why you cannot rely on only a small amount of games to determine who is better.

First I posted this:

Deep Rybka 4 w32 - Stockfish 1.8 JA-----------------186.5--154.5--+114/=145/-82---54.69%
Deep Rybka 4 w32 - Stockfish 1.8 JA TACTICAL----173.0--168.0--+102/=142/-97---50.73%

And then now after 500 games:

Stockfishtest19 2010

Deep Rybka 4 w32 - Stockfish 1.8 JA-----------------260.5 - 239.5---+151/=219/-130---52.10%
Deep Rybka 4 w32 - Stockfish 1.8 JA TACTICAL----259.5 - 240.5---+151/=217/-132---51.90%

Only a 1 game difference over the default. I may do another test of 500 or so games, but I expect that they will eventually even out with the default settings. My settings I think turn off particular extensions in the code, so it does not search deeper when the situation arises. 0 = Off, 1 = Searches some, 2 = Searches extensively.

For instance, when you put Check Extensions to 0 it probably does not search any differently when it's or the opponents king is in check any more than normally. It would be interesting if one of the programmers of this forum can confer with this, as I don't possess any programming skills.

According to swami's STS tests of Default and Tactical it does better than default all around in suites 1,7 and 9. So maybe these settings are only useful over the default in those situations. How much do they arise during the game? I don't know, but turning off pawn extensions seems to really kick up Stockfish in the endgame.

Maybe this should be the next test is to put together various test positions in which a particular side has a slight advantage, and test the settings against each other or against Rybka perhaps.