Just for fun, here's the "smooth scaling" version

[Dann Corbit]

The only binary included in the archive is the 64 bit binary.
http://cap.connx.com/chess-engines/new- ... f_1.6.3.7z

There are UCI options for each of the three new UCI parameters.

[M ANSARI]

Any strength increase over the JA compile? Also why not include SS in the original release?

[UncombedCoconut]

Any strength increase over the JA compile? Also why not include SS in the original release?

1.6.3 is a bugfix release, intended to address all the known/patched bugs but otherwise have no effect on engine strength (unless you wish to test Stockfish 1.6.x 8CPU, which previously did not exist). I assume the answer to "why not include SS in the original release" is "because it isn't ready yet".

I think it's safe to assume that the gain (if any) of SS is the same for 1.6.3 as for 1.6.2.

[David Dahlem]

Will there be a 32-bit SS version?

[Dann Corbit]

Will there be a 32-bit SS version?

If nobody else builds one, I will make one tomorrow.

[Dann Corbit]

Any strength increase over the JA compile?

Nobody knows. The only test with more than a handful of games that I know of between the two versions is here:

Code: Select all

Stockfish 1.6.2s 64-bit 2CPU 3139 +80 −74 77.4% −183.8 32.3% 62 60.3% 
Stockfish 1.6.2 64-bit 2CPU 3114 +29 −29 58.0% −50.7 50.3% 350 

And there are not enough games for statistical significance.

Also why not include SS in the original release?

Because it is nothing by my silly futzing around with the original. Eventually, someone might make something interesting out of the idea, but I do not think we even know if it is a good idea or not yet.

I am doing something different with the way that null move pruning is handled.

[Eelco de Groot]

handful of games

Some CCRL 40/4 results 1.6s vs 1.6.2

I have not checked CEGT but there are some CCRL results Dann. Only for 1.6 with Dann's mods, not 1.6.2 I think. The best comparison maybe is the top 4 CPU version that is one with Smooth Scaling, 1793 games, 3187 elo vs 1795 games 3179 elo

[Dann Corbit]

handful of games

Some CCRL 40/4 results 1.6s vs 1.6.2

I have not checked CEGT but there are some CCRL results Dann. Only for 1.6 with Dann's mods, not 1.6.2 I think. The best comparison maybe is the top 4 CPU version that is one with Smooth Scaling, 1793 games, 3187 elo vs 1795 games 3179 elo

I am somewhat surprised to see the "s" version do well at 40/4, since at very fast time control, the "s" version boils down to the regular version (my additional shaving of the tree happens progressively at greater depths). I also suspect that at 40/2hrs we would see a much bigger difference, but it is pure conjecture on my part.

[Eelco de Groot]

It's certainly possible Dann. Nobody has tested your new UCI settings yet. I haven't either Also the new 1.6.3 may change the picture again, make 8 CPU comparisons possible and good bugfixes can only help in my opinion to bring out the differences more clearly, because bugs just introduce statistical and systematical noise in the results. The long time control testing is anyhow a bit underrepresented in changing Stockfish, there is more room for improvements in that area.

Regards Eelco

[Dann Corbit]

It's certainly possible Dann. Nobody has tested your new UCI settings yet. I haven't either

That's OK. Neither have I!

Also the new 1.6.3 may change the picture again, make 8 CPU comparisons possible and good bugfixes can only help in my opinion to bring out the differences more clearly, because bugs just introduce statistical and systematical noise in the results. The long time control testing is anyhow a bit underrepresented in changing Stockfish, there is more room for improvements in that area.

Regards Eelco