Nirvanachess 2.1 Release

[cetormenter]

Hello all,

Nirvanchess 2.1 has been released.

More details can be found at http://nirvanachess.blogspot.com/2015/0 ... lease.html

The biggest part of this release is the introduction of singular extensions which showed to be a large increase in strength (approximately 20). Everything else is small parameter tweaks using either SPSA or Texel's Tuning method.

Also SMP should be much more stable than before and should always run using all cores.

Some test games run at 60"+0.05"

Code: Select all

Rank Name      Elo    +    - games score oppo. draws 
   1 2.1        39    7    7  5000   63%   -39   48% 
   2 2.0a      -39    7    7  5000   37%    39   48% 

[Graham Banks]

Great engine! Thanks.

[IWB]

Hello Thomas,

Thanks for the new version.
I try to run it, and it seems to run properly, but it gives 4, 5, ..., 8 time in a Ponder on game at 5 + 3 an eval of 124.51 (+/- depending if it is black or white) just for one move. After that it comes back with a normal eval again ...

Don't know if this is relevant, just saying.

Bye
Ingo

[Steve Maughan]

Hi Thomas,

+20 ELO from Singular Extensions seems good. Would you mind outlining the general approach which worked?

- Thanks, Steve

[cetormenter]

@Ingo

Ah! I see the issue it only occurs during pondering. However it should not affect playing strength at all. It only happens when a fail high/low is immediately asked to stop via a ponderhit. If it is bothersome then I can send you a new version as there is only two lines of code changed.

@Steve
There is nothing special here. I used the same approach as stockfish, texel and many others. If a move is not extended, and previously was either a beta cutoff or the best move at an exact node. Then I take the table value and shift it downwards (about 30 CP in my case). Then search the remaining moves at a reduced depth using a null window around this reduced value. If none of the remaining moves fail high then I extend the move.

One thing that I did notice was that exact nodes are much more important than fail high nodes. When I tested using only fail high nodes the results were only slightly positive. This makes sense because at fail high nodes there is probably more than one way to achieve a score >= beta but at exact nodes you know that only one move was able to achieve that value.

[Steve Maughan]

Thanks Thomas - Great explanation.

I'll cobble something together and throw it at Maverick to see if anything sticks.

- Steve

[cdani]

Congratulations for the new version!

I will try again this idea of singular extensions.

[TShackel]

Hello all,

Nirvanchess 2.1 has been released.

More details can be found at http://nirvanachess.blogspot.com/2015/0 ... lease.html

The biggest part of this release is the introduction of singular extensions which showed to be a large increase in strength (approximately 20). Everything else is small parameter tweaks using either SPSA or Texel's Tuning method.

Also SMP should be much more stable than before and should always run using all cores.

Some test games run at 60"+0.05"

Code: Select all

Rank Name      Elo    +    - games score oppo. draws 
   1 2.1        39    7    7  5000   63%   -39   48% 
   2 2.0a      -39    7    7  5000   37%    39   48% 

Hi Thomas. Thanks for this update of your great engine! I was wondering though why this post says it's 39 elo stronger than nirvanachess 2.0a, whereas on your blog post, it says 80 elo stronger. Why the difference?

Sincerely,

Tim.

[gordonr]

@Ingo

Ah! I see the issue it only occurs during pondering. However it should not affect playing strength at all. It only happens when a fail high/low is immediately asked to stop via a ponderhit. If it is bothersome then I can send you a new version as there is only two lines of code changed.

I've got a tournament going on in Arena 3.5 with *no pondering* and I've seen the same issue. It doesn't seem to be affecting the play. No big deal for me.

Thanks for your work on this nice engine.

[cetormenter]

@Gordon
Hmm. Well in any case the issue is fixed. Here is a direct link.

http://www.mediafire.com/download/1ied4 ... ss+2.1.rar


@Tim
Take a look again. In the main post the ratings are 39 and -39 respectively so combining the results gets 78. 80 is a nice round number and is within the error bars so I chose to use that.

Also something else I forgot to mention that I added into Nirvana in this version was a change to move ordering. Previously any searched quite move that did not produce a cut off had its history score reduced if the node turned out to be a cut node. However now if the move failed high during a reduced search and then failed low during the research its score is left alone with the idea being that if a similar position is found in the future this move will need to be researched again if its score is reduced.