Release: Crafty 23.1

[Peter Skinner]

Hello everyone,

We have released Crafty v23.1.

You can find the source here.

And I have built Win32/64 executables here.

Check main.c for changes. Enjoy.

Peter

[Dr.Wael Deeb]

Hello everyone,

We have released Crafty v23.1.

You can find the source here.

And I have built Win32/64 executables here.

Check main.c for changes. Enjoy.

Peter

Thanks Peter
Will be tested regards,
Dr.D

[Vinvin]

I read the changes ... is all written there ? this worth 100 elo ??

Hello everyone,

We have released Crafty v23.1.

You can find the source here.

And I have built Win32/64 executables here.

Check main.c for changes. Enjoy.

Peter

[Peter Skinner]

I read the changes ... is all written there ? this worth 100 elo ??

In our testing, yes we have improved 100+ elo. This is based on literally millions of games via the UAB cluster. Our error bars are +4 -4, and we have improved 14% in scoring above 23.0.

Peter

[bob]

I read the changes ... is all written there ? this worth 100 elo ??

Hello everyone,

We have released Crafty v23.1.

You can find the source here.

And I have built Win32/64 executables here.

Check main.c for changes. Enjoy.

Peter

Every bit of tuning is not written up in main.c. All we add there are any significant changes in overall program structure or logic. There are quite a few changes that you are probably overlooking. The new forward pruning, replacing the old simplistic razoring code / futility pruning is one. It is only briefly mentioned in main.c, but is a significant change. There are quite a few if you read that section pretty carefully, and there are just as many tuning changes that are not mentioned there at all to keep those comments from becoming impossibly large (not to mention laborious to maintain and update for each tuning change made).

[bob]

I read the changes ... is all written there ? this worth 100 elo ??

In our testing, yes we have improved 100+ elo. This is based on literally millions of games via the UAB cluster. Our error bars are +4 -4, and we have improved 14% in scoring above 23.0.

Peter

And as I have noted previously, we are more interested in discovering useful changes and making them work, as opposed to carefully documenting every change in the comments in main.c

If you want to see the totality of the changes, I recommend "diff".

[JohnS]

Bob

Congratulations on a huge strength increase.

In terms of strength on equal hardware, how close do you think you are to the top commercials (Rybka, Shredder etc) and what are the main factors for this difference?

Thanks.

[BBauer]

Thanks for the new release.
From the comments in main.c I see a change in egtb handling.
My crafty built seams to slow down when it comes to tablebase useage.
From 7 mnps to .7 mnps.
I read the 4-men from memory and the 5-men from a fast? USB stick.
Has anyone else noticed this behavior?
Kind regards
Bernhard

[bob]

Bob

Congratulations on a huge strength increase.

In terms of strength on equal hardware, how close do you think you are to the top commercials (Rybka, Shredder etc) and what are the main factors for this difference?

Thanks.

I really don't know. I'd be surprised if we are less than 100 points away, or more than 200. But I don't run Rybka in our testing. When I tried IPPOLIT, it appeared to be maybe +150 over Crafty, if I factored out the losses on time where it crashed. But whether that test was very reliable or not is unknown, it was just a test and when I noticed the time losses I gave up.

BTW, proper testing is having a huge impact on development. I can't begin to count the number of ideas we tested and rejected. Whereas quite a few of them seemed not just reasonable, but intuitively seemed to be better, yet testing proved otherwise. In the past, we were continually taking steps backward often enough that the forward steps were not making an impact. It really helps to be more methodical so that the good ideas are kept, and the bad ideas are either eliminated, or studied and then made to work with more effort.

In the current version, there are some 15-20 Elo changes, and some 1-2-3 Elo changes. But when _all_ are positive, the effect can be remarkable. We are not out of ideas, and already have a slightly stronger 23.2 running.

[bob]

Thanks for the new release.
From the comments in main.c I see a change in egtb handling.
My crafty built seams to slow down when it comes to tablebase useage.
From 7 mnps to .7 mnps.
I read the 4-men from memory and the 5-men from a fast? USB stick.
Has anyone else noticed this behavior?
Kind regards
Bernhard

Not sure what you mean about a change in EGTB handling??? You mean about not probing if there are two pawns on adjacent files (which would mean a possible en passant in the tree?) Can't see how that would change anything except for the positions that start off in the EGTBs at the root, with a potential EP capture, which would just disable egtb's completely. Otherwise they should be no faster/slower. A 90% slowdown is not unusual if you only have 6-10 pieces on the board, particularly if one or more are pawns, since that probes many different tables and will slow down significantly.