Crafty 22.8

[bob]

... had been lost, probably in copying files around to different machines

Slightly off topic but I've been using DropBox to manager my documents and code across different machines. I'm really impressed by it - it also runs on Windows, Apple and Linux. Take a look at www.getdropbox.com.

Best regards,

Steve

Here are some issues to consider. Yes, I am looking hard at some sort of versoin management facility to prevent some of this "lost effort"

1. I do most of my work on my laptop. But for obvious reasons I do not want the "master copy" to reside there, as laptops are stolen or lost or damaged every day. But I want the source handy to work on when the opportunity arises. And it might arise with no way to connect to my office machine if there is no wireless access handy.

2. I do some of my development on my office box, which is a secure machine with redundant everything, although it is fairly old at a 2 x 2.8ghz xeon box, but it has 8 15K SCSI drive and is as fast as they come with respect to disk I/O like tablebases and such. That is the natural place to house the master copy and that is what I currently do.

3. for automated backups, I keep copies of some files on our "nfs-mount home directories" although my office machine does not currently mount those directories as I don't want to depend on them and when they are down I can't do anything at all.

So there are three places I want the files, and therein lies a problem. Checking source out to keep on my laptop is problematic since I can't then modify anything on my office machine, even though no changes have been made on my laptop.

I am looking at options, but want something that doesn't take a lot of effort to make it work, which isn't obtrusive to the development process, etc. My office box is behind a firewall which will not "pass thru" any network accesses that directly go to my office machine, for security reasons. A version management system that works inside a LAN is not so bad, but I am not "inside" a LAN all the time and I don't want to have to jump thru lots of hoops to do a quick change if possible.

Suggestions welcome of course...

[bob]

[So far the testing suggests +1 Elo, so it is not exactly a world-beating bug fix.
You could continue to test the older version, or you could switch to the new version but combine the results so that 22.7 is combined with 22.8 into one 22.8 version. There's not enough difference to measure from this fix...

Hi Bob,
what do you suggest: can we do the same with the results of 22.4 (64bit 1CPU) - or its better (as difference is more than 20 elos) to start a new test series with 2.8 ?

best

The latest is better than 22.4 by some margin that is actually significant, but not huge. It is a tough call, since more games gives a better estimate and my testing does not include opening books so I am interested in results from testing with a book (and we intend to release a better book for testers soon). Based on the 20 Elo improvement (or perhaps a bit better, I would be tempted to just switch to 22.8 but continue lumping all the Crafty games together and perhaps drop the "fractional part" of the versoin for simplicity. If the next change I am working on produces a bigger jump, I will probably re-number to 23.0 anyway, and am going to do that soon as I am fixing to replace the Zobrist numbers with a static set of valeus that have been optimized with respect to pair-wise hamming distances. That is going to wreck the book.bin file anyway making a switch to 23.x necessary at that point. This will happen within 2 weeks or so...

[Zach Wegner]

Git!!! I've been using it for a while. Some say it's complicated, but I find it to be pretty simple once you know the basics. It's very powerful, fast, and unobtrusive. I have some scripts set up so that I just type a quick command, enter a line of description, and it increments a version number, branches/tags appropriately, as well as enters the version into a text database of versions to test by my auto-testing system. Works like a charm.

[Graham Banks]

[So far the testing suggests +1 Elo, so it is not exactly a world-beating bug fix.
You could continue to test the older version, or you could switch to the new version but combine the results so that 22.7 is combined with 22.8 into one 22.8 version. There's not enough difference to measure from this fix...

Hi Bob,
what do you suggest: can we do the same with the results of 22.4 (64bit 1CPU) - or its better (as difference is more than 20 elos) to start a new test series with 2.8 ?

best

http://www.talkchess.com/forum/viewtopic.php?t=25107

[bob]

What is the difference between the full features compiles and the speed compiles ?
Thanks.

With the full featured compiles, you get things like the EPD interface.
It is useful for operations against files of EPD records. For instance, suppose that you have a file of EPD test positions called test.epd. With the full featured compiles you can do this:

epdpfga test.epd test.out

which will analyze test.epd and store the results in test.out.

You can check to see how many crafty got right by doing:

epdscor test.out

For game playing using a GUI, these extra features won't add anything. They are just useful for geeks like me who fool around with crafty using the command line interface.

The 'full feature' compiles also have available the handy new 'Skill' command for chess players so you can adjust Crafty's playing strength.
Adding the extra command-line features does add a small performance hit, hence the two different compiles 'speed' & 'full feature'.

Jim.

I think that is one of the most interesting features in crafty, because it does a nice job of really dumbing the program down, but along both the evaluation axis and the search axis at the same time, so you don't get a positional dummy / tactical genius or vice-versa... I'd like to add some sort of "busy-loop" as well so that as skill goes down, that function burns more and more time in the search to slow the search depth down. But when you get to skill 10 and below, while it searches fast as all hell, it can play some real blooper type moves because of the major random factor. But it never plays total nonsense thanks to the way the tree uses random evaluations to still guide the search toward high-mobility positions.

Feedback from users has been positive, in that they don't see the NPS go in the tank, and the depths drop to 3-4 plies when they dumb it down. It does search less deeply because it throttles null-move and LMR back along with check extensions. But it still "appears" to be searching at full speed, and is still displaying PVs that look like legitimate chess even if there are flaws in the analysis/evaluation. So it doesn't "look" like it is dumbed down which apparently makes it more palatable to play at the lower skill settings.