I will be restarting this study from scratch due to the discovery that
some results are being incorrectly scored due to a bug. (The
autotester for chess is a new product and hasn't been fully debugged.)
I also noticed that the internal rating procedure was not working -
so the matches were being scheduled randomly, which doesn't hurt
anything but is very inefficient.
The same URL will work and anyone interested will be able to follow the
results as the tests proceeds. There will an interruption at some
point as I intend up update my OS but it will then continue.
- Don
Don Dailey wrote:
> A strange place to post this, but the chess study is well underway and
> I have some results.
>
> I have a temporary web page up similar to the go scalability study
> pages. The site is temporary and will probably come and go since I
> will be upgrading my OS soon and I'm running this from my home internet
> connection. It's not very well documented but hopefully enough that
> you can understand and verify/reproduce the results if anyone wants to
> bother.
>
> If someone would like to provide a link to one of the computer chess
> groups, feel free.
>
> http://www.greencheeks.homelinux.org:8015/
>
>
> I decided against using a strict round robin, although eventually this
> would become a round robin. It's clearly a waste of computing
> resources to test a depth 1 search against a depth 9 search. My
> tester uses 200 openings to provide variety which provides up to 400
> games per matchup. The tester ensures that all games between 2
> players are unique. The scheduler favors games between players of
> similar strength, but it will not exceed the 400 games limit and would
> thus revert to a massive multi-game round robin if allowed to run long
> enough.
>
> The lines on the graph seem to indicate that stronger evaluations
> functions improve more with depth.
>
> Here are net differences between the weak and strong evaluation
> functions by depth, after adjusting by the depth 1 difference. You can
> view it as the amount of extra ELO improvement (positive or negative)
> for the strong evaluation over the depth 1 version.
>
> Depth Growth
> ----- -----
> 1 0
> 2 6
> 3 39
> 4 46
> 5 76
> 6 85
> 7 62
> 8 74
> 9 77
>
>
>
>
> - Don
>
> _______________________________________________
> computer-go mailing list
> [email protected]
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
