[rc5] distributed chess

JP Rosevear webmaster at usjc.uwaterloo.ca
Fri Oct 24 16:31:46 EDT 1997


>On Thu, 23 Oct 1997, Alan R. Riley wrote:
>> How about chess.distributed.net vs. Deep Blue?
>
>I think that typical chess programs are intensively iterative, with
>results feeding into the next calculations. This means fat pipes, and no
>progress until your predecessor makes progress. Thus far, we've been
>dividing up a haystack into buckets, and everyone searches their bucket
>(or thimble) for the needle.  Chess is more like everyone taking a whack
>at a block of granite to make a statue. If we all beat on separate blocks
>once, no likeness.
>
>If all possible combinations of the next "n" chess moves from a given
>position onward could be exhaustively described mathematically (spanning
>tree?), they and their children could perhaps then be divided up among

This is a good idea.  It may be possible for the servers to generate only
the positions that need to be checked (or blocks of positions) and have
clients perform the actual evaluation, in addition to capture and check
searches.

>many clients for scoring.  'Course, some of the clients are at the end of
>slow and/or unreliable links, some will respond too slowly, and some will
>never respond. I think it would have to be hierarchical, a la neural net.
>This implies a certain amount of trust that you *will* get a response, or
>the tree gets incorrectly pruned. Or you're smart enough to give up and
>ask another client that has a track record of responding.

Replication maybe useful in increasing the likelihood on response.  I mean
a 3-4 fold decrease may sound like a lot, but for a search space which is
in vast and processing another 2ply means billions more positions, it
doesn't mean a heck of a lot.

>Sounds like a bitchin program! I'd be very surprised if there weren't
>massively parallel chess programs that might be used as starting points.
>On today's 'Net, it might make a killer 'chess by mail' program. I don't
>know about when the chess clock is running...

The person who would be good to contact would be Robert Hyatt at UAlabama
who has programmed Crafty (for many platforms, source code available) as
well as CrayBlitz, a succesful massively parrallel processing program.

One other possibilty is that distributed could try to solve chess for once
and for all (all though I think it would have to use a slightly better
approach than brute force).

-JP


----
To unsubscribe, send email to majordomo at llamas.net with 'unsubscribe rc5' in the body.



More information about the rc5 mailing list