On Fri, Jun 17, 2011 at 4:26 PM, Jean-loup Gailly <[email protected]> wrote:

> > One advantage however to involving a lot of people and their computers is
> that
> > I COULD do the study out to enormous numbers of playouts,  given enough
> help.
>
> Going beyond 8M playouts/move would only be useful, in my opinion, if we
> can test
> stronger programs. It's not as bad as self play, but you don't learn much
> from beating
> a weaker opponent more than 95% of the time. I would love to be able to
> play thousands
> of games against Zen, which is currently about 2 stones stronger than
> Pachi.
> Also it's not enough to double playouts you must also double memory usage,
> and you
> may have trouble finding enough machines with terabytes of ram.
>
> > I would probably have multiple instance of EACH program,  not a single
> fixed
> > opponent for one program but that does require a lot more games.
>
> Same problem. Beating multiple weaker programs won't tell you much. We need
> multiple stronger programs.
>

I agree,  so I don't believe round robins should be played,  but my tester
could be rigged to play swiss system tournaments.   Swiss would be naturally
balancing while still allow some mismatches.    If we have many levels that
span an enormous range,  then the first round of each swiss will still have
a lot of mismatches, perhaps more than we are willing to accept.    However
this can be solved by breaking the tournament into multiple sections by ELO,
for example 4 separate tournaments.    After all 4 tournaments are complete,
  each players rating is updated and we begin again.     There would be
plenty of crossover doing it this way or if we didn't think so we could
occasionally have a big swiss with all players to mix things up.

I don't think we need a lot of players,  just 2-4 of the strongest ones we
can get.  It's counter-productive to have too many, especially if the
weakest is too weak to compete without setting the level really high (it's
just a waste of testing resources.)    Each program is split into N players
where N is the different level settings,  each (perhaps) 2x more playouts
than the previous.   It could 1.4142 times more playouts (so that 2 levels
is 2x) or any factor we choose to get where we want to go.

I personally think just having 2 programs is best with a dozen levels for
each program or so.

I don't really understand how big a factor the memory issue is.   If a
program does not have enough memory to keep the entire tree in memory,  how
quickly does the strength degrade?

Don





>
> Jean-loup
>
>
> _______________________________________________
> Computer-go mailing list
> [email protected]
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to