[Computer-go] Gomill - GTP strength comparison and parameter tuning tools

Matthew Woodcraft matthew at woodcraft.me.uk
Tue Nov 16 07:30:16 PST 2010


Petr Baudis wrote:
>   Some people might also be interested in:
>
> 	http://repo.or.cz/w/pachi.git/tree/HEAD:/t-play/autotest
>
> (It's included in Pachi source tree, but is absolutely independent.)
> Its main advantage is the ability to very easily start off testing
> instances on many machines sharing the same NFS volume. But it's perhaps
> somewhat less user-friendly.

I think farming jobs out to multiple machines is an important feature
that gomill is missing.

But I'm not sure what the most useful setup would be: For example, I
don't know whether it's a good idea to rely on a shared filesystem, or
on having a master machine with ssh access to the others, or whether it
would be a better idea to let the worker machines run independently and
ask a central server for 'work units' to perform.

So if anyone on the list has suggestions for what would be useful for
them, feel free to make them.



>>  - automatically tuning engine parameters based on game results
>>    (EXPERIMENTAL)

> This is very interesting! I'd be intrigued to know how well this works
> for you in practice, e.g. how many games are required for basic tuning
> of few [0,1](R values.

Alas, one thing I have learned is that none of the tuneables my program
supports at present can make it noticeably stronger. The results were
too 'flat' to give much information about the tuner.


I had a go at tuning Fuego's rave_weight_initial and rave_weight_final
parameters using the MCTS tuner, splitting each parameter 5 ways at each
level of the tree.

After a few hundred games, its preferred candidates were perfectly good
ones. But even after a few thousand games it didn't really converge:
there were plenty of relatively unexplored bits of the tree that would
probably be just as good as the ones it liked.

As with MCTS for choosing moves, if you're mainly interested in finding
a good combination rather than the theoretically best one, I think it's
quite practical.


I think to make more progress with the tuner I need something more
sensitive to tune. If you have any suggestions to try for Pachi, I'd be
happy to give them a go. November here is a good time to be running
processor-intensive work; the heat generated does not go to waste.

-M-



More information about the Computer-go mailing list