[Pachi] pachi + dcnn
pasky at ucw.cz
Tue Mar 29 16:10:54 CEST 2016
On Tue, Mar 29, 2016 at 03:37:54PM +0200, lemonsqueeze wrote:
> So far i've tested with 180k simulations, 16 threads:
> pachi -t =180000 threads=16,resign_threshold=0.30,pass_all_alive,pondering=0
> -f extra/opening.dat
> With the opening book it's always playing the same opening but speeds up
> testing =)
> No idea what the right dcnn_eqex might be for different settings, i've only
> tested with this so far. Even there it can still probably be improved (1400
> looks good too, 1200 & 1500 definitely worse ...)
Awesome - I expected testing with much weaker configuration. :) In
that case, these priors are probably pretty good.
> The github branch is a cleaned up version of the tree i'm using for testing,
> hopefully i didn't screw up (haven't tested it yet).
> One thing that's missing in it is PR#26 (check group suicides), it can
> assert out in the endgame otherwise. Also right now dcnn disables pondering,
> would be nice to find a way to make it work again for tournaments.
Oh, I didn't notice that. Would be also good to document.
> I don't have a gpu so hard for me to experiment, but sure would be nice to
> have support for that ! Wow, it would really start to look like darkforest
> then =)
> I'm thinking even multi-threaded cpu support might be good enough to start
> using it further down the tree...
I guess enabling GPU support in caffe per se should be pretty easy.
The more difficult part is implementing asynchronous updates of the tree
after expansion. But that could be also useful for pure CPU mode,
dedicating some threads to just this as you mentioned.
I'm sure we can put all these pieces together. As I said, I'll be
happy to merge this version already as an initial prototype support, as
soon as we also include some documentation (mainly where to get the
models and where to place them).
> Btw do you know what's going on with monkey jumps ?
> For some reason it seems it never plays them, i suspect there's something
> strange going on... Playout issue maybe ?
Monkey jumps are difficult for programs like Pachi, because they are
far away from your stones (which makes for non-enticing priors) and the
pattern required is not circular but long rectangular one. I'd say most
MCTS programs will have the same trouble, and it might need some
special-casing - if DCNN doesn't learn them, which I'd really expect
If you have good ideas, good data and fast computers,
you can do almost anything. -- Geoffrey Hinton
More information about the Pachi