[Pachi] Better semeai handling ?

Petr Baudis pasky at ucw.cz
Mon Apr 25 17:49:34 CEST 2016


  Hi,

On Mon, Apr 25, 2016 at 02:42:06PM +0200, lemonsqueeze wrote:
> I've been wondering how to improve semeais:
> 
> I did some experiments reusing hashed moves from the tree in playouts as in
> the alphago paper, it seems to work but only with early phases of the semeai
> as the tree doesn't go very deep into it usually. So unless playout policy
> can follow up on it doesn't seem to help much (or we'd need the tree to read
> the whole sequence out somehow to have the moves in the hash table ? doesn't
> look straight forward though ...)

  yeah, not too straightforward.  I wrote a Masters Thesis on this. ;-)
But without a large success.  In general, you might expect that the
hashing should however help with the horizon effect anyway...

> I'm playing with a moggy unit test to see what's going on in the playouts
> (t-unit/moggy_semeai.t in my moggy_test branch on github). Looks like it
> gets in trouble as soon as it's more than 2 liberties race:
> 
> X . O X . O
> X . O X . O
> X . O X . O
> . X X O O .
> . . . . . .
> . . . . . .
> 
> So moggy could definitely use some help. The thing i don't understand is how
> it can work for alphago. If i understand correctly their fast policy uses
> probability distribution from pattern features as in Remi Coulom's 2007
> paper, which is also what pachi's pattern.c implements and uses for priors,
> but not for playouts. (I see there was once a playout/elo.c which did that.
> Just wondering, what was the problem with it ?)

  I never got it to work better than Moggy, pure and simple.  Not even
if I gave it equal #playouts (it's quite slower).

> It seems most of the features they use are also in moggy, but there must be
> a big difference also (i doubt moggy gets 24% prediction rate =)
> I guess this must be this one which we don't seem to have :
> 
>   "Response pattern"   Move matches 12-point diamond pattern near
>                        previous move
> 
> So something like this ?
>       *
>     * * *
>   * * . * *
>     * * *
>       *
> 
> Maybe their definition of 12-point / 3x3 pattern is different too ?

  I don't think it's anything but what you have shown. It should
correspond to our d=4 gridcular patterns.

>   "Patterns are based on stone colour (black/white/empy) and liberties
>    (1, 2, ≥ 3) at each intersection of the pattern."
> 
> I don't understand much about the pattern code, but it doesn't look like we
> take liberties into account for 3x3 patterns, right ?

  Nope, just atari status.  Long time ago, I did a quick experiment with
machine learned rather than handcoded patterns and I don't remember the
result clearly anymore, but it couldn't have been a big success.


-- 
				Petr Baudis
	If you have good ideas, good data and fast computers,
	you can do almost anything. -- Geoffrey Hinton


More information about the Pachi mailing list