On Thu, Nov 15, 2007 at 12:13:33PM -0800, Christoph Birk wrote:
> On Thu, 15 Nov 2007, Petr Baudis wrote:
>> This looks like a good technique I should implement too. What "big"
>> values are popular? I'm thinking size*size/3, but maybe that is too
>> conservative?
>
> If there is a capture of more than 1 stone during the random-games then
> count the number of white and black stones on the board.
> If there are more than twice as many stones of one color then
> score current board position
> If this is consistent with the winner of stone counting then
> abort the current simulation
Nice idea. The only catch is that I cannot really properly score a board
position if it has eyes of size larger than 1. ;-)
I guess the usefulness of number of games does not raise linearly? Does
it make much sense to increase the number of playouts over 100,000? Did
anyone try and can estimate how strong would e.g. 200,000 bot be
compared to 100,000? (If not, I can test it... when I unbug my code. :)
The thing is, while my board infrastructure is still order of magnitude
slower than libEgo (oprofile is my only hope now, in case it shows
something very different from gprof), it is already so fast that I could
probably do 3:00 SD 9x9 games rather easily with 100,000 simulations per
move and domain-specific heuristics. So I don't really have that much
motivation for increasing the speed now; I have actually already spent
about 40% of playout time on the heuristics.
(Currently, I have atari heuristic, cut heuristic and "play locally"
heuristic. They seem to have increased the bot's strength a lot - to
maybe about 7k KGS - though it has now the nasty habit to play out
ladders.)
>>> pure MC 10k 1050 ELO -> myCtest-10k
>>> 50k 1350
>>> AMAF 10k 1450 -> myCtest-10k-AMAF
>>> 50k 1450
>>> UCT 10k 1300 -> myCtest-10k-UCT
>>> 50k 1550
>
>>> All algorithms above are "basic" playouts. No Go knowledge, except
>>> the "dont-fill-your-1-pt-eye" rule.
>
> I will put up the bots above for reference again.
I have taken mine down but once the last version plays out few games on
KGS so that I'm sure there's no obvious bug, I'll put the improved
version back with different nicks.
>> Thanks a lot! I'm doing that now and while the ranks are not yet stable,
>> they are all only slightly above 1050 now already. :-( (Even the
>> variants with extra domain-specific knowledge.) I guess I still have
>> some bugs there.
>
> I recommend removing (temporarily) all knowledge and debug the
> basic structure.
I can do that easily in my engine, by just passing it 'pure' as a
parameter. The -pure nicks have the domain-specific heuristics turned
off.
--
Petr "Pasky" Baudis
We don't know who it was that discovered water, but we're pretty sure
that it wasn't a fish. -- Marshall McLuhan
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
