[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] interfacing to gnugo
From: |
Bjørn Magnus Mathisen |
Subject: |
Re: [gnugo-devel] interfacing to gnugo |
Date: |
Mon, 16 Apr 2007 16:46:44 -0400 |
Hi
I will admit that i`m still in the learning stages of c/c++
programming, so
i just realized that i could put the include of gnugo.h in a
namespace tag,
and all my worries about scoping and variable naming would go away :)
my current worry though is the way i rate my neural networks in their
play;
they always play black as of now. i sum of the score for each move made,
multiply it by two and divide it by the number of stones put on the
tables(number of
moves, not including passes). this number is added to the final score.
So this number will grow as black plays bader. So i subtract it from 100
so the value will go up when black plays better :)
The trouble is that my neuroevolutionary algorithm prefers values
that goes
from 0 and up, but the the fitness value above is probably from 50
and up
on a 5x5 board. I could simply subtract, but I would need to do it in
a way
so that i didnt loose any precision downwards to distinguish the bad
from the really bad :)
I think that by somehow using boardsize and komi to normalize the
values would
be better/faster than to go over every network(400+)'s score and
normalize..
The underlying reason for me needing advice in this, is probably the
fact, that I have
more knowledge on neuroevolutionary algorithms than on GO and its
algorithms :)
thanks again, btw i have reverted to 3.6, it seems somewhat faster,
but i will experiment further
with the code that interfaces to the CVS version..
regards
Bjørn Magnus Mathisen