[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] Toying with ideas
From: |
Heikki Levanto |
Subject: |
Re: [gnugo-devel] Toying with ideas |
Date: |
Tue, 22 Jan 2002 00:33:23 +0100 |
User-agent: |
Mutt/1.2.5i |
On Mon, Jan 21, 2002 at 02:22:48PM -0800, Daniel Bump wrote:
> Where would these piles of #ifdef's likely be? That is,
> what files do you imagine would need modifying and what
> would be the nature of the modifications?
Sorry, I should have explained my project before asking.
The idea is to use a neural net to evaluate whole positions, based on a
great number of parameters gnugo is extracting from the position. Once I
have a working model of the positional evaluator, I can take the top N moves
suggested by gnugo, try them out, extract the features again, and evaluate
the positions, choosing the most promising move.
The changes required fall in three groups
1) Extract a lot of features from a position, typically of a loaded game.
The actual extraction can be in its own file, but there will be some
command-line options and a way to call this routine for every position in a
game.
2) The whole neural net (in its own file(s), again with a few command line
options. Also, at this point I need a training program that reads those
feature files and exercises the network.
3) Evaluating resulting positions. This is may require heavier surgery into
gnugo, as I want to try out moves and get all initial information about them
(worms, dragons, lifes, connections...), but do not need proposed moves at
all, nor do I want to pay the price of generating move candidates.
I think you are right that a private fork is the way to start, at least
until step 2 is shown to work. Like I said, this may take some time, as my
availability flunctuates a lot. I might even consider doing the whole thing
in the other end of a gtp pipe, not touching gnugo at all, until much later.
- H
--
Heikki Levanto LSD - Levanto Software Development <address@hidden>