[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: Learning from existing games
On Mon, Jan 13, 2003 at 05:37:35PM +0100, Frank Steinmann wrote:
> My question: To analyze a game, I'd like to evaluate the moves, that have
> been made in that game. The simplest way to do that, is to give every move
> the value of the game result (positive for the moves of the winner, negative
> for the moves of the loser).
That would not work. If you only study professional games, or other games
played by people much stronger than your program, you could assume that
every move was nearly perfect in the position, no matter who played it.
In other words, the chosen move is (almost) guaranteed to be better than
(almost) any other move on the board.
I would rather evaluate positions than moves. Then I could use something not
unlike TD-Learning, where I could know t he score of the final position, and
assume that the beginning (empty board) is even, and interpolate from those
two.
There is lots of interesting potential in learning, but I have the feeling
that the state of art in Go programs is too weak to take much advantage of
it yet - trying to imitate strong players without any underlying
understanding may produce a few good-looking moves, but if only 1 move out
of 20 is a horrible blunder, there is not much hope for a good game.
Best of luck, anyway. I would be delighted to be proven wrong in this.
-Heikki
--
Heikki Levanto LSD - Levanto Software Development <heikki@xxxxxxxxxxxxxxxxx>