Re: Learn Alternative Moves

[Jens Yllman <jens.yllman@xxxxxxxxxxxxxxxxxx>]

Hello,

 I'm working on a program based on Artificial Neural Networks(ANN). Or at
least I'm trying to. Have not had that much time lately. But anyway as I
see this there is never one good/resonable move in one position, except
maybe near the end. And my current ANN go program consist of one(1) ANN
that suggest moves and one(1) that evaluates the board. The ANN that
suggest moves can suggest more then one move. So when I train the ANN I can
give more then one(1) good/bad move in one(1) sample. So what I do is to
merge samples that have the same boardposition. What I also try to make
posible is to get the ANN to understand if I know or don't know if a move
is good or bad. This I hope will help when moves my cancle out each other.
So far I have not had time to test this so much. I can see that the ANN
learns. But I don't have any realy good results yet. And right now I'm
about to redesign my ANN code. It is not flexible enougth. I want to try a
couple diffrent learning methods. And I also want to try diffrent
activation functions. But my schedule does not look to good. If you want to
monitor my progres, visit http://www.uniweb.se/~jens/. There under
Games->GO there is a link to my GO projects. Hopefully I will have time to
update that soon. Hopefully. Right now there is no info about how my
program/ANN is structured and how I teach it. When I have time I will write
some kind of paper on this. But first I need some usefull results. I would
not mind discussing this in more detail if anyone is interested. I hope to
get some ideas from others that is working on this too.

 Jens Yllman

At 10:39 1998-04-26 -0700, you wrote:
>    Hi, there,   Is here some one experienced alternative move problems in
>learning  algorithm? Say as if I have a Go classification machine which is
>taught to  identify board configurations with expected moves. One sample
>said with board  configuration A should move to position a, and another
>said board A, move b.  While a neuronetwork is trying to learn this kind of
>response, (A,a) will tend  to cancel (A,b) and (A,b) will cancel (A,a).
>Here all the samples are  distinctive, so no probability mechanism can be
>applied.
>
>Weimin  Xiao
> 
------------------------------------------------------------
Jens Yllman                       http://www.uniweb.se/~jens