RE: computer-go: Fast scoring program?

["Mark Boon" <tesuji@xxxxxxxxx>]

Sorry to pop in with what may be a trivial question. I haven't followed the
whole discussion but I just read this one.

>From the little that I remember about NNs I think there are estimates to the
amount of things it can learn based on the number of nodes used and the
number of connections between them. 325 nodes with what looks like at most
several thousand connections seems like a very small number to me to learn
anything but the most trivial task.

So my question is: what exactly do you expect the NN to learn to do?

    Mark Boon

> -----Original Message-----
> From: owner-computer-go@xxxxxxxxxxxxxxxxx
> [mailto:owner-computer-go@xxxxxxxxxxxxxxxxx]On Behalf Of Matthew Corey
> Brown
> Sent: Tuesday, August 14, 2001 11:29 PM
> To: 'computer-go@xxxxxxxxxxxxxxxxx'
> Subject: RE: computer-go: Fast scoring program?
>
>
> On Mon, 13 Aug 2001, Grajdeanu, Adrian wrote:
>
> > I've used gnugo for the same purpose as you, Matthew, but I
> gave up on it.
> > Not only it took too long, but sometimes it would never finish scoring a
> > position. So I gave up and implemented the TT scoring. It is not exactly
> > what I want to have eventually, but I figured that I can run an
> experiment
> > in training the NN with TT scoring and if all goes well, I will swap a
> > better scoring engine later. So for now I have something that
> runs fast and
> > gets me ahead. Unfortunately, I tend to think that my nets
> aren't evolving
> > the way I expected. Either I have some more bugs in my system,
> or I am just
> > plainly digging in dry land, I am not quite sure yet. Need more time to
> > figure it out. How's your experience w/ training Go playing nets?
> >
> > Adrian
> >
>
> Not well.. but i am learning lots about neural nets and GA, my latest
> network is the fastest but has a harder time learning when to pass. If it
> wasn't the capibilty for passing in Go it will be easier.
>
> My current architecture is an 81 input 163 (N*2+1) hidden layer and an 81
> output layer, it checks with gnugo (as its the referee program) all legal
> moves. Sorts those moves by the value of its output node and if the top
> legal move (checks super ko rule also as gnugo doesn't) is move then 0
> then it takes it. otherwise it passes. But of course a random net you
> would expect it to pass alot from the beginning. Also if its an
> empty board
> i fill the C3 position with the enemy on the network, cause an empty board
> will get all 0 out puts in the net.
>
> This net is 30-60 times faster then my old way as instead of one net
> computation for each possible move. its just one net computation per
> actual move. But it will take a longer time to evolve i think. And I've
> been playing with how to deal with passing early on. They tend to evolve
> towards always passing the first move as a score of 0 means they are dead
> in the middle. Its only by keeping the mutation rate high enough, that i
> can keep it evolving to a positive score.
>
> Its the fitness function thats the tough problem here, what's the
> trend you
> are having? I've played with deducting points on passing depending on
> empty spaces (was score at that time, but that was too slow) but i rarely
> got above 20 moves per game and averages 8 moves.. way too low on a 9x9
> board. Maybe i can use gnugo's scoreing function to my advantage and
> subtract the time it takes to score a game from the final score. On the
> asumption if its  a long time to score then there is a random
> distrabution. And maybe subtract the time from the loser only. Hmm
> actually thats what i'll do have nothing to lose at this point...
>
>
> Matthew Corey Brown                             bromoc@xxxxxxxxxxxxxxxxx
>                   Happiness is a dry place to live.
>
>
>