Re: [computer-go] Chains and liberties

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

> First of  all, your system appears  to be quite  impressive.

I am stunned by its performance. It now predicts 46% of that particular
game, a rise by 1.7% (the system is still learning, it is not even halfway).

At move #5, the performance is about the same, so it learned how to order
its top-5 moves better.

I can slap myself for having wasted so much time on fruitless
implementations but at least there are results now..


> If you  ARE going to  use pro-prediction (which  I don't think  is all
> bad) then  it makes  sense to make  it work  as well as  possible.  My
> comments were based on the observation that this is a lot of "slop" in
> the  samples.  Your  results could  vary significantly  even  based on
> which players you choose to predict!

Very much so, yes.
When I started these experiments a few years ago, I was lead astray by that.
What used to happen was I fed the system many games, and saw the prediction
average slowly rise.

But then it dipped again, and kept going down for tens of thousands of
games.
because that learning took many days, I always interrupted the process and
tried a new learning algorithm.

But when I came to greatly speed up the learning and let it run on 500,000
games, I discovered that those dips were only temporarily and due to a chunk
of lower-ranking games.


> To my  surprise, the program predicted the  master moves significantly
> better when pawn structure was turned off.


Would you believe it, my pattern system performed much better when I did not
remove captured stones..
I never discovered why.
I am unwilling to test this with the current system :)

But such weirdness is often an indication for a bug in the program and when
you fix it, the performance will improve, sometimes dramatically. I had
myriads of nasty bugs and when I fixed them all, I ended up with this. I
never really changed the system much and I would have had a bit less
performance if I would have stuck with the original learning algorithm, it
would only have taken much more learning time. Because I decided to re-wire
the learning method, I found a massive bug yesterday and fixing that more
than doubled the prediction rate.


> >  Well, the purpose of the pattern system is mainly to be an expert
system for
> >  Fuseki, Joseki and "Good Shape".
>
> I hope it works well for that.


Well I hope to make a Shareware version available ASAP but there still is
loads of work to do.

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/