[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: Perl Module for next move.
On Mon, 4 Jun 2001, Daniel Bump wrote:
>
> Matthew Corey Brown wrote:
>
> > looks like GTP will be the way to go once all the bugs are
> > out. Documenting one bug when you do all_legal on an empty board.
>
> Could you send a bug report to gnugo@xxxxxxxxxxxxxxxxx, please?
I will once i have everything together some time in the next couple of
days.
>
> > it appears the GTP will do score and everything I'll be using
> > that. Hopefully there is a machine readable showboard before gnugo goes to
> > 3.0 I can wait, i'm interested in the network and evolution of them more
> > then go itself.. I know sacriliage, but i have a co-worker that is really
> > into go so i figure i'd turn my AI learning to benefit him.
>
> Are you trying to get a score at the end of the game or
> earlier?
>
At the end of the game is where i need the Go score for the fitness
function. Well the end game Go score is the fitness function. The
showboard is so i can try every legal move and get the results of those
to pipe through the net to see which is the best move to then actually
make/commit a move. Then the next neural net plays the other player doing
the same thing. Each net will play as white and as black against all the
other nets, add up all the scores and the top 50% get to pass their
"genes" to the next generation.
A simplified version of what i plan in the evolution is.. 5 random nets
they all play against each other and the top 2 get crossbreeded to produce
2 offspring. the top one gets cloned. The offspring, the clone and the
number 2 position gets mutated, then the nets go through "natural Go
selection" again to see whom gets to pass on there genes.
in actuality i will probably use 15 nets for each generation
another way to do it is just to have each net only play against gnugo
until the the top 50% always wins then have the nets play each other. But
this way will not be what i'm after. which is "Can the computer learn to
play go well with no human input beyond the rules?" Playing against gnugo
you actually getting human input cause its a human trained/coded
AI.
Matthew Corey Brown bromoc@xxxxxxxxxxxxxxxxx
"Death can not stop true love. All it can do is delay it for awhile."