Re: computer-go: Learning from existing games

[A van Kessel <Adriaan.van.Kessel@xxxxxxx>]

The reward/punish method has been tried before, take
for instance a look at the papers by kojima (sp?).
The real problem of  the approach is the context of the move.
If you want your program to *remember* the move,
it has to remember (part of) the surrounding stones/edges/
liberties. Which part ?
If you remember the whole board, you will never be
able to use the position you stored: it will (allmost)
never occur again.
If OTOH you just store the four surrounding intersections,
you will collect nonsense, and your program will play like
a drunk spider.
Most programs do use some kind of stored knowledge/patterns,
but use it just to give hints to the movegenerator, or to
direct moveordering. I don't think anybody will use stored
moves without subsequent reading. (except for fuseki, perhaps)

Also, if you want to store moves, you should store their
intention/goal, too. fighting moves are different from
quiet moves, opening moves from endgame moves, etc.

If you could defend a group by playing bad shape, what would
you do ?
AvK