On 31-jul-08, at 21:00, Peter Drake wrote:
What you're missing (or so it seems to me) is that it's not to
prevent from running a ladder that is caught.
Really? My motivation has been to prevent my program from
embarrassingly running in just those situations. Is there something
other than a ladder reader used for this?
The embarassing escape from a ladder that doesn't work is a side-
effect resulting from the horizon problem. It's not isolated to
ladders but a general problem of which ladders are the most easily
recognised form. What happens is that a losing position is made to
look better by playing forcing moves that push the moves that make it
see the position is bad over the horizon.
Using a ladder reader does help in this case, as at least it prevents
the ladder from escaping during playout when it shouldn't. So the
ladder reader makes the MC evaluation more accurate but does not
prevent the horizon problem entirely.
It is to ensure to escape when you can or capture when necessary
to prevent an escape. And not only the last stone played of
course, it could be the neighbour of the last stone played as well.
The neighbor point is useful. Of course, as Don pointed out in
another message, there are always additional complexities to add --
what if one of the attacking groups is in atari, or can be put in
atari?
I'm more interested in the bigger issue: exactly what question is
the ladder reader trying to answer? When does it suggest a move and
when doesn't it?
It's simple really. If a big chain is put into atari and it can
easily escape, the MC playouts will still only give it a 50-50%
chance of escaping. This distorts the MC evaluation from what it
should be ideally. So instead of leaving the escape up to chance, the
ladder-reader makes sure that the escaping-move is always tried when
it has been read out it can indeed escape. Whenever a chain is put
into atari, the ladder reader checks if it can escape and suggest the
escaping move instead of a random one.
In case a (random) move is played where a chain was previously in
atari but now has two liberties, the ladder-reader will see if the
chain can be captured and suggest the capturing move instead of a
random one.
The neighbours of the last move come in the picture because usually
it's only the last stone played that can be escaping a ladder and
it's the neighbours of the last move that could have been put into
atari. Nothing to do with the additional complexities Don mentioned.
Hope this helps.
Mark
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
