[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: computer-go: Proof number search



 This is a problem with all types of search.
If you do not visit all the oponent responses then you
can not be certain that about a capture.

 The point I was trying to make was that

 You CAN use heuristics to direct a PN type search. There seemed
to be a perception that you had to ignore any move ordering
information when using PN. This is not the case (maybe this idea
needs/has a different name ?)

 By using ordering you are more likely to visit a node that will
will tell you that there is no point visiting it's siblings. Thus
saving work.

 For your information I do not build a complete tree with the
set of "better" moves. Each node has proof numbers associated with types of
move (like a child node waiting to expand).
By using higher proof numbers for more speculative moves
you avoid visiting these until the alternative branches reach
a level of complexity greater than the assigned value.

 This gives you a search that tries speculative moves when the more
hopeful ones lead to complex (wide) searches. One of the reasons
I implemented a PN type scheme was because I could not see how other searches
could achieved this type of effect.


cheers Paul.





Chris Fant wrote:
 I believe that PN combinbed with heuristics is the best way foward
for solving GO problems. But your proof statement should be in terms
of heuristics goals that prune the search. (e.g. can I capture using
only the suggested moves with a heuristic greater than a given value).
This initially gives you a result that can be refined by lowering
the value to expand the tree.

The thing I don't understand about that is:  If you are only using the
suggested moves, how can you be sure that you do actually capture.  What it
the opponent has a move that was not suggested in the search that can
prohibit the capture?