Re: Shallow Or Deep Search?

["P.J.Leonard" <P.J.Leonard@xxxxxxxxxx>]

Robert Jasiek wrote:
> 
> [Let's get rid of the recent silence and boring ping question marks!]
> 
> In heuristic tree search based on a static evaluation function
> that is rather good it is sometimes best not too search at all
> but to apply the evaluation function only.
> 
> I guess, you have made some experiences when to use shallow
> or deep search, why, and how deep. Which is best in which cases?
> I am eager to listen to your experience and learn from your advice!

 Putting all your effort into the first move is often recommended 
in GO problem books. It is my humble opinion that at present this is
where the most gain is to be made in computer go programs. Since most
searches explode if too many moves are evaluated a program can be
improved
most by putting a lot of heuristics into the static evaluation of a
problem.

 Long term I think the situation will change and the need for accurate
deep
reading will become the deciding factor. Already deep searches are
required
to confirm things like "can my cutting stone live" and life and death
problems. 
It is these type of problems that I have investigated in the pubgo+
program.
In particular I favour the PN search technique because it makes no
assumptions
about the depth of a search. If you have competing subgoals within in a
search
tree then it will solve the easiest (least resources) subgoal first.

 Localising your goals and detecting when that localisation becomes
invalid during
the search (and what to do about it) is very difficult.

  cheers Paul.