Re: Shallow Or Deep Search?

[Scott Dossey <sdossey@xxxxxxxx>]

On Wed, 22 Apr 1998, Dave Dyer wrote:

>
> 
> The nemesis of search engines continues to be positions where a
> solution at depth n is possible, but the problem solver is exhaustively
> evaluating a different branch where the solution doesn't exist within
> the absolute depth limit.
> 
> 
It seems to me that this problem can be mitigated quite a lot by grouping
"sets of interesting moves" as one move.

For instance you can say that any sequence of moves where the opponent
or you are in atari counts as one ply on the search tree.

Thusly, schiho's, and many other maneuvers that extend beyond the
depth of the lookahead tree fall out quite nicely without having to be
special cased.

I also believe that usually the set of all series of moves that keep the
opponent in atari is USUALLY not very large and unmanageable.

So basically a possible criteria for being "interesting" would be
"keeps one side in atari", and the entire set of moves for each possible 
sequence that satifies this condition are treated as one ply.

<shrug>  It's simple, but it seems to me that it'd be somewhat effective.
Most tree pruning algorithms place too much value on "move per move depth"
rather than what I call "perceptual depth" where easy to follow sequences
are lumped together.

I mean, it is often said that master's of the game of go look MANY moves
ahead.  This is not just a finely honed tree-pruning, but also two other
things:
 
 1. An ability to recognize that a move in one place will not effect other
    territory.
 2. An ability to "lump" groups of moves together--IE--such as schiho's.

If you can add the capability to do 1 and 2 effectively, along with doing
effective tree-pruning, you might just have a decent little go program.
Of course, influence and strategy problems are still real sticky points.
 
 -Scott Dossey