Re: computer-go: A problem with understanding lookahead

[Don Dailey <drd@xxxxxxxxxxxxxxxxxxxxx>]

> The first few chessprograms were also very easily to fool
> by feeding them material for obvious positional and strategical
> shots.

Really, it's still the same. This "feeding" strategy still works, it's
just that  the program now see it   farther in advance, and  it's much
harder to get away with it.

> However real soon the selective tactical search depths of chessprograms
> were like 12 to 15 ply.

The truth of the matter,  is that you can subtract  several ply off of
this to get true full width depth.  Even 15  ply FULL WIDTH isn't much
because this does   not see a  slow  attack building,  or a  weak pawn
getting taken.  Sometimes a weak pawn must be lived  with for the most
of the game!

I view modern chess selectivity as just a fancy quies search on top of
a full width search.  It's not called that,  but that's really what it
is.

I think eventually that  will be done  in Go.  It may  be a  long time
before it even  makes sense to think  of full width  layers in Go, but
there comes  a  point where  you cannot   squeeze anything  else  out,
because increases in selectivity equals increase in risk.  You reach a
point of diminishing  returns at some  point  and the only  known sure
thing is to add a full width layer.  (They used to say in chess, "full
width is never having to say I'm sorry!)

I really believe the thing that has to be  worked on right now is lots
of knowledge engineering,  and if we even  do searching, we are really
exploring selectivity near the leaf nodes and trying to see how far we
can push this.


Don