Visual search & finsts

[Mike Byrne]

ACT-folk,

There's a lingering issue with RPM's Vision Module that has come up
now and again and I'd like to get people's thoughts.

A fairly high proportion of the RPM-intensive models out there have
fairly substantial visual search components.  In some cases, this
amounts to an attempt to NOT have a substantial set of ACT-R
productions to manage the visual search component.

The issue resolves around the number of items that can be marked as
having been attended.  I borrowed Pylyshyn's term and these markers
are called "finsts."  Pylyshyn's work suggests that the appropriate
number of these things is four, and that's been the default setting
in RPM.  (I'd like to note that recent work out of Jeremy Wolfe's
lab seems to suggest that there should be none of them whatsoever.)

However, four doesn't seem to be an adequate number for many
models, which causes a lot of people to set the :VISUAL-NUM-FINSTS
parameter to something much higher, like 10.  Or 100.

I can't find anything in the visual attention work that supports
this kind of idea.  Eye-tracking data we've collected in my lab
suggests that people certainly do re-visit seen items from time to
time, which ought not to happen if 100 is a legitimate number for
this parameter.

So, my question is, what breaks such that this number really has to
be so much bigger?  I suspect that one reason people have trouble
with this number is because people write productions in such a way
that they depend on side-effects of a high finst number to do
useful placekeeping work for them.

For instance, in reading a series of words from left to right, one
could use the following call to find the next word:

+visual-location>
   isa          visual-location
   attendend    nil
   screen-x     lowest
  
This works in moving from left to right only if the number of
finsts exceeds the number of words.  If the number of finsts is
exceeded, the leftmost word becomes the item with the lowest
screen-x that has not been attended, so visual attention will snap
to the left word partway through reading.  Obviously, that's not
the behavior one might want from RPM and certainly isn't what
subjects are actually doing.  Thus, to defeat this, the number of
finsts has to be cranked up.

However, it seems to me that there's a better way of doing this
that should both be consistent with visual attention performance
limits as I understand them and should also yield the correct
behavior:

+visual-location>
   isa          visual-location
   screen-x     greater-than-current
   nearest      current
   
[Note:  You could add "attended nil" in there if you wanted but it
would make little difference.]  This will find the next thing to
the right of whatever is currently being attended.

Overcoming problems created by limiting the number of finsts is why
those relative-to-current-location terms were added to RPM in the
first place.  

I suspect that most problems that are now solved by raising the
number of finsts can actually be solved by putting a little more
knowledge into the search control productions, but I might be wrong
about this.  

Those of you who have built models that do even limited visual
search, please let me know what you think.  Do we really need more
than four finsts?  If so, what are the conditions under which more
than four are needed?

Thanks,
-Mike

P.S. to Mike Schoelles:  I really want to hear from you on this,
regardless of how you set this parameter.  You have interesting
data on re-visitation that you've modeled...

===========================================================
Mike Byrne, Ph.D.                             byrne at acm.org
Assistant Professor, Psychology Department
Rice University, MS-25          http://chil.rice.edu/byrne/
6100 Main Street                      +1 713-348-3770 voice
Houston, TX  77005-1892                 +1 713-348-5221 fax