[ACT-R-users] ACT-R and Visual attention problem

[Frank Ritter]

At 09:53 -0400 18/8/10, <db30 at andrew.cmu.edu> wrote:
>--On Wednesday, August 18, 2010 3:42 PM +0900
>"-chang hyun" <y3rr0r at etri.re.kr> wrote:
>  > Dear all,
>>
>>  I have met the ACT-R recently because of my project.
>>  A partial goal of the project is to find an attention region when an
>>  arbitrary picture(such as street, subway, school, etc) is presented to a
>>  monitor.
>>
>>  My question is...
>>  1) Can the act-r be inputted a picture?
>
>No, there are no mechanisms built into ACT-R for processing arbitrary
>images.
>
>ACT-R's interaction with the world occurs through what is called a device,
>and it is the device which generates the features and objects which the
>model can see.  The provided devices allow the model to see some simple
>GUI elements (text, buttons, and lines) which are drawn using either the
>GUI systems built into some Lisps or through a virtual GUI system built
>into ACT-R.  The commands for using the virtual GUI are described in the
>experiment description documents of the tutorial units.
>
>If one wants other input or different features then it is possible to write
>a new device for the system to provide the visual components to a model.
>That new device is then responsible for parsing whatever external
>representation of the world is desired and creating the chunks which the
>vision module will use.  Documentation on creating a new device can be found
>in the docs directory in the presentation titled "extending-actr.ppt".
>
>I know that some researchers have attempted to build more general image
>processing devices for ACT-R, but as far as I know none of those efforts
>are currently available as working systems.
>
>>  2) According to tutorial 2, act-r can find an attended location. What is
>>  the criterion for finding the attended location?
>>  3) Can the act-r find an attended area in the inputted picture?
>>
>
>Sorry, I don't really understand what you're looking for with these
>two questions.  I suspect that given the answer to the first one they
>are not really relevant, but if that is not the case then please
>feel free to elaborate on what you would like to know and ask again.
>
>Hope that helps,
>Dan

Further to Dan's useful comments, there are three ways this can be 
done.  (this is reviewed in a two papers,
http://acs.ist.psu.edu/papers/ritterBJY00.pdf  and 
http://acs.ist.psu.edu/papers/bassbr95.pdf
   1)  you model in ACT-R that it will or has seen something.  This is 
somewhat like doing something in your head.

   2) you use a display built with tools included with ACT-R (the 
former /PM tools, now included with ACT-R).  This subjects can see 
and the model can see, but you have to duplicate the display.  Not 
trivial, but possible.  The tutorials have examples, I believe.  If 
you can find an example in the tutorials related to what you want to 
do, creating a model becomes much easier.

   3) you get SegMan from Rob St. Amant, or recreate it.  Segman 
connects ACT-R to a graphic display by parsing a bitmap.   We've used 
it, and papers at http://acs.ist.psu.edu/papers/  that have St. Amant 
as a co-author use it.  This would be more work, currently, but 
ultimately, I think when combined with ACT-R/PM, more satisfying.


cheers,

Frank