unconstrained retrievals

[Jerry.Ball at williams.af.mil]

model.  To control retrieval of previously processed words, I added a status
slot to the declarative memory chunk created when a word is first processed.
The status slot took on the values "used" "pending" or "available".  During
a retrieval, only DM chunks with status "available" were retrieved.  This
worked OK, but it essentially meant adding processing information to
declarative memory chunks.

There is actually a more general declarative memory retrieval problem with
respect to retrieving the most highly activated DM chunk without regard to
the type of the DM chunk.  I was unable to come up with a solution to this
problem other than forcing all chunks to be of the same type.

My current solution uses a circular stack (limited to 7 elements) to
maintain the prior context and to avoid use of the status slot.  The chunks
on the stack are used to do a retrieval from DM.  That is, they provide a
template for the retrieval (including the chunk type).  I view the stack as
a set of activated links to DM chunks.

In the absence of a goal stack in ACT-R 5.0, there appears to be a tendency
to use DM chunks to achieve similar effects.  That is, the slots of the DM
chunk store goal (or subgoal) chunks that would have been put on the goal
stack in ACT-R 4.0.  The circular stack can be viewed as a variant (and
hopefully an improvement) on this theme.  

Jerry

-----Original Message-----
From: Wolfgang Schoppek [mailto:Wolfgang.Schoppek at uni-bayreuth.de]
Sent: Wednesday, June 05, 2002 2:21 AM
To: ACT-R
Subject: unconstrained retrievals


I'm working on a model that performs "competitive chunking"
(Servan-Schreiber & Anderson, 1990) in ACT-R 5.0. In that model, I use
something like unstructured working memory that contains elements of
the stimulus in arbitrary slots. An example is:

 **Newgoal25    3.006
    isa WM-FREE
    s1 A-0
    s2 B-0
    s3 nil
    s4 D-0
    s5 1-0
    s6 2-0
    s7 3-0
    s8 Pattern15  (represents the former contents of s3 and s8)
    count 7
    ret nil
    task 'learning

In declarative memory, there are chunks ("patterns") that may
represent all possible subsets of the set of elements of Newgoal25.
For example, a pattern like:

 Pattern2    0.251
    isa PATTERN
    pa nil
    pb B-0
    pc nil
    pd D-0
    p1 nil
    p2 nil
    p3 3-0
    p4 nil

would match, but there are lots of other patterns that might match as
well. Because of the large number of possible combinations, I don't
constrain the retrieval of patterns with productions like:

(p retrieve-pattern
  =goal>
    ...
 ==>
 +retrieval>
   isa pattern
   pb B-0
   p3 3-0
...)

Instead, I use unconstrained retrieval that simply returns the most
active chunk:

(p retrieve-pattern
  =goal>
  ...
==>
  +retrieval>
    isa pattern
  =goal>
  ...
)

Now the problem with this kind of retrieval is that it often returns
chunks that are already elements of the goal (because of the strong
self-activation of chunks), which is of course undesired in the
present context. Does anybody know a good solution to that problem or
is there even a "standard" solution?

-- Wolfgang
-----------------------------------------------------------------
Dr. Wolfgang Schoppek          Universitaet Bayreuth
 Tel.: +49 921 554140
 http://www.uni-bayreuth.de/departments/psychologie/wolfgang.html
-----------------------------------------------------------------