[ACT-R-users] matching score

[db30 at andrew.cmu.edu]

--On Monday, November 10, 2003 4:53 PM +0100 Sarah Schimke 
<sarah.schimke at cognition.uni-freiburg.de> wrote:

>
> Hello,
>
> I have the following problem with an act-r model:
> the model works on the symbolic level and also when the
>  subsymbolic computations are turned on by ":esc".
> But when I'm adding noise by ":ans", there are a lot of
> failures of retrieval because the different chunks take
> matching scores that I don't understand.
> In particular, the matching score turns to "nil" for chunks
> the model is trying to retrieve directly (by a request like
> for example "+retrieval> =goal"), so that a failure gets
> retrieved instead of the chunk. This happens often, but not
> all the time.
> Why does this happen?
>

A retrieval failure is caused when the activation
(or matching score) is below the retrieval threshold.
When there is no noise (:ans not set) there is no
variability in the activations, and if a chunk is
above the threshold it will always be above the
threshold.  Once the noise is turned on that's going
to change because there will be a random amount of
noise added to the activation when a retrieval is
attempted, and thus it may end up below the retrieval
threshold.

How you deal with that depends on why you are turning
on the noise and what you're trying to get out of the
model.  If you don't want any failures then one
possibility is to just set the retrieval threshold
(the :rt parameter) very low so that no activation
is likely to ever go below it.

The case with a direct retrieval is a special one
because that is the only chunk attempted.  If its
activation (plus noise) is above the thereshold
it is retrieved, but if not it fails - there is
no partial matching on a direct retrieval.  Seeing
nil for the matching score doesn't sound right, and
I can't seem to find where that would happen because
when I turn on the traces I still see numeric values
during a failure for a direct retrieval:

CG-USER(5): (run .1)
 Matching production Test.
 Matching production Test
 Time  0.000: Test Selected
 Sources of activation are: nil
    Adding noise -0.097
 Partial matching chunk G with activation -0.097
 Matching score of chunk G is -0.097.
 CHUNK G match score -0.097 is below threshold 10.000: failure.

Can you provide a trace or more details as to where you
are seeing nil because that sounds like a bug.

Hope that helps,
Dan