how can ACT-R models age?

[Erik M. Altmann]

At 5:09 PM -0500 1/4/01, Dario Salvucci wrote:
>Might anyone know the current status of work relating ACT-R and 
>aging?  In particular, I'm wondering if anyone has done work toward 
>the following question: Given a "young expert" ACT-R model, is there 
>a general (domain-independent) way of making it an "elderly" model 
>simply by changing appropriate parameters?  For instance, one might 
>imagine that cycle time increases by some percentage causing general 
>slowdown (there seems to be EPIC work suggesting this), or that W 
>decreases, and/or that the latency of certain perceptual-motor 
>parameters increases.  I'm specifically interested in modeling 
>elderly drivers using an existing model of younger drivers, but I'm 
>hoping to carry over any related results / parameter changes from 
>other domains if at all possible.

I've been thinking about a representation of age in which the brain 
fills up with chunks that aren't completely decayed.  For this to 
have typical aging effects, I believe you have to dispense with the 
retrieval threshold and particularly with indexed retrieval (in which 
you force the retrieval of a specific chunk, and give up if that 
specific chunk isn't above the retrieval threshold).  That is, you 
have to be willing to let activation play the role it's meant to play 
under rational analysis, and let it predict the need for a chunk 
based on history and context.  If you do this, then on a given 
retrieval cycle the accuracy of the retrieval (in terms of 
probability of retrieving the target chunk) is determined by the 
chunk choice equation, which factors in the activation of all old 
chunks in memory.  The more old chunks there are, the lower the 
probability of retrieving the correct one, and the greater the 
probability of going off on a tangent as a function of a 
mis-retrieval.  This captures, at an abstract level, the general 
aging phenomenon of decreasing ability to inhibit irrelevant 
information.  This kind of forgetting model (in which the main 
factors are interference and decay and not so much the retrieval 
threshold) also seems quite general, with successful applications to 
the Tower of Hanoi, task switching, order memory, Brown-Peterson, and 
the Waugh and Norman probe digit task.  Some of these applications 
involve long-term as well as short-term memory.

Some other pieces to the puzzle.  First, if cognitive aging is 
represented this way, the existence of vast numbers of old chunks is 
counterbalanced by the fact that each individual one is highly 
decayed and so contributes little to the denominator of the chunk 
choice equation.  So the basic idea seems structurally plausible -- 
background noise in the head increases gradually over a lifetime. 
Second, to approximate the effect of tens or hundreds of millions of 
chunks in memory, one can simply increase activation noise (s).  The 
decline in accuracy (as governed by the chunk choice equation) 
follows a slightly different curve, but is still curvilinear and 
makes for a much more efficient simulation.  In my order memory 
model, I use this trick to account for order reconstruction accuracy 
at a retention interval of 24 hours, in which chunks would ordinarily 
be added to memory at a rate of several per second (assuming that 
people don't turn their brains off when they're not thinking about 
your task).  Third, the advantage of representing aging in terms of 
declarative memory filling up is that it actually specifies an aging 
mechanism, unlike twiddling W, for example.

Erik.

-- 

~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~
Erik M. Altmann
Department of Psychology
Michigan State University
East Lansing, MI  48824
517-353-4406 (voice)
517-353-1652 (fax)
ema at msu.edu
http://www.msu.edu/~ema
~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~