how can ACT-R models age?

[Erik M. Altmann]

At 9:04 PM -0500 1/4/01, Lynne wrote:

>My impression of aging effects is that the *biggest* liability of 
>aging is in acquiring new information not retrieving old 
>information--this occurs  with  poor encoding of simple facts, but 
>is most pronounced in learning new skills and concepts.  It seems 
>that older individuals are most handicapped in domains where their 
>prior knowledge is of least value, e.g., learning new technologies 
>or video games.   It is not obvious to me how Erik's filling up the 
>brain with too many chunks would predict that general 
>pattern/problem with
>aging.  My hunch (and again, since it has not been simulated it is only a
>hunch) is that W would do a better job of explaining that aspect of 
>the demise of intellect with age.

I expect there are many liabilities to cognitive aging, and I can't 
claim to know which is the "biggest", but there is evidence that 
cognitive aging involves a decreased ability to inhibit irrelevant 
information -- that is, an increased susceptibility to interference. 
See May, Hasher and Kane (1999), for some recent examples, and Kane 
and Hasher (1995) for a review.

>Erik's other remark was that his proposal was a natural consequence 
>of aging, while W is a free parameter.  Well, it would at least be
>constrained to go down with age, not up.  Moreover, it is a single parameter,
>while my reading of Erik's proposal involved twiddling several parameters,
>but perhaps I'm mistaken (low W and lots of decayed chunks due to my 
>advanced age, or at least, that's my excuse).

This comment raises a question about how to interpret model 
parameters theoretically.  I'm not sure that number of parameters by 
itself is the best criterion for evaluating the success of a model. 
In the limit, a model with a single parameter set to "do the task" 
would not be particularly interesting.  My concern about W, and also 
about partial matching, relates specifically to this point.  As I see 
W and partial matching often used, they redescribe the data rather 
than explaining the data in terms of underlying or more complex 
processes.

I have little invested in my aging proposal (as yet!) and don't care 
to be dogmatic about it, but it does serve as a counterexample in 
that it involves multiple parameters.  Of course it involves multiple 
parameters, not to mention multiple assumptions -- it's a process 
account, and any complex process will involve many degrees of 
freedom.  This is not by itself a drawback.  The more interesting 
question is how well those degrees of freedom are constrained, both 
externally by other data and theory and internally by mutual 
constraint among mechanisms, and what questions they raise to drive 
further research.

In searching for constraints on this model, I would start by fixing 
the decay rate at d =0.5, which has worked well now in many models. 
The big question would probably be the rate at which chunks are added 
to memory.  There's evidence from many sources, including the ACT-R 
Argus models being developed at GMU, a slough of models in Soar, 
Logan's instance theory, etc., that elements are added to memory at a 
relatively constant rate -- the "fecundity" of learning mechanisms 
leads to pervasive acquisition of episodic memory.  So, pick a 
constant rate.  But then, what's the effect of consolidation, say 
during sleep?  Is this what causes the decay rate to "slow down" as 
John et al recently documented?  More generally, what's the 
physiology of decay?  And what's the interaction with cues spreading 
source activation to really old chunks, reviving those old memories 
and feeding back into their retrieval history?  These questions come 
up because the account is a process model with lots of parameters and 
assumptions.  Stopping with W, like explaining semantic gradients 
with partial matching, cuts off discussion of how the history and 
context of the system might produce or moderate the effect of 
interest.

Cheers,

Erik.

May, Hasher, & Kane (1999).  The role of interference in memory span. 
Memory & Cognition, 27, 759-767.

Kane & Hasher (1995).  Interference.  In Maddox (ed.), Encyclopedia 
of aging (pp 514-516).
-- 

~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~
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
~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~