Learning facts in ACT

[Christian Schunn]

Learning addition (or multiplication) facts is a nice example where 
these issues come up. What one observes empirically is that children 
continue to recompute answers even after they have started to 
sometimes retrieve the answer to a problem. The result of this 
conservative approach is that they don't get locked into overlearning 
incorrect facts. I think data from this domain provides evidence that 
ACT-R's "retrieve if you can" approach to strategy selection is 
likely to be wrong. However, I suppose you might address some of 
these issues by setting the retrieval threshold quite high. See the 
Schunn, Reder et al (1997) JEP:LMC on the SAC model or Bob Siegler's 
(1995) ASCM model for how an alternative model of this process might 
work.

-Chris

At 6:23 PM +0000 1/25/00, Richard M Young wrote:
>ACTors:
>
>I have an embarrassingly simple question, or set of related questions,
>about fact-learning in ACT.  For the purpose of clarity, I'll pose the
>question in the context of learning a set of paired-associates, although I
>think the point is more general.  I suspect the answer already exists in a
>model somewhere, and I just need to be pointed to it.
>
>Let's take as a starting point the (obviously over-simplified) model of
>paired-associate learning and retrieval in the file "paired associate" in
>Unit 6 of the ACT tutorial.  The crucial part is two rules (there are only
>three anyway), one of which retrieves the "pair" if it can, and if it can't
>the other comes into play and "studies" the pair as it is presented.  As is
>pointed out in the tutorial, the retrieval rule serves to reinforce the
>activation of the pair twice, once because it is retrieved on the LHS of
>the rule, and once more when the pair is re-formed from being popped from
>the goalstack on the RHS.  Notice that "studying" only boosts the
>activation of the pair once, when it is formed (or re-formed) on the RHS.
>
>I got to wondering what would happen if the modelled S ever got into
>its/his/her head an INCORRECT pair, i.e. with a valid stimulus paired with
>an incorrect response.  As the model stands, the error would never be
>corrected, because the erroneous chunk would repeatedly be retrieved, and
>would be reinforced (twice) each time.  However, it is probably unrealistic
>to suppose that S doesn't read the feedback just because a response has
>been retrieved, so there is the opportunity to notice that the retrieved
>response is wrong and to "study" the correct response.  However, each time
>that happens, the erroneous chunk gets reinforced twice but the correct
>chunk only once, as we have seen.  So, given that the erroneous chunks
>starts off more active than the correct one, except for a vanishingly low
>probability sequence of events, the correct chunk would never get learned
>to the point of being retrieved.
>
>OK, so it's a crazily over-simplified model, but it does raise the question
>of how *would* ACT learn paired associates given that it starts off with,
>or at any stage acquires, erroneous pairs?  I've thought of a couple of
>ways, but I'm not even sure they'd really work, and they certainly don't
>seem like convincing stories:
>
>(1) Because a retrieval is not guaranteed to be correct, it should not
>automatically be popped on the RHS of a retrieval rule.  If the model waits
>for feedback and makes sure it pops only a correct pair, then a correct
>chunk will be reinforced (once) on each trial.  Unfortunately, the
>erroneous chunk also gets reinforced once, by being retrieved on the LHS.
>Because the correct chunk is reinforced AFTER the erroneous one, it profits
>from recency, and I suppose it's possible that with patience and some luck
>with the noise, on some occasion the two chunks will be close enough in
>activation that the correct pair gets retrieved and therefore twice
>reinforced, and thereafter is likely to win.  But the story doesn't sound
>convincing.  (And solutions which involve the repeated, deliberate,
>multiple rehearsal of the correct chunk sound too contrived.)
>
>(2) When an erroneous retrieval occurs, and the model discovers that it's
>wrong from the feedback, as well as learning a correct pair it could also
>learn the incorrect pair with an additional annotation (attribute) of
>"wrong".  The retrieval would need to become more elaborate: after
>retrieving a pair in reply to a probe with the stimulus, the model would
>check whether it could also retrieve an extended pair marked wrong using
>the stimulus and the retrieved response.  If it couldn't, OK.  If it could,
>then it would need to retrieve another pair, with the previous response
>explicitly negated.  (I think that's possible).  Well, maybe, but again it
>seems rather contrived.
>
>Can anyone tell me how this is done better?
>
>-- Richard

======================================================
   Christian Schunn           Applied Cognitive Program
   Psychology 3F5             cschunn at gmu.edu
   George Mason University    (703)-993-1744  Voice
   Fairfax, VA 22030-4444     (703)-993-1330  Fax
   http://www.hfac.gmu.edu/~schunn
======================================================