[ACT-R-users] New ACT-R software release available

Tue Dec 27 17:00:11 EST 2016

CMU is closed for the Holidays and it might be a while before Dan replies,
so I will take a stab...

The new mechanism came out of conversations between Dan and I which go back
to my thesis work on a lifetime model of cognitive arithmetic (short
version here <http://act-r.psy.cmu.edu/?post_type=publications&p=14049> and
long version here <http://act-r.psy.cmu.edu/?post_type=publications&p=13870>).
It turns out that, under most realistic assumptions, practice does *not*
increase the difference in activation between relevant and irrelevant
chunks. If you assume that the relative frequency of knowledge does not
substantially change over time (e.g., the *relative* frequency of correct
facts such as 3+4=7 vs 3+5=8 remains the same as fixed by the environment)
then the difference in activation between the two chunks remains the same
even though both become more active with practice. That means that the
probability of commission errors (mistakenly retrieving one for the other)
remains constant, which is clearly not inline with the longitudinal data.
My thesis mentioned the theoretical argument that if the noise were to
decrease as a function of the log of practice then the odds of commission
errors would decrease as a power law of practice, which roughly seems to be
as indicated by the data.

That is not the only possible solution to the conundrum. The model actually
used associative learning to achieve that effect, and while it sort of
worked it was also brittle and exhibited a number of quirks. Besides the
mechanism is now deprecated so if you have a new candidate mechanism it
would be interesting to see if it works better. Another possibility would
be if the matching penalty increased with practice, but there are some
issues with getting that to work in the right conditions.

Note that idea of getting the noise to decrease with practice has
interesting connotations. From a rational analysis point of view,
decreasing the noise as a log of practice has been shown in some settings
(e.g., simulated annealing in Boltzmann networks) to yield an optimal
schedule for exploration-exploitation tradeoff. From a neural point of
view, an obvious interpretation would consist of holding the noise
magnitude constant while the size of the weights encoding the memory grows
with amount of practice (e.g., using superposition learning yielding random
walk-type dynamics).

On the latter point, an alternative ACT-R implementation would consist of
moving the noise term from a separate additive term in the activation
equation on a par with history (bll) and matching (al and pm) terms to an
additive term inside the log of the base-level history term. That way, the
log function applied to the sum over past references would quash it as
required. As I recall, Dan thought that it was a more disruptive change and
decided to include the new noise as a separate term instead at this initial
stage. Its parametrization and inclusion as an extra rather than as a
default was meant to let the community experiment with it and decide
whether it made sense before considering further deployment.

Happy holidays everyone.

Christian

On Tue, Dec 27, 2016 at 11:22 AM, Ion Juvina <ion.juvina at wright.edu> wrote:

> Hi Dan,
>
> Thanks for the update.
>
> I have a question: what is the theoretical rationale for adding the
> adaptive noise to declarative memories?
>
> As I understand it now, I think this mechanism is (at best) unnecessary.
>
> It is unnecessary because the effect of noise DOES already decrease with
> practice. Practice increases the difference in activation between relevant
> and irrelevant chunks, which decreases the confusion between the two kinds
> of chunks at retrieval. If noise is set in the appropriate range, its
> effect will diminish and approach zero with practice.
>
> Adding an explicit parameter to achieve something that emerges from
> existing mechanisms may even be detrimental considering the general
> recommendation to keep the number of parameters to a minimum.
>
> I may be wrong, which is why I want to hear the rationale and the data
> that would suggest such an addition to the architecture.
>
> I’d appreciate comments from the community as well.
>
> Thanks,
> ~ ion
>
>
>
> > On Dec 21, 2016, at 10:18 AM, db30 at andrew.cmu.edu wrote:
> >
> >
> > A new version of the ACT-R 7 software is now available from the
> > ACT-R website: <https://urldefense.proofpoint.com/v2/url?u=http-
> 3A__act-2Dr.psy.cmu.edu_software_&d=CwICAg&c=3buyMx9JlH1z22L_G5pM28wz_
> Ru6WjhVHwo-vpeS0Gk&r=vtA7YXGBFnFocyRUneK5pkvCTPZEpjEwzP4TURDawmo&m=
> wAdy6WChw5TkZMFPkFE8jcYsmZfwF2YmsHAsT47QsHc&s=
> 8bxD85GR4gpHnN3lrnNM5SgWDsPGYOLFrlvMrkpwv3I&e= >.  The current
> > version is now 7.3-<2102:2016-12-20>.
> >
> > A few of the notable changes are listed below.  More details
> > can be found in the commit log found on the ACT-R website at:
> > <https://urldefense.proofpoint.com/v2/url?u=http-
> 3A__act-2Dr.psy.cmu.edu_log_actr6log.txt&d=CwICAg&c=
> 3buyMx9JlH1z22L_G5pM28wz_Ru6WjhVHwo-vpeS0Gk&r=
> vtA7YXGBFnFocyRUneK5pkvCTPZEpjEwzP4TURDawmo&m=
> wAdy6WChw5TkZMFPkFE8jcYsmZfwF2YmsHAsT47QsHc&s=
> EKW0s8sMl4vPuMe6ewyS2uJvdAAcfsZA9UOlE2y7zO4&e= >, but only the changes
> > in the actr7 branch are relevant to the current software.
> >
> > Production compilation has been updated to better handle buffers
> > which are strict harvested.  It is now safer because it will not
> > compose productions which have a stuffed chunk between them for
> > goal and imaginal type buffers in cases where strict harvesting
> > was the reason the buffer cleared in the first production.  It
> > also produces better composed productions in situations were
> > strict harvesting clears the buffer and the second production
> > tests that it is empty.  Previously the 'buffer empty' query from
> > the second production was added to the composed production, but
> > that prevents the composed production from competing with the
> > first parent.  Now that query is dropped from the new production.
> >
> > The mod-focus command now makes the modification during the
> > goal-modification event instead of directly.  This allows one to use
> > goal-focus followed by a mod-focus without it complaining that the
> > buffer is empty.
> >
> > Added an extra which implements a modification to the activation
> > calculation for declarative memory.  The purpose of the modifiction
> > is to decrease the effect of noise on the activation of chunks as
> > their activation increases with practice.  The extras/adaptive-noise
> > directory contains the code which adds a parameter called :uan
> > (use adaptive-noise) that can be set to t to enable the mechanism.
> > The documentation included with the extra describes how it works,
> > and there are some sample graphs of the results of some very simple
> > test cases showing the change when :uan is enabled.
> >
> >
> > If you have any questions or problems with the new version please let
> > me know.
> >
> > Dan
> > _______________________________________________
> > ACT-R-users mailing list
> > ACT-R-users at act-r.psy.cmu.edu
> > https://urldefense.proofpoint.com/v2/url?u=https-3A__
> mailman.srv.cs.cmu.edu_mailman_listinfo_act-2Dr-2Dusers&d=CwICAg&c=
> 3buyMx9JlH1z22L_G5pM28wz_Ru6WjhVHwo-vpeS0Gk&r=
> vtA7YXGBFnFocyRUneK5pkvCTPZEpjEwzP4TURDawmo&m=
> wAdy6WChw5TkZMFPkFE8jcYsmZfwF2YmsHAsT47QsHc&s=RnQ_nPNyv_
> IVaNhYKHaMOXOzX4Bs7E8yUgh1ikQdv8M&e=
>
>
> _______________________________________________
> ACT-R-users mailing list
> ACT-R-users at act-r.psy.cmu.edu
> https://mailman.srv.cs.cmu.edu/mailman/listinfo/act-r-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/act-r-users/attachments/20161227/d3aa2bc6/attachment.html>