Parameter settings

[Christian J Lebiere]

Excerpts from mail: 4-Feb-99 Parameter settings by Gary Jones at psychology.no 
> I'm examining the effect that different mechanisms of development have upon
> a model of adult behaviour on a developmental task. Some mechanisms are
> particularly suited to ACT-R (e.g. strategy choice can be operationalised
> using expected gain noise). So for example, as you increase EGN, the
> behaviour of the model is degraded (fitting the reaction time data of seven
> year old's on the task). Parameters such as EGN and retrieval threshold are
> open-ended though - are there any specific ranges that are normally adhered
> to? (Catastrophic changes in behaviour are not seen in the model until
> EGN=5 or 6; I expect the "standard" range for modelling adults would be
> 0-1).

I argue in my thesis (now available in .ps and .pdf format at
http://reports-archive.adm.cs.cmu.edu/anon/1998/abstracts/98-186.html)
that catastrophic behavior happens when the activation noise value t
(see terminology note at the end to distinguish between t, s, and
variance) goes over 1.  The value that best fits the cognitive
arithmetic data (and can be seen as optimal - see Chapter 5 of the
thesis) is s=0.25.  That value has worked well for other simulations (as
has the value s=0.5).  That seems to be the general range for activation
noise.

The game playing model presented by Dieter Wallach at last year's
workshop also used an expected gain noise value s=0.5.  However, in
general the expected gain scale is going to be proportional to the value
of G, and thus one wouldn't a priori expect a narrow range of egs values
to fit for all possible values of G.  The Atomic Components of Thought
book (and the ACT-R web site) presents a number of models that use
expected gain noise (check the index for Noise in production
evaluation), and but for one exception they all have a t value below 1. 
However, they also tend to have very small G values in the range of 1 to
3.

Regarding the retrieval threshold, Chapter 7 of the ACT book on List
Memory uses a wide range of values (from less than 0 to more than 3),
but determines a close speed-accuracy-like relationship between the
retrieval threshold and the latency factor.  Thus even when parameters
vary widely they seem to do so with some sort of regularity, suggesting
possible further restrictions.

Terminology note: :egn (and :an) sets the variance of the noise.  While
those keywords still work, the new standard is to set :egs (and :ans),
the s value of the noise, which is related to the variance by
s=sqrt(3*variance)/pi.  The temperature t used in the boltzmann equation
to determine the probabilities of choice is directly related to s by
t=sqrt(2)*s.