2001 Post-Graduate Summer School

[Wolfgang Schoppek]

I suggest that associative learning should be one of the issues to be
discussed at the post-graduate summer school. I understand that turning
on associative learning often leads to unpredicted (and often
disturbing) effects. But instead of eliminating a core concept of
psychology from ACT-R, we should try to improve the associative learning
mechanism.
 I had a closer look at the details of the current mechanism and found
that one of the biggest problems lies in the estimation of the initial
rji value. Sure, the equation  rji*=m/n is derived from rational
analysis, but this is worthless when you don't have accurate estimates
for m and n ("garbage in - garbage out"). Also, the theoretical
justification why newly created IAs ("soft links", not slot relations)
have an rji*=1 is not clear to me. This "rji*=1 assumption" leads to the
effect that new chunks, when they become associated with old chunks get
more and more negative sjis as the run proceeds - all other things being
equal!
 I tried out some modifications, which are listed below (slot relations
are called "hard links", purely associative Sjis are called "soft
links"):
 
- *wme-number* is a constant (or parameter) now, default value is 250
(interpreted as the "semantic neighborhood")
 
- I removed the default self-activation of chunks (cf. create-wme) but
kept the initial value of fan at 1 (that means that the actual fan is
always 1 less than the calculated) (this change remdies the "retrieving
the same chunk over and over again effect")
 
- Soft links are treated the same way as hard links, i.e. chunk i (the
target) is included in the fan list (wme-fan) of chunk j (the cue) and
the fan is increased by 1.
 
- Also, the initial rji (called rji*) for soft links does no longer
default to 1, but is calculated the same way as for hard links, i.e.
*wme-number* / wme-fan(j).
 
In my tests, the new mechanism produces more plausible sjis, and the
wme-fans don't get unreasonably high. I'll continue my testing. Btw.:
the model of "a pilot flying a 747-400" (which I presented at the 2000
Workshop see:
http://www.uni-bayreuth.de/departments/psychologie/ACTR-2000/index.htm
) uses the modified learning mechanism and runs fine (hundreds of
cycles).
 If anyone is interested in these modifications, please let me know.
 
 Finally one more little ad: I've developed a (mathematical) ACT-R model
of the keeping track task which has some interesting implications about
interference in memory. Don't worry: no associative learning there. You
can read it in the CogSci 2000 proceedings  or download it
(
http://www.ircs.upenn.edu/cogsci2000/PRCDNGS/SPRCDNGS/posters/schoppek.pdf
)
 
Cheers,
-- Wolfgang
--------------------------------------------------------------------
 Dr. Wolfgang Schoppek                <<< Tel.: +1 703-993-4663 <<<
 HUMAN FACTORS & APPLIED COGNITION PROGRAM, George Mason University,
 Fairfax, VA 22030-4444
  http://www.uni-bayreuth.de/departments/psychologie/wolfgang.htm
--------------------------------------------------------------------