ACT-R 2.1 + Summer School material

[Christian Lebiere]

NOTE: I am leaving on vacation immediately and will be back Sunday
night.  I will answer all questions/remarks/complaints first thing
Monday morning.

The latest version of ACT-R and the material used in the just-concluded
1995 Summer School are now available on the usual server(s).  Mac
versions are available on the server largo.psy.cmu.edu in the directory
/public/ACT-R.  Text and/or BinHex versions are available on the server
ftp.andrew.cmu.edu in the directory /pub/act-r/ftp.

A new top-level folder called education has been created in the
repository.  The file GMU ACT-R Curriculum contains the curriculum of an
ACT-R-based course that Wayne Gray taught recently at George Mason
University (Thanks, Wayne!).  The folder 1995 Summer School contains the
summer school material.  It comprises 1) the schedule in MS-Word 2) a
Hypercard stack (sorry, non-Mac people, I will try to see if there is a
Word version) with the text of each lesson, exercises, and assignments
3) eight folders containing for each session one or more examples of
relevant ACT-R models and the solution model (text files).  This
material worked well in the summer school, but still needs improvement. 
We will be going over it this fall and an updated version should be
available this winter.

The folder ACT-R 2.1 contains the new version of the system (the
previous one was ACT-R 2.0, patch1 and patch2).  It is available both as
a BinHex version of a Stuffit Archive containing the whole ACT-R folder
(as released with the ROM book), and as a Code folder containing the
full code in text files.  People who ported ACT-R to other Common Lisp
implementations should reapply the port to this version rather than
attempt to integrate the voluminous change files.  The folder Extensions
contains some (not all) of those change files.  It also contains two
sub-folders, return-extension and new analogy, which contain examples of
the new mechanism which allow subgoals to return values, and the new
analogy mechanism, respectively.  The latter also contains an MS-Word
manual explaining the new mechanism.

ACT-R 2.1 should only display minor differences with ACT-R 2.0 on
existing examples, but incorporates a number of further developments.  I
will attempt to briefly describe them below, while acknowledging that
the existing documentation is insufficient and will be remediated upon
my return.  The best way to learn about them currently would be in the
summer school material (or by contacting me).

* a number of fixes have been incorporated.  The only seriously
noticeable difference is that a production attempting to pop the last
goal on the stack (typically the last instantiation of the run) can now
complete succesfully rather than abort.

* ACT-R is now fully operating on a real-time clock.  That means that
the cycle counter is now a real number starting at 0.0 (instead of 1). 
If Enable Rational Analysis is turned on, then the matching latency is
added to the counter (otherwise nothing).  The latency for the rhs (or
action latency) is the value of the parameter :a (or :effort) for that
production.  So when ERA is turned off and the :a parameters are left to
their default value of 1.0, we get back the discrete time clock.  Our
usual value for :a is 0.05 (50 milliseconds if you will).

* There is now a way to set initial base level activation independently
for each wme, using the SetGeneralBaseLevels and AddGeneralBaseLevels
commands.

* A subgoal can now return a value to its parent goal, using a
generalization of the wme retrieval syntax to the right-hand side.  The
return-extension folder in Extensions contains examples of its use to
compute the factorial and fibonacci function.

* Activation-wise, elements (i.e. slot value which are wmes) of the goal
are now the default sources of activation, instead of the goal itself. 
The total source level of activation (1.0 by default) is divided evenly
among them.  Also, the IA value from a wme to itself is now set to the
same value as the IAs from that element to any other (i.e. direct
retrievals are now treated the same as indirect ones).

* The new analogy mechanism which had been in beta test for several
months has now become standard.  The New Analogy folder in Extensions
contains a brief manual describing its use, and two folders: a folder of
the previously developed analogy models unchanged, and a folder
containing those examples sometimes modified to run under the new
mechanism.

In addition, the partial matching mechanism is now available as an
extension to the standard system (partial-matching.lisp in folder
Extensions) to be loaded as needed, as it is still a bit too unstable to
be integrated in the standard system.  Our Cogsci 94 paper ("Error
modeling in the ACT-R production system", by Lebiere, Anderson and
Reder) introduces the general mechanism, and the summer school material
describes how to use it.

As was discussed extensively at the ACT-R Worskhop, this is the last
major upgrade of the current system, ACT-R 2.x.  We will keep supporting
this system, but have started work on its successor, ACT-R 3.0, which
will incorporate the restrictions and stylistic guidelines that were
discussed at the workshop, in a more efficient and flexible
implementation with a more useful, user-friendly environment.  More
about that later on this list.

Again, feel free to send me any problem/comment/suggestion.  I will be
out until Monday, but will answer my email as soon as I come back.

Christian