punch@cpswh.cps.msu.edu (Dr William Punch) (02/20/90)
Through talking to a number of people (thanks to everyone who
responded) I have two sources of public domain GA/Classifier
software. Below are descriptions of both and the addresses of the
authors. Thanks again all.
>>>bill<<<
punch@cpswh.cps.msu.edu
--------
Number 1
John Grefenstette: gref@aic.nrl.navy.mil, additions by
schraudo%cs@ucsd.edu (Nici Schraudolph).
Our research group here at UCSD has found John J. Grefenstette's GENESIS
genetic algorithm simulator an invaluable tool for GA research, and the
frequency with which we use it in different contexts has motivated us to
put some effort into advancing the GENESIS source code itself.
Starting from GENESIS 4.5 I have therefore developed three new versions of
GENESIS over this summer. The first two of these (which fix a number of 4.5
bugs and improve the user interface, respectively) have been sent to John,
who has kindly agreed to organize their distribution and integration into
mainstream GENESIS.
The third version includes more idiosyncratic improvements and facilities,
and is available as "GENESIS 1.0ucsd" from us via anonymous ftp. Among a
host of minor improvements 1.0ucsd offers the following new features:
* simulations can be killed and restarted at will;
* the setup program dynamically suggests default values for parameters;
* additional termination criteria (Bias and Conv thresholds) are available;
* evaluation functions may operate directly on the packed gene for speed;
* almost any scalar C function can be used directly as evaluation function;
* simulation queues may be executed in parallel on a network of hosts.
Copies of the source code can be obtained as follows:
$ mkdir GAucsd
$ cd GAucsd
$ ftp sdcsvax.UCSD.EDU
Login: anonymous
Password: <anything>
ftp> cd pub/GA
ftp> mget GENESIS.?.Z
ftp> quit
$ uncompress *.Z
$ GENESIS.1
$ GENESIS.2
$ GENESIS.3
$ GENESIS.4
$ lpr UsersGuide
$ view README
WARNING: this is effectively a beta test version, and I can and will not
give any guarantee of correctness. I will maintain a list of users for
distributing bug fixes, updates, and new versions. To get added to that
list, report bugs, or make suggestions send mail to nschraudolph@UCSD.EDU.
GENESIS 1.0ucsd is an important tool for GA research in our group, and we
expect it to continue evolving. For the near future we are planning the
addition of ARGOT-style operation and fine-grain parallelism (using RPC
to distribute evaluations).
--------
Number 2
Rick Riolo: rick_riolo@um.cc.umich.edu
The CFS-C classifier system is a collection of domain independent
subroutines that implement the "major cycle" of a classifier system
(message matching, bidding and competition, message posting,
and bucket brigade re-allocation of strength). The package also
includes some classifier "discovery" routines (including some
"genetic" operators and a rule coupling operator).
I also distribute some subroutines that implement two
example task domains: LETSEQ, a letter prediction task, and
FSW1, a finite state world traversal task.
These can be linked to the CFS-C routines to produce
runnable system, and they can serve as an example
for writing your own task domain dependent subroutines.
The source is 99% machine/compiler independent,
and has been compiled on PC/ATs, Apollo's, SUNs, Vaxes,
IBM 3090's, and Macs. (Of course that means the displays that come
with the code are not graphics oriented, so you would want to write
your own displays, most likely, to fit your machines, task, and interestes.)
The code is written in C, and on 80x86 style machines
has been compiled using Microsoft C (4.0 or 5.0) and CI86.
There are two documents that describe the original CFS-C and
LETSEQ systems, and a draft that describes the FSW1 domain.
Unfortunately the two "official" documents are a bit out of date,
but the distribution includes a NEW.TXT file which describes changes
to the system since the documents were written. The documents
do give an accurate overall view of the system--just some
of the details are out of date (and those are noted in NEW.TXT).
The easiest way for me to distribute it is on
a 1.2MByte, 5.25in disk for ms-dos. Can you read that?
If you need another distribution method, let me know
and I'll see what can be arranged (eg mac 800KByte disks are easy,
too).
I would appreciate a $8 ($20 out of NAmer) "contribution"
to cover media/mailing costs. Please send it to:
Rick Riolo
905 Miller Ave.
Ann Arbor MI 48103
(If you need a tape or something the cost would be more,
and if its international mail, its more.)
And of course please send me your Physical-Mail address.
If you have any problems putting the code up, let me know.
And please let me know of any bugs you find.
Finally, I would appreciate hearing what applications or
research projects for which you plan to use the classifier
system, just so I can learn what people are doing.
And of course I would like to hear of any results you
get (good or bad) or lessons about classifier systems you learn.
Thanks.