[Martin Taylor 970120 11:10]
Bill Powers (970117.1420 MST)]
General note:
Wolfgang Zocher has reminded me of his analog-computer-on-a-digital-
computer, Simcon, which is now in version 5.2. It doesn't have the graphical
layout capability of Stella, but it works great and it doesn't cost hundreds
of dollars (it's free). Furthermore, computing setups are ASCII files that
can be sent directly over the net. Anyone can run the simulations without
having to know how to program them. I have the new postscript version of the
manual, and have asked Wolfgang to send me the latest source code (C) and
runnable code for PCs. When I get it all together and checked out, I'll post
it on my FTP page. Watch for the announcement.
Well, since I can't run any of the SIMCON stuff, I am trying to learn Java
by rewriting SIMCON in Java. Wolfgang has been kind enough to send me the 5.2
source, and I have edited the manual into an HTML version with internal
cross-links that make it (a little) easier to look up points while one
is working through the examples. (Wolfgang sent me an HTML version without
the cross-links). I've sent the revised version to Wolfgang (last night).
By the way, for the person who wanted his e-mail address, it's:
zzzo@apollo.han.de
One point: my revised manual also notes the computational artifact (found
by Bill P. and me some years ago) involved in the calculation of "optimum
gain" for the "Simple Control System" program, and suggests as an exercise
for the reader that they rewrite the program to simulate the same analogue
network with values of the compute iteration cycle (dt) reduced by a factor
of ten. This augments the discussion of computational artifacts later in the
manual.
Such computational artifacts are an ever-present hazard when simulating
analogue machines in discrete time, and as Bill P has so often said, the
only sure(-ish) way of avoiding them is to try the same simulation at
two quite different iteration speeds. I felt that the SIMCON manual ought
to point this out. My own ad-hoc approach is to ensure that no big changes
to a signal value or any of its derivatives happen, and that no signal
influences its own value, in less than two (and preferably considerably
more) compute iteration cycles.
If I ever learn enough Java to be able to produce a full (or even partial)
version of SIMCON, of course I'll post it, either on the CSG Web site
or on my own (<http://www.servtech.com/public/mmt>\). It probably won't
use any of Wolfgang's source code, and will run a lot slower than his
unless it is run on a machine with a Java "Just In Time" (JIT) compiler.
But it will accept any legal SIMCON program--or at least that's a design
objective.
The point of all this is to make the discussions that can be elucidated
by running SIMCON programs available to everyone, not just those with PCs.
I have more ambitious hopes, but SIMCON is quite ambitious enough for now:-)
By the way, Wolfgang told me that the 5.2 version runs only on Unix systems,
and that nobody had volunteered to port it to PCs. Has somebody now offered
to do that port? He says the code has "defines" for PCs, though. Or maybe
I misinterpreted him. I haven't looked through his code yet.
ยทยทยท
-----------------
On the other matter:
I think what's happened is that I have reached my reference level for
speculating about higher levels of organization without data, and want to
get back to what is the main point for me: building the PCT model step by step.
That's a very reasonable position. But it doesn't preclude you from answering
the question I asked twice, which can be boiled down to:
What do you think is the difference between our two views of category
processing, other than that I permit and you prohibit inputs to category
perceptual functions from the higher levels of the hierarchy?
You were quite clear that there are indeed "deep and fundamental" differences,
by which I assume you meant that the differences were important enough to
be mentioned, at least, and probably important enough to be discussed. But
you never said what these difference are. Even if you would prefer to
work only on building the lower levels, I think it a bit unfair of you to
leave me hanging with the assertion that I am making important differences
to the HPCT model on which everyone else is working, without telling me
what those differences are.
A second question that follows is:
Is the difference between permitting and prohibiting category inputs from
higher levels important or trivial? If it's important, why?
(These questions are open to everyone to answer, of course).
Martin