A different idea regarding simulation accuracy

This forum contains all archives from the SD Mailing list (go to http://www.systemdynamics.org/forum/ for more information). This is here as a read-only resource, please post any SD related questions to the SD Discussion forum.
Locked
Bill Harris
Senior Member
Posts: 75
Joined: Fri Mar 29, 2002 3:39 am

A different idea regarding simulation accuracy

Post by Bill Harris »

Long before I discovered system dynamics, I was simulating circuits
with SPICE, home-brew programs, and other such tools. (Before that, I
was in school, solving differential equations on a old EAI, which I
guess is close to SD.) In thinking back to those SPICE experiences, I
recall a few points:

Accurate models were deemed important if I were trying to verify
circuit performance to specs; if I had a bad model of a transistor,
I might not get the results I needed.

I never had quite good enough a model to duplicate reality, so Id
use simulation for other things than proving the circuit met specs:

to get an initial design that was "close enough" more quickly than
I could through prototyping, relying on the actual circuit
prototype (or prototypes) to settle the final details;

to measure the unmeasurable (its hard to measure a current
through a very low impedance, but SPICE could do that easily);

to estimate circuit sensitivity to parameter and environmental
changes (with final verification to come later on the prototype).

Before you get to thinking Im too lazy to make a good model, Ive at
least got some good company. People like Bob Pease
(
http://www.national.com
ap/,
http://www.elecdesign.com/magazine/1998 ... 9bp1.shtml,
and http://www.elecdesign.com/magazine/1999 ... 11bp.shtml
--I never said he wasnt opinionated!), an analog circuit guru at
National Semiconductor, caution new engineers against trusting SPICE
instead of breadboarding circuits--there are just too many parasitic
components to be dealt with in a model.

Okay, is this the SPICE mailing list or what? :-)

No, but I was thinking about what that all might mean to me in the SD
world. If its hard to model transistors accurately (*), how much
harder is it to model people and organizations? We dont have network
analyzers and component testers to plug people and organizations into
which produce nice tables describing their policies.

What if there are two reasonable worlds regarding model accuracy?

For many of us, the goal is making an organization better quickly.
If we can create a reasonable model, demonstrate it can replicate
the salient features of the reference mode, and then demonstrate a
change to the system that moves us reasonably far in the right
direction, maybe weve done enough to warrant moving on to the
breadboard (the real world). Sure, we could do better, but it would
cost us a lot in time to get that accuracy. Well probably learn
still more when we get into implementing the change in the
organization, anyway.

For others of us, accuracy is key. If we are supposed to give
quantitative policies to our client, the accuracy requirements are
tight, and big money is at stake, then its worth the time to do
some of the very good, accurate work certain people have reported.
(Theres also probably enough money floating around these clients to
pay for the development of that sort of model.) The analogy in the
physical world might be the simulation of a circuit one cant test
on a lab bench fairly easily, such as the control of a rocket to
another planet.

Why all this verbiage? Maybe it can help the next time we get into
our periodic discussions on accuracy. The key is purpose: if Im
using SD to improve an organization quickly and effectively but can
make final organizational tweaks on the fly, then the former approach
works. (It represents combining SD with action research
<http://www.scu.edu.au/schools/gcm/ar/arhome.html>, which could be a
whole topic in itself.) If you cant afford the risk of tweaking the
final (real) system and you have to get high accuracy, the second
approach is called for.

Arent I on a slippery slope? Whats to stop me from using CLDs to
describe a system and claim I understand it without simulation? Many
have rightfully pointed out the dangers in that approach.

Going back to the circuit world, I had lots of rules of thumb for
designing circuits. Some were very approximate (and often wrong),
some were better but still not quite right, and some were almost as
good as I could measure. The key to knowing which to apply was
engineering judgment.

So, that could be the answer here, too. Im sure there are simple
situations where a CLD is sufficient to identify possible problems and
cures; there are obviously many situations where simulation is
required to avoid surprises. The key to discernment is judgment,
which comes from experience and, well, judgment.

And the proof of the pudding is in its construction and use. You
question the engineers judgment when the bridge falls down; you can
question the SDers judgment when the organization doesnt respond as
predicted. And you learn as an engineer and SDer by building models
_and_ playing in the real world enough to continue finding where your
blind spots are.

Enough philosophy for one day.

Bill

* It is easier today to create good transistor models cheaply than it
was 20+ years ago.
--
Bill Harris 3217 102nd Place SE
Facilitated Systems Everett, WA 98208 USA
http://facilitatedsystems.com/ phone: +1 425 337-5541
From: Bill Harris <bill_harris@facilitatedsystems.com>
Locked