Feedback and Behavior

[Foster Chindebvu fosterchindebvu]

Posted by Foster Chindebvu <fosterchindebvu@yahoo.com>
Dear folks,

Can I have a complete Systems Dynamics Model without
feedback loops and still be able to demonstrate some
real world behavior? OR you always have to have
feedback loops in order to demonstrate behavior.

Foster
Posted by Foster Chindebvu <fosterchindebvu@yahoo.com>
posting date Sat, 6 Aug 2005 12:26:17 -0700 (PDT)

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>
The 'behavior' that SD models deal with is 'endogenously generated over
an interval of time'. This requires feedback: change in a variable
during a time interval is propagated through the dynamic structure
(stock-and-flow, rates and levels) and modified in the process until its
influence returns to continue to change the variable.

Without feedback, a variable can only be changed exogenously: either it
is described by a time series of data or it is the result of the
influence of one or more variables described by data time series
filtered to produce the 'exogenously driven' behavior of the variable.

There are different algorithms to implement the feedback in order to
arrive at system behavior. The classic System Dynamics approach
represents the feedback explicitly in the dynamic structure.
Joel Rahn

Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Sun, 07 Aug 2005 12:05:05 -0400

[Bob Eberlein bob vensim.com]

Posted by Bob Eberlein <bob@vensim.com>
This is a pretty interesting question, and has a number of answers
depending on how you think about the issue.

Most fundamentally the answer is no, and the reason is that feedback is
absolutely pervasive in everything around us. When we observe
interesting behavior, it is the result of that feedback. It is important
to note that the usual way that we formally model feedback systems,
using stocks and flows, is not the only way to do it. Feedback can be
represented in lots of different mathematical forms from agent based
simulations to markov chains. It is still feedback, and it is still
pervasive.

Stepping back from this there are lots and lots of people who build
dynamic models that do not have any substantive feedback. These models
basicaly take a number of exogenous data series and derive other series
from them. While such models are not robust to significant changes in
assumptions, lacking the feedback the real world has, they are still
much used, and in some cases useful. Many spreadsheet models and
economietric models have this character.

Finally, there is Kim Warren's favorite observation that just making
someone understand the process of integration is often a big step. A
flow of water filling a container does generate dynamics - a constantly
increating water level. Depending on the shape of the container these
dynamics could be interesting, and perhaps a provide lessson about the
relationship between surface area and volume. However, at the end of the
lesson, when the container is full, someone needs to turn off the water
and again we have closed the loop.

I hope those are useful comments. I recommend, as I am sure will many
others, George Richardson's wonderful book ""Feedback Thought in Social
Science and Systems Theory,"" now available from Pegasus Communications

http://www.pegasuscom.com



Bob Eberlein
Posted by Bob Eberlein <bob@vensim.com>
posting date Sun, 07 Aug 2005 13:53:19 -0400

[John Voyer voyer usm.maine.edu]

Posted by ""John Voyer"" <voyer@usm.maine.edu>
Kim Warren's Forrester Award presentation reiterated the point that
interesting stock-and-flow dynamics, which can be VERY real-world, can
emerge in the absence of feedback loops. I think the reason is that
managers so poorly understand the integration that goes on in stocks and
flows that the resulting dynamics are revealing to them.


John J. Voyer, Ph.D.
Interim Dean and
Professor of Business Administration
School of Business
University of Southern Maine
96 Falmouth St.
Box 9300
Portland, ME 04104-9300

voyer@usm.maine.edu
phone: 207-780-4665
fax: 207-780-4662
---------------- Original message (ID=9D4DBB0B) (1313 lines) ------------------
Received: from smtp.albany.edu (mail1.its.albany.edu [169.226.1.105])
by listserv.albany.edu (8.13.0/8.13.0) with ESMTP id j78L05P9023411
for <sdmail@listserv.albany.edu>; Mon, 8 Aug 2005 17:00:05 -0400 (EDT)
Received: from imf19aec.mail.bellsouth.net (imf19aec.mail.bellsouth.net [205.152.59.67])
by smtp.albany.edu (8.13.0/8.13.0) with ESMTP id j78L01df004752
for <sdmail@listserv.albany.edu>; Mon, 8 Aug 2005 17:00:02 -0400 (EDT)
Received: from ibm60aec.bellsouth.net ([70.144.85.44])
by imf19aec.mail.bellsouth.net with ESMTP
id <20050808205957.SJZR28108.imf19aec.mail.bellsouth.net@ibm60aec.bellsouth.net>
for <sdmail@listserv.albany.edu>; Mon, 8 Aug 2005 16:59:57 -0400
Received: from Drew ([70.144.85.44]) by ibm60aec.bellsouth.net with ESMTP
id <20050808205954.TZJM3347.ibm60aec.bellsouth.net@Drew>
for <sdmail@listserv.albany.edu>; Mon, 8 Aug 2005 16:59:54 -0400
From: ""Andrew Jones"" <apjones@sustainer.org>
To: ""SD list"" <sdmail@listserv.albany.edu>
Subject: ANNOUNCE SD job at Centers for Disease Control
Date: Mon, 8 Aug 2005 16:59:50 -0400
Organization: Sustainability Institute
MIME-Version: 1.0
X-Mailer: Microsoft Office Outlook, Build 11.0.5510
X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2800.1106
Content-Type: text/plain;
charset=""us-ascii""
Content-Transfer-Encoding: 7bit

Dear System Dynamics Community -

In April I emailed this list mentioning an upcoming job opening at the
Centers for Disease Control and Prevention (CDC) in the US, working as a
system dynamics specialist on a project addressing Type II diabetes.

Since then, there have been two developments. First, the position has now
been officially announced and a hiring team in Atlanta has formed. Second,
the technical requirements and responsibilities of the position have
increased; they are looking for someone with significant modeling experience
and skills (whereas in April the description was a step less technical).

All emails and CVs submitted in April were forwarded to Dr. Leonard Jack,
who is the point person for the hire. However, feel free to follow up again
with Dr. Jack given the new position description (pasted below).

Hoping our clients find a top-notch system dynamicist to hire!

Drew Jones
apjones@sustainer.org



Position Description


System Thinking and Modeling Fellow
Applied Behavioral Research, Epidemiology, Surveillance and Evaluation Team
Program Development Branch
Division of Diabetes Translation
National Center for Chronic Disease Prevention and Health Promotion
Centers for Disease Control and Prevention
Atlanta, Georgia


The Program Development Branch, Division of Diabetes Translation (DDT) at
the Centers for Disease Control and Prevention (CDC) in Atlanta, Georgia is
seeking a Fellow in Systems Thinking and Modeling. The Systems Thinking and
Modeling Fellow will assist with sustaining, implementing and building on
existing work undertaken by the Program Development Branch. The Program
Development Branch has utilized System Thinking and Modeling to identify and
improve national and state strategies for addressing the burden of diabetes.
This effort is accomplished primarily through the use of a system dynamics
simulation model and the application of other systems thinking tools/methods
as appropriate.

The Systems Thinking and Modeling Fellow will serve as a member on the
Applied Behavioral Research, Epidemiology, Surveillance, and Evaluation
(ABRESE) Team. This team consists of a an interdisciplinary team of public
health professionals with expertise in program evaluation, behavioral and
social sciences, epidemiology, surveillance, data analysis, program planning
and implementation, health education and consultation expertise. The team
serves as a resource to other complimentary teams in the Program Development
Branch that include: 3 Consultation Teams, the National Diabetes Education
Program, and the National Diabetes Wellness Program. In addition to serving
as a resource to other branch teams, the ABRESE team supports the vision and
mission of the Office of the Branch Chief and the Branch Management Team.

Key Responsibilities for the Systems Thinking and Modeling Fellow:

SCIENTIFIC AND TECHNICAL LEADERSHIP
1. Bring to DDT, hand on experience in the development/application of
Systems Modeling/Dynamics
2. With the goal of expanding the science base, identify opportunities
to apply systems dynamics modeling as a method that augments the existing
methods/science base that provides decision support for goal setting,
program development and evaluation.
3. Contribute to the development of DDT's strategy for the application
of simulation modeling to DDT mission and goals
4. Provide the scientific leadership in model expansion and
improvement , i.e. Calibration for scope/scale as appropriate for use at the
state, regional and national level
5. Produce scientific articles and presentations on the DDT
experience/outcomes associated with the development and application of
systems dynamics modeling for improving diabetes strategy and outcomes.

TECHNOLOGY TRANSFER, TRAINING AND CAPACITY BUILDING
6. Assist with the transfer and application of the technology for use
by states by developing the technical assistance capacity that will be
required by DDT staff to provide technical assistance and consultation in
its real world application..
7. Provide facilitation (model assumptions, systems thinking
exercises, framing of policy issues that are derived from the model) for
use/application of the modeling tool in the ""learning lab"" setting among
cross sectoral stakeholders regarding the use of the simulation model as a
tool.
8. Build capacity within DDT for application of systems thinking
tools/approaches, as a strategy for leveraging assets and strengthening
systems (cross sectoral, interdisciplinary,
collaborative/integrated/synergistic) approaches to population level
interventions that result in measurable improvements in health outcomes.
9. Identify opportunities for expanding the constituency (among the direct
and indirect contributors essential to assuring the Ten Essential
Public
Health Services and to improve community health outcomes) for use of
Simulations modeling as a strategy for improving cross sectoral
decision
making and action.
10 Provide training as required.


Other Qualifications and Skills Required:

* Advance Training and Education (minimum Master Degree)
* At least 3 Years Experience
* Causal Mapping
* Simulation Modeling
* Applied Research
* Group Facilitation
* Consultation skills/experience/expertise
* Scientific Writing
* Publication of Scholarly Peer-Reviewed Articles
* Effective oral communication skills

Interested Candidates Should:

All interested individuals should forward an email expressing interest along
with a current resume to: Leonard Jack, Jr., PhD, MSc, Team Leader, Applied
Behavioral Research, Epidemiology, Surveillance, and Evaluation Team
(ABRESE) at: LJack@cdc.gov. Please be sure to provide contact information
in both your email and current resume to include: best telephone number(s)
to reach interested candidate, email and mailing addresses.

Any questions please contact Dr. Leonard Jack, Jr. at 770-488-5942 or
LJack@cdc.gov.

Salary:

Commensurate with experience

Deadline for Submitting Email Expressing Interest and Resume:

September 1, 2005



Andrew Jones
Sustainability Institute
8 Lynmar Avenue
Asheville, NC 28804
work and best voice mailbox: 828-236-0884
mobile: 828-231-4576
home: 828-252-1266
apjones@sustainer.org
www.sustainabilityinstitute.org
Posted by ""John Voyer"" <voyer@usm.maine.edu>
posting date Mon, 08 Aug 2005 10:33:35 -0400

[Bill Harris bill_harris facilita]

Posted by Bill Harris <bill_harris@facilitatedsystems.com>
>> Posted by Joel Rahn <jrahn@sympatico.ca>
>> Without feedback, a variable can only be changed exogenously: either it
>> is described by a time series of data or it is the result of the
>> influence of one or more variables described by data time series
>> filtered to produce the 'exogenously driven' behavior of the variable.


What about the simple function exp(a*t), where a is a constant and t is
the current time? There's no feedback in that expression (although one
can create a feedback model to generate it), and there's no exogenous
data (except for t, but I think we usually exclude t when we're speaking
of exogenous variables, probably because t is so pervasive).

Bill
- --
Bill Harris http://facilitatedsystems.com/weblog/
Facilitated Systems Everett, WA 98208 USA
http://facilitatedsystems.com/ phone: +1 425 337-5541
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
posting date Mon, 08 Aug 2005 14:37:32 -0700

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi Foster.


When I build a model, I always start with a no stock, no loop model, unless
the necessity to have one stock is evident. It looks then like a common
influence diagram. It is only driven by exogenous variables, and that makes
is very easy to understand.

If it is only driven by parameters, it is only necessary to study one
period. If there is no stock the variables keep the same values through all
the periods.

If it is driven by exogenous data varying every time period, the variables
will change depending on the exogenous data only.

I found that trying to get the most of that preliminary model is always
productive.

If I need to add a stock, I try to find the one that has a sufficient
influence and adds a sufficient added value, to justify the overhead added.

Adding a loop or more makes the model much more difficult to build and
understand, multiplies the effect of the parameters approximation depending
on the number of the time periods and the kind of loop (balancing or not).

I use Vensim, and it works very well with no loops, you can add subscripts,
use the sensibility and optimization and is in fact much more practical then
a spreadsheet, because you can better visualize the influences and use
subscripts.

Whether the no loop model, is sufficiently closed to the reality depends on
the amount of resources you have and the objective of the model.

So to my opinion as a USER of SD, the number of loop depends on their
usefulness.

If I could work with a negative number of loops I would.

If I have a problem with many stocks that generate loops of equal
influences, I leave SD and go to another method, because I will have to
build a too costly model to capture a behaviour sufficiently close to
reality.

SD actually lacks a good method to evaluate in advance the justification of
adding stocks when comparing the usefulness and the cost of doing so or if
it exists I would be glad to know where I can learn from it.

Regards.

J.J. Laublé.
Allocar, rent a car
Strasbourg France.
Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Mon, 8 Aug 2005 15:06:34 +0200

[Bliss Judson BlissJ EPI.WUSTL.ED]

Posted by ""Bliss, Judson"" <BlissJ@EPI.WUSTL.EDU>
My ""rookie"" response to the posting regarding the function exp (a*t) is
that I don't think that this expression occurs in the real world. I can
understand the logic of the expression, but I cannot imagine anything
changing over time, disconnected from anything else. I had first
thought of some element undergoing radioactive decay. Although it does
this without external forces, because it is reducing to some other form
the expression would be incomplete. The bottom line is that SD/feedback
systems must exist in the real world.


Thanks,

Judson Bliss, PhD
Post Doctoral Research Scholar
Washington University School of Medicine
Department of Psychiatry
Campus Box 8134
40 N Kingshighway, Suite 4
St. Louis, MO 63108
Email:blissj@epi.wustl.edu
Posted by ""Bliss, Judson"" <BlissJ@EPI.WUSTL.EDU>
posting date Tue, 9 Aug 2005 08:41:48 -0500

[Jay Forrest systems jayforrest.c]

Posted by ""Jay Forrest"" <systems@jayforrest.com>
The exceptionally simple Barry Richmond Digital Equipment model of one
stock/one flow (customers needing support and computer sales, respectively)
strikes me as the near epitome of a system dynamic model even though it has
no feedback. It fit the real world (at the time), generated realistic
output, produced insights that were not obvious, and clarified thinking.

Thus, while I would agree that feedback is generally necessary (if nothing
else to capture the relationships being modeled in complex situations) it is
IMO clearly not mandatory.

RE: the exponential function. If one deconstructs the formula into how it
works it seems to me an exponential functionally incluldes inherent feedback
as the factor acts on a previous value (ex. 2 cubed = 2 x 2 squared).

Jay
Posted by ""Jay Forrest"" <systems@jayforrest.com>
posting date Tue, 9 Aug 2005 08:17:26 -0500

[Bill Braun bbraun hlthsys.com]

Posted by ""Bill Braun"" <bbraun@hlthsys.com>
>From the pen of: Jean-Jacques Laublé jean-jacques.lauble
wanadoo.fr

>> When I build a model, I always start with a no stock, no loop model, unless
>> the necessity to have one stock is evident. It looks then like a common
>> influence diagram. It is only driven by exogenous variables, and that makes
>> is very easy to understand.


As a result of Kim Warren's Forrester Award address and picking up his
""The Critical Path"" book at the conference and reading it, I find myself
headed in the other direction. If I am not able to identify stocks, their
connected flows, and the relationship(s) between stocks (through the
flows) I begin to question the suitability of SD as a means to insight.
This is not the end my inquiry into the problem from an SD point of view,
but it is emerging as an influence.

Bill Braun
Posted by ""Bill Braun"" <bbraun@hlthsys.com>
posting date Tue, 9 Aug 2005 11:25:03 -0400 (EDT)

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>
exp(a*t) is an exogenous function of time; it is equivalent to a Time
Series. System Dynamics models may use time series inputs for various
reasons but they are never part of the 'dynamic core' of the model. The
proof: Given data time series inputs are usually replaced by other,
standard inputs (step, ramp, oscillating functions) during model-testing
phases in order to develop better understanding of the behavior (i.e.
the dynamic behavior) of the model.

A historical note about 'dynamics': I recall from the time of the
release of 'World Dynamics' that some French critics complained that
'dynamics' was the wrong word to use; they preferred 'kinetics' because
they claimed that there were no 'forces' operating in the model. They
may also have been upset by the lack of a proper 'energy' function from
which to derive the equations for the dynamics of the system. It's a
sterile debate; let's agree not to go there.

Joel Rahn
Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Tue, 09 Aug 2005 09:39:01 -0400

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>
Foster's original questions were:

>Can I have a complete Systems Dynamics Model without
>feedback loops and still be able to demonstrate some
>real world behavior? OR you always have to have
>feedback loops in order to demonstrate behavior.

These are not mutually exclusive questions. The first deals only with ""a
complete System Dynamics Model"" and the answer is No.

The second deals with any old model ""to demonstrate behavior"" ('behavior
over time' is meant I believe; otherwise it is of no interest to most of
the readers of this list). The answer to this question is also No
insofar as it deals with all non-SD-compatible methodologies and Yes for
all the others.

Jean-Jacques Laublé apparently prefers to use Vensim (which is not the
same as SD) in order not to develop SD models. I am unsure how working
'with a negative number of loops' will help him (OK, I know it's a joke)
but I am truly mystified about the -dynamic- insights he can generate. I
do of course agree that being able to visualise the effects over time of
applying a complex, static filter to exogenous input(s) that vary over
time can be very instructive. It isn't SD.

Joel Rahn
Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Tue, 09 Aug 2005 10:33:35 -0400

[Nijland lukkenaer planet.nl]

Posted by Nijland <lukkenaer@planet.nl>
What is the essence of feedback?


The sigmoid time-series

y = max / {1+EXP(-a *time) }

,without any loops,

may be reformulated,

as a system with two feedback loops,

a positive one, and a negative one.

dy/dt = max * y * (1 - y)



This system is able to show exponential growth,

exponential decay, it may be in equilibrium, etc.

(also chaotic behaviour, if you choose a too long solution interval)

Many biological, psychological and sociological phenomena in reality

may be described as well as explained by such a feedback system.

Both loops can be interpreted very well as substantial sub-processes.



In my opinion ""feedback"" starts,

where causal interpretation begins.



in the initial equation

y = max / {1+EXP(-a *time) }

y is a function of time, the causal diagram is simple (time >> y),

but time is not the cause of y.



But for any time-series (even a constant horizontal line)

one may state, that the next value depends on the former.

Is ""continuity"" also caused by feedback?



Geert Nijland
Posted by Nijland <lukkenaer@planet.nl>
posting date Wed, 10 Aug 2005 16:31:13 +0200

[Jean-Jacques Laublé jean-jacques]

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Joel Rahn writes



<Jean-Jacques Laublé apparently prefers to use Vensim (which is not the
<same as SD) in order not to develop SD models


I do not see where I wrote that, I wrote exactly the contrary. I start with
a simple model if I can, with no loop, and then choose to add a stock if
necessary and so on.


In fact I do not mind using true or false SD as long as it helps to
understand reality that is always dynamic whatever way you choose to analyse
it. If you suppress time in the world you suppress the world too.

Anybody can go on the forum Vensim site

http://www.ventanasystems.co.uk/forum/ and download a very simple but
fundamental for a rental business static model written in Vensim. There is
too a Word file with explanations and with the sketch, the equations and the
variable comments for those who do not use Vensim.

There are too comments on the way to expand the model by adding stocks.

Regards.

J.J. Laublé Allocar

Strasbourg France.
Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Thu, 11 Aug 2005 12:07:14 +0200

[Tom Fiddaman tom vensim.com]

Posted by Tom Fiddaman <tom@vensim.com>
A model without feedback loops is to System Dynamics what warming up a TV
dinner in the microwave is to gourmet cooking. Sometimes it might be
useful and the right thing to do, but it's a different beast.

>Kim Warren's Forrester Award presentation reiterated the point that
>interesting stock-and-flow dynamics, which can be VERY real-world,
>can emerge in the absence of feedback loops.

Those who don't believe this should read Cloudy Skies at

http://web.mit.edu/jsterman/www/cloudy_skies.html

You don't need a model to discover cases of unperceived stocks, though
unless you're lucky it will take a modeling process to expose such problems.
While stock insights alone may be enough some of the time, as soon as you
start modeling what's going on feedback is needed more or less immediately -
as in Bob's example of someone needing to shut off the tap in the bath tub.

>What about the simple function exp(a*t), where a is a constant and t is
>the current time?

It's possible to imagine much more complex equations representing solutions
to higher-order systems, and these might be surprising, useful, and realistic
in their own right. However, exp(a*t) doesn't tell you much about causality,
the origin of parameter a, or how it might change in response to policies or
uncertainties. In a real project, you get exp(a*t) as a solution that describes
the behavior of a model that includes feedback (if you're lucky; otherwise
you're stuck with simulation). The parameters of such a model are tied to an
operational description of reality and constrained by fit to data. So, in a
technical sense exp(a*t) answers the question in the affirmative, but it
doesn't have much to do with problem-oriented dynamic modeling.

>When I build  a model, I always start with a no stock, no loop model, unless
>the necessity to have one stock is evident. It looks then like a common
>influence diagram. It is only driven by exogenous variables, and that makes
>is very easy to understand.

Of course, the flip side of ""easy to understand"" is ""not likely to provide
surprising insight."" I often start with a feedback-free model for a different
reason: to get all the data around a problem organized and cleaned in a data
model that can then be used to drive the dynamic model that is the real goal.

> I find myself headed in the other direction. If I am not able to identify
>stocks, their connected flows, and the relationship(s) between stocks (through
>the flows) I begin to question the suitability of SD as a means to insight.

Perhaps this highlights the difference between the consulting practitioner,
looking for a good application for SD, and the problem owner, who wants an
answer by any appropriate path of least resistance. Fortunately SD software
has useful applications in both cases.

>Adding a loop or more makes the model much more difficult to build and
> understand ...

I think it depends what you mean by difficult. If a problem is truly dynamic,
non-dynamic models will have a difficult time achieving any kind of fit to
information about the system, though they may be easy to build.

Without some reference to a real problem, this discussion is dangerously
close to a sterile ""dynamic vs. kinetic"" debate. It would be much easier if
we started with ""exactly what are we proposing to model, and why?""

Tom
Posted by Tom Fiddaman <tom@vensim.com>
posting date Thu, 11 Aug 2005 10:04:57 -0600

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>


Jean-Jacques Laublé jean-jacques.lauble wanadoo.fr wrote:

>>I do not see where I wrote that, I wrote exactly the contrary. I start with
>>a simple model if I can, with no loop, and then choose to add a stock if
>>necessary and so on.

My comment was based on the ""if necessary"". In SD, the unit of analysis
is the feedback loop and you can't have a feedback loop without a stock.
The stock is necessary not optional. Until you have a stock, you are not
using SD although you may be using Vensim ar any of the other
programming environments.


>>In fact I do not mind using true or false SD as long as it helps to
>>understand reality that is always dynamic whatever way you choose to analyse
>>it. If you suppress time in the world you suppress the world too.

There is no such thing as ""false SD"". There are SD models that are well
or poorly verified and validated. There are models that change over time
that do not have feedback loops; they are not SD.

Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Thu, 11 Aug 2005 08:15:57 -0400

[Jim Duggan JAMES.DUGGAN NUIGALWA]

Posted by Jim Duggan <JAMES.DUGGAN@NUIGALWAY.IE>

>>Posted by ""Jay Forrest"" <systems@jayforrest.com>
>>The exceptionally simple Barry Richmond Digital Equipment model of one
>>stock/one flow (customers needing support and computer sales, respectively)
>>strikes me as the near epitome of a system dynamic model even though it has
>>no feedback. It fit the real world (at the time), generated realistic
>>output, produced insights that were not obvious, and clarified thinking.

=======================================================================

Jay, I've heard mention of Barry Richmond's one stock model, is there
any further information available on it, and what kind of response
did it get from Digital's management at the time?

I worked for Digital back in the early 1990s, and there was always
""interesting"" dynamics with different internal groups
advocating a range of modelling techniques and
analytical approaches to world wide component sourcing, supply
chain and capacity analysis problems, so I'd be interested to see
how the System Dynamic approach was viewed back then.


regards,
Jim.

Dr. Jim Duggan
Lecturer, Department of Information Technology,
NUI, Galway.

http://corrib.it.nuigalway.ie
Posted by Jim Duggan <JAMES.DUGGAN@NUIGALWAY.IE>
posting date Sun, 14 Aug 2005 21:12:43 +0100

[Jay Forrest systems jayforrest.c]

Posted by ""Jay Forrest"" <systems@jayforrest.com>
Hi Jim!

This is probably a bit long and overly familiar for the whole listserve, but
there are probably people who will benefit, so here goes. Please feel free
to delete the message if you know the DEC/Richardson story. The following is
based on my hearing Barry tell of the experience in classes and in personal
discussions while teaching together. No doubt there are others on the list
who no more than I!

In the late 60's and early 70s DEC's philosophy was that engineering drove
sales so they tried to grow their research and development group to build
sales. And they were relatively successful - sales were growing
exponentially. In the late 60's they hired an outsider to run their service
organization. In the early 70's it became evident that the service group was
growing faster than any other part of the company and that customer service
costs were oupacing growth in research and development. The feeling of
senior management was that he was a kingdom builder. Barry was hired to help
explore the problem - and reading between the lines probably to put him in
his place and squelch the growth of the service department.

Barry spent three days with DEC management trying to build a model that
explained what was happening. Barry indicated that the path to the model was
rough. As in many SD investigations, there is a resistance to simple models
and simple models are often difficult to develop and gain acceptance.

The resulting model was simply a flow of sales to a stock of cumulative
sales (or perhaps more appropriately installed computers). (I can't recall
the precise title they had on the stock.)

It is my impression there was quite a bit of initial resistance to this
model due to its simplicity but management was heavily involved in the
development process and one could feed historic sales in and it gave a very
accurate output of installed computers.

This model defies a few standard SD logics - no feedback, stocks with no
outflow, etc. But at that point in time, the cost of computers was so high
that there was essentially no retiring or replacing of computers (or if they
were replaced the older computer would be used in some other capacity. And
no each computer required seller support - due to the complexity and special
features/internals of each ""brand."" As a result service demand was
essentially directly linked to the number of computers in service.

Now, back to the model. It is not immediately obvious, but for this simple
two element model, the stock of installed computers will grow faster than
sales in all cases except where sales grow supraexponentially. (Building a
model and playing with it will quickly show this! On a more intuitive basis,
the number of installed computers goes up even if sales decline!) Thus the
service group (costs and revenues) should grow faster than sales. Until, of
course, when people begin retiring computers!

The management at DEC was predominatly engineers and once their resistance
was appeased immediately recognized the implications and the ""turf war""
claims were over.

I really like to use this model with students for it really strikes home how
poorly human minds understand stock/flow dynamics and the implications
therein.

Thanks for asking. Hope this is useful!
Jay Forrest
Posted by ""Jay Forrest"" <systems@jayforrest.com>
posting date Mon, 15 Aug 2005 09:08:25 -0500

[Jay Forrest systems jayforrest.c]

Posted by ""Jay Forrest"" <systems@jayforrest.com>
In my email regarding the simple, two-element DEC model I inadvertently typed Richardson instead of Richmond. The story related to Barry Richmond and his work for Digital Equipment. I appologize for any confusio nmy typo may have caused.

Jay Forrest
Posted by ""Jay Forrest"" <systems@jayforrest.com>
posting date Tue, 16 Aug 2005 08:09:37 -0500

[Jim Duggan JAMES.DUGGAN NUIGALWA]

Posted by Jim Duggan <JAMES.DUGGAN@NUIGALWAY.IE>

>>===== Original Message From system dynamics listserve
>>Posted by ""Jay Forrest"" <systems@jayforrest.com>


Many thanks Jay, that was really useful!

I must try and work it into my teaching for next year, particularly because it is related to the computer industry dynamics.

I suppose when you look at the rise and decline of
Digital, the question of the engineering vs marketing
world views was key... but that's another story!

best regards,
Jim.

Dr. Jim Duggan
Lecturer, Department of Information Technology,
NUI, Galway.
Posted by Jim Duggan <JAMES.DUGGAN@NUIGALWAY.IE>
posting date Tue, 16 Aug 2005 12:21:27 +0100

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>
Jay's anecdote is interesting, pedagogically instructive and reveals more about the target audience (DEC managers) than it does about SD.

Jay Forrest systems jayforrest.com wrote:


>>Now, back to the model. It is not immediately obvious, but for this
>>simple two element model, the stock of installed computers will grow
>>faster than sales in all cases except where sales grow
>>supraexponentially.
>>

This is an ambiguous statement. On the face of it, it seems wrong because the stock is 'growing' at the same rate as sales i.e stock-growth-rate = sales. From the following parenthetical remark, it appears that what is meant is that the slope of the stock-as-a-function-of-time graph is greater than or equal to the slope of the sales-as-a-function-of-time graph. This last statement is a truism of integration as long as sales rate is greater than or equal to zero (excluding exceptional cases): the area swept out by a yardstick gows faster than the length of the yardstick would be another example of this truism, which is always true no matter how fast or slow the yardstick moves; i.e. independent of the 'dynamics' or, more explicitly, it is not a 'dynamic' phenomenon.


>> (Building a
>>model and playing with it will quickly show this! On a more intuitive
>>basis, the number of installed computers goes up even if sales
>>decline!) Thus the service group (costs and revenues) should grow
>>faster than sales. Until, of course, when people begin retiring
>>computers!
>>
>>The management at DEC was predominatly engineers and once their
>>resistance was appeased immediately recognized the implications and
>>the ""turf war"" claims were over.
>>
>>

The last paragraph confirms what I have said to people on many
occasions: I learned more about applied math by doing SD than in all of my differential and integral equations courses or, more accurately, I developed better intuition about integration and differentiaion thanks to SD. I hazard a guess that DEC's management suffered from the same 'professional deformation'. In the SD community, our 'professional deformation' often seems to be that if it moves it must be 'dynamic'. Not.


Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Tue, 16 Aug 2005 08:28:46 -0400

[Neil Douglas ndouglas powersim.c]

Posted by ""Neil Douglas"" <ndouglas@powersim.co.uk>

Joel Rahn <jrahn@sympatico.ca> wrote


>>it appears that what is meant is that the slope of the stock-as-a-
>>function-of-time graph is greater than or equal to the slope of the


I was also very interested to hear of Barry Richmond's work at Digital, and agree entirely with Joel's comment about the SD communities tendency to see all problems as dynamic.

However, I'm afraid that either Joel's 'truism' of integration isn't true at all or I'm completely misunderstanding it. The implied statement is that 'the slope of the stock-as-a-function-of-time graph is greater than or equal to the slope of the sales-as-a-function-of-time graph when the slope of the stock-as-a-function of time (the sales rate) is greater than zero' is incorrect.

Firstly when comparing the slope of the stock-as-a-function-of-time graph to the slope of the sales-as-a-function-of-time graph we have a unit problem. These quantities have different units and hence cannot be directly compared. Rephrasing the 'truism' in terms of distances, we can see it says that if an object's velocity is greater than zero, then its acceleration is less than its velocity. The second part of this statement suffers from a bad unit problem.

This problem is particularly acute when comparing the 'length' of a yardstick to the 'area' swept out by the yardstick!

Secondly, even in the perfect mathematical world without units the implied 'truism' is incorrect. The obvious example is if Sales = Stock * 2, in which case the the slope of the sales-as-a-function of time graph is always twice the slope of the stock-as-a-function-of-time-graph .

We don't need feedback for this to happen either. (Sales = Time2) grows polynomially and has slope (Rate of Growth of Sales = 2*Time), and we can see that slope of the stock-as-a-function-of-time-graph is less than the slope of the sales-as-a-function of time graph for all Time less than or equal to 1.

Posted by ""Neil Douglas"" <ndouglas@powersim.co.uk>
posting date Thu, 18 Aug 2005 10:09:01 +0100

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>
I was perhaps a little flippant with my 'slope' comparisons; I am aware that there is a units problem. My comment was on a less-elevated plane and based on eyeballing graphs of behavior over time comparing only such features as the corresponding signs of slopes (+,-,0) and accelerations (rates-of-change of slopes). For the yardstick example, the length is constant (slope=0), the area swept out is increasing (slope is greater than 0); the same holds true if the 'yardstick' is replaced by 'sales rate' and the area swept out is the stock of installed (sold) systems. Thus, ""The stock of installed (sold) systems 'looks like' it is going up faster than the sales (except for some special cases).""

If the sales rate is any monotonic, polynomial function of time, say s0*t, the stock will be a polynomial of the next higher degree, i.e. (s0*t*t)/2. The limit of the ratio of stock to sales rate is an increasing function of time, i.e. t/2. On a graph, except for a negligible initial interval, the stock curve looks like it is rising faster and faster (parabolically in this case) compared to the sales curve that it is rising in a straight line.

If sales are rising exponentially, say s0*exp(t/GT), the stock also rises exponentially, (s0*exp(t/GT)-s0)/(1/GT) or GT*(s0*exp(t/GT)-s0) so the ratio of stock to sales is constant and it is not evident that the stock is growing faster than the sales. This uncertainty appears effectively only if exponential sales growth is sustained for a long period of time. I'm not sure what sales as a 'supra-exponential' function means but I don't think I can integrate it in my head...

If sales are variable (but always greater than or equal to zero), the stock will always be increasing or constant; i.e., the graph of sales may go up (slope positive) or down (slope negative) or be flat (slope =
0) or be zero (also slope = 0) and in all of these cases except the last, the stock will increase, i.e., its slope will be positive and only for the last will the slope of the stock curve be 0 (but that will be the least of management's worries).

So that's it. The stock of installed (sold) systems always 'looks like' it is going up faster than the sales (except for some special cases).

Neil Douglas ndouglas powersim.co.uk wrote:
Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Thu, 18 Aug 2005 11:35:57 -0400

[Tom Fiddaman tom vensim.com]

Posted by Tom Fiddaman <tom@vensim.com>
I think everyone's right here, but the fuzzy language distinction between the fractional growth rate of a stock and the absolute rate of change of a stock is making it hard to see. In Barry's argument, it's the former that matters, because the important thing to DEC managers is the relative magnitude of revenue or cost streams associated with sales and service. The prices of unit sales and services ($/computer and $/computer/year) make the streams of activity comparable in spite of the philosophical incomparability of velocity vs. acceleration.

For DEC:
base = INTEG(sales) [installed base in computers]
service$ = Pserv*base [service revenue, $/year]
sales$ = Pcomp*sales [sales revenue, $/year]
servRatio = service$/sales$ = 1/r * INTEG(sales)/sales [ratio of service revenue to sales revenue, where r is a parameter relating relative prices of sales and service, which could be interpreted as the number of years of service that equals a sale]

The boundary case is (as already pointed out) exponential sales, since
INTEG(sales) is then also exponential at the same rate, and servRatio remains constant. While sales might be supraexponential for a short time early in the lifecycle, eventually they must grow at a declining rate, and the service to sales ratio will rise. For example, if sales follows a ramp (a*time) then servRatio = 1/r*a/2*time2/(a*time) = a/2/r*time, which is increasing.

It's actually a bit more complicated than I have stated, as I neglected initial conditions. Odd initial conditions (e.g. large installed base, low sales) make it possible for simple inputs - like ramp sales - to yield surprising trajectories for the service/sales ratio. The fact that such a trivial model is hard to discuss in words and hard to mentally simulate, though easy to solve in closed form, places it squarely in the realm of SD.

Tom
****************************************************<
Tom Fiddaman
Ventana Systems, Inc.
http://www.vensim.com

Posted by Tom Fiddaman <tom@vensim.com>
posting date Thu, 18 Aug 2005 15:30:59 -0600

[Jay Forrest systems jayforrest.c]

Posted by ""Jay Forrest"" <systems@jayforrest.com>
When Barry Richmond spoke of the growth rate of the stock he was effectively saying the percentage growth rate, and likewise for the flow. Percentages are unitless and therefore there is no unit difference to confuse the comparison. Simply put the growth of installed computers will be greater than the growth in sales for anything less than supraexponential sales growth. (which is feasible only over a short period).

The mechanics of the problem being analyzed was that Customer Service (which was essentially linearly related to computers in service at that time) was growing faster than sales (which was assumed by management to be strictly a function of technology (and thereby engineering.) The elegance of the two element model is wonderful but IMO the model communicates better with a few converters to show the associated assumptions!

A problem arises in arguing against the conclusion that computers in service grows faster than sales in the form of using the formula

Sales = Stock * 2

as this implies that there is a connector from computers to sales. That is NOT an element of the Richmond model and conclusions drawn from that formula do not apply to the model developed by Barry.

Again, I urge you to build the model and experiment with sales growth rate formulas and see how the rate of change of the stock relates to the flow. And...to keep in mind that while actual growth was accelerating it was definitely not supraexponential. As a real world test of the conclusion, keep in mind that slowing sales (which would show a negative growth rate) still result in increasing service demands!

Again, this tiny, exceedingly simple model captured the actual dynamics quite accurately - especially when fed actual sales data. Yet even this tiny, simple model and its insight eluded very intelligent people until confronted by its insights in a form that they could not discount.

Hope this helps!
Jay Forrest
Posted by ""Jay Forrest"" <systems@jayforrest.com>
posting date Thu, 18 Aug 2005 08:34:22 -0500

[Joel Rahn jrahn sympatico.ca]

Posted by Joel Rahn <jrahn@sympatico.ca>
Thanks for clarifying the underlying model and its use of revenue parameters to make the ServRatio dmensionless.

The model described below uses the ratio of a stock, base, and a flow, sales, and it is the absolute rate of change of the stock that is involved, not the fractional growth rate of the stock nor of the flow. From Tom's model below it can be seen that the ServRatio generally grows a long as sales are not negative and ssuming initial base and sales are zero. The result holds for fractional growth rates also as I hope to show.

Lets look at the fractional growth of base and sales. If sales are like a*t^n, then the base goes like a*t^(n+1)/(n+1) and the fractional growth rate of base is positive but decreasing like (n+1)/t. Thus linear growth in sales (n=1) leads to a base that is growing parabolically (n+1=2) but is 'suffering' a decline in fractional growth like 2/t. Note that, aside from the 1/t sort of decline in fractional growth rate here, if sales shift suddenly from cubic growth (n=3) to linear (n=1), the fractional growth rate of base drops to a half or less of its previous value in absolute terms.

The fractional growth rate of sales are positive but decreasing like n/t; in the above example of downshifting sales, the fractional growth rate of sales would drop suddenly to a third or less of its previous value. The comparative growth rates ( the focus of the DEC situation): fractional growth rate of base/fractional growth rate of sales is constant, (n+1)/n, and > 1. Just what we want.

It is easy to show that if sales slow down (but stay greater than or equal to zero) after an interval of growth, the fractional growth rate of base remains positive while the fractional growth rate of sales is negative. Again the fractional growth rate of base > fractional growth rate of sales.

So the general message of the DEC case is maintained: the growth rate of a stock (absolute or fractional) is generally greater than the corresponding (absolute or fractional) growth of its single (or possibly
net) non-negative flow rate.

Finally, I have now to concur with Tom's statement that Richmond's simple model is 'squarely in the realm of SD' because, although without significant feedback, it does show how tricky non-linear functions can be and SD facilitates the exploration of non-linearity in time as well as in relations between other parameters.


Posted by Joel Rahn <jrahn@sympatico.ca>
posting date Tue, 23 Aug 2005 13:10:11 -0400