A Behavioral Approach To Feedba - Desconhecido

A behavioral approach to feedback loop
dominance analysis
David N. Forda
David N. Ford is an Abstract

associate professor in Feedback loop dominance is a critical tool in explaining how structure drives behavior.
the system dynamics Current analytic tools for loop dominance analysis are tacit, not codi®ed, unable to
program at the accurately identify dominant loops or inapplicable to most models. Most loop dominance
University of Bergen analysis tools focus on model structure to link structure and behavior. We use a behavioral
in Norway. He earned
perspective to de®ne dominance, improve descriptions of behavior patterns and identify
his PhD from the
two important and incompletely developed areas of feedback analysis: simultaneous
Massachusetts
Institute of dominance by multiple loops and shadow loop structures. An analytic procedure is
Technology, where he presented, illustrated and compared to an alternative analysis method. An evaluation of
conducted research the behavioral approach is the basis for identifying new issues and future research
on the dynamics of opportunities. Copyright *c 1999 John Wiley & Sons, Ltd.
development Syst. Dyn. Rev. 15, 3±36, (1999)
processes. His current
research interests
include product
development
management, One of the central messages of system dynamics is that structure drives
coordination and behavior. More speci®cally system dynamics explains how structure drives
policy development. behavior (Forrester 1961; Richardson 1991). Linking dominant feedback loops
and shifts in loop dominance to behavior patterns is critical in these explana-
tions. Discovering these links requires an analysis which identi®es the loops
that dominate given time intervals of a simulation. How do system dynamicists
analyze a model to identify the dominant loops? How much con®dence should
be placed in the conclusions of these analyses? A reasonable question to system
dynamicists who explain model behavior with loop dominance is ``How do you
know that the feedback loops which you describe as dominant are the loops
which actually dominate?'' The validity of most current analyses is implicitly
based on the authority of the modeler as a quali®ed system dynamicist, the
plausibility of the loop dominance explanation, the persuasiveness of the
presenter or some combination thereof. These are adequate foundations for
analysis validation in some contexts, such as consulting and teaching, but the
rigorous research required to further solidify system dynamics as an independ-
ent scienti®c domain requires a method of feedback loop dominance analysis
that is independent of the analyst. System dynamics does not currently have
such a method for identifying dominant feedback loops.
System dynamicists have traditionally used experimental model exploration,
model reduction, or both with their understanding of the behavior patterns
typically generated by positive and negative feedback loops to identify
dominant loops (Richardson 1991, 1986). These informal approaches can lead
a
Department of Information Science, University of Bergen, 5020 Bergen, Norway. E-mail: David.Ford@
i®.uib.no, http://www.i®.uib.no/sta/david/
System Dynamics Review Vol. 15, No. 1, (Spring 1999): 3±36 Received March 1997
Copyright *
c 1999 John Wiley & Sons, Ltd. CCC 0883-7066/99/010003±33 $17.50 Accepted January 1998
3
4 System Dynamics Review Volume 15 Number 1 Spring 1999
to errors in identifying dominant loops. Informal analysis procedures are

potentially incomplete and vulnerable to variation and bias due to analyst
expectations and model and project circumstances. Basing the polarity of the
loops that are assumed to dominate on typical loop behavior can cause errors
because feedback loops do not generate unique behavior patterns. Loop
polarity is only loosely coupled to speci®c behavior patterns. Examples of
atypical behavior include positive and negative loops that generate linear
behavior when systems are in equilibrium and Richardson's (1995) description
of Ashby's example of a positive loop that generates convergent behavior.
Graham (1977) provides additional examples. Informal analysis procedures and
limiting candidate loops for dominance to those with only one polarity based
on typical behavior patterns can cause errors in feedback loop dominance
model analysis.
System dynamics researchers have recognized the need for more rigor in
feedback loop dominance analysis in general and the behavioral aspects of
such analysis as a particularly important focus (Richardson 1986). However, no
formal and unambiguous de®nition of behavior as it relates to dominance has
been formulated. In addition, the often criticized (including here) experimental
method currently used for the majority of dominance analyses remains tacit
and uncodi®ed. This failure to map current practice prevents the evaluation
and improvement of those practices, their comparison with recommended
analysis procedures and their use in building improved tools and procedures.
Finally, research has focused on the structural aspects of how feedback
structures and behavior are linked far more than on behavioral aspects.
Richardson (1995) develops mathematically rigorous de®nitions of loop polarity
and system conditions that determine shifts in dominance from loops of one
polarity to the other. Mojtahedzadeh (1996) and Kim (1995) use system
structure to determine loop and gain polarity and strength. Kampmann (1996a;
1996b) extends N. Forrester's (1982) use of feedback loop eigenvalue elasticities
to explain the relative contributions of dierent loops in linear systems. In
contrast Davidsen (1992) focuses on improving an intuitive understanding of
how feedback loops drive behavior with a graphical description of stock and
¯ow behaviors. These structural approaches address only a portion of the
possible feedback structures (e.g. linear or two-loop systems), are dicult to
apply or are impractical for models of signi®cant size. System dynamics needs
an understanding of feedback loop dominance that balances structural and
behavioral perspectives. A behavioral approach to feedback loop dominance
analysis is an important step in developing such a balanced understanding.
To rigorously analyze loop dominance in all but small and simple models
and eectively apply analysis results, system dynamicists need at least two
Ford: Feedback Loop Dominance Analysis 5
things: (1) automated analysis tools applicable to models with many loops and
(2) a clear and unambiguous understanding of loop dominance and how it
impacts system behavior. Other research (Mojtahedzadeh 1996; Kampmann
1996a; 1996b) addresses the development of automated analysis tools, but these
tools are not currently complete or available. This leaves practicing system
dynamicists with only informal analysis methods to identify the potentially
most important links between a system dynamics model and the behavior it is
intended to explain. Even when automatic feedback loop dominance tools
become available, they will be of little value and potentially lead to inaccurate
applications of system dynamics if practitioners lack a fundamental under-
standing of loop dominance. A behavioral approach to feedback loop analysis
can play an important role in identifying the needs of automated tools and
providing an eective means of explaining the relationship between loop
dominance and behavior. In this paper we seek to improve model analysis of
dominance and provide a practical tool for dominant feedback loop identi®ca-
tion by describing a behavior-based approach to feedback loop dominance,
using that approach to formalize an analysis procedure, providing an example
of its application, and evaluating the approach to identify areas for
improvement and further research.
A behavioral definition of feedback loop dominance
Richardson and Pugh (1981, p. 285) provide a useful basis for developing a
behavioral de®nition of loop dominance, ` . . . a loop that is primarily responsible
for model behavior over some time interval is known as a dominant loop.'' We
expand on Richardon and Pugh's de®nition by adding speci®city concerning
three aspects: behavior patterns, location of dominance, and gains. The purpose
of feedback loop dominance analysis is to identify feedback structures that
dominate behavior. A rigorous approach must de®ne the two system features it
relates ( feedback structures and variable behaviors) independently to prevent
circular reasoning such as ``the behavior is the kind generated by a positive
loop, so the dominant loop must be positive.'' Similarly, describing a behavior
pattern in terms of its approach to a system feature such as a goal (as is
commonly done with negative feedback loops) is less rigorous than a descrip-
tion based solely on the behavior of the variable. More elaborate de®nitions of
behavior based on system structure are common in structural approaches to
feedback loop dominance analysis. An improved de®nition of behavior patterns
is required, which relies only on the behavior of the variable of interest itself.
DoÈrner (1996) refers to behavior patterns as `` `shapes' in time'' and considers

our ability to recognize them as essential to understanding complex dynamic
systems. But what are the essential shapes of dynamic behavior? Richardson
(1986) asks the more fundamental question ``What do we mean by a particular
pattern of behavior?'' in his analysis of research needs concerning dominant
structures. To answer these questions we identify three unique behavior
patterns based on the net rates of change of the variable of interest (Figure 1).
The net rate of change of a variable is a characteristic solely of the variable and
can be determined independently of how the change is generated (the structure
and system conditions). The absolute values of these rates describe movement
greater and less than their initial values. Trends in the absolute values of net
rates of change can be used to uniquely identify three atomic behavior
patterns1. The ®rst atomic behavior pattern is linear behavior. When the
absolute value of the net rate of change of a system variable is constant, the
variable grows or declines steadily and the behavior is linear (Figure 1, pattern
1). Equilibrium conditions are a special case of this pattern, in which the net
rate of change is zero. The second atomic behavior pattern is exponential
growth or decay. When the absolute value of the net rate of change of a system
variable increases over time the variable moves away from its initial value
faster over time (Figure 1, pattern 2). The typical behavior generated by positive
feedback loops is exponential. The third atomic behavior pattern is logarithmic
growth or decay. When the absolute value of the net rate of change decreases
over time the variable moves away from its initial conditions at a slower rate
over time (Figure 1, pattern 3).
Fig. 1. Atomic
behavior patterns
The three atomic behavior patterns can be described mathematically with the
derivatives of the values of the variable of interest (denoted x). The net rate of
change is the ®rst derivative of the variable's value (@x/@t). The variable's
second derivative describes the movement of the net rate of change; a positive
second derivative indicates an increasing absolute size of the net rate of
change, while a negative second derivative indicates a decreasing absolute size
of the net rate of change and a second derivative equal to zero indicates a
constant rate of change. Including absolute values to describe behavior greater
and less than initial values produces the following de®nitions:
linear atomic behavior pattern @(j(@x=@t)j)=@t 0 (1)

exponential atomic behavior pattern @(j(@x=@t)j)=@t 4 0 (2)
logarithmic atomic behavior pattern @(j(@x=@t)j)=@t 5 0 (3)
Distinguishing among atomic behavior patterns by visual inspection of variable

plots can be misleading. For example, the early portion of exponential or later
portions of logarithmic patterns can appear linear. However most system
dynamics software includes derivative functions that can model the test for
atomic behavior patterns above for any variable, greatly facilitating the ident-
i®cation of atomic behavior patterns by model analysts.2
Combinations of the three atomic behavior3 patterns can describe most
behavior simulated by system dynamics models. As an example, pattern 1 in
Figure 2 illustrates a combination of two of the three behavior patterns as
generated by a single negative feedback loop. In pattern 2 of Figure 2 the
mathematical de®nition of the atomic behavior patterns identi®es the expon-
ential behavior patterns as those occurring when the atomic behavior pattern
indicator value is positive and logarithmic patterns as those occurring when
the atomic behavior pattern indicator value is negative.
A traditional description of the behavior patterns in Figure 2, pattern 1
would describe them as goal seeking. Our description of behavior patterns
diers from traditional descriptions in its use of only the time-varying value of
the variable of interest and not other system features, such as structure or
system conditions, to describe behavior. Conspicuously absent is any reference
in our de®nition to goals to describe behavior patterns. The three atomic
behavior patterns are also better discriminators of behavior patterns than
traditional means. For example a traditional description of the behavior in
Figure 2, pattern 1 as goal seeking does not distinguish between exponential
and logarithmic behaviors in the pattern, although they are fundamentally
dierent shapes. Scaling and timing changes can make this distinction very
important by causing the growth portion of a negative loop, such as begins at
Fig. 2. Combinations
of atomic behavior
patterns
the minimum value in a limit cycle (e.g. time 8.50±12.75 in Figure 2), to appear
to be exponential growth due to a positive loop thereby leading to errors in
identifying dominant loops.
In addition to an improved de®nition of behavior patterns, a rigorous
behavioral de®nition of dominance requires speci®city concerning the location
of dominance. The location of dominance must be identi®ed more speci®cally
than at the level of a model because dierent variables in a model can have very
dierent behavior patterns in the same time interval. Therefore, the identi®ca-
tion of feedback loop dominance requires the speci®cation of a single system
variable for which dominance is considered important. We refer to this variable
as the ``variable of interest''. Speci®city is also needed concerning the gains that
determine feedback loop strength and therefore determine dominance. The sizes
of these gains can change over the simulation period, potentially changing loop
dominance (Mojtahedzadeh 1997; Kim 1995). Therefore, loop dominance
depends upon the conditions of the rest of the system during the time interval.
The system structure and combinations of parameter values that determine loop
gains in a time interval de®ne the conditions in which the results of dominance
analysis are valid. Completely specifying loop dominance requires specifying
the system structure and conditions under which a given loop dominates.
Based on the above, we de®ne feedback loop dominance as follows: A
feedback loop dominates the behavior of a variable during a time interval in a
given structure and set of system conditions when the loop determines the
atomic pattern of that variable's behavior. This de®nition provides the basis for
an unambiguous and objective test of dominance. As will be illustrated, our
de®nition allows multiple loops to simultaneously dominate a single variable.
This is consistent with an intuitive meaning of dominance (Richardson 1986)
because more than one loop may be required to produce a given behavior
pattern. For example, the success-to-successful system archetype (Senge 1990)
consists of two loops, which are both required to maintain linear behavior
when the system is in equilibrium. While simultaneous multiple dominance is
not precluded in some structural approaches to loop analysis a behavioral
approach expands an analyst's perspective from a search for a single dominant
loop with nondominating in¯uences toward a richer view of how loops drive
behavior that includes multiple dominant loops.
Capturing shadow feedback structures
Richardson describes feedback structures in complex systems as ``a major

source of puzzling behavior and policy diculties'' (Richardson 1991, p. 300).
Forrester (1969) partially attributes this to the in¯uences of multiple feedback
loops on system behavior and the diculties in identifying the actual causes of
behavior. The identi®cation of loop dominance when two or more loops
independently generate similar behavior can create or contribute to these
challenges. These conditions can be generated by simple common structures, as
will be illustrated. The similar behavior of multiple loops can cause an analyst
to misdiagnose which loop dominates the variable of interest or to miss shifts
in dominance among the loops by hiding the in¯uence of some loops behind
the in¯uence of other loops. Such a misdiagnosis can lead to the design and
implementation of policy or system changes to improve performance based on
changing a non-dominant loop. These system changes are likely to fail as a
result of the impacts of the unidenti®ed loops. This confounding behavior and
dominance can be characterized as the hiding of one loop in the shadow of
another loop. Therefore, we refer to these loops as pairs of shadow feedback
loops and this type of structure more generally as shadow feedback structures.
Determining feedback loop dominance in the presence of shadow feedback
structures can be particularly dicult and has not been addressed in previous
research. A rigorous analysis of feedback loop dominance will identify shadow
feedback loops and distinguish them from isolated dominant loops.
One approach to identifying the dominant loop within a pair of shadow
loops is to de®ne loop dominance in broader terms, as in Richardson and
Pugh's (1981) use of the loop with primary responsibility and model behavior
(versus behavior patterns). This approach allows the analyst to distinguish

more ®nely among similar behaviors and potentially identify dominance
changes within shadow feedback structures. But these analyses can be
dicult, because not all behavior changes within a given atomic pattern are
caused by dierences in loop dominance, and are not as rigorous as a binary
test of dominance. Part of the diculty in analyzing shadow feedback
structures is the failure of the removal of the dominant loop in informal
experimental testing to change the behavior pattern and thereby signal the
dominance change from the dominant loop to the shadow loop. Mojtahedza-
deh's (1997) analysis of loop gains may provide a structural solution to dist-
inguishing dominance among shadow feedback loops. A ®rst step in a
behavioral approach to this analysis is the identi®cation of shadow feedback
loop structures. The analysis procedure presented here identi®es the presence
of shadow feedback loops by identifying and deactivating potential shadow
loops, thereby allowing the original loops to signal dominance. Later we will
illustrate the application of the behavioral approach for investigating a shadow
feedback pair in a classic system dynamics model.
The feedback loop dominance analysis procedure
Our analysis procedure, portions of which are currently performed by many

system dynamicists informally, is purposefully behavioral in nature and
structurally simple. We use changes in atomic behavior patterns in the presence
and absence of a feedback loop to signal loop dominance in a chosen time
interval. Shifts in loop dominance across adjacent time intervals are identi®ed
by identifying the dominant loops in each time interval. The procedure
structures the following eight steps into an iterative process for identifying
dominant loops in selected time intervals:
1. Identify the variable of interest4 that will determine feedback loop
dominance and simulate the behavior of the variable of interest over time.
2. Identify a time interval during which the variable of interest displays only
one atomic behavior pattern. This is the reference atomic behavior pattern
and time interval. The system structure and parameter values during this
time interval de®ne the conditions in which dominance is speci®ed.
3. Use the feedback structure of the model to identify the feedback loops that
in¯uence the variable of interest. Select one of those feedback loops as the
candidate feedback loop, beginning with a loop that contains the variable of
interest, if possible.
4. Identify or create a control variable in the candidate feedback loop that is

not a variable in other feedback loops and can vary the gain of the
candidate loop. Use the control variable to deactivate the candidate loop.
5. Simulate the behavior of the variable of interest over the reference time
interval with the candidate feedback loop deactivated and identify the
atomic behavior pattern or patterns of the variable of interest during the
time interval.
6. Identify time intervals, each of which contains a single atomic behavior
pattern.5 Within each time interval:
(a) If the atomic behavior pattern in a time interval generated in step 5 is
dierent than the reference pattern identi®ed in step 2, the candidate
feedback loop dominates the behavior of the variable of interest under
the system conditions during that time interval.
(b) If the atomic behavior pattern of the variable of interest in step 5 is the
same as in the reference pattern identi®ed in step 2, two conditions are
possible: (1) the candidate feedback loop does not dominate or (2) the
candidate feedback loop dominates, but has a shadow feedback loop. To
identify potential shadow loops repeat steps 4 through 6 with the
candidate loop deactivated, eliminating feedback loops individually and
then in sets until the behavior pattern changes. A change in the atomic
behavior pattern identi®es a potential shadow feedback loop or
structure. Deactivate the shadow loop or structure and repeat the test
for dominance of the candidate feedback loop (steps 4±6). If the behavior
pattern does not change with the shadow loop deactivated, the
candidate feedback loop does not dominate (condition 1 above). If the
behavior pattern changes with the shadow loop or structure deactivated,
the candidate feedback loop dominates (condition 2 above). If the
behavior pattern persists through the elimination of all feedback loops,
no feedback loops dominate the behavior of the variable of interest.
7. Repeat steps 3 through 6 with the candidate loop active to test for multiple
dominant feedback loops during the time interval.
8. Repeat steps 1 through 7 for dierent time periods to identify shifts in
feedback loop dominance and feedback loop dominance over other
variables of interest.
Application of the behavioral loop dominance

analysis approach
Three examples using simple structures based on a project model (Ford and
Sterman 1998) will be used to initially illustrate the application of the
behavioral approach to analyzing feedback loop dominance. The equations

which describe the systems used in each of the examples can be found
in Appendix 1. The application of the approach to a more complex model
follows.
A single loop example

The dominant loop in a one-loop system is obvious (assuming feedback
dominates the system behavior). This example illustrates how the approach
elicits a signal of loop dominance and builds a basis for more complex
examples to follow. The ®rst example uses a ®rst-order exponential smooth to
control the completion rate of a set of tasks (Figure 3).
1. Identify the variable of interest that will determine feedback loop domin-
ance and simulate the behavior of the variable of interest over time.
The variable Tasks Completed is selected as the variable of interest. The
behavior of the variable of interest over 40 days is shown in Figure 4.
The time from day 0 through 40 is selected as the time interval. The atomic
behavior pattern indicator is negative, which identi®es a logarithmic
behavior pattern.6 The system conditions under which the dominance
analysis is valid are de®ned by the structure shown in Figure 3, the
equations in Appendix 1 and the parameter value sets that determine the
Fig. 3. Structure
diagram for a single
feedback loop
example
Fig. 4. Reference
behavior for a single
feedback loop
example
gains on the feedback loop. The ranges of these values during the time
interval are:
Tasks Completed 0±1000 tasks

Completion Rate 333±0 tasks per day
Tasks Waiting for Completion 1000±0 tasks
Completion duration 3 days
Scope 1000 tasks.
Feedback loop L1 is selected as the candidate loop.
Feedback loop L1 is deactivated by severing the causal link between the
stock of Tasks Completed and the auxiliary Tasks Waiting for Completion.
This can be done by changing the equation for the Tasks Waiting for
Completion from:
Tasks Waiting for Completion Scope ÿ Tasks Completed

to
Tasks Waiting for Completion Scope.
Fig. 5. Single feedback

loop example;
feedback loop L1
deactivated
time interval.
The behavior of Tasks Completed with loop L1 deactivated is shown in
Figure 5. The atomic behavior pattern indicator is zero during days 0
through 40. This identi®es the behavior pattern as linear.
pattern. Within each time interval:
The atomic behavior pattern in the time interval changes from logarithmic
to linear, indicating that feedback loop L1 dominates the behavior of Tasks
Completed during the time interval from day 0 through day 40 under the
conditions of the system identi®ed in step 2. This helps the analyst
understand that loop L1 constrains the growth of Tasks Completed into a
logarithmic pattern.
Since there are no other feedback loops to consider as candidate loops (step 7)
and no other variables of interest (step 8), the analysis is concluded.
An example of simultaneous multiple loop dominance

To illustrate the application of the procedure to analyze systems that are
simultaneously dominated by multiple loop, a second loop is added to the
Fig. 6. Structure
diagram for a multiple
dominant feedback
loop example
system used in the ®rst example (Figure 6). The new loop describes the avail-
ability of more tasks for completion based on the number of tasks that have
already been completed. In this example a simple but realistic relationship is
assumed in which 10% of the tasks are available initially and the completion of
each task releases another task for completion until all tasks are available.
Such a work availability constraint could describe the availability of ¯oors for
work in a ten-story building project as lower ¯oors are completed (Ford and
Sterman 1998).
1. Identify the variable of interest which will determine feedback loop domin-
The variable Tasks Completed is again selected as the variable of interest.
The behavior of the variable of interest over 40 days is shown in Figure 7.
Fig. 7. Reference
behavior for a
multiple dominant
feedback loop
example
The atomic behavior pattern indicator identi®es that the behavior is linear
from day 0 through 27 and logarithmic from day 27 through 40. Days 0
through 27 are selected as the time interval for analysis. The system
conditions under which analysis is valid are described with the structure
shown in Figure 6, the equations in Appendix 1 and the parameter value
set5 as they vary from day 0 to day 27:
Completion Rate 33.33 tasks per day
Initial Fraction Available 0.10
Rate of increase in Task Availability 1 (percent available per percent
completed)
Fraction Available to Complete 0.10±1.00
Tasks Available to Complete 0±1000 tasks
Fraction Completed 0±0.90
Tasks Waiting for Completion 100 tasks
Scope 1000 tasks.
Fraction Available to Complete and the Tasks Available to Complete
variables. This can be done by changing the equation for the Tasks
Available to Complete from:
Tasks Available to Complete Scope * Fraction Available to Complete
to
Tasks Available to Complete Scope.

time interval.
Deactivating loop L2 reverts the system to the fully activated version of the
system used in the ®rst example and the behavior of Tasks Completed with
loop L2 deactivated is shown in Figure 4. The atomic behavior pattern
during the time frame day 0 through 27 is logarithmic.
The atomic behavior pattern changed from linear to logarithmic over the
entire time interval, indicating that feedback loop L2 dominates the behavior
of Tasks Completed during the time interval 0±27 under the conditions of
the system identi®ed in step 2. This helps analysts understand that loop L2
constrains the system into a linear pattern of completing tasks less quickly
than it would otherwise.
To test for multiple dominant loops feedback loop L2 is reactivated and the
test is repeated with feedback loop L1 (step 3) as the candidate loop. Loop
L1 is deactivated (step 4) by rede®ning the equation for the Tasks Waiting
for Completion from
Tasks Waiting for Completion Tasks Available to Complete ÿ Tasks

Completed
to
Tasks Waiting for Completion Tasks Available to Complete.
The behavior of the variable of interest over the time interval with feedback
loop L1 inactive is shown in Figure 8 (step 5).
The atomic behavior pattern changes from linear (Figure 7) to exponential
(Figure 8) during days 0 through 7 when loop L1 is deactivated (step 6). This
indicates that feedback loop L1 dominates during that time interval (step 6a).
From day 0 through 7 the system behaves logarithmically under the control of
L1 alone (Figure 4) and exponentially under the control of loop L2 alone
(Figure 8), but linearly when both loops L2 and L1 are active (Figure 7). This
indicates that the interaction of the two loops generates the linear behavior and
not either loop alone and therefore they both dominate during days 0±7. The
Fig. 8. Multiple
dominant feedback
loop example;
feedback loop L1
deactivated
procedure and results help the analyst understand that loop L1 restrains the
exponential completion of tasks by loop L2 into a linear pattern and that Loop
L2 restrains the logarithmic completion of tasks by loop L1 into a linear
pattern.
The atomic behavior pattern with loop L1 inactive remains unchanged
(linear) over the time interval 7±27, indicating that loop L1 does not dominate
this time interval. Since there are no other candidate feedback loops that
analysts can use this to understand that during days 7±27 loop L2 dominates
the system behavior alone (step 6b). Applying the analysis to the time interval
27±40 identi®es feedback loop L1 as the only dominant loop. The results of the
analysis can be concisely shown by identifying the time intervals during which
speci®c loops dominate on a graph of the behavior of the variable of interest
with all loops active (Figure 9). Since there are no other variables of interest
(step 8) the analysis is complete.
The multiple dominant loops example demonstrates how the test identi®es
time intervals in which dierent loops dominate the behavior of a variable of
interest, simultaneous multiple loop dominance and the iterative application of
the procedure.
An example with shadow feedback loops

The third example illustrates how the procedure identi®es shadow feedback
loops. A second loop is added to the system used in the ®rst example that
describes the control of work with a ®rst-order exponential smooth to complete
a targeted number of tasks, which is lower than the scope (Figure 10). Such a
Fig. 9. Results of
feedback loop
analysis for a multiple
dominant feedback
loop example
Fig. 10. Structure

diagram for a shadow
feedback loop
example
condition could occur if workers had a completion goal set by labor agreements
as well as a project size.
1. Identify the variable of interest that will determine feedback loop domin-
The variable Tasks Completed is again selected as the variable of interest.
The behavior of the variable of interest over 40 days is shown in Figure 11.
Fig. 11. Reference

behavior for a shadow
feedback loop
example
The time from day 0 through day 40 is selected as the time interval. The
atomic behavior pattern is logarithmic. The system conditions under which
the analysis is valid are described with the structure shown in Figure 10,
the equations in Appendix 1 and parameter value5 set as they vary from day
0 to day 40:
Completion Rate: Scope 333 ÿ 83 tasks per day
Completion Rate: Target 375±0 tasks per day
Completion duration: Target 2 days
Tasks Waiting for Completion 1000 ÿ 250 tasks
Target Completion Gap 750 ÿ 0 tasks
Target Tasks Completed 750 tasks
Scope 1000 tasks.

Completion Rate: Scope and the Completion Rate. This can be done by
setting the Completion Rate: Scope value large enough that it does not
impact the Completion Rate value. The Completion Rate is rede®ned as
follows:
Completion_Rate MIN(Completion_Rate: Scope,
Completion_Rate: Target)
is rede®ned as
Completion Rate Min (1E10, Completion Rate: Target).
time interval.
The behavior of Tasks Completed with loop L1 deactivated is shown in
Figure 12. The atomic behavior pattern during the time interval day 0±40 is
logarithmic.
(b) If the atomic behavior pattern of the variable of interest in step 5 is the
same as in the reference pattern identi®ed in step 2, two conditions are
possible: (1) the candidate feedback loop does not dominate or (2) the
candidate feedback loop dominates, but has a shadow feedback loop. To
identify potential shadow loops repeat steps 4 through 6 with the
candidate loop deactivated, eliminating feedback loops individually and
Fig. 12. Shadow

feedback loop
example; feedback
loop L1 inactive
then in sets until the behavior pattern changes. A change in the atomic
behavior pattern identi®es a potential shadow feedback loop or
structure. Deactivate the shadow loop or structure and repeat the test
for dominance of the candidate feedback loop (steps 4±6). If the behavior
pattern does not change with the shadow loop or structure deactivated,
the candidate feedback loop does not dominate (condition 1 above). If the
behavior pattern changes with the shadow loop deactivated, the
candidate feedback loop dominates (condition 2 above). If the behavior
pattern persists through the elimination of all feedback loops, no
feedback loops dominate the behavior of the variable of interest.
The atomic behavior pattern remained logarithmic over the entire time
interval, indicating that additional testing is required to identify any
shadow feedback loops. Loop L3 is deactivated in addition to loop L1 to test
whether loop L3 is a potential shadow loop (step 4). This can be done by
severing the causal link between the Completion Rate: Target and
Completion Rate: Scope variables by setting the Completion Rate: Target
to a very large value. The Completion Rate is rede®ned as:
Completion Rate Min (1E10, 1E10).
The behavior of the variable of interest over the time interval with feedback
loops L1 and L3 inactive is shown in Figure 13 (step 5). The change in the
atomic behavior pattern from logarithmic to linear indicates that loop L3 is
potentially a shadow loop to loop L1 (step 6). This is tested by repeating
steps 4±6 of the procedure for loop L1 with loop L3 inactive. The results are
shown in Figure 4 (L1 active) and Figure 13 (L1 inactive). The atomic
Fig. 13. Shadow

feedback loop
example; feedback
loops L1 and L3
inactive
behavior pattern with L3 inactive changes from logarithmic to linear when

loop L1 is changed from being active to inactive. The change in the atomic
behavior pattern with loop L3 inactive when there was no change in the
atomic behavior pattern with loop L3 active identi®es loop L3 as a shadow
feedback loop of loop L1. This helps an analyst understand that either loop
L1 or loop L3 can constrain the growth of the Tasks Completed to a
logarithmic pattern.
Repeating the test sequence above with loop L3 as the candidate feedback
loop identi®es loop L1 as a shadow loop to loop L3.
Since the entire simulation time has been analyzed and there are no other
variables of interest (step 8) the analysis is complete.
The ®rst three example applications of the behavioral loop dominance
analysis procedure have illustrated its use in a single-loop system, conditions of
simultaneous loop dominance, and the identi®cation of shadow feedback
structures. Andersen (1998) describes the application of the procedure to a
larger model and the consistency of the results with model sensitivity tests.
A comparison with an alternative analysis method

To compare the behavioral approach with an alternative approach to feedback
loop analysis and to illustrate the approach's application to a larger model the
method will be applied to the portion of a relatively simple (17 equations) model
of the economic long wave by Sterman (1985) analyzed by Kampmann7 (1996a;
1996b) using eigenvalue elasticities. The eigenvalue analysis uses the elasticity
of loop eigenvalues due to changes in loop gains to evaluate contributions of
various loops to behavior. The results of this analysis take the form of
identifying the loop or loops that drive the behavior in each of four phases of
the long wave. While not described in the terms above, the results of this
analysis include descriptions of the dominant loops. We will use the behavioral
approach to analyze the dominance of the loops that the eigenvalue analysis
concludes dominate the referenced long-wave model and compare the
conclusions of the two approaches. For brevity, the steps of the behavioral
approach will be referred to in parentheses and illustrations of behavior and
atomic behavior pattern indicators will be limited to the reference conditions.
To illustrate the eigenvalue analysis approach Kampmann (1996a) analyzes a
third-order model. Figure 14 depicts a simpli®ed structure diagram of the model
with the loops pertinent to this comparison. Complete model equations are
Fig. 14. Simpli®ed

structure diagram of a
long-wave model
(based on Kampmann
1996a)
listed in Appendix 2. The model contains 16 independent feedback loops

(Table 1) which are speci®ed by variable symbol sequence, named and
numbered (Kampmann 1996a, Table 2). We have replaced variable symbols
with descriptive names for clarity, but retain the loop names and numbers to
facilitate comparison of the methods.
Consistent with a structural approach to feedback loop dominance analysis,
Kampmann de®nes dominance with a characteristic of a structural model unit
(eigenvalue elasticities of independent feedback loops). The eigenvalue analysis
identi®es four phases in the behavior of the model (Figure 15) based on four
dominant loops (described below) without focusing on speci®c variables of
interest. Kampmann uses these loops to explain the nominal behavior (rises
and falls) of several important variables. The eigenvalue analysis's structural
de®nition and application to explain nominal behavior contrasts sharply with
our de®nition based on atomic behavior patterns and our use of loop domin-
ance to explain the patterns of single variables of interest. The timing of shifts
in dominance according to the variable-independent eigenvalue elasticity
de®nition of dominance are indicated by the edges of the four phases
Table 1. Feedback
loops in a simple long- Loop no. Loop name Variable sequence
wave model ( from
Kampmann 1996a) L1 Self-ordering/self-eroding Capital orders, Backlog, Desired production, Desired
Capital, Desired Orders, Fractional orders
L2 Hoarding Capital orders, Backlog, Desired Supply Line, Desired
Capital orders, Fractional orders
L3 Supply Line correction Capital orders, Supply, Desired Orders, Fractional
orders
L4 Overtime Backlog, Desired production, Capacity Utilization,
Production
L5 Supply Line-1st order Supply, Acquisitions
control
L6 Capacity correction Capital, Desired orders, Fractional orders, Capital
orders, Supply, Acquisitions
L7 Capital Decay Capital, Depreciation
L8 Capacity expansion Capital orders, Supply, Acquisitions, Capital,
Depreciation
L9 Steady-state Cap. A Capital orders, Supply, Acquisitions, Capital,
Depreciation, Desired orders, Fractional orders
L10 Steady-state Cap. B Depreciation, Fractional orders, Capital orders,
Supply, Acquisitions, Capital
L11 Steady-state Supply line Depreciation, Desired supply, Desired orders,
Fractional orders, Orders, Supply, Acquisitions,
Capital
L12 No easy interpretation Backlog, Acquisitions, Supply, Desired orders,
Relative orders, Orders
L13 No easy interpretation Backlog, Desired production, Capacity utilization,
Production, Acquisitions, Supply, Desired orders,
Fractional orders, Capital orders
L14 Economic growth Production, Acquisitions, Capital, Capacity
L15 Demand balancing Acquisitions, Capital, Capacity, Capacity utilization,
Production
L16 No easy interpretation Backlog, Desired production, Capacity utilization,
Production, Desired supply, Desired orders, Relative
orders, Capital orders
(Figure 15). These times coincide more closely with the apparent times of
dominance shifts, as de®ned by the behavioral approach of some variables (e.g.
Desired Production), than with others (e.g. Capital). These dierences between
the eigenvalue and behavioral approaches to feedback loop dominance analysis
illustrate the critical roles of identifying the purpose of the analysis and the
meaning of dominance and specifying a variable of interest.
Fig. 15. Behavior of a

small long-wave
model (based on
Kampmann 1996a)
Based on the similarities of the model phases with the apparent dominance
shift times of the Desired Production, this variable is selected as the variable of
interest to compare the two analysis methods (step 1). In Phase I the eigenvalue
analysis identi®es the self-ordering feedback loop (L1) as dominant. The atomic
behavior pattern indicator (Figure 16) con®rms a visual inspection of Figure 15,
which indicates that the behavior pattern of the Desired Production in Phase I
is exponential. The eigenvalue analysis describes Phase I as lasting 2±3 years.
A review of the times when the atomic behavior pattern changes polarity
indicates a time interval from year 50 to year 53 (step 2). The end of Phase I
corresponds to the in¯ection point of the ``S''-shaped rise in the Desired
Production. The limits of Phase I coincide with the dominance shift times
indicated by the behavioral method for Desired Production in Phase I.
To test the eigenvalue analysis conclusion that the self-ordering loop domin-
ates Desired Production in Phase I, loop L1 is selected as the candidate loop
(step 3). This loop is deactivated (step 4) by changing the equation of the
Desired Production±Desired Capital link8 from:
Desired_Capital Desired_Production * Capital : Output_ratio

to
Desired_Capital 1.8 * Capital : Output_ratio.
Fig. 16. Reference

behavior for a small
long-wave model
Deactivating the self-ordering loop causes the behavior pattern of Desired

Production to change from exponential to linear throughout Phase I (step 5).
This indicates that the self-ordering loop dominates the Desired Production in
Phase I (step 6a). Our analysis concurs with the conclusion of the eigenvalue
analysis that the self-ordering loop dominates Phase 1.
The eigenvalue analysis describes the continued growth of the Desired
Production in Phase II as being driven by the economic growth and capacity
expansion loops (L14 and L8, respectively) with the economic growth loop
having the larger relevant eigenvalue elasticity component. This implies that
these loops may simultaneously dominate or form a shadow feedback pair. The
atomic behavior pattern indicator helps specify the time interval as years 54
through 62 (step 2), which is consistent with the limits of Phase II as found by
the eigenvalue analysis. The end of Phase II corresponds to the peak of the
initial rise of the Desired Production. The atomic behavior pattern during Phase
II is logarithmic (step 2 and Figure 16). We initially test the economic growth
loop (L14) for dominance (step 3) by cutting the Capacity±Production link (step
4) by changing:
Production Capacity * Capacity Utilization

to
Production 1.8 * Capacity Utilization.
The resulting behavior is linear throughout Phase II (step 5). The change in
atomic behavior pattern from logarithmic to linear supports the eigenvalue
analysis conclusion that the economic growth loop dominates Desired
Production during Phase II (step 6a). To test the capacity expansion loop
(step 3), we reactivate the economic growth loop and deactivate the capacity
expansion loop by cutting the Depreciation±Capital Orders link (step 4) as
follows:
Captial Orders Depreciation * Fractional Fractional orders

is replaced with
Capital Orders 0.16 * Fractional Fractional orders.
This change reduces the length of the entire long wave, but retains the
characteristic shapes of the Desired Production cycle found in the reference
behavior, including the exponential Phase I followed by a logarithmic Phase II,
which ends at the peak of the Desired Production cycle (step 5). Deactivating
the capacity expansion loop did not change the atomic behavior pattern. This
indicates that either the capacity expansion loop does not dominate Phase II or
it dominates with a shadow feedback loop (step 6b). We test whether the
economic growth and capacity expansion loops form a shadow feedback loop
pair by testing the dominance of the capacity expansion loop (L8) with the
economic growth loop (L14) inactive. As described above, the behavior of the
Desired Production with loop L14 inactive and loop L8 active is linear during
Phase II. Deactivating both loops also results in linear behavior. Since the
dominance test for the capacity expansion loop does not change depending on
the economic expansion loop, we conclude that the capacity expansion loop
does not dominate Phase II and that the capacity expansion and economic
growth loops are not a shadow feedback structure (step 6b). This does not
contradict the results of the eigenvalue analysis, which assesses the relative
in¯uence of dierent loops on behavior. These two loops can be the most
in¯uential feedback structures in causing the Desired Production to rise (the
eigenvalue approach conclusion) and have only one of those loops determine
the shape of that rising behavior (the behavioral approach conclusion).
The eigenvalue analysis identi®es Phase III with the precipitous decline of
the Desired Production (years 63 through 66) and Phase IV with the gradual
recovery of the Desired Production (years 67 through 95). The self-eroding
feedback loop (L1) is considered to dominate Phase III. The reference atomic
behavior pattern indicator (Figure 16) suggests that two loops dominate this
phase, one generating exponential decay in the early portion of Phase III and
another generating logarithmic decay in the later portion of Phase III. A repeat
of the analysis steps described above for Phase I results in linear behavior
throughout Phase III, indicating that the self-eroding loop dominates Phase III.
The behavioral analysis also identi®es the capital decay loop (L7) as
simultaneously dominant with loop L1 in Phase III and dominant in Phase IV.
An increasing strength of the capital decay loop in Phase III can explain the
logarithmic shape of the behavior pattern in the later portion of Phase III. The
dominance of the capital decay loop in Phase IV identi®ed with the behavioral
approach concurs with the conclusions of the eigenvalue analysis. As in Phase
II, the eigenvalue and behavioral analyses dier in their conclusions but are
both correct because of the dierence in their de®nitions and uses of loop
analysis.
In summary, the eigenvalue and behavioral analyses reach the same general
conclusions concerning which loops dominate the long-wave model but dier
somewhat in when shifts in loop dominance occur. Dierences can be
attributed to dierences in analysis purposes, de®nitions of dominance and
possibly the linearization required for eigenvalue analysis. Kampmann
discusses several of these issues as they relate to the eigenvalue analysis
approach.
Evaluation of the behavioral loop dominance

analysis approach
Based on the preceding and other applications of the behavioral approach to

feedback loop dominance analysis we make the following assessments.
1. The method generates explicit and precise results and the conditions in
which those results are valid.
2. The procedure is explicit and applicable to any system dynamics model for
the analysis of any variable with simple modeling tools and methods.
3. The method can isolate speci®c model variables and feedback loops for
analysis. This allows the analyst to perform partial model analysis and
investigate feedback loops of particular interest (e.g. loops believed to be
high-leverage or loops that can be in¯uenced).
4. The approach rigorously separates structure and behavior. Behavior
patterns and analysis results do not depend on the polarity of feedback
loops or behavior pattern estimates based on loop polarity. The procedure
does not require distinctions among structural components such as stocks,
¯ows, and auxiliaries.
5. The procedure identi®es simultaneous domination by multiple loops. This
provides a more accurate description of how structure drives behavior than
procedures that identify only one dominant loop. However, this ability
may increase the challenges of communicating analysis results since
simultaneous multiple loop dominance may be more dicult to understand

and describe than the results of tests that identify only one dominant loop.
6. The procedure identi®es shadow feedback structures, but does not provide
an explicit and completely rigorous method for identifying the dominant
loop within shadow feedback structures. Shadow feedback structures in
which more than two loops create similar behavior are more dicult to
identify and analyze, potentially causing an analyst to conclude that there is
a shadow feedback loop pair when in fact there is a shadow feedback loop
triplet or larger structure.
7. The procedure fully validates its results only through the testing of many
loop combinations. Its rigor is therefore limited by the diculty of
identifying and isolating feedback loops. This can be partially addressed
through a partial-model testing approach (Homer 1983) to feedback loop
analysis, but does not resolve the challenges of analyzing large complex
models and loops that integrate model subsystems.
8. The diculty of specifying the system conditions under which loop analysis
results are valid increases with model size.
9. The procedure allows non-unique methods of deactivating feedback loops.
Dierent deactivation methods may generate dierent behavior patterns.
Our assessment indicates that the behavioral approach to feedback loop
dominance analysis provides clear advantages over informal and undocumen-
ted analysis procedures, but has weaknesses that can act as warnings for its
application and guidance in the development of automated feedback loop
analysis tools.
Conclusions
The rigorous identi®cation of dominant feedback loops in models is required to

explain how structure drives behavior. Our behavioral approach de®nes domin-
ance with atomic behavior patterns and identi®es simultaneous dominance by
multiple loops and shadow feedback structures. Our analysis procedure uses
changes in atomic behavior patterns to signal dominance. Illustrations of the
application of the procedure show its ability to identify and distinguish among
dominant structures better than informal analysis methods. A direct com-
parison with an alternative analysis approach supports the validity of analysis
results using our method and emphasizes the critical roles of analysis purpose
and the de®nition of dominance. Our approach contributes to the analysis of
feedback loop dominance by improving the de®nition of loop dominance and
de®ning multiple dominance and shadow feedback structures, two important

types of feedback structures. We have also codi®ed an applicable analysis
procedure. This will allow analysis methods to be implemented, evaluated and
improved by many dynamists simultaneously and, as suggested by Richardson
(1995), bridge between intuition-based analysis and more formal solutions.
Our behavioral approach to feedback loop dominance analysis shows
promise as a tool for the analysis of small-to-medium size and large simple
models and for the development of improved understanding and explanation of
how feedback drives behavior. However, the limitations of its manual
application and the complexity of dominance in shadow feedback structures
emphasize the need for automated model analysis tools based on rigorous
mathematical de®nitions of both feedback loop dominance and behavior
patterns. Future research can further validate our procedure, expand our initial
investigations of simultaneous dominance and shadow feedback structures
and integrate our behavioral perspective with structural approaches to
feedback loop dominance analysis.
By proposing and testing a more speci®c and explicit de®nition of feedback
loop dominance and analysis method, our work raises new questions that must
be resolved to fully address the feedback loop analysis of system dynamics
models. Is the purpose of feedback loop analysis to diagnose nominal model
behavior, explain the shapes of behavior plots, educate about how structure
drives behavior or some combination thereof? Are dierent types of dominant
loops, such as those that determine patterns versus those that determine the
timing of dominance shifts, more or less important for dierent purposes? Are
dierent de®nitions, tools, and methods needed for dierent loop analysis
purposes? How can feedback loop analysis bridge the gap between the com-
plexity of most feedback models and the limited capacity of humans to
comprehend dynamic complexity? The continued development of analysis tools
and methods can provide the basis for improving our understanding and use of
feedback loop dominance to improve system performance.
Appendix 1: Equations of models used in simple

application examples
Models use timestep 0.0625 and the Euler integration method. The equation
for the atomic behavior pattern indicator used in all models is:
Atomic_Behavior_Pattern_Indicator
DERIVN(ABS(DERIVN(Variable_of_Interest,1)),1)
Single loop system

Tasks_Completed(t) Tasks_Completed(t ÿ dt) (Completion_Rate) * dt
INIT Tasks_Completed 0
Completion_Rate Tasks_Waiting_for_Completion/Completion_duration
Completion_duration 3
Scope 1000
Tasks_Waiting_for_Completion Scope ÿ Tasks_Completed
Simultaneous multiple loop dominance

Completion_Rate Tasks_Waiting_for_Completion/Completion_duration
Fraction_Available_to_Complete MIN(1,Initial_Fraction_Available
Rate_of_increase_in_Task_Availability * Fraction_Completed)
Fraction_Completed Tasks_Completed/Scope
Initial_Fraction_Available 0.10
Rate_of_increase_in_Task_Availability 1
Tasks_Available_to_Complete Scope * Fraction_Available_to_Complete
Tasks_Waiting_for_Completion
Tasks_Available_to_Complete ÿ Tasks_Completed
Scope 1000
Shadow feedback structure

Completion_Rate MIN(Completion_Rate : _Scope,Completion_Rate : _Target)
Completion_Rate : _Scope Tasks_Waiting_for_Completion/
Completion_duration
Completion_Rate : _Target Target_Completion_Gap/
Completion_duration : _Target
Completion_duration : _Target 2
Tasks_Waiting_for_Completion Scope ÿ Tasks_Completed
Target_Completion_Gap Target_Tasks_Completed ÿ Tasks_Completed
Scope 1000
Target_Tasks_Completed 750
Appendix 2: Equations of a simple long-wave model
Timestep 0.25 years; start time year 4; Euler integration method.
Stocks
Backlog(t) Backlog(t ÿ dt) (Capital_Orders : _backlog Goods_Orders ÿ
Production) * dt
INIT Backlog Normal_delivery_delay
Capital(t) Capital(t ÿ dt) (Acquisitions ÿ Depreciation) * dt
INIT Capital (Capital : Output_ratio * Average_Life_of_Capital)/
(Average_Life_of_Capital ÿ Capital : Output_ratio)
Supply(t) Supply(t ÿ dt) (Capital_Orders : _supply ÿ Acquisitions) * dt
INIT Supply (Backlog/Production) * Depreciation
Flows
Capital_Orders Depreciation * Relative_Orders
Acquisitions Supply * (Production/Backlog)
Depreciation Capital/Average_Life_of_Capital
Production Capacity * Capacity_Utilization
Goods_Orders 1
Auxilliaries
Capacity Capital/Capital : Output_ratio
Capital_adjustment (Desired_Capital ÿ Capital)/Capital_adjust_time
Desired_Capital Desired_Production * Capital : Output_ratio
Desired_Orders Depreciation Capital_adjustment Supply_adjustment
Desired_Production Backlog/Normal_delivery_delay
Desired_Supply Depreciation * (Backlog/Production)
Supply_adjustment (Desired_Supply ÿ Supply)/Supply_adjust_time
Capacity_Utilization GRAPH(Desired_Production/Capacity)
(0.00, 0.00), (0.2, 0.3), (0.4, 0.6), (0.6, 0.8), (0.8, 0.9), (1, 1.00), (1.20, 1.03),
(1.40, 1.05), (1.60, 1.07), (1.80, 1.09), (2.00, 1.10)
Relative_Orders GRAPH(Desired_Orders/Depreciation)
( ÿ 0.5, 0.00), (0.00, 0.2), (0.5, 0.5), (1.00, 1.00), (1.50, 1.50), (2.00, 2.00),
(2.50, 2.50), (3.00, 3.00), (3.50, 3.50), (4.00, 4.00), (4.50, 4.40), (5.00, 4.80),
(5.50, 5.20), (6.00, 5.50), (6.50, 5.65), (7.00, 5.70), (7.50, 5.75), (8.00, 5.80)
Average_Life_of_Capital 20
Capital : Output_ratio 3
Capital_adjust_time 1.5
Normal_delivery_delay 1.5
Supply_adjust_time 1.5
Notes
1. The three pattern names chosen re¯ect the general shapes and not
mathematical structures. See Mojtahedzadeh (1996. p. 54) for an alternative
set of de®nitions and pattern names.
2. Due to the use of dierence equations by simulation software to perform
calculations, two timesteps are required to generate an accurate second
derivative. Therefore, the ®rst two timesteps may generate false positive
values for the atomic behavior pattern indicator. These should be ignored.
3. One notable exception is discrete ¯ows such as those caused by pulses or
the instantaneous release of work (Ford and Sterman 1998).
4. Any model variable may be the variable of interest.
5. Deactivating some candidate loops that do not include the variable of
interest alters when dominance shifts between loops that dominate adjacent
time intervals but not which loops dominate the majority of those time
intervals. Deactivating these ``dominance timing'' loops can generate two
behavior patterns in a single reference time interval as well as candidate
loops that change the loops that dominate. The relative importance of
identifying which loops dominate and when dominance shifts determines
the importance of these dominance timing loops. See the ®nal example
application for an example.
6. The atomic behavior pattern indicator on the right side of Figure 4
illustrates how gradual logarithmic and exponential patterns can generate
atomic behavior pattern indicator values that closely approximate, but are
not equal to, zero. Inspecting several signi®cant digits of indicator values or
magnifying the size of the indicator value with a positive multiplier
facilitates distinguishing these conditions from linear atomic behavior
patterns.
7. The model analyzed by Kampmann diers slightly from the model in
Sterman (1985) but retains the important structures. The model analyzed by
Kampmann has been used here to facilitate comparison of the analysis
methods.
8. The second simulated wave and parameter values approximating steady-

state or initial values are used to deactivate feedback loops throughout the
comparison, so as to avoid results generated by the system's transitional
response to initial conditions.
References
Andersen, P. 1998. Water Allocation in Arid and Semi-Arid Regions. Masters thesis.
University of Bergen, Bergen, Norway.
Davidsen, P. 1992. The structure±behavior diagram: understanding the relationship
between structure and behavior in complex dynamic systems. In Proceedings of
System Dynamics Conference 1992, ed. J. A. M. Vennis. Utrecht: The Systems
Dynamics Society, 127±140.
DoÈrner, D. 1996. The Logic of Failure, Recognizing and Avoiding Errors in Complex
Situations. Reading, MA: Addison-Wesley.
Ford, D. N. and J. D. Sterman. 1997. Dynamic modeling of product development
processes. System Dynamics Review 14(1): 31±68, Portland, OR: Productivity Press.
Forrester, J. W. 1961. Industrial Dynamics. Portland, OR: Productivity Press.
Forrester, J. W. 1969. Urban Dynamics. Cambridge, MA: Productivity Press.
Forrester, N. B. 1982. A Dynamic Synthesis of Basic Macroeconomic Theory:
Implications for Stabilization Policy Analysis. Report D-3384. Systems Dynamics
Group, Sloan School of Management, Massachusetts Institute of Technology,
Cambridge, MA.
Graham, A. K. 1977. Principles on the Relationship between Structure and Behavior of
Dynamic Systems. Ph.D. thesis. Massachusetts Institute of Technology, Cambridge,
MA.
Homer, J. 1983. Partial-model testing as a validation tool for system dynamics.
Proceedings of 1983 International System Dynamics Conference. Boston, MA:
Systems Dynamics Society.
Kampmann, C. E. 1996a. Feedback loop gains and system behavior. Unpublished draft.
Kampmann, C. E. 1996b. Feedback loop gains and system behavior. Proceedings of the
1996 International System Dynamics Conference. Cambridge, MA: Systems
Dynamics Society.
Kim, D. H. 1995. A new approach to ®nding dominant feedback loops: loop-by-loop
simulations for tracking feedback loop gains. System Dynamics, an International
Journal of Policy Modeling 7(2): 1±15.
Mojtahedzadeh, M. T. 1996. A Path Taken: Computer-Assisted Heuristics for
Understanding Dynamic Systems. Ph.D. thesis. Rockefeller College of Public Aairs
and Policy, University at Albany, State University of New York, Albany, NY.
Richardson, G. P. 1986. Dominant structure. System Dynamics Review 2(1): 68±75.
Richardson, G. P. 1991. Feedback Thought in Social Science and Systems Theory.
Philadelphia, PA: University of Pennsylvania Press.
Richardson, G. P. 1995. Loop polarity, loop dominance, and the concept of dominant
polarity (1984). System Dynamics Review 11(1): 67±88.
Richardson, G. P. and A. L. Pugh III. 1981. Introduction to System Dynamics with

Dynamo. Portland, OR: Productivity Press.
Senge, P. M. 1990. The Fifth Discipline, The Art and Practice of the Learning
Organization. New York: Doubleday.
Sterman, J. D. 1985. A behavioral model of the economic long wave. Journal of
Economic Behavior and Organization 6: 17±53.

A Behavioral Approach To Feedba - Desconhecido

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

A Behavioral Approach To Feedba - Desconhecido

Caricato da

Copyright:

Formati disponibili

A behavioral approach to feedback loop

David N. Ford is an Abstract

to errors in identifying dominant loops. Informal analysis procedures are

A behavioral definition of feedback loop dominance

DoÈrner (1996) refers to behavior patterns as `` `shapes' in time'' and considers

linear atomic behavior pattern @(j(@x=@t)j)=@t  0 (1)

Distinguishing among atomic behavior patterns by visual inspection of variable

Capturing shadow feedback structures

Richardson describes feedback structures in complex systems as ``a major

(versus behavior patterns). This approach allows the analyst to distinguish

The feedback loop dominance analysis procedure

Our analysis procedure, portions of which are currently performed by many

4. Identify or create a control variable in the candidate feedback loop that is

Application of the behavioral loop dominance

behavioral approach to analyzing feedback loop dominance. The equations

A single loop example

Tasks Completed  0±1000 tasks

Tasks Waiting for Completion  Scope ÿ Tasks Completed

Fig. 5. Single feedback

An example of simultaneous multiple loop dominance

atomic behavior pattern or patterns of the variable of interest during the

Tasks Waiting for Completion  Tasks Available to Complete ÿ Tasks

An example with shadow feedback loops

Fig. 10. Structure

Fig. 11. Reference

Feedback loop L1 is deactivated by severing the causal link between the

Fig. 12. Shadow

Fig. 13. Shadow

behavior pattern with L3 inactive changes from logarithmic to linear when

A comparison with an alternative analysis method

Fig. 14. Simpli®ed

listed in Appendix 2. The model contains 16 independent feedback loops

Fig. 15. Behavior of a

Desired_Capital  Desired_Production * Capital : Output_ratio

Fig. 16. Reference

Deactivating the self-ordering loop causes the behavior pattern of Desired

Production  Capacity * Capacity Utilization

Captial Orders  Depreciation * Fractional Fractional orders

Evaluation of the behavioral loop dominance

Based on the preceding and other applications of the behavioral approach to

simultaneous multiple loop dominance may be more dicult to understand

The rigorous identi®cation of dominant feedback loops in models is required to

de®ning multiple dominance and shadow feedback structures, two important

Appendix 1: Equations of models used in simple

Single loop system

Simultaneous multiple loop dominance

Shadow feedback structure

Appendix 2: Equations of a simple long-wave model

Timestep  0.25 years; start time  year 4; Euler integration method.

8. The second simulated wave and parameter values approximating steady-

Richardson, G. P. and A. L. Pugh III. 1981. Introduction to System Dynamics with

Potrebbero piacerti anche

linear atomic behavior pattern @(j(@x=@t)j)=@t 0 (1)

Tasks Completed 0±1000 tasks

Tasks Waiting for Completion Scope ÿ Tasks Completed

Tasks Waiting for Completion Tasks Available to Complete ÿ Tasks

Desired_Capital Desired_Production * Capital : Output_ratio

Production Capacity * Capacity Utilization

Captial Orders Depreciation * Fractional Fractional orders

simultaneous multiple loop dominance may be more dicult to understand

Timestep 0.25 years; start time year 4; Euler integration method.