Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
dominance analysis
David N. Forda
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
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.
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.
6 System Dynamics Review Volume 15 Number 1 Spring 1999
Fig. 1. Atomic
behavior patterns
Ford: Feedback Loop Dominance Analysis 7
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:
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
Ford: Feedback Loop Dominance Analysis 9
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.
Fig. 3. Structure
diagram for a single
feedback loop
example
Ford: Feedback Loop Dominance Analysis 13
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:
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.
Feedback loop L1 is selected as the candidate loop.
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.
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:
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.
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.
6. Identify time intervals, each of which contains a single atomic behavior
pattern. 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.
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.
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-
ance and simulate the behavior of the variable of interest over time.
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
16 System Dynamics Review Volume 15 Number 1 Spring 1999
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.
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:
Tasks Completed 0±900 tasks
Completion Rate 33.33 tasks per day
Completion duration 3 days
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.
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.
Feedback loop L2 is selected as the candidate loop.
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.
Feedback loop L2 is deactivated by severing the causal link between the
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.
5. Simulate the behavior of the variable of interest over the reference time
interval with the candidate feedback loop deactivated and identify the
Ford: Feedback Loop Dominance Analysis 17
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
18 System Dynamics Review Volume 15 Number 1 Spring 1999
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.
Fig. 9. Results of
feedback loop
analysis for a multiple
dominant 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-
ance and simulate the behavior of the variable of interest over time.
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.
2. Identify a time interval during which the variable of interest displays only
one atomic behavior pattern. This is the reference atomic behavior pattern
20 System Dynamics Review Volume 15 Number 1 Spring 1999
and time interval. The system structure and parameter values during this
time interval de®ne the conditions in which dominance is speci®ed.
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:
Tasks Completed 0±750 tasks
Completion Rate: Scope 333 ÿ 83 tasks per day
Completion Rate: Target 375±0 tasks per day
Completion duration 3 days
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.
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.
Feedback loop L1 is selected as the candidate loop.
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.
Ford: Feedback Loop Dominance Analysis 21
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
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.
26 System Dynamics Review Volume 15 Number 1 Spring 1999
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:
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
28 System Dynamics Review Volume 15 Number 1 Spring 1999
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:
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
Ford: Feedback Loop Dominance Analysis 29
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.
Conclusions
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)
32 System Dynamics Review Volume 15 Number 1 Spring 1999
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)
34 System Dynamics Review Volume 15 Number 1 Spring 1999
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.
Ford: Feedback Loop Dominance Analysis 35
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.
36 System Dynamics Review Volume 15 Number 1 Spring 1999