The Hard Core of the System Dynamics Research
Programme
Mohamed Saleh
University of Bergen
Introduction
In this research, we attempt to identify the set of basic assumptions underlying the
system dynamics research programme. This work isnspired by Lakatos’ theory of
science (Lakatos, Musgrave 1970). Lakatos calls the set of basic assumptions of any
programme, the hard core of the programme. The hard core of a programme is the
defining characteristic of a programme. According to Lakatos, “the hard core of a
programme is rendered unfalsifiable by the methodological decision of it
protagonists” (Chalmers, 1982, p. 81). One is unwilling to give up this hard core,
even if one meets resistance in the form of empirical counteevidence, anomalies and
scientific arguments. The research to explain the anomalies is sought in the other
assumptions, auxiliary hypothese, initial conditions or what Lakatos calls the
protecting belt of the programme.
I think that the research to identify the hard core is very crucial in understanding and
evaluating the system dynamics research programme. Many of the criticisms (which
will be illustrated later, when discussing the validation issue) of the system dynamics
esearch programme stemmed from a misconception of the hard core of the
programme. To understand the origin of such a misconception, we have to recall
Kuhn's claim (Kuhn, 1970) that rival theories are "incommensurable" (cannot be
measured by the same standards). This stems from the fact that proponents of rival
theories will subscribe and adhere to different sets of basic assumptions. Judged by its
own set of standards (that is a direct result of its set of basic assumptions) theory A
may be judged superior to theory B, while if the standards of theory B are used as
premises, the judgment may be reversed. The conclusion of an argument is
compelling only if its premises areaccepted. Supporters of rival theories will not
accept each other's premises and so they will criticize each other from different
perspectives.
As Meadows puts it:
Different modeling worldview, or in Thomas Kuhn terminology,
paradigms, cause their practitioners to define different problems,
follow different procedures, and use different criteria to evaluate
their results. In a very real sense the paradigm biases the way the
modeler sees the world, and thus influences the content and shape
of his models. "If the only tool you have is a hammer, you tend to
treat everything as if it was a nail". ... It often leads to sterile
arguments across paradigms, each school criticizing the problems,
assumptions, and standards of the other from the biased
perspectives of its own problems, assumptions, and standards.
(Meadows, 1980, p.24)
In the next section, the historical roots of system dynamics will be briefly illustrated.
According to Lakatos, the discussion of a specific programme should include a
description of the historical context in which it emerged. Then the most important
point in this paper, that is the identification of the hard core of system dynamics, will
be explained in detail. The last section in this paper is the conclusion.
The Historical C ontext of System Dynamics
System Dynamics is a computer-aided approach to policy analysis and design. It
applies to dynamic problems arising in complex social, managerial, economic,
ecological systems, or any dynamic systems characterized by interdependence, mutual
interaction, information feedback and circular causality.
System dynamics was originally developed by Jay Forrester at MIT in the 1950s. His
classical book "Industrial Dynamics" (Forrester, 1961) is still an important statement of
the philosophy and methodology in system dynamics. Initially, system dynamics was
named "Industrial Dynamics", as it was targeted to corporate and industrial problems.
Only later, the name was changed (evolved) to system dynamics to reflect the breadth
and universal nature of the research programme. The name system dynamics can
suggest the presence of a link between the system dynamics research programme and
other system-based research programmes. In fact, the links are weak and misleading. It
will be shown later that the system dynamics research programme is different from the
general system theory and cybemetics.
Engineering control theory “servomechanisms” (Ogata, 1997) is the parent research
programme of system dynamics. System dynamics took the concepts of engineering
control theory and applied it to social, managerial and other domains. In the system
dynamics research programme, the methods of management have been extended and
improved by methodologies developed in engineering (Forrester, 1961; Richardson,
1981; Richardson, 1991; Richardson 1996). What emanated most strongly to the social
science from engineering control were:
¢ Control or management requires the feedback of information. In particular, the
necessary information is the “error signal", which is the difference between the goal
and the actual condition of the system.
¢ Drawing visual diagrams is vital for the conceptualization of a feedback model.
Control theory introduced the "block diagram" which was the first formal visual
feedback loop diagram ever. As we will see later, system dynamics adopted the
stock-and-flow diagram to portray feedback.
Nowadays many managers and consultants use system dynamics as an effective tool to
redesign their organizations’ infrastructure. System dynamics is used by managers that
want to solve an array of problems, which are resilient to traditional ways of thinking;
managers who want to make their organizations more effective, and see their visions
come through.
According to Jay Forrester’s philosophical view, corporate executives should be
corporate designers (engineers); they should create the policies for their middle
managers to be able to make informed decisions. The problems of the corporation will
be eliminated by changing the structure, and the policies of the corporation; and the
only people in a position to do so are the executives (Forrester, 1961).
The ultimate goal of managers and consultants is to design a leaming organization
(Senge, 1994). An organization that is capable of solving current problems, and seizing
current opportunities. At the same time, the organization invests in its capacity to
obtain and keep the leading edge in the future. The power of the learning organization
stems from the ability of its members to enhance and expand their collective
awareness and capabilities.
The Hard Core Pyramid
From my point of view (my own reconstruction of the programme based on my
experience, education and intuition) the system dynamics’ hard core consists of a
multi-level pyramid of basic assumptions. Each level sets a foundation on which the
upper levels res. This pyramid consists of six levels. At the first level is the causality
assumption. Then comes the assumption that the accumulation process is the
fundamental process in nature. Then the basic assumption regarding the "endogenous
origin of cause" at the next higher level. Moving up to the fourth level, we find the
fluid representation metaphor. Then there is the assumption that human mental
inference has a limited capacity. Finally, we reach the assumption that each model by
itself is a theory. The folbwing figure illustrates the hard core pyramid of the system
dynamics research programme.
A Model is a Causal
Human Mental Inference
Fluid Representation Metaphor
Endogenous Origin of Cause
Accumulation Process
Causality
Fig. 1 The hard core pyramid of the system dynamics research programme.
The Causality Assumption
The System dynamics research programme assumes the existence of causal
relationships governing the behavior of social systems. There are many questions and
doubts about the concept of causality in social systems (Lessnoff, 1974; Elster 1983).
From a methodological philosophical perspective, it is impossible to absolutely affirm
the presence of any universal causal relationship between two variables. In the System
Dynamics programme, it is assumed that concept of causality has a meaning in social
science. It is also assumed that we can form closed loops of circular causality (this
point will be explained in more details in the endogenous origin of cause assumption).
The following Forrester’s quote briefly explains the causality assumption in the system
dynamics programme. As Forrester puts it:
The nature of causality should perhaps be a subheading under
philosophy. In system dynamics, causality is implied by the
structure of feedback loops. Causality is unidirectional from
element to element around a loop. Causality is not ambiguous or
reversible. From a system dynamics viewpoint, water flowing
from a faucet causes the water level in a glass to rise. And the
level of water in the glass can cause a person to shut off the
faucet. But I have had people argue that no such sense of
causality is justified, that it is just as correct to assert that the
water flows from the faucet because the water level in the glass is
rising. I do not understand the basis for such a perspective, but
confusion does seem to exist regarding causality and its
relationship to modeling. An indifference to the importance of
causality must underline search in the social sciences for
statistical correlations. Statistical correlations can show that one
variable "explains" another in the sense that changes in the two
accompany each other. But such correlation needs not to imply a
direction of causality between the two variables, or even that a
direct connection exists. They may both vary as a consequence of
a third influence... The social sciences need more emphasis on
causality and less on correlation and statistical "explanation".
(Forrester, 1980, p. 15)
The Accumulation Process as the Fundamental Process in Nature
Forrester argues that: "Nowhere does Nature differentiate; in real systems, dynamic
change arises only from accumulation, that is, integration’ (Forrester, 1980, p.15).
This is an ontological assumption about reality. In the modeling of real systems,
Forrester’s focus is on the accumulation (integration) process, rather than on the
differentiation process. Y et, the main focus of mathematicians is on the differentiation
process. This focus on the differentiation process can be traced back in history to
"Isaac Newton", the founder of the differential calculus field. There are two reasons
behind the mathematicians’ choice to focus on differentiation. First, the direction of
causality is not important to mathematicians. Second, the differentiation process is
mathematically simpler than the integration process.
To illustrate this point, I will retum to the "water' example given in the previous
section (in Forrester’s quote); according to Forrester the water level in a glass rises
due to the accumulation process of the water flowing from the faucet. But if we
focused (as many mathematicians do) on the flow (the mathematical derivative), that
will imply that the flow exists because the level of water is changing. This
interpretation can be a valid mathematical statement, but is totally disassociated from
the real dynamics of our world.
Forrester believes that many students become confused, when being taught dynamics
through the mathematics of differential equations, and that the proper way of
understanding dynamics is through the accumulation processes that characterize
change.
In reality, the accumulation process is the most fundamental process from which the
dynamic behavior of systems originates. Without this process, there will be no
dynamics at all; you could not have a dynamic system that does not encompass this
process. Yet, for the very reason that it is so fundamental, many authors oversee it
when listing the factors that are responsible for the dynamic behavior of complex
systems.
The Endogenous Origin of Cause
The endogenous origin of cause assumption is one of the most fundamental concepts in
system dynamics. It implies the existence of some other assumptions such as: closed
causal boundaries, feedback, nonlinearity, etc.
The Endogenous origin of cause assumption dictates that the root causes of a "solvable
problem" associated with a particular system are contained within the internal structure
of the system itself. A solvable problem is defined as a problematic behavior that can
be adjusted (controlled) by human intervention. As long as the problem of interest is a
solvable one (and many of the modem world's serious problems are solvable), then the
root causes of the problematic behavior are not a consequence of unavoidable
exogenous disturbances, but rather arise from the complex relationships of the
structure of the system. It is an inward-looking perspective. It assumes that the
essential dynamic behavior of a system results from the inner workings of the system
itself. The focus is intemal, not extemal. The endogenous point of view creates a
dramatically different problem focus from that of the extemal view. While the extemal
view places us at the mercy of some unavoidable exogenous events, the internal view
inspires us to search for causes (that can be solved) from within our own system. While
the external view directs us to identify an uncontrollable source outside the system to
blame, the intemal view forces us to look for the reason within the system. Thus the
system that created the problem can be modified by man, and there is hope to solve the
problem. This is the major benefit of the endogenous point of view.
To fully understand the previous paragraph, we need to further understand the
relationship between man and system. As Skinner describes it:
Man himself may be controlled by his environment
[system], but it is an environment [system] which is
almost wholly of his own making ... When a person
changes his physical or social environment [system]
"intentionally". that is, in order to change human
behavior, possible including his own _ he plays two roles:
one as a controller, as the designer of a controlling
culture, and another as the controlled, as the product of
the culture. There is nothing inconsistent about this; it
follows from the nature of the evolution of a culture, with
or without intentional design. (Skinner, 1971, p. 196)
The endogenous assumption necessitates the presence of a closed causal boundary for
the system. This closed boundary separates the dynamically significant inner workings
of the system from the dynamically insignificant external environment. While, in
general system theory a closed system is defined as a materially closed one, in system
dynamics a closed system is defined as a causally closed one. In system dynamics, we
are focusing on the problem and its causes, rather than on a materially closed "whole"
system. A system dynamicist looks for the boundary that encompasses the smallest
number of components within which the dynamic behavior of the problem is generated.
Those components are capable by themselves, without exogenous aid, to reproduce
the essential characteristics of the problematic behavior. A problem focus acts as a
critically important filter that screens out unnecessary details and centers our attention
on the significant aspects of the system. The typical case is to have one problem per
system dynamics model (where a system dynamics model is a representation of a
closed causal system), with the implication that the modeler should exclude everything
(from the model) that does not contribute to the problem behavior. The basic idea in
modeling is, therefore, to represent a portion of reality, sufficiently large to represent
the current problem, yet small enough to exclude everything that is of no significant
relevance.
The modeler should also choose the appropriate level of aggregation for the model.
The level of aggregation is directly related to the time-horizon of the model. In a long
term horizon several smaller components of a system get so thoroughly intertwined,
that it makes more sense to aggregate them. The time horizon is in turn dependent on
the problematic behavior we want to understand. In system dynamics, we focus on a
certain mode of behavior, and deliberately exclude some of the other modes of
behavior present. For example a model of long term economic growth may exclude the
business cycle fluctuations. The time horizon must be adequate to the mode of
behavior we are interested in. The golden rule in aggregation is that the modeler
should aggregate to the greatest extent possible, while making the policy variables (the
variables the manager can use to solve the problem), the goal variables (the variables
describing the problematic behavior), and the causal relationships between them
visible in the model (Wils, 1976).
The feedback view of system dynamics can be seen as a consequence of the closed
causal boundary and the endogenous assumption. If feedback loops were not present in
a closed boundary system, then all causal links would at the end have to be connected
to exogenous factors outside the boundary of the system. This will make the behavior
of the system a result of those exogenous factors. Since system dynamicists have an
intemal perspective to problems (as illustrated before), the existence of the feedback
loop concept is inevitable.
George Richardson demonstrates in his book titled: "Feedback thought in Social
Science and Systems Theory" (Richardson, 1991), that feedback thinking is one of the
most penetrating patterns in all social science, and that it is embedded in the very
foundations of social science as a fundamental building block. As Richardson puts it
"Great social scientists are feedback thinkers; great social theories are feedback
thoughts" (Richardson, 1991, p. 2). There is, however, no consensus among social
scientists about that. Y et, Richardson has cited outstanding quotes of several famous
social thinkers (like Adam Smith, Hegel, Marx, Thomas Malthus, John Dewey, John
Staurt Mill, etc.), which illustrate that the feedback concept was largely implicit in their
works. For example, the feedback concept is largely implicit in the following quote by
Adam Smith, in which he explains the effect of price on production:
If.....the quantity brought to market should at any time
fall short of the effectual demand, some of the component
parts of its price must rise above their natural rate. If it is
rent, the interest of all other landlords will naturally
prompt them to prepare more land for the raising of this
commodity; if it is wages or profits, the interest of all
other labourers and dealers will soon prompt them to
employ more labour and stock in preparing and bringing
it to market. The quantity brought thither will be
sufficient to supply the effectual demand. All the different
parts of its price will soon sink to their natural rate, and
whole price to its natural price. (Smith, 1776, p. 57)
In system dynamics, a feedback loop is defined as a closed sequence of causes and
effects (circular causality), a closed path of action and information. Also, in system
dynamics, a feedback system is defined as a one of an interconnected set of feedback
loops.
Feedback thinking can be seen as contrary to open-loop thinking. In a traditional open-
loop thinking, after we discover a problem, we design a policy plan to deal with it, and
then act on our plan, and that is the end of the problem-solving process. It is usually
forgotten that our action will alter the state of the system, cause secondary effects, and
thus may generate a new set of problems (for which new policies must be made), or
may redefine the problem in a new context. The feedback thinking approach aims at
closing such open loops, and completing the link between our actions and the problem
that initiated our actions.
Feedback loops are naturally divided into two distinct categories, which are labeled
positive and negative. This label is called the polarity of the feedback loop. The
concept of polarity gives the feedback loop analytical and explanatory power.
Negative feedback loops are stabilizing loops, which oppose or negate any
destabilizing disturbances. Negative feedback loops are also called stabilizing,
equilibrating, or self-correcting loops.
In contrast, positive feedback loops are destabilizing loops, which amplify any
disturbances. Positive feedback loops are also called destabilizing, self-reinforcing
loops. Sometimes they are called vicious circles.
A common feature of a feedback loop is the presence of delay in the flow of
information and material throughout the loop. An order of goods does not immediately
result in a delivery of goods, crops planted cannot immediately be harvested, the
process of averaging sales takes time, new ideas take time to spread, etc.
Delays have the tendency to dramatically change the behavior of a model. According
to Jay Forrester, “delays are crucial in creating the dynamic characteristics of
information feedback systems” (Forrester, 1961, p. 86). Negative feedback loops can
exhibit oscillating behavior if there is a major delay in it. Also, delays usually attenuate
the amplification power of a positive feedback loop.
Given the behavior characteristics of the problems addressed by system dynamics, then
the endogenous origin of cause assumption forces system dynamicists to include
nonlinearities in the structure of their models. To understand the reason, we must first
understand the behavior characteristics of the problems that the system dynamics
programme attempts to solve. Many problems that the programme addresses are
characterized by an unstable, non-linear, self-limiting behavior, or in short, complex
10
behavior. Complex behavior is a typical characteristic of many natural phenomena in
the universe. From a pure mathematics point of view, Kovach observed:
Strange that these nonlinear phenomena that abound so widely
in nature should be so interactable. It is almost as if Man is to
be denied a complete knowledge of the universe unless he
makes a superhuman effort to solve its nonlinearities... ..So
far, our efforts to scale the nonlinear barrier have consisted of
chiseling a few footholds. (Kovach, 1960, pp. 218-225)
Also systems in social sciences, including economics have nonlinear behavior. Jay
Forrester in his paper titled: "Nonlinearity in high-order models of social systems"
(Forrester, 1987), has clearly demonstrated and has given several examples to this fact.
Complex behavior can never be generated, without exogenous aid (the endogenous
origin of cause assumption), by a linear feedback model. A typical control model
(recall that the control theory is the parent programme of system dynamics) will be a
linear feedback one. That is because most engineering applications do not exhibit such
kind of complex behavior. Y et, Jay Forrester had to step out of the linear world into
the nonlinear universe to be able to address the problems in social science exhibiting
complex behavior.
In such kinds of complex behavior, no static view of the feedback loops of the model is
therefore sufficient. It is necessary for the feedback loops to change endogenously their
relative strength of influence as conditions (states) change to generate such complex
behavior. In system dynamics terminology, we call this the ability of the model to
endogenously shift its dominant loops. Dominant loops are loops that are primarily
responsible for the behavior of the model over an interval of time. Loop dominance
usually shifts among a number of loops in the course of time. For example, the self-
limiting behavior (the so-called S-shaped behavior) can be generated by two coupled
feedback loops. One loop is positive and the other one is negative. In the beginning the
positive feedback loop is dominant and this generates the exponential growth behavior
and then, as the model changes its state, the negative feedback loop dominates and the
saturation behavior results.
11
The endogenous shift in loop dominance takes place as a consequence of nonlinearities
in the equations defining the model's structure. If these equations were linear, no such
shift in loop dominance would occur, and only one fixed set of loops will continuously
dominate the model. In system dynamics, we expect models to change their dominant
structure over time. Consequently, we focus on non-linear models. As Forrester puts
it: “Indeed, from a feedback loop perspective, this ability to shift loop dominance is the
fundamental reason for advocating nonlinear models of social system behavior”
(Forrester, 1987, p. 105).
The concept of shifting loop dominance and the presence of nonlinearities are vital
supporting pillars for the feedback concept. As George Richardson puts it:
In fact, without the concept of shifting loop dominance,
without nonlinearities in formal models of complex systems,
the feedback concept is justifiably perceived to be a weak tool
incapable of capturing important dynamics in real systems.
(Richardson, 1996, pp. 149-150)
That is the reason for the emphasis system dynamicists put on the concept of the
feedback loop dominance. As John Sterman puts it:
The concept of feedback-loop dominance is central to the system
dynamics paradigm. In complex systems _high order, multi-loop
nonlinear feedback systems behavior over time depends on
which of the many feedback processes in the system dominates.
At any moment in the evolution of the system, some feedback
loops will be highly influential and other will be inactive.
Because of nonlinearities, loop dominance often shifts as the
behavior unfolds. As the system enters new regions of state
space, latent loops that have had no role in the behavior may
suddenly become active, causing qualitative changes in the mode
of behavior. Endogenous shifts in loop dominance are
responsible for shifts from, for example, exponential growth to
decline or from stability to instability_ they are responsible for
bifurcations. These nonlinear interactions lie at the heart of
many important behaviors observed in models. Understanding
the nature of loop dominance and the factors that cause it to
shift endogenously is critical to the design of robust policy.
(Richardson, 1984, p.67)
Structural changes are considered to be of significance and are handled quite
differently in cybernetics and in system dynamics programmes. In traditional
cybemetics, structural changes are captured linguistically and sometimes
12
diagrammatically, for example by redrawing the system structure. In the more
quantitative programme, the system dynamics programme, structural changes are
represented by endogenous shifts in loop dominance, which are capable of changing
the active structure over time (Forrester, 1987).
Finally, I would like to point out the principle outstanding problem in the system
dynamics programme. In the system dynamics programme, it is assumed that there is a
closed feedback loop (a hidden meta-loop) between structure and behavior (Davidsen,
1991). Shift in loop dominance is a result of the dynamic behavior of the model itself.
These shifts in loop dominance change the active structure of the model. As the active
structure changes, so does the dynamic behavior of the model, which in tum further
shifts the dominance of the loops, and so on. Understanding the mechanisms of this
meta-loop is the principle outstanding problem in the system dynamics programme.
The goal is to understand the mechanisms of this hidden meta-loop in the model first
(ie. assuming that the model is our virtual world), then to reflect our understanding of
the model on the real system. System dynamics has a great potential in describing the
complexity of the real world (something that many modeling programmes lacks), yet
currently the explanatory power of system dynamics is not adequate. The major
challenge, in future, for the system dynamics programme, is to develop new concepts
and tools that can enhance the explanatory power of the programme through
innovative ways of "tracing" in depth the mechanisms of this hidden meta-loop;
otherwise the programme is under the threat of reaching a “crisis” state.
The Fluid Representation Metaphor
From a pure mathematical point of view, the basic structure of a formal system
dynamics computer simulation model is a system of coupled, nonlinear first-order
integral equations. Higher-order equations can be constructed by cascading several
first-order ones. A system that can be represented this way is called a state-determined
system in the engineering literature.
Jay Forrester has invented the fluid representation metaphor to hide the mathematical
complexity of Integration. Recall that (as discussed before) the integration
(accumulation) process is the most fundamental process from which the dynamic
behavior of systems originates. System dynamicists take the simplifying view that
13
feedback systems involve continuous, fluid-like processes. The idea of Jay Forrester
was to describe in nonmathematical terms the engineer's notions of a state-determined
system (Forrester, 1961).
Jay Forrester has used the term stock as analogous to the term state variable in the
engineering field. The level of the stock is analogous to the value of the state variable.
The term level is intended to invoke the image of the level of a liquid accumulating in a
tub. The variables increasing and decreasing a stock are called flows. The speed of a
flow is called the rate. In engineering terminology, rates are the components of the
derivatives of the variable states. Jay Forrester has also invented the stock-and-flow
feedback diagram to facilitate the conceptualizing of the model's structure and the
communicating of the model-based insights (Forrester, 1961). Flow diagrams portray
flows and stocks as valves and tubs, thus further emphasizing the analogy between the
integration process and the fluid accumulation process. The Flow diagram enables the
modeler to visually distinguish integral equations from algebraic equations. Integral
equations are represented as conservative fluid flows, while algebraic equations are
represented as non-conservative information links. The distinction between information
and conserved flows is a crucial insight stemming from our new information era.
I would like to note that in system dynamics, a stock can represent measurable, as well
as non-measurable (soft) variables. Some soft non-measurable variables can be very
crucial to understand and solve our problems. Examples of soft variables are moral,
love, happiness, etc. The usual practice in system dynamics is to numerically scale
those soft variables. Currently, there is research to use fuzzy logic to represent soft
variables (De Kok, 1997).
The clarifying power of levels and flows increases when we take a continuous view of
a system. Although a discrete view, focusing on specific events and decisions, is
entirely compatible with the endogenous assumption, Jay Forrester, by choice, adopted
a continuous point of view, despite the fact that a continuous view of a system is
usually hard to achieve. One usually sees discrete events and makes discrete decisions.
Jay Forrester in his seminal book "Industrial Dynamics" argued that a continuous view
14
helps us to focus on the most important aspects of the systems, and that the dynamics
of a continuous flow model are easier to understand. As Forrester puts it:
A continuous flow model helps to concentrate attention on the
central framework of the system. This framework is more
orderly and unchanging than usually expected. Diversion of
attention toward separate isolated events tends to obscure the
central structure of the system that we are trying to define.
(Forrester, 1961, p. 65)
According to Jay Forrester, in the dynamic thinking process one should focus on
patterns over time rather than on discrete events, one should group individual events
to form a pattern of behavior. To be consistent with this continuous point of view,
Forrester chose to view decision making as a continuous process.
[Decision-making] is a conversion mechanism for changing
continuously varying flows of information into control signals
that determine rates of flow in the system. The decision point
is continually yielding to the pressures of the environment.
(Forrester, 1961, p. 96)
Forrester’s focus was on policy rather than discrete decisions.
Policy is a formal statement giving the relationship between
information sources and resulting decision flows. It is what
has often been referred to in the literature as a decision rule. In
physical systems, particular in the field of servomechanisms,
the corresponding term is “transfer function". (Forrester,
1961, pp. 96-97)
Forrester saw both an overt and an implicit policy structure in an organization. Overt
decisions are the conscious decisions people make, while implicit decisions are
unavoidable results of the physics of the system (Forrester, 1961, pp. 102-103).
The net result of Forrester’s view of the decision process was the connection of the
feedback concept with the decision making process at a certain level of aggregation
determined by the emphasis on continuity.
After designing the fluid representation metaphor, Jay Forrester elevated it from a
representation schema to the level of principles of a feedback system. He presented
several principles to indicate the way feedback-loops must be captured. Nowadays, the
system dynamics modeling packages enforce system dynamicists to adhere to
Forrester ’s principles. Forrester’ s principles distinguished feedback loops as portrayed
15
in system dynamics models from other types of feedback loops. Some examples of
Forrester’s principles are listed below. The reader can refer to Jay Forrester book
"Principles of Systems" (Forrester, 1968) for a more complete list of principles.
1) A feedback loop consists of two distinctly different types of variables: the levels
(states) and the rates (decisions or actions). Except for constants, these two are
sufficient to represent a feedback loop. Both are necessary.
2) Levels integrate (or accumulate) the results of actions in a system.
3) Levels are changed only by rates.
4) Levels and rates are not distinguished by units of measure. The identification must
recognize the difference between a variable created by integration and one that is a
policy statement in the system.
5) Rates are not instantaneously measurable. No rate can, in principle control another
rate without an intervening level variable.
6) Level variables and rate variables in a feedback loop must altemate.
7) Levels (state variables) completely describe the system condition (state)
8) Levels exist in conservative subsystems; they can be changed only by moving the
contents between levels (or to or from a source or a sink).
9) Information is not a conservative flow; information is not depleted by its use.
10) Decisions (rates) are based only on available information.
The Limited Capacity of Human Mental Inference
This assumption states that unassisted human-beings cannot infer the behavior of
complex dynamic models. In a complex model, feedback loops are constantly
interacting, gaining, or losing strength relative to each other (shift in loop dominance),
and thus creating a behavior that cannot be inferred by the unaided human mind.
Man cannot assimilate and process all the information required to mentally generate
the behavior of a complex dynamic model. Also, man can at most look a few steps
ahead in future.
This assumption claims that computers must be used to simulate the behavior of such
models. System dynamics allows the application of logic to reveal essentially
tautological, but in practical terms, hidden results.
16
As Jay Forrester puts it
The fourth foundation for industrial dynamics progress is the
electronic digital computer that became generally available
between 1955 and 1960. Without, the vast amount of work to
obtain specific solutions to the characteristics of complex
system would be prohibitively expensive ... The appearance of
high-speed electronic digital computers has removed the
practical computational barrier. (Forrester, 1961, p.18)
The feedback structures of real problems are often so complex so that the behavior
they generate over time can be only traced by simulation. The Latin verb simulate
means to imitate or mimic. The idea behind simulation is to mimic the real system so
that the behavior can be identified.
One can argue that this assumption is just a technical issue. Y et, in my point of view,
this assumption points to an important philosophical difference between system
dynamics and the general "System Thinking" programme. General system thinkers use
causal loop diagrams and generic structures to illustrate common behavior, but they
imply that a system's behavior can be inferred from the feedback loops alone. Or they
advocate "fitting" a problem to a generic structure and then suggest a behavior that
allegedly will originate from the generic structure. In the same spirit, they suggest how
the structure can be modified to alleviate the problem. For example, some system
thinkers in cybemetics (e.g., Beer, 1981; Beer, 1995) map the parts of the company to
the parts of the body and then giving a generic "answer" accordingly. General system
thinkers do not necessarily use computer-based simulation models to relate structure to
behavior. The System dynamics programme, on the other hand, takes the information
about a system's structure that normally remains hidden in mental models and
formalizes it into a computer model. The behavior generated by that particular
structure is only revealed when the model is made subject to simulation. Model
simulation is a prerequisite in the system dynamics research programme.
System Dynamics Models as Causal Theories
Each model by itself is a theory. A model is a concatenation of causal laws offering a
plausible representation of a real system and hence an explanation of a certain
behavior.
17
The general goal of a system dynamics model is understanding. The aim is to
understand how the intemal structure of the system generates the problematic
behavior.
The dynamic hypothesis in a system dynamics based theory is a statement describing
the smallest system structure that is postulated to endogenously generate the
problematic behavior. The dynamic hypothesis is usually given as a descriptive diagram
of interconnected closed causal loops. To test the dynamic hypothesis, a formal
quantitative system dynamics model is built (based on the dynamic hypothesis) and
then through comparing the model with reality, we can validate the model and hence
the dynamic hypothesis.
After understanding the cause of the problem, the modeler should try to understand
what policies could improve the problematic behavior and for what reasons. The final
goal of a modeling study is the application of model insights; the aim is to help some
people manage some activities better in future. Policies could be formal or informal.
Informal policies are consequences of habits, intuition, personal interests, and social
pressures and power within the organization. On the other hand, a formal policy is
explicit and enjoys a formal awareness of the reasons for action.
In system dynamics, it is not enough to know that a particular policy improves model
behavior. The critical question is why. No manager or executive will be convinced to
implement a policy in the real world just because it made the model behave better. The
manager must have a clear understanding of the reason of the problem, and a clear
understanding of the reason that makes a particular policy solves the problem. What
managers really need is a change in their mental model. This is the essence of building
a system dynamics model: that is to elicit the mental model of a manager and to built a
formal model based on this model; the computer simulations can consequently be used
in testing this formal model (which can never be done on the manager’s mental model
as we explained before), for the purpose of detecting contradictions and flaws in the
manager’s mental model, hence correcting them and discovering the real reason of the
18
problem. Then we can experiment with the formal model (using computer simulations)
to design policies that solve the problem.
Management flight simulators and interactive leaming environments are largely used to
communicate policy insights. An interactive leaming environment is not the same as a
management flight simulator. A management flight simulator can be considered a part
of an interactive learning environment. While a management flight simulators engages
the user in simple interactions with a simulation, an interactive leaming environment is
firmly established on instructional design theory, which is based on the cognitive
leaming theory (Davidsen & Spector 1997). Interactive leaming environments are
transparent boxes that demonstrate to the leamers the relationships between key
system variables and the most influential delays and feedback loops in the system. It is
important for the learners to acquire an understanding of the structure of the system, if
the learning goal is to understand the behavior of the system.
Archetypes are also used to summarize and communicate generic insights about
complex system dynamics models. The word "archetype" comes from the Greek word
“archetypoes" meaning the first of its kind. System archetypes were originally
developed at Innovation Associates in the mid-1980s. The aim of archetypes is to help
catalogue the most commonly seen behaviors as being generated by simple generic
structures. As Peter Senge puts it:
Archetypes are accessible tools with which managers can
quickly construct credible and consistent hypotheses about the
governing forces of their systems. Archetypes are natural
vehicle for clarifying and testing mental models about those
systems. They are powerful tools for coping with the
astonishing numbers of details that frequently overwhelm
beginning system thinkers. As you work with archetypes, and
they become second nature, they will become part of your
diagnostic repertoire. You will be able to talk about
systematic issues at a surprisingly sophisticated level. (Senge,
1994, p. 121)
Not all system dynamicists have exactly the same set of archetypes in their mind, but
there is a common set of archetypes shared among most system dynamicists. This
common set evolves and changes as the system dynamics programme matures.
According to Peter Senge, “some archetypes, including ‘Limits To Growth’ and
19
‘Shifting the Burden’, where translations of ‘generic structures’ mechanisms which
Jay Forrester and other systems thinking pioneers had described in the 1960s and
1970s” (Senge, 1994, p.121).
After presenting and communicating the model insights (which will hopefully change
the mental models of the managers), comes the hardest part in any system dynamics
project, that is the implementation of the proposed policy in the real system. It is really
the most crucial and difficult phase. Policy studies are usually faced with many
uncertainties in the real environment, and many implementation obstacles in the real
system. Therefore the modeler has to identify the extent to which the policy outcomes
are sensitive to changes in the uncertain factors in the environment (i.e. the degree of
robustness of the policy), and has to identify the extent to which the policy
recommendations are implementable.
Finally, I would like to point out here a controversial issue in the various modeling
programmes for many years; an issue that has witnessed a heated debate over the last
40 years. This issue is model "validation". From this long debate, one can summarize
the various critiques of the system dynamics validation methodology, in one common
general critique: that system dynamics does not employ objective, quantitative model
validation tests (Barlas, 1989; Barlas, 1990; Barlas, 1996; Bell & Senge, 1980;
Forrester & Senge, 1980; Sterman, 1984).
I will base my defense of the system dynamics validation methodology on three
philosophical arguments. The first argument is that rival programmes are
“incommensurable" (see the introduction section). The second argument is that the
system dynamics programme adopts a sophisticated falsificationism philosophy, rather
than a naive falsificationism one. The third argument is that a system dynamicist does
not assume that his model is the only objective representation of a real system, but
rather assumes that his model is one of many possible ways of describing the real
system. I will explain each argument in details in the following paragraphs. Yet, I
would like to point out here, that in my opinion (I will give reasons below for that
opinion) the system dynamics validation methodology is "bolder" than the validation
methodologies of other rival modeling programmes. The word "bolder" here means
20
that the number of "potential falsifiers" of a system dynamics model are more than the
number of "potential falsifiers" of other types of models.
To understand the first argument (rival programmes are incommensurable), we have to
recall Kuhn's claim that in some cases the disagreement between scientists as to which
of a pair of theories to prefer arises from disagreements about the purpose of the two
theories, or in other words the relative importance of various values by which one can
evaluate the two theories. As Kuhn describes it:
What it should suggest, however, is that such reasons
{accuracy, simplicity, fruitfulness and the like} function
as values and they can thus be differently applied,
individually and collectively, by men who concur in
honouring them. If two men disagree, for example, about
the relative fruitfulness of their theories, or if they agree
about that but disagree about the relative importance of
fruitfulness and, say, scope in reaching a choice, neither
can be convicted for a mistake. Nor is either being
unscientific. (Kuhn, 1970, pp. 199-200)
This is the case in the different modeling schools; the disagreement is about the
purpose of various types of models. To understand the origin of this disagreement we
have to distinguish between two basic types of models: pure correlation models and
pure causal models (Meadows, 1980). Pure correlation models are simply constructed
from observed statistical correlations among various elements of a real system. A pure
correlation model is an empirical data-driven model; one cannot compare or (validate)
any relationship in the model against any known theory that explains this relationship.
Moreover, correlation models usually use dummy variables that do not exist in reality.
The pure correlation model is said to be valid if the model output behavior matches the
observed data to a certain degree of numerical (quantitative) accuracy. This is
primarily because the only purpose of a pure correlation model is to predict the short-
term future. Normally nobody expects the pure correlation model to explain the causes
of the behavior (short-term or long-term). On the other hand, pure causal models
(much like system dynamics models) are much like a scientific theory in the sense that
they attempt to explain how the real world really works (a realism philosophy). That is
the primary purpose of a system dynamics model. Thus it makes sense that the essence
of model validity (in system dynamics) lies always in structural validity. That is, "the
21
right behavior for the right reasons", not the "right behavior for the wrong reasons"
(like in pure correlation models where you introduce imaginary dummy variables to get
the right behavior).
Now, I will explain why the validation methodology of causal models is "bolder" than
that of correlation models. As a system dynamics model is a concatenation of many
causal relationships, then, if a single causal relationship is rejected (because it
contradicts reality), then the whole model must be rejected. Also any variable in the
model must have a real world counterpart. Thus the potential falsifiers of a system
dynamics model are much more than those of a pure correlation model. In system
dynamics, the term “structure validation” is used to describe the process of
corroborating the model through checking those potential falsifiers. Moreover the
output behavior of a system dynamics model has to qualitatively match the observed
data. In system dynamics, the term “behavior validation” is used to describe this
process. To end this argument, I would like to point out that system dynamicists rely
on a broader range of information sources for creating, and validating the model. While
pure correlation models are based only on numerical information, system dynamics
models are based on mental (knowledge stored mentally in people's heads), written
(information stored descriptively in writings) and numerical (numerical databases)
information. As Forrester suggests: "The amount of available information declines,
perhaps by a factor of a million, in going from mental to written information and again
by another factor of a million in going from written to numerical information"
(Forrester, 1980, p. 555).
The second argument was that system dynamics adopts a sophisticated falsificationism
philosophy rather than a naive one. According to the naive falsificationism philosophy,
all system dynamics models (created till now) are false models, and all of them must be
rejected. There is no such thing as a "perfect" system dynamics model. Perfection here
means that the model behavior is an exact duplicate (point by point) of the real system
behavior, and that there is no discrepancy between the output behavior of the model
and the system real behavior. In fact, understanding a "perfect" model would be as
difficult as understanding the real world. Modeling, by virtue of its name, however, is a
simplification process applied to real world phenomena in order to understand the
22
important aspects of world. Any system dynamics model has some simplifying
assumptions that lead us to ignore some minor effects. Also many system dynamics
models deliberately exclude some modes of behavior and focus only on the problematic
mode of behavior (as we mentioned before). According to the sophisticated
falsificationism point of view, the discrepancy between the output behavior of the
model and the historical data can be traced back to those simplifying assumptions, and
the model can be accepted as long as the model can be corroborated (structurally and
behaviorally validated), and as long as the model can predict novel phenomena.
The third argument is that a system dynamicist does not assume that his model is the
only objective representation of a real system, but rather assumes that his model is one
of many possible ways of describing the real system. A system dynamics model has
some subjective nature, in the sense that it is influenced (biased) by the modeler’s
background knowledge, his ideology, etc. This does not mean that system dynamics
models are not built through direct observations of the real world, (it neither means
that models are not validated through direct observations of the real world), but rather
means that "neutral observations" are something impossible in this world. The fact that
system dynamics rely heavily on mental information (rather than just numerical
information as many other programmes) makes system dynamics models more
susceptible to "subjective observations". Y et, as we described before, the essence of
system dynamics modeling is to map the subjective mental model of the manager
(which must be subjective, otherwise he should not be a manager) to a formal
computer model, that can be made subject to experimentation through various
simulations.
Conclusion
In this paper, we identified the hard core of the system dynamics program, and
illustrated its coherence; through presenting the hard core as a multi-level pyramid of
basic assumptions, where each level in this pyramid sets a foundation on which the
upper levels rest. We then discussed each assumption in details.
We discussed the most controversial issue in the programme, which is the validation
issue. We found that it is directly related to the incommensurablility concept (as
23
defined by kuhn's theory of science). One must understand the basic assumptions and
values of the system dynamics programme to adequately evaluate a system dynamics
based theory.
We also identified the principle outstanding problem in the system dynamics
programme, that is to fully understand the mechanisms of the hidden meta-loop
between structure and behavior.
Acknowledgments
The author wishes to thank Pal Davidsen and Anders Molander for their inspiring
discussions, and illuminating ideas. They both gave valuable comments and objective
critiques on the preliminary drafts of this paper.
The author also wishes to thank George Richardson, David Lane, John Barton, and all
other system dynamicists who actively participated in a “brain storming” discussion
(that was initiated by the author in the electronic mailing list of the system dynamics
society) about the hard core of the system dynamics research programme.
24
References
Barlas, Y . (1989): Multiple Tests for Validation of System Dynamics Type of
Simulation Models. European Joumal of Operational Research.
Barlas, Y .; Carpenter, S. (1990): Philosophical roots of model validation two
paradigms. System Dynamics Review 6, no. 2.
Barlas, Y . (1996): Formal Aspects of Model Validity and Validation in System
Dynamics. System Dynamics Review Vol 12.
Beer, S. (1981): The Brain of the Firm (2nd Edition). Chichester, England. John Wiley.
Beer, S. (1995): The Heart of Enterprise: The Managerial Cybemetics of Organization.
John Wiley & Sons.
Bell, J.; Senge, P. (1980) : Methods for Enhancing Refutability in System Dynamics
Modeling. Tims Studies in the Management Sciences 14. North-Holland Publishing
Company.
Bjomestad, S. (1991): A Research Programme for Object-Orientation. Obligatorisk
innlegg I vitenskapsteori for Dr. Polit. Programmet. University of Bergen, Norway.
Chalmers, A. (1982): What is this thing called Science? Second edition. Open
University Press.
Cover, J. (1996): Introduction to System Dynamics. Powersim Press.
Davidsen, P. (1991): The Structure-Behavior Graph. System Dynamics Group. MIT,
Cambridge.
Davidsen, P.; Spector, M (1997): Cognitive Complexity in System Dynamics Based
Learning Environments. 15" International System Dynamics Conference. Istanbul,
Turkey.
De Kok, J.; Wind, H.; Coffa, A.; Van Densen, W.; Pet-Soede, L. (1997): Fuzzy Logic
as a Method for the A pplication of Qualitative Concepts in a Quantitative System
Framework. 15" International System Dynamics Conference. Istanbul, Turkey.
Elster, J. (1983): Explaining Technical Change. A Case Study in the Philosophy of
Science. Cambridge University Press.
Forrester, J. (1961): Industrial Dynamics. Productivity Press
Forrester, J. (1968): Principles of Systems. MIT Press.
Forrester, J. (1980): System Dynamics _ Future Opportunities. TIMS Studies in the
Management Sciences; pp. 7-21. North-Holland Publishing Company.
25
Forrester, J. (1980): Information Sources for Modeling the National Economy. Journal
of the American Statistical Association Vol. 75; pp. 555-574.
Forrester, J.; Senge, P. (1980): Tests for Building Confidence in System Dynamics
Models. TIMS Studies in the Management Sciences. North-Holland Publishing
Company.
Forrester, J. (1987): Nonlinearity in High-order Models of Social Systems. European
Journal of Operational Research Vol. 30; pp. 104-109.
Kovach, L. (1960): Life can be so Nonlinear. American Scientist Vol. 48; pp. 218-225.
Kuhn, T. (1970): The Structure of Scientific Revolutions, Second Edition. The
University of Chicago Press.
Lakatos, J.; Musgrave, A. (1970): Criticism and The Growth of Knowledge.
Cambridge University Press.
Lessnoff, M. (1974): The Structure of Social Science. London: Allen and Unwin.
Meadows, D. (1980): The Unavoidable A Priori. In Randers, J. (ed.), Elements of the
System Dynamics Method. Productivity Press.
Ogata, K. (1997): Modern Control Engineering. Third Edition. Prentice-Hall
Intemational.
Randers, J. (1980): Elements of the System Dynamics Method. Productivity Press.
Richardson, G.; Pugh, A. (1981): Introduction of System Dynamics Modeling with
Dynamo. Productivity Press.
Richardson, G. (1984): Loop polarity, loop dominance, and the concept of dominant
polarity. System Dynamics Review Vol. 11; pp. 67-88.
Richardson, G. (1991): Feedback Thought in Social Science and Systems Theory.
University of Pennsylvania Press.
Richardson, G. (1996): Modelling for Management. University Press, Cambridge.
Saul, I.; Carl, M. (1996): Encyclopedia of operations research and management
science. Kluwer Academic Press.
Senge, P.; Ross, R; Smith, B.; Roberts C.; Kleiner, A. (1994): The Fifth Discipline
Fieldbook. Nicholas Brealey Publishing.
Skinner, B.F. (1971): Beyond Freedom and Dignity. Bantum Books, New Y ork.
26
Smith, A. (1776): An Inquiry into the Nature and Causes of the Wealth of Nations.
Edwin Cannan. New Y ork.
Sterman, J. (1984): Appropriate Summary Statistics for Evaluating the Historical Fit of
System Dynamics Models. Dynamica Volume 10, Part II.
Wils, W. (1976): Aggregation. 5" International System Dynamics Conference. Geilo,
Norway
27
