SD uniqueness and the challenge presented by complex systems

This forum contains all archives from the SD Mailing list (go to http://www.systemdynamics.org/forum/ for more information). This is here as a read-only resource, please post any SD related questions to the SD Discussion forum.
Locked
Shelley Michael L Civ AFIT/ENV <
Junior Member
Posts: 2
Joined: Fri Mar 29, 2002 3:39 am

SD uniqueness and the challenge presented by complex systems

Post by Shelley Michael L Civ AFIT/ENV < »

Alan Graham posted a recent "PS" asking for dialog on "what makes SD unique
and how do we preserve that while avoiding insularity?" In my view, SDs
unique contribution is the ability to formulate the correct system boundary
for addressing the question at hand. Classically, the modeler will pose a
reference behavior for a system entity that the client agrees is a good
representation of the problem being presented. The experienced modeler then
quickly represents the influence structure that is known to give rise to
that behavior, both "good" behavior and "bad" behavior depending on which
loop is dominating under certain system state conditions. After correlating
the structure components with actual system entities that make intuitive and
mechanistic sense and formulating the model to demonstrate reference
behavior, the modeler then "sells" the client on the notion that this simple
structure with aggregated (but real) entities (not generally referred to in
practice) and void of many actual system entities (deemed unimportant to the
objective) is the model formulation that should be used to most effectively
address system optimization toward the stated objective. Refinement and
control policy recommendations then focus on improving the "behavioral
pattern". Other modeling methods cannot approach this level of efficiency
and achieve this quality of policy solutions because they are using the
wrong model structure in the attempt, regardless of how mechanistically
correct the structure is. They tend to flounder in the dark, looking for
solutions by trial and error, where as the SD modeler can see how the system
structure must be modified to achieve desired behavior because he or she
knows why the system is behaving the way it is in the first place.

In the case of business and organizational management, I think it is easier
to preserve this unique approach and bring the client along because SD
practitioners have a good collection of success stories, and these systems
tend to lend themselves well to the approach in the availability of simple
system structure that is easily recognized by the client as "real". The
modeler simply needs to move the client beyond or around certain sticky
points like his need to see a statistical analysis of model output
comparison to empirical system data. John Sterman suggests some ways to do
that in one of his earlier papers (Dynamica 10, 1984). I believe such
practices should be avoided if at all possible and used only if necessary to
move the client around this point in order to get him to focus on what he
needs to focus on (improving general system behavior patterns using a model
structure that should already be sufficiently validated for the purposes at
hand by classical SD validation tests).

In disciplines looking at different kinds of systems like complex engineered
or natural physical systems, there is much more of a challenge. These
scientists and engineers have a very difficult time accepting a simple
system boundary that addresses the question being posed. They will general
hold that any model that fails to include all known mechanisms relevant to
the appropriate temporal and spatial scales as a weaker model, regardless of
its purpose. Furthermore, it is often very difficult for the modeler to
express the objective in terms of reference behavior. There are a number of
papers being generated in the literature (several from my own students!)
that claim "a system dynamics approach to...<<some natural or engineered
ecosystem performance analysis>>". Most of these models are not really
formulated in the classical SD manner but are simply straight model
formulation of known system mechanisms with careful attention to
representing feedback relationships where we can envision their existence.
In the end I generally feel that we have failed to envision the proper
system boundary or gained any insight into a higher level system structure
that elucidates a better way to view the system toward effective management
policies. I fear that these papers do not represent the SD method well and
are doing a disservice to the SD community in helping to define and preserve
the methods uniqueness (regardless of the value and contribution of the
work otherwise).

I presented an example at last years conference, talking about landfill
dynamics. The desired behavior is simply quicker stabilization of the
landfilled waste (no particular behavior pattern). Good empirical data was
available on time phased emissions of gases from the landfill that could
potentially be correlated to certain kinds of microbial activity. But
patterns were complex, and many kinds of processes were known to be
overlapping to produce these patterns. The model was formulated by coding
sequential microbial population growth with limiting food sources, organic
byproducts, and inhibition feedback, trying to scope the effort down to
general major categories of biodegradation processes (fermentation,
acetogenesis, methanogenesis, etc). The complex model was able to match
empirical data patterns, but I felt that, in order to bring the power of
system dynamics to bear, I had to somehow discern a higher order structure
in the flow diagram, formulate an initial influence diagram of higher level,
perceive associated behavior patterns in the structure, and iteratively
derive a simpler system boundary ultimately based on a "bootstrapped"
reference behavior and influence structure. Initial attempts to do so
failed.

Although there were elements of the SD method in the above effort, it
clearly did not meet the SD uniqueness criteria that I espouse and teach to
my students without employing some modified backwards iterative approach to
get to the "right" system boundary. Such a method is not clearly delineated
in the literature to my knowledge. Dynamic environmental modeling texts do
not present such a method and the environmental modeling cases presented
compromise on the classic SD "unique" method (if I may respectfully suggest)
and present something less than what I would call SD. Yet, environmental
systems intuitively ought to be of a nature that is well suited for SD with
the elegant feedback stabilization mechanisms inherent to these systems. I
reject the notion that these systems must simply be approached differently.
They are simply more challenging, and I am frustrated with the lack of a
method to facilitate employing classic SD to these more complex systems to
elucidate the high level system boundary appropriate to the objective.
Until that method exists and is employed, these complex modeling efforts
should not be labeled as SD. It dilutes the definition of the method and
hurts the SD community. Is anyone else frustrated on this? I have enjoyed
some dialog with Andy Ford on this issue and would like to hear from him and
others in this forum.

Humbly,
Mike Shelley
Air Force Institute of Technology
Michael.Shelley@afit.edu <mailto:Michael.Shelley@afit.edu>
Shelley Michael L Civ AFIT/ENV <
Junior Member
Posts: 2
Joined: Fri Mar 29, 2002 3:39 am

SD uniqueness and the challenge presented by complex systems

Post by Shelley Michael L Civ AFIT/ENV < »

In my suggestion that SD is unique in its ability to formulate the correct
system boundary for addressing the problem at hand, Bill Harris inquires:

"Whats the difference between that and, for example, a Colored Petri Net
model I might make? I still determine the model boundary to suit my goals."

I do not mean to imply that other analysis methods have no sense of a system
boundary appropriate to the goals of the analysis. My point is that, for
problems that are concisely expressed by a characteristic continuous
behavior pattern over time (a problem amenable to SD analysis), the SD
practitioner formulates the system boundary very concisely and perfectly
matched to the problem (the behavior pattern). A Petri Net (which I
understand to be a network of nodes having "states" that are stochastically
defined that influence the states of other nodes through Boolean logic) can,
I presume, be aggregated into simpler form for a particular analysis when
such aggregation is perceived to have no effect on the analysis. But Im
not sure how that judgment is made with high confidence. I am also unclear
that Petri Nets well represent continuous dynamic systems in time and that
well established methodological rules are established that dictate Petri Net
structure corresponding to particular behavior patterns. If this is true,
then it is simply another SD simulation tool that is used within an SD
analysis displaying that unique character of SD. Otherwise I think we may
be comparing apples and oranges in terms of the class of problems being
discussed.

In response to my suggestion that some scientists and engineers have
difficulty accepting a simple model that does not contain all known
mechanisms, Bill also inquires:

"Is that a function of the type of system, as you suggest, or its purpose or
the required accuracy? Ive heard people here speak of very complex SD
models to calculate results accurate to 0.1% or better. I was an engineer
in the past, and Ive modeled systems/circuits there with exceedingly simple
(non-SD) models and found them useful in achieving my purpose."

Again, I recognize that good analysts can simplify their system
representations in serving a clear goal. But these adjustments of system
boundary are generally derived from the bottom up, that is, from a
perspective of the detailed system and how certain components can be rolled
up and represented in less detail. The SD practitioner presents a model
structure derived from the top down, that is, from a metric behavioral
pattern without knowledge of the details of the system. This representation
is often foreign to the non-SD client, not intuitively seen as a proper
system boundary. (I would also suggest that "complex SD models...accurate
to 0.1%" are actually very detailed numerical simulations of ordinary
differential equations in time that are used outside the bounds of what I
would regard as a system dynamics method. Such predictive accuracy is
generally not the purpose of an SD analysis.)

Mike Shelley
Air Force Institute of Technology
Michael.Shelley@afit.edu
Locked