Using Complex Systems Theory to Understand Emergent Behavior

Walter Warwick
Michael Matessa
Alion Science and Technology
MA&D Operation
4949 Pearl E. Circle, Suite 200
Boulder, CO 80301
303-442-6947
wwarwick@alionscience.com, mmatessa@alionscience.com

Keywords:
cascade, network theory, cognitive modeling


1. Introduction

Assessing the goodness-of-fit between a computational
model’s predictions and the phenomenon it purports to
represent is hardly ever straightforward. This is
especially true when the phenomenon being modeled is
the emergent behavior of a complex socio-
technological system. In this poster, we will describe
our efforts to model the flow of chat messages through
a command and control system. Our efforts are
anchored empirically by log files that were taken during
a three-day exercise that simulated a sustainment and
stability operation that involved some five dozen
different operators and officers. Although those log
files do yield some data (e.g., message counts) that
could be used to quantify some degree of fit between
the performance of the model and the actual command
and control structure, our initial attempts to fit the
model to such data suggest that chat messages tend to
cascade through the command structure and
summarizing such behavior in terms of means or
aggregate rates seems inappropriate. For this reason,
we have retreated from assessing the goodness-of-fit in
the traditional manner (e.g., as suggested by Schunn
and Wallach, 2001) and have instead turned to more
formal accounts of cascading behavior (e.g., Watts,
2002) to see if the model behavior can be understood in
terms of more general features that characterize
complex systems.

This poster will describe our initial attempts to model
the chat behavior with the command and control system
and a second attempt to build a more abstract model
that might be aligned with the formal characterization
of cascading behavior.

2. C3TRACE Simulation

Using an augmented version of C3TRACE (Warwick et
al., 2008), a task network modeling tool developed for
the US Army Research Laboratory to study the impact
of technological and organizational changes on
communication flow within complex command and
control (C2) structures, we represent the personnel
present during a Battalion-level exercise. The model
focuses on chat as the mode of electronic
communication (though placeholders for representing
email have been built into the model). The model
explicitly represents individual personnel grouped into
chat rooms. Activity in each chat room is stimulated by
the periodic arrival of a ―seed‖ message. The single
recipient of that seed message then decides whether to
respond to everyone else in the chat room; if that initial
decision is to respond, then every other person in the
chat room will receive a chat message and will, in turn,
decide whether to respond to that message, sending a
new message to everyone in the chat room. The model
produces, as output, a log file of time-stamped chat
messages indicating sender, receiver and message
content by type.

Our initial attempt to assess the goodness-of-fit
between the model and the empirical data focused on
the comparison of the arrival rates over time for
messages within each chat room. In many cases in the
empirical data there were clear and sudden spikes in
chat activity that would then quickly diminish.
Individual runs of the computational model were able
to capture this effect qualitatively, but when looking
over multiple runs, it was clear that the precise timing
of these spikes was quite variable; some runs matched
the empirical data very well and others didn’t. Further,
―averaging‖ across runs simply washed out the spikes.

Of course, at this point, it might seem easy to conclude
that the mode just wasn’t reproducing the empirical
behavior. At the same time, however, given the
unpredictable nature of the computational model in
producing qualitatively similar cascades in chat
behavior, we were unsure whether we might see exactly
the same variability in the human behavior were the
live experiment to be repeated. That is, we were no
more confident in the robustness of the specifics of the

human behavior observed in a single trail of the
exercise than we were in the results of a single
(carefully chosen) run of the model. So we turned to
network theories of emergent phenomena to give us a
framework to address qualitative results.

3. Simplified Simulation

Watts (2002) presents a possible explanation of
cascading in terms of a sparse, random network of ten
thousand interacting agents whose decisions are
determined by the actions of their neighbors according
to a simple threshold rule. Watts describes three
general features that shape the dynamics of cascades—
local dependencies, fractional thresholds, and
heterogeneity. Unfortunately, none of these features
were faithfully represented in the C3TRACE model;
rather than condition the decision to reply to portion of
neighbors responding, the decision to respond to a chat
message was determined independently by each
modeled operator according to a probability of reply
parameter.

In order to align our modeling efforts more clearly
with the formal account of cascades we developed a
simplified simulation in Lisp. The network consists of
ten nodes representing people in a chat room, and is
fully connected, representing common read and write
access.

In a manner similar to that in Watts (2002), actions in
the model consist of a binary decision (replying to a
chat). Also, heterogeneous threshold values are used so
that decisions are triggered when the percentage of
acting neighbors are greater than the agent’s threshold.
Unlike the situation in Watts (2002), decisions are
probabilistic in that noise is dynamically added to
threshold values, and the number of neighbors is held
constant as the total number of actors.

In the simulation results presented here, threshold
values vary between 0.0 and 0.4 and noise varies
between 0.0 and 1.9. On average, in only one in ten
runs does the decision to reply cascade to all ten agents.
Note that the noise parameter can sometimes be greater
than one. These values can be thought of as a
representing a general inhibition to respond. Threshold
values less than 0.4 can be thought of as representing
the need of seeing less than half of other agents
responding in order to decide to respond. In a cascade
situation, the increasing number of replying agents
increases the probability of other agents replying.

With regards to the C3TRACE model, we plan to use
the Human Behavior Architecture (HBA -- Warwick et
al., 2008) to represent more detailed agent cognition.
This cognition will need to represent the general
inhibition and low threshold of observed replying found
in the simplified simulation.

4. Conclusions

When trying to compare our initial modeling results to
human behavior we found that both were a result of
emergent phenomena. We are currently using the
complex systems literature (e.g., Watts, 2002) to find a
framework for understanding this emergent behavior.
Indeed, without traditional approaches to fit to support
theory confirmation, we see no other alternative to
account for the behavior we see in the human and
model data. Useful concepts found with this
exploration include inhibition of response and a low
threshold for triggering behavior based on observed
behavior. These concepts will be used in developing
more cognitively plausible agents and, at the same time,
should provide a principled account for characterizing
the emergent behaviors those agent models produce.

5. References

Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992).
A theory of fads, fashion, custom, and cultural
change as informational cascades. Journal of
Political Economy, 100(5), 992–1026.
Schunn, C. D. & Wallach, D. (2001) Evaluating
goodness-of-fit in comparisons of models to data.
Online manuscript.
http://lrdc.pitt.edu/schunn/gof/index.html
Watts, D. J. (2002). A simple model of global cascades
on random networks. Proceedings of the National
Academy of Sciences of the United States of
America, 99(9), 5766-5771.
Warwick, W., Archer, R., Hamilton, A., Matessa, M.,
Santamaria, A., Chong, R., Allender, L., &
Kelley, T. (2008). Integrating Architectures:
Dovetailing Task Network and Cognitive Models.
Proceedings of the 17th Conference on Behavior
Representation in Modeling and Simulation.
SISO.

Author Biographies

WALTER WARWICK is a Principal Systems Analyst
at Alion Science and Technology. He has worked on
several projects having to do with the modeling and
simulation of human behavior.

MICHAEL MATESSA is a Lead Cognitive Scientist
at Alion Science and Technology. He has over a
decade of experience with cognitive architectures and
models of communication.