Situated Socio-Cognitive and Ecological Simulation of the Triple-Helix Theory
of Regional Innovation Dynamics
by
James R. Morris-King
A dissertation submitted to the Graduate Faculty of
Auburn University
in partial ful llment of the
requirements for the Degree of
Doctor of Philosophy
Auburn, Alabama
May 3, 2014
Keywords: triple-helix, agent, simulation
Copyright 2014 by James R. Morris-King
Approved by
Levant Yilmaz, Chair, Associate Professor of Computer Science
John A. Hamilton Jr., Professor of Computer Science
David Umphress, Associate Professor of Computer Science
Abstract
This dissertation focuses on applying the concept of agent-based ecological simulation
to the problem of developing and maintaining regional innovation economies. A simulation
model based on a popular model of innovation, the Triple-Helix theory of public-private
partnership, is presented. This theory guides the alignment of the three principle sectors of
regional and national systems of innovation; academic research, industrial development, and
government control. The development and maintenance of systems of innovation has been
recognized to be a key driver of economic growth, but the underlying mechanisms which
govern these systems have proven di cult for pure statistical analysis to capture.
We introduce an agent-based modeling and simulation approach, which leverages eco-
logical concepts better suited for describing the behavioral mechanisms of multi-agent,
environmentally-situated complex adaptive systems. This model extends the basic formu-
lation of the triple-helix to include a new environmentally situated actor, the researcher,
which uses indirect signaling to discover and exploit grant-funding. This extension uses
principles from classical Predator-Prey interaction and a variant of ant colony optimization,
a non-deterministic search algorithm.
The work presents a didactic framework for understanding the role government policy
has on innovation-driven productivity and the con guration of resources around centers
of innovation. We  nd that innovation is principally driven by concentrations of mature
academic research institutions and is mediated by consistent government support and highly
active industrial partners. We also present a uni ed innovation resilience metric for future
cross-model comparison.
ii
Acknowledgments
For my parents; Donald King and Vivien Morris, who taught me to value my mind and
my self.. For my  anc e; India Napier, who stood by my side through all the upheavals of
graduate life and managed to stay (relatively) sane despite  elding  ve years of my incoherent
ramblings. For my friends; we?re still digging in the sand, dreaming about robots. I think
we did something good here.
iii
Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
1 Innovation Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Scope of the Thesis & Motivations . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Modeling the Innovation Economy . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Goal & Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Achievements & Contributions . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Outline of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Literature Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1 Knowledge and Innovation in Economic Systems . . . . . . . . . . . . . . . . 12
2.2 The Knowledge Economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 The Triple-Helix of Public-Private Partnership for Innovation . . . . . . . . 15
2.3.1 Government Intervention . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.2 Appropriability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.3 Government Grant System . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Modeling Innovation Economies . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Top-down modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.2 Bottom-up modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.3 Agent-based modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.4 Ecological modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 Verifying & validating simulation models . . . . . . . . . . . . . . . . . . . . 33
iv
2.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Agent-based Perspective for Simulating Innovation Dynamics . . . . . . . . . . 38
3.1 An Ecological Approach to ABM . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 The Agent-based Ecosystem Model . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.1 Resource Management Models . . . . . . . . . . . . . . . . . . . . . . 42
3.2.2 Resource Management Instruments . . . . . . . . . . . . . . . . . . . 44
3.3 Previous Work in Agent-based Innovation Simulation . . . . . . . . . . . . . 47
3.3.1 Behavioral models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.2 Explicit NSI Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4 THESIM: Triple-Helix Ecosystem Simulation . . . . . . . . . . . . . . . . . . . 58
4.1 Triple Helix Ecosystem Interaction Model . . . . . . . . . . . . . . . . . . . 58
4.2 Conceptual Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3 Behavioral Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3.1 Agent Behavior Rules . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4 Overview and Design Concepts of THESIM . . . . . . . . . . . . . . . . . . 70
4.4.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4.2 State Variables and Scale . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4.3 Process Overview and Scheduling . . . . . . . . . . . . . . . . . . . . 71
4.4.4 Design Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4.5 Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.4.6 Submodels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5 Model Implementation & Basic Relationships . . . . . . . . . . . . . . . . . . . 83
5.1 THESIM-Repast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2 Class Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3 Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
v
5.3.1 Activity Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3.3 Preliminary Experiments . . . . . . . . . . . . . . . . . . . . . . . . 88
6 THESIM Experimental Design & Analysis . . . . . . . . . . . . . . . . . . . . . 96
6.1 Introduction & Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.2 Systematic Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.3 Sub-DOE THESIM-Sensitivity 1st-iteration . . . . . . . . . . . . . . . . . . 105
6.3.1 Main E ect Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.3.2 Sub-DOE THESIM-Sensitivity 1st-iteration Summary . . . . . . . . . 115
6.4 Sub-DOE THESIM-Sensitivity 2nd iteration . . . . . . . . . . . . . . . . . . 117
6.4.1 Response Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.4.2 Sub-DOE THESIM-Sensitivity 2nd-iteration Summary . . . . . . . . 123
6.5 Sub-DOE THESIM 3rd-iteration . . . . . . . . . . . . . . . . . . . . . . . . 125
6.5.1 Sub-DOE THESIM-Sensitivity 3rd-iteration Summary . . . . . . . . 127
6.6 DOE Analysis & Observations . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.7 THESIM-ER Ecological Resilience Experiment . . . . . . . . . . . . . . . . . 129
6.7.1 THESIM-ER Analysis & Summary . . . . . . . . . . . . . . . . . . . 132
6.8 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7 Conclusion & Recommendations for Future Work . . . . . . . . . . . . . . . . . 138
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
vi
List of Figures
2.1 Three functional spheres of patenting . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Representation of the Triple Helix Model . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Architecture of a grant system. . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 An agent interacting with an environment. . . . . . . . . . . . . . . . . . . . . . 27
2.5 Simple version of the modeling process (Sargent, 2005) . . . . . . . . . . . . . . 34
3.1 Structure diagram of the standard Nelson{Winter model of industrial dynamics.
The diagram puts an emphasis on a particular  rm, while aggregates are placed in
the last row. An arrow from x to y should be read ?x codetermines y? (Andersen,
2001). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.1 Model of a simple NSI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Triple-helix ecosystem interaction model. . . . . . . . . . . . . . . . . . . . . . 59
4.3 Conceptual model of the ?innovation food web? . . . . . . . . . . . . . . . . . . 62
4.4 Basic resource cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.5 Relational Map of agents in the Triple Helix ecology. . . . . . . . . . . . . . . . 64
4.6 The legend of environment (a), agent?s movement directions (b), coding sequence
for local perception (c), example of local perception (d) . . . . . . . . . . . . . . 67
vii
4.7 University-Industry-Government relations in a three-dimensional space. . . . . . 76
4.8 Von Neumann Neighborhood for varying values of r. . . . . . . . . . . . . . . . 79
5.1 UML model of THESIM class relationships . . . . . . . . . . . . . . . . . . . . 84
5.2 University?s local grant neighborhood t = 0 and t = 150 respectively. . . . . . . 86
5.3 THESIM Repast sample output showing grant clustering and exploitation regions
in a four university system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.4 System productivity via gross domestic product over time (t) . . . . . . . . . . 90
5.5 Behavior and value of technology acquisition among industry agents over time (t) 91
5.6 Researcher agent population growth over time (t) . . . . . . . . . . . . . . . . . 92
5.7 Shannon-Wiener index of researcher agent population (HR?) over time (t) . . . 93
5.8 The Success Rate (Sr) and Average Con dence of industry-university partnership
(Ca) under conditions of risk-aversion (Er) over time (t). . . . . . . . . . . . . 94
6.1 Goals: Relationship Mapping of xi to Ri . . . . . . . . . . . . . . . . . . . . . . 102
6.2 Parameters? main e ects on System GDP (R1). . . . . . . . . . . . . . . . . . . 109
6.3 Parameters? main e ects on researcher population count (R2). . . . . . . . . . . 110
6.4 Parameters? main e ects on researcher population diversity (R3). . . . . . . . . 111
6.5 Parameters? main e ects on average university size (R4). . . . . . . . . . . . . . 112
6.6 Parameters? main e ects on university-industry trust (R5). . . . . . . . . . . . . 113
6.7 Parameters? main e ects on average industry GDP (R6). . . . . . . . . . . . . . 113
viii
6.8 Parameters? main e ects on industry partnership success rate (R7). . . . . . . . 114
6.9 Parameters? main e ects on technology transfer count (R8). . . . . . . . . . . . 115
6.10 A hypothetical well-being function for an idealized system. . . . . . . . . . . . . 130
6.11 Linear- t of Y under increasing system stress. . . . . . . . . . . . . . . . . . . . 133
6.12 Y with respect to Df. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.13 Y with respect to Dl. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.14 Overview of disturbance characteristics across various combinations of Df and Dl. 135
7.1 Sample output of THESIMDX?s force-directed graph running in a web browser. 141
2 Interaction E ects on R1 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 163
3 Interaction E ects on R2 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 164
4 Interaction E ects on R3 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 165
5 Interaction E ects on R4 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 166
6 Interaction E ects on R5 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 167
7 Interaction E ects on R6 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 168
8 Interaction E ects on R7 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 169
9 Interaction E ects on R8 for THESIM Sub-DOE  rst-iteration. . . . . . . . . . 170
10 Full data table for THESIM-ER Including all values for DRTR and the Linear
Model Yle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
11 Sample THESIMDX graphml output from a single-simulation trial. Note the
node and edge schema as well as the govenrment node representation (FA1) . . 172
ix
Chapter 1
Innovation Dynamics
1.1 Introduction
Innovation economics is a growing economic doctrine which places knowledge, technol-
ogy, innovation, and entrepreneurship at the center of economic theories of policy planning.
This is a departure from the classic view of these forces, which is often described as exter-
nal and independent to the formulation of policies and thus unable to be a ected directly
through their application. Innovation economics rests soundly on two basic tenants: that the
core goal of economic policy should be to increase productivity through increased innovation,
and that markets relying on resource and price signaling will not always be e ective in in-
creasing productivity, and thereby economic growth. Porter (1990) observed that a nation?s
primary economic goal is to produce a \high and rising standard of living" for its citizens.
Predictably, this ability depends on the productivity with which a nation?s resources can be
exploited and deployed. To this end, a great deal of energy has been devoted to the study
of public policies for sustained economic growth.
One well-established engine of productivity is the notion of technological change. Edquist
(1997) argues that it is almost universally accepted that technological change, in tandem with
other kinds of innovations, are the most important sources of productivity growth and in-
creased material welfare and that this has been the case for centuries. Such change has
provided  rms with the power to circumvent scarce factors and market disruptions via new
products, processes, and services. E ciency, the reduction of materials, energy and other
resource-based inputs, is often a natural byproduct of this change.
The discovery or production of modern materials, inventions, and processes all arise
from innovation activities. Likewise, the steady monetization of the product of this creative
1
re nement process results in improved productivity and pro t for the country or region.
For example, automation and process innovations resulted in a reduction of the labor costs
within many industries, implying that access to high technology is more important than low
local wages (Porter, 1990). This realization has led many to conclude that it is in a country
or region?s best interest to promote this parallel system of innovation in order to promote
competitiveness.
1.2 Scope of the Thesis & Motivations
The exploration of alternative strategies for the establishment and long term stability of
innovation economies and the testing of di erent ideas and theories before implementation in
the real world can result in major economic bene ts. The implementation of poorly designed
policies or the long-term mismanagement of a region?s innovation infrastructure can often
lead to social and economic deprivation through the loss of local or global competiveness.
Unwinding bad policies or restructuring the built knowledge infrastructure of a region can
be a costly endeavor. Thus, it is important to have tools through which one can evaluate
the viability of policies which can a ect a region?s system of innovation.
In this dissertation, we make use of agent-based ecological simulation modeling to study
cooperative R&D among public and private  rms, represented as university and industry
collectives. These collective  rms are modeled as agents with heterogeneous behavior when
cooperating. To de ne  rms? behavioral micro-foundations (i.e. how  rms decide whether
to cooperate or not, which kind of partner they would prefer to choose, and the level of trust
present in that partnership), we follow the interaction rules laid down in the triple-helix
model of public-private partnership (Leydesdor , 2000). The agents in this model
are environmentally situated in a two-dimensional physical environment where monetary
resources are distributed and managed by a centralized government agent, and the behavior
of  rms can be observed by the growth and dissolution of clusters. Additionally, we propose
a new agent class, the ?researcher?, which is monitored and controlled by research  rms and
2
serves the dual role of contributing to R&D e orts and discovering environmentally situated
resources via swarm mechanics (Ant-colony optimization). De ning the determinants of
cooperation to be used between the agents in our simulation system required e orts on
collecting appropriate theoretical background as well as analyzing past studies regarding the
formation and maintenance of innovation economies.
From the aforementioned de nitions of each agent?s behaviors we have implemented
the model as a two-dimensional landscape where both resources and  rms \attract" each
other based on their individual characteristics. This physical landscape is then overlaid with
several ?virtual? (social) layers representing each additional sphere of the triple-helix. These
layers are represented abstractly as asymmetric networks, where  rms are represented as
nodes connected by their economic relationship with one another. The notion of \attrac-
tion" is measured in both physical proximity, network centrality, and a trust metric, where
high values eventually result in strong collaboration between  rms during R&D partnership
(Dongsheng & Yongan, 2008).
Thus described, the motivation for this thesis is twofold. On the one hand, we are
interested in answering the basic scienti c question of how the interaction of  rms shapes the
environment within an innovation ecosystem. Previous research focused on the development
of computational models which explore individual mechanisms of such a change, but only
in models restricted to a single dimension of interactivity (physical or social). Studies have
focused strictly on the behavior of individuals in research collaboration (Jordan et al., 2005),
the clustering of  rms (Butel & Watkins, 2006), the interaction of industrial entrepreneurs
(Feldman et al., 2010), or local search for optimal resource siting (Mutiah et al., 1996).
However, few computational models have been put forward that capture all of these dynamics
simultaneously, and none which investigate their interaction. As a result, the activities of
the three helices of innovation production have not been examined in a holistic fashion.
On the other hand we are developing a practical tool, which allows us to test the
triple-helix theory and provide insight into the e ects of various policies and conditions on
3
the formation and production of stable innovation economies. Its use to simulate these
systems|even before the actual creation of policies or  rms|makes it possible to determine
where a relevant regional management & planning agency may face di culties, why they
face them, and how policies or system features may have to be considered more fully, or
changed, to avoid such di culties. Simulation of human-like behavior in this space is a
powerful research method to advance our understanding of the interaction between people,
 rms, and their environment. It allows for both the examination and testing of models
and their underlying theory as well as the observation of the system?s macroscopic behavior
(Gimblett et al. 1997).
The validation of this approach is done by performing a basic model reduction activ-
ity focused on limiting the parameter space to those which are considered germane to the
development of triple-helix structures. This reduction process is followed by a sensitivity
analysis step. Sensitivity analysis begins with a level reduction on the values of the remain-
ing parameters, whereby three values are selected which represent the upper, lower, and
median conditions. Experiment design theory is then used to reduce the parameter combi-
nations to a reasonably (e.g. computable) orthogonal subset, which allows us to explore the
search space in a directed fashion. Finally, the results of sensitivity experiments are used to
identify the most signi cant parameters. Signi cance is determined by the relationship to
known empirical works. We are particularly interested in identifying empirical regularities
(power laws) concerning the productivity of  rms and their relative technological maturity,
the growth of trade relationships between research and commercial  rms, and the impact of
 scal allocation policies on system growth, diversity, and stability. We compare these results
to those distinguished by industry studies in order to validate the accuracy of the model
against real-world problems. We believe that the equivalent quantitative outcomes help val-
idate the accuracy of the simulation model while allowing us to test more proscriptive policy
strategies.
4
1.3 Modeling the Innovation Economy
The underlying networks present in an innovation economy have been approached and
emphasized in many studies, especially in public policy and evolutionary economics literature
(Forfas, 2004; Yang et al., 2009; Gollmitzer & Murray, 2008). However, few works have been
done, toward modeling the processes by which these systems are formed and estimation of
the outcomes and e ects on innovative products, processes, and  rm dynamics as they relate
to exogenous parameters and cooperative strategy. The complexity of the network dynamics
involved and the heterogeneity of the actors make it hard to model this problem using
traditional techniques (Metcalfe, 1995). Moreover, modeling techniques such as agent-based
cooperative economics often aggregate or ignore cooperation behaviors present in modern
knowledge economy con gurations (Zhu et al., 1999).
A variety of deterministic (Haynes et al., 1976; Teece, 1980; Giovanis & Skiadas, 1999;
Qin & Ljung, 2003; Yildiz et al., 2011) and non-deterministic (Downs & Mohr, 1976; Markus
& Robey, 1988; Holland & Miller, 1993, van Laarhoven & Aarts, 1987; Goldberg, 1998,
Kau man et al., 2000) approaches have previously been applied to modeling systems of
innovation. Deterministic models guarantee the same solution at di erent runs with the
same parameter values, and the non-deterministic models such as genetic algorithms and
simulated annealing generate di erent solutions due to their randomness (see Chapter 2 for
details of previously applied approaches). Butel & Watkins (2006) reveal that ant colony
search (ACS) can be applied to problems where actors operate in conditions of dynamic
uncertainty, particularly in the business environment, where entrepreneurs must identify
and exploit opportunities, and form cooperative clusters.
Agent-based modeling (ABM), which includes  ocking models and ant colony optimiza-
tion (ACO), uses heuristic algorithms to determine solution steps. Heuristic algorithms are
especially useful when  nding optimal solutions within a limited time is likely to be impossi-
ble. These algorithms  nd results of \acceptable quality" in a short period of time (Maniezzo
and Carbonaro, 2001). ACO is a novel optimization method which has been used to solve
5
various real-world problems such as the travelling salesman problem (Dorigo et al, 1996;
Corloni et al., 1991; Stutzle & Hoos, 1997), the quadratic assignment problem (Maniezzo et
al. 1994), the Job Shop Scheduling Problem (Dorigo et al, 1996), telecommunication routing
and load balancing (Sim & Sun, 2003), and others. ACO has been shown to perform well
compared to other non-deterministic algorithms such as genetic algorithms and simulated
annealing (Dorigo & Stutzle, 2004). Despite this, there have been few applications of this
class of algorithm applied to the  eld of philosophy of science, particularly problems related
to information di usion (Pirolli, 2009) and the development of regional innovation systems
(Zhi & Gang, 2006).
Furthermore, little attention has been paid to the granular activity of researchers within
the R&D process. Studies on innovation di usion tend to aggregate individuals in to  rms
or enterprises, and often ignore the role indirect funding plays in the establishment of new
ideas. Government intervention is given equally short shrift in simulation literature, which
often focuses on the self-organizing nature of  rms in an economy, despite the fact that
intervention strategies are well-accepted in economics literature to be a major driving force
in building and maintaining regional innovation systems.
Thus, we believe that injecting these two notions is not only important in order to create
a holistic simulation of innovation economies, but also a ords the application of ecological
modeling to the conceptualization and design of the system. This ecological, or ecosystem,
view of the system provides a more accurate and holistic architectural framework than an-
alytical and statistical models. Finally, the relation between actors occupying the various
(trophic) roles in the ecosystem model provide new analogies for analyzing group behavior,
such as foraging, reproduction, and predation.
6
1.4 Goal & Hypothesis
The goal of this thesis is to develop a computational theory of the triple-helix theory of
public-private partnership, which explains how government, entrepreneurial  rms, and aca-
demic research organizations collaborate both directly and indirectly to build and maintain
knowledge economies. The theory uses an agent-based approach and focuses on knowledge in
the world in the form of a ordances and information, and their utilization by representative
agents during the process of innovation creation and sharing. Firms are involved in vari-
ous activities during this process; therefore, the agent model needs to include di erent roles
whose interplay allow for simulating these activities. For this purpose we turn to ecologi-
cal modeling, a conceptual technique used to describe the interactions of complex systems
with multiple classes of interdependent actors. Hence, we aim to answer the following two
questions from the ecological view of system modeling:
1. What is the minimum set of roles necessary to capture the interactions present in the
triple-helix model of the innovation economy?
2. What are the critical parameters|a ordances and information|necessary to ensure
system stability and diversity?
We do not focus on aspects of individual learning and lasting cognitive-map-like rep-
resentations of the environment, instead we turn to decentralized strategies of information
transmission in order to address issues of complexity. Our central hypothesis is:
The development and maintenance of innovation economies can be explained on
the basis of an ecologically driven agent-based model of the triple-helix model of
public-private partnership. This model simulates the minimum set of interactions
between government, academia, and private industry in a fashion more consis-
tent with their behavioral roles than other traditional agent-based and statistical
modeling techniques.
7
The main hypothesis can be further detailed by the following two sub-hypotheses:
1. Agents need a minimum set of interacting components|knowledge about the world|
to meet speci c role-de ned goals in an unfamiliar environment. These components
are its observation schema, agent?s state, exploration & exploitation strategies, and
rules for interacting with other agents.
2. Knowledge in the world can be represented by ecological a ordances and environmental
information. The process of innovation creation and transport works on the basis of
the interplay between the two.
1.5 Achievements & Contributions
The signi cance of this research lies in the potential of the developed simulation model,
which leverages spatial environments, relational networks, and swarm-based optimization to
compose a tool useful for analyzing the e ect of resource allocation and intervention policies
on the growth of regional knowledge economies arranged in the triple-helix con guration.
Further, despite the recent advancements in the  elds of innovation systems and agent-based
simulation, the following issues have not been directly addressed by other researchers (See
chapters 3 for details of previously applied approaches).
 There has been relatively little agent-based examination of the so-called ?triple helix?
of public-private partnership.
 Few of the existing simulation models of regional innovation have taken into account
an ecological framework for designing agent interactions.
 Existing approaches have addressed neither the activity of individual researchers in the
typical RSI, nor the notion of resource foraging among researchers and  rms.
 There have been few examples of swarm dynamics applied to address complexity issues
inherent to the problem space.
8
 Most of the simulation models are restricted to innovation di usion or economic growth.
 Innovation products are almost often treated as homogeneous units, rather than as
artifacts with speci c attributes.
 Moreover, few of the existing approaches take into account collective social pressures
(institutional trust, peer pressure, explorative dogma, etc.) which steer innovation,
speci cally the creation of funding regimes which favor one or more  eld of innovation
to the detriment of others.
In this thesis, the above issues are addressed. In addition to these issues, the thesis suggests
multiple modi cations to ant colony optimization and a more formal adherence to ecological
modeling principles in the development of future simulations.
First Contribution
Initially, an ecological model of the triple-helix theory is developed taking in to account
the presence of individual researchers as well as the interaction rules between multiple actors.
From this, a general agent-based conceptual model is formalized following the ecological
framework of producer-consumer relationships. An environmental model is also developed
as an analogue of the ?funding space? provided by the regional government. A modi ed
version of ant-colony search is developed as a component of the agent behavior in order to
provide researcher agents the capability to explore and exploit this theoretic landscape, and
di use knowledge about its contents and topography to other researchers.
While developing the ant-colony algorithm for the environmentally situated research
agents, three modi cations have been proposed to better  t the problem domain.
 The  rst modi cation is the segregation of communication between swarm populations
belonging to each institution. This enables the creation of several pseudo-layers of
communication and encourages competition.
9
 The second modi cation reduces the size of the search space for each agent to a radius
around its originating institution and promoting the convergence of each swarm to
local optimum solutions.
 In order to combat the potential stagnation problem brought on by the previous two
modi cations, the resource pool in the environment may occasionally experience per-
turbation. This e ect is managed by the governor agent.
Second Contribution
In addition to the swarm-based search, a simple marketplace model is introduced which
supports policy-based intervention on the part of the government. The relationship rule-
set models the trust between university and industry  rms. This model takes into account
past cooperation in the negotiation of prices and costs for innovation di usion. In addition,
a single government agent monitors the market and engages various strategies to promote
productivity, while taxing the pro ts of enterprises in order to fund basic research. Within
the market, each  rm attempts to  nd the price that maximizes its pro t through bartering,
the government is allowed to intervene in the form of variable subsidies that can ease pressure
on both parties and increase the amount of long-term cooperation in the system.
This approach has a long history in real-world scenarios. Governments often subsidize
research partnerships between  rms in order to promote cooperation or ease market pressures
that might hinder new development of promising technologies. Additionally, taxation and
indirect research funding act as levers through which a government can enact long-term
policies. The introduction of these factors into the broader model of the innovation economy
results in an ecological feedback loop which more closely models real systems.
1.6 Outline of the Dissertation
This dissertation is organized as follows: in chapter 2 we provide the theory and concepts
of cooperation in innovation, including background on knowledge economies and a brief
10
overview of the triple-helix model and the government grant system. Chapter 3 presents
the justi cation for an ecological approach to agent-based modeling introduced in chapter 1
along with supporting empirical analysis drawn from literature, and serves as the foundation
of our design model. In chapter 4 we present our model design, following a formal template
to ensure reproducibility. We also provide an initial view of the various behavior parameters
and rule-sets that govern agent roles and and activities.. In chapter 5 we present preliminary
observations and  ndings used to justify the feasibility of our modeling approach. In chapter
6 we present a formal experimental design and analysis of our model?s behavior with respect
to quantifying error and answering questions about its functionality with respect to our
initial assumptions. Finally, in chapter 7 we discuss our overall  ndings as well as future
extensions to our work.
11
Chapter 2
Literature Survey
2.1 Knowledge and Innovation in Economic Systems
In engineering and economic literature the notion of \innovation" comprises the creation
or signi cant technological changes of new or novel products and processes. The early work
of Schumpeter seized on the idea that innovative activity was a critical driver of economic
growth (Schumpeter, 1942). The act of developing innovation is regarded as a creative \de-
struction process" because it caused constant disruption to economic systems in equilibrium.
In adjusting to equilibrium, other innovations are conceptualized and spun-o , leading to yet
more innovation. Through this self-reinforcing process, Schumpeter?s theory predicts that an
increase in the number of innovations leads to an increase in economic growth. This theory
also implies that an increase in the number of centers of innovation will have a similar e ect
on an economy.
In modern economic parlance, it is accepted that innovation is not strictly generated
in the boundaries of a  rm or an organization, cooperation between multiple entities with
varying degrees of specialty in the activity of generating innovation is an accepted and often
desirable strategy for success. Thus,  rms are not expected to develop all the relevant tech-
nologies without accessing external knowledge sources or entering cooperative relationships
with one another. Patents are an example of a \boundary product" between  rms (Leydes-
dor and Meyer, 2006); while they are largely recognized due to the fact that they o er legal
protection, they can also be used as indicators of knowledge production and/or economic
value (Fig. 2.1).
As products and processes have become more complex, there has been a steadily in-
creasingly requirement for technological, organizational and marketing search involving many
12
Figure 2.1: Three functional spheres of patenting
players such as suppliers, universities, research institutes, nonpro t organizations and so on.
Cooperative innovation is widely considered an e cient means of industrial organization of
complex R&D processes (Dachs et al., 2008). The sources of valuable knowledge for inno-
vation may be found anywhere on a  rm?s production chain and accessing them may be
crucial to maintain competitiveness. The positive impact of cooperation in innovation is
strongly supported by the extensive literature (Cassiman & Veugelers, 2002; Doo & Sohn,
2008; Miotti & Sachwald, 2003).
While a growing body of research literature supports the notion that cooperation be-
tween  rms is crucial for innovation and economic development, many studies have identi ed
barriers, which may impede this behavior (Forfas, 2004). Indeed, It is incorrect to assume
that this bene cial view of the cooperative behavior is enough for knowledge economies to
emerge spontaneously. In order to promote formation and growth of such inter- rm coop-
eration, governments must be aware of the barriers and play a major role in fostering and
creating positive conditions for successful partnership to emerge.
13
2.2 The Knowledge Economy
A knowledge economy can be de ned as the use of knowledge technologies (innovation)
to produce economic bene t (Druckter, 1969). While many models for successful knowledge
economies have been proposed, perhaps the most in uential was the National System of
Innovation (NSI) proposed by Freidrich List (1841). The NSI can be described as: the set
of distinct institutions which jointly and individually contribute to the development and
di usion of new technologies and which provides the framework within which governments
form and implement policies to in uence the innovation process (Metcalfe, 1995). The growth
in size, capitalization, capability, and e ectiveness of these institutions comprises the process
known as industrialization.
This process accounts for the emergence of the cleavage between developed and un-
derdeveloped countries (Cooper, 1992). Many national systems, which were not among the
early pioneer processes of industrialization have continued to struggle in their search for de-
velopment. Despite this, industrialization after the Second Industrial Revolution has proved
to be a signi cantly di erent process than that which preceded it at the turn of the century.
At the core of the speci cities of late industrialization there is a particular focus on technical
change processes.
As noted by Porter (1990), innovation is considered the principle engine of capitalist
development. Nevertheless, processes of technical change led by innovations are often a
privilege of industrialized regions and nations. In industrializing economies these processes
are usually limited to the absorption and improvement of innovations produced in other
industrialized countries, rather than in support of local  rms.
However, the identi cation of the fundamental distinctions between the processes of
technical change of economies at various stages of development has rendered largely uncer-
tain results. This is due to the fact that these investigations are usually hindered by the
current tendency towards an increasing conceptual imprecision in the literature on innovation
systems. One reason that is often pro ered as an excuse for this imprecision is the growing
14
Figure 2.2: Representation of the Triple Helix Model
awareness of the complex interactive relationships that exist between invention, innovation,
di usion, and capitalization (Bell & Pavitt, 1993).
Despite this rich interaction, each of those forms of technical change presents enough
di erences to justify their independent existence in the conceptual framework of most knowl-
edge systems. Indeed, the relationships between these notions have vast and varied meanings
for the competitiveness of  rms, industries and nations. Hence, the preservation of the iden-
tity of each one of those concepts becomes crucial. It is particularly important, though, to
recognize each when building a speci c framework for the analysis of technical change in
NSIs.
2.3 The Triple-Helix of Public-Private Partnership for Innovation
To more concretely describe the processes through which the co-evolution between tech-
nological developments and their cognitive and institutional environments change the knowl-
edge infrastructure, the notion of a triple-helix con guration of public institutions, private
enterprise, and government regulatory systems was proposed (Freeman and Perez, 1988;
Leydesdor , 2000).
There are three selection environments, or spheres, speci ed in the Triple Helix Model
(Leydesdor . 2000):
15
1. Wealth generation (industry)
2. Novelty production (academia)
3. Public control (government)
A model which captures the cooperation, di usion of resources, and transport of prod-
ucts across each sphere can be referred to as a Triple Helix Cooperative Innovation (THCI)
model. The con guration of successful self-sustaining knowledge economies has been rec-
ognized as a key driver of economic growth at the regional and national levels (Tesfatsion,
2002). Evolutionary analysis of the interactions between each sphere in the triple helix fo-
cuses on outputs. This can be contrasted with historical analysis, which informs us about
how institutions and institutional arrangements carry certain functions (Leydesdor and
Meyer, 2006).
The triple-helix model, like most representations of an innovation economy, is composed
of many actors whose collective behavior enables the  ow of technology and information
among people, enterprises, and institutions (Prevost et al., 2004). In this way it is not
dissimilar from an ecosystem, a notion which we will revisit in later chapters. When viewed
through the lens of economic theory, an innovation economy is an economic model that
e ectively promotes the collection and di usion of knowledge product throughout the system
(Holling, 1987). This particular view a ords the notion that an innovation economy can be
represented principally as a resource management problem. Under neoclassical economics,
the exploitation of resources is explained using microeconomic theory (Kreps, 1990). Under
this theory various categories of actors make decisions based on the scarcity of available
resources (Ferguson, 1969). Models based on this theory have been used to study the way
changes in resource density modify the behavior of rational actors (Akaishi & Arita, 2001).
A part of economic literature on resource management is based on a normative approach;
a methodology that concerns itself with the measures that must be taken to cause the
behavior of the producer or consumer of a resource to reach an optimal solution for resource
16
gathering. This approach implies that some system of supervision or control be imposed by
a central entity (Sutinen, 1995). These measures comprise a suite of management tools and
controls for restricting access to resources or in uencing the costs associated with production.
This particular belief is echoed within innovation systems literature, which recognize the
importance of this central authority in developing and maintaining the three spheres of the
triple-helix. However, the complexity of this model leads to uncertainty with regard to
the relationship between environmental and exogenous parameters that a ect economic and
technological growth. Various methods for addressing this complexity have been proposed,
but one recent approach that shows particular promise is known as ecological modeling
(Yawson, 2008), which we will discuss in the later sections.
2.3.1 Government Intervention
The role government control plays in the development and long-term management of
economies has long been controversial. However, various strategies of government partnership
and control are common in most NSI models. These include policy instruments such as:
 Internal R&D.
 Trade taxes
 Grants & direct subsidies
These policy initiatives have the potential to change relative product and factor prices
as well as generate shifts in economic resources between consumption and R&D activities.
The long term impacts of a government?s use of policy instruments to a ect growth of a
regional economy are well documented. In fully endogenous growth models, a policy-induced
shift in per-capita resources towards R&D activities has been shown to permanently accel-
erate the rates of innovation and growth. In semi-endogenous growth models this resource
shift generates a temporary increase in the rate of innovation. This phenomenon was more
17
formally analyzed in the case of structurally identical industries and countries by Dinopoulos
& Segerstrom (1999) and Segerstrom et al. (1990).
The complexity of such systems cannot be ignored. Indeed, the interaction of related
economic policy mechanisms can have wildly unpredictable combinatorial e ects. Grossman
& Helpman (1991) state clearly that, in the presence of asymmetric industries and countries,
general equilibrium interactions can reverse the desired e ects of several policy interventions.
In one example, an R&D subsidy may cause a country to export fewer R&D intensive goods
and import more of them, thus creating a dampening e ect on local technology  rms. In an-
other example, industrial policies that subsidize manufacturing activities in high-technology
sectors may have detrimental e ects on global long-run innovation and growth because they
could raise the aggregate costs of R&D.
2.3.2 Appropriability
Interventions are not limited to cost adjustment mechanisms. Government policy can
also a ect appropriability or the environmental factors that govern an innovator?s ability
to capturepro tsgenerated by aninnovation. The portability of new knowledge across an
innovation economy has as much to do with the social structure of the system as it does the
prevalent economic incentives present. This distinct property is the basis for cooperation
between the three helices of an innovation economy. As a result, policies related to the
appropriability of innovation not only have drastic economic impacts, but also change the
social behaviors, which play a signi cant role in the economic interaction between  rms.
Social intervention strategies can generally be thought of as government?s economic \soft
power" to indirectly manipulate or change behavior by altering the risk-reward calculus in
the innovation process without the use of direct capital or labor expenditure.
There is widespread agreement that in a perfect competition setting, that is, a situation
in which, among other assumptions, no producer has market power, there is no product
di erentiation and all  rms have immediate and perfect access to the same technologies, the
18
rate of innovation in a market economy would be very low. Thus, the creation and regu-
lation of mechanisms like intellectual property rights (IPRs) is seen as the most prominent
pillar of government appropriability policy. IPRs have attracted increasing attention both
in academic circles as well as in public policy debates over the past decades. This has gone
hand-in-hand with their increasing use, particularly (but not only) patents, reforms in the
national and international legal frameworks that have resulted in the strengthening of IPRs
and the fast growth of sectors in which knowledge, innovation and appropriability play a
key role (e.g. biotechnology, information and communication technologies and the cultural
industries).
IPRs, including patents, copyright, trademarks, industrial designs, utility models and
plant breeders? rights, are some of the appropriability mechanisms that may be regulated by
the government. However, as is well known, there are other available mechanisms, including
policies related to activities such as:
 exploitation of lead time
 moving rapidly down the learning curve
 use of complementary manufacturing capabilities
 secrecy
A more complete summary is available in Cohen et al. (2000).
Since labor mobility is also a form of technology imitation, labor legislation, contracts
and human resource management practices are also very relevant appropriability mecha-
nisms (Hurmelinna & Puumalainen, 2007), although some of those mechanisms could be
included under the heading of secrecy. Much of the literature on non-market related govern-
ment intervention is devoted to issues regarding appropriability of innovation via the patent
system.
Some models studying growth of innovation economies have analyzed the dynamic ef-
fects of stronger protection of IPRs. In these models, strengthening IPR is captured as either
19
an increase in patent length as shown in Segerstrom et al. (1990) or a reduction in the rate of
imitation as shown in Krugman (1979). Both models rely on a North-South Schumpeterian
growth model, where two distinct economic zones compete with one another by either inno-
vating internally or imitating innovation in the rival community. Stronger IPR protection
was shown to reduce the rate of international technology transfer from innovating North to
imitating South. This in turn raised the North-South wage gap, and had an ambiguous e ect
on the rate of innovation and global growth.
Helpman (1993) showed that even if stronger IPR protection increases the rate of North-
ern innovation in the short run, it could reduce the welfare of Southern consumers and could
raise the welfare of Northern consumers by shifting production from low-price South to
high-price North. The IPR model has been shown to be relevant not only in understanding
the relationship between industrial  rms, but also to the behavior of academic institutions
engaged in R&D.
Recognition of the value of public-private partnerships led to the creation of Institu-
tional Patent Agreements (IPA), which, among other things, allowed universities and non-
pro ts with approved patent policies to retain title to inventions developed with government
funding. This would later be followed by the adoption of thePatent and Trademark Law
Amendments Act or theBayh{Dole Act, which expanded these rights. This policy legis-
lation is largely credited for paving the way for the later success of  rms transforming of
grant-funded scienti c research into pro table consumer products, evidenced by the rapid
expansion of the country?s technology sector.
2.3.3 Government Grant System
Grants are a method of funding whereby one party (Grant Issuers), often a government
agency, fund, trust, or other interested  rm disburses funds to another entity, business, or
individual. While the notion of a grant support at the federal level in the United States
has existed in some form going back as far as Federal Tax Act of 1913, regional targeting
20
of grants in support of scienti c research is a relatively new concept. The system itself is
composed of various classes of issuers and applicants whose relationship generates product
in the form of innovation and capital. This architecture of the US government grant system
(GGS) is laid out in the Federal Grant and Cooperative Agreement Act of 1977, which
states that the principal purpose of the relationship between the GGS and researchers is to
transfer a \thing of value" to the government or other recipient ( rm) for the public purpose
of support or economic stimulation.
Government programs that subsidize commercial R&D are justi ed on the grounds that
pro t-maximizing  rms underinvest in R&D. Government funding of private R&D projects,
however, can increase R&D e ort only if the subsidies cause  rms to undertake projects that
would otherwise be unpro table. Empirical studies to measure the e ect of these government
grants typically regress some measure of innovation or  rm productivity (e.g., R&D spending
or employment) on the subsidy. Many of these studies  nd a positive correlation between
government R&D funding and private R&D e ort and employment (e.g., Levy and Terleckyj,
1983; Robson, 1993; Nadiri, 1993; Irwin and Klenow, 1996; and Lerner,1996).
Until recently, most federally funded industrial R&D was directed at government needs,
such as large weapon systems. While federal support for private R&D dates back to the 19th
century, government-industry R&D programs have become increasingly popular in recent
years. Many of these programs aim to help  rms commercialize innovations by subsidizing
their R&D. Cohen and Noll (1995a) and Nelson and Romer (1996) document the increasing
popularity of these programs and the reasons for their rise. Others examine individual
government-industry R&D programs. The economic justi cation for these programs is clear:
The social returns to private R&D are often higher than the private returns, some research
projects would bene t society but would be privately unpro table. By lowering the cost to
the  rm, a subsidy can make these projects privately pro table as well.
21
Figure 2.3: Architecture of a grant system.
For the purposes of our investigation, it is useful to regard a grant as a replenishing
resource that exists in a notional environment whose connection with institutions and in-
dividuals is governed by suitability, or theoretical closeness. This closeness is de ned as a
relationship gradient between the grant?s issuer and potential institutional applicants. The
distance between a grant and an institution is analogous to the amount of ?work?, which
must be done to create a successful application.
In a more concrete sense, grants are both strict resources as well as renewable environ-
mental properties. If one were to represent the grant system as a natural ecosystem, grants
would be the equivalent of producers, such as algae or grass. The entities which collect and
transform grants are analogous to ecological consumers, which transform environmentally sit-
uated resource into higher order products. In this way, the relationship and behavior of the
artifacts and  rms in the GGS approximate a simple energy  ux. This  ux can be captured
in a variety of natural models, explicitly those that map to grazing or foraging activities.
The more general form of this interaction is known as the predator-prey relationship.
2.4 Modeling Innovation Economies
The goal of system modeling is to create a realistic abstraction of the system in question.
Due to inherent complexity of real-world socio-economic systems, it is generally not possible
for a model to capture every detail. Innovation economies present additional challenges in
22
that their behavior is often unpredictable by virtue of being driven by processes of creative
destruction, complex psycho-social phenomenon, and exogenous environmental conditions.
Thusly, most models in this domain must be ne-tuned and designed based on a specic
hypothesis to be tested for the corresponding system. The basic implication is that models
based on the same system can be structured in a myriad of di erent con gurations based on
the abstraction chosen by the designer.
In a general sense, modeling has been shown to be very useful in understanding and
predicting various properties of complex systems. Stareld et al. (1990) refers to models
as \purposeful representations" and \tools for problem solving" they enable the central
tenants of a problem to be captured. Most importantly, models help facilitate prediction
and understanding of system-level properties. In suitably complex systems these properties
may be obfuscated by the behavior of constituent mechanics. The modeling process involves
a thorough analysis of the key parameters a ecting the behavior of the main components
of the system and their interactions with one another. In return, this provides the modeler
with valuable insights on the overall patterns of organization within the system.
Robinson (2006) makes the claim that the output product of a model is more than a set
of results that match some experimental data. Indeed, the approach as a whole allows us to
test the role of mechanisms which may be impossible to manipulate, or even capture properly
in the existent systems from which the model in question is derived. Moreover, using models
allows for repeated simulation and testing of various mechanisms and scenarios which can
be constructed merely by tuning parameters, thus avoiding di cult and time-consuming
real-world experimentation.
Deterministic modeling usually involves manually analyzing data and deriving di eren-
tial equations which approximate and \average: behavior of the system. These approaches
often fail to capture a large number of interactions between components which could include
factors leading to population-level behavior. Consequently, methodologies focused on the
creation of individual-based models have become more en-vogue across the many analytical
23
disciplines which concern themselves with complex adaptive systems (CAS). These models
are often considered more realistic and better suited to capture interactions at the individual
level and provide deeper insights into the reasons behind population level behavior.
An important aspect which a model designer should remain cautious about is the number
of iterations required for a model to be simulated. It is important to have a su cient number
of runs for the system to evolve from smaller parts, otherwise community-level behaviors may
not be observed e ciently or e ectively. There are two main approaches to modeling complex
adaptive systems. These are the top-down, and bottom-up approaches which are discussed
in the following sections. Though omitted for brevity, it should be noted that a combination
of the two-approaches, referred to as \integrated" or \middle-out" modeling, also enjoys
wide use.
2.4.1 Top-down modeling
Based on mainly ordinary or partial di erential equations, subject systems are analyzed
in a top-down manner in terms of a population of identical individuals. Individual di er-
ences in population are ignored and the underlying behavioral rules are usually treated as a
\black box", or approximated at the best case. Essentially, this means that the designer is
less concerned with the individual mechanics that govern the behavior of a population and
instead focuses on the observable changes. The methods through which top-down models,
also referred to as state-variable models, are essential and have proven useful in formula-
tion of general theories. One popular example is that of the Lotka-Volterra model of the
predator-prey interactions. Despite this usefulness, biological and, most importantly, eco-
nomic systems tend to represent a great challenge for mathematical modelers, due to their
inherent complexity and non-linear relationships among individual components. In the study
of the social insects, successful applications can be found in the works of (Bonabeau et al.,
1998). These models do often successfully describe the behavior of some homogeneous pop-
ulations, but they also limit the understanding of how patterns are formed. Such models
24
may apparently work well, but in order to understand the exibility and diversity of colony
activity, it is also essential to investigate the agents at the individual level. Rather than car-
rying out a full investigation on the underlying mechanisms of the system, top-down models
average individual behavior over the whole population.
2.4.2 Bottom-up modeling
In bottom-up modeling, the system is modeled at the individual level, aiming to study
the mechanisms for underlying behavior as well as local interactions between the individuals
in a population (homogeneous or heterogeneous). Bottom-up models tend to emerge from
low-level components of the system, and are studied in order to understand how the overall
pattern of the system emerges. This method emulates the behavior of the system under
study in the most realistic way, but can be taxing when used to analyze scale systems with
large populations. Recently, bottom-up approaches have started to attract more interest
in modeling complex economic systems. This is due to researchers realizing the potential
advantages of bottom-up approaches over top-down. In a top-down model, it is not possible
for a user to trace back the system properties to the behavior of an individual actor, which is
critical when modeling economic systems. In contrast, this is easily achieved in bottom-up
models.
Cellular Automata (CA) is a popular method in this regard. CA emulate real physical
laws using a small set of simple rules by limiting all properties to a few states, frequently no
more than two (Wolfram, 2002). Developed by Ulam and von Neumann, CA was founded
upon the notion of one robot building another robot, aiming to address the problem of
self-replicating systems. One such classic example is John Conway?s ?Game of Life? by
which a CA model operating on four simple rules of generations allow individual cells to
react and change their states based on their interactions with surrounding neighboring cells.
CA has been e ectively used to explore many problems in  elds such as biology, sociology,
and psychology, but su ers when applied economic system models because it treats  rms
25
as homogeneous, without any exibility or variability, which are key characteristics of such
systems.
Traditional bottom-up models focus on the individual components of a system. Con-
sequently, the set of methodologies conforming to this approach are often referred to as
individual-based modeling. Each individual system component is treated as an agent employ-
ing certain behavioral rules (agent-based modeling), where over time interactions between
these agents will give rise to characteristic patterns at the population level. In modeling so-
cial groups (such as insects, societies, etc), successful implementations of such models could
be found in the works of Deneubourg et al. (1989). Agent-based modeling is discussed in
detail in the next section
2.4.3 Agent-based modeling
Agent-based models (ABM) provide a platform upon which simulations of large number
of agents can be constructed. These models encourage bottom-up approaches by allowing
the designer to focus on the interaction between individual elements of the system, rather
than trying to describe behaviors at the population-level. The characteristic behaviors of
the individuals are assigned to agents as simple rules. These rules govern how and where
each agent is able to make decisions, as well as providing additional constraints based on
the problem domain. As a result of various interactions between the agents through the
environment, certain population-level dynamics emerge. The most common use of agent-
based models involves mimicking complex systems, in which theoretical lessons could be
learned according to the results of the simulations.
In the most basic sense, an agent is a computer system that is situated in some envi-
ronment and that is capable of autonomous action in this environment in order to meet its
design goals (Wooldridge and Jennings, 1995). Once the lowest-level components of a system
are identied as agents, simple rules governing their real behavior are applied. These rules
26
Figure 2.4: An agent interacting with an environment.
then facilitate interaction between the agents and leads to emergent behavior over time. An
agent may act both on and within anenvironment.
A system composed of multiple autonomous reasoning agents is commonly referred to as
a multi-agent system (MAS). In typical multi-agent systems, an agent?s environment includes
other agents with which interaction may take place. In this context, an agent together with
its environment is called aworld.
Some clear advantages of ABM over traditional modeling approaches, including CA, are
the following:
 Asynchrony: Agents do not need to simultaneously perform actions at constant time-
steps, rather they can follow discrete event queues or a sequential schedule of interac-
tions.
 Spatial: The environment does not necessarily need to be grid-based, nor do the
agents need a direct tie to the environment, which allows cohabitation of agents with
di erent environmental experiences.
These features are powerful as they allow  exibility in representing systems via abstraction.
This is relevant when modeling real-world socio-economic systems which often consist of
layered physical and relational (meta-physical) environments which may or may not be linked
27
directly. Biological models also bene t from this loose environmental coupling, as these
systems are often decentralized. For example; Ant colonies are known for their decentralized
communication system, as they do not have a leader inuencing the colony members. Rather,
their colony level behavior emerges from the interactions between the individual ants, which
is mainly via pheromones deposited in the environment. It is therefore more appropriate
to apply bottom-up agent-based modeling to ant behavior. This way, interesting features
emerge over time, which are not pre-determined in the design at all. (Deneubourg et al.
1990).
While many types of agent modeling focus on cognition, the situated multi-agent system
(situated MASs) modeling avoids individual intelligence in favor of environmental interac-
tion. Contrary to cognitive approaches of agency where a great deal of e ort is devoted to
the formalization of agent concepts, relatively little work has been done on the formalization
of situated multi-agent systems (situated MASs).
Situatedness is a property of agents adopted by most researchers in the domain of MAS.
It is described by Woolridge & Jennings (1995) in their de nition of an agent: ?an agent is a
computer system that is situated in some environment, and that is capable of autonomous
action in this environment in order to meet its design objectives?. In this case, the notion
of situatedness expresses the fact that an agent is not an isolated entity but exists within a
conceptually abstract environment. This de nition deliberately does not prescribe explicitly
what it means for an agent to be situated in an environment, e.g., nothing in the denition
explicitly refers to the fact that the existence of an agent in an environment entails a social
component.
In situated MASs, agents are considered as social entities. Put more directly, the em-
phasis in situated MASs is on ?M? in MAS, rather than on ?A?. This view expresses the
belief that agents and environment constitute complementary parts of a multi-agent world.
In that sense, situatedness expresses the local relationships between agents and objects in
the environment rather than strict location. These relationships give the system meaning
28
and structure, driving the evolution of the MAS as it persists. Through its situatedness an
agent is placed in a context that it is able to perceive and in which it can interact with other
agents. Thus, intelligence in a situated multi-agent system is derived from these interactions,
rather than from capabilities of individual agents.
Despite its relative lack of attention within the domain of ABM, the situated approach to
MAS design has a long history. Brooks (1987, 1991) identied two central ideas of situatedness;
embodiment and emergence of intelligence, while Steels (1990) and Deneuborg et al. (1986)
introduced basic interaction mechanisms for agents to communicate and coordinate through
the environment: gradient  elds, and marks. The work of Maes (1990) introduced robotic
principles of reactivity to the  eld of software MAS. Dorigo (1992), Corloni et al. (1993),
and many other researchers drew inspiration from social insects and adopted the principles
in situated MASs.
Situated ABMs have been applied with success in numerous practical applications over
a broad range of domains, some of which will be explored in greater depth in later sections.
Examples of such applications include: manufacturing scheduling (Parunak et al., 1998)
and supply chains systems (Sauter & Parunak, 1999), articial worlds and social simulation
(Macy & Willer, 2002), network support (Bonabeau et al., 1998) and peer-to-peer systems
(Babaoglu et al, 2002). Situated MAS o er very clear bene ts, the most important being
 exibility, robustness, and e ciency. Despite these a ordances, this approach comes with
a number of drawbacks which can be attributed to the inherent qualities of the approach,
e.g. decision making of situated agents based on local, current information. This means
that situated agents by de nition have a short-term view of the state of the world. Other
unsolved problems stem from complexity issues regarding the engineering of situated agents
with respect to desired overall behavior (Rouchier, 1999). Due to the intensely decentralized
nature of the system, and the sheer volume of concurrent, often interdependent, interactions
between agents, addressing this problem has continued to pose di culty to ABM designers.
29
2.4.4 Ecological modeling
Ecology is the study of living organisms and their connection with the environment.
These organisms, called species, are divided into semi-homogeneous groups called popula-
tions and classi ed by their general role in the larger ecosystem. Communities are formed
from various populations of distinct species that all inhabit the same environment or operate
in a dependent fashion. These dependency cycles are often referred to as \lifecycle processes"
which, when collected together and viewed in a top-down fashion describe one of ecology?s
most important and fundamental concepts; the ecosystem (Bosquet & Le Page, 2003).
Tansley (1935) characterized ecosystems as being composed of both the organisms
present in an ecological unit and the \... e ective inorganic factors of its environment".
A major strength of the ecosystem concept is that it is appropriate for any situation in
which the biological and physical interact, leading to the broad utility of the concept across
many domains (Stepp et al., 2003). Some authors and researchers have sought to stretch the
de nition of the ecosystem to suit particular problem spaces, leading to concepts such as the
human and economic ecosystem views. To perform this transformation, it is often the case
that the common de nition is expanded with domain-speci c concepts. However, because
we wish to develop a more fundamental ecosystem model we are principally concerned with
the attributes these varying interpretations share.
In general, an ecological system can be said to have several key properties that helps
distinguish it from other complex systems:
With these attributes in mind, we can construct a general model of an ecosystem in
terms of a lifecycle which connects the members of a community to one another and the
environment through a variety of mechanisms and rules which are governed by the nature of
the system and trophic levels (roles) available (Costanza et al., 1993). The interplay between
the trophic levels may produce cycles that allow energy or material to be transferred through
the environmental strata. This transference can also result in transformation activity, which
30
Table 2.1 Properties of an ecosystem.
allows high-level products to be synthesized through the cooperation of many low-level actors
in a completely undirected fashion.
The concept of energy  ow is probably the most central to the ecological model as it
acts as the basis for making connections between all the constituent components (species
and environment) (McCormack, 2007). The  ux of energy, organic, and inorganic material
forms the crux of the relationships between trophic levels and is generally regarded as the
mechanism through which all systemic change can be measured.
Most ecosystem models adhere to the preceding classi cation of roles, naturally organiz-
ing the dependent relationships in a graph known as a food web (Prevost et al., 2004). Links
in a food-web illustrate direct trophic relations among species, but there are also indirect
e ects that can alter the abundance, distribution, or biomass in the trophic levels. Examples
of this can be found in natural systems such as when the over-predation of a primary con-
sumer (such as deer or bison) by an apex consumer results in the aggressive expansion of a
previously stable species of plant. In this case, the apex consumer was indirectly in uencing
the regulation of the plant species through its action on the primary consumer. The net
e ect of these direct and indirect relationships is called trophic cascade and is separated into
species-level and community-level cascades.
The potential for radical environmental change, whether through stochastic models or
the activity of communities, introduces a level of uncertainty not often tolerated in engineered
31
systems. This fact, coupled with the lack of a central agency driving the system towards a
particular set of goals, allows ecological systems to shift between various regimes while still
maintaining equilibrium conditions.
Patten (1978) made a similar observation in his expansion to the Tansley model of
ecosystems. However, the Patten view includes both input and output models in the formu-
lation of the environment such that;
Input Environment + System + Output Environment = Ecosystem
The resulting model proves to be more easily generalizable beyond the ecological domain.
The concepts from which it is derived places an emphasis on the signi cance of indirect
causality, which is realized through complex networks of species interactions and energy
 ows. This notion permits the exploration of the role of information in ecosystems and
more concretely de nes the parameters for the feedback cycle which lies at the heart of all
ecological systems.
To properly de ne principles of ecological management, we turn to Holling (1987)
who lays out three central concepts that have dominated causality in ecological systems
and that demarcate the principles for the management of ecosystems. The  rst concept is
based on the notion of equilibrium (balance of nature). The second concept de nes several
states of stability (nature engineered or nature resilient). This second perception is primarily
concerned with dynamics caused by variance (stochasticity), usually attributed to events that
occur at small scales.
The third concept is that of organizational change (natural evolution), wherein he notes
that not only does the system undergo change (external events lead to perturbation of
the system), but also the actors of the bespoke system may, by themselves, change its
organization. This is especially true when human interference is taken into account as
a feature of the system. This  nal concept corresponds to the approach adopted by the
sciences of complexity: i.e. the general state of a set of interacting entities may converge
32
toward attractors, may be disordered, or may exhibit patterns of organization that change
from one to another in an unpredictable way (Wolfram, 1984; Langton, 1992). The study
of these systems relies on observations that focus on the connectivity of the ecosystem?s
elements, their interactions, and their organization across various scales.
2.5 Verifying & validating simulation models
The proliferation of simulation models as problem-solving and decision-making tools
has caused serious questions to be raised about the accuracy and usefulness of the informa-
tion obtained from their operation. Ensuring ?correctness? is paramount. This concern is
addressed through model veri cation and validation (V&V) processes. Model veri cation
can be de ned as \ensuring that the computer program of the computerized model and
its implementation are correct" (Sargent, 2009), which we will use here. Schlesinger et al.
(1979) o ers a similarly suitable de nition of validation as \substantiation that a comput-
erized model within its domain of applicability possesses a satisfactory range of accuracy
consistent with the intended application of the model". These de nitions are symbiotic and
imply that a model should be developed for a speci c application and its veri cation and
validation criteria be evaluated with respect to that application. In the case of validity,
a model is considered valid for a set of experimental conditions if the model?s accuracy is
within its acceptable range. This satis es the notion that absolute validity is often too
time-consuming, expensive, or computationally impossible given development or operational
constraints. Formalized model validation approaches vary depending on the domain. Where
development teams are involved, the model development team itself may make the decision
as to whether a simulation model is valid. In other instances expected users supply valida-
tion criteria or provide the validation themselves through direct observation of simulation
output. The use of outside observers is common in \independent veri cation and validation"
(IV&V) where an independent third party determines validity (Wood, 1986). Finally, scoring
models, such as those proposed by Balci (1989), where weights are determined subjectively
33
Figure 2.5: Simple version of the modeling process (Sargent, 2005)
when conducting various aspects of the validation process and then combined to determine
category scores and an overall score for the simulation model.
Following the ?simple? modeling process ( gure 2.4), model validation can be further
re ned into two main components (Sargent, 2005);
 Conceptual model validation: Determining that the theories and assumptions underly-
ing the conceptual model are correct and that the model representation of the problem
entity is \reasonable" for the intended purpose of the model.
 Computerized model veri cation: the computer programming and implementation of
the conceptual model is correct.
Valid simulation models are constructed iteratively, with each successive model subjected
to V&V checks to ensure correctness (Sargent 1984). There exists to-date no algorithm or
procedure to select which techniques to use during the V&V stage of each iteration. Instead,
this is left to the discretion of the researcher or development team. Some attributes that
a ect which techniques one might use to make this decision are discussed in Sargent (1984).
In general, agent-based simulation model V&V is di cult. However, those situated in
the biological and socio-economic domains present particular validation challenges. Grimm
34
(1999) reviewed 50 individual-based animal population models, providing a detailed compar-
ison between individual-based and statistical modeling approaches. According to Grimm;
the expectations from individual-based models were not fullled. This is because it was ex-
pected that once the rules and characteristics of individuals have been assigned to agents,
the population and community-level consequences would emerge naturally (DeAngelis et al.,
1994). However, out of the 50 individual-based animal population models reviewed, only
36% of them were found satisfactory and the remaining 64% were found to be generally in-
adequate. Grimm (1999) explains the reason for this disparity was the absence of a general
strategy for building and analyzing individual-based models. His suggestion upon this was
to start with a very coarse model that reproduces a pattern, followed by re ning the model
step-wise, and  nally checking and testing the model rigorously with each renement. It was
concluded that bottom-up approaches alone will not lead to theories at the system level, due
to the need for top-down approaches to provide an appropriate integrated view. However, it
was also underlined that it will never be possible to fully understand a system?s properties
unless the mechanistic interactions through which they emerge are well understood.
The solution to this V&V challenge is to express individual-based models in a formal
framework (Kefalas et al., 2003), as well as adopt formalized methods for model creation
similar to those described in Sargent (2005). The formal specication of a model facilitates
model checking, which in turn guarantees completeness of a model with respect to require-
ments. Merely testing a model via simulation may only reveal surface-level inconsistencies or
misconceptions in the model and is insu cient grounds to concretely validate or invalidate
a design.
From this survey we conclude that model V&V is performed to assure quality as well
as theoretical soundness. Carley (1996) grouped quality into the  ve categories seen below:
 Theoretical validity: Adequacy of the underlying theoretical model in characterizing
the real world.
 Internal validity: Whether the underlying assumptions are consistent.
35
 Operational validity: The adequacy and accuracy of the model results when compared
with experimental or system data (Sargent, 2005).
 Cross-model validation: Comparison of two models against each other.
 Data validity: Accuracy of the data (real and computer generated) and the data?s
adequacy for addressing the issue of concern.
In this research, attempts are made to ful ll each category by taking the following factors
into consideration:
 Face validity: The model results or simulations ?appear? to be consistent with reality.
 Parameter validity: The parameters of the model match available experimental data.
 Process validity: The process described by the model corresponds to the observed or
implied mechanisms of the socio-economic system (Yilmaz, 2006).
 Pattern validity: The pattern of results generated by the model matches the patterns
of results from similar studies (where available).
In addition, a number of validation techniques were used. Adapted from Carley (1996) these
can be summarized as follows:
 Grounding: The simplications made in the model design should not a ect providing
important insights. Parameters and states should be set in accordance to real rules,
and wherever possible it should be demonstrated that the model results and behavior
are consistent with the real behavior.
 Calibrating: Models should be extensively tuned to the mechanisms of the system
under study by parameter estimation and/or altering algorithms and rules within the
model.
 Verication: The model results are compared graphically or statistically with experi-
mental data and previous related  ndings.
36
It should be noted that while model validation requires a close accounting of the parameters
and mechanisms of the system, our veri cation process is constrained to those areas where
empirical data is available. Since the domain of innovation system dynamics is new, there
will be inevitable gaps in data coverage. To this end we supplement our veri cation process
with a focus on repeatability.
2.6 Chapter Summary
In this chapter we provided a summary of relevant works in innovation economics and
agent-based modeling. In section 2.1 we discussed the history of regional and national sys-
tems of innovation. In section 2.2 we outlined the role knowledge plays in productivity and
competitiveness of economies. In section 2.3 we outlined the triple-helix theory of public-
private partnership, paying particular attention to the properties and mechanisms which
govern the behavior of the three regimes of control; academia, private industry, and govern-
ment. We also explored the theoretic environments within which triple-helix actors exist and
interact with one another. In section 2.4 we discuss a basic framework for understanding
holistic modeling of innovation ecosystems, introducing the principles of ecological modeling
as an alternative to top-down or bottom-up interpretations of complex socio-economic sys-
tems. In addition we cover the basic tenets of agent-based modeling that we will revisit in
chapter 3. Finally, in section 2.5 we address various validation criteria for models and clarify
the standards we will use to instill con dence in the conceptual design and experimental
operation of our simulation model.
37
Chapter 3
Agent-based Perspective for Simulating Innovation Dynamics
3.1 An Ecological Approach to ABM
In order to e ectively map multi-agent system design methodologies to the ecological
model of innovation economies it is important to delineate the criteria through which we
cast an ecosystem as a complex adaptive system (CAS). Indeed, Levin (1998) identi ed
three important criteria which a system must satisfy in order to qualify as a CAS:
1. Sustained diversity and individuality of components.
2. Localized interactions among those components.
3. An autonomous process that selects from among those components, based on the results
of local interactions.
These properties are simple, yet expressive, and allow us to observe that CAS arise from
the conjunction of three main processes: one that creates and supports diversity, one that
supports interactions between components (species), and one that selects on the gradient
between form and function (Levin, 1998). The  nal product is the community of compo-
nents which continuously changes towards a dominance of those best suited to deal with the
selective forces and limiting features of the environment.
This mapping also helps us ground the contention that ecosystems are prototypical
examples of CAS that not only exhibit a dispersed and autonomous selection process but
also continual adaptation (Brooks, 1991). The maintenance of diversity and individuality
of components implies the generation of perpetual novelty. This process closely resembles
the cycle of innovation creation described in by List (1841). Thus, the preceding listed
38
set of ecosystem processes cleave to the essential de nition of features of a CAS while also
supporting the central intent and processes of an NSI.
Following on, we also observe in literature that the development of MAS has been intrin-
sically linked with both the ecological and economic domains for a long time (Baray, 1998).
The legacy of decades of cross-pollination between ecology and agent simulation is realized
in some of the earliest formative work in the  eld, such as the ant-hill metaphor (Babao glu
et al., 2002); which provides an oft used illustration representing concepts of reactive agents
and emergent behavior. Later applications would be found in Hogeweg & Hesper?s (1983)
work on bee colonies and the development of nature-based swarm dynamics such as ant-
colony simulation (Dorigo et al., 1999) and Reynolds? (1987) work simulating the migratory
behavior of birds which predated organized notions of MAS or Arti cial Life. Indeed, the
adoption of agent-based simulation techniques was picked up early by the ecological com-
munity and mostly used in environmental applications which involved interactions between
natural and social dynamics such as land-use management or  sheries (Beckenbach et al.,
2008).
From this we can gather that there are properties of agent simulations which a ord a
mapping between MAS architectures and ecological NSI models. In the context of innovation
individual actors and organizations are typically characterized as having bounded rationality,
presumed to be acting in what they perceive as their own interests, such as reproduction,
economic bene t, or social status (Axtel et al., 2000) using heuristics or simple decision-
making rules. From a MAS perspective, some agents may experience learning, adaptation,
and reproduction.
The individual agents within MAS have several important characteristics that a ord the
distributed nature of the problem domain addressed in this thesis, the NSI. These detailed
below:
 Structural Decentralization: The system displays collectivist behavior, eschewing
central behavioral management in favor of local autonomy.
39
 Self-Organization: The system allows agents to de ne their own hierarchal or rela-
tional structure based on their interaction rules.
 Limited View: Agents do not possess a global view of the system. This is often a
practical concern as the domain is often too complex for a single agent to make sense
of in a reasonable period of time.
 Autonomy: Each agent in the system acts in pursuit of its own agenda over time and
exhibits goal-seeking behaviors.
Since we are observing MAS applied to an ecological model, despite its grounding in
economics, we must also concern ourselves with interaction style. There are three major
types of interactions in an ecological system:
 Interactions via agent communication.
 Physical interactions (grow, eat, push, etc).
 Interactions mediated by the environment.
Unlike in other scenarios (social, cognitive, etc) direct interaction through the exchange
of messages is exceedingly rare in ecological applications. However, some examples of this
can be found in literature such as contract negotiations for goods and services among human
agents or the communicative predator-prey model put forward by Baray (1998). These
behaviors are also prevalent in con ict simulations (Carley & Tsvetovat, 2004). The second
type of interaction, physical, is also the most commonly found interactions in an ecological
model and is typi ed through inter-agent behaviors such as predation.
The third type of action, environmental, is a reaction to questions raised by the Arti cial
Life community, namely; what is the relationship between an object and its environment?
In their view, the environment cannot simply be taken to mean all other agents; instead
the environment means the physical space and the resources it contains. This notion dove-
tails with the concept of externality used by economists. The results of an agent?s actions
40
transform the environment with a retroactive e ect on other agents. For certain mechanisms
(stigmergy, etc) the dynamics of the environment become a medium for collective adaptation
(Dorigo et al., 1999).
Having generally described the relational mapping between the MAS and ecological NSI
model architectures, we proceed on to examine work done to concretize the notion of the
agent-based innovation ecosystem model.
3.2 The Agent-based Ecosystem Model
The multi-agent methodology has been successfully applied to problems in a variety of
domains through the development and analysis of simulation models (Antona et al., 1998).
ABS are especially useful for examining the emergence of macroscopic properties of related
communities of actors (Baray, 1998). This analysis is often approached in a bottom-up
fashion, with close attention paid microscopic interactions between agents and their collective
behavior over time. This more holistic view, which combines top-down and bottom-up
strategies, is known as agent-based ecosystem modeling (Picket & Cadenasso, 2002).
Ecosystem modeling has been applied to the analysis of a broad range of natural and
arti cial systems through the application of theories derived from such disparate areas as non-
linear physics and dynamical systems (Parrott & Kok, 2000). The  eld is, therefore, highly
multidisciplinary, bringing together researchers in all specialties, ranging from economics and
social policy to biology, physics and modern visual arts (Klomp & Green, 1996). Despite
this nexus of expertise, e ective applications of ecosystem theories to practical  eld problems
su er from the lack of knowledge-models and their parameters, as well as uncertainty. Model
parameters are often di cult, if not impossible, to measure directly because of time, capital,
methodology, scale, or conceptual constraints. Despite this limitation, the development of
agent-based simulation models which capture the critical functions of target systems can be
used to both discover and quantify parameters of interest (Assad & Packard, 1992).
41
Table 3.1 Comparison of the natural ecosystem and the innovation ecosystem
The work of Leydesdor & Etzkowitz demonstrated that, at its core, a knowledge
economy can be reduced to a theoretic model whose interlocking components display the
characteristics of a complex-adaptive system (Leydesdor & Etzkowitz, 2005). Further in-
vestigation has indicated that the relationship of the three spheres in the triple-helix model
discussed earlier can be captured in a MAS model (Dongsheng & Yongen, 2008). We also
see that the actors within innovation economies can be represented in ecological terms (Dai
et al., 2007), and that the macroscopic relationships between classes of THCI system actors
approximate that of the classic predator-prey model (Moore, 1993).
3.2.1 Resource Management Models
In the literature on resource management, a common representation of an economic
environment is based on an ecosystem model. Under this model, resources are represented
globally, and the actions of actors, or agents, upon those resources are represented via a
cost function (Antona et al., 1998). Ecological and socioeconomic variables such as resource
renewability, demand, harvest intensity and its cost to various agents are synthesized. In this
scenario, the harvesting of a resource is the sole objective of the agent. Transformation and
42
exchange are rendered secondary and their rules are often unspeci ed (Rouchier et al., 2001).
Thus the assumption then is that these behaviors are only necessary for setting equilibrium
prices and creating synthetic resource artifacts and that the scale at which management
operates is at the harvesting stage, directing the activities of uniform agent (Mathevet et al,
2003; Mathevet et al, 2003b).
Alternatively, there are other resource models which are interested in more than just
primary harvesting (Janssen & Carpenter, 1999). These models consider that the resource
is subject to the derived demand behaviors of  nal users rather than that of the primary
harvester. Moreover, these demands may be for secondary resources which must be syn-
thesized from primary resources. Thus, this representation takes into account not only the
heterogeneity of agents, but also that of resources (Costanza et al, 1993). Additionally, this
view takes into account agent location with respect to resources and the act of secondary
production.
For our model, we chose to represent resource exploitation in the form of innovation
communities, i.e. agents with various tasks such as harvesting, consuming, and transfor-
mation resources (Beckenbach et al., 2008). These communities roughly equate to regional
industrial conglomerates in economic parlance. Using this form of representation allows us to
formalize exchange and transformation of resources through various stages of the innovation
economy. Within economic literature on renewable resources, this representation is typically
found in cases where this  ow of resources between levels is represented as a series of cost
functions (Charles, 1988).
In this model, the method for coordination of the resource model comes through  xed
cost functions associated with each agent, with global management policy a ecting the place-
ment and amount of resources available at each location in the environment. Because pricing
is loosely  xed, we do not observe the market-based dynamics that often act as a balancing
mechanism for system behavior at the same scale as have been observed in existing systems
43
(Rouchier et al., 2001). The precise mechanics of these functions will be discussed in later
sections.
Despite the above limitation, the proposed model is relevant to the discussion of man-
agement controls as a set of collective rules even if the selection is somewhat limited. The
economics of resource management describes two main categories of harvest tools (Nag et
al., 2007):
 Those that reduce harvesting intensity or inputs (number of harvesters, input costs,
etc)
 Those that directly limit the amount harvested (outputs)
Inputs are reduced by imposing a limit on the number of harvesters or via some taxation
scheme imposed on the harvested resource. These use-conditions are particularly relevant
in the discussion natural resources ( shery, forest, and aquatic environs), but may be less
accurate when dealing in purely economic constructs.
3.2.2 Resource Management Instruments
We will brie y discuss three main instruments in the study; resource spatial positioning,
taxation and harvest quotas as they pertain to our model.
Resource positioning provides a simple ecological analogue to environmental con gu-
ration. The physicality of resources in relation to actors can be a major factor dictating
exploitation in renewable resource models (Tansley 1935). In theoretic models this notion is
often described via density ratios, with low density implying local scarcity, and high resource
density implying local abundance. This is particularly pronounced in ecological models per-
taining to land-use, as the primary driver of action is degree to which environmental resources
are utilized by agents (Sutinen, 1993). In economic literature, the cost function associated
with resource harvesting often contains some aggregate calculation for work, which accounts
for all the activities performed by an actor for the purpose of securing and transporting a
44
resource. In human endeavors, and more speci cally in our model, the placement of resources
within the grid space gives provides us a metaphor through which to translate this equa-
tion (Stepp et al., 2003). Spatial distance within the environment translates to the relative
amount of work a harvester must spend to exploit a particular resource. The actual cost
associated with extracting resource is a function of distance,
From a management perspective, the modi cation of this property is akin to diminishing
or increasing the coe cient for exploitation (work) in the cost function for harvester agents.
As a macroeconomic tool, policies organized around spatial location have profound e ects
on the behavior of the agent community as well as the nature and timing of regime changes
within the system (Pottage et al., 2004). Within the literature pertaining to innovation
economies this equates to the strategic allocation of funding around particular industries
and institutional programs (Axtel et al., 2000).
Taxation provides a direct means for a central authority to suppress or promote a partic-
ular kind of resource utilization or transformation process (Teece, 1998). This action usually
comes through the imposition of rates of exchange coupled with some sort of independent
loss equation. This formulation di erentiates itself from simple waste mechanics in that the
product is often collected (rather than dispersed) and can act as a driving mechanic for other
activities by the management entity that initiated it (Stiglitz, 1999).
In economic models taxation is often a method for resources to  ow out of the system,
while in innovation models it is the means by which the governing actor extracts value from
inter-agent trade (Cook, 2001). In the latter example taxation is an integral component of
the system?s renewal subsystem and acts as a direct reward to the system for cooperative
behavior.
Our model is principally concerned with examining the direct impacts varying tax rates
may have on the system rather than determining an actual e ective tax policy. As such, the
notion of tax is treated an incentive value which directly a ects the individual behavior of
agents. Tax is principally active during transactions downscale transactions (i.e. leveraging
45
of products within the market), thus tax is a share of the gain from the sale of product the
product of innovation (technology) and is taken from the seller (Parisi, 1997). This share is
proportional to the quantity sold. Thus it is possible to examine, through simulation, the
e ects of redistributing the product of this tax at di erent stages of the system.
Harvest quotas o er a simple measure through which actor interaction with the envi-
ronment can be directly regulated by a central authority (Steiglitz et al., 1996). Similar to
spatial positioning, quotas act as a barrier to speci c exploitative behaviors. By limiting
both the amount that can be harvested at once and the amount available to be harvested
from a particular spot, the central authority has a mechanism through which it can guide
the activity harvesters indirectly. Additionally, quotas can also set minimum bid amounts,
ensuring that particular market behavior occur (sales at  xed lot sizes, etc.).
In economic models, quotas exist as hard supply and transport limits that place a
ceiling on resource acquisition functions and essentially behave as  ow rate limiters across
the system (Stigler, 1941). This notion also extends to ecological models in the concept of
resource carrying capacity between trophic zones. Within the innovation economy, quotas
take on a similar role as seen in pure economic models and are a useful tool for guiding the
behavior of consumption and production across multiple communities (Glazer & Hirshleifer,
2005).
In our model quotas merely de ne the maximum amount product a  rm can collect from
a single source in its local environment. The interpretation of the quota varies between the
helices, but is enacted at the global, rather than individual level. More speci cally, this means
that a research  rm may be limited in the amount of money it may draw from a particular
funding source, and that industrial  rms may be limited in the number of technological
partnerships they are allowed to have, or are capable of pursuing simultaneously.
46
3.3 Previous Work in Agent-based Innovation Simulation
Because we are interested in the cooperative and competitive behavior of several co-
dependent species in an ecological framework, we now turn our attention to examples of
relevant developments in the  eld of agent-based innovation simulation.
Previous e orts in this  eld largely been grounded in domains such as cognitive theory
and social network dynamics. Ecological models of regional ecosystems have been proposed
which incorporate innovation concepts in order to capture the e ects of human behavior
(Hare & Deadman, 2004). However, there have been few studies to date which extend inno-
vation models with ecological concepts, and none which apply ecological modeling method-
ology to NSI simulation. Despite this, we explored studies relevant to both domains in order
to inspire con dence in the applicability of said approach.
Innovation transport and di usion models
Several investigations have examined the link between transport costs (risk), the like-
lihood of  rms to form partnerships (inter- rm trust), and proclivity to carry innovation
forward. Findings point toward several best-practice conditions that must be met in order
to maximize di usion. These included incentivizing individuals and  rms with high spatial
centrality to buy in, lowering initial cost barriers to tempt wary individuals, and maxi-
mizing the visibility of  rms within the notional marketplace to extend their trading power
(Terna, 2009; Ma & Nakamori, 2005). These results follow the common wisdom that directly
addressing risk aversion among  rms is a good strategy to promote innovation.
Others suggest that arti cially suppressing risk can be costly and may lead to market
instability, especially in scenarios where trust between  rms is low (Dongsheng & Yongan,
2008). These  ndings could guide planning agencies in designing intervention strategies
that allow public agencies to assume some of the risk in the development of technology,
either through seed-grants or partnering with private institutions seeking to bring useful
technologies to market. Similar guidance has been o ered by Canils & Van den Bosch
47
(2010), who noted that the partnership between higher education institutions (HEIs) and
regional government can help signi cantly reduce costs during the process of transferring
technology to private industry.
Ma & Nakamori (2005)describe this transfer process in more detail via a multi-agent
model intended to simulate the low-level mechanics of technological innovation. This model
is based on the theory that technological innovation could be viewed as an evolutionary
process using Kau man?s NK (hill-climbing) model to handle the mapping from the design
parameter space to the performance parameter space (Ma & Nakamori, 2005; Kau man,
1993). Through the interaction of this evolutionary process and environmental selection
criteria, the simulation model is able to inscribe the concept of technological innovation.
The environment is represented in the simulation model as a simple market network.
In the NK model, N represents the number of genes in a haploid chromosome and K
represents the number of linkages that each gene has to other genes in the same chromo-
some. The description of the NK model follows Altenberg?s (1994) more formalized de -
nition with several slight modi cations, namely that the genes are not binary-valued, but
rather Hi-valued. This means that an individual gene can take on values contained in set
Hi (i = 1;:::;n). This represents the notion that there exist multiple designs for a single
technology (such as engines, in the automobile example).
In the paper, two factors which prevent producer monopoly are identi ed through anal-
ysis of results obtained from simulation; consumers? incomplete information and diversity of
consumers? demand. The author?s also show mainly through intuition that industry actors
likely operated under sub-optimal technological solutions.
Berger (2001) proposes a spatial multi-agent programming model for assessing policy
options in the di usion of innovations and resource use changes. The model applies a multi-
agent cellular automata (MAS-CA) approach by developing and using several heterogeneous
farm-household models and capturing their social and spatial interactions explicitly. The
MAS-CA model consists of two main components: an economic sub-model and a hydrologic
48
sub-model, which are both tightly connected to a consistent spatial framework represented
via a bi-level toroidal lattice. A recursive linear programming approach is implemented
wherein each farm-household in the study area is captured in the model and solves its
decision problems (whether to adopt or reject innovation) over time autonomously. Results
demonstrate that structural change will occur as traditional actors (farmers) will experience
persuasive ?pull e ects? to leave the farm business as costs and other barriers increase.
Terna (2009) proposes an agent-based framework for examining how innovation and
new ideas di use among a community, or act to conserve the status quo, in an epidemic
fashion. The notion of contagion, or susceptibility to infection of novelty, is borrowed from
the studies of infectious disease and is modeled using the chameleon metaphor. While the
emergent structure of this model is primarily based on population density, its behavior was
show to vary dramatically if agents were allowed to evolve some intelligent behavior, such
as a-priori action planning. This evolutionary process is implemented using arti cial neural
networks supported by reinforcement learning techniques. The development of the structure
of the neural network relied on the Cross Target (CT) method which corrects both the
guesses about the actions to be done and those about the related consequences.
The chameleon metaphor is described as follows: in the starting phase a certain number
of chameleon actors are present in each of three colors: red, green and blue. When two
chameleons of di erent colors meet, they both change their color, assuming the third (non-
present) color. If all chameleons in the environment display the same color, a steady state
solution is reached. In the context of innovation, the author interprets the metaphor in the
following fashion: an agent di using innovation (or political ideas) can change itself through
the interaction with other agents. Conversely, an agent di using epidemics modi es the
others without changing itself. Additionally, each agent class? principle intent is to preserve
its own color, and attempts to implement di erent strategies to do so based on the reward
structure of the model. The simulation environment is modeled as a, N xN lattice, of which
agents are allowed a 3x3 local viewing area.
49
The reinforcement learning model was constructed using the Swarm Like Agent Protocol
in Python (SLAPP) tool and follows the methodological structure laid down by Sutton and
Barto (Sutton & Barto, 1998). In this structure S denotes the set of states related to an
environment, A denotes the set of possible actions, and R represents the set of scalar rewards
for given actions. At any time t an agent exists in state of st and can choose action a in set
A(St). After the action, the agent exists in state st+1with a reward given byrt+1 . Rewards
are summed over time with a discount factor. Through this transformation the agent is able
to map all the possible actions A in a state S to all the related rewards.
The author?s results indicate that agents who displayed aggressive strategies (red chameleons)
were most successful in preserving their color, implying that innovation di usion bene ts
from dynamic, rather than passive activity. This dovetails with earlier  ndings regarding
the importance of  rst-action in successful clustering.
Knig et al. (2010) demonstrate via empirically-based simulation on a micro-founded net-
work model that small-world networks could be used to model the search for collaboration
across a theoretic region. The environmental structure is a relational network of researchers
and patents based on Watts and Strogatz (1998) small-world networks and  ndings gath-
ered empirical surveys of regional collaboration networks. Analysis of the simulation was
conducted using the notion of small-world regimes and normalized clustering divided by
normalized path length for the largest connected component in each geographical region.
The author?s results indicate that that links between inventors are established mainly by
performing a local search and this e ect has become more prevalent over time.
Padgett et al. (2005) propose a hypercycle model of productive systems where goods  ow
through  rms as chemicals through reactors. Eventually, clusters form around production
loops, which are reminiscent of Marshallian industrial districts (Belussi, 2005). Interestingly,
economies with more than 4 goods require the existence of clusters in order to sustain
production.
50
Page & Tassier (2003) propose an agent-based model where local economies are super-
seded by chained stores. They observed that the formation of chains arose because they
exploit a niche that is pro table at several locations. Subsequently, wherever these chains
arise they homogenize the economic structure and beget other chains in (potentially) other
sectors. Thus, they observed the process to be a cumulative phenomenon. In addition, they
state that  nal con guration is likely to be suboptimal despite the fact that it is a local
optimum in the environment created by the existing chains. They surmise that decay may
be the fate of clusters that do not attain world-wide recognition in the globalized economy.
3.3.1 Behavioral models
Dawid et al. (2001) analyze the evolution of output decisions of adaptive  rms in
an environment of oligopolistic competition via an agent-based market simulation. The
model is composed of  rms which, at each time step, make a decision regarding whether to
produce one of several existing product variants or try to establish a new product variant
on the market. This decision is driven by aggregate local demand as well as each  rm?s
heterogeneous abilities to develop products and imitate existing designs. Certain estimation
rules such as market potential and market founding potential are used along with a measure
of stochastic decision-making. Further, the characteristics of the  rms are allowed to evolve
via \learning by doing" e ects.
After sensitivity analysis the simulation model was tested by running the simulation
over 64 batches of 100 runs, followed by an additional 36 runs with tuned parameters. The
results indicate that the abilities for technological improvement are the dominant factor for
the determination of a good innovation strategy. Moreover,  rms should choose optimal
strategies on their production e ciency relative to the market?s need.
Agent-based Nelson-Winter models (NW) have seen limited use with respect to com-
munication mechanisms. A basic NW model is outlined in  gure 4.1 and is more fully
described in Anderson (2001). Zhang (2003) proposed a NW model which was paired with
51
Figure 3.1: Structure diagram of the standard Nelson{Winter model of industrial dynamics.
The diagram puts an emphasis on a particular  rm, while aggregates are placed in the last
row. An arrow from x to y should be read ?x codetermines y? (Andersen, 2001).
an explicitly de ned (physical) landscape to study the formation of high-tech industrial clus-
ters. The extended model focuses on social, rather than economic dynamics in site selection
and innovation activity between actors and uses cluster metrics to evaluate agent con gura-
tion and performance. The problem is de ned in terms of Schumpeterian entrepreneurship
(Schumpeter, 1934), where economic development is assumed to driven by the activities of a
certain class of highly motivated creators. The author?s results indicate that entrepreneurial
clustering occurs even without the bene t of industrial clusters, such as knowledge spillover.
Moreover, the  ndings reinforce the importance of pioneering actors in the establishment of
industrial clusters.
Spatially situated ecosystem models
Case studies involving the application of agent-based models to ecological domains have
become more common now that simulation toolkits have become more user-friendly to non-
technical researchers. These works have yielded interesting results which incorporate mech-
anistic themes such as resource exploitation, trophic interaction (Schernewski & Neumann,
52
2005), spatial con guration (Lansing & Kremer, 1994; Jannsen et al., 2000), and energy
transfer processes (Williams et al, 2001) into the multi-agent paradigm.
These simulation models also provide capture the emergent behaviors which drive ecosys-
tem development. The importance of properties such as population mobility (ability for
agents to relocate in search of better conditions) and density (agent clustering) were ex-
plored by Bosquet et al. (2002), Rouchier et al. (2001), Ho mann et al. (2002), and many
others. These properties were found to be directly correlated with behaviors such as resource
exploitation as well as the di usion of new cooperative and competitive behaviors. This is
by no means surprising, as prior work using numerical simulation produced similar results
(Mertens & Lambin, 1997).
Of equal importance is the ability of ecosystem simulation to explicitly represent hu-
man decision making paradigms. In this context, decision-making refers to the ability of an
agent to decide its behavior at any point in time. Agents are able to apply psycho-social
knowledge of actual decision-making to agent design that may contrast with the rational
homo-economicus of classical economics. This feature is highly relevant with regard to in-
novation di usion studies, such as those conducted by Balmann et al (2002) and Berger
(2001). These studies pro er model the proliferation of new agricultural and land-use prac-
tices through networks of farmsteads. Albino et al. (2004) proposes an agent-based sim-
ulation model to investigate how innovation processes in industrial districts (ID) have to
be modi ed to assure their survival in a highly competitive environment. The environment
is modeled as a dynamic network where linkages between agents signi ed relational com-
mitments between various classes of supply and transformational agents. The model was
analyzed after being run on four experimental scenarios which varied between low and high
demand & and innovativeness parameters. The results indicate that the number of  rms
within an ID is a proxy of success in a particular competitive context.
These works, and others, show that ecosystem simulation brings forward powerful fea-
tures such as the ability to capture complex population dynamics, leveraging of individual
53
and group decision-making, systemic resilience, and layered spatial (and non-spatial) con-
texts.
Otter et al. (2001) developed an agent-based model where  rms and households decide
where to locate within a multi-layered toroidal N xN lattice according to availability of labor,
services, natural resources and recreation areas. The model has two principle agent types,
households and  rms. The latter of which distinguishes between enterprises operating in
heavy industries, in light manufacturing and in services. Both  rms and households have
imperfect information and can only observe other agents that are within a visibility range.
The environmental model is composed of a ?land-use? and ?attraction? layer. The each
cell in the land-use layer has  xed attributes such as ?land?, ?natural area?, and ?sea?. The
initial grid is con gured such that no agents are present and each cell registers as empty.
Agents are only allowed to locate in empty cells and make their relocation decisions based
on the values in the underlying attraction layer. These values are de ned by the number
of agents and  rms present in nearby cells. This interplay is simulates agglomeration phe-
nomena and is similar in execution to the concept of pheromone deposition in ant-colony
optimization strategies.
The authors observed the emergence of clusters of  rms and households of various size
and composition depending on exogenous parameters as well as the initial con guration.
Visibility range was shown to be a large determinant in both the size of clusters as well as
the time needed to reach equilibrium.
3.3.2 Explicit NSI Simulation
Dongsheng & Yongan (2008) propose an agent-based model for studying the evolution
of University-Industry Cooperative Innovation (UICI). Three agent types are identi ed and
modeled following the UICI metaphor; enterprise, university, and government. Enterprise
agents attempt to cooperate with university agents to produce technologies, which they then
leverage to earn pro t. The government agent taxes the earnings of enterprise agents and
54
uses these taxes to fund university agents who promote new research initiatives. The en-
vironment consists of a 45 x 45 lattice, within which both university and enterprise agents
reside. Universities are only allowed to partner with neighboring enterprises. Both univer-
sity and enterprise agents are allowed to migrate if local conditions are not optimal (low
cooperation rate). In order to evaluate the evolution of innovation between these agents,
the authors introduce two variables, named as Success Rate (SR) and Average Con dence
(CA) respectively. The SR is the ratio of successful cooperation to all cooperation attempts.
The CA is the arithmetic mean of the con dences of every enterprise to every university.
The results indicate that when the Expected Pro t (EP) for university agents is su ciently,
a collapse in trust and cooperation success inevitably results. This implies that strategies
which promote more modest pro t expectations lead to stable long-term partnership between
university and enterprise  rms.
Butel & Watkins (2006) propose a model for identifying cluster dynamics for entrepreneurs
using an ant-colony optimization strategy of indirect communication. The environment con-
sists of a 50 x 50 lattice, with two types of agents; resources and researchers. The research
agents are initialized with no prior knowledge regarding the location of resources of interest
and are allowed to operate for a limited number of turns. After each run, successful agents
(those that discovered a resource patch) are allowed to increase pheromone along their search
path. The results showed that after a training period, clustering occurred quickly, even when
unskilled agents were added to the population or the individual search space was reduced.
Moreover, agents spent less time searching unpro table areas. This indicates that under this
methodology, the activity of successful agent clusters can negate much of the waste (in terms
of unproductive activity) of untrained agents.
3.4 Chapter Summary
This chapter has reviewed concepts relating to the development of our model of the
agent-based innovation ecosystems. In section 3.1 we outlined an agent-based approachto
55
ecological and followed with more speci c submodels and mechanisms in agnt-ecosystems
(section 3.2). Many of the previous agent-based innovation studies discussed in section
3.4 are brie y summarized in table 3.2. The "knowledge-based" column states whether
the algorithm is embedded or not with a knowledge-based (e.g. rule-based, etc.) system for
selecting a path from the solution set. The last column shows the capability of the algorithm
of producing multiple paths in parallel.
We can make several key observations from the algorithms and literature cited above.
First, while statistical models of innovation can capture general trends, the behavior centric-
ity of agent-based models has been used to gain insight into the complex forces which drive
the creation of knowledge market places. Second, we observe that social factors can play a
large role in governing the formation and success of innovation economies. Third, we observe
that network and spatial models display di erent relational characteristics between agents
and o er a ordances for di erent types of analysis. This also survey obviates the fact that
very few models have been developed which combine these varied approaches, namely using
a market-based interaction strategy paired with relational networks in a spatial environment.
Due to complexity and hardware limitation issues the computational e ectiveness of these
approaches varies and in the real world they require many hours of computational time to
achieve a satisfactory solution.
In light of the above observations it is notable that few approaches to this problem
take advantage of the inherent power of swarm systems to address issues of spatial search.
Some simulation models use used classical multi-objective techniques, such as the weighted
sum approach or a single-objective function to obtain locally optimum solution, while others
leverage neural networks or genetic algorithms. Since there is more than one optimum
solution for this sort of problem, many of these approaches are not capable of generating
these con icting optimum solutions (or trade-o solutions). Those that are capable of doing
so must operate under severe constraints.
56
Table 3.2 Previous works on simulating innovation economies
57
Chapter 4
THESIM: Triple-Helix Ecosystem Simulation
4.1 Triple Helix Ecosystem Interaction Model
Ecological modeling provides a novel framework for representing the relationships be-
tween the regimes of the innovation economy. It represents the system as an ecosystem and
utilizes the concept of energy  ux as a mechanism to capture the transport and transforma-
tion of innovation products through the system. This agent-based ecological view exposes
the relational components of real world innovation systems, and allows individuals within
various communities to be included as explicit actors in the model instead of having their
collective activity represented deterministically (Grabher & Powell, 2005).
To study such interactions between industry, university, and research  rms on the per-
formance of triple-helix aligned NSI through computer simulation, a model of the simplest
NSI in which the three control regimes have no links to one another was constructed.
This simple NSI model (Figure 4.1) consists of two of the three helices of control,
industry, university, and government. In the classic triple helix model described by Etzkowitz
& Leydesdor (1997) the government is described as a separately partitioned entity, similar
to industry and university. However, in keeping with the ecological view of the system we
may also view the government as an actor as well as controlling the environment within
which the other helices operate.
After establishing the simpli ed model of the triple helix NSI, we need to add the inter-
actions among the three regimes. Interpreted from the aforementioned literature, in a triple
helix system it is the government?s duty to drive industry and university  rms to innovate
through means such as lawmaking and  scal measures. Universities provide human R&D
58
Figure 4.1: Model of a simple NSI.
Figure 4.2: Triple-helix ecosystem interaction model.
59
resources, basic research, as well as technical support to industries engaged in product de-
velopment (Ja e & Trajtenberg, 1989). The pro ts generated by industrial  rms are used
to provide  scal support to universities for R&D e orts (Nelson, 1994). Government plays a
mediation role between the other regimes by providing funds in support of research, negoti-
ating university-industry partnerships, and imposing control mechanisms on entrepreneurial
 rms. These relationships form the core of the triple-helix simulation and are shown in
Figure 4.2.
Given the mapping between ecological and economic concepts we explored in chapter 3,
we observe that co-evolution between the control regimes and their institutional environments
change the knowledge infrastructure in unpredictable ways (Uyarra, 2009). Government and
industry regimes aim to maximize revenue, while minimizing cost, while the university (or
research) regime aims to capture as much funding as possible in order to support large popu-
lations of researchers and create new technologies. Public administrators and planners must
simultaneously balance these competing agendas to avert economic stagnation or collapse.
Policies provide a rule set for governance which is enacted through mechanisms avail-
able to the government. Those policies which drastically alter resource provisions, such as
increasing or decreasing funding for basic research or altering return on industrial technolog-
ical investment through tax policies often have mixed or counterproductive results (Cooke
& Morgan, 1998). This relationship is also observable in biological ecosystems, where land
management decisions often lead to unintended (and often undesirable) environmental per-
turbations. In a climate of economic uncertainty where  scal supply is constrained, policy
makers must  nd other ways to change system behavior to increase productivity. Using eco-
logical modeling to capture the behavioral relationships that modify resource transfer may
help obviate the hidden causes of con ict created by such decisions.
Following this basic framework we are able to construct a generalized simulation of the
triple helix aligned with NSI. However, more speci city is needed in order to derive speci c
agent rules necessary to create a complete ABEM model.
60
4.2 Conceptual Model
In this section we outline a conceptual model for the purpose of de ning an agent-based
ecological view of innovation economics. The conceptual model is composed of the set of
ontologies describing the actors of the system (researchers, universities, etc.), the objects
they are acting on (resource patches), the actions carried out by the actors on the objects
(foraging, transforming, selling, etc.) and the regulations imposed on these behaviors (tax,
waste, etc.).
In our model, an actor may be a member of a large number of formal (university sys-
tems, industries, governments, etc.) or informal (research community) institutions. Each
institution de nes the functioning of its constituents, including its ontologies and norms (Os-
trom, 1990). An actor may also be situated geographically within a set of areas that can be
embedded within one another. However, we only concern ourselves with areas upon which
regulations can be applied, thus implying that formal and informal institutions control these
areas. Actors, by virtue of being environmentally situated, are subject to the regulation
of these institutions. Thus, an actor is permanently subject to these formal and informal
regulations which can contradict one another.
There are two kinds of actors: the individual actors and the collective actors. In our
model, individual actors are the researchers, who exhibit communal foraging behavior and
are identi ed by their a liation to a particular university. These actors take on the role
of harvesters, foraging in the environment for resources which are returned to their univer-
sity. The act of exploration and exploitation of environmentally situated resources (grants)
represents the  rst stage of ecological energy  ux. Collective actors include the universities,
entrepreneurial  rms representing the industry, and the regional government. Each inter-
acts with the preceding stage by bartering for resources and converting that resource into a
higher order product for sale to actors in later stages. Within this model grants take on the
role of producers; researchers, universities, and industry actors behave as consumers; while
government actors exhibit the characteristics of a decomposer, returning resources back to
61
Figure 4.3: Conceptual model of the ?innovation food web?
Figure 4.4: Basic resource cycle
the environment. The relationship between the environmental features and various actor
classi cations forms a resource web, which is displayed in Figure 4.3.
In multi-agent systems, an institution is de ned in terms of a set of roles together with
the speci cation of norms, which govern the environmental and social behavior for each
role. Using the norms for regulating the interactions with environment as well as other
situated agents is a natural extension when dealing with resources management. These
norms also de ne which objects count as a good, a product, etc. Therefore, the norms also
de ne the categories (or roles) in which the objects can be classi ed (Pottage et al., 2004).
Thus, having de ned our relational model, we now concern ourselves with the behaviors
and norms (regulations) which govern the behavior of our institutions by de ning particular
actor classi cations.
62
Agents: Each agent belongs to a particular stage of the triple helix, collects or buys
resources of a particular type, transforms and sells the resource, or interacts with the environ-
ment through directed activity. The  rst stage (S1) is the harvester agent which collects the
raw environmentally situated resource (currency, in the case of our metaphor). At this stage
there is no transformation. At the second stage (S2) the agent pays a liated S1 agents and
collects their resources, partially converting it to knowledge product (second-order resource)
which it sells to third stage agents. The third stage agent (S3) buys knowledge product
from S2 agents, converting into currency and innovation product (third-order resource) and
currency which is taxed by fourth stage agents. This fourth stage agent collects a tax on
all sales as well as monitors the creation of technologies. It receives an imposed quantity of
money per round which it disperses back to the environment in the form of grants. This
cycle is illustrated in  gure 4.4.
4.3 Behavioral Model
Cooperation between universities and industries is one of the most important forms
of innovation activity. Besides the two principle actors (universities and enterprise con-
glomerates, or industries), innovation involves many other agents, such as venture capital,
individual researchers, and of course government agents. In the most common model, little
attention is paid to the granular actions of constituent agents of the Triple Helix entities
(university, industry, and government). It is more common to model their interactions in the
form of direct relationships, where super-classi cations of entities act on or in concert with
one-another to approximate group behaviors. In this model we extend this paradigm to in-
clude several lower-order agents (researchers and grants) with the intention of demonstrating
the e ect of distributed social learning on the behavior of the system.
To facilitate this analysis, we modify the hypotheses presented by Dongsheng & Yongan
(2008) to include these social activities:
63
Figure 4.5: Relational Map of agents in the Triple Helix ecology.
64
 Government actors have no direct connection to universities. Funding for research must
be derived from the environment via the collection and distribution of grant resources.
 The actions of researcher agents (grant-seeking) within the environment approximate
the function of discovering funding opportunities, making grant submissions, and col-
laborating with other researchers.
 Each university manages its own distinct population of grant-seeking agents by con-
trolling their population and paying for their upkeep.
 When an agent discovers a grant resource, it makes an attempt to acquire it, and the
result is stochastically driven. If successful, it collects funds and then deposits them
with its parent university.
 Grant-seeking agents use exogenous communication to facilitate learning and discovery.
 Cooperative innovation is always initiated by industry-level actors. Universities decide
to accept or reject these requests. The total number of requests allowed per round is
limited to 1.
 Each cooperative request involves only one university and one industry partner, thus
there is no direct cooperation between separate industries.
 Each industry has a con dence value for every university. After each cooperative
innovation, the value can be changed according to the result of the venture, and the
changed value will a ect the future cooperation of the agents.
Based on the above analysis and hypotheses on the model, we can summarize the behavior of
each agent within the triple-helix cooperative innovation framework. Grant-seeking agents
search the environment for funding opportunities and return them to their parent universities.
University agents generate technologies, and handle a constant stream of requests as industry
agents develop. Industries expect to cooperate with universities rather than innovate by
65
themselves, as this is a dominant strategy for growth. Industries search the local environment
for universities to cooperate with. After  nding a suitable candidate, the industry actor will
propagate a cooperation request to the selected university, according to determinates such
as the R&D cost of the technology and con dence (record of past cooperation) between
the two agents. The university decides to accept or reject based on its own bene t. If the
request is accepted, the technology will be transferred to the industry. Otherwise, no transfer
occurs and the industry continues searching for another university to partner with. If the
industry fails to  nd a partner, it will innovate by itself. During the course of innovation,
the government involves itself via taxation of industry pro ts and by the maintenance and
replenishment of resources within the environment.
Following this structure we proceed to more fully de ne the rules which govern agent
behavior.
4.3.1 Agent Behavior Rules
The behavior rules of agents of di erent types are as follows:
1. Rules of grant agents.
(a) Movement: At the end of each round, grants may migrate to a new (unoccupied)
position within the environment. The rules of migration are set by the government
agent at the start of the simulation run. Unused grants use random (8-way)
movement, while grants which have been utilized will move in the general direction
of the university which visited it most often. This creates gravity around well
optimized universities.
(b) Signaling: Grants signal their position in the local environment by broadcasting
their position to other agents within a local region initialized by the system. When
a grant?s resource supply is depleted, it will diminish its local signals to divert
agents from searching in its local area.
66
Figure 4.6: The legend of environment (a), agent?s movement directions (b), coding sequence
for local perception (c), example of local perception (d)
(c) Supply: Grants which see high tra c will make requests to the government to
increase funding, whereas grants which see little tra c are penalized by having
their total funding cut in future rounds and distributed to other sources.
1. Rules of research agents.
(a) Movement: Research agents may make a single (8-way) move each turn to traverse
their local environment. Agents are allowed only a  xed number of moves before
they must return to their parent university.
(b) Signaling: Researchers communicate through environmental signals based on the
deposition of a virtual pheromone over the path it follows, marking a trail. This
pheromone trail is enhanced by repeat traversals but diminishes at a proportional
rate. Agents may use this signal as a determinant in movement decisions or in
detecting resources.
(c) Innovation: Research agents are paid by their parent university, and in turn con-
tribute to the development of technologies related to their  eld of specialization.
67
(d) Resource Collection: If physically adjacent in the environment, researchers may
attack grant agents. The success rate of the attack is determined by the di er-
ence in specialization  eld between the grant and the researcher. The larger the
di erence, the lower the possibility of success. If the agent successfully collects
resources from a grant which di ers from its base specialization, the researcher
has a small chance to change its base research classi cation to match that of the
grant which it consumed.
1. Rules of university agents.
(a) Management: The University determines the number of research agents it can
support by analyzing its current stored value, income, and expenses. If its pool of
researchers is too small, it hires new researchers. The specialization type of these
new researchers is determined by proportional probability based on the amount
of funding the university has collected from each specialty area.
(b) Innovation: Universities invest money in R&D activities each period. The amount
of growth of each technology being researched is determined by the research level
of the university. Higher level universities research at a faster rate than those of
lower level.
(c) Starting new innovation: At every round, the university creates a certain number
of research projects according to its research level and current funding. The
parameters of these functions are initiated by the current research level of the
university, the distribution of grant funding between the various specialization
 elds, and random functions.
(d) Dealing with requests: When a university receives cooperation request from an
industry, the university decides whether to accept it or not according to its ex-
pectation of pro t (Ep). If accepted, it transfers the technology at the price
proposed by the enterprise. Otherwise the university sends a notice of rejection.
68
1. Rules of industry agents.
(a) Updating technology: Industries submit multiple technology requests each round.
The number of requests is governed by the value ( nancial scale) of each industry.
Generally speaking, the bigger the scale of an industry, the more cooperation
requests it promotes.
(b) Innovation: Industries invest money in R&D activities at every period. The
success rate of R&D is determined by the research level of the industry. Higher
level industries research at a faster rate than those of lower level.
(c) Production: Industries produce and gain pro t by monetizing the  nished tech-
nologies they have, including both those innovated by themselves and purchased
from universities. After tax is removed, the remaining pro t is added to the
industry?s total money in order to support future purchases and development.
(d) Searching for Partners: When an industry has a technology request, it searches
the environment for university partners. The industry then selects the university
with the highest con dence value and makes a request for technology. If the
university has a suitable technology, the industry agent calculates the transaction
cost according to the di culty of the technology, R&D level, con dence between
the two parties, and random functions. If the promoted request is denied, the
industry continuse to search. If no suitable partner is found, it conducts the
innovation by itself.
1. Rules of government agent.
(a) Grant management: At the end of each round, the government evaluates the
usage of each grant in the environment and may modify the grant?s maximum
funding amount, location in the environment, type, and replenishment amount.
These decisions are determined by the current goal, strategy, and parameters set
at the beginning of the simulation.
69
(b) Taking tax: At the end of each round, the government examines the pro ts of
each industry and takes a corresponding amount of tax. The relative tax rate is
speci ed at the beginning of the simulation.
(c) Setting innovation goals: Periodically, the government may set innovation targets
in the form of \Create N new technologies in  eld M".
(d) Implementing innovation strategy: During the grant management phase, the gov-
ernment may apply a strategy in concert with its current innovation goal. Strate-
gies may include modifying industry tax rates, or changing the distribution and
type of resources present in the environment.
4.4 Overview and Design Concepts of THESIM
A simulation model was developed based on a mapping of the triple helix theory of
public-private partnership using the ABEM framework. The resultant conceptual model is
described using theOverview, Design concepts, and Details protocol (ODD) proposed by
Grimm et al. (2006). This protocol is aimed at facilitating the description of individual-
based models in ecology, with attention paid toward its application in agent-based social
simulation and other disciplines. The details of this model and of the core structure of the
innovation economy theory upon which it is based are presented in the proceeding sections.
4.4.1 Purpose
THESIM (Triple Helix Ecosystem SIMulation) is a multi-context agent-based model
designed to explore the e ects of positive and negative government spatial externalities
and psycho-social outcomes. Spatial externalities refer to exploitative land-use activities
by university and industrial  rms which change the pay-o of other  rms in an abstract
environment modeled after on a regional government-grant system. THESIM is an abstract
model designed for theoretical exploration and hypothesis testing. Speci cally, THESIM was
designed to extend existing analytical innovation economy theory to examine relationships
70
Table 4.1 Grant patch agent model
between externalities, social behavior, and the e ciency of simple government intervention
policies.
The THESIM model functions as an observational tool for identifying the parameter
spaces and conditions under which an innovation economy is resilient to environmental per-
turbation. The policy mechanisms and agent behavior in the model mimic that of the three
helices identi ed in the Mode 2 innovation model (Gibbons et al., 1994), with a technol-
ogy market model based on the work of Dan et al. (2009) where  rms over and under
bid for technology products based on trust and risk aversion. The magnitude of these
over-and-undershoots diminishes as the model reaches equilibrium but can be disrupted by
perturbations in state of each helix.
4.4.2 State Variables and Scale
The THESIM Model consists of the following core entities, which are described in the
following set of  gures (table 4.1 through table 4.5). A run is composed of multiple rounds
with each round corresponding to 30 time steps. A single time step in the model is non-
analogous to any real-life time scale. Simulations typically run 200 rounds.
4.4.3 Process Overview and Scheduling
A high-level diagram showing an overview of the schedule in THESIM is given in Figure
4. More detail on each section in the diagram is provided afterwards. THESIM considers
71
Table 4.2 Researcher agent model
Table 4.3 University agent model
Table 4.4 Industry agent model
72
Table 4.5 Government agent model
several main processes in a loop. At the beginning of each step, each agent evaluates its
state and then chooses a set of actions based on that state. At the end of each round, each
agent performs a diagnostic and may change certain operational parameters as a result of
an optimization process. Researcher agents seek to maximize Success, university & industry
agents seek to maximize income while minimizing cost, and the government agent seeks to
maximize GDP while minimizing waste.
Figure 4.5 illustrates the processes of THESIM at an aggregate level. Exogenous ele-
ments (initialization parameters and environmental conditions set by the user) combine with
the market model to form various decision cycles within each context. Each agent then se-
lects the highest-valued decision, and additional derived endogenous elements are calculated
based on these choices. Resulting changes in social network and landscape patterns then
initialize the next time step.
THESIM uses a  xed event-scheduling mechanism, which allows every agent to make an
activity decision in each time step (usually consisting of a single move and a resultant action,
where possible). This mechanism prevents oscillation, while introducing a minimal amount
of additional path-dependency due to stochasticity. Agents are processed sequentially, but
updating is synchronous, after each active agent has reached the end of its activity loop.
73
4.4.4 Design Concepts
Emergence: THESIM was designed to explore the relationship between two related
emergent phenomena: grant clustering and public-private con dence driven market dynam-
ics. The funding landscape patterns and associated market productivity measures are emer-
gent in the sense that they are the result of the decentralized decisions of distinct populations
of autonomous agents.
Adaptation: THESIM models a speci c kind of ecological adaptation known as en-
gineered ecological resilience. Each of the individual agent classes modify their endogenous
parameters based on their perception of the state of the world. At the macro-level this can be
observed in  uctuations in populations, productivity, trade, and social adherence phenom-
ena. Extensions to this model could include more explicit machine learning techniques such
as hypothesis generation via genetic algorithms, or a more robust reinforcement learning
framework for the decision-making processes governing each agent.
Fitness: At the individual agent-level,  tness is measured di erently across role and
context. Agents operating within the market context measure  tness by pro tability, while
those operating in the spatial context use utility. Global  tness is measured through economic
welfare statistics (GDP, total surplus, total con dence, and total population), which depend
on the degree of resource utilization and market dynamics (price, trust, etc.) as described
in previous sections.
Prediction: In the spatial context THESIM agents are generally reactive and do very
little to predict the outcome of future events. However, within the market context participat-
ing agents predict future outcomes in very simple ways. University agents form an Expected
Value for completed technology patents based on sunk costs, maturity level, novelty, and
the anticipated  tness (pro t) maximizing decisions of Industry agents. In this calculation
they assume to know the pro t calculations of all other agents. However, this knowledge
is incomplete and does not take into account ?hidden? metrics such as expected pro t and
con dence. In this sense they are much more myopic than real-world agents are likely to be.
74
Interaction: Agents interact indirectly at both a spatial (local) and market (global)
level. At a local level, the resource-use choice of one agent a ects the behavior and income of
its neighbors. At a global level, the agent?s choice a ects pricing, trust, and competiveness
of its partners and competitors.
Sensing: Di erent classes of agent possess di erent collections of spatial and relational
information. Grant agents only possess knowledge about the 8 neighboring spaces in their
local environment. Researcher agents have the same viewing area, but can pass information
to one another through environment via stygmergic processes. University agents can only see
their space (similar to grant agents), but have a market view that is complete. Government
agents are assumed to know the current market network and resource landscape pattern,
as well as the positions and endogenous parameters of all other agents. In this instance,
information is complete and certain.
Stochasticity: THESIM is subject to a degree of stochasticity that is prevalent across
many multi-context ecosystem models. Despite this, e orts were made to minimize path
dependence, and the main source of stochasticity is the ability of the end user to generate
an initial random input parameters and resource-landscape.
Observation: Observations include graphical display of environmental patterns (spa-
tial and network), metrics measuring the fragmentation and productivity characteristics of
the resource- landscape, and metrics that re ect economic outputs (outputs from each uni-
versity &  rm, technology prices, tax revenue, and multiple economic  tness measures).
4.4.5 Details
Initialization: The state of the model world is initialized by setting an initial parameter
con guration, including a prede ned spatial model.
75
Figure 4.7: University-Industry-Government relations in a three-dimensional space.
Input: Following initialization, environmental conditions remain constant over space
and time in the THESIM model. There is no imposed or modeled spatial or temporal het-
erogeneity, aside from the endogenous variations in market productivity, landscape pattern,
and agent choices that are the focus of the model.
4.4.6 Submodels
In developing the simulation model, it is critical to identify metrics through which
we can evaluate global behavior. To this end we focus on the various intersecting ar-
eas of cooperation in order to more fully understand the output of the entire simula-
tion; university-government (UG), university-industry (UI), industry-government (IG), and
university-industry-government (UIG). We examine grant clustering as a measure of the de-
gree of UG trust via spatial occupancy of a local Von Neumann neighborhood. We de ne
technology transfer capacity and rates as measures of UI partnership. Individual tax output
serves to identify IG success, and we de ne a global productivity score based on production
analysis.
Innovation Production: The creation of technology in both universities and indus-
trial  rms are modeled as learning functions. Knowledge artifacts (technology patents) are
76
created at the beginning of each round and incremented until they reach completion. CN is
the cost of incrementing technology N at time t. The equation for is;
CN = ComplexityNR
level
 Rcost Eq. 4.1
Where ComplexityN is the complexity order for technology N, Rlevelis the research level
of the institution, andRcost is a constant representing the operational cost of innovation.
Innovation Value: The valuation of technology after completion is represented by a
Gaussian function:
f (V ) = Novelty e (v b)
2
2d2 Eq. 4.2
Where Novelty represents the age and ?freshness? of a technology, brepresents innovative
maturity, and dis a constant representing di usion time of new ideas. Under this submodel,
the competitive advantage o ered by newly completed technologies is low immediately after
adoption, slowly rises in value as their application becomes more well-understood by the
 rm, then diminishes as competitors become aware of it.
We de ne return on assets (ROA) as the ratio of after-tax income mode of all available
assets, as a percentage. This value is given by the following equation;
ROA = NOP ATModeofTotalASsets Eq. 4.3
We use ROA to determine the value of technology available to industrial  rms for
monetization. This value is held indivudally by each  rm and contributes to the market
model. When RoA is low,  rms are pressured to increase their value by acquiring more
technologies.
Acqusisition Process: To understand the dynamics of university-industry (UI) part-
nerships, it is critical to derive metrics which go beyond trade volume. One such method is
to analyze industrial investment in technology and the decisions & processes which drive it.
The transfer of technology from research to industry provides us a useful window through
which we can quantify the state of the system at a given point in its development.
Transaction Cost Theory states that industries generally incur the cost of communica-
tion, negotiation, coordinating, monitoring and enforcing the contract. The communication
77
cost (C) is de ned as the temporal and  nancial spending on searching for cooperative part-
ners. However, due to the fact that industries in this model are not geographically situated
and instead can operate across the entire environment, the communication cost is represented
by a weighted random function (Vij), which represents the non- xed cost of communication
between industry i and university j. The cost of negotiating and enforcing is enacted because
the partners must pay some premium to avoid the risk in the cooperation. In this analysis,
we associate the cost with the con dence between enterprise and university (Cij), which
takes a value between 0 and 1, where 0  Cij  1. The larger the con dence value, the
lower the cost of negotiating and enforcement that must be paid by the two agents. More-
over, there will still be some uncertain environmental factors that will in uence the value of
transaction costs. These are represented by a random function (fenv). The transaction cost
can be expressed as a function of three variables, as shown:
T = f (Vij;Cij;fenv) Eq. 4.4
In order to evaluate the evolution of innovation between these agents, two variables are
introduced; Success Rate (SR) and Average Con dence (CA) respectively. The SR is the
ratio of times of successful cooperation to total number of cooperation attempts. The CA is
the arithmetic mean of the con dence values of every industry to every university.
SR = TSuccess)T
Success+Tfailure
CA =
PNE
i=1
 PN
i
j=1 Cij
Ni
!
NE Eq. 4.5
Where Ni represents the total count of universities that industry i can cooperate with.
Cij represents the con dence of industry i to university j.
Transactional Codependence (Clustering): Technology-led economic development
has historically been driven by major research universities, often in partnership with gov-
ernment either through co-development or targeted grant support. This relationship is the
bedrock of university-government partnerships By analyzing the growth patterns of grant
clusters around universities we can make some analogies to the formation and maintenance of
78
Figure 4.8: Von Neumann Neighborhood for varying values of r.
research communities through partnership with the government via direct funding dispersal
(grants). Since the targetting of funds through cluster initiatives and emerging technology
funds are accepted and well wrought partnership strategies between universities and fed-
eral/state government, this particular metric allows us to quantify the degree of cooperation
between those classes of actors in our model.
We represent a cluster as the number of cells occupied by grants in a  xed Von Neumann
neighborhood around each university site. The equation for this neighborhood is given as;
Nvx0;y0 =f(x;y) :jx x0j+jy y0j rg Eq. 4.6
The number of cells in this range S is given as;
Stotal = 2r(r + 1) + 1 Eq. 4.7
To determine the cluster value (K) of a given university we calculate the ratio of occupied
to cells to the total number of cells in the neighborhood.
Kn = SoccupiedS
total
Eq. 4.8
Thus, the cluster value for the entire model is given as the non-linearized average of
K for all universities in the simulation, where N represents all universities, following the
formulation model provided by Watts & Strogatz.
 K =
PNi
i=1 Kn
Ni Eq. 4.9
79
Government Taxation: Generally, we can evaluate the throughput of the system
by measuring the downstream returns on initial government investment by following the
growth patterns of the industry?s taxable income. Speci cally, we de ne return as the net
tax extracted from each industry at round n. The marginal tax rate is given by the constant
r, which takes a value between 0.1 and 0.5, 0:1 r 0:5. The total tax (W) is given by the
following equation;
W = PNii=1 incomen r Eq. 4.10
Productivity Heuristic: A  nal metric for validating the entire relationship between
universities, industry, and government actors (UIG) can be found the accumulation of earn-
ings from the sale of  nal products derived from technologies transferred between university
and industry partners. To this end we use the common Gross Domestic Product formulation
given as;
GDP (Y ) = p:cons: + gross:inv: + g:spend: Eq. 4.11
or
Y = P + I + G
In our model, private consumption (P) is assumed to be the value of all derivative sales
made by each industry actor i at the end of round n. Gross investment is the total amount
spent by all industries to acquire technology artifacts from university partners during each
round. Government spending is the  xed cost of investment per round. It is calculated by
summing the value of all replenishment to the grant pool. Sector contributions to GDP
are also tracked in order to determine the a ect of perturbations to the population of grant
objects in the environment.
Diversity Index: One of the more interesting corrolaries to our selection of the ecosys-
tem view of the system is that we can investigate the impact of species diversity on overall
system behavior. In order to quantify diversity it is often useful de ne a measure which
goes beyond species richness (the number of distinct species in a particular ecosystem). The
relative abundance of species gives the investigator an indication of the relationship between
80
environmental conditions and particular ecological con gurations. The relative abundance
of rare and common species is refered to as evenness. Species diversity takes into account
both richness and evenness, where communities with large numbers of species that are evenly
distributed are the most diverse, while those with few species dominated by one species are
the least diverse. Various techniques exist to measure species diversity. These techniques are
commonly referred to as diversity indices. In the ecological domain, two of the most common
diversity indices are Simpson?s Index and the Shannon-Weiner Index. For the purpose of
this study, we will use the Shannon-Weiner Index as it is more ideal for cases where a greater
deal of empirical data is present.
The Shannon-Weiner index was developed from information theory and is based on
measuring uncertainty. The degree of uncertainty of predicting the species of a random
sample is related to the diversity of a community. If a community is dominated by one
species (low diversity), the uncertainty of prediction is low; a randomly-sampled species is
most likely going to be the dominant species. However, if diversity is high, uncertainty is
high. In our case, we have perfect information regarding the species sample, so we can make
stronger claims about the accuracy of the product.
Calculating the Shannon-Weiner Index is straight forward. The index of diversity H0 is
given by the equation;
H0 = N lnN 
Pn
i lnni
N Eq. 4.12
where N is the total number of individuals of all species andniis the number of individ-
uals of species i. While not part of our basic hypothesis, there is strong correlation between
commercial and academic diversity and long-term RSI resilience in the face of perturbation.
Thusly, we intend to use this index as a part of our measure of ecological resilience following
the conclusions of Holling (1973).
81
4.5 Chapter Summary
In this chapter we presented a conceptual framework and formal model for THESIM, an
agent-based simulation model of the triple-helix theory of regional and national innovation
system organization. In section 4.1 we described the interaction philosophy and model of
the three control regimes of the triple helix; academia (university), private  rms (industry),
and government. In section 4.2 we laid out our agent-based conceptual model of the three
helices, by de ning categories of actors and establishing the resource cycle. In section 4.3
we more fully described agent behaviors and classi ed them into speci c roles based on their
ecological characteristics (producer, consumer, decomposer, etc.). These speci c roles were
then organized in a relational map (Figure 4.3) and a simpli ed rule set was developed based
on their expected behavior. In section 4.4 we formally propose a simulation model based on
the conceptual framework laid out in section 3 using a modi ed Overview, Design concepts,
and Details protocol (ODD) similar to the methodology proposed by Grimm (2006). Using
ODD we formally present the criteria of parameter selection, behavioral guidelines, and sub-
models from which THESIM will be developed. These submodels will provide the statistical
output for use during experimental veri cation and validation.
82
Chapter 5
Model Implementation & Basic Relationships
5.1 THESIM-Repast
In order to test the feasibility of the simulation model proposed in chapter 4 we im-
plemented THESIM using the current version of the Recursive Porous Agent Simulation
Toolkit, Repast Simphony 2.0. Repast Simphony is the successor of Repast. We selected
this toolkit over other similar packages (SWARM, NetLogo, JADE, etc.) because it provides
a user-friendly nterface that allows inexperienced users to rapidly design and implement
agent logic and simulation behavior in a  owchart-like interface. Simulation settings are
similarly accessible via a simple menu system.
Repast Simphony provides a wide selection of built-in tools for users to track the course
of a simulation. The simulation environment also provides customizable graphical output
system, access runtime controls (step-wise execution, debugging, live modi cation of agent
setting, etc.), and has interfaces to export data to a broad range of statistical and math-
ematical programs including MatLab and other Simulink products. Repast Simphony also
provides interfaces to the Terracotta Java clustering infrastructure and GRASS GIS.
Repast Simphony models can be developed in several di erent forms including ReLogo
(a dialect of Logo), point-and-click  owcharts, Groovy, or Java. Java was selected as the
development language of choice for implementing THESIM due the investigator?s familiar-
ity as well as the  exibility provided. The Repast toolkit is integrated with Eclipse, an
open-source development platform comprised of extensible frameworks, tools and runtimes
for building, deploying and managing software. Eclipse has a well-supported development
community and is actively updated. The resultant Repast implementation is referred to as
THESIM-Repast.
83
Figure 5.1: UML model of THESIM class relationships
5.2 Class Structure
We begin by identifying the class structure and mapping the relationships between each
class to the relational diagram of the model. The resulting class structure can be seen in
Figure 5.1. We proceed to label and de ne the input parameters necessary to initialize the
simulation. Finally, we outline and de ne the simulation?s output parameters.
Based on analysis of  ve di erent types of agents: government, industry, university,
researcher, and grant, we developed  ve classes describing the characteristics and behaviors
of these agents. In addition there are two classes directly associated with the running of the
model named THEModel and THEEnviron. The THEModel class handles the creation of
84
indispensable variables, agent scheduling, control of the simulation (start, stop, display, etc.),
and the exogenous parameters accessed by active agents. The THEEnviron class contains all
the parameters and data-structures associated with the discrete environmental layers visible
to agents (resource terrain and market network).
5.3 Simulation Environment
The resource terrain is composed of several discrete two-dimensional grid layers. The
physical layer contains a number of grant, researcher, university, and industry agents. Each
additional layer is a communication layer corresponding to a speci c university. Each grid
position in the physical layer may contain several agents simultaneously, with university and
industry agents restricted to cells which do not contain resource agents. Cells in communica-
tion layers contain no agents, but rather pheromone data-structures used by the ant-colony
search variant used by researcher agents to explore the environment.
The grid layers of THESIM-Repast?s environment are initialized using variables Gridxand
Gridy to set the dimensions. Agents and decorators are then added to the physical layer
based on the input parameters speci ed by the user via the interface. The visualization of
the grid interface is color and shape-coded, red circles represent universities, yellow circles
represent industry agents, each small colored square represents a grant agent whose color
corresponds to one of  ve types representing di erent science and technology sectors. The
background color of each cell is determined selecting the pheromone deposition (Px;y;i) with
the largest non-zero value among all the communication layers.
Using this visualization schema we can see the con guration of grants within the envi-
ronment overlaid atop the growth and maintenance of the dominant search space for each
university?s population of research agents. Additional visualization schemas allow us to view
more detailed information, such as topological intensity graphs of university search space,
market networks (three-dimensional), and charts associated with overall simulation behavior.
85
a) t = 0
b) t = 150
Figure 5.2: University?s local grant neighborhood t = 0 and t = 150 respectively.
5.3.1 Activity Scheduling
Agent step activation order is determined by the Repast activity scheduler, which it-
erates with a simulated concurrency (i.e. in random order). At the beginning of each step
agents evaluate their current condition, location, and surroundings then take appropriate
action based on their type and status. Time periods are broken up into rounds consisting of
a  xed number of time steps speci ed by the user. Some agent classes (such as the researcher
class) perform operations at each time step, while others only perform operations at the start
of a new round.
Researcher agents activate once per time step. These agents traverse the environment
using movement rules speci ed in section 4.3 and perform search and resource gathering
activities. University agents activate once at the start of the round. They monitor their
status, perform management activities on their population of researcher agents, and iterate
technological development. Industry agents also activate at the start of the round and
develop research partnerships with university agents as well as iterate technologies. The
government singleton performs management activities on grant agents on a variable schedule.
86
Table 5.1 De nition of the basic input parameters within in THESIM-Repast
In the standard con guration, this maintenance step is performed at the start of the round,
however the user may specify alternative timing or activity conditions.
5.3.2 Parameters
THESIM-Repast takes a variety of user-speci ed input parameters which allow for var-
ious experimental conditions to be examined. For the purpose of determining feasibility, we
performed a down-select on the simulation conditions in order to determine those variables
most critical to the development of consistent system stability. The basic initialization inputs
can be observed in table 5.1. The simulation requires grid dimensions, information about
round duration, simulation length, the size of the academic and industrial bases, and the
level of government funding and intervention.
Optional parameters include risk-aversion thresholds, public-private growth target in-
formation, and resource proliferation modi ers. These optional parameters are not necessary
to initialize a standard run but they allow us to initialize a broad variety of basic condition-
testing experiments to validate the individual and system-level behaviors and verify whether
the simulation model put forward in chapter 4 indeed results in a system whose characteristics
match known innovation economies or are consistent with existent simulation experiments.
87
Table 5.2: De nition of the output parameters involved in THESIM-Repast
As outlined in chapter 4, THESIM-Repast provides a variety of output parameters used
to analyze the behavior of the system. These parameters allow us to validate the activity of
the various sub-models of the simulation as well as providing data for experimental veri ca-
tion. Table 5.2 contains a selection of the most salient the output parameters gathered from
a typical simulation run. These include; innovation production score, GDP, risk aversion,
and diversity calculations. Measures such as resilience and robustness are evaluated from
the aggregate data collected over the course of an entire experiment. The full set of in-
put and output parameters are discussed at length in chapter 6, including scenario-building
parameters used for stress testing the model.
5.3.3 Preliminary Experiments
Following the initial hypotheses proposed in chapter 1 we constructed an experimental
pro le intended to investigate two structural conjectures;
1. That the collection of behavioral rules put forward in the THESIM model resulted
in agent population and technology production behaviors similar to those observed in
statistical models of innovation economies.
2. The social trends present in inter-agent communication are consistent with the expec-
tations of the relevant sub-models (stygmergy, cooperative development, etc.).
88
3. Parameters associated with environmental  tness are closely tied to system stability
and resilience in the face of perturbation.
To evaluate the  rst objective, we pay close attention to innovation and value produc-
tion metrics (Y , G, and R respectively). We are also interested in patterns of technology
production on a per-round basis, though this is not captured in the simulation output pa-
rameters. The second objective is a consistency check against the  ndings of Dongsheng
& Yongen (2008) which is the most relevant model demonstrating socio-economic dynamics
between the various production stages within the triple helix. We use the global transac-
tion success rate (Sri) and cooperation con dence interval (Cai) as measures to evaluate
simulation performance in these areas. Finally, the third objective can be captured in the
experimental design.
Holling describes the reaction a resilient system displays to perturbation as \positive
feedbacks which overcome instabilities and variability experienced by the system, if not of
a magnitude su cient to exceed the systems? resilience." (Allen & Holling, 2008). By
instilling regular disturbance into the conditions of the experiment, we can determine the
e ective carrying capacity of the system by observing the reactions of the agent population.
This is accomplished by introducing a randomized resource upheaval at a regular interval
during the simulation. By focusing on ecological population metrics such as species richness
and diversity (Shannon-Weiner Index) we can strengthen claims about agent-based ecological
modeling. Speci cally, the notion that using ecosystem models for the behavior of agents and
the conceptualization of the environment allows us to better adapt concepts from population
dynamics for use in veri cation and validation.
Additionally, we were interested in determining whether the behavior of THESIM-
Repast?s sub-models remained consistent with known work in innovation economic simu-
lation despite the di erences in the social and environmental models. To this end we chose
to focus on broad markers of production, cooperation, and competition. The tests were run
89
Figure 5.3: THESIM Repast sample output showing grant clustering and exploitation regions
in a four university system.
Figure 5.4: System productivity via gross domestic product over time (t)
with ?small-world? initialization parameters and a  xed number of university and industry
agents.
The simulation was run for 50 trials; each consisting of 10000 ticks (335 rounds) under
default parameters. The averaged productivity results are shown in shown in  gure 5.5. The
GDP curve begins at turn 10 due to the absence of  rst-products in the system. Universities
and industries are forced to innovate alone for the early cycles before they have innovation
products to trade and monetize. We can see from the curve that production ramps up early
before the growth stabilizes. This is due in part to residual technologies, which depreciate in
value as their novelty erodes but nevertheless provide consistent income for industrial  rms.
One outcome worth noting is the consistency between the global value of all technolo-
gies and the productivity of industry agents seen in  gure 5.6. Due to the slow speed of
90
(a) Average global value of technology products over time (t)
(b) Average technology acquisitions by a single industry agent over time (t)
Figure 5.5: Behavior and value of technology acquisition among industry agents over time
(t)
91
Figure 5.6: Researcher agent population growth over time (t)
solo-innovation, early technology acquisitions are highly valued and provide signi cant com-
petitive advantage. However, as industries mature and their relationships with university
agents solidify, the need to make large, risky acquisitions is overcome by  nancial constraints
and general satisfaction with their level of competitive advantage over their rivals. These
results vary somewhat across the agent population but their aggregate behavior is consistent.
For the purposes of instilling con dence in the population model as it relates to ecological
domains we turn to metrics for analyzing communities of researcher agent. Looking at the
curve of Figure 5.7 we can observe a fairly typical population curve whose growth bounds
are reasonably well de ned and resemble real-life examples. Agents well optimized to exploit
the types of natural resources present in the environment persist and encourage the hiring of
additional agents of similar quality, while those that do not are eliminated from the workforce.
This workforce growth is limited by the amount of resources available in the environment.
As the capacity is reached, the population stabilizes (with minor  uctuations). While this
competitive mechanism is simple, and does not capture more complex notions such as tenure,
university-level research diversity initiatives (for example; it is unlikely a university would
eliminate the mathematics program simply because its research sta was less successful
acquiring federal grants), and long term research initiatives it does provide a proxy through
which ecological questions can be asked.
92
Figure 5.7: Shannon-Wiener index of researcher agent population (HR?) over time (t)
We look at the changes in HR?(where R is the population of researcher agents) to
give us some insight into the population dynamics at work within the simulation model.
Researcher-agent hiring and  ring (equivalent to ecological birth and death rates) are tied to
search performance. Biologically realistic H? values range from 0 (only one species present
with no uncertainty as to what species each individual will be) to about 4.5 (high uncertainty
as species are relatively evenly distributed). In theory, the H? value can be much higher than
4.5, although most real world estimates of H? range from 1.0 to 3.0. H? values close to 2.0
are considered normal and indicate a moderate amount of volatility in species density and
diversity. In general, it is thought that more disturbed and less stable environments should
have lower H? values. Looking at the curve of HR? in  gure 5.7 we see very little  uctuation,
with averaged values ranging from 1.281 to 1.53. This is in line with the resource model,
which is initially evenly distributed across all types, and expectations of agent population
dynamics.
Finally, we are interested in correlating our work with existing studies of innovation
economy simulation. To this end we made a deliberate e ort to pattern our technology
acquisition process on the work done by Dongsheng & Yongen (2008), who proposed a
speci c model for evaluating trade deals between university and industry agents whose output
could be measured as the ratio of successful to unsuccessful transactions and a trust metric
which in uenced likelihood to  nalize deals. Our model extends this theory by allowing the
93
(a) Er = 2:0
(b) Er = 4:0
(c) Er = 5:0
Figure 5.8: The Success Rate (Sr) and Average Con dence of industry-university partnership
(Ca) under conditions of risk-aversion (Er) over time (t).
94
opportunity for limited government intervention to help ?sweeten? deals which risk-averse
partners might otherwise decline. However, in this  rst experiment we have disabled this
functional interaction in order to gauge the consistency of our sub-model with the results
gathered in the 2008 study.
Much like the earlier work, we can observe that the curves of Srg and Ca vary with the
industry agent?s expectation of return (given by Er = 2.0, 4.0, and 5.0, respectively). The
rapid increase of cooperation during the early phase of the simulation can be attributed to
the need on the part of industry  rms to address their competitive needs by acquiring new
technologies (as they start out with none). However, these early ventures are often costly, and
depending on the level of risk-aversion may sour these  rms to future cooperation or at least
reduce the amount they are willing to bid. Low bids, combined with high risk aversion results
in fewer successful transfers which in turn depresses the success rate. A feedback reaction
exists between acquisition success or failure and adjusting of trust values (successful bids
raise trust, unsuccessful bids lower trust). Thus we observe an e ective sensitivity threshold
in Er such that high risk aversion engenders consistently lower trust between  rms. Lower
values of Er result in higher degrees of consistent and stable cooperation, while higher values
cause this relationship to fall to ruin. We also see that Sr reaches a stable value gradually
after a period of evolvement. From this we draw a similar conclusion with the authors of
the similar study, which is that industry and university agents are able to form a stable
cooperative relationship.
95
Chapter 6
THESIM Experimental Design & Analysis
6.1 Introduction & Objectives
Our experimental goal is to explore four facets of knowledge ecosystem development;
system productivity, carrying capacity, resilience, and inter-agent relational structure. This
follows the overall aim of our thesis; the development a computational theory of the triple-
helix theory of public-private partnership. To this end, we can revisit the initial hypothesis
guiding the work as stated in chapter 1;
Ha: The development and maintenance of innovation economies can be explained on the
basis of an ecologically driven agent-based model of the triple-helix model of public-private
partnership.
 Ha1: Agents need a minimum set of interacting components to meet speci c role-de ned
goals in an unfamiliar environment.
 Ha2: Knowledge in the world can be represented by ecological a ordances and envi-
ronmental information. The process of innovation creation and transport works on the
basis of the interplay between the two.
While the development of the THESIM model following ecological agent-based modeling has
provided satisfactory resolution to Ha1, Ha2 requires dedicated experiments designed to more
fully investigate the claim.
Principally, we can validate our resolution of Ha2 by providing clear analysis of our
computational model that indicates how the transformation of a set of system variables by our
implementation of the simulation model results in the innovation creation and conservation
96
behavior described by triple helix theory. In particular, if we can show that there is some
small set of critical variables that control key model output parameters we can falsify the
null hypothesis that the system?s behavior is indistinguishable from a random distribution
of values (i.e. all relevant system output is noise). It is therefore instructive to develop a
comprehensive experimental model to structure our inquiries regarding simulation behavior
and performance.
For the purpose of completeness and replicability we follow the guidelines and recom-
mendations given in support of a systematic design of experiments (DOE) as detailed by
Lorscheid et al. (2011). DOE increases the transparency of simulation model behavior
and the e ectiveness of reporting simulation results. This technique is well suited to socio-
economic simulation and has been applied successfully to multi-agent market models (Meyer
& Troitzsch, 2012; Klingert & Meyer, 2012).
6.2 Systematic Model Analysis
This section illustrates the application of DOE principles for structured experimental
analysis of the THESIM model. We proceed from the standard reporting template provided
by Lorscheid et al. (2011), paying speci c attention to the variable description and selection
process due to the exploratory nature of the proposed experimentation. For the communica-
tion of results, we make use of standardized output formats described in the aforementioned
work and detailed in later sections.
As described in Chapter 5, the THESIM model is used to examine the structure and
functionality of Triple-helix aligned knowledge ecosystems. Therefore, the performance of
the system under simple policy and environmental constraints are analyzed and compared.
The simulation experiment has three objectives: First, the e ects of inter-agent innovation
trade on system productivity are analyzed through a relative comparison of the performance
several experimental models with di ering endogenous environmental parameters. Second,
the optimum carrying capacity in terms of innovation production, agent-population, and
97
Table 6.1 Simulation DOE full parameter selection.
98
system con guration in the given model scenario is determined to enable the comparison.
Third, we examine the system?s overall resistance to shock by introducing environmental
stress and calculating a resilience threshold for the base con guration of the model.
To prepare our experiment, the variables of the THESIM model are classi ed in two
steps. We begin by enumerating all the independent, dependent, and control variables
present in our experimental model. These values are taken directly from the source-code
de ning THESIM operation and are more comprehensive then the basic set presented in
chapter 5. These simulation parameters are used as reference points to identify all model
variables (Table 6.1).
We proceed to identify all parameters which in uence model behavior and include them
in the DOE as independent variables. Parameters deemed necessary to evaluate the simula-
tion model are entered as dependent variables. Support parameters are included as control
variables as they are necessary for calculating the simulation outcome but do not vary be-
tween simulations. As they are required for implementation purposes only, they are not
included as model variables.
Independent Variables (xi)
 Government Annual Funding: x1)
 Researcher Annual Pay Rate: (x2)
 Industry Knowledge Value: (x3)
 Industry Maintenance Rate: (x4)
 Industry Innovation-to-Money Conversion Rate: (x5)
 Government Tax Rate: (x6)
 Expected Pro t (Er): (x7)
 University Count: (x8)
99
 Shake-Up Time: (x9)
Control Variables (Ci)
 Economic Model (C1)
 Grid XY-dimension (C2)
 Update Style (C3)
 Patch Density (C4)
 Patch Cull Probability (C5)
 Patch Distribution Policy (C6)
 Patch Maintenance Policy (C7)
 Minimum Patch Energy (C8)
 Maximum Patch Energy (C9)
 Patch Migration Policy (C10)
 University Random Distribution (C11)
 Run Length (C12)
 Growth Conservation (C13)
 Additive Market (C14)
 Government Initial Funding (C15)
Dependent Variables (Ri)
 System GDP: (R1)
 Researcher Population Count: (R2)
100
 Researcher Population Diversity Index: (R3)
 University Size: (R4)
 University Innovation Creation: (R5)
 Industry GDP: (R6)
 Industry Partner Count: (R7)
 Technology Transfer Count: (R8)
 Network Structure Graph: (R9)
After classi cation, these variables are then collected in order to develop an overview of
the model components (Table 6.2). This allows us to determine areas for potential experi-
mentation on the model.
Based on the classi cation of variables, the investigated relation in the model can be
described as the e ect that di erent (x1-x9) environmental and behavioral parameters have
on the (R1, R6) GDP, (R2-R5) population dynamics, (R5, R7-R8) innovation carrying ca-
pacity, and (R9) network structure graph of the triple-helix knowledge ecosystem. This can
generally be represented in the diagram below below (Figure 6.1).
The variation of potential independent variables (C1) economic model, (C4) patch den-
sity, and other policy-based options are not of interest for the given exploratory exercise. As
a result, these parameters remain  xed as control variables in the model. The variable (C1)
economic model contains all endogenous agent decision parameters associated with  scal dis-
tribution and that form the basis of the experimental model. Environmental variables (C2
-C11) can be combined into a more general environmental model for the experiment. These
features remain constant across the lifetime of the experimental run and can be referenced
by agents for the purpose of local competitive decision-making.
The variable (C13) growth conservation sets a maximal distribution of new researcher
agent creation and destruction by university agents based on their current size. This tuning
101
Table 6.2 Formal classi cation of THESIM DOE variables.
Figure 6.1: Goals: Relationship Mapping of xi to Ri
102
parameter allows for the introduction (or reduction) of population volatility across the system
by limiting the rate at which universities can add, remove, or replace researchers in their
pool. Run length (C12) de nes the number of time steps per simulation. In our experiment,
it states that the number steps in a single run is initialized to 1000. The variable (C14)
additive market allows an experiment-de ned fraction of pro ts to be return to the system
at the end of each round. Finally, government initial funding (C15) is held constant across
each simulation run.
The dependent variables are (R1) system GDP, (R2) researcher population count, (R3)
researcher population diversity, (R4) university size, (R5) university innovation creation,
(R6) industry GDP, (R7) industry partner count, (R8) technology transfer count, and (R9)
degree centrality. System GDP denotes the total economic productivity of the system and is
calculated by combining the total value of all technological assets and pro ts generated by
innovative behavior.
Researcher population count is running total of all active researcher agents active in the
system at the end of each time step. Meta information for each agent is retained for the
purposes of more granular investigation into the properties of sub-populations. Addition-
ally, researcher population diversity is the Shannon-Weiner diversity index of the researcher
population at the end of each time step. University size is a list of all active universities
within the system and the size and composition of their researcher population. University
innovation creation tallies the total number of new technologies completed by all universities
in the system.
Industry agent-speci c productivity information is collected by the variable industry
GDP which indicates the total productivity of industries represented in the model. Indus-
try partner count represents the total number of unique universities that have completed
successful transactions. Following on, technology transfer count indicates the total number
of successful technological transactions between universities and industry across the lifetime
103
Table 6.3 Goals: Description of DOE factors and factor levels.
of the simulation. The network structure graph provided by the simulation isan xml doc-
umentrecording the links incident upon each university and industry agent in the graph
representation of the system along with performance metadata for each agent. These graphs
are stored for future network structure analysis and data visualization e orts.
Within the simulation experiment, the described model variables are analyzed as re-
sponse variables with distinct factor levels. To accomplish this, the model variables need to
be transformed into factors whose factor level ranges are de ned following factorial reduction
strategy (Table 6.3). To accomplish this, we begin with a basic 2k experimental design. In
this case, for the set of response variables Ri we outline an experimental case designed to
evaluate the combinatorial e ect of each independent factor xi (as indicated in Figure 6.1).
This exhaustive approach to the evaluation of the Top-Level-DOE represents the  nal step
towards answering the research questions;
1. What is the minimum set of factors that describe the response variance of the model?
2. What are the critical thresholds of these factor-levels that de ne di erent response
con gurations?
For the purpose of developing a tractable set of early experiments to evaluate these questions
we temporarily contract our independent variable range by pegging the variable shake-up
time (x9) to its  rst factor level for the sake of simplicity. This variable is only relevant for
104
Table 6.4 Top-level DOE chart.
Table 6.5 DOE E ect Matrices.
speci c experiment types (system reaction to perturbation) and can thusly be treated as a
control variable for the purpose of early experimentation. The resulting model gives us a 2k
model where (K = 8) that can be represented using a Top-level DOE chart. This model can
be represented by an e ect matrix (table 5).
6.3 Sub-DOE THESIM-Sensitivity 1st-iteration
In order to proceed toward answering the  rst question ( nding a minimum set of
factors), basic estimation of error variance and variable sensitivity must be performed on
the model in order to reduce the number of factors in our experimental design to a more
tractable set. Sensitivity analysis of a model is performed to determine how outputs vary
with changes of input parameters, in order to identify the relative importance of parameters
105
and to help in optimization of the model. Traditionally, this is performed using the one-
factor-at-a-time (OFAT) method. However, the results of OFAT are unreliable if there
are signi cant interactions between parameters (Mikhael & Funnell, 2005). Therefore, we
incorporate the Taguchi method of sensitivity analysis in to our experimental design as an
alternative to OFAT or full-factorial selection. This method employs only a small number
of all the possible combinations of model parameters to estimate the main e ects and some
interactions. An orthogonal array (OA) is used to reduce the number of simulations (Taguchi,
1987) but still obtain reasonable information.
The procedure for applying the Taguchi method is as follows. Minitab 16 was used for
the calculations.
1. Select parameters and interactions of interest.
2. Select parameter levels.
3. Find a suitable OA with the smallest number of runs. This normally involves looking
up a prede ned OA based on the numbers of parameters, interactions and levels.
4. Map the factors and values to the OA.
5. Run simulations based on the OA.
6. Analyze the simulation results.
The eight relevant independent variables (x1-x8) were selected for this analysis, with
shake-up time (x9) omitted for the sake of initial relevance. This process can be summarized
in table 6 via an OA L8 212 , representing 8 two-level parameters, and a total of 12 simulation
con gurations.
The eight outputs investigated (R1-R8) are summarized in the previous section. A total
of 12 static  nite-element simulation models were input and each was performed 30 times in
REPAST; Their results are summarized in Table 6.6. The experimental designation for this
trial is recorded as TKGL12-2.
106
Table 6.6 OA L8 212 table for investigation and simulation results.
Next, the e ect analysis of all factors is conducted for the simulation data. For a
systematic analysis of the factor e ects, the analysis proceeds in two steps. First, factors
are tested for signi cant e ects on each simulation response. Second, the strength of all
identi ed signi cant factor e ects is calculated. For the  rst step, testing the factor e ects
for signi cance, a generalized linear model (GLM) was developed for the model outputs based
on the L8 Taguchi test matrix. Table 6.7 shows the p-value output of the GLM with respect
to each response. In statistical signi cance testing p represents the probability of obtaining
a test statistic at least as extreme as that which was actually observed in the simulation.
Because this level of testing will be focusing on main e ects in a fairly complex model,
we relax the con dence interval for signi cance by apply additional gradations of p. We
establish that weak signi cance is attributed to p  0:25 (75%), moderate signi cance is
attributed to p 0:1 (90%), and strong signi cance is attributed to p 0:05 (95%). The
strong signi cance class follows the general ANOVA model for p and will be most relevant
in later iterations which will be more tightly focused.
Due to the size of the model and the number of initial response variables, only  rst-
order response interactions were performed. By summing the p-values for each factor across
107
Table 6.7 p-value results from GLM of TKGL12-2.
the respective response parameter in this model we immediately see that there are clear
disparities in predictive signi cance. This signi cance of factor e ects can be read from the
GLM response table. While there is no clear main e ect on all responses from any single
factor, it?s clear that x4, x5, x6,and x7 display relative importance indicating potentially
hidden interaction e ects for each response. We also see that several factors have main e ects
approaching or within the strong interval (x2, x4, and x6 respectively). This immediately
indicates that we might be well served focusing our investigation on these factors.
6.3.1 Main E ect Analysis
In order to reduce our model?s parameter space to make it more tractable for traditional
analysis of variance we  rst examine the factor e ects in more detail by examining e ects
plots for both individual factors. First-order interaction e ects for each factor were also
calculated and charts containing this information are available in the appendix. For the
purpose of this exercise we will restrict our analysis to main e ects. Further DOE iterations
will examine speci c interaction e ects of interest.
This form of analysis di ers from the p-test included with the GLM in that we are
interested in the slope of the graph. In the main-e ects graphs, nearly-horizontal lines
indicate little e ect, while the degree of slope indicates the strength and direction of main
e ect. Comparisons of ?weak?, ?moderate? and ?strong? are relative to the set of response plots
108
Figure 6.2: Parameters? main e ects on System GDP (R1).
and do not imply correlation with particular p-values. This correlation step is restricted
to higher-order analysis with respect to factor interactions. We also use some analytical
nomenclature to describe the e ects in terms of the domain, innovation ecosystems for the
purpose of providing a summary of the trends in main e ects.
Parameter main e ects on system GDP (R1) Tested parameters have strong
confounding main e ects on system GDP (Figure 6.2). No clear dominant set of factor
responses appear present among main e ects.
 Strong negative e ect from higher taxation (x2).
 Strong negative e ect from higher expected pro t (x8).
 Strong positive e ect from higher industry maintenance overhead (x6).
Parameter e ects on researcher population count (R2)
109
Figure 6.3: Parameters? main e ects on researcher population count (R2).
 Compared with the other 6 parameters, government tax rate (x2), university count
(x4) and knowledge value (x5) have the greatest main e ects on system productivity
(Figure 6.3).
 Strong positive e ect from increased taxation.
 Strong positive e ect from increased university count.
 Moderate negative e ect from increased knowledge value.
Parameter main e ects on population diversity (R3)
 Compared with the other 6 parameters, university count (x4) has the greatest main
e ects on system productivity (Figure 6.4).
 Strong positive e ect from increased university count.
 Moderate negative e ect from increased knowledge value.
Parameter main e ects on average university size (R4)
110
Figure 6.4: Parameters? main e ects on researcher population diversity (R3).
 Compared with the other 6 parameters, researcher annual pay rate (x3) and industry
IMC rate (x7) have the greatest main e ects on system productivity (Figure 6.5).
 Strong negative e ect from increased researcher salaries.
 Strong positive e ect from increased industry IMC.
 Moderate negative e ect from increased government funding.
 Moderate positive e ect from increased risk sensitivity (x8).
Parameter main e ects on university-industry trust (R5)
 Compared with the other 5 parameters, government funding (x1), industry knowl-
edge value (x5), and industry maintenance rate (x6) has the greatest main e ects on
university-industry trust (Figure 6.6).
 Strong positive e ect from increased government funding.
 Strong positive e ect from increased knowledge value.
111
Figure 6.5: Parameters? main e ects on average university size (R4).
 Moderate positive e ect from reduced industry maintenance e ciency.
 Moderate negative e ect from increased university count.
Parameter main e ects on average industry GDP (R6)
 Compared with the other 6 parameters, university count (x4) and industry IMC rate
(x7) have the greatest main e ects on average industry GDP (Figure 6.7).
 Strong positive e ect from increased university population.
 Strong positive e ect from increased industry IMC.
 Moderate negative e ect from increased taxation.
Parameter main e ects on average on industry partnership success rate (R7)
 Compared with the other 5 parameters, government funding (x1), university count
(x4), and industry maintenance rate (x6) have the greatest main e ects on industry
partner count (Figure 6.8).
112
Figure 6.6: Parameters? main e ects on university-industry trust (R5).
Figure 6.7: Parameters? main e ects on average industry GDP (R6).
113
Figure 6.8: Parameters? main e ects on industry partnership success rate (R7).
 Strong positive e ect from increased government funding.
 Strong negative e ect from increased university count.
 Strong positive e ect from reduced industry IMC.
 Moderate positive e ect from increased industry knowledge value.
 Negligible e ect from increased risk sensitivity (x8).
Parameter main e ects on average on technology transfer count (R8)
 Compared with the other 5 parameters, university count (x4), industry knowledge value
(x5), and industry IMC rate (x7) have the greatest main e ects on industry partner
count (Figure 6.9).
 Strong positive e ect from increased university count.
 Strong negative e ect from increased industry knowledge value.
 Strong positive e ect from increased industry IMC.
114
Figure 6.9: Parameters? main e ects on technology transfer count (R8).
6.3.2 Sub-DOE THESIM-Sensitivity 1st-iteration Summary
While a broad range of interesting e ects are presented, a key  nding of this initial
experimentation indicates that while most responses are tightly bound to an identi able set
of interactions between factors, several (R1, R4, and R5) display no strong factor e ects or
confounding e ects. This implies that there is some other, hidden e ect (or mixture e ects)
governing their behavior. Though useful for our analysis, the interaction graphs for each
response have been omitted from this chapter for brevity, they are available in the appendix
(Appendix 1).
In light of this observation, it is helpful to represent this data in terms of general e ects
on response data. By isolating each interaction and applying a minimum exclusion threshold,
we can build a table of e ectiveness to give us a rank-order notion of the interactions taking
place within the model (table 6.8).
In the above  gure ?+? represents a positive e ect, ? ? represents a negative e ects, and
?=? represents a null e ect. Darkened cells represent moderate to strong e ects, while empty
115
Table 6.8 Sign representation of main-e ects of factors in GLM of TKGL12-2.
cells represent weak e ects. Combining this view with the p-values in Table 6.7 we can make
the case for reducing our the factor-scope of our experimental design down to x4, x5, and x6
as these factors seem to carry the strongest e ect across the response variables of interest.
While e cacious, x7 seems to operate as a scalar calibrating e ect and could be  xed for
greater experimental simplicity. Following this observation we can fully omit those factors
which have only weak or calibrating e ects on model response.
At this point, the analysis process is pursued iteratively. With these factors  xed, the
analysis can be continued by running more simulation experiments with further factor level
combinations on a smaller variable-set to  nd best possible performance for each response
of interest.
It should be noted that this  rst step satis es our  rst question, namely that we can
represent a demonstrably large set of model response output through activity on this reduced
set of input factors. However, in order to answer the second question we must isolate the
factors of interest and determine discrete thresholds for system behavior across our responses.
116
Table 6.9 Factor levels for 2k3 factorial design of Sub-DOE.
6.4 Sub-DOE THESIM-Sensitivity 2nd iteration
As stated, the goal for this iteration is to identify criticalities among the remaining
factors of interest. We do this partially in order to instill con dence in our model, but also
to determine if the interaction is consistent with empirical knowledge regarding real-world
systems.
The starting point for this step is once again selecting an appropriate factorial design.
New factor levels around the previously successful levels are de ned for the next simulation
experiment iteration with the ultimate goal of identifying the most successful factor levels
in the factor screening process. As per our initial analysis in the main DOE we were able to
reduce the scope of our DOE to three factors; x4, x5, and x6 (Table 6.9).
Due to the mixed e ects of each factor on our response variables we peg the control
factor levels at the mean values described in Table 5 to mitigate variance.
The combinations of the new factor levels yield a total of eight (23) new design points.
They de ne the simulation experiment as shown in the new design matrix (Table 6.10) which
is referenced as 2k3E. The values depicted in the design matrix enable the two-step process
outlined above for analyzing all factor e ects and their e ect strengths. The e ect matrix
captures the outcome of the e ect analysis.
117
Table 6.10 2k3 Factorial Sub-DOE simulation p-value results.
6.4.1 Response Analysis
In this iteration we eschew main e ect plots in favor of more traditional e ect matrices
due to the fact that we want to di erentiate between speci c main e ects and interactions
across the experiment. In the prior iteration such plots were not tractable as there were far
too many factor interactions to appropriately attribute e ects in this fashion. Each e ect
matrix displays the crosswise comparison of main e ects and interaction e ects (p-values)
in a fashion similar to the example (Table 6.11). In the discussion of each table, ?positive?
and ?negative? e ects refer to sign (direction) of the slope of the e ect plots. The individual
slope computations for each e ect are omitted from this section for the sake of brevity, but
are included in the discussion via short-hand nomenclature.
For this iteration we tighten our con dence interval to the range more traditionally
associated with ANOVA, with weak e ects attributable to p  0:15, moderate e ects at-
tributable to p 0:05, and strong e ects attributable to p 0:01. To avoid confusion with
prior terminology, we will speak generally about signi cance and attempt (where possible)
to link this analysis to prior  ndings in the experiment set.
Analysis of 2k3E-R1
The high factor level of x4 stands out with its positive e ect on R1 (Table 6.11). Thus,
a further testing of factor x4 around the high factor level (?+? = 20) might lead to a better
118
Table 6.11 2k3E R1 e ect matrix.
Table 6.12 2k3E R2 e ect matrix.
system productivity than R1 = 1782911:864 which is the best value observed so far (see Table
6.10, design point 8). Weak interactions are observable but fall outside of our signi cance
threshold. The factors x5 and x6 have only minor e ects on R1 within the given factor level
interval. Hence, for further study on R1 they should be  xed as control at their mean values.
Analysis of 2k3E-R2
Similar to the  ndings from R1, we see that x4 has a substantial e ect on R2 (Table
6.12). The e ects of x5 and x6 respectively have minor impact on R2 within the given factor
level. A p-value less than 0:01 indicate that the performance of the R2 is tightly correlated
with x4. This supports the prior recommendation that further analysis around the high
factor level (?+? =25) might lead to higher performance of R2, the best being R2=114:1078.
The factor x5 appears to have little impact on R2 and it is recommended that any further
investigation on this response  x x5 at its mean factor-level.
Analysis of 2k3E-R3
While there is no clear correlation within the established con dence intervals, R3 (pop-
ulation diversity) seems to operate as a function of the interaction between x4 and x5 (Table
6.13). Returning to table 6.7, we initially saw that the behavior of this response appeared
to be correlated with the main e ects of all three factors. However, further study seems to
119
Table 6.13 2k3E R3 e ect matrix.
Table 6.14 2k3E R4 e ect matrix.
indicate only moderate complimentary interactions between these factors. Indeed, several
factors not contained in this iteration may play a larger role in the outcome of R3 than was
indicated by the initial this test. This suggests that further experimentation focusing on
these omitted factors would be prudent. These  rst-order interactions can be observed in
interaction graph for R3 (Appendix).
Analysis of 2k3E-R4
The high factor level of x4 (Table 6.14) stands out with its positive e ect on R4 (univer-
sity researcher-population size). This result is actually unsurprising, as increasing the num-
ber of universities in an environment with a  xed resource quantity would suggest greater
competition and a general reduction in the average population size per university. This
reduction appears to behave in a scalar fashion (Table 6.10) which could be determined by
further testing around factor x4. Interestingly, while the values lie outside strict con dence
intervals set by this iteration level, there appears to be a weak interaction (negative) interac-
tion between x4 and x6 within the given factor level. Finally, x5 seems to have little bearing
on the response. In further studies it is suggested that x5 be  xed as a control to its mean
factor-level.
Analysis of 2k3E-R5
120
Table 6.15 2k3E R5 e ect matrix.
Table 6.16 2k3E R6 e ect matrix.
While x4?s contribution to R5 (inter- rm trust, Ca) is near enough to the con dence
interval (95%, or p=0:05) to imply strong correlation, it should be noted the factor x1
(annual government funding) which was included in the DOE iteration as a control displays
a signi cant positive main-e ect on the outcome of this response (Table 6.9, design point
4). Despite this, increasing x4 has a predictably negative impact on R5 (Table 6.15). This
indicates that increased competition via the addition of more universities has a curiously
negative e ect on the mean trust across the system. However, this outcome may belie more
dynamic trends within the model that are not captured in the response, as R6 measures
the mean of Ca rather than the distribution. Further re nement of model response output
may be necessary to fully extract the granular behavior of the system under the factor-levels
included in this iteration. Due to the relatively minor impact of interaction e ects, it is
suggested that future iteration on R5  x x5 and x6 and their respective mean factor-levels.
Analysis of 2k3E-R6
Observing the results in Table 6.16, x4 displays a strongly positive main e ect on R6
(average industry GDP). While the a ect appears to be scalar, further analysis around the
high factor level (?+? = 25) might lead to higher performance of this response, the best being
R6=2059:386 (Table 6.10). A minor interaction is observable between x4 and x6, but its
121
Table 6.17 2k3E R7 e ect matrix.
Table 6.18 2k3E R8 e ect matrix.
e ect seems to be dominated by the action of x4 on the response. Due to the relatively
minor impact of interaction e ects, it is suggested that future iteration on R5  x x5 and x6
and their respective mean factor-levels.
Analysis of 2k3E-R7
Interestingly, while x4 and x5 exhibit strongly positive main e ects on R7 (inter- rm
trade success rate, Sr), countervailing negative interaction e ects exist between x4 and x5/x6
(Table 6.17). This suggests a more complicated relationship between the three factors. The
response data in Table 6.6 and Table 6.7 indicates a correlation between R5 (Ca) and R7 (Sr),
as pro ered by Dongsheng & Yongan (2007). and may indicate that whatever mechanisms
impact one response may indeed work upon the second in similar fashion. Because of this
observation, it may be worth re-examining the interaction e ects between x4, x5, x6 and x8
(Er) which was identi ed in prior as a bounding factor governing the behavior of R7.
Analysis of 2k3E-R8
Both x4 and x6 display strongly positive main e ects on R8 (technology transfer count),
as well displaying a moderate positive interaction e ect (Table 6.18). , further analysis
around the high factor level (?+? = 25) might lead to higher performance of this response,
the best being R8=6106.2. This positive interaction is of particular interest when combined
122
with observations on R1, as it strongly supports the main supporting argument of innovation
ecosystem development; the development and transfer of new technology from academia to
private industry is a key driver of economic growth. The factor x5 appears to have little
impact on the response and it is recommended that future studies  x this factor to its mean
value.
6.4.2 Sub-DOE THESIM-Sensitivity 2nd-iteration Summary
A key  nding of this iteration is that, unlike the  rst iteration, the impact of x4 on
the broad range of system responses is strong. The notion that increased regional university
count (or by proxy, density) is a strong indicator of innovation production capacity is well
documented and is supported by literature (Cooke et al., 1997; Varga, 1998). There is an
obvious cleavage in the response data recovered from the experiment set that indicates that
increasing the number of universities in the model has a positive impact on virtually all
factors and may dominate the contribution (positive or negative) from most other factors.
In a majority of the experimental cases we recommend  xing the values of x5 and x6,
leading us to the general conclusion that the factor level of x4 accounts for a majority of the
variation across most of our included response variables. However, the degree and direction
of this variance does not appear to be scalar, and in fact there are several interesting response
interactions that would bene t from further DOE iteration.
The most curious result can be seen in the behavior of total researcher population (R2),
which appears to be the response most tightly bound to x4. While relatively pronounced
with respect to the other responses, this  nding is interesting when observed in context
with system GDP output (R1) and average university size (R5). One would expect, given
the negative impact of inter-university competition, that R3 would be relatively static as
a 2-university system would result in two very large agent populations of similar total size
to that in a system with more universities (i.e. that researcher population is a function of
123
resource quantity). Instead, increasing x4 to its ?+? factor level results in a signi cantly
larger total population. In context, this suggests the presence of ?hidden? interactions.
The  rst relates to the local-search dynamic of each researcher agent, which has a set
search time (round length) in which it can make only a single move. This e ectively limits
the visible area of researchers belonging to a particular university to a radius potentially
smaller than the bounds of the environment. This means that in a 2-university system with
randomly distributed resources, it is possible that many grants are simply impossible to
reach, resulting in bounded growth.
The second may be more subtle, as we also see an increase in the total number of
technology transfers in the system. This implies that R3 is actually a function of available
environmental resources, university search space, and (most critically) industry participation
in technology trade. Determining the validity of this observation requires more focused
testing on independent values of x4 and perhaps observation on a di erent set of response
variables than those selected for the general model analysis.
Another interesting observation comes from our analysis of population diversity (R3).
It is clear from results that the factors included in this iteration do not appear to strongly
impact the outcome of this response. This is actually expected, as the diversity index closely
re ects the fact that the government agent selects grant funding types using a stochastic
function which follows a normal distribution. As a result, there is no inherent favoritism of
one  eld of funding over another over the course of a simulation run resulting in an evenly
balanced population.
Indeed, a closer look at granular data from individual runs indicates that our researcher
agent model quickly trains itself to match even minute changes in grant distribution, and this
training speed can be roughly treated as a resilience measure. However, to fully quantify the
behavior of this response requires speci c experimentation which can be pursued in a separate
iteration once the main factor interactions are explained. In particular, this experimental
model should include the omitted input factor ?shake-up time? (x9).
124
The  nal exception to the primacy of university count?s response dominance can be
observed in the negative impact on university-industry trust (R5). This  nding at  rst
seems counter-intuitive from a pure innovation perspective, but upon examination actually
 ts with the concept of economic and ecological competition. Consider that the competitive
space (grant environment size) and total available resources remains  xed between the test
groups, but x4 is increased. This leads to greater competition for resources which in turn
reduces the average researcher population per university in the model.
Since technology production and iteration operates as a function of the maturity level of
the university which itself is tied to its researcher population size, this means that in denser
models each individual university actually produces fewer technologies more slowly. Because
our implementation of Transaction Cost Theory introduces downward pressure on universi-
ties to accept technology transfer requests at reduced rates when their store of technologies
and iteration speed is higher, scenarios which negatively impact those two internal factors
result in more rejections. In order to quantify the particular dynamics of this relationship
further iteration on this design point is necessary.
In summation, this iteration step obviates the notion that the university count is re-
sponsible for a majority, but not all, of the model?s response variance. In order to quantify
the speci c e ect, we perform a  nal iteration before looking at the special cases discussed
previously.
6.5 Sub-DOE THESIM 3rd-iteration
As explained by prior analysis, factor x4 has been shown to be the dominant e ect on
response variables R1, R2, R4, R5, R6, and R8. In the previous iteration of the model?s
DOE, we also observed only weak interactions which were outside the chosen signi cance
thresholds. Therefore, in this iteration new factor levels are de ned for x4 for the subsequent
analysis with all other factors designated as control variables and restricted to their mean
factor-value (Table 6.19).
125
Table 6.19 Sub-DOE third iteration factorial design.
Thirteen new factor levels around the former successful high factor level of x4 (?+? = 25)
are chosen (Table 6.20). This gives us a fairly wide window to compare e ects of incrementing
this value step-wise.
Because we are looking for patterns rather than maximizing any single response, we can
immediately identify two countervailing trends in response behavior. Maximal values for
average research population size (R4) and average trust (R5) can be seen about DP2, with
very clear fallo as university count (x4) increases. Maximal values for system productivity
(R1), researcher population (R2), average industrial  rm productivity (R6), and technologies
transferred (R8) emerge around DP11. This indicates that for future experiments looking
to maximize quantities in R4 and R5, low values close to DP2 should be considered. Alter-
natively, experiments looking to maximize quantities for R1, R2, R6, and R8 higher values
of x4 close to DP11 should be considered.
Given that policy-makers are mostly interested in boosting economic productivity, grow-
ing more robust business communities, and increasing technological competitiveness, we can
strongly recommend DP11 as the best factor level for x4. Moreover, an analysis of these
results indicates that above DP5, R1 is a semi-scalar predictor for R2, R6, and R8.
126
Table 6.20 Sub-DOE third iteration analysis of selected response variables versus change in
x4.
Table 6.21 Optimal parameter combination.
6.5.1 Sub-DOE THESIM-Sensitivity 3rd-iteration Summary
In this iteration we focus on the behavior of the university count factor (x4) and establish
a bidirectional trend in the response factors of interest. We were able to establish the tight
coupling of x4 to the behavior of R1, R2, R4, R5, R6, and R8 and isolate two design points
for x4 which promote optimality of two subsets of response data.
We establish the primacy of maximizing R1, R2, R6, and R8, and we also observe that
R1 is a predictor for the remaining factors. In the given scenario of R1 = 1812086:911, 35
emerges as the best factor level for maximizing relevant simulation output. The  nal step for
factor screening is to set the factor level for all sub-factors at the highest measured response
for R1 (Table 6.21). The response variable values of this parameter combination will serve
127
as the performance value for the simulation model in the Top-level DOE. All Sub-DOE of
our sample model have been analyzed in the described way. On this basis, the calibrated
sub-factors with the best performance can be entered in the Top-level DOE as the factor
level values for the subsequent experimental scenarios.
6.6 DOE Analysis & Observations
The observed trends lead us to the following set of conclusions about the model;
The set of input factors can be initially reduced to the following set;
 Government Annual Funding: (x1)
 Researcher Annual Pay Rate: (x2)
 Industry Knowledge Value: (x3)
 Industry Maintenance Rate: (x4)
 Industry Innovation-to-Money Conversion Rate: (x5)
 Government Tax Rate: (x6)
 Expected Pro t (Er): (x7)
 University Count: (x8)
 Shake-Up Time: (x9)
The set of response factors can be reduced to the following set;
 System GDP: (R1)
 Researcher Population Count: (R2)
 Researcher Population Diversity Index: (R3)
 University Size: (R4)
128
 University Innovation Creation: (R5)
 Industry GDP: (R6)
 Industry Partner Count: (R7)
 Technology Transfer Count: (R8)
A large class of these response factors is signi cantly impacted by variation on a single factor,
university count (x4).
There are, however outliers to this trend. The behavior of response factor R3 is largely
a function of the distribution of grant-types in the environment and therefore is independent
of the factors chosen for experimentation. x4?s impact on R1 is signi cant but complex and
may require additional study to accurately quantify.
x4 acts strongly on responses R2, R4, R5, R6, and R8 which are the responses of interest
in this model. The action of x4 on the remaining variables is bidirectional in nature:
 High values of x4 corresponding with high performance of R1, R2, R6, and R8.
 Low values of x4 corresponding with low performance of R4 and R5.
Because R1, R2, R6, and R8 are of principle importance to policy makers and inves-
tigators, the top-level DOE design should  x these factors at the values recommended in
table 6.21. Further experiments should now focus on environmental con guration or policy-
designs.
6.7 THESIM-ER Ecological Resilience Experiment
Following our formal description of a Top-level DOE we can now initiate experiments
designed to better quantify the ecological factors in model behavior alluded to in Chapter
1. Of principle interest is the notion of structural ecological resilience in complex socio-
economic systems driven by human activity (Alberti & Marzlu , 2004). In the DOE, we
129
Figure 6.10: A hypothetical well-being function for an idealized system.
establish the relationship between the input parameters and establish tuning thresholds for
system performance based on a set of response variables of interest.
In this regard, we develop an experiment to measure the ecological resilience of our model
using the technology yield Y of a system experiencing periods of stress. In this approach
stress is represented as an across-the-board resource disturbance and reduction approximat-
ing drought-like conditions. The two environmental variables of interest are disturbance
frequency (Df) and disturbance duration (Dl). This sub-model is an attempt approximate
the impact of shifting policies on a regional innovation ecosystem and brings forward vul-
nerability metrics as described by Luers et al. (2003) and the general idea of Total Regional
Technological Resilience (TRTR) which has been adapted from Rose (2004).
To develop our factor space we return to the  ndings from sub-section 6.5.1 which allows
us to establish baseline input values for our model (Table 6.21). We proceed on the basis
that this con guration is representative of a stable, productive system from which will vary
environmental con guration in order to induce periodic system stress.
Thus we can de ne the factors of this sub-model as an extension of the prior experi-
mental framework with two additional stress factors. We de ne system stress as the periodic
occurrence of negative externalities on our system. From the ecological parlance we choose
130
Table 6.22 Factorial Deisgn of THESIM-ER.
the ?drought? metaphor, in which government support for technological research is reduced
by a static factor (50%) for a set period of time, For this purpose we introduce two new
environmental control factors to our model:
 Disturbance Frequency, Df { How often environmental stress occurs in a single run.
 Disturbance Duration, Dl { How long (in rounds) a typical disturbance lasts.
The total time of disturbance becomes a useful catch-all expression to describe change
in overall system stress. (Dt) can be found by the following equation:
Dt = Df  Dl Eq. 6.1
While we will use Dt to describe overall system stress, it should be noted that unpacking
this function into its constituent components helps di erentiate between scenarios. For
example a scenario with high disturbance frequency and low duration may have a similar
Dt, but engender vastly di erent internal behavior.
To evaluate model output, we will focus on the total number of technology transfers (R8)
and rede ne it as our yield, Y . To build a general resilience metric, we analyze the ability
of the system to absorb shock and mitigate stress. Rose (2004) provides several  exible
resilience metrics which we paraphrase for this work. Direct regional technological resilience
(DRTR) is the extent to which the estimated direct output reduction deviated from the
131
likely maximum, which is the equivalent to the percentage technology transfer disruption in
a linear approach to baseline estimation of model output:
DRTR = % DQ000 % DQ% DQ000 Eq. 6.2
Where %  DQ000 is the maximum percentage change in direct model output with respect
to it?s optimal conditions (as provided in Table 6.21) and %  DQ is the estimated percent
change in direct output. Estimates are based on a linear  t of simulation output
DRTR gives a general resilience rating for an individual simulation runs which we modify
to provide a generalized evaluation measure for technological resilience. Our measure of total
regional technological resilience (TRTR) to environmental disruption of funding allotment
is described by the following equation:
TRTR = % TQ000 % TQ% TQ000 Eq. 6.3
Where %  TQ000 is the maximum percent change in total output and %  TQ is the
estimated percent change in total output. Thus we can provide a holistic response metric
that can give us an approximation of our model?s behavior that can be compared to future
re nements or new implementations.
6.7.1 THESIM-ER Analysis & Summary
The model was run for each orthogonal permutation of Df and Dl 30 times each and
their yield output was compared to a general compared to the linear model of expected
yield, Y?. This operation resulted in 27 independent data-points for Y (Figure 6.11), with
additional materials available in the appendix.
This initial view provides us con dence that the system?s reaction to disturbance falls
within believable bounds. Signi cantly, we see the expected shift in response on Y with
respect to increased funding scarcity. The most relevant observation from this data set
con rms the maximum yield at the lowest disturbance levels, both in terms of overall time
as well as frequency and duration.
132
Figure 6.11: Linear- t of Y under increasing system stress.
Figure 6.12: Y with respect to Df.
133
Figure 6.13: Y with respect to Dl.
A more interesting set of features can be observed when we separate the yield response
with respect to disturbance frequency. We see here that there are clear response bounds that
delineate the reaction of the system under low, medium, and high-stress con guration. Each
frequency class displays regimes of e ectiveness which overlap at regular intervals. While
the plots are not remarkably di erent, it becomes clear that the model is more vulnerable
to disturbances which are infrequent but longer in duration those that are shorter and more
frequent, even in the case where the total length of disturbance is identical. This can be
seen even more clearly in the individual plots of technology transfers (Figure 6.14a-c).
Finally, we devleop a  nal calculation of TRTR, giving us an estimated simulation
resilience threshold of 0:405421, or 40:54%. This threshold represents the the upper stress-
bound of the base simulation model after which negative regime change is likely to occur.
This can generally be associated with percentage of time Dl when the system is under stress.
In a vacuum, this value is relatively unremarkable. However it does provide a baseline for
cross-policy comparison in future experimentation and more critically allows model output
134
(a) Average technology value for high frequency, short duration disturbance
(b) Average technology value for medium frequency, medium duration disturbance condi-
tions.
(c) Average technology value for low frequency, high duration disturbance conditions
Figure 6.14: Overview of disturbance characteristics across various combinations of Df and
Dl.
135
to be directly compared with empirical studies which utilize direct regional economic re-
silience (DRER),indirect regional economic resilience (IRER), and total regional economic
resilience (TRER) with signi cantly less transformational work than existing innovation
simulations (Rose, 2004).
6.8 Chapter Summary
In this chapter we developed a series of experiments based on our simulation model with
the intention of validating our core hypothesis:
\The development and maintenance of innovation economies can be explained on the
basis of an ecologically driven agent-based model of the triple-helix model of public-private
partnership."
In order to do so we proceeded with the construction of a formal Design of Experiments
with the paired goals of quantifying our simulation model?s performance with respect to a
minimum set of input parameters and establishing a reusable resilience metric for facilitating
comparisons between innovation simulations and empirical studies. To accomplish the  rst
goal, we performed a sensitivity analysis and reduced our parameter set to a tractable set of
variables that had the broadest impact on the response variables of interest to policy makers.
Further re nement led us to the discovery that number of university?s present in our
model had the strongest impact on innovation production and industrial productivity of
the entire set of input parameters present. However, the impact of increasing this variable
resulted in a bifurcation in the response data, with larger university populations exerting
signi cant negative competitive pressure on one another and eventually sti ing growth eco-
nomic and academic growth. This  nding suggests that while increasing the number and
maturity of academic institutions within an innovation ecosystem leads to robust economic
gains, there is a maximum carrying capacity within the novelty production sector of the
triple helix that when exceeded turns the system over into negative terrain.
136
Following the sensitivity analysis we conducted a study of the ecological impact of en-
vironmental stress on the resilience of our model with the express aim to develop an output
metric that would allow future innovation ecosystem simulations to be directly compared
and contrasted with each other and empirical studies. This preliminary experiment intro-
duced several new environmental input factors to simulate system stress through a ?drought?
mechanic. The model was tested under a set of independent stress conditions the results
were compared linear model of expected system output. This comparison resulted in the
computation of total regional technological resilience (TRTR), an adapted variant of a total
regional economic resilience (TRER), a metric proposed by Rose (2004) for the study of
regional economic system resilience in the face of natural disaster. A TRTR index value
was computed from the experimental results that compared favorably with results from em-
pirical studies. Validation of this technique will rely on future studies that perform direct
comparisons between real innovation systems and computational models.
137
Chapter 7
Conclusion & Recommendations for Future Work
In this work we present an agent-based computational model of the triple helix innova-
tion ecosystem supported by formal sensitivity analysis and eclogical experimentation. Our
model leverages the ecoogical paradigm and ant-colony optimization toward the development
of lightweight agent models, as well as presenting a novel environmental representation of
the government grant system. This model supports the contention that ecological modeling
is a useful methodology for developing tractable agent-based simulation models of systems
whose complexity de es traditional techniques. Our contribution to the  eld come in three
parts. First, we present a holistic model of the regional innovation ecosystem that encom-
passes all three control regimes of the triple helix and at multiple levels of scale. Second, we
introduce the cooperative and competitive activity of individual grant-seeking researchers
within a physical environment and connect to innovaiton development and trade network
between academia and private industry. Third, we implemented a simulation tool based
on model discussed above and demonstrated that the behavior of the system is consistent
with both prior simulation e orts as well as empirically-supported relationships between the
control regimes of innovation systems.
The development and testing of THESIM-REPAST satis es both of our initial goals,
namely that a computational model of the triple helix innovation system does indeed produce
results that are both consistent, dynamical, and sensitive to environmental factors. We  nd
that the productivity triple-helix innovation systems is intrinsically tied to the maturiy and
size of its population of universities. This relationship is bounded by the economic resources
available to the system as well as inter- rm competition and overall risk-aversity. The result
of our sensitivity analysis allowed us to determine an optimal-performance parameter set
138
for use as a benchmark (control) in future experimentation and suggested several interesting
sub-model interactions that could spur future research (section 6.6).
With respect to the ecological metaphor, we made several intriguing  ndings. The
most interesting of these  ndings can be found in section 6.7, where we introduce a new
metric for calculating and innovation system?s resilience to environmental disturbance; Total
Regional Technological Resilience (TRTR). Aside from exposing the critical threshold for
our simulation model?s performance, this metric can be used in future studies to compare
 ndings between other simulation models and emprical data.
As we summarize above, the rigorous e ort applied to executing a Repast variant of
the simulation model has been signi cant, and has yielded interesting results that appear
to instill con dence in our initial hypothesis that ecological modeling could be applied to
the creation of future agent-based simulations of complex, adaptive socio-economic systems.
Through this work we have demonstrated an understanding of the problem domain, made
the case that ecological concepts are indeed valuable tools which have applications beyond
the modeling of biological systems, and provided a novel simulation application with which
to test the previous assertions.
From its inception, THESIM was conceived as a long-term project that would start
with the development and deployment of the initial simulation tool as a proof-of-concept and
would continue to be expanded in scope and detail. Two avenues for future development
remain; the refactoring of the simulation to incorporate the testing of speci c innovition-
supporting policies, and the development of more robust simulation visualization and analysis
tools.
The notion of direct policy testing was proposed as part of the model in chapter 4, though
only partially implemented (model input parameters set both environmental con guration
and agent behavior). To more fully realize this concept we propose the development of
explicit policy objects, which consist of a set of environmental and behavior schema which
may alter the rules, composition, and activity of agents. These objects would derived from
139
science-of-science liteature and be modi able by policy makers using an outside tool (or policy
builder interface). The simulation model would then be altered to require both environmental
parameters (as it currently operates), and a policy (or set of policies) upon instantiation.
This expansion would provide a more robust and useful tool to decision-makers, as was the
initial goal of the work.
Providing clearer data visualization may help address this, as well as more fully devel-
oping the social metrics such as measuring cooperation and innovation contribution through
social network graphs. While our model does provide a  nal graph representation of the
system at the end of a simulation run, we did not include any rigourous testing of network
composition. Despite this, we have made some preliminary e orts to extend THESIM?s data
analysis and visualization capabilities in order to provide graph-centric capabilities.
THESiM-DX represents one such e ort, and provides a force-directed view of the socio-
economic structure which results from individual experimental runs. This secondary tool was
developed in python and utilizes a custom mash-up of two software package; NetworkX and
D3. NetworkX is a python software package which provides a simple interface for the creation,
manipulation, and study of the structure, dynamics, and functions of complex networks. D3
is a javascript library used to build data-driven documents in web interfaces. Additionally,
a Java exporter was developed to convert raw graph data from THESIM-REPAST into
a format readable in python. This new tool, parses simulation into a NetworkX graph,
performs basic network analysis operation, and displays the relational structure of the data
in a force-directed graph. The output of THESIMDX is stored in a custom graphml format
which is readable by most open-source graph analysis software packages. An example of this
 le output can be found in the appendix. Sample output of THESIMDX?s interactiveforce-
directed graph view can be seen in  gure 7.1.
The design and development of experimental cases which test con gurations, rules, and
parameters of relevance to policy makers (or drawn from known data sets related to real-
world RSIs) will be critical to determining the value of this model. In addition, we can
140
Figure 7.1: Sample output of THESIMDX?s force-directed graph running in a web browser.
use well-de ned experiments to answer questions of reproducibility and consistency across
sub-models as well as better identify incongruence between our assumptions and actual sim-
ulation behavior. Many of the sub-models are relatively abstract interpretation of complex
systems whose validity should be individually con rmed before broader assumptions about
correctness can be made about the simulation model. Experiments should be designed which
target critical parameters relevant to these systems to determine the model?s sensitivity.
The advantage of the agent-based approach is the capability to implement sub-models at
varying scales of complexity. It might become useful to implement more robust market mod-
els, social interaction mechanisms, or environmental features. Re ning or expanding these
sub-models could be useful for more closely integrating the simulation tool to a particular
problem or set of explicit experimental conditions. For example, implementing a cooperation
model between industrial  rms which allows them to form private research consortiums, or
introducing work-force dynamics into the productivity of private  rms.
141
Introducing concepts from cognitive science, risk assessment, or game theory might prove
to be both fruitful and interesting expansions to the model. Though we have selected species
diversity and richness as metrics for substantiating our claim that innovation ecosystems
display the same markers of ecological resilience as are found in natural ecologies, it is
possible to further expand this analogy in our model by borrowing concepts of evolutionary
biology and predation rules. It is also the intention of this project to pursue reproducibility
standards in both model speci cation and data-collection. To this end, providing an online
resource where the source code and sample data sets generated by the simulation could
also be useful extensions of the project. In all, we believe that this project will not only
yield scienti cally signi cant data, but also provide context for further exploration of the
agent-based ecological modeling paradigm.
In addition, the original problem of creating a tool useful to policy investigators remains.
While the current Repast model is certainly a useful starting point for future development, it
is critical to obtain the input of public-policy makers, academics, and professionals involved
in the domain of innovation and technology to determine what capabilities would be most
useful for generating data relevant to addressing real-world problems. This objective will
likely prove di cult to achieve, namely because the  eld itself is relatively immature and
there is little agreement in literature regarding the speci c capabilities desired in future
tools. For the moment, future model development will focus on broad-level interpretations
of systems, mechanisms, and parameters. In all, we believe that the THESIM project o ers a
strong contribution to the domains of computational modeling, agent simulation, economics,
and innovation science.
142
Bibliography
[1] Akaishi, J., Arita, T. 2001. Multi-agent Simulation Showing Adaptive Property of Mis-
perception
[2] Albino, V., Carbonara, N., Giannoccaro, I., 2004 Innovation in industrial districts: An
agent-based simulation model, International Journal of Production Economics, Volume
104, Issue 1, Strategic Issues and Innovation in Production Economics, November 2006,
Pages 30-45.
[3] Altenberg, L., 1994. Evolving better representations through selective genome growth,
Proceedings of the IEEE World Congress on Computational Intelligence (1994) 182{187.
[4] Alupoaei, S., Katkoori, S. 2004. Ant colony system application to marcocell overlap
removal. IEEE Trans Very Large Scale Integr (VLSI) Syst 2004;12(10):1118{22.
[5] Antona, M., Bousquet, F., Le Page, C., Weber, J., Karsenty, A., Guizol, P., 1998.
Economic theory of renewable resource management: a multi-agent system approach
[6] Archibugi, D. and Lundvall, B.-. 2001, The globalising learning economy: Major socio-
economic trends and European innovation policy, Oxford:Oxford University Press.
[7] Arditi R. and Saiah H. 1992. Empirical evidence of the role of heterogeneity in ratio-
dependent consumption, Ecology, 73, pp. 1544-1551, 1992.
[8] Arditi R., Akcakaya H. R. 1990. Underestimation of mutual interference of predators,
Oecologia, 83, pp. 358-361, 1990.
[9] Arditi R., Berryman A. A. 1991. The biological control paradox, Trends Ecol. Evol., 6,
p. 32, 1991. 7. R.
[10] Arditi R., Ginzburg L. R., 1989 Coupling in predator-prey dynamics: ratio-dependence,
J. Theor. Biol., 139, pp. 311-326
[11] Arditi, L. R. Ginzburg, L. R., Akcakaya, H. R. 1991. Variation in plankton densities
among lakes: a case for ratio-dependent models, Am. Nat., 138, pp. 1287-1296, 1991.
[12] Argyres, N. S., Liebeskind, J. P.1999. Contractual Commitments, Bargaining Power,
and Governance Inseparability: Incorporating History into Transaction Cost Theory.
The Academy of Management Review Vol. 24, No. 1 (Jan., 1999), pp. 49-63
143
[13] Assad, A. M. and Packard, N.H. (1992) Emergent colonization in an arti cial ecol-
ogy, in Toward a Practice of Autonomous Systems: Proceedings of the First European
Conference on Arti cial Life, (ed.) F.J. Varela and P. Bourgine, 143-152, MIT Press,
Boston.
[14] Axtel, R. L., Andrews, C. J., Small, M. J., 2000. Agent-Based Models of Industrial
Ecosystems.
[15] Azadeha, A., Skandarib, M. R., Maleki-Shojaa, B. 2009. An integrated ant colony op-
timization approach to compare strategies of clearing market in electricity markets:
Agent-based simulation
[16] Babaoglu, O., Canright, G., Deutsch, A., Di Caro, G. A., Ducatelle, F., Gambardella,
L.M., Ganguly, N., Jelasity, M., Montemanni, R., Montresor, A., Urnes, T. 2006. Design
patterns from biology for distributed computing.ACM Trans Auton Adapt Syst1: pp.
26{66
[17] Balci, O. 1989. How to assess the acceptability and credibility of simulation results.
In Proc. 1989 Winter Simulation Conf., ed. MacNair, E. A., Musselman, K. J., and
Heidelberger, P. 62-71. Piscataway, New Jersey.
[18] Baray, C. 1998. E ects of population size upon emergent group behavior. Complex. Int.
6,
[19] Bautista, J., Pereira, J. 2002. Ant algorithms for assembly line balancing. In: Dorigo,
M., Di Caro, G., Sampels, M., editors. Ant algorithms|Proceedings of ANTS 2002|
Third international workshop. Lecture Notes in Comput Sci, vol. 2463. Berlin: Springer;
2002. p. 65{75.
[20] Beckenbach F., Briegel R., Daskalakis M. 2008. Behavioral foundation and agent-based
simulation of regional innovation dynamics
[21] Beddington J. R. 1975 Mutual interference between parasites or predators and its e ect
on searching e ciency, J. Anim. Ecol., 45, pp. 331-340, 1975.
[22] Bell, M., Pavitt, K. 1993. Accumulating Technological Capability in Developing Coun-
tries. Proceedings of the World Bank Annual Conference on Development Economics -
1992 (a supplement to The World Bank Economic Review and World Bank Research
Observer). Washington: World Bank, 1993.
[23] Belussi F. 2005. Industrial district/local production systems as hyper-networks: a neo-
Marshallian interpretative frame, in Guidi M. and Parisi D. (eds.), The Changing Firm.
Contributions from the History of Economic Thought, Franco Angeli, Milan, p. 385-418.
[24] Berger, T. 2001. Agent-based spatial models applied to agriculture: a simulation tool for
technology di usion, resource use changes and policy analysis, Agricultural Economics,
Volume 25, Issues 2-3, September 2001, Pages 245-260
144
[25] Bernstein C., Auger P., Poggiale J. C. 1999. Predator migration decisions, the ideal free
distribution and predator-prey dynamics, Am. Nat., 153, pp. 267-281.
[26] Bianchi, L., Gambardella, L. M., Dorigo, M. 2002. An ant colony optimization approach
to the probabilistic traveling salesman problem. In: Merelo, J. J., Adamidis, P., Beyer,
H-G., Fernndez-Villacanas, J-L., Schwefelm H-P., editors. Proceedings of PPSN-VII,
seventh international conference on parallel problem solving from nature. Lecture Notes
in Comput Sci, vol. 2439. Berlin: Springer; 2002. p. 883{92.
[27] Blaine T. W., DeAngelis D. L. 1997. The e ects of spatial scale on predator-prey func-
tional response, Ecol. Model., 95, pp. 319-328.
[28] Blesa, M., Blum, C. 2004. Ant colony optimization for the maximum edge-disjoint paths
problem. In: Raidl, G. R., Cagnoni, S., Branke, J., Corne, D. W., Drechsler, R., Jin, Y.,
Johnson, C. G., Machado, P., Marchiori, E., Rothlauf, R., Smith, G. D., Squillero, G.,
editors. Applications of evolutionary computing, proceedings of EvoWorkshops 2004.
Lecture Notes in Comput Sci, vol. 3005. Berlin: Springer; 2004. p. 160{9.
[29] Blum, C. 2005. Beam-ACO|Hybridizing ant colony optimization with beam search: An
application to open shop scheduling. Computers & Operations Res. 2005; 32(6):1565{91.
[30] Blum, C., Dorigo, M. The hyper-cube framework for ant colony optimization. IEEE
Trans Syst Man Cybernet Part B 2004;34(2):1161{72.
[31] Blum, C., Socha, K. 2005. Training feed-forward neural networks with ant colony opti-
mization: An application to pattern classi cation. HIS, pages 233{238.
[32] Bonebeau E, Theraulaz G, Aron S, Camazine S. 1997 Self organization of social insects.
Trends in Ecol Evol, 1997, 12:188-93
[33] Bousquet, F. Cambier, C., Morand P., Quensire J., Mullon, C. A. P. 1993. Simulating
the interaction between a society and a renewable resource.
[34] Bousquet, F., Le Page, C. 2003. Multi-agent simulations and ecosystem management:
a review
[35] Brauer, F., Castillo-Chavez, C. 2000. Mathematical Models in Population Biology and
Epidemiology, Springer-Verlag,
[36] Briscoe G., 2011. Complex Adaptive Digital EcoSystems
[37] Brooks, R. A. 1991. \Intelligence Without Representation", Arti cial Intelligence 47
(1991) 139-159.
[38] Bui, T. N., Rizzo, J. R. 2004. Finding maximum cliques with distributed ants. In: Deb
K, et al., editors. Proceedings of the genetic and evolutionary computation conference
(GECCO 2004). Lecture Notes in Comput Sci, vol. 3102. Berlin: Springer; 2004. p.
24{35.
145
[39] Bullnheimer, B., Hartl, R., Strauss, C. 1999. A new rank-based version of the Ant Sys-
tem: A computational study. Central European J Operations Res Econom 1999;7(1):25{
38.
[40] Butel, L., Watkins, A. 2006. Clusters of entrepreneurs: The application of ant colony
optimisation modelling, Journal of Modelling in Management, Vol. 1 Iss: 3, Pp.255 {
269
[41] Callon, M. 1991. \Techno-economic networks and irreversibility." Pp. 132-165 in A
Sociology of Monsters: Essays on Power, Technology and Domination, edited by John
Law. London: Routledge.
[42] Carley, K. M. 1996. Validating computational models. Working paper, Carnegie Mellon
University, Pittsburgh, PA.
[43] Cassiman B., Veugelers R. 2002. R&D cooperation and spillovers: some empirical evi-
dence from Belgium, American Economic Review 92 (2002) (4), pp. 1169-1184.
[44] Charles, A. T. 1988. Fishery socioeconomic: a survey. Land Economics, 64, 276-295.
(Charles, 1988)
[45] Chen, L., Shen, J., Qin, L., Fan, J. 2002. A method for solving optimization problem
in continuous space using improved ant colony algorithm. In CASDMKM, pages 61{70,
2004. Previously published in 2002.
[46] Cherrett, J. M. 1988. Key concepts: the results of the survey of our member?s views.
Bulletin of the British Ecological Society
[47] Cohen, L., Noll, R. 1991. The Technology Pork Barrel. Washington, D.C.: The Brook-
ings Institution.
[48] Cohen, W., Nelson, R., Walsh, J. 2000 Protecting Their Intellectual Assets: Appropri-
ability Conditions and Why US Manufacturing Firms Patent (or Not), Working Paper
N. 7552, National Bureau of Economic Research, Cambridge, US, 2000.
[49] Collings J. B. 1997 The e ects of the functional response on the bifurcation behaviour
of a mite predator-prey interaction model, B. Math. Biol., 36, pp. 149-168.
[50] Colroni, A., Dorigo, M., & Maniezzo, V. 1991. Distributed Optimization by Ant
Colonies. In proceedings of ECAL91 { European Conference on Arti cial Life, Else-
vier Publishing, pp. 134-142.
[51] Commins, H. N., Hassell, M. P. 1996. Persistence of multispecies host-parasitoid inter-
actions in spatially distributed models with local dispersal. J. Theo. Biol. 183 (1996),
19{28.
[52] Cook, P. 2001. Regional Innovation Systems, Clusters, and the Knowledge Economy.
Industrial and Corporate Change, Vol. 10, No. 4
146
[53] Cooke, P., Uranga, M. G., Etxebarria, G. 1997. Regional innovation systems: Institu-
tional and organisational dimensions. Research policy 26.4 p. 475-491.
[54] Cooper, C. 1992. Are Innovation Studies on Industrialized Economies Relevant to Tech-
nology Policy in Developing Countries? In UNU/INTECH Working Paper No. 3, June
1992. Maastricht, Netherlands: The United Nations University, Institute for New Tech-
nologies (mimeo).
[55] Corloni A., Dorigo M., Maniezzo V., Trubian M. 1993. Ant system for job-shop schedul-
ing. Belg J Operation Resources, Sat Computer Science.
[56] Corry, P., Kozan, E. 2004. Ant colony optimisation for machine layout problems. Com-
put Optim Appl 2004;28(3):287{310.
[57] Cosner C., DeAngelis D. L. 1999. J. S. Ault, D. B. Olson, E ects of spatial grouping on
the functional response of predators, Theor. Popul. Biol., 56, pp. 65-75.
[58] Costa, D., Hertz, A. 1997. Ants can color graphs. J Oper Res Soc 1997; 48:295{305.
[59] Costanza R, Waigner L, Folke C, Maler, K-G. 1993. Modeling complex ecological eco-
nomic systems: towards an evolutionary dynamic understanding of people and nature
[60] Dachs, B., Ebersberger, B and Pyka, A. 2008. Why do Firms Co-operate for Innovation?
A comparison of Austrian and Finnish CIS 3 results. International Journal of Foresight
and Innovation Policy, Vol. 4, Number 3-4, pp.200-229.
[61] Dai. W., Chen, M., Ye, N. 2007. Research on the Formation of Innovation Ecosystem
in Software Industry and Its Development Strategy. Proceedings of the Sixth Wuhan
International Conference on E-Business 2211-2216.
[62] Dawid, Herbert, Reimann, Marc and Bullnheimer, Bernd (2001). To Innovate or Not
to Innovate? IEEE Transactions on Evolutionary Computation (special issue on Agent-
Based Computational Economics) 5 (5): 471{481.
[63] de Mottoni, P., Rothe, F. 1979. Convergence to homogeneous equilibrium state for
generalized Volterra-Lotka systems with di usion. SIAM J. Appl. Math. 37 (1979), no.
3, 648{663.
[64] DeAngelis D. L., Goldstein R. A., O?Neill R. V., 1975. A model for trophic interaction,
Ecology, 56, pp. 881-892, 1975.
[65] DeAngelis, D.L., Rose, K.A. and Huston, M.A. 1994. Individual-oriented approaches
to modelling ecological populations and communities. In Levin, S.A. (ed.): Frontiers in
Mathematical Biology. Springer, Berlin. 390-410.
[66] Deb, K., Thiele, L., editors. Proceedings of the second international conference on
evolutionary multi-criterion optimization (EMO 2003).
[67] De uant, G., 2001. Improving agri-environmental policies: a simulation approach to the
cognitive properties of farmers and institutions
147
[68] den Besten, M. L., Sttzle, T., Dorigo, M. 2000. Ant colony optimization for the total
weighted tardiness problem. In: Schoenauer, M., Deb, K., Rudolph, G., Yao, X., Lutton,
E., Merelo, J. J., Schwefel, H.-P., editors. Proceedings of PPSN-VI, sixth international
conference on parallel problem solving from nature. Lecture Notes in Comput Sci, vol.
1917. Berlin: Springer; 2000. p. 611{20.
[69] Deneubourg, J.L., Goss, S., Franks, N.R. and Pasteels, J.M. 1989. The blind leading
the blind:modelling chemically mediated army ant raid patterns. Journal of Insect Be-
haviour. 2: 719-725.
[70] Dinopoulos E., Segerstrom, P. S. 1999. \A Schumpeterian Model of Protection and
Relative Wages", American Economic Review 89(3): 450-472.
[71] Dinopoulos, E., Thompson, P. 1999. \Scale E ects in Schumpeterian Models of Eco-
nomic Growth", Journal of Evolutionary Economics 9(2): 157-185.
[72] Doerner, K., Gutjahr, W. J., Hartl, R. F., Strauss, C., Stummer, C. 2004. Pareto ant
colony optimization: A metaheuristic approach to multiobjective portfolio selection.
Ann Oper Res 2004;131:79{99.
[73] Dongsheng, Y., Yongan, Z., 2008. Innovation Based on Multi-Agent Method. Interna-
tional Conference on Computer Science and Software Engineering 2008.
[74] Doo, M., Sohn, S 2008. Productivity Improvement of Manufacturing SMEs via Technol-
ogy Innovation in Korea, 7th WSEAS Int. Conf. on Arti cial Intelligence, Knowledge
Engineering and Daabases (AIKED?08), University of Cambridge, UK, Feb 20-22.
[75] Dorigo, M, Gambardella, L. 1997. Ant Colony System: A Cooperative Learning Ap-
proach to the Travelling Salesman Problem, IEEE Transactions on Evolutionary Com-
putation, 1:1. pp. 53-66.
[76] Dorigo, M. 1992. Learning and Natural Algorithms. Doctoral Dissertation, Politecnico
di Milano, Italie.
[77] Dorigo, M., Bonabeau, E., Theraulaz, G., 1999. Swarm Intelligence.
[78] Dorigo, M., Maniezzo, V., & Colorni, A. 1996. The Ant System: Optimization by a
Colony of Cooperating Agents. IEEE Transactions on Systems, Man, and Cybernetics-
Part B, Vol. 26, No. 1, pp. 1-13.
[79] Dorigo, M., Stutzle, T. 2004. Ant Colony Optimization, MIT Press, Cambridge, MA.
[80] Downs, G. W., Mohr, L. B. 1976. Conceptual issues in the study of innovation. Admin-
istrative Science Quarterly, Vol. 21, 1976. Pp. 700-714.
[81] Dro, J. 2001. Modlisation de la mobilisation chez les fourmis, Mmoire de DEA, Universit
Paris7 & Universit Libre de Bruxelles, 2001.
148
[82] Dro, J., Siarry, P. 2002. A new ant colony algorithm using the heterarchical concept
aimed at optimization of multiminima continuous functions. In M. Dorigo, G. D. Caro,
and M. Sampels, editors, Ant Algorithms, Third International Workshop, ANTS 2002,
Brussels, Belgium, September 12-14, 2002, Proceedings, volume 2463 of Lecture Notes
in Computer Science, pages 216{221. Springer, 2002.
[83] Dro, J., Siarry, P. 2004. Continuous interacting ant colony algorithm based on dense
heterarchy. Future Generation Computer Systems, 20(5):841{856, June 2004.
[84] Drogul A., Ferber, J. 1992. Multi-Agent Simulation as a Tool for Modeling Societies:
Application to Social Di erenciation in Ant colonies. In Arti cial Social Systems, ed.
Cristiano Castelfranchi et Eric Werner. 3-23. 830. Berlin
[85] Edquist, C.1997. Systems of Innovation: Technologies, Institutions, and Organizations.
A Cassel imprint. Wellington House, London. eds., Technology, R&D, and the Economy.
Washington, D.C.: The Brookings Institution and The American Enterprise Institute,
1996.
[86] Ferguson, C.E., 1969. The Neoclassical Theory of Production and Distribution. Cam-
bridge. Description, ch. 1 excerpt, pp. 1-10
[87] Folke, C., 2006. Resilience: The emergence of a perspective for social-ecological systems
analyses.
[88] Forfs, A., 2004. Innovation Networks, The National Policy and Advisory Board for
Enterprise, Trade, Science, Technology And Innovation; Dublin, June 2004.
[89] Freeman, C. 1987. Technology Policy and Economic Performance: Lessons from Japan,
Pinter, London.
[90] Freeman, C. 1995. The National System of Innovation in Historical Perspective, Cam-
bridge Journal of Economics, No. 19, pp. 5-24
[91] Freeman, C. 2002. Continental, National and Sub-national Innovation Systems - Com-
plementarity and Economic Growth, Research Policy 31, 191-211.
[92] Freeman, C., Perez, C. 1988. Structural Crises of Adjustment, Business Cycles and
Investment Behaviour. in DOSI et al., eds., pp.38-66
[93] Fuentes-Fern andez, R., Hassan, S., Pav on, J., Gal an, J. M., L opez-Paredes, A. 2012.
Metamodels for role-driven agent-based modelling, Computational & Mathematical Or-
ganization Theory, vol.18 n.1, p.91-112, March 2012
[94] Gagn, C., Price, W. L., Gravel, M. 2002. Comparing an ACO algorithm with other
heuristics for the single machine scheduling problem with sequence dependent setup
times. J. Oper. Res. Soc. 2002;53:895{906.
149
[95] Gambardella L. M., Dorigo M. 1995. Ant-Q: A reinforcement learning approach to the
traveling salesman problem. In: Prieditis A., Russel S., editors. Proceedings of the ML-
95, 12th International Conference on Machine Learning, 1995; Morgan Kaufmann, San
Francisco, CA. p. 252-60.
[96] Gambardella, L. M., Dorigo, M. 1996. Solving symmetric and asymmetric TSPs by ant
colonies. In: Baeck, T., Fukuda, T., Michalewicz, Z., editors. Proceedings of the 1996
IEEE international conference on evolutionary computation (ICEC?96). Piscataway, NJ:
IEEE Press; 1996. p. 622{7. [41] Gambardella, L. M., Dorigo, M. Ant Colony System
hybridized with a new local search for the sequential ordering problem. INFORMS J
Comput
[97] Gambardella, L. M., Taillard, . D., Agazzi, G. 1999 MACS-VRPTW: A multiple ant
colony system for vehicle routing problems with time windows. In: Corne. D., Dorigo,
M., Glover, F., editors. New ideas in optimization. London: McGraw-Hill; 1999. p.
63{76.
[98] Gandibleux, X., Delorme, X., T?Kindt, V. 2004. An ant colony optimisation algorithm
for the set packing problem. In: Dorigo, M., Birattari, M., Blum, C., Gambardella, L.
M., Mondada, F., Sttzle, T., editors. Proceedings of ANTS 2004|Fourth international
workshop on Ant colony optimization and swarm intelligence. Lecture Notes in Comput
Sci, vol. 3172. Berlin: Springer; 2004. p. 49{60.
[99] George J. Stigler 1941, 1994. Production and Distribution Theories. New York: Macmil-
lan.
[100] Gibbons, Michael; Camille Limoges, Helga Nowotny, Simon Schwartzman, Peter Scott,
& Martin Trow (1994). The new production of knowledge: the dynamics of science and
research in contemporary societies. London: Sage
[101] Giovanis, A. N., Skiadas, C. H. 1999 A Stochastic Logistic Innovation Di usion Model
Studying the Electricity Consumption in Greece and the United States, Technological
Forecasting and Social Change, Volume 61, Issue 3, July 1999, Pp. 235-246.
[102] Goldberg, D. E. 1998. The Race, the Hurdle, and the Sweet Spot: Lessons from Genetic
Algorithms for the Automation of Design Innovation and Creativity. IlliGAL Report No.
98007. April, 1998.
[103] Goldwyn, E. E., Hastings, A. 2008. When can dispersal synchronize populations?
Theo. Popu. Biol. 73 (2008), 395{402.
[104] Goldwyn, E. E., Hastings, A. 2009 Small heterogeneity has large e ects on synchro-
nization of ecological oscillators. Bulletin of Mathematical Biology 71 (2009), 130{144.
[105] Gollmitzer, M., Murray, C. 2008, From Ecology to Economy: a policy framework for
creative labour. Canadian Conference of the Arts
[106] Goodwin, R.M. 1967. \A Growth Cycle", Socialism, Capitalism and Economic Growth,
Feinstein, C.H. (ed.), Cambridge University Press,
150
[107] Gottlieb, J., Puchta, M., Solnon, C. 2003. A study of greedy, local search, and ant
colony optimization approaches for car sequencing problems. In: Cagnoni, S., Romero
Cardalda, J. J., Corne, D. W., Gottlieb, J., Guillot, A., Hart, E., Johnson, C. G.,
Marchiori, E., Meyer, J.-A., Middendorf, M., Raidl, G. R., editors. Applications of
evolutionary computing, proceedings of EvoWorkshops, 2003. Lecture Notes in Comput
Sci, vol. 2611. Berlin: Springer; 2003. p. 246{57.
[108] Grimm, V. 1999. Ten years of individual-based modelling in ecology: what have we
learned and what could we learn in the future? Ecological Modelling. 115: 129-148.
[109] Grossman, G. M. and Helpman, E. 1991. Innovation and Growth in the Global Econ-
omy, Cambridge, MA, MIT Press.
[110] Guret, C., Monmarch, N., Slimane, M. 2004. Ants can play music. In: Dorigo M,
Birattari M, Blum C, Gambardella LM, Mondada F, Sttzle T, editors. Proceedings of
ANTS 2004|Fourth international workshop on Ant colony optimization and swarm
intelligence. Lecture Notes in Comput Sci, vol. 3172. Berlin: Springer; 2004. p. 310{7.
[111] Guntsch, M., Middendorf, M. 2001. Pheromone modi cation strategies for ant algo-
rithms applied to dynamic TSP. In: Boers, E. J. W., Gottlieb, J., Lanzi, P. L., Smith,
R. E., Cagnoni, S., Hart, E., Raidl, G. R.
[112] Guntsch, M., Middendorf, M. 2003. Solving multi-objective permutation problems with
population based ACO. In: Fonseca, C. M., Fleming, P. J., Zitzler, E.,
[113] Gutierrez A.P. 1992. The physiological basis of ratio-dependent predator-prey theory:
a metabolic pool model of Nicholson?s blow ies as an example, Ecology, 73, pp. 1552-
1563, 1992.
[114] Guttman, R. H., Maes, P. 1998. Cooperative vs. Competitive Multi-Agent Negotiations
in Retail Electronic Commerce.
[115] Hanski I., The functional response of predator: worries about scale, Trends Ecol. Evol.,
6, pp. 141-142, 1991.
[116] Haque M., Li B. 2008. A ratio-dependent predator-prey model with logistic growth for
the predator population, in: Proceedings of 10th International Conference on Computer
Modeling and Simulation, University of Cambridge, Cambridge, UK, 2008, pp. 210-215.
[117] Hardin, G. 1968. The Tragedy of the Commons. Science 13 December 1968: 162 (3859)
[118] Hassell, M. P., Varley G. C.,1969. New inductive population model for insect parasites
and its bearing on biological control, Nature, 223, pp. 1133-1137, 1969.
[119] Haynes, K. E., Mahajan, V., White, G. M. 1977. Innovation di usion: A determin-
istic model of space-time integration with physical analog, Socio-Economic Planning
Sciences, Volume 11, Issue 1, 1977, Pp. 25-29.
151
[120] Helpman, E. 1993. \Innovation, Imitation and Intellectual Property Rights", Econo-
metrica 61(6): 1247-1280.
[121] Hirshleifer, J., Glazer, A., and Hirshleifer, D. 2005. Price theory and applications:
Decisions, markets, and information. Cambridge University Press, 7th Edition: 2005.
[122] Hogeweg, P., Hesper, B. 1983. The ontogeny of the interaction structure in BumbleBee
colonies: a MIRROR model. Behav. Ecol. Sociobiol. 12: 271-283
[123] Holland, J. H., Miller, J. H. 1991. Arti cial Adaptive Agents in Economic Theory. The
American Economic Review, Vol. 81, No. 2, Papers and Proceedings of the Hundred and
Third Annual Meeting of the American Economic Association. May, 1991. Pp. 365-370
[124] Holland, M. D., Hastings, A. 2008. Strong e ect of dispersal network structure on
ecological dynamics. Nature 456 (2008), 792{795. (Holland & Hastings, 2008)
[125] Hlldobler, B., Wilson, E. O. The Ants, Springer, Berlin, 1990.
[126] Holling, C. S. 1973. Resilience and Stability of Ecological Systems.
[127] Holling, C. S. 1965. The functional response of invertebrate predators to prey density,
Mem. Ent. Soc. Can., 45 (1965), pp. 3-60.
[128] Holling, C. S. 1987. Simplifying the Complex; The paradigms of ecological function
and structure. European Journal of Operational Research 30 139-146.
[129] Holling, C. S., Folke, C., Gunderson, L., Maler, K. G. 2000. Resilience of ecosystems,
economic systems, and institutions. Final Report submitted to the John D. and Cather-
ine T. MacArthur Foundation, April 30.
[130] Holt, R. D. 1984 Spatial Heterogeneity, indirect interactions, and the coexistence of
prey species. The American Naturalist 124 (1984), no. 3, 377{406.
[131] Horsthemke, W., Moore, P. K. 2004 Turing instability in inhomogeneous arrays of
di usively coupled reactors. J. Phys. Chem. A 108 (2004), no. 12, 2225{2231.
[132] Hurmelinna, P., Puumalainen, K. 2007. Nature and Dynamics of Appropriability:
Strategies for Appropriating Returns on Innovation, R&D Management, Vol. 37, N.
2, 2007.
[133] Irwin D., Kneow, P. 1996. High-Tech R&D Subsidies: Estimating the E ects of Se-
matech. Journal of International Economics, Vol. 40 (1996), pp. 323-344.
[134] Janssen, V. A. A. 1995. Regulation of predator-prey systems through spatial interac-
tions: a possible solution to the paradox of enrichment. Oikos 74 (1995), no. 3, 384{390.
[135] Janssen, V. A. A. 2001. The dynamics of two di usively coupled predator-prey popu-
lations. Theoretical Population Biology 59 (2001), 119{131.
152
[136] Janssen, V. A. A., Carpenter, S.R., 1999. Managing the resilience of lakes: a multi-
agent modeling approach.
[137] Karpenko, O., Shi, J., Dai, Y. 2005. Prediction of MHC class II binders using the ant
colony search strategy. Arti cial Intelligence in Medicine 2005;35(1{2):147{56.
[138] Kau man, S. 1993. The Origins of Order: Self-Organization and Selection in Evolution,
Oxford University Press (1993).
[139] Kau man, S., Lobo, J., Macready, W.G., 2000. Optimal search on a technology land-
scape. Journal of Economic Behavior and Organization 43, Pp. 141{166.
[140] Kau man, S., Macready, W., 1995. Technological evolution and adaptive organizations,
Complexity 1 (1995) (2), pp. 26{43.
[141] Kefalas, P., Holcombe, M., Eleftherakis, G., Gheorghe, M. 2003. A formal method for
the development of agent-based systems. In Intelligent Agent Software Engineering, (ed.
Plekhavona, V.), Idea Group Publishing Co. 68-98.
[142] Klingert, F. M., Meyer, M. 2012. E ectively combining experimental economics and
multi-agent simulation: suggestions for a procedural integration with an example from
prediction markets research, Computational & Mathematical Organization Theory, v.18
n.1, p.63-90, March 2012
[143] Klomp, N. and Green, D. (1996) Complexity and connectivity in ecosystems, Com-
plexity International, 3.
[144] Konig, M. D., Mller, M., Pich, C., von Pro , S. 2010. Networks of Inventors. Imperial
College London Business School, June, 2010.
[145] Kreps, David M. A Course in Microeconomic Theory. Princeton University Press: 1990
[146] Krugman, P. 1979. \A Model of Innovation, Technology Transfer, and the World Dis-
tribution of Income", Journal of Political Economy 87(2): 253-266.
[147] Kurose, J. F., Simha, R. 1989. A microeconomic approach to optimal resource alloca-
tion in distributed computer systems. IEEE Transactions on Computers, 38, 705-717.
[148] Langton, C.G. 1992. Life at the edge of chaos. In: Langton, C.G., Taylor, C., Farmer,
J.D., Rasmussen, S. (Eds.), Arti cial Life II. Addison-Wesley, pp. 41{91.
[149] Lecture Notes in Comput Sci, vol. 2636. Berlin: Springer; 2003. p. 464{78.
[150] Lengyel, I., Epstein, I. R. 1991. Di udion-induced instability in chemically reacting
systems: Steady-state multiplicity, oscillation and chaos. CHAOS 1 (1991), no.1, 69{75.
[151] Lerner, J. 1999. The Government a Venture Capitalist: The Long-Run Impact of the
SBIR Program. Journal of Business Vol. 72 (1999), pp. 285-318.
153
[152] Levin, S. A., 1998. Fragile Dominion: complexity and the commons. Reading (MA):
Addison-Wesley.
[153] Levy, D., Terleckyj, N. 1983. E ects of Government R&D on Private R&D Invest-
ment and Productivity: A Macroeconomic Analysis. Bell Journal of Economics, Vol. 14
(1983), pp. 551-561.
[154] Leydesdor , L. 2000. The triple helix: an evolutionary model of innovations. Research
Policy-29. 2000., pp.243-255
[155] Leydesdor , L., Etzkowitz, H. 2005. Universities And The Global Knowledge Economy:
A Triple Helix Of University-Industry-Government Relations. Continuum International
Publishing Group.
[156] Leydesdor , L., Meyer, M. 2006. Triple Helix indicators of knowledge-based innovation
systems - Introduction to the special issue. in Research Policy Vol. 35 (2006)
[157] List, F. 1841. Das Nationale System der Politischen konomie, Basel: Kyklos, (trans-
lated and published under the title: The National System of Political Economy? by
Longmans, Green and Co., London 1841).
[158] Lopez-Ibanez, M., Paquete, L., Stutzle, T. 2004. On the design of ACO for the biob-
jective quadratic assignment problem. In: Dorigo, M., Birattari, M., Blum, C., Gam-
bardella, L. M., Mondada, F., Sttzle, T., editors. Proceedings of ANTS 2004|Fourth
international workshop on Ant colony optimization and swarm intelligence. Lecture
Notes in Comput Sci, vol. 3172. Berlin: Springer; 2004. p. 214{25.
[159] Lundvall, B. A., 1985. ?Product innovation and user-producer interaction, industrial
development?, Research Series 31, Aalborg: Aalborg University Press.
[160] Lundvall, B.-. 1992. National Systems of Innovation: Towards a Theory of Innovation
and Interactive Learning, London: Pinter Publishers.
[161] Lux, T., Marchesi, M. 1999. Scaling and criticality in a stochastic multi-agent model
of a  nancial market.
[162] Ma, T., Nakamori, Y. 2005. Agent-based modeling on technological innovation as an
evolutionary process, European Journal of Operational Research, Volume 166, Issue 3,
Advances in Complex Systems Modeling, 1 November 2005, Pages 741-755.
[163] Maiti, A., Pathak, S. 2009. A Modi ed Holling-Tanner Model in Stochastic Environ-
ment.
[164] Maniezzo, V. 1999. Exact and approximate nondeterministic tree-search procedures
for the quadratic assignment problem. INFORMS. J. Comput. 1999; 11(4):358{69.
[165] Maniezzo, V., Boschetti, M., Jelasity, M. 2004. An ant approach to membership overlay
design. In: Dorigo, M., Birattari, M., Blum, C., Gambardella, L. M., Mondada, F.,
Sttzle, T., editors. Proceedings of ANTS 2004|Fourth international workshop on Ant
154
colony optimization and swarm intelligence. Lecture Notes in Comput Sci, vol. 3172.
Berlin: Springer; 2004. p. 37{48.
[166] Maniezzo, V., Colorni, A. 1999. The Ant System applied to the quadratic assignment
problem. IEEE Trans. Data. Knowledge. Eng. 1999; 11(5):769{78.
[167] Maniezzo, V., Dorigo, M., & Colorni, A. 1994. The Ant System Applied to the
Quadratic Assignment Problem. Technical Report IRIDIA/94-28, Universit Libre de
Bruxelles, Belgium.
[168] Markus, M. L., Robey, D. 1988. \Information Technology and Organizational change:
Casual Structure in Theory and Research. Management Science, Vol. 34, No. 5, 1988,
Pp. 583-598.
[169] Mathevet, R., Bousquetb, F., Le Page, C., Antona, M. 2003. Agent-based simulations
of interactions between duck population, farming decisions and leasing of hunting rights
in the Camargue (Southern France). Ecological Modelling 165(2-3): 107-126.
[170] Mathevet, R., Lifran, R., Mauchamp, A., Lefebvre, G., Poulin, B. 2003. ReedSim:
Simulating Ecological and Economical Dynamics of Mediterranean Reedbeds
[171] May, R. M. 1974. Stability and Complexity in Model Ecosystems, Princeton University
Press, Princeton, 1974.
[172] McCormack, J. 2007. Arti cial Ecosystems for Creative Discovery, in Thierens D. et.
al. (eds) Genetic and Evolutionary Computation Conference (GECCO 2007) 7-11 July
2007, London, ACM, New York, Volume 1, pp. 301-307.
[173] Metcalfe, S. 1995. \The Economic Foundations of Technology Policy: Equilibrium
and Evolutionary Perspectives", in P. Stoneman (ed.), Handbook of the Economics of
Innovation and Technological Change, Blackwell Publishers, Oxford (UK)/Cambridge
(US).
[174] Meyer, M., Troitzsch, K. G. 2012. Epistemological perspectives on simulation: overview
and introduction, Computational & Mathematical Organization Theory, vol.18 n.1, p.1-
4, March 2012
[175] Michlmayr, E., Graf, S., Siberski, W., Nejdl, W. 2005. Query Routing with Ants. In
Proceedings of the 1st Workshop on Ontologies in P2P Communities, ESWC2005, May
2005.
[176] Miotti, L., Sachwald, F. 2003. Co-operative R&D: Why and with whom? An integrated
framework of analysis. Research Policy, 32, 8, 1481-1500.
[177] Monmarche, N., Venturini, G., Slimane, M. 2000. On how pachycondyla apicalis ants
suggest a new search algorithm. Future Gener. Comput. Syst., 16(9):937{946, 2000
[178] Moore, J. F. 1993. Predators and Prey: A New Ecology of Competition. Harvard
Business Review
155
[179] Moss, J. D,, Johnson, C. G. 2003. An ant colony algorithm for multiple sequence
alignment in bioinformatics. In: Pearson, D. W., Steele, N. C., Albrecht, R. F., editors.
Arti cial neural networks and genetic algorithms. Berlin: Springer; 2003. p. 182{6.
[180] Murray J. D. 1993. Mathematical Biology, Springer-Verlag, New York.
[181] Nadiri, I. 1993. Innovations and Technological Spillovers. NBER Working Paper no.
4423
[182] Nag, R., Hambrick, D. C.; Chen, M.-J, 2007. What is strategic management, really?
Inductive derivation of a consensus de nition of the  eld. Strategic Management Journal.
Volume 28, Issue 9, pages 935-955, September 2007.
[183] Nelson, R. R., Winter, S. G. 1982, An Evolutionary Theory of Economic Change,
Cambridge MA: Belknap Press of Harvard University Press.
[184] Nelson, R. R., Romer, P. 1996. Science, Economic Growth, and Public Policy.In Smith,
B., Bar eld, C.
[185] Norberg, J., Swaney, D. P., Dusho , J., Lin, J., Casagrandi, R., Levin, S. A. 2001.
Phenotypic diversity and ecosystem functioning in changing environments: A theoretical
framework Proc Natl Acad Sci U S A. 2001 September 25; 98(20): 11376-11381.
[186] Ostrom, E. 1990. Governing the Commons: The Evolution of Institutions for Collective
Action. Cambridge University Press.
[187] Otter, H. S., van der Veen, A., de Vriend, H. J. 2001. Abloom: Location behaviour,
spatial patterns, and agent-based modelling. Journal of Arti cial Societies and Social
Simulation, 4(4), 2001.
[188] Padgett, J.F., Lee, D., Collier, N. 2004. Economic production as chemistry. Industrial
and Corporate Change, 12:843{877, 2004.
[189] Page, S. E., Tassier, T. 2003. An ecological model of chains. University of Michigan,
Center for Political Studies, working paper
[190] Parisi. E., 1997. What do to with a surplus. Lecture notes in economics and mathe-
matical systems, 456.
[191] Parker, D.C., Manson, S.M., Janssen, M.A. 2003. Multi-agent systems for the simula-
tion of land-use and land-cover change: a review.
[192] Parpinelli, R. S., Lopes, H. S., Freitas, A. A. 2002. Data mining with an ant colony
optimization algorithm. IEEE Trans Evolutionary Comput 2002; 6(4):321{32.
[193] Parrott, L., Kok, R. 2000. Incorporating Complexity in Ecosystem Modelling.
[194] Patten, B. C. 1978. Systems approach to the concept of environment. Ohio J. Sci.
78:206-222.
156
[195] Pirolli, P. 2009. An elementary social information foraging model. In Proceed-
ings of the 27th international conference on Human factors in computing systems
(CHI ?09). ACM, New York, NY, USA, 605-614. DOI=10.1145/1518701.1518795
http://doi.acm.org/10.1145/1518701.1518795
[196] Pickett, S.T.A., Cadenasso, M.L. 2002. Ecosystem as a multidimensional concept:
meaning, model and metaphor. Ecosystems 5:1-10
[197] Poggiale J. C., Michalski J., Arditi R.1998. Emergence of donor control in patchy
predator-prey systems, B. Math. Biol., 60, pp. 1149-1166.
[198] Por, G. 2000. Nurturing systemic wisdom through knowledge ecology. The Systems
Thinker 11(8): 1-5
[199] Porter, M. 1990. Competitive Advantage of Nations. Free Press: New York.
[200] Pottage, A., Murdy, M., 2004. Law, anthropology, and the constitution of the social:
making persons and things, Lavoisier, 2004. (Pottage et al., 2004)
[201] Prevost, G., Tranouez, P., Lerebourg, S., Bertelle, C., Olivier, D. 2004. Ecosystem
complexity described with ontological tool for a multi-scale, multi-model approaches in
distributed environment
[202] Qin, S. J., Ljung, L. 2003. Closed-loop Subspace Identi cation with Innovation Esti-
mation. Proceedings of the 13th IFAC SYSID Symposium, 2003.
[203] Reimann, M., Doerner, K., Hartl, R. F. 2004. D-ants: Savings based ants divide and
conquer the vehicle routing problems. Comput Oper Res 2004;31(4):563{91.
[204] Reynolds, C. 1987. Flocks, herds and schools: A distributed behavioral model. SIG-
GRAPH ?87: Proceedings
[205] Robinson, E.J.H. 2006. Self-Organisation in Ants: A Behavioural & Theoretical Study
of Foraging and Nest Excavation. PhD Thesis. University of She eld, She eld, U.K.
[206] Robson, M. 1993. Federal Funding and the Level of Private Expenditure on Basic
Research. Southern Economic Journal, Vol. 60 (1993), pp. 63-71.
[207] Rouchier, J., Barreteau, O., Bosquet, F., Proton, H. 1998. Evolution and co-evolution
of individuals and groups in environment. Proceedings of the third international con-
ference on multi-agent systems. AAAI Press.
[208] Rouchier, J., Bousquet, F., Requier-Desjardins, M., Antona, M. 2001. A multi-agent
model for transhumance in north Cameroon.
[209] Rovinsky, A., Menzinger, M. 1992. Interaction of Turing and Hopf bifurcation in chem-
ical system. Phys. Rev. Lett. 46 (1992), no.10, 6315{6322.
[210] Sahal, D. 1985. "Technology Guide-Posts and Innovation Avenues\, Research Policy,
14:61-82
157
[211] Sargent, R. G. 1984. Simulation model validation. in Simulation and Model-Based
Methodologies: An Integrative View, ed. Oren, T. I., Zeigler, B. P., Elzas, M. S. pp.
537-555. Heidelberg, Germany.
[212] Sargent, R. G. 2005. Veri cation and validation of simulation models. In Proceedings
of the 37th conference on Winter simulation (WSC ?05). Winter Simulation Conference
130-143.
[213] Sasmojo S., Tasrif M. 1991. CO2 emissions reduction by price deregulation and fossil
fuel taxation: A case study of Indonesia. Energy Policy. Volume 19, Issue 10, December
1991, Pages 970-977
[214] Schlesinger, S., 1979. Terminology for model credibility. Simulation, 32: 103-104.
[215] Schumpeter, J. A., 1942, Capitalism, Socialism and Democracy, New York: Harper
and Row.
[216] Segerstrom, P. S., Anant, T.C.A., Dinopoulos, E. 1990. \A Schumpeterian Model of
the Product Life Cycle", American Economic Review 80 (5), 1077-1091.
[217] Shmygelska, A., Aguirre-Hernndez, R., Hoos, H. H. 2002. An ant colony optimization
algorithm for the 2D HP protein folding problem. In: Dorigo, M., Di Caro, G., Sampels,
M. editors. Ant algorithms|Proceedings of ANTS 2002|Third international workshop.
Lecture Notes in Comput Sci, vol. 2463. Berlin: Springer; 2002. p. 40{52.
[218] Silva, C. A., Runkler, T. A., Sousa, J. M., Palm, R. 2002. Ant colonies as logistic pro-
cesses optimizers. In: Dorigo, M., Di Caro, G., Sampels, M., editors. Ant algorithms|
Proceedings of ANTS 2002|Third international workshop. Lecture Notes in Comput
Sci, vol. 2463. Berlin: Springer; 2002. p. 76{87.
[219] Sim, K.M., Sun, W.H. 2003. Ant Colony Optimization for Routing and Load-
Balancing: Survey and New Directions. IEEE Transactions, Man, and Cybernetics {
Part A: Systems and Humans, Vol. 33, No 5, Sep. 2003. pp. 560- 572.
[220] Socha, K. 2004. ACO for continuous and mixed-variable optimization. In Dorigo, M.,
Birattari, M., Blum, C., Gambardella, L. M., Mondada, F., Stutzle, T., editors, Ant
Colony Optimization and Swarm Intelligence, 4th International Workshop, ANTS 2004,
Brussels, Belgium, September 5 - 8, 2004, Proceedings, volume 3172 of Lecture Notes
in Computer Science, pages 25{36. Springer, 2004.
[221] Socha, K., Dorigo, M. Ant colony optimization for continuous domains. European
Journal of Operational Research, 2006. In press.
[222] Socha, K., Sampels, M., Manfrin, M. 2003. Ant algorithms for the university course
timetabling problem with regard to the state-of-the-art. In: Cagnoni S, Romero
Cardalda JJ, Corne DW, Gottlieb J, Guillot A, Hart E, Johnson CG, Marchiori E,
Meyer A, Middendorf M, Raidl GR, editors. Applications of evolutionary computing,
proceedings of EvoWorkshops 2003. Lecture Notes in Comput Sci, vol. 2611. Berlin:
Springer; 2003. p. 334{45.
158
[223] Solnon, C. 2002. Ant can solve constraint satisfaction problems. IEEE Trans Evolu-
tionary Comput 2002;6(4):347{57.
[224] Star eld, A.M., Smith, K.A. and Bleloch, A.L. 1990. How to model it: problem solving
for the computer age. McGraw-Hill, New York.
[225] Steiglitz, K., Honig, M., & Cohen, L. 1996. A Computational Market Model Based
on Individual Auction. In: Clearwater, S. (ed), Market-Based Control. A Paradigm for
Distributed Resource Allocation. World Scienti c.
[226] Stepp, J. R., E. C. Jones, M. Pavao-Zuckerman, D. Casagrande, and R. K. Zarger.
2003. Remarkable properties of human ecosystems. Conservation Ecology 7(3): 11.
[227] Stiglitz, J. E., 1999. Public policy for a knowledge economy. Department for Trade and
Industry and Center for Economic Policy Research. London, U.K.
[228] Stutzle, T. 1998. An ant approach to the  ow shop problem. In: Proceedings of the
6th european congress on intelligent techniques & soft computing (EUFIT?98). Aachen:
Verlag Mainz; 1998. p. 1560{4.
[229] Stutzle, T., Hoos, H. 1997. MAX-MIN Ant System and Local Search for the Travelling
Salesman Problem, IEEE International Conference on Evolutionary Computation, pp.
309- 314
[230] Sutinen, J. 1993. Cadre detude sur la gestion e cient des ressorces biologiques de la
mer. The study on the e cient management of biological resources of the sea (trans.
1995) fs (Sutinen, 1993) AGR/FI/ED 93 (3). OCDE
[231] Sutton, R. S., Barto, A. G. 1998, Reinforcement Learning: An Introduction, Cambridge
MA, MIT Press.
[232] Tanner J. T. 1975. The stability and intrinsic growth rates of prey and predator pop-
ulations, Ecology, 56, pp. 855-867.
[233] Tansley, A. G. 1935. The use and abuse of vegetational terms and concepts. Ecology
16, 284-307
[234] Teece, D. J., 1998. Capturing Value from Knowledge Assets: The New Economy,
Markets for Know-how and Intangible Assets. California Management Review, Vol. 40,
No. 3, Spring 1998
[235] Teece, D.J. 1980. The Di usion of an Administrative Innovation. Management Science
Vol. 26, No. 5. May, 1980, pp. 464-470
[236] Terna, P. 2009. The Epidemic of Innovation { Playing Around with an Agent Based
Model, Economics of Innovation and New Technology, 2009, vol. 18, issue 7, pages
707-728.
[237] Tesfatsion, L. 2002. Agent-based computational economics: Growing economies from
the bottom up
159
[238] Tijink, H., editors. Applications of evolutionary computing: Proceedings of EvoWork-
shops, 2001. Lecture Notes in Comput Sci, vol. 2037. Berlin: Springer; 2001. p. 213{22.
[239] Tschirhart, J. 2009. Integrated Ecological-Economic Models. Annual Review of Re-
source Economics. Vol. 1: 381-407 (Volume publication date July 2009)
[240] Van Dyke Parunak, H. 1998. What Can Agents Do in Industry, and Why? An Overview
of Industrially-Oriented R&D at CEC. CIA
[241] van Laarhoven, P. J. M., Aarts, E. H. L., 1987. Simulated Annealing: Theory and
Applications. Reidel, Dordrecht.
[242] Varga, A. 1998. University research and regional innovation: a spatial econometric
analysis of academic technology transfers. vol. 13. Springer.
[243] Watts, D.J., Strogatz, S.H. 1998. Collective Dynamics of ?Small-World? Networks.
Nature, 393 (1998): 440-442; D.J. Watts, Small Worlds: The Dynamics of Networks
between Order and Randomness (Princeton, NJ: Princeton University Press, 1999);
M.E.J. Newman, A.L. Barabasi, and D.J. Watts, eds., The Structure and Dynamics of
Complex Networks, Princeton, NJ. Princeton University Press, 2003.
[244] Wolfram, S. 2002. A new kind of science. Wolfram Media, Champaign, IL.
[245] Wolfram, S., 1984. Cellular automata as models for complexity. Nature 311, 419{424.
[246] Wollkind D. J., Collings J. B., Logan J. A. 1988. Metastability in a temperature-
dependent model system for predator-prey mite outbreak interactions on fruit trees, B.
Math. Biol., 50, pp. 379-409.
[247] Wollkind D. J., Logan J. A., 1978. Temperature-dependent predator-prey mite ecosys-
tem on apple tree foliage, J. Math. Biol., 6, pp. 265-283.
[248] Yang, S., Chae, S., Kwak, W., Kim, S., Kim, I. 2009. Agent-Based Approach for
Revitalization Strategy of Knowledge Ecosystem J. Phys. Soc. Jpn. 78.
[249] Yildiz, E., Acemoglu, D., Ozdaglar, A. E., Scaglione, A. 2011. Di usion of Innova-
tions on Deterministic Topologies. forthcoming in Proceedings of the 36th International
Conference on Acoustics, Speech and Signal Processing, ICASSP 2011, Praque, Czech
Republic, May 22-27, 2011
[250] Yilmaz, L. 2006. Validation and veri cation of social processes within agent-based
computational organization models. In Computational & Mathematical Organization
Theory, CMOT 2006, 12-4. Pp. 283-312
[251] Zhang, J. 2005. Growing Silicon Valley on a landscape: an agent-based approach to
high-tech industrial clusters. Entrepreneurships, the New Economy and Public Policy.
pp. 71-90. Springer Berlin Heidelberg
160
[252] Zhang, J., neng Chen, W., hui Zhong, J., Tan, X., Li, Y. 2006. Continuous function
optimization using hybrid ant colony approach with orthogonal design scheme. In Simu-
lated Evolution and Learning, 6th International Conference, SEAL 2006, Hefei, China,
October 15-18, 2006. Proceedings, Lecture Notes in Computer Science, pp. 126{133.
Springer, 2006.
[253] Zhang, J., neng Chen, W., Tan, X. 2006. An orthogonal search embedded ant colony
optimization approach to continuous function optimization. In Ant Colony Optimization
and Swarm Intelligence, Lecture Notes in Computer Science, pages 372{279. Springer,
2006.
[254] Zhi, X, Gang, Z. 2006. A combination of forecasting method in our R&D investment
of funds based on Ant colony algorithm . Science of Science and Science and Technology
Management, 2006,(09): 19-2 1.
[255] Zhu, Y., Li, S., Liu, S., Xu, Y., 2009. Agent-based Cooperative Analysis and Assis-
tant Decision-Making Method for Regional Agricultural Economic Information. Sixth
International Conference on Fuzzy Systems and Knowledge Discovery.
161
Appendices
162
Figure 2: Interaction E ects on R1 for THESIM Sub-DOE  rst-iteration.
163
Figure 3: Interaction E ects on R2 for THESIM Sub-DOE  rst-iteration.
164
Figure 4: Interaction E ects on R3 for THESIM Sub-DOE  rst-iteration.
165
Figure 5: Interaction E ects on R4 for THESIM Sub-DOE  rst-iteration.
166
Figure 6: Interaction E ects on R5 for THESIM Sub-DOE  rst-iteration.
167
Figure 7: Interaction E ects on R6 for THESIM Sub-DOE  rst-iteration.
168
Figure 8: Interaction E ects on R7 for THESIM Sub-DOE  rst-iteration.
169
Figure 9: Interaction E ects on R8 for THESIM Sub-DOE  rst-iteration.
170
Figure 10: Full data table for THESIM-ER Including all values for DRTR and the Linear
Model Yle.
171
Figure 11: Sample THESIMDX graphml output from a single-simulation trial. Note the
node and edge schema as well as the govenrment node representation (FA1)
172