1. INTRODUCTION
Design is considered a creative activity. It is also considered a
source of innovation and a foundation for social change (Gero, 2000). However, our current understanding of
creativity in general and in relation to design, innovation, and social
change is rather limited. The last 50 years of research in creativity have
been highly speculative with a vague level of theorizing, and inconclusive
empirical evidence (Sternberg, 1999). The
present challenge in the study of creativity is to use a combination of
research methodologies to move from speculation to specification and
explanation.
The term creativity is polysemous and ambiguous. In the
literature it often refers to different ideas including aesthetic appeal,
novelty, quality, unexpectedness, uncommonness, peer recognition,
influence, intelligence, learning, and popularity (Runco & Pritzker, 1999). A useful and
well-accepted definition is that of historical or H-creativity (Boden, 1994) or creativity with a big C (Gardner, 1993). This definition highlights the
importance of social evaluation. Namely, historical creativity refers to
“the generation of ideas that are both novel and valuable; and
values are negotiated by social groups” (Boden,
1999). Social consensus constitutes a kind of Turing test, where
creativity is determined by an evaluative group interacting with the
designer over time, and is not limited to the isolated generative process
or its product.
In design, creativity can be determined by the relation between the
design process and a set of complementary social factors including
evaluation by a target population, selection by opinion leaders, and
colleague recognition. Innovation can be defined by the diffusion of
design solutions across a social group (Rogers,
1995). Thus, as a precursor of innovation, creativity can be
defined as a property socially ascribed to individuals that generate
solutions that are considered to be novel and useful by members of their
society.
The canonical approach to the study of creativity in design to date is
based on an individualistic premise under which creativity is assumed to
be an extraordinary capacity, trait, or generative process. However, such
an approach has yet to present evidence of a single or a consistent set of
individual characteristics associated with creative designers or creative
design solutions. In everyday discourse, these types of explanations are
often circular: people tend to attribute exceptional performance to talent
and to explain talent by exceptional performance (Howe
et al., 1999). An increasingly accepted approach of inquiry focuses
on the relation between individual–generative and
group–evaluative processes. Under this view, creativity is seen as a
social construct or a communal judgment (Feldman et al. 1994), where a creative designer is necessarily
defined in relation to an environment of social and epistemological
dimensions.
This paper presents an experimental test bed where qualitative
generalizations about the nature of creative behavior in design can be
explored. This framework is based on the
domain–individual–field interaction model (DIFI; Feldman et al., 1994), which locates creativity in the
interrelations of three main parts of a system: domain, field, and
individual. This supports a view of creativity as a systems property in
the same way that other constructs such as consensus or negotiation cannot
take place within a single person; they are a result of group interaction.
In the DIFI model, a domain consists of the set of solutions, knowledge,
techniques, and evaluation criteria shared by the members of a given
community. Fields include groups of individuals who share a common domain.
The key potential implication of the DIFI model is that, situated in a
dynamic environment, creative designers are those who generate “the
right product at the right place and at the right time,” where
“rightness” is largely defined by evolving social
standards.
The motivation of the research presented in this paper is to extend
our understanding of how certain individual actions in design can be
determined by collective conditions and, as a result, trigger social
changes.
2. METHOD OF STUDY
One way to investigate creative design as a social construct is to
define and implement computer simulations of the different actors and
components of a system, and the rules that may determine their behavior
and interaction. This enables the systematic study of characteristics,
conditions, and effects of interest as the simulation unfolds. By
manipulating the experimental variables of the system at initialization,
the experimenter is able to extract patterns from the observed results
over time and build hypotheses in relation to the target system.
Multiagent-based simulation of social phenomena is the primary method
of inquiry used with these types of systems (Gilbert & Troitzsch 1999). In this paper, we define a framework of social
agency based on the DIFI model, which includes a small number of competing
designer agents, a social group of clients or adopter agents, and a
cumulative repository of design solutions or artifacts that represent the
design domain, as shown in Figure 1. This
architecture supports experimentation with the types of interactions
between these system components, which have been described in general as
transmission from domain to individual, variation from individual to
field, and selection from field to domain, drawing from evolutionary
systems (Feldman et al., 1994).
Creativity as a system's property: the DIFI map. [A color
version of this figure can be viewed online at
www.journals.cambridge.org]
The canonical architecture of rational agency divides a system into
two explicit parts: agent and environment (Wooldridge,
2000). For that type of agent, changes in the environment reflect
the impact of actions by other agents or external effects. According to
the interpretation of individual autonomy in rational agency, social
interaction is limited to indirect communication via an external state.
However, the behavior of socially aware individuals cannot be expected to
be hardwired as a reaction to environmental stimuli. In social agency,
individual determinants can be expected to be complemented by interaction
with a social environment (Castelfranchi,
2001).
In our framework of design, social agency implies that the evaluation
and adoption of design solutions by members of a social group are not
entirely determined internally, but are subject to social influence. In
other words, designer and adopter agents are represented as socially
interdependent. Agents in this system, adapt their behavior to continuous
changes triggered by the generation of new solutions and by an iterative
process of social influence (Sosa & Gero,
2003). In this paper, this generative–evaluative coupling is
analyzed by manipulating some of the characteristics of social interaction
and observing consistent effects on design behavior and domain
configurations. To support a social system of design, our agent
architecture includes individual and social behaviors, as illustrated by
the mechanisms of collective agency of the field, depicted in Figure 2: the architecture of a multiagent
implementation based on the DIFI model of creativity.
A multiagent implementation based on the DIFI model of
creativity.
The social behavior of an agent (M) in an environment
(E), can be defined as the following:
where agent behavior (M) is determined by the sum of internal
state components (mn) and construed situation
(S). Internal state components (mn)
may include goals, perceptions, preferences, skills, knowledge, and
actions. Environment (E) is perceived by a bounded agent as
interpreted external state (E′). Situation (S) is,
in this sense, a function of the combination of internal and interpreted
states.
For a social agent, an external state (E) may be, for
example, a measure of group pressure to adopt an innovation. Group
pressure by itself would not determine individual behavior (M)
because it is a passive contextual feature. It becomes perceived group
pressure (E′) when it is part of a social situation, if
construed in combination with a relevant internal state (m), such
as a degree of certainty, or a preference threshold to express an opinion
or take an adoption decision (Asch, 1955). In
this way, equivalent group pressure perceived by a group of social agents
with different internal states would lead to the construction of different
social situations (S), for example, compliance or assertiveness
with others' decisions. Thus, the same context may generate different
evaluation and adoption outcomes within different social situations. When
situational factors are strong determinants, agent behavior can be
expected to be normalized; in systems where personal factors dominate,
behavior would be more differentiated across a population.
A situation can be defined at the individual level, and it can also be
shared by a group. A shared situation is perceived by a group of agents as
a result of the combination of internal states and a shared perceived
state. Extending the previous example, at the individual level, a social
situation could be one of compliance, while at the group level it could be
one of unanimity (Asch, 1955). The latter
requires, by definition, the aggregate action of a group. These two levels
of social situations are corresponding effects of one common contextual
structure, that is, group pressure.
Individual behavior under this view is defined as a function of the
agent and the situation (Ross & Nisbett,
1991). The main implication of this approach is that it supports
equivalent agents acting differently within different situations, and
different agents acting similarly within similar situations. In a system
of designer and adopter agents, this implies that given a common design
solution, different decisions to adopt are possible from members of a
social group. Conversely, given a state of social adoption, different
design decisions are possible. The multiagent system is described in
detail in the next sections. A complete list of variables is given in
Table 1.
3. ADOPTION FRAMEWORK
Adopter behavior consists of evaluating solutions generated by
designers, and deciding to adopt or abstain. Design solutions or artifacts
are described in a simple, two-dimensional linear representation, as shown
in Figure 3a. This representation is chosen
because it supports evaluation functions based on intuitive visual
geometric features. It also supports multiple interpretations by adopters
and shape emergence. As a result, it enables experimentation with some of
the key aspects of design problems in multiobjective decision making. The
particular mechanisms and values implemented in this system do not
replicate previous findings; they are mostly hypotheses based in the
literature and practice. However, this is often rather ambiguous because
observations of social phenomena are usually not specific enough to be
directly implemented. In such cases, we set parameters to explore a range
of possible options and their effects. When these parameter ranges are set
to extreme values, we do not assume them to be necessarily realistic; we
are more interested in the transition values, which are more likely to
capture real situations.
(a) Sample artifact representation and (b) a range of features
(shapes) that adopters may perceive. [A color version of this figure
can be viewed online at www.journals.cambridge.org]
Clients or adopters of design solutions in this system evaluate them
according to individual thresholds of perception and preference. Variation
of perception across a population is used to support different
interpretations of artifacts, as shown in Figure
3b, where a number of shapes (features) can be extracted
(perceived) from the sample artifact. Likewise, variation of preferences
enables different adoption decisions based on shared interpretations. In
other words, differences of perception support adoption decisions based on
different interpretations of an artifact, while differences of preference
support adoption decisions based on different evaluations of an
artifact.
The process of perception by adopter agents is implemented here by a
shape–recognition algorithm executed by every adopter with a side
limit called perception threshold (μ). Starting from every vertex in
the artifact representation, adopters conduct a search for artifact
features (shapes) following all possible paths until a number (μ) of
points is reached. This search produces a set (σ) of singular closed
shapes of (μ) number of segments, which stands for the artifact's
features as perceived by each adopter. As Figure
4 illustrates, the artifact shown in Figure
3(a) can be perceived differently when perception is based on
different thresholds. The artifact perception shown in Figure 4a is conducted by adopters with a threshold
μ = 10; the feature in Figure 4b is
perceived by adopters with μ = 9. However, to allow for a more
realistic overlap of perceptions in a society, a tolerance (μ =
±2) is defined.
Perception of features (shapes) of artifacts based on thresholds of
(a) μ = 10 and (b) μ = 9.
A μ is assigned to every adopter agent from a Gaussian
distribution at initial time (t0). As the process of
perceiving artifacts is computationally expensive, it is scheduled at
intervals of adoption. We assume that, although adopters take decisions
continuously, they only update their perceptions periodically. This is
consistent with the notion that social agents base their decisions on
approximations that they update regularly. The perception of artifacts in
our system refers to the idea that in human populations there may be a
number of distinct but overlapping views of a design artifact's
features, a notion illustrated by market segmentation.
Variation of μ across a population is controlled by the standard
deviation of the percept distribution. Different studies may consider
different percept variance assuming more subjective or more normalized
interpretation across a target population.
3.1. Adoption decision
The adoption decision process consists of a multivariate evaluation
function where adopters seek to maximize conflicting geometric objectives.
These criteria include number of shapes, shape alignment in horizontal and
vertical axes, preferred number of sides, overlapping of shapes, and shape
bounds. The evaluation or performance (ρ) of a design artifact is
individually estimated by adopters:
where artifact evaluation is based on an individualized set of
geometric relations (φ) between pairs of perceived features
(σi, σj). Evaluation
ratings (ρ) of artifacts are compared by each adopter to determine an
adoption decision, where the criterion for adoption
(φmax), is defined by the difference between evaluations
along each criterion. Namely,
where the criterion for adoption (φmax) refers to the
geometric relationship with the largest difference from the mean
(ρmax − ρmean) weighted by an
individual preference or bias (βi) between 0.0 and
1.0 assigned to adopters at t0. This adoption decision
process captures novelty preference because adopters in this system tend
to choose artifacts that they perceive to have the highest differentiation
from the rest. Adoption in this system is, therefore, a function of how
competing features of design solutions compare at a given time. To be
adopted, a design solution or artifact needs to perform better in a
criterion (φ) that other artifacts do not meet; and it helps if that
criterion is positively biased by adopters' preferences.
The set of adoption preferences (B) of an agent evolves over
time following a mechanism of habituation where the
βi for each criterion increases marginally as a
function of φmax. Namely, as adopters decide to adopt
artifacts, their preference for the geometric criterion best satisfied by
an artifact is gradually increased. This mechanism “pulls”
group preferences toward criteria that artifacts best satisfy. The set of
B of an agent is defined as the set of βi
for each evaluation criterion:
3.2. Adoption satisfaction
Adopter satisfaction is computed in this system, as a postadoption
coefficient of quality. It indicates agreement between adopters'
preferences and artifacts' features. In the adoption decision, if the
choice criterion (φmax) equals the leading preference of
an adopter (βmax), its satisfaction level is set to a
maximum in a scale of discrete values that represents “very
satisfied with the current adoption decision.” If the choice
criterion is 1 standard deviation above the mean of the adopter's
preferences, then the satisfaction level is set to a medium level or
“satisfied with current adoption decision.” Otherwise, the
adoption has been based in a criterion that is of little relevance to the
adopter, and its satisfaction level is set to a minimum or “not
satisfied with adoption decision.”
Finally, an adopter may abstain from adoption if no difference is
perceived between artifacts, that is, if
ρi−n is equal for all artifacts and
φmax = 0.
3.3. Social interaction
To represent social groups, adopters are defined in a number of social
spaces or configurations, that is, they form neighborhoods, or have
adjacency relations to other agents in simultaneous social environments.
For instance, individuals have different positions in kinship and work
structures within a society. Other approaches, such as cellular automata
of social influence, tend to conflate physical and social location into a
notion of 2-dimensional neighborhoods. In this system, agents have
n sets of neighbors in m social spaces. Such spaces are
modeled with different parameters: social tie strength and number of ties
are two structural properties defined in this paper.
At every iteration step (t), adopters rely on social
interaction to validate their perceptions, spread their preferences, and
in general to conduct their adoption decisions. To this end, different
social spaces (Π) are defined where adopters interact. In this system,
a social space is implemented in a social network where nodes represent
adopter agents of a social group, and links represent between them (Wasserman & Faust, 1994). At
t0, adopter agents are randomly assigned a location on
each social net. These social spaces have different rules of interaction
and development. Two aspects addressed in this paper are social tie
strength (T) and neighborhood size (H). Ties are defined
as interaction links between nodes in a social network and represent the
relationship between adopter agents (nodes) in a social space (Wasserman & Faust, 1994).
The T is associated with the probability that connected nodes
may interact over a period of time (Granovetter,
1973). Strong social ties usually exist between nodes in a kinship
network, while weak ties characterize networks where casual encounters
occur between strangers or acquaintances. The H is determined by
the number of links from a node, also called ego-centered networks (Wasserman & Faust, 1994).
In our framework, we implement a basic notion of tie strength as a
probability 0.0 ≤ T ≤ 1.0 that any possible pair of adopter
agents will remain in adjacent positions at the next time step (t
+ 1). When a social space has a strength T ≈ 0.0, it supports
higher social mobility. This means that adopter agents are shuffled more
often and have more opportunity to interact with different adopters over
any given period. In contrast, when strength is T ≈ 1.0,
relations between adopters remain unchanged, causing a decrease in social
mobility, that is, adopters interact with the same neighbors for long
periods of time.
3.4. Influence dominance
A social space (Π1) is set in this framework where
adopters exchange preferences (β). Within a second social space
(Π2), perception thresholds (μ) are traded. A third
space (Π3) is set where agents exchange
φmax. The strength of social ties (T) is a
property of social spaces, and it is the main control variable discussed
in this paper. Neighborhood size (H) has a constant initial value
of 2 at t0, and it varies during a system run
according to the influence that each adopter exerts on others. In this
system, more influential adopters increase the size of their
neighborhoods. These assumptions can be changed by experimenters according
to the hypothesis under inspection. For instance, the purchase of cars may
be shaped by influence interaction in small kinship networks, while the
adoption of products like mobile phones may be strongly influenced by
large peer networks.
Figure 5 illustrates different structures of
social influence in this system. Adopters are represented by circles and
influence of preferences, percepts, or decisions by arrows. Vertical axis
plots relative influence dominance; neighborhood size increases with
influence. Figure 5a shows a possible influence
structure where a few adopters have high influence and large
neighborhoods. In Figure 5b, all adopters have
similar influence and neighborhood sizes.
Influence structures in societies where (a) a few concentrate high
levels of influence and (b) influence is distributed and neighborhoods are
small.
The distribution of influence dominance (d) in a social space
is measured in this framework by the Gini coefficient, which is a summary
statistic of inequality. The Gini coefficient (γ) is used in studies
of wealth distribution, where limited group resources are exchanged among
members of a population. Influence can be seen here as analogous to
wealth, in that it is generated by the interaction between two agents
where one may increase its share at the expense of another. The Gini
coefficient ranges from a minimum value of γ = 0.0, where influence
between all individuals is equal, to a theoretical maximum of γ = 1.0,
in a population where one individual concentrates all influence dominance.
The Gini coefficient is calculated by the following:
where the average difference of every possible pair of influence
values (di −
dj) is divided by two times the average
squared (n2) of the mean group size (N; Dorfman, 1979). The larger the coefficient is, the
higher the degree of dispersion.
To determine the direction of influence between neighboring adopter
agents, they are assigned random extroversion thresholds (χ) in every
social space at t0. An adopter agent is assigned
different χ values in different social spaces
(Πn). Extroversion values are not fixed during a
system run, but change as a result of exerting influence over other
agents.
Exchange between any pair of adopters starts by a comparison of their
χs. In the social space where βs are exchanged, the adopter agent
with the higher extroversion of the pair influences the less extrovert
adopter on the criterion with the highest preference. A negotiation
process occurs by which the influenced adopter increases its preference by
a ratio of the difference between their preferences. However, if the
chosen artifact of both adopters is the same and their preferences are too
similar, the more extrovert adopter changes its own focus of attention by
shifting its preference to another criterion. This is a way to implement
uniformity-avoidance and novelty-seeking behavior, that is,
“βi is an adopter's top preference
until it perceives that βi is commonplace.”
Within other social spaces, different content is exchanged following a
similar approach. Influence (d) between adopters i and
j is of the form
where the more extrovert adopter χi
influences the less extrovert χj. Negotiation
occurs as the target preference β of agent j approaches agent
i by a ratio of their difference, in this case 0.5. The exchange
of μs and φmaxs in their corresponding social spaces
takes place in the same form.
Although the details of this interaction can be fine tuned to match
other assumptions, the key idea is that adopter agents exchange building
blocks of their adoption process. However, even if an influential adopter
is successful in spreading its preferences or percepts, the adoption
decisions of a group need not converge. Namely, adopters with equal top
preferences may still perceive artifacts differently and therefore reach
different adoption decisions.
In ergodic systems such as 2-dimensional cellular automata, a
population converges from any initial random configuration. In contrast,
when exchange occurs in more than one social space, the population may not
converge as t0 → t∞ due
to random walks being transient (Sosa & Gero,
2003).
3.5. Opinion leadership and gatekeeping
As a result of social interaction, adopter populations form aggregate
hierarchical social structures. In this framework, these structures are
determined by exchanges of preferences, percepts, and adoption decisions.
Opinion leaders are defined as adopter agents with high influence
(d) values as a result of social interaction. At
t0, the set of opinion leaders is empty. The role of
opinion leader is given to adopters whose influence is greater than 1
standard deviation above the mean of group influence. The role of opinion
leaders in this framework is to enable interaction between adopters and
designers. First, leaders serve as adoption models providing designers
with positive feedback for reinforcement learning. Second, they become
gatekeepers of the field by selecting artifacts for entry into
the domain or repository, that is, a collection of artifacts that defines
the material culture of a population (Feldman et al.,
1994).
Because the number of opinion leaders is, by definition, a small ratio
of the adopter population, they are more likely to spend more real and
computational resources in analyzing available artifacts. With an adopter
background, leaders follow the standard adoption decision process
described above, complemented by additional geometric evaluation criteria
including rotation, reflection, and uniform scale.
The domain of solutions, or artifact repository, is initialized with
an entry threshold (ε) of 0. During a system run, ε is
increased, enabling a notion of group progress by which the entry bar is
raised with every entry. Two possible entry modes are addressed in this
paper. Opinion leaders in their role as gatekeepers select artifacts that
either increase the population's threshold of ε or perform well
in different evaluation criteria (φ) than existing entries.
The nomination of artifacts by gatekeepers occurs at a control rate
specified by the experimenter. Figure 6 shows
sample repository entries as selected based on their geometric
relationships. Geometric relationships can be recognized within these
artifact shapes including scale, rotation, and symmetry. Entry threshold
(ε) to repositories has a decay mechanism (α) of the form
where ε decays marginally over time when gatekeepers fail to
nominate qualified entries above ε.
Sample entries to the repository and their perceived geometric
functions: (a) uniform scale, (b) rotation, (c) rotation, (d) reflection,
(e) uniform scale, and reflection. [A color version of this figure
can be viewed online at www.journals.cambridge.org]
The last domain mechanism described in this section is a measure of
difference between artifacts as perceived by adopters. The strategic
differentiation index (SDI) is an index estimated collectively by adopters
that reflects the perceived differentiation across the available artifacts
(Nattermann, 2000). With a design system
initialized in a converged state, SDI = 0.0. As designers seek to generate
artifacts that differ from other available artifacts, SDI > 0.
where SDI is the mean performance variance for all evaluation criteria
as estimated by every adopter agent in the population.
Adopters and opinion leaders provide the first elements for the
definition of creativity in this system. In this framework, a creative
design must exhibit novelty and fitness values to a social group at a
certain period of time, and be adopted as a result. Cumulative adoption of
artifacts addresses a notion of popularity (Simonton,
2000). An artifact must also be selected by gatekeepers, that is,
experts representative of their social group (Amabile
& Hennessey, 1999). This selection is based on rules of entry
that evolve as artifacts and societies change. Critics' choice
addresses the idea that creativity is judged by relevant arbiters (Gardner, 1993). Adoption categories enable
classification on the basis of when in the diffusion process, adopters
choose an artifact (Rogers, 1995).
4. DESIGN BEHAVIOR
The size of a group of designer agents is determined by the
experimenter as a ratio of the adopter population. At
t0, artifacts are configured and assigned to each
designer. Designer agents are given a set of standard constraints to which
their artifacts must comply. Designers' knowledge and adopter bases,
recognition levels, and repository entries are all set to zero at the
beginning of a system run. Knowledge base refers to simple domain rules
that designer agents apply during a simulation. Adopter base is defined by
cumulative adoption. Recognition is given by peer designers that imitate
features of an existing solution.
The role of designer agents in this system is to generate and present
their artifacts for assessment by adopters and gatekeepers. The details of
the design task are determined by the adopter group decisions, and by the
ability of competing designers to generate solutions. The goal of
designers in this system is to consistently generate artifacts that are
chosen by adopters, are selected by critics, and are imitated by their
peers.
In this framework, design update and adoption rates are assumed to be
periodic. Design takes place in these experiments at constant intervals
during which adopters execute their decisions and interact socially.
Variations of these assumptions are required to model different product
markets and industries, requiring particular experimentation
scenarios.
Designers may engage in different types of behavior, depending on a
number of internal and external factors. Contingent design strategies can
be seen as the product of the confluence of these conditions. The term
strategy is used as adaptation of behavior that appears to serve
a function in achieving the goal of generating artifacts that are adopted,
selected by experts, and influential. As determined by a strategy, design
behavior seeks to increment adopters' satisfaction levels and extend
adopter base by capitalizing on relative superiority (competition), or by
maximizing differences to other artifacts (differentiation).
Designer agents seek a type of contingent strategy where they learn a
design rule, that is, an instance of domain knowledge tied to the artifact
representation. In this case, condition → action rules are made by
artifact feature → target criterion. Domain rules are generated based
on the designer's model of the population's adoption process,
construed by retrieving preferences and choices of opinion leaders. This
is a way to implement positive feedback, because otherwise, a designer
would not have access to target criteria and target perception, that is,
an opinion leader may be an adopter of a competing artifact or may be
abstaining from adopting. A designer can emulate the collective decision
process by generating hypotheses of possible alternative artifacts (i.e.,
informed random changes to existing solutions).
Designers formulate hypotheses in this system, by evaluating and
changing the configuration of their artifacts to improve performance along
the modeled adoption criteria retrieved from opinion leaders. Namely,
designers sort the lines of their artifacts according to their
contribution to the formation of perceived shapes (σ). Designers are
able to delete or generate new lines as a function of adopter perception
(μ). Hypotheses consist here of rules to change a current artifact.
Features that do not contribute to good performance are replaced at
random. They are then evaluated following the multicriteria adoption
function of Eqs. (2) and (3) above.
A design rule (λ) consists of artifact changes that increase its
performance along a target criterion.
where a hypothesized feature (φh) results in an
increment of artifact performance (ρ). A positive value of Δ
stands for the improvement ratio of λ.
Individual differences between competing designers are addressed as
differences in processing and synthetic abilities assigned at
t0. Processing refers to the capacity of designers to
generate and retrieve domain rules; synthesis stands for the number of
hypotheses that designers can generate before having to transform their
artifacts. In this paper, designers are assigned constant abilities at
t0. However, abilities gradually increase as a
function of design behavior. This enables experimentation with the impact
of individual factors on creativity, which is beyond the scope of this
paper.
If during the design of a new artifact a designer is not able to
generate new domain knowledge, it seeks a strategy to apply existing
λs. Here, two assumptions can be explored: domain knowledge may have
private or public access. If private, every designer agent only has access
to their own rules, while in public mode all designers have access to all
existing rules. In this paper, public access is constant across all
experiments. Existing knowledge is applied by the following:
where an existing rule λ that improves performance (ρ) in a
target criterion (φ) is applied to an artifact.
If a designer is not able to generate or apply relevant knowledge, the
last option is to imitate other designers. Imitation is the simplest form
of collective learning, that is, blind learning, because information about
features, criteria, and perception is missing. Imitation is defined here
as the transfer of random artifact features. Imitation is chosen as the
last option because designers seem to be fixated to “reinvent the
wheel” (Purcell & Gero, 1996). This
type of imitation is rather simplistic in our model; it would be of
interest to explore in the future how relaxing this assumption would
change the observed effects.
Designers whose artifacts have low adoption rates imitate the features
of artifacts with higher rates. This is acknowledged by a mechanism where
peer recognition is given to the designer of the source artifact.
Recognition from colleagues indicates the influence of a designer.
Designers may address the perceived group's choice criterion or
they may determine an alternative target criterion. This choice is a
function of perceived adopter preferences (β′) and estimated
artifact performance (ρ′). If a designer
“determines” that its artifact's performance is
competitive (defined as equal or above mean adopter preference),
capitalization is chosen and design rules are built or applied to improve
performance on the choice criterion (exploit relative superiority). If
estimated performance is instead low on perceived adopter preferences,
then designers seek to differentiate their artifacts in a highly
competitive industry by selecting their best performing criterion.
Strategies of competition and differentiation are defined as
where ρ′ and β′ are estimated by the designer
agent.
Designer agents in this system are not equipped with creative
abilities per se. The aim is not to introduce special traits to assess the
effects of agents' creativeness as defined by the experimenter (Heck & Ghosh, 2000). Instead, all designers are
given equivalent sets of mechanisms. No extraordinary process within the
individual is hardwired, but in time, agent interaction renders a social
self-organized construct of how a designer may exhibit behavior considered
creative within its society.
These framework mechanisms encapsulate in a simple way some of the
characteristics of design problems, including poor structure and
interpretation; incremental solutions, hypothesis generation, nomological
constraints, no right or wrong answers, and delayed feedback (Goel, 1994).
Design behavior complements the definition of creativity in this
system. Adoption rate is a trend measure used to determine what designer
is imitated at a particular time step. Peer recognition is considered a
key element in the creativity literature (Runco &
Pritzker, 1999). The contribution of each designer to domain
knowledge is interpreted as transformation of the design space (Gero, 2000), learning, and experience (Runco & Pritzker, 1999). The number of hypotheses
generated resembles idea productivity. The number of entries selected by
gatekeepers gives a measure of a designer's contribution to the
repository or domain (Feldman et al., 1994).
Experimentation with this framework consists of exploring the effects
that the described individual and situational factors have on determining
the creativity of designers. A designer is considered creative by its
social group in this framework when its artifacts reach large adopter
groups, its artifacts are entered into the repository, other designers
imitate its artifacts, it transforms the design space by formulating
knowledge, and its adopters have high satisfaction levels.
The framework has been implemented in a system built in Java 1.4.2
using the following libraries: Colt 1.3 (Hoschek,
2002) for array operators and random number generators, Jxl (Khan, 2004) for output data management, and JGraph
(Alder, 2004) for visualization.
5. EXPERIMENT: SOCIAL TIES
The aim of this experiment is to assess the potential effects of
different types of social interaction in the definition of creativity in
design. It addresses the role of social ties in the formation of influence
structures and the associated effects on design behavior. Tie strength
(T) is implemented as the frequency of contact between adopters
(Marsden & Campbell, 1984). A series of
simulations are run where the initial configuration of adopters and
designers is kept constant and the strength of social ties (T) is
the experimental variable. Monte Carlo runs are conducted to explore the
range 0.0 ≤ T ≤ 1.0 in populations of 100 adopter agents
and 3 designer agents, which represents the range where adopters remain in
their social location at all times, to where they change locations on
every step, respectively.
In social networks with weak ties (T ≈ 0), connections
between adopters are reconfigured more often and they have the opportunity
to interact with different adopters over a period of time. In contrast, in
social groups where agents have strong ties (T ≈ 1), adopters
are bound, causing a decrease in social mobility, that is, adopter agents
interact within the same groups for longer periods.
Cases are run over 7500 iterations, as preliminary runs showed that
dependent variables stabilize between around 5000 iterations in most
cases. The resulting data set is filtered to exclude outliers, defined
here as 1.5 standard deviations from the mean. All the following results
represent means of 30 simulation runs. Each simulation run is initialized
in a converged state to avoid biases in the form of random initial
artifact configurations. Therefore, at t0, adopters
perceive no differentiation between artifacts and all abstain from
adopting. It is only after designers first modify their artifacts that
adoption commences.
5.1. Influence hierarchies results
The result of varying (T) from 0.0 to 1.0 shows that
influence concentration increases with social tie strength. In societies
with strong ties (T ≈ 1), a few opinion leaders become
dominant (higher γ). In contrast, as social ties become weaker
(T ≈ 0), social mobility increases and agents have contact
within a varying neighborhood causing structures of influence to be more
distributed (lower γ). Figure 7 shows a
scatter plot on a logarithmic scale of the relation of T and
γ with fitness = 0.92. Although cases with very strong social ties
yield a high γ, most strength values in the range yield comparatively
low results. It is particularly interesting to obtain an exponential
distribution by linear increments of an experimental variable, a type of
pattern that is prevalent in biological and social phenomena (Barabasi et al., 2001).
Exponential function for tie strength (T) and the Gini
coefficient (γ). [A color version of this figure can be viewed
online at www.journals.cambridge.org]
This result suggests that in most cases in these types of systems,
influence hierarchies can be expected to be rather flat or egalitarian,
the exception being only when adopter agents tend to remain in stable
social positions over long periods. In social groups with strong ties,
there is lower mobility and hierarchical structures of influence exist
between adopters. As a result, in such groups influence hierarchies
guarantee that a few individuals become dominant in the spread of adoption
opinions. In contrast, in weaker social settings, adopters can be expected
to influence their peers to a lesser degree. Influence is more diffused in
these groups.
Small amounts of social mobility in societies of strong ties rapidly
reduce disparities. As tie strength decreases further, influence becomes
more egalitarian up to a point at which even large changes in social tie
strength and mobility do not have a significant impact. Figures in the
following sections plot only the ends (T = 0) and (T =
1), because most (T) values yield similar results until
(T ≈ 1), at which point effects vary significantly.
5.2. Gatekeeping effects results
At the domain level, the formation of structures of influence has
effects that may be unexpected: an inverse correlation is shown between
the T and number of entries to the repository. In other words, in
cases where influential opinions concentrate in a few adopters, designer
agents are able to generate less creative solutions. Lower values of
T are correlated with larger repositories as shown in Figure 8 (Pearson = 0.67, N = 30, p
= 0.001). In societies with weak social ties (T ≈ 0), a mean
of 97 artifacts with a standard deviation of 43.4 are selected by
gatekeepers. In societies with strong social ties (T ≈ 1), a
mean of 16 artifacts with a standard deviation of 11.7 are selected.
Social spaces with high tie strengths tend to produce smaller
repositories. [A color version of this figure can be viewed online at
www.journals.cambridge.org]
From the result discussed previously it can be seen that in societies
with strong ties, a constant set of adopter agents tends to remain in the
role of gatekeepers. Namely, gatekeeping is more stable and controlled by
a small unchanging group of influential experts. Therefore, evaluation
criteria remain constant over time. As a consequence, domains or
repositories tend to be smaller. In contrast, in societies with lower tie
strength and therefore where influence is distributed rather than
concentrated, there is a higher change rate of gatekeepers. The gatekeeper
group is constantly composed of different adopters. Consequently, more
diverse evaluations support a larger number and a higher variety of domain
artifacts.
5.3. Differentiation effects results
The differentiation of design artifacts is measured by the SDI as an
aggregate measure of differences perceived by adopters. These experiments
show that SDI is inversely correlated with the strength of T, as
seen in Figure 9. Designer agents operating on
strong social spaces where influence structures are stable, tend to
generate more similar artifacts. The same designers operating on wider
distributed influence social spaces, have a tendency toward higher
differentiation (Pearson = 0.57, N = 30, p = 0.004).
The effects of the social tie strength (T) in the strategic
differentiation index. [A color version of this figure can be viewed
online at www.journals.cambridge.org]
This effect on design behavior can be explained by the normative
nature of strong social ties. In societies where a few influential opinion
leaders exist, adoption choices can be expected to be more similar. As a
result, designers repeatedly engage in competition to improve their
artifacts. In contrast, in societies with weaker links, adoption opinions
are expected to diverge and provide designers with a wider range of
preferences. In such cases, different artifacts are adopted.
5.4. Prominence effects results
Last, effects on the size and nature of adopter groups are addressed.
Results show that tie strength (T) is positively correlated with
adopter group size (Pearson = 0.608, N = 26, p = 0.001).
The standard deviation of adoption in weak ties (1718) is also
significantly higher than in strong ties (726). This illustrates that weak
social ties increase abstention and make adoption less predictable. This
is a consistent result with the notion that in more rigid societies there
is a higher agreement of adoption opinions.
Adoption variance, on the other hand, is given by the distribution of
adopters by designer agent. When adoption variance is high, most adopters
choose the artifacts of one designer, whereas a low adoption variance
indicates that adopters distribute their choices among all designers. The
strength of social ties (T) is correlated with adoption variance
as shown in Figure 10 (Pearson = 0.68,
N = 26, p = 0.001). Namely, in social spaces with weak
ties adoption choices tend to be more distributed across designers. In
contrast, strong ties (T ≈ 1) increase total adoption and
concentration of choices around a few designers.
Strong ties (T ≈ 1.0) produce higher variation of adoption
between adopter groups. [A color version of this figure can be viewed
online at www.journals.cambridge.org]
This result has an interesting potential implication from the
designers' point of view. Designer agents with the same individual
characteristics but operating in two extremes of social tie strength can
expect different outcomes. When within a society with weak links
(T ≈ 0), their popularity is likely to be lower and more
unstable, although prominence among peers is harder to obtain. In this
framework, the popularity of designers is given by the size of their
adopter groups and prominence by the distribution of adoption choices. In
contrast, when the same designers operate within a society with strong
ties (T ≈ 1), one should expect higher and more consistent
popularity levels, and a higher concentration of prominence, that is, a
few designers concentrating most adoption choices.
5.5. Summary of results
According to the shape of the relationship between tie strength
(T) and influence distribution (γ), lower popularity and
lower concentration of prominence can be expected to be the norm in these
types of generative–evaluative social systems. Under exceptional
social conditions, the effects of otherwise equivalent designer agents
have a sudden change, as the critical point at which influence
concentrates is reached. Within such rare situational conditions, one
designer agent is likely to concentrate the choices of a majority of
adopters.
Within a society with strong ties, significant effects occur
throughout the system. Adopters converge in their decisions, artifacts are
perceived as more different, and domain sizes are smaller and more
predictable.
5.6. Verification
A key aspect of this experiment has been replicated using an
alternative implementation, to demonstrate this framework's validity
beyond a specific computational representation. This replication makes use
of the social net library in a well-known multiagent simulation toolkit
for Java (Collier, 2004). A social network is
defined here by configurations of nodes connected by directional links. In
a simple example of how the structure of a social network is created,
random links or ties are added between nodes. The nodes are placed in an
array, and their ties constitute the social network. Link directionality
makes it possible to dispense with the extroversion threshold in the
mechanism of social influence between adopters. With 1-directional link,
source nodes always exert influence over destination nodes. A density
parameter (f) is included where 0.0 ≤ f ≤ 1.0,
which determines the ratio of connections between nodes in the social net.
Figure 11a depicts a social net of 25 agents and
f = 0.5 in a circular configuration (Freeman,
1998).
Replication of influence hierarchies using social nets. (a) The marker
represents link directionality between nodes. (b) Most tie strength values
(T) generate low Gini coefficients, except when T ≈
1. [A color version of this figure can be viewed online at
www.journals.cambridge.org]
The behavior of this social net is limited to the exchange of
influence between adopter agents described in our framework. At initial
time, a value of influence is given to all nodes, equal to zero. During a
simulation run, source nodes increment this value by one unit as they
exert influence over those nodes to which they are linked. The probability
of tie replacement between a node pair in this network is a function of
T, specified as an experimental condition for each case. In
networks with strong ties (T ≈ 1), ties remain constant during
a simulation (they change with probability = 0); in contrast, in networks
with weak ties, these are changed more often until (T ≈ 0),
when ties between nodes are reconfigured at random at every iteration step
(with probability = 1). Monte Carlo runs are conducted to explore the
range 0.0 ≤ T ≤ 1.0 in social nets of 100 agents,
f = 0.5. The distribution of influence in a social net is again
measured by the Gini coefficient.
Results are consistent with those shown in Figure
7: as the scope of contact between nodes extends, influence
distribution rapidly decreases, up to a level after which large
differences of tie strength have only marginal effects on influence
distribution. As Figure 11b shows, most tie
strength values generate low Gini coefficients. This reinforces the
suggestion that social situations where the decisions of a few opinion
leaders affects the creators' patterns are rather unlikely in systems
of collective evaluation.
6. DISCUSSION
In this paper, a social framework for the study of creativity and
innovation in design has been introduced and used to experiment (to a
limited extent) with a specific situational factor of creativity in
design. Factors that regulate aggregate behavior of a population of
evaluators are shown to affect the way creators operate, and their impact
as change agents of their societies.
The results presented in this paper illustrate one way in which
creativity can transcend the individual domain. This is a significant
contribution to the ongoing discussion on the relationship between
creators and their societies. For instance, based on a biographic and
historical analysis of several creative individuals, Gardner (1994) suggests that in more hierarchical fields (i.e.,
“where a few powerful critics render influential judgments about the
quality of work”), it has been easier for a small number of creators
to gain recognition and influence. These studies of creative figures
suggest that social characteristics may determine who is considered
creative, and when. Individuals who are characterized as extraordinary
creators, may exhibit similarities of personality traits and abilities, or
there may be similarities between the structures of the fields within
which they operate. While in the cases analyzed by Gardner (1994) personalities vary significantly across
creators, in most cases a few powerful critics rendered influential
judgments about the quality of their work.
This paper has shown in a computational simulation of design as a
social activity that agent societies with strong social ties are likely to
develop uneven hierarchies that support powerful opinion leaders. As a
result, a few prominent creators are likely to emerge in more static
societies. In contrast, in social networks with weak ties, influence is
distributed, expert judgments tend to vary over time, and they tend to
have a lesser impact in the evaluation of new works. Likewise, creators
will be less differentiated in dynamic societies where exchange of
opinions is open and frequent between different members of society. These
findings are limited to the assumptions and restrictions of the system
implemented, but they show consistency with Gardner's (1994) observations. They reinforce the idea that
social structures of evaluation can significantly affect the distribution
of prominence, and therefore, how individual creators are perceived by
their societies.
These experiments illustrate a key idea about the likely nature of
creativity and innovation: a situational factor that regulates the way in
which adopters interact may have a significant effect on how both
designers and social groups operate. Nonetheless, there is an alternative
interpretation of the relation between authority and creativity. Rudowicz
(2003) presents a review of several empirical
studies that support the idea that educational practices in hierarchically
organized societies tend to promote behavior that is incompatible with
creativity, that is, conformism and conventional thinking. As a number of
evolutionary models have shown, novelty in a society may be facilitated by
a balance between dissent or diversity, and the converging effect of
imitation (Boyd & Richerson, 1995; Gabora, 1995). This discrepancy indicates that further
research is necessary to fully understand the role of structures of
authority in creativity and innovation.
The concept of situations seems an adequate unit of analysis to model
the link between design cognition and social change. A creative situation
(i.e., one within which designers with different characteristics are
likely to trigger a social change), could be typified in design to
complement the dominance of studies that focus on the creative
personality.
The key potential implication of this research is that it may not be
possible to put forward conclusions about human designers or computational
design generators by limiting inquiry to their characteristics in
isolation. A number of situational factors depict a close interdependence
of designers with their social groups. A corollary of these types of
studies is that the understanding of creativity will require the extension
of the unit of study outside the cognitive realm of the design process,
and into the social psychology of design. Computational simulation has a
fundamental role in supporting experimentation of sociocognitive
interactions. These in silico studies can provide new ways to
think about the phenomenon of creativity and new ways to study it with
in vitro (laboratory) and in vivo (biographic) tools of
inquiry.
ACKNOWLEDGMENTS
This research is supported by an International Postgraduate Research
Scholarship and a University of Sydney International Postgraduate Award.
Computing resources are provided by the Key Centre of Design Computing and
Cognition (www.arch.usyd.edu.au/kcdc). We thank Andy Dong, Guest
Editor, for his valuable assistance.
