Creation of Creative Work Teams using Multi-Agent based Social
Simulation
Adrián Bresó, Alfonso Pérez, Javier Juan-Albarracín, Juan Martínez-Miranda, Montserrat Robles and
Juan Miguel García-Gómez
Grupo de Informática Biomédica, Instituto ITACA, Universitat Politècnica de València, València , Spain
Keywords: Individual creativity, Group creativity, Repast, Social simulation, Multi-Agent System, Teamwork.
Abstract: Over the past decades, advances in Artificial Intelligence (AI) techniques have investigated the modelling of
complex systems. In particular, the use of Multi-Agent Systems (MAS) opened new possibilities for
studying different domains using social simulation. In the present work we have implemented and
empirically evaluated a Multi-Agent Based Social Simulation (MABSS) system to support the formation of
creative work teams. Based on existent psychological and organizational creativity studies, we have
modelled a set of personal characteristics and contextual factors to represent and analyse their influence on
creativity at both: the individual and the group level. The obtained initial results were significantly better
than the results obtained with a pure stochastic model (average improvement of 8.2%). Additionally, we
empirically confirm some hypothesis about group formation from the organizational studies.
1 INTRODUCTION
Simulation systems have been applied since 1950’s
in several research domains, such as Political
Sciences (Yamakage et al., 2007), Economics and
Social Sciences (Phan & Varenne, 2010),
Environmental Science (Gernaey et al., 2004) and
Natural Resource Management (Galán et al., 2009)
to name a few.
A particular simulation technique that has been
widely used in recent years is the known as Multi-
Agent Based Simulation (MABS). Some of the main
reasons behind the growing popularity of this kind
of systems are that they offer (1) the possibility to
carry out “what-if” scenarios to better understand the
domain under analysis at lower costs and, (2) the
flexibility to simulate wide combinations of
behaviours observed in the real world. In MABS, an
agent represents a real world entity that perceives
events (such as interactions with other agents) and
autonomously reacts according to their mental state.
One of the domains where MABS is more
frequently applied is Social science, resulting in the
so-called Multi-Agent Based Social Simulation
(MABSS) (Davidsson et al., 2006). The MABSS
field is frequently used to analyse the properties and
effects at social level of a set of attributes modelled
in the individuals within an organizational structure.
The agent-based model described in this paper
makes use of MABSS to analyse the creativity
process in a work team. In particular, our model is
focused in the abstract representation of a work team
to better understand the influences of the individual
team members’ characteristics and their interaction
on the group creativity.
The rest of this paper is organized as follows:
Section 2 presents the related work developed in the
fields of creativity and agent-based Social
Simulation Models. Section 3 describes our
proposed model whereas Section 4 describes its
implementation. Section 5 discusses the initial
evaluation of the model through the obtained results
of the simulations. Finally, Section 6 presents the
conclusions and the future work.
2 RELATED WORK
The design and implementation of MABSS
simulations for the analysis of complex behaviours
have been used over quite different domains such as
military applications (Luscombe & Mitchard, 2003),
police and criminal behaviours (Melo et al., 2006)
211
Bresó A., Pérez A., Juan-Albarracín J., Martínez-Miranda J., Robles M. and García-Gómez J..
Creation of Creative Work Teams using Multi-Agent based Social Simulation.
DOI: 10.5220/0004240302110218
In Proceedings of the 5th International Conference on Agents and Artificial Intelligence (ICAART-2013), pages 211-218
ISBN: 978-989-8565-38-9
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
and management of health emergencies (Benkhedda
& Bendella, 2012) among others.
A common characteristic of these works is the
study of human behaviours within a group. In this
line some MABSS have been specialised in the
simulation of teamwork characteristics. Some
authors have simulated teamwork behaviours
through the representation of specific features such
as shared mental models, collaborative behaviours,
communicative behaviours and others. A close
related work to our proposed model is the Dong et
al. (2008) approach, where they evaluated the
relationships between members of a workgroup
based multi-agent model. Whereas this model is
focused on the study of group efficiency, our model
is on the analysis of group creativity.
From all the existent models, just a few have as
the main objective the analysis of individual, social
and contextual factors behind creativity behaviours.
One of them is the model developed by (Sosa &
Albarran, 2008) which is focused on team formation
based on the social tie strength to improve teamwork
practices in creative activities. With their model, the
authors tried to respond if teams with strong ties
(teams of friends) and teams with weak ties (teams
of strangers/partner) produce different creative
processes and solutions. The results concluded that
teams of friend produced more quality solutions and
teams of strangers produced more creative solutions.
A similar work is described in (Martínez-
Miranda, 2010) where the development of an agent-
based simulation model to support the formation and
configuration of work teams is presented. This
model represents and analyses the performance of
the team as a consequence of four human attributes:
personality type, emotional state, social-related
skills and cognitive abilities (including creativity).
This model considers the individual creativity as an
influential variable on work team performance but it
is static and the model does not calculate or modify
it, a key difference with our proposed model.
3 MODELING CREATIVITY IN A
WORK TEAM
3.1 Scenario Description
The scenario is composed of a Manager agent and N
Worker agents. Each of these Worker agents
represents a different role within the work team:
Director, Assistant, Technician and Scholar.
Additionally, they have individual characteristics
(see section 4.2) that lead to significant differences
between them. Worker agents aspire to have a job on
the working group that the Manager agent is
forming, whereas the Manager agent receives a
proposal to create a working group as creative as
possible.
The objective of the Manager is to form a work
group (initially empty) with the maximum creativity
constrained to the limitations of the proposal (e.g.
budget of the project or team size). Using a
configurable selection criterion, the Manager agent
performs the negotiation by selecting the most
promising candidate. After creating the working
group, the Manager agent could make replacements
in order to achieve a more creative group.
The main hypothesis that our model aims to
study is whether the inclusion of a highly creative
individual outperform directly the work team’s
creativity.
3.2 Creativity Assessment
The function that models the creativity of the group
(CF_Group) is the most important function in the
model and it is the objective function to be
maximised by the simulation. This evaluation
function depends on the individual creativity
function (CF_Individual), the group characteristics
(GroupFactors), and the relationships of the
members in the group (RelationalFactors).
3.2.1 CF_Individual
The calculation of individual creativity
(CF_Individual) is based on six positive and two
negative (creativity hinders or as defined in Amabile
(1998) creativity killers). The negative factors are
those identified by (Batey et al., 2010) and the
positive factors are taken from (Carroll et al., 2009)
summarised in the Table 1. The eight factors are
classified into cognitive capabilities, social skills,
emotional states and personality traits similarly as in
(Martínez-Miranda, 2010).
Using the factors listed in Table 1 we define the
individual creativity function as a weighted sum of
the positive and negative individual factors (see
equation 1).
The range of the factors is [0, 1] and the range of
weights (randomly assigned) is [0, 7]. The sum of
the weights must be 28. In the code, we apply a
normalization to the CF_individual function so the
range is [0, 1].
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
212
Table 1: Factors to calculate the individual creativity.
Index Factor Category Effect
1 Exploration Cognitive-
Related
Capabilities
Enabler
2 Immersion Enabler
3 Results Worth
Effort
Enabler
4 Collaboration Social-Related
Skills
Enabler
5 Expressiveness Enabler
6 Enjoyment Emotional
State
Enabler
7 Agreeableness Personality
Traits
Killer
8 Conscientiousness Killer
(1
)
3.2.2 CF_Group
For calculating group creativity (CF_Group) we
have based our model on a simplification of the
work presented in (Woodman et al., 1993)
concentrating our model only at the individual and
group level and discarding the organisational level.
The GroupFactors function refers to aspects of
composition / characteristics of the group as number
of leaders, longevity, composition, cohesion or
structure (Payne, 1990), (King, 1990). In our model,
we always create new teams (i.e. without longevity)
and the user sets the team size and the required roles.
So the only variable that we use to calculate the
GroupFactors function is the cohesion. The
cohesion is the commitment of the group members
to work together to complete their shared tasks and
accomplish their goals (Guzzo & Salas, 1995).
The last factor used is the RelationalFactors
function which refers to aspects of communication
and relationships between the team members. This is
an important factor identified by several authors
such as (Payne, 1990), (King, 1990), (Carroll et al.,
2009) or (Woodman et al., 1993). Our model
establishes randomly good or bad individual
perceived relationship between all team members.
Hence, we define the group creativity as a weighted
sum of the N individual creativities, the group
factors and the relationship factors (equations 2, 3, 4,
5). We normalize to the CF_Group function in order
to obtain a bounded range [0, 1]:
∗
∗
∗
(2
)
_
(3
)
(4
)
∗
∗
(5
)
3.3 Agent Communication
Communication takes place only between the
manager and the worker agents. The communication
process is supported by a negotiation protocol
(Figure 1) in order to construct the better work team
in terms of creativity.
The simulation starts with a work team proposal
that the manager should form. The work team
proposal is defined by a set of requirements such as
the number of agents to form the group, the number
of different roles within the group and the available
budget for hiring agents. Following these
requirements, the manager generates an offer and
sends it to the workers. An offer has two attributes:
the wage and the current team creativity. The wage
is set by the manager depending on the role for
which the offer is made up. The current team
creativity indicates the creativity of the team in
which the agent will become a member (initially
CF_Group=0).
Figure 1: Negotiation protocol for a work team creation.
According to its personal preferences, each agent
accepts the offer or ignores the proposal based on
CreationofCreativeWorkTeamsusingMulti-AgentbasedSocialSimulation
213
the satisfaction degree related with the offer (see
details in section 4.2.3). Each agent has different
preferences: while some agents are more motivated
by the economic aspect, others give more
importance to the team creativity level. Thus, a score
is computed for each agent, indicating the degree of
satisfaction with the offer. This score is then used in
the CF_Group, particularly in equation (4), as
IndividualCohesion.
After the offer evaluation, a list of candidate
agents is filled with those agents that accept the
offer. Then, the manager selects the highest
creativity agent and includes it into the team.
4 MODEL IMPLEMENTATION
From the available platforms used
for the
development of agent-based models and simulations,
we selected the Repast Simphony software
(http://repast.sourceforge.net/repast_simphony.html)
which is an open source software with a good
documentation and support. The following
subsections describe the classes implemented in the
model.
4.1 General Purpose
4.1.1 Work Team
The work team class represents the current work
team in every simulation step. This class stores the
list of agents belonging to the work team and
provides the method to compute the team creativity
(equations (2), (3), (4) and (5)). Weights used in the
equations are summarized in Table 2.
Table 2: Weights for the work team creativity attributes.
Weight Value
Individual weight 0.3
Group weight 0.4
Relational weight 0.3
Good relations weight 0.4
Bad relations weight 0.6
4.1.2 Offer
The offer class represents an instance of a job offer
that the manager sends to the workers. The offer has
two attributes:
1. Proposed wage
2. Current work team creativity
This class is used as a message within the
communication protocol during the negotiation
process. Each agent evaluates the content of this
class in order to accept or reject the offer.
4.1.3 Project
The project class represents the work team
restrictions that the manager should met when
forming the team. The attributes of this class which
can be set through the developed GUI are:
1. Current budget
2. Initial budget
3. Number of work team members
4. Required roles
4.2 Worker Agent
Worker Agents store a set of methods and attributes,
each of them representing different features and
behaviours that workers must have. These attributes
and methods are divided in the following groups:
1. Creativity skills.
2. Agent relations.
3. Negotiation.
Different types of agents have been implemented
in terms of the role it plays in the system (Director,
Assistant, Technician and Scholar). Each role
inherits from the Agent class and represents a type
of worker with a similar behaviour and
characteristics. Differences between roles rely in the
values that each attribute can get. These attributes
can be modified (using the GUI) to analyse how
these changes affect the global behaviour of the
system.
4.2.1 Creativity Skills
An agent has an individual creativity based on a
combination of its creativity attributes and the
weights attached to each of them. We implemented
these attributes as floating point variables in the [0,
1] range whit 0 referring to the absence of the skill
represented by the attribute, and 1 representing the
maximum ability in that skill (Table 3). Moreover,
several ranges are modified to generate different
states from a set of initial conditions.
Thus, an agent randomly initializes its attributes
in its specific range defined by ourselves. This
implementation allows us to generate a large
population of agents of different types, with a high
intra-role and inter-role variability. A linear
combination of the creativity attributes and their
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
214
corresponding weights provides the final individual
creativity of the agent (see equation 1).
Table 3: Attribute ranges for each agent’s role.
Attributes Ranges
Director Assistant Technician Scholar
Exploration [0, 1] [0, 1] [0.8, 1] [0.8, 1]
Immersion [0.8, 1] [0, 0.7] [0.8, 1] [0, 0.7]
Results Worth
Effort
[0.8, 1] [0, 0.7] [0.8, 1] [0, 1]
Collaboration [0, 0.5] [0.8, 1] [0, 0.8] [0.8, 1]
Expressiveness [0.8, 1] [0.8, 1] [0, 0.5] [0.5, 0.8]
Enjoyment [0, 0.4] [0, 1] [0.5, 1] [0.8, 1]
Agreeableness [0, 1] [0.8, 1] [0, 1] [0.9, 1]
Conscientiousn
ess
[0.8, 1] [0.4, 1] [0.5, 1] [0, 1]
4.2.2 Agent Relations
Social relations are one of the most important factors
influencing the work team creativity. Several studies
claim that relations within a team can become more
relevant to the team creativity than the individual
creativity of the members.
In our model, the agent relationships are
implemented as individual hash maps in which the
key indicates the agent identifier and the value
indicates the type of relationship. A positive value
means a good relationship; while a negative value
indicates a bad relationship. Each agent has its own
hash map initialized with the identifiers of all the
agents in the system and a random relation for each
of them. Therefore, relations may not be reciprocal.
Agent A1 may have a good relationship with agent
A2, but agent A2 may have a bad relationship with
agent A1. This approach allows us to model the
social interactions between agents in order to
measure the impact of relationships in the
development of worker teams, which is part of the
hypothesis that our system tries to corroborate.
4.2.3 Negotiation
The negotiation process is based on two aspects: the
economic and the creativity motivations. Economic
motivations are based on a desired wage and the
proposed wage of the offer. The desired wage
attribute is randomly initialized in a specific range
depending on the role, while the proposed wage is
set by the Manager. Creativity motivations are
influenced by the current team creativity in which
the agent will become a member. Both parameters
(economic and creativity) are weighted by each
agent, so the offer evaluation process is performed
as follows:
∗
_
∗
_
(6)
(7)
0.5
→
→
(8)
Both wageRatio and workgroup Creativity are
represented in the [0, 1] range with 0 representing
the worst value and 1 referring to the better value.
Moreover, these values are weighted with different
weights for each agent, modelling the personal
preferences in the negotiation process. Finally, if the
result of the evaluation is greater than 0.5, (i.e. is
more positive than negative), the offer is accepted.
Otherwise, the offer is rejected.
4.3 Manager Agent
The manager is a special type of agent whose
purpose is to conduct the simulation. Unlike worker
agents, the manager has no creativity attributes and
only interacts with the agents during the negotiation
process. The manager behaviour follows the next
steps:
1. Discard an agent (if the group is full)
2. Propose offer
3. Select best candidate
4. Update work team
The manager owns the work team proposal
represented by a project object. The work team is
initially empty but in posterior steps it may be full,
so it is needed to check the number of members in
the work team, in order to discard an agent if
needed. We followed a discard criteria based on the
work team creativity loss when an agent is
discarded. Thus, each agent is virtually excluded
from the work team in order to measure the amount
of creativity loss. Following this approach, the worst
agent is the one that generates the least loss in the
work team creativity. Moreover, it is possible that
creativity increases when a member is excluded, e.g.
when the member has bad relations within the group.
Once a vacant is generated in the work team, a
new agent must be hired. Therefore, the manager
proposes an offer for the required role and sends it to
the agent population. The offer is then evaluated by
CreationofCreativeWorkTeamsusingMulti-AgentbasedSocialSimulation
215
the agents and the manager receives a list of
candidates. The main hypothesis that our system is
trying to verify is whether the inclusion of an agent
with a high creativity always increases the team
creativity, so the manager must select the best
candidate based solely in the individual creativity,
ignoring relationships or other attributes. The agent
with highest creativity level is then included into the
work team.
5 RESULTS
The empirical evaluation of the system have been
performed by defining three set of simulations which
creates work teams with 50, 10 and 5 workers
respectively. For each set of simulations we
compared the group creativity obtained by our
model algorithm vs. a pure stochastic selection
algorithm using a high (500), medium (100) and
small (50) agent populations of each role. A
summary of the simulations can be seen in the
Tables 4-6.
Table 4: Simulation results summary of SET 1 (50 Work
Team Size: 15 assistants, 5 directors, 10 scholars and 20
technicians).
Population size
500 (x4) 100 (x4) 50 (x4)
Our Model 0,718 0,712 0,64
Random Model 0,612 0,605 0,611
Improvement
10,6% 10,7% 2,9%
Table 5: Simulation results summary of SET 2 (10 Work
Team Size: 3 assistants, 1 director, 3 scholars and 3
technicians).
Population size
500 (x4) 100 (x4) 50 (x4)
Our Model 0,763 0,719 0,685
Random Model 0,648 0,64 0,62
Improvement
11,5% 7,9% 6,5%
We used the same seed in all the simulations for
reproducibility. The number of iterations used for
calculating the results was 200.
Table 6: Simulation results summary of SET 3 (1 Work
Team Size: 1 assistant, 1 director, 1 scholar and 1
technician).
Population size
500 (x4) 100 (x4) 50 (x4)
Our Model 0,799 0,754 0,768
Random Model 0,724 0,682 0,671
Improvement
7,5% 7,2% 9,7%
As shown in Table 4, in all cases our algorithm
get better creativity levels than the stochastic
selection algorithm. Our algorithm runs better with
high populations and small work teams. However,
the stochastic selection algorithm had similar results
in all the tests.
As Figure 2.a shown, in the initial iterations, the
evolution of work team creativity is highly variable.
At this step, since the work team is not full, the
manager only hires agents considering their
individual creativity. This can result in work teams
where agents have enough bad relations and then
unstable work team creativity is obtained. Once the
work team is full, the manager should discard the
worst agent and include a new one in order to
observe the creativity evolution. The discard
criterion takes into account the GroupFactors and
RelationalFactors, besides CF_Individual.
As also represented in Figure 2.a, the work team
creativity increases significantly if these aspects are
considered (iterations 50 to 200), which is the main
hypothesis we want to corroborate. The inclusion of
agents with highest individual creativity do not
ensures the highest work team creativity. In the
simulation (b) no clear differences were seen as in
the previous case. All the time the CF_Group
remains similar because it is only influenced by the
CF_Individual.
6 CONCLUSIONS
The presented model could be used to empirically
support, hypothesise, train and analyse how factors
such as cognitive-related capabilities, emotional
state, personality and social-related skills can affect
individual and group creativity.
As confirmed by the results, our model reaches
values higher than a pure stochastic model obtaining
a significant difference (until 11.5%). Additionally,
we have confirmed the hypothesis that when
incorporating an agent with a high of creative
individual to a working group, the group creativity is
not always positively affected. This is because in the
calculation of creativity there are other (group and
relational) factors besides the individual factor.
In further improvements, we would follow the
approach proposed by Woodman et al. (1993),
which states that when an individual becomes part of
a group and the group creativity is affected, the
individual creativity is also influenced immediately
by the group conditions.
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
216
Simulation (a): 100 (x4) agents, 50 team members,
algorithm selection implemented.
Simulation (b): 100 (x4) agents, 50 team members,
stochastic selection.
Figure 2: Two representative simulation examples. The
green line shows the total number of agents belonging to
the group. The blue line shows the group creativity and the
red line shows the individual creativity of the hired agent.
Another necessary improvement is to add an
organizational layer designed to assess aspects such
as the organizational structure of the working group,
an important influence on creativity.
REFERENCES
Amabile, T. (1998). How to kill creativity. Harvard
Business Review, Sept. 1998 pp. 77-78.
Batey, M., Furnham, A., & Safiullina, X. (2010).
Intelligence, general Knowledge and personality as
predictors of creativity. Learning and Individual
Differences Vol. 20(5), pp. 532–535.
Benkhedda, S., & Bendella, F. (2012). Multi-Agents
Simulation of Human Behavior in a Situation of
Emergency. Advanced Computing: An International
Journal ( ACIJ ), Vol.3, No.1, pp. 35-30.
Carroll, E. A., Latulipe, C., Fung, R., & Terry, M. (2009).
Creativity factor evaluation: towards a standardized
survey metric for creativity support. Proceedings of
the seventh ACM conference on Creativity and
cognition, pp. 127-136. New York.
Davidsson, P., Holmgren, J., Kyhlbäck, H., Mengistu, D.
& Persson, M. (2006). Applications of Agent Based
Simulation. Proceedings of the 2006 international
conference on Multi-agent-based simulation, pp. 15-
27. Berlin: Springer-Verlag.
Dong, S., Hu, B., & Wu, J. (2008). Modelling and
simulation of team effectiveness emerged from
member-task interaction. Proceedings of the 40
th
Conference on Winter Simulation, pp. 914-922.
Galán, J., López-Paredes, A. & del Olmo, R. (2009). An
agent-based model for domestic water management in
Valladolid metropolitan area. Water Resources
Research Vol. 45 W05401.
Gernaey, K., van Loosdrecht, M., Henze, M., Lind, M., &
Jorgensen, S. (2004). Activated sludge wastewater
treatment plant modeling and simulation: state of the
art. Environmental Modelling and Software, Vol. 9(9),
pp. 763-783.
Guzzo, R., & Salas, E. (1995). Team Effectiveness and
Decision Making in Organizations. Wiley, ISBN: 978-
1-55542-641-5.
King, N. (1990). Innovation at Work: The Research
Literature. In M. A. West and J. L. Farr (eds).
Innovation and Creativity at Work: Psychological and
Organizational Strategies, Wiley: pp. 15-80.
Luscombe, R., & Mitchard, H. (2003). Exploring simple
human behaviour representation using agent based
distillations. Proceedings of Simulation Technology
and Training (SimTecT), pp 26-29. Australia.
Martinez-Miranda, J. (2010). Modeling Human Behaviour
at Work: an Agent-Based Simulation to Support The
Configuration of Work Teams. Madrid: PhD. thesis.
Melo, A., Belchior, M. & Furtado, V. (2006). Analyzing
Police Patrol Routes by Simulating the Physical
Reorganization of Agents. Multi-Agent-Based
Simulation VI, pp. 99-114. Springer.
Payne, R. L. (1990). The effectiveness of research teams:
A review. In M. A. West & J. L. Farr (eds.),
Innovation and creativity at work: Psychological and
organizational strategies, pp. 101 – 122. UK: Wiley.
Phan, D., & Varenne, F. (2010). Agent-Based Models and
Simulations in Economics and Social Sciences: From
Conceptual Exploration to Distinct Ways of
Experimenting. Journal of Artificial Societies and
Social Simulation. Vol. 13(1), 5.
Sosa, R., & Albarran, D. (2008). Supporting idea
generation in design teams. Proceedings of EDPE 08
Anna Clarke, Mike Evatt, Peter Hogarth, Joaquim
Lloveras, Luis Pons (eds.), pp. 61-66. Barcelona.
Woodman, R., Sawyer, J., & Griffin, R. (1993). Toward a
Theory of Organizational Creativity. In Academy of
Management Review, Vol. 18(2), pp. 293-321.
CreationofCreativeWorkTeamsusingMulti-AgentbasedSocialSimulation
217
Yamakage, S., Hoshiro, H., Mitsutsuji, K., Sakamoto, T.,
Suzuki, K., & Yamamoto, K. (2007). Political Science
and Multi-Agent Simulation: Affinities, Examples and
Possibilities. In Agent-Based Approaches in Economic
and Social Complex Systems IV, pp. 165-182.
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
218
