systems, like planets rotating around the sun, balls running down an inclined plane, or
strictly engineered systems such as cars and televisions (and even those break down at
moments they are not supposed to). Most systems of real-world importance, such as
organisms, organizations, markets, brains, ecosystems, or the weather, are extremely
complex and open to a variety of perturbations coming from outside the system under
observation. This makes it intrinsically impossible to establish the initial conditions of
the system with any degree of accuracy. Moreover, because they are non-linear, they
are particularly prone to various “butterfly effects” that magnify tiny errors or
fluctuations in these initial conditions [Heylighen, 2009].
These observations have given rise to a novel paradigm for modeling complex
phenomena, under the label of “multi-agent systems/simulations” (MAS)
[Wooldridge, 2002; Miller, Page & LeBaron, 2007]. Instead of producing an
exhaustive mathematical description of the system in question, the MAS approach
starts by identifying the most active components of the system, the so called agents.
In an organization or society, the agents are individual people, in an ecosystem they
are organisms, in an organism they are cells, in a market they are firms, etc. Agents
are assumed to be goal-directed: they try to maximize their “utility”, “benefit” or
“fitness”. However, because of the complexity of the system and their intrinsic
cognitive limitations, agents are by definition uncertain: they only have local
knowledge, and cannot foresee the global or long-term effects of their actions. Agents
act on their environment and on each other according to certain rules, determined by
their goals and knowledge. Formalizing these rules makes it possible to write a
computer program that simulates the interaction between the different agents, and its
further evolution. The typical result from such a simulation is that novel, global
patterns emerge out of local interactions, a phenomenon termed self-organization
[Heylighen, 2009]. If the rules and initial conditions of the simulation are well chosen,
the emergent patterns correspond to real, recognizable phenomena, such as coalitions,
conflicts, cycles and clusters. While such MAS in general cannot make reliable
quantitative predictions, they often succeed in producing amazingly accurate
qualitative predictions and, more importantly, explanations.
While agent-based modeling has become a very popular, flexible and useful
method, its conceptual foundations remain vague. The paradigm of “complex adaptive
systems” (CAS) [Holland, 1992; Miller, Page & LeBaron, 2007] provides a first
justification for why systems of interacting agents are so fundamental. It takes its
inspiration from the biological theories of evolution and ecosystems, together with
social science theories of markets and societies. In this paper, I wish to take the CAS
paradigm one step further, by analyzing the concept of agent more deeply, with a
focus on the agent’s interactions with its complex and uncertain environment.
The Behavioral Sciences
The disciplines that study human behavior, including psychology, sociology,
economics, anthropology, history, and media studies, tend to be particularly