Agent-based modeling for organizational design:
three real-world case studies
Multi-agent models of human organizations are ideally
suited to play what-if scenarios and organizational designs. Drawing from
three real-world case studies from the consumer goods, pharmaceutical and
chemical industries, we will show how to use agent models to simulate the
operations of an organization and how it can help to test new processes,
organizations and incentives. Example 1: A leading consumer goods manufacturer
wanted to improve its ice cream business supply chain efficiency by aligning
its local country managers’ incentives with the good of the whole company;
because of the incentive structure, each country manager was focusing on
maximizing certain metrics, which led to adverse emergent properties. By
understanding the link between each manager’s behavior and the collective
behavior of the entire organization through an agent-based model, it was
possible to design and test incentives that maximized the entire organization’s
efficiency and not just the fitness of country business units. Example 2:
One of the major bottlenecks in the pharmaceutical industry is early drug
development. Two major issues there are to align employees’ incentives with
the interests of the company and the design of decision making processes
that are in phase with how work really happens. By building a highly detailed
agent model of the activities of early drug development, it was possible
to gain insights into what kind of incentives was needed in each phase of
drug development and how to induce people to make the “right” decision; the
model was also used to test new ways of organizing work to realign work with
decision making. Example 3: An agent-based model of the corporate ventures
group of leading chemical firm produced insights into the design of the appropriate
incentives for the various phases of venture development and was used to
test a variety of management and decision making rules.
Eric Bonabeau
Ian Fenty
Belinda Orme
Oliver Bandte
Joe DeAngelo
Agent-based model of abnormal behavior
A number of social and economic systems are affected by
people who try to harm or exploit the systems, either legally or illegally.
Examples include money laundering, computer attacks, and more generally terrorist
activities. Detecting abnormal or fraudulent behavior when there is a high
level of background noise is a daunting task as weak abnormal signals can
be lost in an ocean of normal behavior –detecting abnormal behavior is the
equivalent of looking for a needle in a haystack. Real detected cases, in
addition to probably having become irrelevant, provide insufficient information
to draw statistically significant conclusions from existing data. This is
a major issue for designing security systems, such as intrusion detection
systems in computer networks or anti-money laundering surveillance systems.
Multi agent-based modeling provides a powerful answer to the problem of lack
of data: by modeling normal as well as abnormal behavior it is possible to
capture how data and evidence is generated from the bottom up in a way that
is realistic. The agent model provides an explicit mapping between individual
behavior and the evidence generated, thereby enabling to create massive amounts
of relevant synthetic data that can then be used to teach statistical tools
ranging from simple correlation analysis to more complex tools such as neural
or bayesian networks: the model is used to generate many needles in the haystack.
The agent approach does not attempt to build a statistical model of abnormal
behavior from existing data but rather a behavior model that leads to the
observed data. Constructing the model is based mostly on expert descriptions
of abnormal behavior, including the motives potentially governing that behavior,
and qualitative as well as quantitative analysis of normal usage patterns.
Three examples will be given, drawn from computer security, financial markets
and terrorist activities.
Eric Bonabeau
Ben Shargel
The econometrics of agent-based models
One component that seems to be missing from a lot of agent
modeling is rigorous model estimation. How to connect agent models to the
real world and in particular how to use existing data (often available only
at the aggregate level) to estimate models of individual behavior is a difficult
issue but one that needs to be addressed if agent models of human behavior
are to move beyond their current status of insightful but poorly predictive
tools. Agent models constitute an important conceptual advance to the simulation
of human behavior and have a number of features that make them more powerful
than traditional utility-based econometric approaches. Such features include:
explicit disaggregate behavioral modeling, decision making under conflicting
constraints, strong non-linearity, natural framework for dealing with irrationality/bounded
rationality, natural framework for dealing with social interactions between
people, and many more. A combination of domain expertise, discrete choice
survey data and conjoint analysis, and, when applicable, social network analysis,
provides the input to define the model structure or model space. Because
there is no general mathematical framework for estimating agent-based models,
we will present heuristic machine learning-based estimation techniques used
to identify both the model AND its parameters. In this presentation we will
draw from real world examples to show how agent models of human behavior
can be estimated and made predictive. Three examples will be presented: (1)
agent model of how people select an insurance policy; (2) how physicians
decide to prescribe certain drugs based on marketing channels and peer-to-peer
social interactions; (3) estimating social network models of terrorist activity.
In the last two examples we will show how agent models can be used to determine
which data should be collected to make the model predictive.
Eric Bonabeau
Tobin Van Pelt
Illy Khatib
Alexis Arias
Corresponding Author:
Eric Bonabeau
Icosystem Corporation
www.icosystem.com
eric@icosystem.com