Background Report

Bibliography for Background Report

Case Studies of Industrial Ecosystems Overview

Eco-Efficiency Report

Powerpoint Presentations

Useful Links

Latest News!!

View U.S. Environmental Protection Agency project summary


Agent-Based Models of Industrial Ecosystems

Robert L. Axtell, The Brookings Institution (

Clinton J. Andrews, Rutgers University (

Mitchell J. Small, Carnegie Mellon University (

Consider millions of individuals making decisions on which products to buy, whether and how to recycle, and whether to drive or take public transportation to work. Each decision affects the use and flow of energy and materials in society. Some of these individuals may work for firms that design and manufacture the products, or for government agencies that plan, motivate or approve public infrastructure elements. Other individuals may conduct research on more sustainable products, processes, and infrastructures. Still others may work to restructure institutions, modify organizational rules and procedures, or change reporting requirements. How are these decisions made? How is information generated, received and processed to support such decisions? How do these decisions interact to determine the aggregate features of social behavior, and the resulting economic and environmental profile and performance?

Historically, economists, engineers, and planners have addressed these questions with aggregate mathematical models, assuming informed, rational behavior leading to equilibria that are societally “optimal” (Nordhaus, 1992). Similarly, to date much of the work in industrial ecology has focused on aggregate flows of resources, waste, people, and capital. While this work is essential to understanding the broad dimensions and impacts of pollution generation by industry, it is almost certainly an inadequate basis for proceeding to policy, given its highly retrospective nature. Rather, in assessing the comparative efficacy of distinct policies one must have a thorough understanding of the values, knowledge and incentives faced by various actors in the both the demand and supply sectors in question.

In recent years a disaggregate approach to analysis, referred to as agent-based modeling, has been motivated by new insights on the limitations of traditional economic assumptions and approaches as well as computational advances (Axtell, 2000). In an agent-based model individual software objects are instantiated to represent the targeted subjects. The program is 'run' or executed by simply letting the individual objects interact. The resulting evolution of the system of interacting objects is then studied both from the perspective of the aggregate population of agents as well as with respect to individual agent behavior. No assumptions concerning the attainment of equilibrium are necessary, as in the social sciences where, for example, Nash equilibrium is seen as the main 'solution concept' in game theory. Rather, equilibrium either obtains or not, and this one assesses from the actual model execution.

Indeed, the general problem of agency—the behavior of individuals in the face of imperfect incentives—lies at the core of industrial ecology (IE). The phenomena studied in IE originate in classical economic externalities and the departures from socially optimal behavior by industry and consumers due to distorted prices, incomplete markets and imperfect regulations. An important role for agent-based modeling in IE should be, therefore, to explicitly treat the incentives that face behaviorally realistic agents in empirically-credible environments. Suggestions along these lines have recently been made in this journal by two of us (Fischhoff and Small, 2000; Andrews, 2001). Here we focus more closely on the methodological aspects of agent models and ways in which industrial ecologists may begin to exploit their capabilities for systems modeling and insight.

The Intellectual Pedigree of Agents

There is a close relation between agents and so-called particle and cellular automata methods in physics and computer science. These methods have been used, for example, to represent fluid flows and have proven especially useful in the simulation of turbulent regimes, precisely where the mathematical representation of flow via the Navier-Stokes equations breaks down (Doolen et al., 1990).

In the social sciences there existed a close connection between game theory and the digital computer from early on in the post World War II era (Mirowski, 2001). Beginning at the RAND Corporation and continuing with developments in early behavioral economics research at Carnegie Mellon in the 1950s (March and Simon, 1958), early models were so heavily constrained by limited computing technology that they of necessity focussed on two or at most a few individuals.

The modern conception of agents is often credited to Schelling (1978). His model of urban segregation situated several dozen purposive-behaving individuals with explicit behavioral rules on a spatial landscape and studied the typical configurations of the model. An important contribution of this work was his demonstration that a population of individuals, none of whom prefers segregated outcomes to integrated ones, can nonetheless end up in segregationist configurations by virtue of system effects—essentially, if agents of the same type wish to have some fraction of their neighbors of their same type, this leads to clustering of like-agents at the aggregate level, thus producing segregation despite the innocuous preferences of the individuals. Interestingly, Schelling produced most of his results by hand, moving pennies and nickels around on a chess board according to rolls of dice.

Today, agents are a rapidly growing research area in the social sciences as well as within computer science (Weiß, 2000). In the latter field there has been a progressive movement of research in artificial intelligence toward the multi-agent perspective, first in the guise of distributed artificial intelligence (DAI), and most recently as multi-agent systems (MAS).

A final dimension of the agent pedigree derives from research in artificial life (ALife), often associated with the Santa Fe Institute (Langton, 1992). This line of research, pursued both by computer scientists and biologists, seeks to create 'life forms' within software through genetic, evolutionary and other means (Holland,1976). Agent-based models in ecology derive much of their impetus from this tradition (Grimm, 1999).

Modern Agent Technology

Recent generations of agent models have been made possible by technical developments in both hardware and software. On the software side, the advent of object-oriented programming (OOP) has proved crucial for this development. In older programming languages such as FORTAN, BASIC or C, a program consists of data structures and procedures for modifying the data in those structures. OOP merges the notion of data and algorithm into a single structure. An object's data are its instance variables, while its procedures for modifying these variables are its so-called methods. This ability to encapsulate data and procedures into a single object facilitates object replication and reuse, two primary advantages of OOP.

One natural by-product of object technology was the ability to give each object 'self-interest.' This might be accomplished by explicitly defining a utility function for each object and permitting the object to seek, through its behavior, improvements in its utility. Alternatively, in simpler environments where the tradeoffs implicit in the utility calculus are not present, the objects would display their self-interest through merely purposive behavior. Software objects endowed with such self-interest are known as agents.

On the hardware side, the exponential growth in computer memory, processing speed and display resolution permitted the creation on personal computers of models in which a large number of objects could be instantiated and their behavior visualized in real-time. While the first generation of agent-based models of a decade ago employed a few simple agents, today one can deploy some tens of thousands of sophisticated agents, or even millions of simple agents, on single machines. Typically, each agent in such models is characterized by the same behavioral rules, but heterogeneous data. Thus, the actual behavior of the individual agents can be quite variegated. Because agents are similar, the actual code that describes the typical agent can be quite short, while a computer's entire memory can be populated with agents by cloning. Thus, agent models tend to have a small source (compile-time) code, but a large execution (run-time) profile that can be reduced by parallel processing.

An agent model is built simply by specifying the agent population, including all the ways that agents interact. Then one creates the environment on which the agents 'live.' This might be a physical landscape or a social network, and too can be object-oriented. The timing of agent activation is specified—agents can operate in parallel or serially, and if the former then either partially or fully asynchronously. Finally, once the agent population is instantiated and permitted to interact, data gathering objects—sometimes agents— can query the population. This is done so that the user might better understand the behavior of the overall system of agents, but could also be used to simulate firm or societal data collection and information dissemination. This information can be used by selected agents to assess system performance – and thus influence their subsequent decisions.

To implement working models there exist several agent-based modeling software packages. SWARM (, Ascape (, and RePast ( are among the best known of those suitable for constructing research-quality models. Each makes use of - Java - and is in the public domain, i.e., is freely downloadable for non-commercial use. StarLogo is another well-known agent modeling environment (Resnick, 1994), in our opinion somewhat better suited to instruction than research.

Application to Industrial Ecology

Agent models have been applied in a variety of contexts of some relevance to IE, including models of resource extraction and trade (Epstein and Axtell, 1996) and organizational dynamics (Prietula, Carley, and Gasser, 1998). A recent dissertation uses agents to model the ways in which firms adapt to changed regulatory environments (Teitelbaum, 1998). Future models will investigate conflicting incentives within firms as barriers to adopting efficient technologies. Agents can also be used to model inter-firm interactions within an industry.The role of the media and other opinion-affecting agents such as citizen groups could be added to such a model, in order to consider the full range of organizational stakeholders that influence private and public decisions. Individual consumers that learn to recycle, to buy green products, and vote on environmental referenda also could be modeled. Such models will give industrial ecologists a needed test bed for safe, low-cost management and public policy experiments. For that purpose, models (so long as their predictions can be validated) will surely beat real life.


Andrews, C.J. 2001. Building a micro foundation for industrial ecology. Journal of Industrial Ecology, 4(3): 35-51.

Axtell, R.L. 2000. "Why Agents? On the Varied Motivations for Agents in the Social Sciences." In C. M. Macah and D. Sallach, eds. Proceedings of the Workshop on Agent Simulation: Applications, Models, and Tools. Argonne National Laboratory: Argonne, Illinois. Available online at

Doolen, G.D., et al, eds. 1990. Lattice Gas Methods for Partial Differential Equations. Santa Fe Institute Studies in the Sciences of Complexity, Proceedings Volume IV. Addison-Wesley Publishing: Reading, Mass.

Epstein, J.M. and R. Axtell. 1996. Growing Artificial Societies: Social Science from the Bottom Up. MIT Press/Brookings Institution Press: Cambridge, Mass./Washington, D.C.

Fischhoff, B. and M.J. Small. 2000. Human dimensions of industrial ecology. Journal of

Industrial Ecology, 3 (2 & 3): 4-7.

Grimm, V. 1999. "Ten years of individual based modeling in ecology: What have we learned and what could we learn in the future? " Ecological Modelling, 115: 129-148.

Holland, J.H. 1976. Adaptation in Natural and Artificial Systems. University of Michigan Press: Ann Arbor, Mich.

Langton, C., ed.. 1989. Artificial Life. Santa Fe Institute Studies in the Sciences of Complexity, Volume VI. Addison-Wesley Publishing: Reading, Mass.

March, J.G. and H.A. Simon. 1958. Organizations. Wiley: New York.

Mirowski, P. 2001. Machine Dreams: How Economics Became a Cyborg Science. Cambridge University Press: N.Y.

Nordhaus, W. 1992. The ecology of markets. Proceedings of the National Academy of Sciences, 89: 843-850.

Prietula, M.J., K.M. Carley and L. Gasser (Eds.) 1998. Simulating Organizations: Computational Models of Institutions and Groups, Menlo Park, CA: AAAI Press / The MIT Press.

Resnick, M. 1994. Turtles, Termites, and Traffic Jams. MIT Press: Cambridge, Mass.

Schelling, T. 1978. Micromotives and Macrobehavior. Norton: N.Y.

Teitelbaum, D. 1998. Technology Change and Pollution Control: An Adaptive Agent-Based Analysis. Ph.D. dissertation, Department of Engineering and Public Policy, Carnegie Mellon University: Pittsburgh, Penn.

Weiß, G., ed. 2000. Multi-Agent Systems. MIT Press: Cambridge, Mass.

Young, H.P. 1998. Individual Strategy and Social Structure. Princeton University Press: Princeton, N.J.