Co-Evolution of Social Sciences and Engineering Systems
ROBERT L. AXTELL
George Mason University
Since the birth of engineering, social and behavioral scientists have played an important role in bringing new technologies to market and designing user interfaces. Many of these technologies have also proven to be invaluable to social and behavioral scientists in their efforts to understand people. In other words, there is a kind of co-evolution of engineering systems and social sciences.
Multi-agent systems (MAS), in which a population of quasi-autonomous software objects (agents) interact directly with one another and with their environment for purposes of simulation, control, and distributed computation, is poised exactly at this interface. MAS can be considered social systems, in which each member of a heterogeneous population pursues its own objectives, constrained by its interactions with others. Indeed, ideas from the social sciences, including game theory (e.g., mechanism design), economics (e.g., auction theory), and sociology (e.g., social networks), are increasingly being incorporated into such software systems.
At the same time, social scientists are increasingly using software agents to model social processes, where the dominant approach is to represent each person by a software agent. Such models yield high-fidelity depictions of the origin and operation of social institutions (e.g., financial markets, organizational behavior, and the structure of social norms). They can also be used to understand the differential effects of alternative policies on such institutions.
In short, social systems are systems with multiple agents, and MAS are (increasingly) social systems. The co-evolution of social and technological sys-
tems means that advances in one field lead to progress in the other, nucleating further improvements in the original field, and so on. This interface between these broadly defined fields of knowledge opens up many opportunities for new research.
SOCIAL SYSTEMS AS MULTI-AGENT SYSTEMS
Building models in which purposive software objects represent individual people is a way around two classical problems—aggregation and the necessity to assume equilibrium—within conventional mathematical modeling in the social sciences. Because social systems are typically composed of a large number of heterogeneous individuals, mathematical models in the social sciences have been one of two types: (1) aggregate models, in which the heterogeneity of the population is either assumed away (e.g., representative agent models) or averaged away by looking only at mean behavior (e.g., systems dynamics models); and (2) models written at the level of individuals, in which “solution” of the models involves all agents engaging only in equilibrium behavior (e.g., Nash equilibria in game theory, Walras-Arrow-Debreu equilibria in economics) and all dynamic paths by which such equilibria might be achieved are neglected. It is clear how the agent approach fixes aggregate models by fully representing individuals. The agent-based approach also remedies the second problem by letting agents interact directly (in general these are out-of-equilibrium interactions); equilibrium is attained only if a path to it is realized from initial conditions.
MAS grew up in the mid-1990s and combined with so-called artificial life (ALife) models, giving rise to agent-based approaches in the social sciences. As the capacity of computer hardware increased exponentially, more sophisticated agent models could be built, using either more cognitively complex agents or a larger number of simple agents, or both. Thus, large agent populations were soon realized in practice leading naturally to the metaphor of an artificial society (Builder and Bankes, 1991).
In modeling an artificial society, a population of objects is instantiated and permitted to interact. Typically, each object represents one individual and has internal data fields that store the specific characteristics of that individual. Each object also has methods of modifying its internal data, describing interactions, and assessing its self-interest (i.e., it can rank the value to itself of alternative actions). This quality of self-interestedness, or purposefulness, makes the objects into agents.
Conventional mathematical models in the social sciences rely heavily on a suite of heroic assumptions that are false empirically and, arguably, do more harm than good as benchmarks. There are four main ways agent-based computing can be used to relax these assumptions. First, mainstream economics makes much of a “representative agent,” conceiving the entire economy as simply a scaled-up version of a single decision maker. This specification is easy to relax computationally. Second, economics models normally consider only rational
agent behavior, whereby optimal behavior can be deduced by all agents for all time. Not surprisingly, in a MAS of any complexity, such deductions are computationally intractable and cannot be implemented in practice. Thus, models often resort to bounded rationality. Third, modeling conventions have often dictated that agents not interact directly with other individuals but interact either indirectly through aggregate variables or perhaps through some idealized interaction topology (e.g., random graph, lattice). In agent-based computing, however, any topology, including empirically significant networks, can be easily implemented to mediate agent interactions. Finally, equilibrium is the focal point for all solution concepts in the social sciences. Whether equilibrium obtains or not in an agent-based system, the dynamics matter and are fully modeled.
All of the social sciences—anthropology (Axtell et al., 2002; Diamond, 2002, 2005; Kohler and Gumerman, 2000); geography (Gimblett, 2002); social psychology (Kennedy et al., 2001; Latane et al., 1994; Nowak et al., 2000); sociology (Gilbert and Doran, 1994; Gilbert and Conte, 1995; Flache and Macy, 2002; Macy and Willer, 2002); political science (Axelrod, 1984; Kollman et al., 1992; Cederman, 1997; Lustick et al., 2004); economics (Arifovic and Eaton, 1995; Arifovic, 1996; Kollman et al., 1997; Tesfatsion, 1997; Kirman and Vriend, 2000; Luna and Stefansson, 2000; Allen and Carroll, 2001; Arifovic, 2001; Bullard and Duffy, 2001; Luna and Perrone, 2001; Tesfatsion, 2002, 2003; Arifovic and Masson, 2004; Axtell and Epstein, 1999; Axtell et al., 2001; Young, 1998); finance (Palmer et al., 1994; Arthur et al., 1997; Lux, 1998; LeBaron et al., 1999; Lux and Marchesi, 1999; LeBaron, 2000, 2001a, 2001b, 2002, 2006); organizational science (Carley and Prietula, 1994; Prietula et al., 1998); business (Bonabeau and Meyer, 2001; Bonabeau, 2002); many areas of public policy (Axtell and Epstein, 1999; Moss et al., 2001; Saunders-Newton, 2002; Bourges, 2002; Janssen, 2002; Rauch, 2002); transportation science and policy (Nagel and Rasmussen, 1994; Nagel and Paczuski, 1995; Nagel et al., 1998; Gladwin et al., 2003); public health/epidemiology (Wayner, 1996); demography (Kohler, 2001); and the military (Ilachinski, 2004)—have more or less active research programs using agent computing. Although the nature of these applications is idiosyncratic within particular fields, they are unified methodologically in the search for agent specifications that yield empirically observed (or at least empirically plausible) social behavior.
MULTI-AGENT SYSTEMS AS SOCIAL SYSTEMS
Not only has agent computing changed the practice of the social sciences, but the social sciences have altered the face of MAS. Certain social science methods have been adopted by computer and information scientists not only at the research frontier, but also in commercial systems. In the same way that social scientists have reworked the MAS paradigm for their own ends, developers have adapted extant social science methods to specific problems in their domains.
The role of agents within computer and information science has been primarily to enhance the function of distributed, decentralized systems. For example, in designing a new computer network, each individual node in the network might be given the ability to manage its own resource allocation, based on information about the overall load on the network. This might be done cooperatively or competitively (Huberman, 1987; Miller and Drexler, 1988). Similarly, a well-designed network should function properly regardless of the topology of how the machines are hooked together. Thus, ideas from graph theory and social network theory—each node can be thought of as socially interactive—have been relevant and put to good use.
Basic research on agent systems has been amplified and extended beyond the academic community by the exigencies of e-commerce. The prospect of automated bargaining, contracting, and exchange among software agents has driven investigators to explore the implications of self-interested agents acting autonomously in computer networks and information technology servers.
Because decentralization is an important idea within MAS, ideas from microeconomics and economic general equilibrium theory that focus on decentralization were incorporated into MAS under the rubric “market-oriented programming” (Wellman, 1996). Mechanism design is an approach to the synthesis of economic environments in which the desired performance of a mechanism is specified, and one then figures out what incentives to give the agents in a way that the equilibria (e.g., Nash equilibrium) that are individually rational and incentive compatible achieve the objective. This formalism was developed largely in the 1980s and is today viewed by some as a viable way to design MAS (Kfir-Dahav et al., 2000).
In distributed control, market metaphors have been replaced with actual market models (Clearwater, 1996). Temperature control of a building is an example application (Huberman and Clearwater, 1995) of a MAS that makes explicit use of concepts from economic general equilibrium (e.g., Mas-Collel et al., 1995).
In automated negotiation (e.g., Rosenchein and Zlotkin, 1994), MAS researchers have made extensive use of game theory. In automated contracting, the Contract Net protocol (Weiss, 1999) was an early example of a high-level protocol for efficient cooperation through task sharing in networks of problem solvers. Since then, much more work of this type has been done. More recent work has taken an explicitly social stance, looking for an emergent social order, for example, through the evolution of social orders and customs, institutions, and laws.
AGENT-BASED TECHNOLOGY AND CO-EVOLUTION
I am aware that portraying agent-based computing as a bridge between engineering and the social sciences may be risky. By touting the apparent effective-
ness of a new methodology, there is always a risk that “hype” will overshadow substance and raise unrealistic expectations.
The alternative approach is to paint an evolutionary picture, in which today’s new methodologies are seen as logical extensions of adequate but dated conventional methodologies. Thus, progress appears to be natural, with no abrupt changes. This view can be “sold” more easily to existing research communities and is easier to insinuate into conventional discourse.
Evolution or revolution? Continuous change or abrupt change? Smooth transition or phase transition? One is tempted to invoke Kuhn (1962) at this point, but it may be enough to point out that the technical skills required for those who are fomenting change are quite different from those of many current faculty members and those who teach current graduate students. Only a very small subset of social science researchers knows enough about computer science to perform agent-based modeling in their areas of expertise. This is also the major barrier to the systematic adoption of these new techniques—and proof that agent-based modeling constitutes a discontinuous advance.
Assuming that Moore’s law will continue to hold true for the next generation (20 to 30 years), the capabilities of agent computing will double every 18 to 24 months, increasing by an order of magnitude each decade. From the social science perspective, this technological revolution will permit the construction of increasingly large models involving greater numbers of progressively more complex agents. When one contemplates the possible desktop hardware of 2020, one can imagine hundreds of gigabytes of ultrafast RAM, fantastic clock and bus speeds, and enormous hard disks. The continuing computer revolution will fundamentally alter the kinds of social science models that can be built. It will also alter the practice of social sciences, as equations give way fully to agents, empirically tested cognitive models arise, and decision models grounded in neuroscience emerge.
It is anyone’s guess where co-evolution will lead. Co-evolutionary systems have the capacity to fundamentally alter one another and their environments in novel, creative ways. Thus, speculations for the medium term and long run may look and sound a lot like science fiction.
Allen, T.M., and C.D. Carroll. 2001. Individual learning about consumption. Macroeconomic Dynamics 5(2): 255–271.
Arifovic, J. 1996. The behavior of exchange rates in the genetic algorithm and experimental economies. Journal of Political Economy 104(3): 510–541.
Arifovic, J. 2001. Evolutionary dynamics of currency substitution. Journal of Economic Dynamics and Control 25: 395–417.
Arifovic, J., and C. Eaton. 1995. Coordination via genetic learning. Computational Economics 8: 181–203.
Arifovic, J., and P. Masson. 2004. Heterogeneity and evolution of expectations in a model of currency crisis. Nonlinear Dynamics, Psychology, and Life Sciences 8(2): 231–258.
Arthur, W.B., J.H. Holland, B. LeBaron, R. Palmer, and P. Tayler. 1997. Asset Pricing Under Endogenous Expectations in an Artificial Stock Market. Pp. 15–44 in The Economy as an Evolving Complex System II, edited by W.B. Arthur, S.N. Durlauf, and D.A. Lane. Reading, Mass.: Addison-Wesley.
Axelrod, R. 1984. The Evolution of Cooperation. New York: Basic Books.
Axtell, R.L., and J.M. Epstein. 1999. Coordination in Transient Social Networks: An Agent-Based Computational Model of the Timing of Retirement. Pp. 161–183 in Behavioral Dimensions of Retirement Economics, edited by H.J. Aaron. Washington, D.C.: The Brookings Institution Press.
Axtell, R.L., J.M. Epstein, J.S. Dean, G.J. Gumerman, A.C. Swedlund, J. Harburger, S. Chakravarty, R. Hammond, J. Parker, and M.T. Parker. 2002. Population growth and collapse in a multiagent model of the Kayenta Anasazi in Long House Valley. Proceedings of the National Academy of Sciences of the U.S.A. 99(supp 3): 7275–7279.
Axtell, R.L., J.M. Epstein, and H.P. Young. 2001. The Emergence of Classes in a Multi-Agent Bargaining Model. Pp. 191–211 in Social Dynamics, edited by S.N. Durlauf and H.P. Young. Cambridge, Mass./Washington, D.C.: MIT Press/Brookings Institution Press.
Bonabeau, E. 2002. Agent-based modeling: methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences of the U.S.A. 99(supp 3): 7280–7287.
Bonabeau, E., and C. Meyer. 2001. Swarm intelligence: a whole new way to think about business. Harvard Business Review 79: 106–114.
Bourges, C. 2002. Computer-Based Artificial Societies May Create Real Policy. Washington Times, May 12, 2002.
Builder, C., and S. Bankes. 1991. Artificial Societies: A Concept for Basic Research on the Societal Impacts of Information Technology. Santa Monica, Calif.: RAND Corporation.
Bullard, J., and J. Duffy. 2001. Learning and excess volatility. Macroeconomic Dynamics 5(2): 272– 302.
Carley, K.M., and M.J. Prietula. 1994. Computational Organization Theory. Hillsdale, N.J.: Lawrence Erlbaum Associates.
Cederman, L.-E. 1997. Emergent Actors and World Politics: How States and Nations Develop and Dissolve. Princeton, N.J.: Princeton University Press.
Clearwater, S.H. 1996. Market-Based Control. Hackensack, N.J.: World Scientific.
Diamond, J.M. 2002. Life with the artificial Anasazi. Nature 419: 567–569.
Diamond, J.M. 2005. Collapse: How Societies Choose to Fail or Survive. New York: Allen Lane.
Flache, A., and M.W. Macy. 2002. Stochastic collusion and the power law of learning. Journal of Conflict Resolution 46: 629–653.
Gilbert, N., and R. Conte, eds. 1995. Artificial Societies: The Computer Simulation of Social Life. London: UCL Press.
Gilbert, N., and J. Doran, eds. 1994. Simulating Societies: The Computer Simulation of Social Phenomena. London: UCL Press.
Gimblett, H.R., ed. 2002. Integrating Geographic Information Systems and Agent-Based Modeling Techniques for Simulating Social and Ecological Processes. Santa Fe Institute Studies in the Sciences of Complexity. New York: Oxford University Press.
Gladwin, T., C.P. Simon, and J. Sullivan. 2003. Workshop on Mobility in a Sustainable World: A Complex Systems Approach, Ann Arbor, Michigan, June 20–22, 2003.
Huberman, B.A., ed. 1987. The Ecology of Computation. New York: North-Holland.
Huberman, B.A., and S.H. Clearwater. 1995. A Multi-Agent System for Controlling Building Environments. First International Conference on Multi-Agent Systems, San Francisco, Calif. Cambridge, Mass.: AAAI Press/MIT Press.
Ilachinski, A. 2004. Artificial War: Multiagent-Based Simulation of Combat. Singapore: World Scientific Publishing.
Janssen, M.A., ed. 2002. Complexity and Ecosystem Management: The Theory and Practice of Multi-Agent Systems. Northampton, Mass.: Edward Elgar Publishing Inc.
Kennedy, J., R.C. Eberhart, and Y. Shi. 2001. Swarm Intelligence. San Francisco, Calif.: Morgan Kaufmann.
Kfir-Dahav, N.E., D. Monderer, and M. Tennenholtz. 2000. Mechanism Design for Resource Bounded Agents. Pp. 309–315 in Proceedings of the Fourth International Conference on Multi-Agent Systems. Boston, Mass.: IEEE Computer Society.
Kirman, A.P., and N.J. Vriend. 2000. Learning to Be Loyal: A Study of the Marseille Fish Market. Pp. 33–56 in Interaction and Market Structure: Essays on Heterogeneity in Economics, edited by D. Delli Gatti and A.P. Kirman. Berlin, Germany: Springer-Verlag.
Kohler, H.-P. 2001. Fertility and Social Interactions. New York: Oxford University Press.
Kohler, T.A., and G.J. Gumerman, eds. 2000. Dynamics in Human and Primate Societies: Agent-Based Modeling of Social and Spatial Processes. Santa Fe Institute Studies in the Sciences of Complexity. New York: Oxford University Press.
Kollman, K., J.H. Miller, and S.E. Page. 1992. Adaptive parties in spatial elections. American Political Science Review 86: 929–937.
Kollman, K., J.H. Miller, and S.E. Page. 1997. Political institutions and sorting in a tiebout model. American Economic Review 87(5): 977–992.
Kuhn, T.S. 1962. The Structure of Scientific Revolutions. Chicago, Ill.: University of Chicago Press.
Latane, B., A. Nowak, and J.H. Liu. 1994. Measuring emergent social phenomena: dynamicism, polarization and clustering as order parameters of social systems. Behavioral Science 39: 1–24.
LeBaron, B. 2000. Agent-based computational fiannce: suggested readings and early research. Journal of Economic Dynamics and Control 24: 324–338.
LeBaron, B. 2001a. A builder’s guide to agent-based financial martkets. Quantitative Finance 1: 254–261.
LeBaron, B. 2001b. Evolution and time horizons in an agent-based stock market. Macroeconomics Dynamics 5: 225–254.
LeBaron, B. 2002. Short-memory traders and their impact on group learning in financial markets. Proceedings of the National Academy of Sciences of the U.S.A. 99(supp 3): 7201–7206.
LeBaron, B. 2006. Agent-Based Financial Markets: Matching Stylized Facts with Style. Pp. 221–238 in Post Walrasian Macroeconomics: Beyond the Dynamic Stochastic General Equilibrium Model, edited by D.C. Colander. New York: Cambridge University Press.
LeBaron, B., W.B. Arthur, and R. Palmer. 1999. Time series properties of an artificial stock market. Journal of Economics Dynamics and Control 23: 1487–1516.
Luna, F., and A. Perrone, eds. 2001. Agent-Based Methods in Economics and Finance: Simulations in Swarm. Boston, Mass.: Kluwer Academic Publishers.
Luna, F., and B. Stefansson, eds. 2000. Economic Simulations in Swarm: Agent-Based Modeling and Object Oriented Programming. Advances in Computational Economics, volume 14. Boston, Mass.: Kluwer Academic Publishers.
Lustick, I.S., D. Miodownick, and R.J. Eidelson. 2004. Secessionism in multicultural states: does sharing power prevent or encourage it? American Political Science Review 98(2): 209–229.
Lux, T. 1998. The socioeconomic dynamics of speculative markets: interacting agents, chaos and the fat tails of return distributions. Journal of Economic Behavior and Organization 33: 143–165.
Lux, T., and M. Marchesi. 1999. Scaling and criticality in a stochastic multi-agent model of a financial market. Nature 397: 498–500.
Macy, M.W., and R. Willer. 2002. From factors to actors: computational sociology and agent-based modeling. Annual Review of Sociology 28: 143–166.
Mas-Collel, A., M.D. Whinston, and J.R. Green. 1995. Microeconomic Theory. New York: Oxford University Press.
Miller, M.S., and K.E. Drexler. 1988. Markets and Computation: Open Agoric Systems. Pp. 231–266 in The Ecology of Computation, edited by B.A. Huberman. New York: North-Holland.
Moss, S., C. Pahl-Wostl, and T. Downing. 2001. Agent-based integrated assessment modelling: the example of climate change. Integrated Assessment 2(1): 17–30.
Nagel, K., and M. Paczuski. 1995. Emergent traffic jams. Physical Review E 51: 2909–2918.
Nagel, K., and S. Rasmussen. 1994. Traffic at the Edge of Chaos. Pp. 224–235 in Artificial Life IV, edited by R.A. Brooks and P. Maes. Cambridge, Mass.: MIT Press.
Nagel, K., R. Beckman, and C.L. Barrett. 1998. TRANSIMS for Transportation Planning. Technical Report. Los Alamos, N.M.: Los Alamos National Laboratory.
Nowak, A., R.R. Vallacher, A. Tesser, and W. Borkowski. 2000. Society of self: the emergence of collective properties in self-structure. Psychological Review 107: 39–61.
Palmer, R.G., W.B. Arthur, J.H. Holland, B. LeBaron, and P. Tayler. 1994. Artificial economic life: a simple model of a stock market. Physica D 75: 264–274.
Prietula, M.J., K.M. Carley, and L. Gasser, eds. 1998. Simulating Organizations: Computational Models of Institutions and Groups. Cambridge, Mass.: MIT Press.
Rauch, J. 2002. Seeing around corners. Atlantic Monthly 289: 35–48.
Rosenschein, J.S., and G. Zlotkin. 1994. Rules of Encounter: Designing Conventions for Automated Negotiation among Computers. Cambridge, Mass.: MIT Press.
Saunders-Newton, D. 2002. Introduction: computer-based methods: state of the art. Social Science Computer Review 20(4): 373–376.
Tesfatsion, L. 1997. How Economists Can Get a Life. Pp. 533–564 in The Economy as an Evolving Complex System, Volume II, edited by W.B. Arthur, S. Durlauf, and D.A. Lane. Menlo Park, Calif.: Addison-Wesley.
Tesfatsion, L. 2002. Agent-based eomputational economics: growing economies from the bottom up. Artificial Life 8(1): 55–82.
Tesfatsion, L. 2003. Agent-based computational economics: modeling economies as complex adaptive systems. Information Sciences 149(4): 262–268.
Wayner, P. 1996. Computer Simulations: New-Media Tools for Online Journalism. New York Times, October 9, 1996.
Weiss, G., ed. 1999. Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence. Cambridge, Mass.: MIT Press.
Wellman, M. 1996. Market-Oriented Programming: Some Early Lessons. Pp. 74–95 in Market Based Control: A Paradigm for Distributed Resource Allocation, edited by S.H. Clearwater. Hackensack, N.J.: World Scientific Press.
Young, H.P. 1998. Individual Strategy and Social Structure. Princeton, N.J.: Princeton University Press.