« Complexity Theory & Political Science | Main | The Fifth Discipline »

Adaptive agent Modeling in a Policy Context

Gulden, T. R. (2004). Adaptive Agent Modeling in a Policy Context. Unpublished Dissertation, University of Maryland, College Park.

This dissertation attempts to add to the empirical functionality of adaptive agent (or agent based) models for use in policy analysis. It makes some self-admittedly modest contributions to both theory and methodology.

Following Axtell (2000), Gulden claims that adaptive agent models may be useful in three distinct situations. Firstly, these models can be used to analyze systems which can be modeled with equations that are solvable. Examples of these systems abound in economics, when the neoclassical assumptions of perfect rationality, perfect information flow, decreasing returns, etc. are held to be true. Although adaptive agent models are obviously not necessary for modeling such systems, Gulden claims they may be useful in providing novel ways to structure the analytical issue and in allowing the modeler to relax assumptions that may not reflect the true dynamics of the system.

Secondly, adaptive agent models can be used to analyze systems that can be described by equations that are not easily solvable either analytically or numerically. Gulden says that "these include models with badly behaved equilibria, particularly models where the features of interest are not equilibrium states, but rather the fluctuations that the system goes through on its path to equilibrium." Analytical intractability in these systems may be due to the heterogenaity of agents, spatial dependence between agents within the system, or complex internal states of agents.

Thridly, adaptive agent models may be useful in analyzing systems for which formulating numeric equations are not analytically feasible and may not be theoretically productive. Such systems often feature spatial heterogenaity between agents and bounded rationality of agents.

During the course of his dissertation, Gulden applies an adaptive agent model to three different problems, each of which is typical of one of the three systems classes. The first problem that Gulden tackles is how the assumption of increasing returns may affect international trade policy. As a key feature of macroeconomic theory, international trade has long been integrated within macroeconomic models. However, this means that neoclassical assumptions have also been rigorously applied to trade theory. The assumption of decreasing returns, in particular, has led to the supremacy of free trade in todays global economy. Using adaptive agent modeling, Gulden is able to relax the assumption of decreasing returns in favor of increasing returns that have been observed in some national industries. He is confident that this model accurately handles increasing returns and imperfect capital mobility because the adaptive agent model produces the same results as the neoclassical model when neoclassical assumptions are given in the model's parameters. Under the assumptions of decreasing returns and perfect capital mobility, Ricardian trade theory predicts that nations will most efficiently produce the goods for which they have a comparative advantage and will trade for all other goods. However, the assumptions of increasing returns and imperfect capital mobility may result in a nation gaining a competitive advantage in an industry for which it is not particularly well suited and may keep producing this industry's goods, even when other nations could potentially produce the good more efficiently. Practically speaking, a developed nation with intensive capital investment may gain supremacy in an industry that may be better suited for a particular developing nation. Gulden argues that trade protectionism on the part of the developing nation would be preferable to free trade.

The second problem that is analyzed using an adaptive agent model is the observed Zipf distribution of city sizes in many nations. The Zipf distribution is a particular type of power-law described as rank/size. In the case of cities, this means that a city's size rank compared to all other cities in the nation-state is inversely proportional to its population compared to all other cities. The log-log distribution of these variables is linear with a slope of -1. The interesting thing about the Zipf distribution of cities in a nation is that researchers have been able to mathematically model this distribution, but not in a way that is theoretically meaningful. The general assumption is that cities assume a Zipf distribution as a result of the economic dynamics which disperse populations among cities. Gulden compares the Zipf-distributed cities of France and the United States, then applies the same model to the distribution of Russian cities which are not Zipf distributed. There is, in fact an overabundance of middle-sized cities compared to a Zipf distribution. Guldens is careful to affirm that a Zipf distribution is not normative but is positive. Different dynamics and different national objective s have merely resulted in different rank/size distributions between the US and France on one hand and Russia on the other. What is important is that Gulden's model is able to replicate the distributions of cities in all three countries fairly well, and that the parameters of the model which lead to these results are theoretically meaningful. The policy implications of these results are twofold. They give an insight into how city sizes may be distributed in Russia should the national political leaders ever choose to stop subsidizing population maintenance in the medium-sized cities. Understanding how this change would accord with the geographical dependency of economic markets could help leaders decide how to shift subsidies to develop infrastructure in certain cities. This model also interjects insight into trends in city distribution that we can continue to expect in developing countries that are continuing to urbanize. Countries that have an urban population that exceeds the number of cities necessary to maintain a Zipf distribution can expect the development of mega cities. Simply attempting to incentivize relocation to medium-sized cities will not work. If countries can expect the evolution of these mega cities, they will need to address potential deficiencies of critical infrastructure and other issues that are related to large populations and high population densities (e.g., pollution).

The final problem that Gulden tackles with an adaptive agent model is spatial and temporal patterns of armed conflict. According to Gulden, "much of the existing literature examining quantitative aspects of civil violence concentrates on risk factors and and searches for correlation between these factors and various indicators of violence." This type of analysis is obviously limited in its ability to analyze the nuances of conflict. The strength of an adaptive agent model in this context is that it has the power to analyze the internal dynamics of conflict. Gulden applies this model to a detailed set of data from the Guatemalan civil conflict that was compiled throughout the conflict's duration (1960-1996). Because conflict dynamics are so complex, Gulden's analysis does not form a comprehensive explanation for the conflict; he seeks, rather to demonstrate the appropriateness of this methodology for analyzing armed conflict. In his model, Gulden uses a ten-year subset of the data from 1977-1986. The violence in Guatemala was mostly purely civil in its nature, but some of the violence was genocidal. When the genocide killings were disaggregated from the broader civil conflict, The killings in the broader conflict followed a Zipf distribution. This disaggregation is theoretically justifiable because different dynamics underlie these two types of violence. The model that Gulden employed was developed by scholars at the Brookings Institute. Although the model is broad, it does show some success at depicting the dynamics of conflict. The important policy implications are that modeling conflicts can help decision makers know which areas would benefit from peace keeper presence and which areas would require other interventions.

Regarding the Zipf distributions within the city and conflict data, Gulden is careful to point out again that there is nothing inherently normative about the Zipf. Some scholars have asserted that the presence of a power law in itself is proof of complex behavior, but Gulden claims that the presence of this particular distribution only indicates that large incidents will be very large, and small incidents will be very small. This is because "a Zipf distribution can, in general terms, be produced by a phenomenon which balances positive feedback (making large events larger) and negative feedback (keeping most events small)." Thus, a Zipf distribution tells us little beyond the broad dynamics of a system.
In conclusion, Gulden says that the adaptive agent approach is especially appropriate for (and I quote verbatim):
• Modeling path dependent processes where the history of the system matters. (Particularly relevant in
the chapter on trade)
• Modeling individual based processes where the heterogeneity of actors matters. (Particularly relevant in
the chapter on civil violence)
• Modeling situations where bounded rationality and imperfect information are fundamental to the
process under study. (Particularly relevant in the chapters on cities and civil violence)
• Managing conserved quantities. (Relevant in all three cases)
• Examining distributional impacts of changes in process or policy. (Relevant in all three cases)

Most importantly, in expanding the analytical tools afforded to researchers, adaptive agent modeling "[expand] the way that problems can be conceived." Because traditional econometrics forces researchers to restrict their formulation of a model, they severely limit how researchers are able to define the problem. Adaptive agent modeling "allows for a richer pre-analytic vision which takes account of history, social organization, and human diversity." In the social sciences, this may represent a huge conceptual leap. Because social scientists are very rarely able to conduct randomized, controlled experiments, their traditional methodology is largely concerned with controlling for the bias that is thus introduced. By expanding their analytical toolbox, researchers may be able to achieve a far more complete view of social phenomenon. This could lead to a far more comprehensive range of policy responses.


Axtell, Robert. (2000). Why Agents? On the Varied Motivations for Agent Computing in the Social
Sciences, CSED Working Paper No. 17.

Epstein, Joshua M., John D. Steinbruner, Miles T. Parker 2001. “Modeling Civil Violence: An Agent-Based
Computational Approach.? Brookings Institution Center on Social and Economic Dynamics Working
Paper No. 20.