1998 Graduate Workshop in Computational Economics

Student Projects

Each student began a research project during the two-week workshop. Below are brief descriptions of these various projects. These projects will form the basis for dissertation chapters and/or journal articles.

Ravinder Bhavnani, University of Michigan (rvib@umich.edu).

Ravi would like to understand how institutions can emerge and promote cooperation among agents, and he is investigating a variety of pathways for such institutional formation. For example, entrepreneurs may make investments into "trading clubs" in which agents, after paying a fee, are allowed to interact with other members of the club. The model indicates that the emergence of such institutions can allow high levels of cooperation to develop in the system.

Sylvie Geisendorf, Kassel (geisendorf@wirtschaft.uni-kassel.de).

Sylvie is exploring the behavior of adaptive agents who must compete with one another in acquiring a renewable, but fragile, resource. The model considers a fishery where adaptive agents decide on the size of the boat (catching potential) they wish to acquire based on past experience. Analysis of the behavior of the model should increase our understanding of tightly coupled economic and ecological systems, illuminate the conditions under which agents learn to sustainably use resources, and provide insight into the effectiveness of various policy options.

Iris Ginzburg, C&L (iris.ginzburg@us.coopers.com)

Iris has modeled a system in which firms have the option of sharing information with other firms in the industry. The flow of such information may be carried by agents who have knowledge of each firm's operations, for example consultants. The model implies that the sharing of information is most likely in environments where firms are frequently exposed to catastrophic events.

Asim Khwaja, Harvard University (akhwaja@kuznets.harvard.edu).

Asim has developed a model that can be used to understand how explicit and implicit contracts can emerge in society. In the model, information about each agent's past behavior flows through the network of friendship relations. This information allows each agent to form beliefs about the other agents, and based on these beliefs decide on how to form contracts with one another. Depending on certian parameters, either contractual form may arise and support different informational regimes.

Susan Lee, University of Wisconsin (susanlee@students.wisc.edu).

Sue has analyzed a model in which agents undergo assortative matching before playing a coordination game---thus allowing interaction networks to form endogenously. The underlying analytic and computational models allow her to investigate central issues surrounding income inequality. The inequality that results from the model can be linked to a variety of exogenous factors, for example, the strength of the assortative matching mechanism, the memory of the agents, etc.

Nienke Oomes, University of Wisconsin (noomes@ssc.wisc.edu).

Nienke has built a dynamic, discrete choice model with local interactions. Agents, based on information about their neighbors' and their own past behavior, decide whether or not to seek work during a given time period. Employed agents in a particular area earn money and contribute to the local economy. In general, the model generates classes of persistent, well defined employment patterns. The actual characteristics of the final pattern can be linked to just a few key parameters in the model.

Alexander Outkin, Virginia Tech, (outkin@vt.edu).

Sasha is attempting to understand how information flow across local interactions can result in alliance formation. Firms, each with a particular spatial location, attempt to compete with one another by offering goods that better match the preferences of their local customers. Even though neighboring firms compete for the same customers, they still may find it in their interest to share information with their joint customers since such information allows them to better compete against their other neighbors. Links between local and global processing are emerging as key to understanding this type of system.

Dawn Parker, University of California---Davis (parker@primal.ucdavis.edu).

Dawn is investigating how local spatial externalities impact land use choices. In the model, generators (for example, pesticide-based agriculture) impose a negative externality on their immediate neighbors (for example, organic farmers). The model allows bargaining between immediate neighbors, and preliminary results indicate that the dynamic behavior and ultimate land-use patterns emerging from the model can be tied to exogenous price and productivity parameters as well as relative bargaining power. Moreover, it appears that decentralized markets and bargaining often imply very inefficient outcomes.

John Patty, Cal Tech (jwpatty@steuben.caltech.edu).

John wants to understand how political parties and institutions emerge and evolve. In the model, political actors have limited information (in the form of agent-specific information partitions) about the key, underlying characteristics of the system. If an agent trades this information with another agent, both can benefit as the information partition becomes more refined. The model links the likelihood of trade to the similarity of the agents' past behavior. These information exchange networks are sufficient to promote party formation.

Jianping Shen, C&L (jianping.shen@us.coopers.com).

Jianping developed a genetic algorithm that is able to sort arrays of agents attempting to solve a coordination problem. The algorithm, relying on the spatial structure imposed by the underlying problem, uses a spatially oriented cross over operator, and was able to very quickly find ideal sortings while simultaneously moving all of the agents.

Justin Smith, Oxford (justin.smith@economics.ox.ac.uk).

Justin designed a program that attempts to beat a human player in a game of matching pennies. The program exploits patterns in the play of the human to gain an advantage in the game. Preliminary results show that if humans have no feedback on the play of the game, they can be easily exploited by the machine. However, even when the humans have full feedback, the machine can often defeat them. The model is being extended so that issues of human pattern recognition and learning can be investigated further.

James Thomas, Carnegie Mellon University (jthomas@cs.cmu.edu).

James is analyzing how decentralized organizations can solve non-decomposable problems. He has developed a simple framework from which to study how problems of various difficulties can be solved by both centralized and decentralized entities. The model indicates that even small levels of noise in the evaluation function result can imply a big advantage to decentralization. The basic framework should have ready applications to both human and artificial systems.

William Vachula, Wharton (vachula@wharton.upenn.edu).

Bill created an abstract model of contract negotiations. In the model, buyers and sellers must agree on 10 negotiation issues---each of which can be resolved in a way that will favor either the buyer or the seller in the event of an exogenous failure on the issue in question. Initial results indicate that as the cost of issue failure increases, the number of agreed upon contracts declines as agents take bargaining stances that become more rigid.

William Walsh, University of Michigan (wew@engin.umich.edu).

Bill is attempting to solve task allocation problems that arise in computer systems by creating a decentralized market mechanism in which computational agents bid for resources, and in so doing create multi-stage supply chains. Agents must confront both local knowledge and communication, as well as resources and final goods that are time dependent. He finds that the market mechanism appears to be both quicker and more efficient than techniques that are currently used to solve such problems.

James Warnick, University of Pittsburgh (jcwst22+@pitt.edu).

Jim uses symbolic regression techniques (based on a genetic program) to analyze the role of expectation formation in games. In the model, agents attempt to induct the behavior of opponents through regression and then, using this information, best respond. Though agents use deterministic predictors, behavior that resembles mixed strategy emerges on long time scales. He also found that when one agent is allowed to use more elaborate expectation formation mechanisms significant performance advantages can accrue. However, even small amounts of noise in the system dramatically mitigate this advantage.

John H. Miller , miller@zia.hss.cmu.edu.