2003 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.

Elenna Dugundji, Social and Behavioral Sciences, U. of Amsterdam (edugundji@fmg.uva.nl).

Elenna is exploring discrete-choice, nested-logit models on networks. Using computational methods, she finds that various network characteristics can be tied to the dynamics of the estimation outcome in the nested logit. She is currently investigating the impact of parameters on forcing phase transitions, adding nonlinear feedback, and means to couple spatial and aspatial models.

Matt Golder, Political Science, New York University (mrg217@nyu.edu).

Matt is investigating voter choice and party system diversity. He is using a computational model of adaptive parties that attempt to win votes in a world populated by various social groups. Each member of a given social group has a similar point of view across the various issues that confront the electorate. He finds that social groups have interesting dynamic consequences for the behavior of the system. For example, as social groups become more diverse, the system is able to find better platforms more quickly. His future work, along with more fully developing the above model, will link the computational results to ongoing empirical work.

Laszlo Gulyas, Informatics, Lorand Eotvos U., Budapest (laszlo.gulyas@sztaki.hu).

Gulyas is implementing endogenous social networks in a model of discrete choice. The basis of the model is that "connected" individuals influence each other's choices, and that these connections may be reconfigured over time. When connections get positively reinforced, he finds that low-density networks tend to get disconnected, segregating the agents. When networks have initial biases, then positive feedback prevents tipping; Under negative feedback, the system also remains resilient to tipping under a variety of reasonable parameter values.

Deddy Koesrindartoto, Dept. of Economics, Iowa State U. (deddypri@iastate.edu).

Deddy employs agent-based modeling to understand better market design in the context of electric power auctions. He considers agents that trade capacity contracts in a discrete, double-auction market. Agents use reinforcement learning, and are confronted by a variety of environments across fundamental market conditions and market structure. He links key elements of market structure to the resulting efficiency of the market and the likelihood of capacity withholding by the agents. Such a methodology, along with improving our understanding of basic market behavior, can also be used to design markets to achieve key social goals.

Nandi Leslie, Dept. of Ecology and Evolutionary Biology, Princeton U. (nleslie@princeton.edu).

Nandi is analyzing the role of profit in land-use change. Her focus is on land use in areas such as the Amazon basin---an area that has recently undergone rapid deforestation. She has taken a basic model of land-use change, based on a probabilistic cellular automata, and introduced profitability of farming as an additional factor in the dynamics. She looks at both exogenous and endogenous profit drivers. Under both of these regimes, she finds that the land-use dynamics can take on interesting, non-periodic oscillatory patterns. In future work she will refine the model using more accurate empirical estimates of profit functions as well as link the results to data from geographical information systems.

Kevin M. Lochner, AI Lab, U. of Michigan (klochner@umich.edu).

Kevin is developing agents for automated price prediction and optimal bidding in multi-auction environments. Agents must trade in a "travel shopping game," in which they need to arrange a complicated package of travel, lodging, and entertainment options, across various auction mechanisms and dynamics. Using agent modeling techniques, his algorithm relies on estimates of the competitive equilibrium. While this technique works well initially, over the course of the game, as other agents acquire goods, it begins to fail. Future work will employ more advanced estimation techniques, such as neural networks, to better infer the hidden states of the other agents---here, previously acquired goods---to improve the performance of the algorithm.

Scott Moser, Social and Decision Sciences, Carnegie Mellon U. (smoser@andrew.cmu.edu).

Scott is investigating the origins of strategic situations, that is, how do the games that agents play arise. To model this situation, he uses computational methods to co-evolve games and strategies: agents play a series of games and then vote for the game that they most prefer. His initial findings indicate that of all possible games in a taxonomy, there is a large class that persists after the coevolution. This class includes some obvious candidates, for example, dominance solvable games, but it also contains some more surprising outcomes. In future work, he will be imposing more constraints on the allowable game structure to better reveal how games, and at a more deeper level, institutions, arise and persist in social worlds.

Suresh Naidu, Economics, U. Mass. at Amherst (snaidu@econs.umass.edu).

Suresh is modeling group formation and dissolution in open source software development projects. He looks at the interaction between project governance and software architecture, and models the system using agent-based techniques. He finds that the resulting project structure can be closely tied both to governance and the intrinsic motivation of the agents. Along with refining the model, he intends to test the predictions using data drawn from archives of open source development trees.

William Rand, Center for the Study of Complex Systems, U. of Michigan (wrand@umich.edu).

Bill is looking at how to evolve heterogeneous, decentralized systems that can perform productive computations. His base model considers a one-dimensional, non-uniform cellular automaton. He considers the density classification problem, and uses a genetic algorithm to evolve the rules over time. His initial findings indicate that heterogeneity does not appear to improve performance in this problem, except under some very special conditions. Notwithstanding this initial result, the work does hint at key features that may favor heterogeneity in decentralized problem solving.

Eugen Tereanu, Dept. of Economics, Johns Hopkins U. (eugen@jhu.edu).

Eugen is studying expectations formation, driven by a contagion process, in macro-economic models. Agents get "infected" with key economic information when they come in contact with the media, and based on this information update their behavior. His main interest is in the social transmission of expectations, which he explores by implementing social networks ranging from random mixing to purely local neighborhoods. He finds that different network structures can lead to very different empirical implications for the model. He intends to apply these results to issues of growth and development in transitional economies.

C. Jason Woodard, Harvard Business School, Harvard U. (jwoodard@hbs.edu).

Jason is evolving networks based on agents attempting to improve their payoffs by reconfiguring their connections to other agents. The key question he wishes to answer is how effectively can agents form networks in situations of strategic interdependence. In the model, agents adapt their connections to other agents in an attempt to improve their payoffs. By altering the payoff function, a variety of different dynamics and final network structures emerges. Moreover, intriguing differences in final network structure can be associated with the directionality of connections and the number of agents in the world.

John H. Miller , miller@santafe.edu.