We are very happy to announce that the Open Economics Working Group is releasing the datasets of the research project “Small Artificial Human Agents for Virtual Economies“, implemented by Professor David Levine and Professor Yixin Chen at the Washington University of St. Louis and funded by the National Science Foundation [See dedicated webpage].

The authors who have participated in the study have given their permission to publish their data online. We hope that through making this data available online we will aid researchers working in this field. This initiative is motivated by our belief that in order for economic research to be reliable and trusted, it should be possible to reproduce research findings – which is difficult or even impossible without the availability of the data and code. Making material openly available reduces to a minimum the barriers for doing reproducible research.

If you are interested to know more or you like to get help in releasing research data in your field, please contact us at: economics [at] okfn.org

List of Datasets and Code

Project Background

An important need for developing better economic policy prescriptions is an improved method of validating theories. Originally economics depended on field data from surveys and laboratory experiments. An alternative method of validating theories is through the use of artificial or virtual economies. If a virtual world is an adequate description of a real economy, then a good economic theory ought to be able to predict outcomes in that setting. An artificial environment offers enormous advantages over the field and laboratory: complete control – for example, over risk aversion and social preferences – and great speed in creating economies and validating theories. In economics the use of virtual economies can potentially enable us to deal with heterogeneity, with small frictions, and with expectations that are backward looking rather than determined in equilibrium. These are difficult or impractical to combine in existing calibrations or Monte Carlo simulations.

The goal of this project is to build artificial agents by developing computer programs that act like human beings in the laboratory. We focus on the simplest type of problem of interest to economists: the simple one-shot two-player simultaneous move games. There is a wide variety of existing published data on laboratory behavior that will be our primary testing ground for our computer programs. As we achieve greater success with this we want to see if our programs can adapt themselves to changes in the rules: for example, if payments are changed in a certain way, the computer programs will play differently: do people do the same? In some cases we may be able to answer these questions with data from existing studies; in others we will need to conduct our own experimental studies.

There is a great deal of existing research relevant to the current project. The state of the art in the study of virtual economies is agent-based modeling (Bonabeau (2002)). In addition, crucially related are both the theoretical literature on learning in games, and the empirical literature on behavior in the experimental laboratory. From the perspective of theory, the most relevant economic research is Foster and Vohra’s (1999) work on calibrated play and the related work on smooth fictitious play (Fudenberg and Levine (1998)) and regret algorithms (Hart and Mas-Colell (2000)). There is also a relevant literature in the computational game theory literature on regret optimization such as Nisan et al. (2007). Empirical work on human play in the laboratory has two basic threads: the research on first time play such as Nagel (1995) and the hierarchical models of Stahl and Wilson (1994), Costa-Gomes, Crawford, and Broseta (2001) and Camerer, Ho, and Chong (2004). Second are the learning models, most notably the reinforcement learning model of Erev and Roth (1998) and the EWA model (Ho, Camerer, and Chong (2007)). This latter model can be considered state of the art, including as it does both reinforcement and fictitious play type learning and initial play from a cognitive hierarchy.