RUI: Experiments in Learning from Others in Games

Williams College, Williamstown MA

Investigators

Abstract

It is a robust experimental finding that agents do in fact use information about the play of similar agents when this information is available, and the amount and type of information available can have a dramatic impact on outcomes. This research project uses experimental and simulation methods to systematically examine the effects of the ability to learn from others on convergence to Nash equilibrium. In addition to opening the possibility to imitative behavior, information about the play of others may alter both the propensity to experiment with new strategies and the strategies experimented with. Whereas knowledge only about the payoffs of others or only about the actions of others may induce or discourage experimentation, the ability to imitate successful others requires both action and payoff information. By systematically changing the amount and type of information available about the play of similar players, this experimental study enables the isolation of the effects from experimentation with new strategies arising from aspiration or conformity from those of imitation of successful strategies. The effects of the ability to learn from others may depend on the underlying structure of the competitive environment. In particular, games with strategic complementarities have robust stability properties, and often have multiple, Pareto-ranked equilibria. In games with a unique Nash equilibrium, learning algorithms consistent with adaptive learning converge to the Nash equilibrium. Insofar as learning from others improves convergence to the Nash equilibrium, this research project assesses whether the ability to learn from others is a complement to, or a substitute for, strategic complementarities. In games with multiple, Pareto-ranked equilibria, this program assesses the impact of learning from others on equilibrium selection. Finally, the project complements experimental studies with the calibration of standard learning models modified to account for the possibility of learning from others. The outcome of this exercise enables researchers to assess the likely effect of learning from others in other environments using data generated from subjects without access to this information. The program has substantial broader impacts. Most learning experimental studies consider a very particular information condition: agents have complete information about the underlying structure of the game, but receive feedback only from their own actions. The results of this program will assist research and policy makers in extrapolating these results to environments and institutions with more realistic information conditions. Additionally, games with strategic complementarities encompass important economic applications such as macroeconomics under imperfect competition, bank runs, and network and adoption externalities. Many games with strategic complementarities are marked by multiple, Pareto-ranked equilibria. This program will assist policy makers in identifying likely outcomes based on information available to participants in the environment. Finally, the successful and timely completion of the experimental program is only possible with the involvement of undergraduate students. Undergraduate students will have the opportunity to participate in most phases of the project, including literature review, assistance in the conduct of experiments and preliminary data analysis. Top undergraduates, particularly those with an interest in continuing their training, will learn the entire process of experimental research, from experimental design, execution, and data analysis to writing and presentation.