Predicting behavior in new behavioral experiments: Outcomes of a modeling competition

This paper reports the results of the inaugural modeling competition sponsored by the Network for Computational SocioEcological Sciences (CoMSES Network). Competition participants were provided with a dataset collected from human-subjects experiments and were asked to develop an agent-based model that replicated behavioral patterns reflected in the data with the goal of using the model to predict behavioral changes in a slightly modified experimental treatment. The data were collected in a resource foraging experiment in which human subjects moved avatars on a computer screen to harvest tokens in a common pool resource. In the original experiments, on which the competition participants based their models, the subjects possessed full information about the state of the resource and the actions of the other group members sharing the resource. The competition challenged participants to predict what would happen if the experimental subjects had limited vision. Using only the data from the original experiment, participants had to design a model that would predict the behavioral changes that would be observed in the new experiment treatment. We compared the models on their assumptions about speed, direction, and harvesting decisions agents make. All the submitted models underestimated the amount of resources harvested. The best performing model was the simplest model submitted and had the best fit with the original dataset provided.