Replicating agent-based models: Revisiting March’s exploration–exploitation study

Abstract

To validate prior agent-based models, public disclosure of model code is necessary, but not sufficient. Conceptual model replication, involving independent reproduction of a model without referring to the originator’s code, is crucial for detecting nonconformity between the publication text and the model implemented. We frame and evaluate four scenarios to expose the shortfalls of replication efforts that use the original study’s program code and neglect the published model description. We identify, evaluate, and offer recommendations regarding disparities between the text and the program code that generated the results reported in March’s classic study of exploration and exploitation in organizational learning. Based on our experience replicating his model, we suggest a two-step method of model verification—beginning with replicating the model from the published description, then turning to the program code of the original implemented model to account for divergent results. Exclusive reliance on either the published model description or the original program code does disservice to theory advancement, because disparities between the original conceptual and implemented models go unaddressed.