Agent-Based Models Assisted by Supervised Learning: A Proposal for Model Specification

Abstract

Agent-based modeling (ABM) has become popular since it allows a direct representation of heterogeneous individual entities, their decisions, and their interactions, in a given space. With the increase in the amount of data in different domains, an opportunity to support the design, implementation, and analysis of these models, using Machine Learning techniques, has emerged. A vast and diverse literature evidences the interest and benefits of this symbiosis, but also exhibits the inadequacy of current specification standards, such as the Overview, Design concepts and Details (ODD) protocol, to cover such diversity and, in consequence, its lack of use. Given the relevance of standard specifications for the sake of reproducible ABMs, this paper proposes an extension to the ODD Protocol to provide a standardized description of the uses of Machine Learning (ML) in supporting agent-based modeling. The extension is based on categorization, a result of a broad, but integrated, review of the literature, considering the purpose of learning, the moment when the learning process is executed, the components of the model affected by learning, and the algorithms and data used in learning. The proposed extension of the ODD protocol allows orderly and transparent communication of ML workflows in ABM, facilitating its understanding and potential replication in other investigations. The presentation of a full-featured agent-based model of tax evasion illustrates the application of the proposed approach where the adoption of machine learning results in an error statistically significantly lower, with a p-value of 0.02 in the Wilcoxon signed-rank test. Furthermore, our analysis provides numerical estimates that reveal the strong impact of the penalty and tax rate on tax evasion. Future work considers other kinds of learning applications, e.g., the calibration of parameters and the analysis of the ABM results.