Interaction Biases in Multi-Agent Simulations: An Experimental Study

摘要

How to ensure that two different implementations of a simulation will produce the same results? In order to assure simulation reproducibility, some domain-independent functional unit must be precisely described. We show in this paper that the management unit that rules the participation of an agent in simultaneous interactions is one of them. Usually, many choices concerning this unit are made implicitly, even if they might lead to many simulation biases. We illustrate this issue through a study of biases that appear even in simple cases, due to a specification lack, and we propose as a solution a classification of interactions that makes those choices explicit.