Imitation Learning for Social Simulation

Justin Downes, Hamdi Kavak

Abstract: Modeling the behavior of complex agents is challenging. In social systems, agents have different motivating factors and goals that drive each decision. In some situations, though, we can observe the agent behavior and the outcomes while being at a loss on how to quantify the decision-making process. Imitation learning is a powerful tool to learn behavior without understanding reasoning. In this paper, we explore modern machine learning techniques to train models that imitate agents’ behavior in social environments. This work continues and builds off of emerging work that merges social simulation and modern machine learning. We have shown that such surrogate models can learn heuristically-driven agent behavior. We note that, however, these models do show fragility to changes in environmental dynamics.

External IDs:dblp:conf/sbp-brims/DownesK22