Optimizing Transitions between Abstract ABM Demonstrations

Agent-based models (ABMs) involve large numbers of individual agents, each governed by a common behavior program (Agent-Level Parameters, or ALPs), whose collective behavior (System-Level Parameters, or SLPs) is emergent due to interactions among the agents and the environment. Applications of ABMs include modeling the spread of epidemics, supply chain optimization, and representing the dynamics of financial markets. A typical application involves specifying one ALP to get a desired SLP. In this work, we explore emergent behavior sequences, such as a swarm of drones transitioning from broad area search to focused search to airlifting disaster victims. The central question is how one achieves graceful and ef?cient changes between SLPs by manipulating ALPs. We explore three different ways of transitioning between ALPs and observe their behavior on SLPs, with the goal of fast and stable convergence on the desired SLPs. All of the empirical work is done in an existing framework that allows users to specify ALPs by demonstrating desired SLPs, thereby removing the need for deep ABM knowledge on the part of users.