Guided Evolutionary Games for Influencing Interacting Agents' Behavior in Large Populations

点击翻译按钮获取中文摘要

Swarm robotics is the use of a large number of small, simple robots to accomplish common goals, mimicking the cooperative task sharing observed naturally in hives and colonies. As this technology becomes a reality, there is an increasing need to develop methods for guiding and controlling the swarm. The standard approach to controlling a system of multiple agents relies on treating them as approximately separate, however other paradigms will be needed as the number of these agents becomes unmanageably large, and their interactions dominate their behavior. The growing success of evolutionary game theory in biology and ecology suggests a mechanism for control at the population level, in which individual interactions are manipulated to drive the agent community towards global goals. This award is concerned with developing and studying these guided evolutionary games, and testing their consequences in simulations of interacting agents. Applications include the control of microscopic robot swarms in medical applications and the guided management of online social phenomena, including the development of active influence mechanisms for decreasing negative online behaviors like cyberbullying. This project will provide valuable interdisciplinary training for young researchers, and target outreach activities towards high school students, and undergraduates.This project takes the first steps in bridging the gap between evolutionary game theory, massively distributed mechanism design, and the design of control systems. The work focuses on semi-autonomous agents who interact with each other through localized communication, and describes their behavior via evolutionary games using either replicator or discontinuous imitation dynamics. Communications play out either on a network or in Euclidean space; both topologies increase the problem complexity and can lead to organized spatial patterns. Connections will be made systematically between microscale, agent-based simulations, and macroscale, density-based evolutionary equations, such as the replicator and other alternatives. The dynamics of hybrid controllers, that modify the game played by the interacting agents, will also be considered. Control of the population is accomplished by periodic actuation of the game governing the interactions. The goals of this project are to understand the control-theoretic preliminaries necessary to allow equilibrium shaping in order to control populations, to determine the theoretical limits of control in this setting, and finally to apply this approach to the control of autonomous agents in high-fidelity simulations.