Multi-Objective Optimization in Multiagent Systems

Graduation date: 2015Cooperative multiagent systems are used as solution concepts in many application domains including air traffic control, satellite communications, and extra planetary exploration. As systems become more distributed and complex, we observe three phenomena. First, these systems cannot be accurately modeled, rendering traditional model based control methods inadequate. Second, system parameters are highly coupled in a nonlinear manner, making it difficult for humans to develop heuristic based control policies. Finally, these systems are distributed to the point that a centralized controller is either impractical or infeasible. These types of systems are often inherently multi-objective; unfortunately, they are not treated as such in most multiagent research. To date, there has been little research attention given to multi-objective multiagent systems. \ud \ud This dissertation addresses these systems from a learning-based approach to optimize system performance in four ways: (i) deriving a form of credit assignment compatible for use with multi-objective problems (ii) deriving multiagent equivalents to state-of-the-art multi-objective evolutionary algorithms (MOEAs); (iii) developing a fast, effective multiagent multi-objective algorithm that outperforms state-of the art MOEAs in as little as one tenth of the computation time; and (iv) integrating the previously developed algorithm into a multiagent system