Considerations for Multiagent Multi-Objective Systems

Abstract

Cooperative multiagent systems are an important area of research, with many applications including air traffic control, satellite communications, and extraplanetary exploration. These applications also lend themselves to multiple simultaneous objectives, but multiobjective multiagent systems have garnered little attention. We make contributions to three distinct aspects of multi-objective multiagent systems: integrating credit assignment into multi-objective systems; providing a computationally efficient technique for dealing with non-convex Pareto fronts; and providing a framework for outside knowledge to be integrated into the agent's reward structure in multi-objective problems. These three contributions together break down some of the largest barriers that have prevented multiagent multi-objective research from gaining traction.