Autonomous Agents and Distributed Intelligence Laboratory

Determining the contribution of an agent to a system-level objective function (credit assignment) is a key area of research in cooperative multiagent systems. Similarly, multi-objective optimization is a growing area of research, though mostly focused on single agent settings. Yet, though there is little work on their intersection, many real-world problems are multiagent \em and multi-objective (e.g., air traffic management, scheduling observations across multiple exploration robots).In this work, we leverage recent advances in single-objective multiagent learning to address multi-objective domains. We focus on the impact of difference evaluation functions (which extracts an agent's contribution to the team objective) on the Non-dominated Sorting Genetic Algorithm-II (NSGA-II), a state-of-the-art multi-objective evolutionary algorithm. We derive multiple methods for incorporating difference evaluations in to the NSGA-II framework, and test each in a multiagent rover exploration domain, which is a good surrogate for a wide variety of distributed scheduling and resource gathering problems. We show that how and where difference evaluations are incorporated in the NSGA-II algorithm is critical, and can either provide significant benefits or destroy system performance, depending on how it is used. Median performance of the correctly used difference evaluations dominates best-case performance of NSGA-II in a multiagent multi-objective problem.