Multi-objective evolutionary algorithms for distributed tactical control of heterogeneous agents

Abstract

Controlling large numbers of heterogeneous agents in dynamic environments has a number of civilian and defense applications, but presents challenges in cooperation among agents and decision making in uncertain environments. This paper investigates a new problem representation using genetic algorithm tuned potential fields and a new multi-objective problem formulation that evolves distributed control for large numbers of cooperating and competing heterogeneous agents in dynamic environments. Using real-time strategy game-like simulation as a test-bed, results show that the proposed approach scales to controlling a number of and different types of agents. Our representation uses influence maps to choose a target and a set of potential fields to control the maneuverability of agents in real-time. We formulated this problem as a multi-objective optimization problem and used an evolutionary multi objective optimization technique to maximize two conflicting objectives in simulation skirmishes. Results indicate that our evolutionary algorithm based representation produces good cooperative behavior and generalized well across groups composed from several different types of agents.