Multi-agent Target Defense Game with Learned Defender to Attacker Assignment

Abstract

This paper considers a variant of pursuit-evasion games where multiple attacker unmanned aerial vehicles (UAVs) are trying to converge on a target. The goal is to use a set of defender UAVs to save the target by ensuring they converge on the attackers before the latter converge to the target. The core challenge lies in appropriately assigning a particular defender to an attacker. The simple heuristic assignment based on Euclidean distance between the attacker and defender performs poorly. This paper presents a data-driven solution assuming that the attacker uses a known optimal control policy. We show how massive offline simulations can be leveraged to predict the optimal cost/value function incurred by the defender to converge on an attacker for a given target trajectory. We use this optimal cost/value function as a true measure of separation between an attacker and a defender. We use it as the guiding heuristic in the Hungarian algorithm for computing defender-attacker assignments. We perform extensive simulations to validate our approach wherein we couple the learned assignment with a non-linear model predictive controller to perform realistic simulations. We show that our assignment approach outperforms that based on the Euclidean heuristic in terms of the number of successful attempts by the defenders.