FED-UP: Federated Deep Reinforcement Learning-based UAV Path Planning against Hostile Defense System

Abstract

In military operations, unmanned aerial vehicles (UAVs) have been heavily utilized in recent years. However, due to the antenna installment regulation, UAVs cannot be controlled by human operators in a restricted area. Hence, artificial intelligence (AI)-driven UAVs are the practical solution to this out-of-coverage problem. With the increased use of autonomous UAVs in military applications, defense systems are deployed to target and shoot down the enemy UAVs in operation. Thus, UAVs are needed to be trained, not only to achieve goals but also to avoid static and dynamic hostile defense systems. In this work, we propose FED-UP, a federated deep reinforcement learning (DRL)-based UAV path planning framework, that enables UAVs to carry out missions in a hostile environment with a dynamic defense system. The federated learning (FL) based training accelerates the reinforcement learning process and improves model performance. We additionally introduce significant reply memory buffer (SRMB) to quicken the training process more, by selecting the crucial experiences during the training period. The experimental results validate the efficiency of the proposed model in controlling UAVs in dynamic, hostile environments.