Drone Swarming Tactics Using Reinforcement Learning and Policy Optimization
Abstract
This project aims to develop defensive drone swarming tactics using reinforcement learning (RL). Swarming is a military tactic where many individually operated units maneuver as one mass to attack an enemy. Defensive swarm tactics are current topics of interest for the US military as other countries and non-state actors are gaining advantages because swarm agents are usually simple, inexpensive, and easy to implement. Current work has already developed the means of flying (drones), communicating, and swarming. However, swarms do not yet have the ability to coordinate an attack against an enemy swarm. We simulated drone battles between two swarms of military fixed wings drones using pre-programmed tactics. Even when outnumbered by up to 100%, there were effective tactics that could overcome the difference in size. When used in defense of a ship, these programmed tactics, on average, allowed between 0 and 0.5 drones to pass the defense and hit the ship which outperforms the current defenses on an Arleigh Burke Class Destroyer and other researched drone swarm defenses. This research shows that it is possible to gain a tactical advantage over an enemy swarm using certain maneuvers and tactics. In order to develop even more effective tactics, we trained an "Agent" tactic using RL. RL is a branch of machine learning that allows an agent to learn an environment, train, and learn which actions that will result in success. The "Agent" tactic does not exhibit emergent behavior yet, but it does kill some enemy drones and outperform other researched RL trained drone swarm tactics.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 12, 2021
- Accession Number
- AD1149672
Entities
People
- Elizabeth K. Gergal
Organizations
- United States Naval Academy