Coordinating Team Tactics for Swarm-vs.-Swarm Adversarial Games

Abstract

While swarms of UAVs have received much attention in the last few years, adversarial swarms (i.e., competitive, swarm-vs.-swarm games) have been less well studied. In this dissertation, I investigate the factors influential in team-vs.-team UAV aerial combat scenarios, elucidating the impacts of force concentration and opponent spread in the engagement space. Specifically, this dissertation makes the following contributions: (1) Tactical Analysis: Identifies conditions under which either explicitly-coordinating tactics or decentralized, greedy tactics are superior in engagements as small as 2-vs.-2 and as large as 10-vs.-10, and examines how these patterns change with the quality of the teams' weapons; (2) Coordinating Tactics: Introduces and demonstrates a deep-reinforcement-learning framework that equips agents to learn to use their own and their teammates' situational context to decide which pre-scripted tactics to employ in what situations, and which teammates, if any, to coordinate with throughout the engagement; the efficacy of agents using the neural network trained within this framework outperform baseline tactics in engagements against teams of agents employing baseline tactics in N-vs.-N engagements for N as small as two and as large as 64; and (3) Bio-Inspired Coordination: Discovers through Monte-Carlo agent-based simulations the importance of prioritizing the team's force concentration against the most threatening opponent agents, but also of preserving some resources by deploying a smaller defense force and defending against lower-penalty threats in addition to high-priority threats to maximize the remaining fuel within the defending team's fuel reservoir.