Date of Award

6-26-2026

Date Published

July 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Qinru Qiu

Keywords

Communication Efficiency;Explainable Artificial Intelligence (XAI);Machine Unlearning;Multi-agent Systems (MAS);Reinforcement Learning (RL)

Subject Categories

Computer Sciences | Physical Sciences and Mathematics

Abstract

The rapid proliferation of autonomous systems, such as Unmanned Aircraft Systems (UAS) transitioning to large-scale Beyond Visual Line of Sight (BVLOS) operations, demands a paradigm shift toward reliable, transparent, and resilient autonomy. While Deep Reinforcement Learning (DRL) and Multi-Agent Reinforcement Learning (MARL) have demonstrated exceptional capabilities in complex decision-making and coordination, their real-world deployment in safety-critical domains is hindered by two fundamental challenges. First, the opaque, "black-box" nature of DRL models prevents human operators from understanding and trusting the agents' underlying logic. Second, as multi-agent operations scale, they become severely constrained by the stochastic link qualities and strict bandwidth limitations of physical communication infrastructures. To overcome these barriers, this dissertation proposes a comprehensive algorithmic ecosystem built upon two core pillars: Trustworthy Reinforcement Learning and Communication-Efficient Multi-Agent Systems. Under the first pillar, Trustworthy Reinforcement Learning, this research bridges the interpretability gap to establish human trust in autonomous decision-making. We introduce VisionMask, a novel, agent-agnostic contrastive explanation framework. Unlike traditional perturbation-based Explainable AI (XAI) methods, VisionMask leverages self-supervised action-wise and feature-wise contrastive learning to generate highly faithful, robust, and sparse action-specific saliency maps for discrete environments. To accommodate the complex, continuous control mechanisms required for realistic flight dynamics, this methodology is extended via Continuous VisionMask (cVM). By employing state vector quantization and optimizing Kullback-Leibler divergence for continuous policy distributions, cVM effectively isolates the critical visual and state features driving an agent’s behavior. Together, these frameworks provide transparent rationales, empowering operators to intuitively understand the "why" behind specific autonomous maneuvers. The second pillar ensures the robust deployment of these transparent agents by establishing Communication-Efficient Multi-Agent Systems. Grounding autonomous behaviors in physical network realities, we develop an integrated co-simulation platform that couples air traffic mobility logic (MATRUS) with high-fidelity, 3GPP-compliant cellular network protocols (ns-3). Leveraging this platform, we propose the SINR Aware Predictive Planning (SAPP) framework and the CommA* algorithm. SAPP utilizes neural networks to accurately predict signal-to-interference-plus-noise ratios (SINR), enabling CommA* to treat dynamic link quality as a routing cost and proactively navigate agents through secure communication corridors. Furthermore, to optimize multi-agent coordination under strict bandwidth constraints, we pioneer MUTE (Message Unlearning for Targeted Efficiency). MUTE reframes communication reduction as a targeted machine unlearning problem. By estimating the Counterfactual Message Value (CMV) of shared signals, MUTE systematically unlearn redundant communication channels via behavioral anchoring. This return-preserving approach achieves up to a 99% reduction in network bandwidth while maintaining near-optimal cooperative task performance. Ultimately, this dissertation provides a unified methodology that seamlessly integrates explainable AI with rigorous, bandwidth-aware networking. By ensuring that autonomous decisions are both logically transparent to human operators and efficiently coordinated within physical communication constraints, this work lays the critical foundation for the safe, scalable, and trustworthy deployment of multi-agent systems in complex real-world environments.

Access

Open Access

Download

Share

COinS