University of Michigan – Dearborn

The Dynamics of Driver–System Interaction in Advanced Driver Assistance Systems

Duha Alkurdi, Industrial and Systems Engineering

Advanced Driver Assistance Systems (ADAS) play a critical role in vehicle control, but their effectiveness depends on how drivers understand, trust, and engage with these systems. Even when functions perform as intended, mismatches between system behaviour and driver expectations can undermine trust and disrupt coordination. This dissertation investigates driver-system interaction as a dynamic process shaped by mental models, trust calibration, and mode awareness. Using a multi-phase empirical approach, the research combines naturalistic driving studies and controlled simulator experiments to explore how drivers interpret system capabilities, adapt to different ADAS designs, and manage control transitions. Behavioral, visual, and self-report data are analyzed to identify real-time indicators of uncertainty and coordination breakdowns. The findings contribute to the development of human-centered interface concepts. These concepts provide evidence-based design principles that improve clarity, communicate system state, and support intuitive collaboration between drivers and automation, ultimately promoting safe and effective interaction.

AI-Enhanced UAV-Assisted NextG Networks: Orchestration, Resource Management, and Resilience

Yuhui Wang, Electrical, Electronics, and Computer Engineering

The emergence of computation-intensive and latency-sensitive applications such as autonomous driving, smart agriculture, and extended reality in 5G and beyond (NextG) networks presents unprecedented challenges for real-time data processing, network scalability, and service reliability. Unmanned Aerial Vehicles (UAVs), with their flexible deployment and high mobility, have gained attention as agile aerial base stations and mobile edge computing (MEC) nodes capable of supporting user equipment (UE) in dynamic and infrastructure-scarce environments. However, orchestrating UAV-assisted networks in a manner that balances coverage, computation, and connectivity remains an open problem due to UAVs’ limited onboard resources, fluctuating user demands, and the need for resilient coordination across network layers.

This thesis investigates AI-enhanced orchestration and optimization techniques for UAV-assisted NextG networks, focusing on three core aspects: intelligent UAV placement, joint optimization of positioning and computation offloading, and distributed, dependency-aware cloud-edge collaboration frameworks. First, the thesis examines UAV placement for network formation, exploring both control-theoretic and learning-based approaches. Distributed swarming algorithms are introduced to enable resilient UAV flocking for coverage and inter-UAV connectivity, particularly in spatially dispersed user environments. In contrast, model-free deep reinforcement learning (DRL) strategies are evaluated for adaptive UAV placement under integrated access and backhaul (IAB) constraints, enabling UAVs to autonomously balance fronthaul and backhaul performance. Simulation results demonstrate that our approach successfully establishes a robust UAV network capable of providing seamless coverage for complex user configurations and also ensuring comprehensive inter-cluster connectivity among dispersed user clusters. Additionally, the network exhibits strong resilience against random failures, swiftly recovering from disruptions even when UAVs are compromised.

Building on this foundation, the thesis advances to the joint optimization of UAV positioning and computation offloading in UAV-enabled multi-access edge computing (MEC) networks. This stage addresses the coupled challenge of minimizing end-to-end service latency while respecting UAVs’ limited battery and compute capacity. A bi-level architecture is proposed in which UAV positions are dynamically optimized to reduce path loss to users, and computational tasks are distributed among UAVs using proximal policy optimization (PPO). The learned policies effectively reduce processing latency and improve load balancing across UAVs. Extensive simulations validate each proposed component across realistic UAV-network scenarios, showing significant improvements in user satisfaction, latency reduction, resilience to node failures, and adaptability to dynamic user and task distributions.

To address the complexity of task dependency and scalability of collaborative cloud-edge computing systems, the thesis first proposes a multi-agent task offloading and resource coordination algorithm (ORCA) to enable scalable, distributed offloading and resource provisioning among edge-cloud layers. ORCA leverages multi-agent proximal policy optimization (MAPPO) to enable decentralized, real-time decision making. These agents learn to optimize task distribution, bandwidth allocation, and compute provisioning using only local observations, while coordinating to achieve global system efficiency. The second proposed solution introduces a dependency-aware task offloading framework via Quantum Graph Attention Network-based Deep Reinforcement Learning (QGAT-DRL). Tasks are modeled as directed acyclic graphs (DAGs) to capture subtask dependencies, while system states—including device load, link capacity, and queue status—are encoded through a quantum-enhanced graph attention network that leverages quantum entanglement for expressive feature representation and improved generalization. The offloading decision process is formulated as a Markov decision process and optimized through a PPO algorithm to accelerate convergence and enhance policy exploration. Extensive simulations demonstrate that the proposed frameworks dynamically adapts to varying task demands and server capabilities, and significantly outperforms baseline approaches in terms of energy efficiency, latency reduction, and system scalability.