MDP-Based Optimization and Risk-Aware Control of Edge Offloading and Adaptive Inference with Early Exiting for Connected Vehicles

Intelligent Transportation Systems (ITS) require mobile nodes such as vehicles increasingly rely on computationally intensive perception pipelines, mostly based on Deep Neural Networks (DNNs), to support safety-critical functionalities such as object detection, collision avoidance, and cooperative driving. As vehicles operate in highly dynamic and resource-constrained environments, meeting stringent latency and accuracy requirements remains a significant challenge. Edge Computing allows vehicles to offload complex tasks, such as the execution of DNNs, to servers at the edge of the network, thus reducing computing times and energy consumption at the mobile devices. Early Exiting (EE) is an emerging paradigm in deep learning that equips DNNs with intermediate classifiers, enabling a trade-off between inference accuracy and latency. In this work, we investigate the integration of EE mechanisms into edge computing architectures, focusing on use cases involving task execution in resource-constrained computing and communications environments for Connected and Automated Vehicles (CAVs). In particular, we present a unified framework based on Markov Decision Processes (MDPs) for modeling, analyzing, and optimizing the interaction between early exiting, edge offloading and resource allocation in dynamic vehicular environments. Specifically, this thesis presents two main contributions. The first contribution focuses on analyzing the performance of the integrated system, demonstrating the benefits of combining early exiting with edge computing. We formulate a control problem that jointly decides the computation location and the DNN exit points, and we obtain an optimal policy by solving a linear program that captures the trade-offs between latency, accuracy, and task discarded rate. The second contribution extends the considered scenario to a more complex environment, in- cluding different mobile users and explicit radio and computational resources allocation. In this setting, we employ Distributional Reinforcement Learning (DistRL), which allows not only to learn adaptive control policies but also to quantify the uncertainty associated with policy outcomes. This approach provides a more robust framework for decision-making under uncertainty, enabling performance guarantees that account for the variability inherent in real-world ITS environments.