With the rapid growth of the number of Internet of Things (IoT) devices, efficient resource allocation becomes increasingly critical to ensure reliable, scalable, and energy-efficient communication. However, heterogeneity of IoT devices, dense network deployments, and the presence of hardware impairments (HWIs) pose significant challenges to conventional resource allocation strategies. IoT networks require a distributed and intelligent platform capable of adapting to dynamic network conditions while managing execution of complex tasks based on diverse application requirements.
In this thesis, we advance the state-of-the-art in radio resource allocation for IoT networks by addressing the combined impact of interference and HWIs. We propose novel optimization frameworks that enhance resource management in IoT networks under time-varying channel conditions and HWI-induced distortions caused by non-ideal devices. Conventional deterministic optimization approaches available to date fail to provide optimal solutions in the presence of such stochastic and unknown impairments that are hard to model and predict. Classical machine learning techniques are also limited, as they typically rely on supervised learning and static datasets, making them unsuitable for sequential decision-making in dynamic environments. Reinforcement learning (RL) overcomes this limitation by enabling agents to learn optimal policies through interaction with the environment without requiring explicit models. However, standard RL methods do not scale well to high-dimensional state and action spaces encountered in large scale IoT networks. To address this, we adopt deep RL (DRL), which leverages deep neural networks as function approximators to capture complex, non-linear relationships and enable scalable decision-making. This makes DRL essential for learning adaptive and efficient resource allocation policies in realistic IoT scenarios.
In Chapter 2, we address the power allocation problem in downlink fog radio access networks-enabled IoT networks, taking into account both HWIs and co-channel interference. Specifically, we propose a distributed resource allocation framework in which each fog access point operates as a DRL agent. These agents dynamically adjust the transmit power of associated devices based on observed network states, with the objective of maximizing overall spectral efficiency. To further enhance learning performance and decision making accuracy, we integrate an ensemble learning strategy enabling the selection of the best-performing model among a set of trained DRL policies. This integration significantly improves convergence speed and robustness in dynamic network environments.
However, while Chapter 2 focuses on maximizing the network capacity of IoT networks under the assumption of transmitting long (theoretically infinite-length) codewords, in delay-constrained IoT networks, the message lengths are typically short, making this assumption impractical. Accordingly, Chapters 3 and 4 focus on the joint clustering and power allocation problem in Industrial Internet of Things networks, accounting for finite blocklength constraints, HWIs, and co-channel interference. To this end, we propose a two-step distributed framework that integrates clustering and power management. In the first step, a greedy clustering algorithm is introduced to group devices into multiple clusters. Greedy clustering is adopted due to its low computational complexity, scalability, and ability to operate with limited local information, making it well-suited for dynamic IoT environments. To address the power allocation problem, we then develop a multi-agent DRL (MADRL)-based algorithm where each cluster operates as an independent agent. This decentralized approach is found to enhance network scalability and performance while ensuring adaptability to dynamic and heterogeneous wireless environments.
Finally, Chapter 5 explores resource allocation in multi-radio access technology IoT networks, addressing both adjacent channel interference and HWIs. Motivated by the decentralized integration of artificial intelligence in sixth-generation IoT systems, we develop a MADRL framework where each IoT device acts as an autonomous agent. These agents continuously learn optimal resource allocation policies—including modulation, transmit power, and coding rate—based on local observations. To facilitate safe and efficient training, we employ a digital twin network enabling the agents to learn in a simulated, risk-free environment. Unlike existing approaches that treat access points or clusters as agents, our framework models each device as an independent learner, thereby enhancing flexibility, scalability, and adaptability. This contribution supports the evolution of intelligent, self-organizing IoT systems and advances the application of distributed learning algorithms in complex, heterogeneous wireless environments.
| Date | 26 Apr 2026 |
|---|
| Original language | American English |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | Georges Kaddoum (Supervisor) |
|---|