Résumé:
There is a growing need for efficient computational offloading technologies that can handle large data flows while adhering to strict response time
and power consumption constraints as a result of the explosive growth of
Internet of Things (IoT) devices. Real-time IoT applications require scalability and responsiveness, which traditional cloud computing cannot provide. To address these issues, this thesis presents a task offloading framework based on Deep Q-Network (DQN) in a collaborative architecture between IoT, fog computing, and cloud computing levels for intelligent and
adaptive decision-making that maximizes system performance. The DQNbased strategy achieves superior performance, with reductions of up to 50%
in total cost, 33% in latency, and 25% in energy consumption compared to
competing methods, according to extensive simulation experiments against
state-of-the-art algorithms, such as bat, DJA, and DDPG-based approaches.
The DQN algorithm also shows strong convergence behavior, low variance,
and high reliability across multiple scenarios, confirming its robustness and
adaptability in distributed computing environments. In order to improve
Quality of Service (QoS) and Quality of Experience (QoE) in IoT–fog–cloud
ecosystems, this work offers a scalable, data-driven offloading solution, which
advances the expanding field of intelligent task management. The suggested
framework opens perspectives for more independent and energy-conscious
computing paradigms by laying the groundwork for future studies on multiagent deep reinforcement learning, federated offloading techniques, and optical network-enhanced IoT infrastructures.
