Core Concepts
Reinforcement learning (RL) can effectively optimize resource allocation and utilization in mobile edge computing (MEC) networks by enabling intelligent decision-making and adaptation to dynamic environments.
Abstract
This paper provides a comprehensive survey on the application of reinforcement learning (RL) in mobile edge computing (MEC) networks. It begins by introducing the concept of MEC and the role of RL in addressing the key challenges faced by MEC, such as low latency, high data rate, massive connections, dynamic uncertainty, and security.
The paper then presents a detailed overview of RL techniques, covering fundamental concepts like Markov decision processes (MDPs), partially observable MDPs (POMDPs), and multi-agent MDPs (MAMDPs). It discusses the basic classification of RL algorithms, including model-based vs. model-free, on-policy vs. off-policy, and online vs. offline approaches. The characteristics, advantages, and limitations of single-agent RL (SARL) and multi-agent RL (MARL) are also examined.
The core of the paper focuses on the application of RL in three key aspects of MEC: task offloading, content caching, and communication. For each aspect, the paper outlines the specific problems, reviews the RL-based solutions proposed in the literature, and highlights the advantages of using RL compared to traditional optimization methods.
The paper then explores several popular application scenarios for RL-empowered MEC systems, including industrial IoT, autonomous driving, robotics, virtual/augmented reality, healthcare, and the tactile internet. These case studies demonstrate the versatility and potential of RL in addressing the diverse requirements and challenges of MEC.
Finally, the paper discusses various future research directions and open challenges in deploying RL in MEC, such as software and hardware platforms, representation of the MEC network, robustness against uncertainties, large-scale resource scheduling, safe RL, generalization and scalability, and privacy concerns. It proposes specific RL techniques to mitigate these issues and provides insights into their practical applications.
Stats
"MEC provides real-time, low-latency, and high-throughput servers, supporting many novel services such as autonomous driving, stream gaming, virtual reality (VR), augmented reality (AR), remote healthcare, etc."
"The requirements of applications are low latency (sub-millisecond level), high data volumes (>1 Gbps), scalability, traffic loads, and security in MEC networks."
"RL can handle complex and dynamic environments, which may be challenging for classical black-box optimization methods, such as genetic algorithms, simulated annealing, semidefinite relaxation, etc."
Quotes
"RL agents can conduct decisions adaptively and apply them to dynamic and uncertain MEC networks (i.e., dynamic task requests and time-varying channel gain)."
"RL executes the end-to-end optimization solving non-convex optimization problems directly. It uses a black-box solver integrating state information blocks into neural network architectures as input layers so that the output of the resource allocation policy."
"RL-empowered mobile devices can adapt to the changing environment and dynamically optimize resource allocation with excellent performance."