Deep Reinforcement Learning for Computation Offloading and Dynamic Resource Allocation in Mobile Edge Computing

Abstract

Mobile edge computing extends computational and storage capabilities toward the network edge to support latency-sensitive and resource-intensive applications running on mobile and Internet of Things devices. A central challenge is determining when computation should be processed locally, transferred to an edge server, or allocated across available network resources under rapidly changing wireless and computational conditions. Conventional optimization methods may become difficult to apply when system dynamics are unknown, high dimensional, or continuously varying. Deep reinforcement learning provides a framework in which agents learn sequential resource-management policies through interaction with the environment. This review examines deep-reinforcement-learning approaches to computation offloading and dynamic resource allocation available through 2021. Foundational value-based and actor-critic techniques - including deep Q-networks, asynchronous actor-critic methods, deterministic policy gradients, proximal policy optimization, and soft actor-critic - are reviewed in relation to edge-computing problems. Applications involving task offloading, wireless channel assignment, power management, vehicular edge computing, and multi-user resource allocation are subsequently examined. The review analyzes state-space design, reward formulation, discrete and continuous action spaces, single- versus multi-agent environments, convergence, energy consumption, service latency, and quality-of-service constraints. Persistent limitations include training instability, exploration cost, non-stationary network environments, generalization across deployment conditions, and the difficulty of validating learned policies in real infrastructure. Nevertheless, by 2021 deep reinforcement learning had emerged as an increasingly important mechanism for adaptive edge-resource orchestration.