Deep Reinforcement Learning for Dynamic Resource Allocation in Multi-Access Edge Computing Environments

International Journal of Computer Science and Technology | Vol. 1, No. 7, July 2020 | pp. 49–56

Research Article

Title: Deep Reinforcement Learning for Dynamic Resource Allocation in Multi-Access Edge Computing Environments

Names of Authors: Kenji Sato¹, Hiroshi Tanaka², and Yuki Takahashi³

Authors’ Affiliations:
¹Department of Computer Science, University of Tokyo, Tokyo, Japan
²Department of Information Physics and Computing, Kyoto University, Kyoto, Japan
³Department of Communications Engineering, Osaka University, Osaka, Japan

Abstract: Multi-Access Edge Computing (MEC) brings computational and storage resources close to mobile users at the network edge, supporting delay-sensitive services like augmented reality, autonomous driving, and industrial automation. However, managing highly fluctuating user workloads and intermittent wireless channel conditions requires adaptive, real-time resource allocation strategies that avoid manual tuning or rigid static provisioning. Traditional optimization techniques struggle with the combinatorial complexity and non-linear dynamics of multi-user MEC systems. This paper investigates a deep reinforcement learning (DRL) framework based on the Deep Q-Network (DQN) algorithm to optimize dynamic CPU frequency allocation and transmission power control in multi-user edge environments. The system models the resource allocation problem as a Markov Decision Process (MDP) where the state space comprises queue backlogs, channel gains, and task execution deadlines, while the action space dictates CPU core allocation levels and uplink power scaling values. Through continuous interaction with a simulated dynamic wireless environment, the agent learns optimal allocation policies that minimize a composite cost function balancing task latency violation penalties and total energy consumption measured in joules (J). Simulation results indicate that the proposed DRL approach achieves a 31.8% reduction in overall system energy consumption and lowers task drop rates by 14.5% compared to greedy and heuristic baseline allocation strategies. The study demonstrates that reinforcement learning provides robust adaptation to non-stationary mobile traffic patterns.

Keywords: Multi-access edge computing, Deep reinforcement learning, Resource allocation, Edge intelligence, Markov decision process, Power control

Manuscript Timeline: Received: April 12, 2020; Revised: May 20, 2020; Accepted: June 15, 2020; Published: July 1, 2020

Citation: Sato, K., Tanaka, H., & Takahashi, Y. (2020). Deep reinforcement learning for dynamic resource allocation in multi-access edge computing environments. International Journal of Computer Science and Technology, 1(7), 49–56. DOI: 10.46882/2020/IJCST/000007

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read