Work

Centralized dual-agent DRL for joint region and power resource optimization in mMTC

Year: 2026

Type: article

Abstract: Ensuring energy-efficient transmission while maximizing successful packet decoding is a critical challenge in Massive Machine-Type Communication (mMTC) with Non-Orthogonal Multiple Access (NOMA), due to the high device density, sporadic traffic patterns, and severe interference inherent in grant-free transmissions. This paper proposes a Centralized Dual-Agent Deep Reinforcement Learning (CDA-DRL) framework for joint region partitioning and transmit power pool optimization in uplink grant-free NOMA. The framework employs two cooperative agents located at the base station, one that dynamically selects the number of spatial regions, while the other determines the number of discrete power levels per region. Each agent is trained independently using a recurrent Deep Q-Network (DQN) under a Centralized Training and Decentralized Execution (CTDE) paradigm, enabling scalable and efficient learning with reduced action space complexity. Results show that the proposed CDA-DRL framework outperforms a single-agent baseline and existing state of the art in terms of learning stability, SIC success rate and achieving up to 114\% higher energy efficiency compared to geometric gap-based baselines and 33\% compared to fixed power allocation schemes.

Source: Computer Networks

DOI: 10.1016/j.comnet.2026.112262

Authors: Nasim Ravi, Nuno Lourenço, Marilia Curado

Institution: University of Coimbra

Language: English

Topic: Advanced Wireless Communication Technologies

Subfield: Electrical and Electronic Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,460