Quantum-Driven Deep Reinforcement Learning for Energy-Efficient and Interference-Aware NOMA-Based D2D Groups Underlaying Cellular Networks in 6G

With the rapid proliferation of energy-constrained applications and ultra-dense user deployments in sixth-generation (6G) networks, improving energy efficiency (EE) has become a critical and challenging problem in uplink non-orthogonal multiple access (NOMA)-based device-to-device (D2D) communication underlaying cellular networks. The increasing D2D density, dynamic interference patterns, and heterogeneous service requirements demand intelligent, scalable, and decentralized resource management solutions. To address these challenges, this paper proposes a quantum-driven deep reinforcement learning (QDDRL) framework for joint subchannel (SC) and power allocation aimed at maximizing EE in D2D-enabled 6G systems. The resource allocation problem is formulated as a decentralized Markov decision process, where each D2D transmitter operates as an autonomous learning agent. By integrating quantum amplitude encoding (QAE) and variational quantum circuits (VQCs) within an actor–critic architecture, QDDRL efficiently handles high-dimensional state spaces, accelerates policy convergence, and improves learning stability under dynamic and interference-limited environments. Unlike conventional deep reinforcement learning and optimization-based approaches, QDDRL exploits quantum superposition and parallelism to evaluate multiple state–action combinations simultaneously, reducing exploration overhead and enhancing adaptability. Extensive simulation results demonstrate that QDDRL consistently outperforms existing baseline schemes, including hybrid quantum deep learning (HQDL), quantum reinforcement learning-based dynamic spectrum access (QRL-DS), quantum learning with relay and power selection (QL-RPS), and Quadratic Unconstrained Binary Optimization (QUBO)-based optimization. In particular, QDDRL achieves up to 7.1%, 11.1%, 25%, and 95% higher EE compared to HQDL, QRL-DS, QL-RPS, and QUBO, respectively in dense network scenarios, while also exhibiting faster convergence, lower reward variance, improved robustness to imperfect CSI and residual SIC errors, and reduced control overhead.