<p>Cognitive device-to-device (cD2D) communication is considered a promising solution for improving spectrum utilization and supporting spectrum-efficient Internet of Things services in next-generation wireless networks. This paper investigates a multi-agent deep reinforcement learning (DRL)-based framework for joint channel and power allocation in underlay cD2D cellular networks. The resource allocation problem is formulated as a mixed-integer nonlinear programming problem due to the discrete resource block allocation variables, non-convex signal to interference plus noise ratio constraints, and dynamic wireless channel conditions. To address these challenges, a proximal policy optimization (PPO)-based multi-agent DRL framework is proposed to enable distributed resource allocation under both perfect and imperfect channel state information (CSI) conditions. Two multi-agent learning architectures are investigated, namely decentralized training with decentralized execution (DTDE) and centralized training with decentralized execution (CTDE). The proposed framework is evaluated against several benchmark approaches, including genetic algorithm, random search, deep Q-network, and soft actor-critic. Simulation results demonstrate that the proposed multi-agent PPO framework achieves improved sum-rate performance, stable learning behavior, and robustness under CSI uncertainty. In addition, fairness performance is analyzed using Jain’s fairness index to investigate the impact of different MARL learning architectures on resource distribution among cD2D users. The fairness results indicate that CTDE generally achieves more balanced resource allocation compared with DTDE due to the shared training information available during centralized learning. The results further show that CTDE achieves faster and more stable convergence, whereas DTDE provides fully distributed learning capability suitable for decentralized wireless environments.</p>