Resilient Resource Allocation in OFDMA Systems: A Scalable Recurrent MARL Framework for Channel Aging Compensation

Abstract

In downlink orthogonal frequency division multiple access (OFDMA) systems, efficient joint optimization of user scheduling and link adaptation is crucial for meeting the high data rates and connectivity demands of future wireless communications. However, the inherent feedback delays in practical systems cause channel state information (CSI) acquired by base stations to become outdated, a phenomenon known as channel aging. This temporal state mismatch leads to erroneous selection of modulation and coding schemes (MCS) and ineffective user scheduling, resulting in significant degradation of system throughput and user fairness. To address this challenge, this paper proposes a Resilient Resource Allocation Strategy based on a scalable recurrent multi-agent reinforcement learning (MARL) framework. To ensure scalability, we formulate the problem as a decentralized partially observable Markov decision process (Dec-POMDP), where a multi-agent approach is used to decompose the global action space, allowing for distributed scalability and avoiding the exponential complexity of centralized optimization. Temporal processing capabilities are integrated into the agents via gated recurrent units. This framework implicitly predicts real-time channel states using historical observations, effectively overcoming the challenge of partially observable states. Simulation results demonstrate that the proposed approach maintains robust performance under varying channel aging conditions. Compared to traditional baseline algorithms, the proposed algorithm achieves an average 14.47% improvement in proportional fairness utility, with performance gains reaching up to 22.31% under severe channel aging conditions. This framework effectively balances user fairness while significantly enhancing system throughput and transmission reliability.