Bandwidth-aware federated reinforcement learning for real-time multi-agent autonomous systems under dynamic environments

Abstract The expansion of autonomous systems across fields such as transportation, robotics, and industrial automation demands dependable real-time decision-making, particularly in environments with limited communication resources. The core problem lies in traditional Federated Reinforcement Learning (FRL) frameworks, which often disregard bandwidth constraints, leading to inefficiencies, delays, and degraded performance in dynamic settings. To address this, this study proposes a novel Bandwidth-Aware Federated Reinforcement Learning (BA-FRL) approach tailored for real-time multi-agent autonomous systems functioning in dynamic and bandwidth-variable environments. The solution plan integrates an adaptive synchronization strategy that modulates model sharing based on environmental fluctuations and network capacity, combined with gradient sparsification for efficient updates. A rigorous theoretical foundation is presented, establishing convergence guarantees even under restricted communication conditions. Experimental validation using complex vehicular and aerial multi-agent simulations reveals that BA-FRL can reduce communication overhead by up to 43% while sustaining high task performance. These findings underscore the framework’s practicality and promise for scalable, efficient deployment of autonomous systems in communication-limited real-world scenarios.