|
Full title—Deep Reinforcement Learning-Enabled Performance Gap-Driven Load Balancing in Air-Ground Integrated Networks
Load balancing (LB) has been widely investigated in air-ground integrated networks (AGINs), which integrate the terrestrial layer (TL) with the aerial layer (AL); however, the existing literature largely ignores the fact that some offloaded users may have already met their performance requirements at the TL, making their offloading unnecessary or inefficient.
In this paper, we propose a performance gap-driven cross-layer LB scheme for AGINs, where the gap is quantified as a dissatisfaction ratio between the achieved performance and the target requirement for three different service types: wide-coverage service, high-throughput service, and low-latency service. In this LB scheme, when the TL is overloaded, users exhibiting the largest performance gaps are prioritised for offloading. A system-wide utility maximization problem is formulated to jointly optimize the throughput, delay, and coverage of the three considered service types, respectively.
To solve this problem, we devise a novel conditional independent deep reinforcement learning algorithm, in which two independent agents perform resource allocation for the TL and AL, respectively, and the AL agent is activated only upon detection of TL overload. Numerical results show the advantages of our proposed LB scheme over the benchmarks in terms of throughput, delay, and coverage.
Full Article: IEEE Transactions on Vehicular Technology, Early Access
|
