Generalized Coordinated Learning for Radio Resource Management in Dense D2D Networks
Dense device-to-device (D2D) communication networks with a huge number of directly communicating devices necessitate an efficient radio resource management to maximize network performance. To this end, we first propose a novel intelligent channel reuse based on deep deterministic policy gradient (DDPG) to maximize efficiency of the radio resource usage. In the real-world, any channel reuse inevitably imposes additional interference requiring to i) allocate properly transmission power at individual reused channels and ii) knowledge of a high number of interference channels among devices. Such practical challenges are addressed in existing literature commonly via deep neural networks (DNNs). However, simple coexistence of multiple naturally sub-optimal machine learning models, such as DNN for channel quality prediction, DNN for transmission power allocation, and DDPG for channel reuse leads to a problem with a propagation of inevitable small errors in the prediction by individual models. Consequently, even a small error in the decision taken by one model can mislead decisions taken by other models. To solve this challenge, we further propose a low-complexity Generalized Coordinated Learning (GCL) allowing a coordination of multiple light-weight machine learning models. The GCL employs feedback loops among machine learning models to jointly optimize multiple radio resource management parameters in a coordinated way with awareness of decisions taken by other models. Simulation results demonstrate that the proposed GCL reaches near-optimal performance even in dense D2D networks and improves sum capacity and ratio of users meeting their required communication capacity by up to 69.6% and 22.1%, respectively, compared to state-of-the-art works.