MBPANet: Solving multiple power allocation optimization problems by a universal neural network architecture

Deep learning (DL)-based methods have shown great potentials to solve power allocation optimization problems in the wireless network. However, most of the existing DL-based works are dedicated to satisfy a specific network performance metric. Therefore, for various metrics, different neural network models need to be trained and stored, which cause parameters redundancy and large storage space occupation. For solving the above issues, we propose a universal neural network architecture called MetricBank Power Allocation Neural Network(MBPANet), which can realize power allocation for multiple metrics with a single neural network model. Furthermore, a two-branch training strategy that integrates supervised learning with unsupervised learning is also introduced to improve training efficiency. Simulation results show that our proposed MBPANet can perform power allocation optimization for multiple metrics with fewer parameters compared to the existing dedicated neural network model. Meanwhile, MBPANet can achieve almost a similar or even higher performance than traditional iterative power allocation algorithm with much lower time complexity.