A Diversity-Enhanced Neural Network Approach for Wireless Transmission Power Control

Neural network (NN) approaches can accelerate real-time wireless power control by producing instantaneous, iteration-free solutions. However, their performance degrades when a single input configuration admits multiple (sub-)optimal solutions. Recent NN-based methods often reduce this one-to-many input–solution relationship to a one-to-one mapping by selecting a single representative solution per input, thereby failing to capture the diversity of high-quality (near-optimal) solutions associated with the same input. We propose a diversity-enhanced neural network (DENN) that addresses this limitation by transforming the one-to-many input-solution mapping into a constructed mapping from an augmented input, combining the original features with auxiliary information, to the solution space. A regularization term further penalizes solutions inferior to the initial one, steering the model toward higher objective values. Simulation results show that DENN substantially outperforms state-of-the-art NN-based methods in both solution quality and solution diversity.