This study introduces a new Deep Unsupervised Learning (DUL) approach based on an optimization problem with box constraints coupled with polytope constraints for maximizing the sum rate in Device-to-Device (D2D) networks, a key factor in enhancing network capacity and efficiency. Current deep learning methods struggle with managing resource-intensive projection steps and need multiple iterations to optimize the sum rate in varying D2D environments. The proposed approach overcomes these challenges by minimizing the loss function and satisfying constraints when dealing with a monotone matrix. The novel approach controls transmit power through a fully connected, multi-layer Deep Neural Network (DNN), solving the complex, non-convex optimization problem associated with optimizing the sum rate in a symmetric interference channel model. The result shows that this method outperforms other power control methods regarding average sum rate, hit rate, and complexity when applied to a standard symmetric K-user Gaussian interference channel.
