Optimization of Deep Learning based Tone Reservation

Tone reservation is a key technique to solve the high peak-to-average power ratio in orthogonal frequency division multiplexing systems. There are two important problems in tone reservation, i.e., the generation of peak-reducing signal and the selection of the reserved subcarriers. In this paper, the neural network pruning technique is investigated to optimize the indexes of reserved subcarriers. In details, the reserved subcarriers are regarded as neurons in neuron network, and their importance are evaluated to select better subcarriers for tone reservation. Besides, several enhancements for deep learning based model are presented to avoid the overfitting and convergence problems in the training for peak-reducing signal generation. Simulation shows that the proposed methods are effective in improving the performance of deep learning based tone reservation.