本論文將深度神經學習(DNN)結合於正交分頻多址(OFDMA)系統的子載波配置中，透過適當的學習來提升分配的效率。在正交分頻多址系統中，假設通道增益為已知的，分給不同數量的人數，比較其所需的傳送功率。論文中提出的方法可以大幅地提升效率和減少運算複雜度，我們將一組通道增益分給使用者視為一個批量去學習，透過一定數量的迭代及重複學習，成本函數會收斂至一穩定的數值，並且在滿足位元錯誤率的限制下最小化與ESA演算法的均方誤差差異。我們同時也比較不同的最佳化演算法的收斂速度，最後透過驗證來設定任何可能的控制參數，使用測試集來評估分配的精確度與效能。本論文提出的方法提供了更高的效率於分配子載波中，效能也與ESA演算法相近。;In this paper, we propose a deep neural networks (DNN) structure to allocate subcarrier for orthogonal frequency-division multiple access (OFDMA). Assuming that the channel gains of all subcarriers are known, and allocate to different number of users respectively. The proposed method can be dramatically increased the efficiency. We trying to minimize the mean squared error (MSE) between ESA algorithm while satisfying the bit error rate constraint. We suggest a deep learning architecture in which each group of allocation as a batch, after an appropriate number of iterations and epochs, the loss will tend to converge to a constant value. We also discuss different optimizer to compare their convergence rate. The proposed scheme offers better efficiency of allocating subcarrier and the performance is close to ESA algorithm.
