請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78415
標題: | 藉由深度學習方法找出非正交多重存取之碼書 Finding NOMA Codebook with Deep Learning-Aided Approaches |
作者: | Hsiang-Yu Liu 劉翔瑜 |
指導教授: | 蘇炫榮 |
關鍵字: | 非正交多重存取,深度學習,稀疏碼多重存取,自編碼器,碼書建立,深度神經網路, Nonorthogonal Multiple Access,Deep Learning,Sparse Code Multiple Access,Autoencoder,Codebook Constuction,Deep Neural Net work, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 為了處理當今各種通訊應用對於極高傳輸速度的要求,非正交多重存取這項概念 在第五代行動通訊受到了極大的關注,這個概念大幅增加了頻譜效率。在近期的 研究中,其中一種非正交多重存取技術-稀疏碼多重存取,取得了出眾的表現, 然而,這項技術表現出色的關鍵在於碼書的建立,若要連同現實中出現的各種環 境因素一併考慮的話,碼書的設計將會變得複雜且困難。因此,我們尋求了深度 學習的方法,其中一種就是利用自編碼架構與通訊模型相似的特性,將其套入我 們的系統中,訓練出區域最佳解的碼書及解碼器。在這篇論文中,我們將會更進 一步探討這個架構,並提出幾種簡化模型的想法,於模擬結果中能發現我們的簡 化不僅降低計算量,也達到了一樣低的錯誤率。在論文的最後,我們簡化了在編 碼器中的深度神經模型,將其用一個線性層取代,發現其仍然能達到相同的表 現。從這個奇特的結果,我們推斷由深度模型組成的編碼器(碼書),在訓練的過 程中並沒有好好的利用到非線性的淺力,我們需要重新思考更適合這類通訊系統 的深度架構或是訓練方式。 To cope with the excessive data rate demands of future multimedia ap- plications, nonorthogonal multiple access (NOMA) techniques has drawn sig- ni ciant interests in improving the spectral e eiciency (SE). Recently, sparse code multiple access (SCMA), one of the NOMA schemes, achieves outstand- ing performance. However, the SCMA performance highly depends on the condebook construction, and it is di ucult to construct an optimal codebook for di erent situation in a handcrafted manner. To address this task, there has been some solutions with the help of deep learning. One of the solutions is constructing the encoder codebook and the decoder with the autoencoder structure. In this thesis, we make the performance of autoencoder superior to that of handcrafted codebook by applying some methods and ne tuning the hyperparameters. Then we propose two schemes, the lookup table based encoder and two-stage encoder. The lookup table based encoder not only highly reduces the computational complexity during training but achieves slightly better performance. The two-stage encoder is designed for the situ- ation where there is large number of symbol combinations of di erent users, and it converges more rapidly comparing to the lookup table based encoder. Finally, based on our results, we simplify the deep neuron network (DNN) based encoder of the autoencoder model to a linear operation, and nding that it also has the similar performances with the DNN based encoder. From this peculiar result, we conjecture that the DNN encoder may not exploit the non-linear capacity well, we must explore some other particular structures to suit the NOMA system. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78415 |
DOI: | 10.6342/NTU202000746 |
全文授權: | 有償授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-R06942095-1.pdf 目前未授權公開取用 | 2.3 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。