面向频谱大数据处理的机器学习方法
Machine Learning Methods for Big Spectrum Data Processing

随着移动互联网与物联网的迅猛发展，个人线设备的数 量呈现指数级增长，随之产生的海量频谱数据与日俱增，频谱大数据的存在已成事实。同时 ，频谱赤字也日益严峻。为提高频谱利用率，有效的频谱大数据处理显得十分重要。本文从 线通信的角度，首先给出了频谱大数据的定义并分析了它的基本特征；然后总结了一些 对于频谱大数据分析与利用颇具前景的机器学习方法，如分布式和并行式学习、极速学习、 核学习、深度学习、强化学习、博弈学习和迁移学习；最后给出了几个开放性话题和研究 趋势。
With the rapid development of the mobile Internet a nd the Internet of Things, the number of personal wireless devices has grown exp o nentially, result ing in the increase of massive spectrum data. Therefore, the bi g spectrum data are literally formed. Meanwhile, the spect rum deficit is also increasingly precarious. Effective big spectrum data process ing is significant in improving the spectrum utilization . Firstly, fr om a perspective of wireless communication, a definition of big spectrum data is presented and its characteristics are also analyzed. Th en, p romising machine learning methods to analyze and utilize the big spectrum data are summarized, such as, the distributed and parallel learning, extreme lea rning machine, kernel b a sed learning, deep learning, reinforcement learning, game learning, and transfer learning. Finally, several open issues and research trends are addressed.