Soft Decision Decoding For Block Codes Using Artificial Neural Networks

[216686] Wicker 20世纪80年代初，人们对人工神经网络（ANN’s）的兴趣复兴，刺激了对其在许多工程领域应用的研究，包括数字信号处理和通信。人工神经网络的模式识别能力允许它们模拟各种功能。这种仿真通过使用训练算法（如反向传播）得到增强，这种算法允许某些类型的人工神经网络“学习”所需的函数。本文研究了人工神经网络在分组错误控制码软判决译码中的应用。在过去的两年里，实验研究表明，某些人工神经网络可以对一些分组码进行错误控制解码。实验结果表明，人工神经网络可以进行硬决策解码，但软决策解码的结果非常有限。在本研究中，误差控制问题首先转化为泛函分析的术语。解码被看作是从连续接收信号空间到离散信息字空间的映射。解码器的设计因此变成了一个函数近似的问题。下一步是使用一种非常有前途的人工神经网络，前馈神经网络（FFNN），以一种有效的方式实现解码功能。FFNN的操作可以从几何角度来看待，网络中的每一层都将接收到的信号空间分割成越来越低维的子空间，最终形成一个输出决策层，提供对传输信息位的估计。因此，函数逼近方法允许理想的软判决解码，而FFNN设计使用块码中的代数-几何冗余来降低解码器的复杂性。这些解码器的性能将通过使用人工神经网络训练技术进一步增强，该技术将在解码器运行时将输入层决策指标与信道条件相匹配。其结果将是在各种应用中为各种分组码提供一系列高效的软判决解码器。＊＊＊

9216686 Wicker The revival of interest in artificial neural networks (ANN's) in the early 1980's has spurred research into their application in many engineering fields, including digital signal processing and communications. The pattern recognition capabilities of ANN's allows them to emulate a variety of functions. This emulation is enhanced through the use of training algorithms, such as back-propagation, that allow some types of ANN's to "learn" desired functions. This research is an investigation into the application of ANN's to the soft decision decoding of block error control codes. In the past two years, experimental studies have demonstrated that certain ANN's can perform error control decoding for some block codes. The experimental results have so far shown that ANN's can perform hard decision decoding, but soft decision decoding results have been extremely limited. In this investigation, the error control problem is first translated into the terms of functional analysis. Decoding is viewed as a mapping from a continuous received signal space onto a discrete information word space. Decoder design is thus translated into a problem of functional approximation. The next step is to use a highly promising class of ANN's, feedforward neural networks (FFNN's), to implement the decoding function in an efficient manner. The operation of FFNN's can be viewed geometrically, with each layer in the network carving up the received signal space into increasingly lower dimensional subspaces, culminating in an output decision layer that provides estimates of the transmitted information bits. The functional approximation approach thus allows for ideal soft decision decoding, while the FFNN design uses the algebro-geometric redundancy within the block code to reduce the decoder complexity. The performance of these decoders will be further enhanced through the use of ANN training techniques that will match input layer decision metrics to channel conditions w hile the decoder is operating. The result will be a series of efficient soft decision decoders for a variety of block codes in a variety of applications. ***