Abstract:

Aiming at the defect of the dependence of BP neural network on label data, a situation assessment method based on deep automatic coding network is proposed. The model uses depth auto-encoder as the basic unit to construct depth auto-coding network, and trains depth auto-coding network with expert experience and hierarchical evaluation method. The network is pre-trained by unsupervised layer-by-layer algorithm using unlabeled data to determine the range space of parameters and weights of each layer of the network. Based on this, the network is fine-tuned by using labeled samples with supervised algorithm so that the parameters and weights of each layer are optimized. As a result a model with the ability to accurately evaluate the input situation data is formed. Compared with BP neural network, the deep auto-coding network model is less affected by labels, which significantly reduces the dependence on expert experience, and has a higher overall evaluation accuracy.

Key words: neural network, situation prediction, deep automatic encoder, deep auto-encoders network, tag

摘要：

针对BP神经网络类方法对标签数据的依赖性缺陷，提出了一种基于深度自动编码网络的态势评估方法。模型应用深度自动编码器作为基本单元构建深度自编码网络，结合专家经验和层次化评估的方法训练深度自编码网络。利用无标签数据采用无监督逐层算法对网络进行预训练，确定网络各层参数及权值的范围空间。在此基础上，采用有监督算法使用有标签样本对网络进行微调，对各层参数及权值进行优化，最终形成具有对输入态势数据进行准确评估能力的模型。多种样本数量条件下的对比实验表明，相对于BP神经网络类方法，基于深度自动编码网络模型受标签的影响较小，明显减少了对专家经验的依赖，并且具有整体上较高的评估精度。

关键词: 神经网络, 态势评估, 深度自动编码器, 深度自编码网络, 标签