基于无监督特征学习的演化计算行为分析

发布时间：2017-12-31 16:04

本文关键词：基于无监督特征学习的演化计算行为分析　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：演化计算作为一类启发式优化方法,其在解决真实世界中的复杂优化问题时的良好性能已经在过去的几十年中得到了很好的验证。但是演化计算自身复杂的随机行为导致对其进行理论分析异常困难,时至今日,仍然难以找到一种有效的方法来对演化算法在不同环境下的行为进行学习和分析。为了更好地理解演化计算的行为,本文尝试采用无监督特征学习的方法,对演化计算在搜索过程中的一代群体行为进行分析。首先对所研究的演化计算行为数据进行定义,然后从基于自组织映射的演化计算行为数据预处理、基于慢特征分析的演化计算行为数据特征提取和基于深度信念网络的演化计算行为数据特征提取三个方面入手,对演化计算的行为数据进行了特征提取和分析。具体工作如下:1)研究了基于自组织映射的演化计算行为数据预处理方法。研究了基于t分布随机邻域嵌入(t-SNE)的自组织映射网络预训练方法,从而将自组织映射网络的训练分为二个阶段:预训练、粗训练和微调三个阶段,使得网络能够收敛到最好的状态。然后使用训练好的自组织映射神经网络将原始高维空间中的演化计算行为数据映射到二维平面上,实现高维数据集的归一化表示,为后续使用无监督特征提取算法对演化计算行为数据进行分析做好数据准备。2)研究了基于慢特征分析算法的演化计算行为数据特征提取算法。首先对慢特征分析算法应用到无监督模式识别问题时的时间序列结构调整进行了研究,同时对需要保留的慢特征维数也进行了分析和计算。然后针对演化计算行为数据的特点,设计了基于二阶非线性扩展慢特征分析算法的特征提取框架,对演化计算行为数据进行特征提取。最后设计了多组对比实验,分别研究了不同演化算法在同样的landscape下的行为特征差异,以及同一演化算法在不同的landscape下的行为特征差异。实验结果表明,慢特征分析算法可以提取到不同演化算法之间具有判别性的稳定特征。3)研究了基于深度信念网络的演化计算行为特征提取算法。首先对深度信念网络的基本组成单元——受限玻尔兹曼机,进行了详细研究。然后针对要分析的演化计算行为数据,设计了一个包含有七层受限玻尔兹曼机网络的深度信念网络框架。最后设计实验得到了不同演化算法在同一个测试函数下的行为数据经过深度信念网络提取到的特征分布结果,将该结果与慢特征分析提取到的特征进行对比,对选用的四种演化算法的行为进行了分析。
[Abstract]:Evolutionary computing is a kind of heuristic optimization method. Its good performance in solving complex optimization problems in the real world has been well verified in the past several ten years. However, the complex stochastic behavior of evolutionary computation makes it extremely difficult to theoretically analyze it. . Up to now, it is still difficult to find an effective way to study and analyze the behavior of evolutionary algorithms in different environments, in order to better understand the behavior of evolutionary computing. This paper attempts to use the unsupervised feature learning method to analyze the behavior of the generation of evolutionary computing in the search process. Firstly, the data of evolutionary computing behavior are defined. Then we preprocess the evolutionary behavior data based on self-organizing mapping. There are three aspects: feature extraction of evolutionary computing behavior data based on slow feature analysis and feature extraction of evolutionary computing behavior data based on deep belief network. The behavior data of evolutionary computing are extracted and analyzed. The main work is as follows: 1) the preprocessing method of evolutionary computing behavior data based on self-organizing mapping is studied, and the random neighborhood embedding based on t-distribution is studied. T-SNE-based self-organizing mapping network pretraining method. Thus, the self-organizing mapping network training is divided into two stages: pre-training, rough training and fine-tuning. The network can converge to the best state, and then use the trained self-organizing mapping neural network to map the evolutionary computing behavior data in the original high-dimensional space to the two-dimensional plane. The normalized representation of high-dimensional data sets is realized. Prepare the data for the analysis of evolutionary computing behavior data using unsupervised feature extraction algorithm. 2). The feature extraction algorithm of evolutionary computing behavior data based on slow feature analysis algorithm is studied. Firstly, the time series structure adjustment of slow feature analysis algorithm is studied when it is applied to unsupervised pattern recognition problem. At the same time, the dimension of slow feature which needs to be preserved is also analyzed and calculated. Then, a feature extraction framework based on second-order nonlinear extended slow feature analysis algorithm is designed according to the characteristics of evolutionary computing behavior data. Finally, we design a number of comparative experiments to study the behavior characteristics of different evolutionary algorithms under the same landscape. And the behavior characteristics of the same evolutionary algorithm under different landscape are different. The experimental results show that. Slow feature analysis algorithm can extract stable features with discriminant property between different evolutionary algorithms. 3). An evolutionary behavior feature extraction algorithm based on deep belief network is studied. Firstly, the constrained Boltzmann machine, which is the basic component of the deep belief network, is studied. Then the evolutionary behavior data to be analyzed are studied in detail. In this paper, we design a framework of deep belief network with seven layers of constrained Boltzmann machine network. Finally, we design experiments to obtain the behavior data of different evolutionary algorithms under the same test function, which are extracted by the deep belief network. Characteristic distribution results. The results are compared with the features extracted by slow feature analysis, and the behavior of the four evolutionary algorithms is analyzed.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【参考文献】