基于超限学习机的改进算法研究

发布时间：2018-04-15 01:10

本文选题：神经网络 + 超限学习机　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：近年来,基于神经网络的智能算法由于其在深度学习、智能数据处理及大数据等领域的应用而被广泛的研究。其中,超限学习机与稀疏表示混合分类算法(ELMSRC)在数据识别方面比超限学习机更具有优势,在模型训练时间方面比稀疏表示分类算法更具有优势。然而,仅依赖于超限学习机的输出向量中最大值与次大值之差做分类器筛选在某些应用中具有一定的不可靠性。针对这个问题,本论文提出了竞争机制的超限学习机与稀疏表示分类算法(En-SRC),即在数据分类阶段采用竞争机制的超限学习机(VELM)代替超限学习机(ELM)。另外,基于差异度自调节算法在优化权重向量w阶段采用竞争机制的超限学习机,提出了基于差异度自调节的超限学习机算法(SPLD-ELM);采用正则化竞争机制的超限学习机,提出了基于差异度自调节的正则化超限学习机算法(SP-RELM)。本论文的主要贡献如下:(1)提出了基于竞争机制超限学习机与稀疏表示混合算法(En-SRC)。ELMSRC算法在分类器筛选阶段采用了超限学习机算法,而超限学习机算法在处理某些噪声较高的样本时会面临较差的结果,意味着分类器筛选阶段的可靠性有待提高。因此在分类器筛选阶段,采用竞争机制的超限学习机代替超限学习机进行分类器筛选,进而有效提高分类器筛选的准确性。经实验验证表明:与ELM算法相比,En-SRC算法的识别率提高2%到7%左右;与ELMSRC算法相比,En-SRC算法在取得相同甚至更高识别率的同时,所需测试时间更短。(2)提出了基于差异度自调节的超限学习机算法(SPLD-ELM)。基于差异度自调节算法在模型训练过程中优化权重向量w阶段采用了传统的迭代学习方法,而迭代学习方法存在模型训练时间较长的缺点。为了解决这个问题,本论文提出了SPLD-ELM算法,即在优化权重向量w阶段采用超限学习机算法代替传统的迭代方法。经实验验证表明:与ELM相比,SPLD-ELM算法的计算复杂度有所增加,但识别率得到了有效的提高。
[Abstract]:In recent years, neural network-based intelligent algorithms have been widely studied because of their applications in the fields of deep learning, intelligent data processing and big data.Among them, the hybrid classification algorithm of out-of-limits learning machine and sparse representation algorithm (ELMSRC) has more advantages in data recognition than out-of-limit learning machine, and in model training time, it has more advantages than sparse representation classification algorithm.However, only relying on the difference between the maximum and the second in the output vector of the learning machine has some unreliability in some applications.In order to solve this problem, this paper proposes a new algorithm, En-SRCU, which is used to replace the over-limited learning machine (ELMU) in the stage of data classification by using the over-limited learning machine and the sparse representation classification algorithm (En-SRCU) in the process of data classification.In addition, based on the difference degree self-regulation algorithm, the over-limit learning machine with competition mechanism is adopted in the stage of optimizing the weight vector w, and the algorithm of the over-limit learning machine based on the difference degree self-regulation is put forward, and the regularized competition mechanism is adopted in the over-limited learning machine.In this paper, a regularization algorithm based on differential self-regulation is proposed.The main contributions of this paper are as follows: (1) A hybrid algorithm, En-SRCU. ELMSRC, based on competition mechanism and sparse representation, is proposed.However, the over-limit learning machine algorithm will face poor results when dealing with some samples with high noise, which means that the reliability of the classifier selection stage needs to be improved.Therefore, in the stage of classifier screening, the competition mechanism is used to replace the out-of-limit learning machine for classifier screening, and the accuracy of classifier screening is improved effectively.The experimental results show that the recognition rate of En-SRC algorithm is about 2% to 7% higher than that of ELM algorithm, and the recognition rate of En-SRC algorithm is the same or higher than that of ELMSRC algorithm.In this paper, a new algorithm, SPLD-ELMU, based on self-regulation of difference degree, is proposed.The traditional iterative learning method is used to optimize the weight vector w stage in the process of model training based on the self-adjusting algorithm of difference degree, but the iterative learning method has the disadvantage of long training time of the model.In order to solve this problem, SPLD-ELM algorithm is proposed in this paper, that is, in the stage of optimizing the weight vector w, the over-limit learning machine algorithm is used to replace the traditional iterative method.Experimental results show that the computational complexity of the SPLD-ELM algorithm is higher than that of the ELM algorithm, but the recognition rate is improved effectively.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【参考文献】