Improved extreme learning machine based on quantum genetic algorithm and its application
Li Xueyan1，Liao Yipeng2
(1.College of Artificial Intelligence,Yango University,Fuzhou 350015,China; 2.College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China)
Abstract：Artificial neural network is an important learning method of machine learning,and this paper mainly studies the optimization and improvement of the new training method of neural networkthe algorithm of extreme learning machine.This paper firstly studies traditional neural network algorithms,introduces the main ideas and processes of the algorithm, and compares it with the traditional algorithm to show its characteristics and advantages.Secondly,due to the fact that the algorithm has no small flaws in the accuracy of the prediction and the stability of the application,by describing several intelligent optimization algorithms and comparing their advantages and disadvantages, it introduces the focus of this article quantum genetic algorithm,and uses this algorithm to select the optimal weight and threshold to give the test network,to achieve good results.Finally,the steps and processes of the improved limit learning machine algorithm for experimental simulation and result analysis on MATLAB are introduced.The experimental results show that the improved algorithm has an advantage over the classical algorithm in the prediction of regression problems,with higher prediction accuracy and more stable results.The accuracy of classification is also overwhelming.
Key words :extreme learning machine;quantum genetic algorithm;regression fit;classification;artificial neural networks
0 引言
人工神经网络是机器学习的一种重要学习方式，而对神经网络的研究已经很久了，有些训练算法已经非常成熟，如经典的多层前馈(Back Propagation，BP)神经网络等，应用已经非常广泛，大量地应用于回归拟合和分类问题之中。但是这种被广泛应用于多层前馈神经网络的经典训练算法，大多是基于梯度下降的方式来调整权值和阈值。这类算法的训练速度慢、有可能得到的不是全局最优而是陷入局部最优，还有着参数调整复杂的问题。HUANG G B等人在2004年提出了一种新型的前馈神经网络即极限学习机(ELM)。极限学习机(Extreme Learning Machine,ELM)是用于单隐层神经网络(Single hidden LayerFeedforward Neural networks,SLFNs)训练的一种高效的训练算法。ELM不同于经典的神经网络，它不需要梯度下降算法中繁琐的迭代过程去调参而耗费很多时间。其随机产生所有的权值和隐层节点阈值。并且它在训练中一直不变，需要人为设定的只有节点个数，然后求逆矩阵得到输出权值，便能计算得到最优值。相较于传统的SLFNs，ELM的训练速度显著提升，效率远高于之前算法，且泛化性能好。ELM作为优秀的分类器，拥有良好的应用前景。但是在实际应用中，尤其是在处理回归拟合的问题上，它的效果并不好，准确度一般。为了达到理想的误差精度，ELM需要庞大的隐含层神经元。而由于它的输入权值和阈值是随机设定的，这导致庞大的基数中有很多神经元是无效的，即存在随机出的输入权值和阈值为0。
