【作者】 李欣；

【导师】 程春田；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2010， 博士

【摘要】 量子计算是信息科学和量子理论相结合的新兴交叉学科。依托量子计算基本原理产生的量子算法,以其独特的优化性能受到世界各国学者的普遍关注,并以显示出十分广阔的应用前景。过程神经元网络是根据生物神经系统信息处理机制并结合实际问题的应用背景提出的一种新的人工神经网络模型。网络的过程式输入放宽了传统神经元网络模型对输入的同步瞬时限制,是传统神经元网络在时间域上的扩展,是更一般化的人工神经元网络模型。本论文重点研究了量子计算与神经网络和智能优化算法的融合机制，以及过程神经网络训练算法,主要研究内容可归纳如下：第一,通过将量子计算和神经网络理论相融合,提出一种基于量子旋转门和受控非门的量子BP网络模型及算法,证明了该模型的连续性,通过太阳黑子数预测的实验结果表明,该模型的非线性预测能力明显优于普通BP网络；通过模拟生物神经元的信息处理机制,提出一种具有量子权值和量子活性值的神经网络模型及超线性收敛学习算法,通过模式识别和函数逼近验证了模型的有效性；通过分析量子门物理意义,提出一种量子门节点神经网络模型及算法,该模型具有隐含的多重吸引子,可明显增加收敛概率,仿真结果表明其收敛速度和预测能力明显优于普通神经网络。第二,针对过程神经网络模型的训练问题,提出一种基于勒让德正交基函数的过程神经网络学习算法,该方法可有效解决复杂的时空聚合运算问题；针对海量样本的并行处理问题,提出了并联过程神经网络预测模型及算法,该方法既能分散网络负载,又能提高单个网络的预测能力。第三,在分析目前量子搜索算法存在问题的基础上,通过构造新的量子搜索引擎,提出了两种改进算法。首先,提出了一种目标加权搜索算法,该算法可使得到每个目标的概率与目标的重要程度相一致,使越是重要的目标,成功概率越大；其次,提出了一种基于小相位旋转的量子搜索算法,通过将旋转相位固定为0.01π,可使算法的成功概率最低为99.99%。第四,通过分析目前量子进化算法存在的问题,提出一种基于量子位球面坐标的量子遗传算法,设计了两种新的量子门算子。在该算法中,通过将量子位不同维的坐标均视为优化问题的近似解,该算法可以增加对解空间的遍历性,提高收敛的概率。通过直接将量子位的相位视为待优化个体的编码,分别提出了基于相位编码的量子遗传算法、量子免疫算法、量子蚁群算法和量子粒子群算法。在这些算法中,由于优化过程统一在[-1,1]n或[0,2π]n进行,与具体问题无关,因此对不同尺度空间的优化问题具有良好的适应性。通过将这些算法融合到过程神经网络的训练中,可明显提高模型的计算效率和预测精度。第五,对比研究了基于相位编码的量子遗传算法、量子蚁群算法、量子粒子群算法、带精英保留策略的遗传算法、量子BP神经网络、量子权值神经网络、量子门节点神经网络、普通BP网络等8种预测模型在径流中长期预报中的建模应用。通过对漫湾水电站52年月径流系列及洪家渡水电站55年月径流系列的预测结果表明,量子智能优化算法的预报性能明显优于普通遗传算法,量子神经网络的预报性能明显优于普通BP网络。从而验证了量子机制的引入可以改善传统模型的预报性能,并提高月径流时间序列的预报精度。最后,对全文的研究工作进行了总结,并对有待于进一步研究的问题进行了展望。更多还原

【Abstract】 AbstractQuantum computation is a novel inter-discipline that includes quantum mechanics and information science.The quantum algorithm based on the basic principles of quantum computation has been widely concerned by scholars around the; world,and has shown the very broad application prospects.Process neural network (PNN) is a new artificial neural network (ANN) model proposed at the beginning of this century according to information processing mechanism of the biological nervous system in conjunction with the application of practical problems.The process, input of the PNN remove the instantaneous synchronization constraints for input in traditional ANN model. The process neural network is an extension model of traditional ANN in the time domain, and it is a more generalized ANN model. The research content of this thesis can be summarized as follows.Firstly, a quantum BP neural networks model and algorithm based on quantum rotation gates and quantum controlled-not gates are proposed, and then the continuity of the model is proved The experimental results of the sunspot number prediction show that the predictive power of this model is superior to common BP networks. By simulating biological neural information processing mechanism, a quantum weight neural network model is presented with both the quantum linked weight and the quantum activation value.Using gradient descent algorithm, a super-linearly convergent learning algorithm of this model is designed. The availability of the approach is illustrated by two application examples of pattern recognition and function approximation; by analyzing the physical meaning of quantum gates, a quantum gates neural network model and algorithm are introduced with quantum gate nodes.This model can significantly increase the probability of convergence because of its implicit multi-attractors.Simulation results show that the convergence speed and prediction capabilities are significantly better than that of the ordinary ANN.Secondly, the learning algorithms of PNN are developed based on Legendre orthogonal basis functions, which can effectively solve complex computing problems of spatial and temporal aggregation. For the issue of parallel processing of massive samples, a parallel process neural networks model and algorithm are constructed. This method can not only decentralized networks load, but also improve the predictive power of a single networks.Thirdly, on the basis of analyzing the existing problems in current quantum search algorithm, two improved algorithms are proposed by constructing the new quantum search engines.A quantum search algorithm is presented based on weighted targets, in which the successful probability of each marked item is consistent with the corresponding weight coefficient. Namely, the greater successful probability is gotten for the more important target. An improved quantum search algorithm with small phase rotations is proposed. When the size of phase rotations are fixed at 0.01π, the success probability, at least 99.99% can be obtained.Fourthly, by analying the problems existing in current quantum evolutionary algorithm, a quantum genetic algorithm is proposed based on the spherical coordinates of quantum bit, and two new quantum gate operators are designed. In this algorithm, by regarding coordinates of the qubit as approximate solutions of the optimization problem, this algorithm can increase the solution space traversal arid the probability of convergence.By directly taking the qubit phase as a gene on chromosome,four quantum-inspired optimization algorithms are respectively presented. In these algorithms, the optimization process is performed in [-1,1]n or [0,2π]n,which has nothing to do with specific issues, therefore, the proposed methods have good adaptability for a variety of optimization problems.By combining these algorithms into the process of neural network training, the computational efficiency and prediction accuracy of model can be significantly improved.Fifthly, on the basis of the Mid-and-Long term forecasting of monthly discharge time series, the simulation comparisons of eight intelligent optimization models are studied. Through applying these model to the two monthly discharge time series of hydropower station of the Manwan and the Hongjiadu, the prediction results show that the quantum-inspired optimization algorithms are superior to ordinary genetic algorithm,and quantum neural networks are superior to common BP neural networks, which verifies the introduction of quantum mechanisms can improve the forecasting performance of the traditional model and algorithm,and can increase the forecasting accuracy of monthly discharge time series. Finally, all the studies in this text were summarized, and some new topics are discussed.更多还原