【作者】 宁爱平；

【导师】 张雪英；

【作者基本信息】 太原理工大学 ， 电路与系统， 2013， 博士

【摘要】 群智能优化算法是受自然界中群居生物生活习性的启发而提出用来解决复杂优化问题的算法,目前己成为解决优化问题的研究热点。人工蜂群算法就是受蜜蜂采蜜行为的启发而提出的一种新型群智能优化算法。人工蜂群算法由于在寻优过程中能很好地兼顾食物源的开采和探测过程,能在一定程度上跳出局部最优,以较大的概率找到全局最优解,因而深受广大学者的关注语音识别技术是实现人与计算机能很好交流的关键技术,而语音识别模型是语音识别的重要模块,因此如何使识别模型更优是语音识别领域的重点课题。本文在对人工蜂群算法的性能分析和理论研究的基础上,提出了两种改进的人工蜂群算法,并探讨了人工蜂群算法在语音识别模型中的应用,识别结果显示了人工蜂群算法及改进的人工蜂群算法良好的性能。本文的主要内容和创新性成果如下：(1)阐述了人工蜂群算法的产生背景及研究现状,深入研究了人工蜂群算法的产生机理,分析了人工蜂群算法与其他智能优化算法的共性和特性及算法的时间复杂度,用标准函数测试了人工蜂群算法的优化性能。(2)对人工蜂群算法进行了理论研究。对人工蜂及人工蜂群的状态及状态空间进行了严格的数学定义,给出了人工蜂群的一步转移概率,证明了人工蜂群状态转移过程是有限齐次的Markov链。根据随机优化的收敛准则,分析了人工蜂群算法的全局收敛性。利用鞅定义和下鞅收敛定理,分析了人工蜂群算法搜索过程中适应度的变化过程实质上是一个下鞅随机过程,证明了算法的处处收敛性。(3)针对人工蜂群算法易早熟和收敛速度慢等缺陷,本文从两个角度出发提出了两种改进的人工蜂群算法。①混沌时变人工蜂群算法。采用了混沌映射初始化种群来增加种群的遍历性,在跟随蜂搜索方程中加入时变参数因子,使跟随蜂在搜索过程中根据迭代次数的变化不断改变其搜索空间,加快搜索效率,同时为了使算法能更好的跳出局部最优,在侦查蜂搜索阶段引入混沌搜索。②排序分裂选择的人工蜂群算法。跟随蜂选择食物源时先依据排序选择策略对适应度值排序,然后把排列序号按分裂选择的思想计算选择概率,在维持种群多样性的同时,提高了搜索精度。(4)针对离散隐马尔可夫(DHMM)孤立词语音识别系统中由矢量量化引起的量化误差造成识别率低的问题,在分析LBG算法的基础上,提出了人工蜂群算法和混沌时变人工蜂群算法的码书设计方法。算法中食物源代表码字,适应度函数用矢量量化的失真测度表示,算法的搜索过程就是最优码书的生成过程。将人工蜂群算法和混沌时变人工蜂群算法应用到DHMM的语音识别系统中,识别结果与粒子群初始化码书的LBG算法及LBG算法的DHMM语音识别系统的识别结果相比,显示出这种方法的可行性及CTABC算法很好的寻优能力。(5)为了解决模糊神经网络的传统训练算法BP算法不能很好地找到全局最优及易陷入局部极小值等问题,提出了采用人工蜂群算法和排序分裂选择的人工蜂群算法优化模糊神经网络的隶属度中心,宽度由中心周围的样本点确定,伪逆法计算模糊神经网络权值矩阵的混合优化算法。将优化后的模糊神经网络用于语音识别,识别结果与粒子群算法和BP算法优化的模糊神经网络相比,表明这种混合学习算法的有效性,同时显示出排序分裂选择的人工蜂群算法算法容易跳出局部极值,能很好地找到全局最优,且提高了模糊神经网络的鲁棒性和识别率。(6)由于支持向量机的核函数及参数对支持向量机的分类性能有很大的影响,因此选择好的参数优化方法对支持向量机来说是非常必要的。针对传统支持向量机参数选择方法易陷入局部极值的问题,提出了人工蜂群算法的支持向量机参数优化方法。该方法中用食物源位置表示支持向量机的惩罚因子和核参数,适应度函数用分类正确率的函数来表示,人工蜂群算法的搜索最优食物源的过程就是支持向量机寻找最优参数的过程。用优选参数后的支持向量机对三种数据库进行了语音识别,识别结果与粒子群算法优化的支持向量机进行比较,表明了人工蜂群算法是一种很好的参数寻优方法,不仅克服局部最优解、提高了语音识别率,还增强了支持向量机的鲁棒性和推广能力。更多还原

【Abstract】 Swarm intelligent optimization algorithm is inspired by the living habits of gregarious biology, and currently it has become a hotspot to solve complex optimization problems. Artificial Bee Colony (ABC) algorithm is a new type of swarm intelligent optimization algorithm, which is inspired by the behavior of bees finding honey. In the optimization process, ABC algorithm can well balance the exploitation and detection processes of food source. It can escape from the local optima to a certain extent, and find global optimal solution with the larger probability. Thus more and more researchers pay attention to it.Speech recognition technology is a key technology to exchange information between people and computers. Speech recognition model is an important speech recognition module, and therefore how to make better recognition model is a key subject in the field of speech recognition. In this thesis, two improved ABC algorithms are proposed on the base of the performance analysis and theoretical study of artificial bee colony algorithm, and then the applications in speech recognition model are discussed too. The results of speech recognition show that the ABC algorithm and improved ABC algorithms have good optimization performance. The main works and innovative results of this thesis are listed as follows:(1) The background and research status of ABC algorithm are described. The generation mechanism of ABC algorithm is studied in depth. The commonalities and characteristics of ABC algorithm with other intelligent optimization algorithms and the time complexity of algorithm are analyzed. The optimized performances of ABC algorithm are tested by optimizing standard functions.(2) The theoretical research is performed on ABC algorithm. The strict mathematical definitions are given for the state and state-space of the artificial bee and artificial bee colony and one-step transition probability of artificial bee colony. The state transition process of artificial bee colony is proved to be a finite homogeneous Markov chain process. According to the convergence criteria of stochastic optimization, global convergence of artificial bee colony algorithm is analyzed. The martingale definition is used to analyze the search process of ABC algorithm where the changing process of fitness is a submartingale stochastic process. Then the everywhere strong convergence of the algorithm is proved by submartingale convergence theorem.(3) According to the analysis result of ABC algorithm which is prone to premature and slowly convergence speed, this paper proposed two improved method from two points of view which are Chaos time Variant Artificial Bee Colony (CTABC) and Ranking Disruptive Selection Artificial Bee Colony Algorithm (RDABC).①CTABC:Firstly chaotic map is used to initialize the population for increasing population ergodicity. Secondly the Time Variant parameters factor is joined in search equation of onlooker bees which change the search space and speed up the search efficiency according to the change of iterative times in the search process. Finally in order to make the algorithm better escape from local optima, chaotic search is applied in the scout bees search stage.②RDABC:When onlooker bees select food source, the fitness value is ranked based on the rank fitness selection strategy. Then the selected probability is calculated by rank ordinals using disruptive selection method. This selection aigorithm maintains the diversity of the population and improves search precision.(4)In the Discrete Hidden Markov Model (DHMM) isolated word speech recognition system, the problem of low speech recognition is caused by the quantization error of vector quantization. The paper proposes the two modified DHMM speech recognition algorithms which use ABC algorithm and CTABCalgorithm to cluster speech feature vector and generate the optimal codebook in the DHMMspeech recognition system. In ABC and CTABC algorithms, each food source indicates a codebook. The optimal codebook is obtained by using bee evolution ways to iterative initial codebook. The optimal codebook enters the DHMM to be trained and recognized. The experimental re sults show that the modified DHMM speech recognition algorithm has higher recognition ratio and better robustness.(5) In order to overcome the shortages that BP algorithm does not well find the global optimum and is easy to fall into local minimum values, a novel hybrid learning algorithm for fuzzy neural network(FNN) parameters is proposed, hybrid learning algorithm is that apply ABC and RDABC clustering algorithm to determine the centers of the membership function and use the sample points around centers determine the widths of the membership function and use pseudo-inverse method determine the weight between normalization layer and output layer. The FNN trained by hybrid learning algorithm is used in speech recognition system which improves the ability of generalization and self-learning of FNN and is able to determine the fuzzy rule numbers according to the vocabulary to be recognized. The experimental results show that the FNN optimized by hybrid learning for speech recognition system have faster convergence, higher recognition ratio and better robustness than FNN trained by PSO algorithm and BP algorithm.(6) Because of the support vector machine (SVM) kernel functions and parameters have a large impact on classification performance of SVM,the research of parameter optimization method of SVM is very necessary.In order to solve the problem of falling into local optimal solution of all the common SVM parameters selection methods, a novel SVM parameters optimization method based on ABC algorithm is proposed. In this method, food-source position is the penalty factor and kernel parameter, and fitness value is classification accuracy function of SVM. The process of ABC searching optimal food source is the process of SVM parameters selection. Compared speech recognition results with PSO algorithm, the proposed algorithm is a good parameters optimization method of SVM, which not only can overcome the local optimal solution problem and increase speech recognition ratio but also enhance robustness and generalization ability of SVM.更多还原