【作者】 谭左平；

【导师】 王士同；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2009， 博士

【摘要】 智能可以看作是知识积累和知识运用的综合能力反应,主要是认识客观事物、掌握客观规律,以及运用知识去解决实际问题的能力。人工智能就是用人工系统来模拟人的问题求解、推理、学习等方面的能力。人工智能经过近半个世纪的发展,形成了多个研究发展方向,其中智能计算方法是智能科学当前研究的重要方向之一。在过去几十年的时间里,智能计算得到了广泛的研究和迅猛的发展,并在信号处理、模式识别、系统辨识、发酵控制、生物信息学、食品和医疗以及商业等领域取得了丰硕的成果。本文针对智能计算方法在生物发酵过程以及生物信息学中的相关应用进行了研究,包括Cascaded Centralized Takagi-Sugeno-Kang(CCTSK)模糊神经网络,多层感知器(Multi-Layer Perceptron)、径向基函数(Radial Basis Function,RBF)神经网络、Takagi-Sugeno-Kang (TSK)模糊系统等相关智能计算方法在谷胱甘肽、乳酸杆菌的发酵过程中的应用。另外,对生物信息学中基因表达调控网络的重构也做了一点工作,具体来说,本论文的创造性研究成果主要有:第一部分,对模糊系统进行了研究,建立了不确定的高斯混合模型和具有可加性的二型Takagi-Sugeno-Kang模糊系统之间的对应关系。然后利用Cascaded Centralized Takagi-Sugeno-Kang(CCTSK)模糊神经网络对谷胱甘肽(glutathione,GSH)的发酵过程进行建模。在实际的发酵生产过程中,由于菌体发酵过程自身的复杂性以及实验中的客观条件的限制,使得实验数据不可避免的含有一定量的噪音,从而使得传统的神经网络所建立的模型的收敛速度和精度明显下降,且建模结果缺乏可解释性。CCTSK模糊神经网络由于采用了中心化的级联的网络结构,实验结果表明,运用CCTSK模糊神经网络对谷胱甘肽发酵的过程所建立的模型具有良好的鲁棒性和更高的可解释性。第二部分,研究了鲁棒的基于熵准则的RBF谷胱甘肽发酵建模。利用信息论中关于熵的概念,从概率密度的角度出发,采用判别熵构造出一个新的误差准则函数——基于熵准则的误差准则函数,并将其应用于谷胱甘肽的发酵建模过程中。由于新的误差准则函数能够从训练样本的整体分布结构来进行模型的参数学习,从而有效地避免了传统的基于Mean Square Error(MSE)准则的RBF神经网络的过学习和泛化能力弱的缺陷。第三部分,进一步研究了基于熵准则的误差准则函数的特点。将新的准则函数用于多层感知器模型、径向基函数神经网络模型以及Takagi-Sugeno-Kang模糊系统模型中,然后对乳酸杆菌发酵生成多糖(exopolysaccharide,EPS)的过程进行建模,实验结果表明:新方法具有较高的预测精度、泛化能力以及良好的鲁棒性。最后一部分,研究了目前生物信息学中的热点——基因表达调控网络的重构问题。针对传统的线性组合模型只考虑了基因之间的线性调控关系的缺陷,引入了能量因子的概念,从而使得模型具备了分析基因间的非线性调控关系的特性。更多还原

【Abstract】 Intelligence can be looked as an compositive reflection of the accumulation and the application of the knowledge. It refers to the cognition of objectively existing things, the mastery of objective regulation and the abilitiy of making use of knowledge to solve practical problems. Artificial Intelligence (AI) just try to simulate the abilities of solving, inferring and learing as human beings by synthetic systems.After developing for nearly half an century, Artificial Intelligence contains many reasearch fileds. And the Intelligence Computation Method is one of the significant important fields. Thus many researchers have paid much attention on this field.In the past decades, Intelligence Computation has been studied wildly and achieved rapid developments. Morever, it has been arrived at fruitful achievements in the application of so many fields, such as signal processing, pattern recognition, system identification, fermentation control, bioinformatics, food science, medical treatment and business.In this dissertation, Intelligence Compution Methods and their applications in fermentation process modelling and bioinformatics have been deeply investigatied. They include Cascaded Centralized Takagi-Sugeno-Kang (CCTSK) fuzzy neural networks, Multi-Layer Perceptron, Radial Basis Function (RBF) neural networks, Takagi-Sugeno-Kang (TSK) fuzzy system and their applications in the fermentation process of Glutathione (GSH) and exopolysaccharide (EPS). Among them, we also do some reasearches on the reconstruction of the gene express regulatory network of E. coli in bioinformatics. To be concrete, the contributions of this dissertation are as follows:(1) We do some reasearches on fuzzy logic system and translate Uncertain Gaussian Mixture Model to an additive type-2 Takagi-Sugeno-Kang fuzzy logic system, and then we introduce a fuzzy-inference based fuzzy neural network, called the Cascaded Centralized Takagi-Sugeno-Kang fuzzy neural networks, for GSH fermentation process. It is well known that the data obtained from experiments unavoidly contain noises in the practical fermentation production, because of the complexity of fermentation process, the shortage of chemical apparatus and the limitation of the experimetal situation. Therefore, the covergence performance and the prediction accuracy of the GSH fermentation process modelling are often deteriorated by the noise existing in the experimental data. Moreover, the traditional model for the GSH fermentation process is usually lack of the interpretation. Since the CCTSK fuzzy neural network introduces the syllogism inference and centralized strategy, it is demonstrated that the modelling has a good robustness to the noise data and a high interpretation for the GSH fermentatiom process modeling, compared with the traditional fuzzy neural networks.(2) A robustness Radial Basis Function neural network model based on entropy criterion for the Glutathione fermentation process has been deeply investigated. Originated from the Parzen window density estimator and relative entropy for the sampling set, we propose a new criterion function, called entropy-based criterion function. Then the new criterion function is applied in the Radial Basis Function neural network model for the Glutathione fermentation process. Since the novel entropy-based criterion can be used to train the parameters of the RBF neural network model from the whole distribution structure of the training data set, which results in the fact that the Radial Basis Function neural network model method can have global approximation capability. Compared with the Mean Square Error criterion, the advantage of this novel criterion exists in that the parameter learning can effectively avoid the over-fitting phenomenon, therefore the proposed criterion based RBF neural network model have much better generalization ability and robustness for the Glutathione fermentation process.(3) We attempt to do furthur researches on entropy-based criterion function, and then incorporate this entropy-criterion based objective function into Multi-Layer Perceptron (MLP) model, Radial Basis Function (RBF) neural network model and Takagi-Sugeno-Kang (TSK) fuzzy system model for exopolysaccharide (EPS) fermentation from Lactobacillus. Our experimental results indicate that the three modeling methods mentioned above with entropy-criterion based objective function have obvious advantages over these with Mean Square Error criterion based objective function in the sense of approximation/generalization capability and robustness.The reason leading to such results may be that entropy-criterion based objective function is derived from the Parzen window density estimator and relative entropy, and it considers the whole distribution structure of the training set in the parameter’s learning process, which is quite different from the MSE-criterion based objective function.(4) We also do some researches on the hot topic in bioinformatics fields, which is named as the reconstruction of gene express regulatory networks. Reconstruction of gene regulatory network is much significant to explore the essence of life. It is well known that the Linear Combination Model has been successfully applied to the reconstruction of the gene regulator network for its easy and fast solving. However, this model just takes the linear relationships between genes into account. In order to circumvent this problem, energy factor has been added in the Linear Combination Model, thus the model can be used to analysis the nonlinear relationships between genes. Then the proposed model has been applied to reconstruct the gene regulatory network of Escherichia coli on SOS DNA repair process. Our result demonstrates that the proposed model can reconstruct the SOS DNA repair process well and improve the accuracy.更多还原