基于群智能优化的机器学习方法研究及应用

发布时间：2018-03-05 18:30

本文选题：随机森林　切入点：支持向量机　出处：《吉林大学》2017年博士论文　论文类型：学位论文

【摘要】：当前,机器学习技术已广泛应用于工业、农业、交通、环境等各个领域。特别在农业生产领域,由于农业信息化和精准农业的推广和实施,农业生产数据不断迅速增长并大量积累,这大大增加了对信息技术的需求,尤其对机器学习技术的需求更加显著。然而,由于农业生产问题的复杂性,对决策问题直接采用传统的机器学习方法往往达不到理想的效果,因此如何构造最优的基于机器学习的农业智能决策方法是当前亟待解决的问题。本文重点研究利用群智能优化技术对现有若干机器学习技术进行改进,构建基于机器学习技术的农业智能决策新方法,进而将这些新方法用于解决实际农业生产问题。我们首先对随机森林、多目标聚类、支持向量机和核极限学习机等方法在计算精度、适用性和稳定性方面存在的问题进行了探讨,提出了三维混沌果蝇优化技术、改进粒子群优化技术、改进灰狼优化技术和多种群灰狼优化技术,然后分别对上述机器学习方法进行改进,进而提出三维混沌果蝇优化的随机森林预测模型、改进粒子群的动态多目标优化诊断模型、改进灰狼优化的支持向量机诊断模型、多种群灰狼智能演化核极限学习机预测模型,分别用于水稻的虫害预测、病害诊断、缺素诊断和产量预测,较好地解决了水稻生产决策难题。具体说明如下:(1)针对随机森林模型预测过程受其参数影响的问题,提出一种三维混沌果蝇优化的随机森林预测模型。首先,将原始果蝇优化算法从二维搜索空间扩展到三维空间,同时引入混沌理论对种群进行初始化操作,避免陷入局部最优,提出一种改进的三维混沌果蝇优化算法。选取多个测试函数进行实验仿真,实验结果表明,提出的方法与原始的果蝇优化、粒子群优化等群智能算法进行对比,不仅解的质量更好,在收敛速度方面也更快。然后,我们将该算法引入到随机森林模型中,利用三维混沌果蝇优化算法对随机森林进行训练,从而建立最优的计算模型。最后,将该方法在水稻虫害数据集上进行测试,并与其它算法进行对比。实验结果表明,该模型具有更好的预测精度,能更有效的实现水稻虫害的预测。(2)针对机器学习方法在单目标粒子群优化中解的局限性问题,提出一种改进粒子群的动态多目标优化诊断模型。首先,对原始粒子群算法进行改进,包括环境变化因子、惯性因子和变异因子的改进。然后,将该方法与动态多目标技术结合,选取两个聚类方法作为目标函数,采用背景差分法设计环境变化因子及规则,建立了改进粒子群的动态多目标优化诊断模型,从而实现图像识别算法的优化。最后,对预处理的水稻病害图片进行特征提取后,应用该模型对病害特征集测试,并与其他方法进行了对比。实验结果表明,该模型可获得数量多、质量高且分布均匀的帕累托(Pareto)解集,较之单目标方法有更高的病害识别精度。(3)针对支持向量机的模型选择问题,提出一种改进灰狼优化算法的支持向量机诊断模型。首先,在灰狼优化中引入新的种群初始化机制,为灰狼优化生成更合适的种群位置,使之避免陷入局部最优,获得更好的解,并提高算法的收敛速度。在多个单峰和多峰函数上对该算法进行了测试,结果表明,改进灰狼优化算法在解的质量及收敛速度上都优于已有的灰狼优化算法。然后,我们将该策略引入到支持向量机中,对该模型中的惩罚因子和核宽进行动态选择和调整,从而得到最优识别模型。最后,利用该模型对水稻缺素情况进行诊断。实验结果表明,该模型识别准确率超过95%,识别精度优于基于原始灰狼优化的支持向量机方法、基于网格搜索的支持向量机和神经网络方法,实现了水稻缺素问题的精准决策。(4)针对核极限学习机在预测问题中受关键参数影响的问题,提出一种利用多种群灰狼智能演化核极限学习机的预测模型。首先,充分利用多种群智能演化方法的优势,将灰狼算法的种群和搜索空间同步多元化,并采用精英机制在多个种群中进行信息共享,从而获得全局最优解。然后,我们将该策略引入核极限学习机中,对该模型中的惩罚系数和高斯核宽进行动态调整,从而建立最优的预测模型。最后,对水稻产量数据进行测试,实验结果表明,与基于原始灰狼优化的核极限学习机、支持向量机以及神经网络等方法对比,该模型不仅提高了水稻产量的预测准确率,还获得了更稳定的预测结果。这意味着本文提出的水稻产量预测模型能够较好的对粮食产量进行预测,可作为农业生产中一种重要的辅助决策工具。
[Abstract]:At present, machine learning techniques have been widely used in industry, agriculture, transportation, environment and other fields. Especially in the field of agricultural production, the promotion and implementation of agricultural informatization and agriculture, agricultural production data continues to grow rapidly and accumulated, which greatly increased the demand for information technology, especially the technology of machine learning needs more significant. However, due to the complexity of agricultural production problems, learning methods often fail to achieve the desired result directly using the traditional machine of decision problem, so how to construct the optimal agricultural intelligent decision method of machine learning based on the current problems to be solved. This paper focuses on the use of existing swarm optimization several machine learning techniques the improved intelligent technology, a new method for construction of machine learning technique for agricultural intelligent decision based on, and then the new method is used to solve the actual agricultural production. The problem. We first of random forest, multi-objective clustering, SVM and kernel extreme learning machine and other methods in calculation accuracy, discussed the applicability and stability of the existing problems, puts forward a three-dimensional chaotic Drosophila optimization technology, improved particle swarm optimization technology, improved optimization technology and multi group wolf wolf optimization technology. Then the machine learning method is improved, and then puts forward the three-dimensional chaotic random forest fruit fly optimization prediction model, improved particle swarm optimization for dynamic multi-objective optimization diagnosis model, improved gray wolf optimized support vector machine multi swarm intelligent diagnosis model, the gray wolf evolution kernel extreme learning machine model, respectively for rice pest forecast, disease the diagnosis, diagnosis and prediction of lack of hormone production, solved the rice production decision problem. Details are as follows: (1) the prediction process for random forest model The influence of parameters, this paper proposes a random forest 3D chaotic Drosophila optimization prediction model. Firstly, the original fruit fly optimization algorithm from two-dimensional search space is extended to three-dimensional space, while the introduction of chaos theory for initialization of population, avoid the local optimum, this paper proposes an improved chaotic optimization algorithm. Selection of Drosophila melanogaster a plurality of test function simulation, the experimental results show that the proposed method with the original fruit fly optimization, particle swarm optimization and swarm intelligence algorithm are compared, not only better quality solutions, in terms of the rate of convergence is also faster. Then, we apply the algorithm to the random forest model, using three-dimensional chaotic optimization algorithm of Drosophila the training of random forest, so as to establish the optimal model. Finally, the method is tested in the data set on the rice pests, and compared with other algorithms. The experimental results show that the prediction accuracy of the model is better, can realize the prediction of rice pests effectively. (2) aiming at the limitations of machine learning method in single objective particle swarm optimization solutions, this paper proposes an improved model for dynamic multi objective particle swarm optimization diagnosis. First, the original particle swarm optimization algorithm improved, including environmental factors, improvement of inertia factor and variation factors. Then, combining the method of multi-objective and dynamic technology, selects two clustering method as the objective function, the background difference change factor and rule design environment, established the modified particle swarm dynamic multi-objective optimization model of diagnosis, in order to achieve optimization of image recognition algorithm. Finally, the feature extraction of image pretreatment of rice diseases, to test the characteristics of the disease in the application of the model, and other methods were compared. The experimental results table Ming, the model can be obtained in quantity, high quality and uniform distribution of Pareto (Pareto) solution set, the single target method compared with disease identification more precise. (3) for the problem of selecting support vector machine model, put forward a kind of improved support vector machine algorithm fault diagnosis gray model. First of all, the introduction of the new mechanism in the gray wolf population initialization optimization, more suitable for the generation of population location wolf optimization, so as to avoid falling into a local optimum, get a better solution, and improve the convergence speed. In a number of unimodal and multimodal functions of the algorithm were tested, the results show that the improved gray optimization algorithm a optimization algorithm is superior in both quality and convergence of the solution existing. Then, we will the strategy into support vector machine, dynamic adjustment and the choice of the model of the penalty factor and kernel width, from the optimal recognition mode Type. Finally, using the model of rice nutrient deficiency for diagnosis. The experimental results show that the model recognition accuracy rate is more than 95%, the recognition accuracy is better than the original method of support vector machine optimization based on gray grid search, neural network and support vector machine method based on rice nutrient deficiency problems to achieve accurate decision-making. (4) according to the kernel extreme learning machine by the key parameter in the prediction of the problem, put forward a prediction model of evolution kernel extreme learning machine using a variety of swarm intelligent wolves. First, make full use of the advantages of swarm intelligence evolution method, gray wolf population algorithm and search space synchronization and diversification, information sharing mechanism with the elite in multiple populations, so as to obtain the global optimal solution. Then, we will introduce the strategy of kernel extreme learning machine, the model of penalty coefficient and width of Gauss kernel in dynamic adjustment Thus, the optimal prediction model is established. Finally, test the rice yield data, the experimental results show that machine learning and Optimization Based on the original wolf nuclear limit, compared to support vector machine and neural network method, this model not only improves the prediction accuracy rate of rice yield, but also prediction results more stable. This means that the proposed rice yield prediction model can better predict the grain yield, agricultural production can be used as an important auxiliary decision-making tool.

【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】