【作者】 李枚毅；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 模式识别与智能系统， 2005， 博士

【摘要】 移动机器人技术在空间探测、国防、工业、农业、医学及社会服务业等领域显示了越来越广泛的应用前景,已成为国内外移动机器人学术界研究的热点。本文针对大范围环境、变化环境和未知环境,以免疫进化和多示例学习作为支撑技术,围绕移动机器人在它的运动过程中始终需要解决的定位与规划二个关键问题进行了比较深入的研究,其研究内容涉及基于多图像的定位、并发定位与建图、路径规划、进化与免疫计算和多示例学习等。 本文完成的主要工作和取得的创新性成果如下。 通过分析现有采样方法和多模函数优化过程中典型的早熟收敛现象,认为进化计算和免疫算法等智能优化算法需要一种以父体两旁为着重点搜索区域的采样范式,并提出使用逆正态分布进行采样。通过理论分析和实验,证实了这是一种在跳出局部极值方面有明显优势的采样方式。 提出了一种多样度和适应度联合引导的选择、交叉与变异概率适应性计算策略;通过实验和分析,总结出了选择压力与进化算法性能之间的密切关系。通过数学分析,证明了二进制遗传算法的变异概率和多样度之间的数学关系,为多样度引导的遗传算法提供了一种变异概率调整理论和计算方法。 在进化计算框架下,融合全局并行搜索的克隆选择和启发式局部搜索的免疫疫苗接种,设计了一类免疫克隆进化算法,并证明了该算法的全局收敛性。以该算法为核心,针对大范围环境和变化环境下的路径规划及未知环境下并发定位与建图等任务,分别提出了四种嵌入了领域知识的启发式免疫操作和进化操作的免疫克隆进化算法,并研究了各种任务下隐藏领域知识的提取问题。由于这些算法嵌入了从领域知识中提取出来的启发式操作算子,而且核心算法具有全局搜索能力,因此,明显地提高了这些算法处理相应问题的能力。 结合多示例学习问题的固有特性,提出了一种监督-非监督多示例神经网络,该网络以负示例作为监督学习的教师信息,以正包进行非监督自组织聚类学习。该网络训练速度比较快、准确性较高和具有多概念学习能力。利用这些性质,设计了基于多示例学习的图像多候选目标识别方法和一种移动机器人导航方法,并通过移动机器人进行了导航实验。 针对未知大范围环境下移动机器人定位问题采用了多图像表达场景信息,并在此基础上,提出了使用多示例学习的自动发现能力识别不同场景来进行移动机器人定位的方更多还原

【Abstract】 Mobile robot technique has shown broader application prospect in many fields, such as space exploration, national defense, industry, agriculture, medicine, and social service etc. And the technique has become a research focus in the academic field of robotics globally. Basing on the techniques of immune evolution and multi-instance learning and focusing on changing environments, large scope environments and unknown environments, this dissertation revolves the localization and path planning problems, which are crucial and need to be solved in the motion of the mobile robot. The main research efforts of this dissertation are about localization and path planning of the mobile robot in large scope environments, changing environments and unknown environments, which include: localization based on multiple images, concurrent mapping and location (CML), path planning, evolutionary computation and immune computation, multi-instance learning, etc.The primary work and the contributions in this dissertation are as follows: By means of analysis of current sampling methods and typical phenomena in the optimization process of multimodal functions, it is discovered that there is a demand of the sampling paradigm to search mainly at the both near sides of the parent and an anti-normal distribution is presented as its implementation. Employing theoretic consequences and experiments, it is proved that the sampling method using the anti-normal distribution has greater ability of escaping from local optima than other probability distributions such as normal distribution, Cauchy distribution, etc. Methods of calculating adaptively probabilities of selection, crossover and mutation operators with the combination of both diversity and fitness are proposed by means of a simple method of machine learning. By means of experiments and analyses, relations of selection pressure and performances of EC are summed up. Employing mathematical reasoning, mathematical relations of mutation probabilities and diversity in binary genetic algorithms are proved which provide theorems and computation methods for adjusting adaptively mutation probabilities with diversity. Under the framework of evolutionary computation (EC), immune clonal evolutionary algorithms are constructed, which integrate the vaccination with heuristic local search and clonal selection with parallel global search. And the global convergences about the algorithms are proved. On the foundation of the algorithms, four algorithms are designed individually in which immune opreators and evolutionary opereators with domain knowledge-based heuristic rules are embedded for tasks of path planning under large environments, changing environments and unknown environments, and for ones of CML under unknown environments. Since these algorithms are combined with domain knowledge-based opreators and the basal algorithm can explore globally, the ability of dealing with corresponding problems by these algorithms has been obviously improved.Combining the inherent characteristics of the multi-instance learning (MIL), supervised-unsupervised MIL neural networks are proposed, in which negative instances are taken as teacher’s signs and positive instances in positive bag are clustered with self-organization. The training of the networks will be finished quickly. The predictive accuracy of the networks is high and the networks have the ability of learning multi-concepts. An approach to recognize candidate from multiple targets in an image based on the multi-instance learning (MIL) and a method of mobile robot localization are designed to utilize these natures, further the navigation experiments are finished by the mobile robot. Aiming at problems of mobile robot in large unknown environments, the information of expressing scenes by multiple images is adopted. Further, the method of mobile robot localization is presented and implemented using the ability of automatic discovery of multi-instance learning to identify different scenes. We have realized to extract the hidden features automatically from the multiple images, further realized the mobile robot localization in large and unknown environments by the mobile robot.In a word, by means of immune evolutionary computation with immunity and multi-instance learning, problems of mobile robot localization and path planning in large scope environments, changing environments and unknown environments have been investigated, the thoeries and methods of evolutionary computation, immune evolutionary computation and multi-instance learning have been investigated deeply, the sampling method using anti-normal distribution is presented, strategies of calculating adaptively probabilities of selection, crossover and mutation operators with the combination of both diversity and fitness and the supervised-unsupervised MIL neural networks with self-organization are proposed. The theorems put forward in the dissertation are proved and analyzed on mathematics. The efficiency, reliability and practicability of the methods in this dissertation have been validated for the mobile robot navigation through theoretic analysis, simulations and physical experiments. The research work of this dissertation has offered some new thinkings of settlement for mobile robot navigation in complicated environments. The localization and planning methods proposed have directive significance in theory and reality for mobile robot to carry out the navigation task in complicated environments.更多还原