【作者】 刘娟；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 模式识别与智能系统， 2003， 博士

【摘要】 本论文的着眼点不仅限于对计算机科学和控制工程领域的技术与算法研究，还在于对人类智能和认知心理的探索。考虑到运动能力是智能机器人系统的关键能力，而时序记忆也在人类智能活动中扮演了重要角色，因此论文以移动机器人导航问题为载体，综合信息的时间与空间关系，探讨人工系统中非符号表示的内在知识的学习和运用方式，寻求实现智能的新途径。 论文工作主要围绕移动机器人空间认知问题展开，以时空序列模式处理为技术主线，从机器人的视角出发，着重分析空间知识学习和表示方法、环境与机器人之间的交互、感知与行为的耦合，研究在缺乏坐标信息和先验地图的条件下、适应于实际物理系统能力的导航策略，在智能机器人系统范式、内部认知模型、时序信息处理、传感信息处理技术、导航机制的设计与实现等方面提出了相应的理论和方法。本文的研究立意、分析视角和实现技术与传统方法均有所不同，其主要工作和成果包括： 1．在深入研究已有系统结构和特点的基础上，改变以感知、规划、行为三者关系分析机器人体系结构的传统方法，从知识类型这一独特视点，建立了包含外在世界模型、内部认知模型和隐含行为知识这三个知识层面的智能机器人范式(paradigm)理论。新范式更完整地描述和区分了机器人系统的智能水平，明确提出内在知识的形成和利用问题，将学习因素有机地融入系统框架中。 2．从机器人的视角出发，分析系统能力、导航任务和工作过程，首次提出空间知识在系统认知中的时空转换过程，阐明了时序记忆在空间认知过程中的重要性，为实现移动机器人内部认知模型提供了可行途径。 3．深入分析现有连接机制在处理序列数据能力方面的缺陷，提出非对称、增量式联想网络——时间序列处理网络(Temporal Sequence Processing Network，TSPN)，实现对复杂序列的记忆与重建，包括对序列项间隔时间的记忆。通过对网络性能的分析与实验，验证了该方法具有良好的记忆深度和容量，在计算复杂性、记忆内容的可访问性、灵活性、容错性等方面优于已有方法。 4．针对移动机器人感知信息的不确定、不完整、多噪声等特点，提出了一类传感信息动态处理方法，包括以超声波测距信息为基础的结构特征提取方法、以图像信息为基础的路标主动选择方法，以及融合距离、视觉信息的区域识别方法。上述方法充分利用当前信息和历史数据，可为导航过程提供简洁、丰富的环境信息，实际系统实验和仿真验证了这些方法的可行性和有效性。 5．提出基于时空经验的移动机器人导航机制，这种以内部认知模型为基础的导航方法消除了对机器人的传感能力的不合理假设，没有传统导航方法中对环境先验地图、初始和全局中南大学刘娟博士学位论文 坐标以及定位算法的要求。系统所获得的具有一定距离信息的环境拓扑知识，完全建立 在机器人自身经验基础上，以TSPN中的细胞时域特性表示，可在外部环境和个体意图 刺激下重现，指导机器人行为。这种与生物行为类似的导航机制，可使机器人在行为过 程中，快速建立对环境的基本认识，并不断学习。文中讨论了新方法与其他方法的关系 和特点，并提出非理性智能(non一ratfonal intelligcnce)是人工智能研究中不可忽视的内容。 仿真和实际实验表明，该导航策略使机器人对环境结构变化具有较强的适应性，而对先 验知识的需求显著降低。 论文的最后一章在总结论文工作的基础上，对进一步的研究方向进行了分析和展望。 本研究不仅对于解决应用问题有实际意义，更希望能为智能机器人方向拓展更为宽广的研究空间。更多还原

【Abstract】 Intelligent robotics, the product of advanced manufacturing technique, modern electronic and information science, has penetrated into almost every domain of our daily-life. Although they have been applied in various domains and some can communicate with us by simple sentences and sign language, as well as face expression, they are not endowed with human intelligence and adaptability. They just carry out routine programs or work reactively in a specified world. Artificial Intelligence (AI) aimed at thinking machine is still far from its ultimate goal after decades’ efforts. Some researchers come to the conclusion that intelligence can never be generated from algorithms. Do we miss something in our exploration of AI and intelligent robots? Are we on the wrong direction? After reflection, people begin to look for the outlet from the nature, biology and ethology. The theme of this dissertation is proposed under this circumstance. It is not only a research on techniques and algorithms applied in computer science and control engineering, but also an adventure in the domain of human intelligence and cognitive psychology. Considering that motion ability is critical for intelligent robotic systems and temporal memory also plays an important role in human intelligence, the dissertation is focused on mobile robot navigation problem, synthesizing the spatial and temporal relationships of various data. By probing into the learning and utilizing process of non-symbol represented knowledge in artificial system, it is aimed to exploit a new path towards intelligence.The study in this dissertation is carried out around the spatial cognition problem of mobile robot systems, taking spatio-temporal pattern processing as the core of its methods. The problem is analysed from the viewpoint of the robot, with emphases on the learning and representation of spatial knowledge, the interaction between the environment and the agent, and the coupling of sensing and acting. Its immediate purpose is to find a practical navigation strategy that fit in with the perception and motion capabilities of real robots when there is lack of coordinate information and a priori map. A set of self-contained theory and methods are proposed, including the paradigm of intelligent robot .systems, the implicit cognitive map, temporal sequence processing network, the techniques for sensory data processing, the design and realization of a novel navigation strategy. This dissertation differs from other researches in its objective, viewpoint and approaches. Its main contributions are as follows:1. A set of knowledge-oriented paradigms is established after the comprehensive overview of existing system architectures and traditional paradigms of robot systems that discriminate from each other by the relationship between the Sense, Act and Plan primitives. The new paradigmsare described in the way world knowledge is organized in an intelligent robot system. Three knowledge levels are introduced, i.e. external world model, internal cognitive model and tacit knowledge. The intelligent level of robot systems are completely described and clearly discriminated in the new paradigms. The formation and utilization of implicit knowledge are explicitly stated. As the forth primitive, learning is seamlessly contained in the new framework.2. Based on the analysis of system capabilities, navigation tasks and working process from the standpoint of the robots, it is proposed that a spatio-temporal transformation procedure happens during the learning of spatial knowledge, which clarifies the importance of temporal memory in spatial cognition and elicits an approach to implement cognition maps of mobile robots.3. A non-symmetrical incremental associative network-Temporal Sequence Processing Network (TSPN) is proposed to implement the encoding and retrieval of complex sequences and the delay between sequence items. The analyses and experiments prove that its memory depth, capacity, content accessibility, flexibility, computational complexity and fault tolerance performanc更多还原