Self-Constructing Systems and their Applications to System Identification and Control

自建系统在观测数据的基础上进行自我构建和自我修复。对于描述非数字现象的系统，开发了一种程序，该程序自动构建一般规则集模型并检测数据中的变化，其中适当控制双边信息流、从数据到概念的泛化和与数据相矛盾的概念的专门化。给出了两类神经网络的自构造方法：一类是可以统一表达实例和概念的神经网络，另一类是可以用输入门显式表达概念的神经网络。此外，提出了一种易于处理数据变化的神经网络结构设计方法和一种新的优化方法，该方法可用于神经网络训练并获得创造性的新函数；作为自构造系统的通用框架，提出了两种网络型框架：用于表示用实值变量描述的连续系统的通用学习网络和用于变量取离散值的离散事件系统的自动机学习网络。两者都具有通过学习进行自我建构和自我改变的能力。并设计了高效的学习算法。提出了一种自动确定网络结构的方法，自构造系统的主要目标之一是开发出控制系统，该控制系统通过自我改变而表现出良好的性能，而不受周围环境的变化。因此，在通用学习网络的框架下，以起重机的控制为例，开发了鲁棒控制系统，该系统能够很好地处理有效载荷、载荷初始位置和目标位置以及外部扰动的变化。

Self-Constructing systems construct and after themselves based on observed date. For systems describing nonnumerical phenomena, a procedure was developed that automatically constructs general rule-set models and detects changes in the data, where bilateral flows of information, generalization from the data to concepts and specialization of concepts that contradict to the data, is appropriately controlled. Also, self-constructing methods were developed for two types of neural networks : neural networks that can express both instances and concepts in a uniform manner, and those which can express concepts explicitly with their input gates. Moreover, a structure design method was proposed for neural networks which can easily cope with changes in the data, and a new optimization method was devised that can be employed in neural networks training and leads to acquiring creatively new functions.As a general framework for self-constructing systems, two network-type frameworks were proposed : Universal Learning Networks for representing continuous systems whose behaviors are described by real-valued variables, and Automaton Learning Network for discrete event systems where the variables take discrete values. Both of them have self-constructing and self-altering capability by learning. And efficient learning algorithms were devised. Also, an automatic structure determination method for the networks was developed.One of the major aims of self-constructing systems is to develop control systems that exhibit, by self-alteration, good performance despite of changes in their surrounding situations. Therefore, in the framework of Universal Learning Network, with control of a crane as an example, robust control systems were developed which perform satisfactorily well coping with changes in payload, initial and target positions of the load and external disturbances.