Control System Design of Neuro-controller Using Genetic Algorithms for Non-holonomic Systems

本课题的主要研究目标是利用基于遗传算法的神经控制器（NC）构建非完整系统的控制系统。非完整系统控制器设计的一种方法是利用链式形式转换的时间状态控制形式。链式结构对于非完整控制系统的设计具有强大的实用价值。但是，时间状态控制形式由于转换的原因，在可控范围内存在一定的局限性。在本研究中，我们提出了一种非完整系统的状态反馈控制器的设计方法。NC通过遗传算法进行训练。在控制器设计中，充分利用了神经网络的模式识别能力和泛化能力。在遗传算法过程中，根据控制性能对nc进行评估，其中计算从所有初始状态开始的控制仿真结果的平方误差。基于控制性能，通过遗传过程对nc进行演化。仿真结果表明，用遗传算法训练的神经网络对非完整系统的一些样例对象具有良好的控制性能。数控系统的控制策略之一类似于时间状态控制形式。该方法在初始状态的可控范围内不受限制。

The main object of this research project is to construct control system for non-holonomic systems using neurocontroller (NC) based on a genetic algorithm (GA). One method for the nonholonomic system controller design is the time-state control form that utilizes a chained form conversion. The chained forms are powerful and useful for designing the nonholonomic control system. However, the time-state control form has some limitations in the controllable ranges due to the conversion. In this research, we propose a design method of a state feedback controller for a nonholonomic system using an NC without chained forms. The NC is trained by a genetic algorithm. In the controller design, the abilities of pattern recognition and generalization of the neural network are utilized. In the GA process, NCs are evaluated on the basis of control performance in which the squared errors that result from the control simulations starting from all the initial states are calculated. Based on the control performance, NCs are evolved through the GA processes. Results of simulations show that the NCs trained using a GA exhibit good control performance of some example objects of the nonholonomic systems. One of the control strategies of the NC resembles that of time-state control form. The proposed method has no limitations in the controllable ranges in the initial states.