Joint design of model set identification and learning type control

对于模型集辨识和学习型控制的联合设计，我们得到了以下结果：对于模型集辨识，难点之一是模型集辨识的框架与传统的随机参数辨识方法不一致。因此，所得到的模型集趋于保守。提出了通过考虑噪声和输出信号之间的独立性等随机特性来获得模型集的辨识方法。这部分克服了这一缺点。其次，我们考虑了一类称为哈密顿系统的非线性系统。这门课包含机械系统和电气系统的组合。对于学习控制，现有方法的主要缺点是在不了解被控对象的精确知识的情况下，需要对误差信号进行微分。我们已经通过几种方式解决了这个问题。一种是采用迭代学习控制时，将输入空间限制在给定的有限输入信号内，而迭代学习控制与模型辨识密切相关。另一种是利用哈密顿系统的伴随系统的I/O信号来计算其相对于给定代价函数的梯度。这里的重点是，我们可以在没有任何模型参数的情况下实现这一点。通过使用非线性机械手的实验，证明了这两种方法的有效性。我们还开发了一种新的对摩擦具有鲁棒性的迭代反馈整定方法，将学习控制与模型集辨识相结合的更有效方法的开发还有待于进一步的研究工作。

As for the joint design of model-set identification and learning type control, we have obtained the following results.Concerning to the model-set identification, one of the difficulties is that the framework of model set-identification is not consistent with the traditional stochastic approach of parameter identification. As a result, the obtained model set tends to be conservative. We have proposed identification methods which obtain model sets by taking the stochastic properties such as independency between noises and output signals into account. This overcomes the shortcoming partially. The effectiveness is evaluated through experiments using flexible structures.Next, we have considered a class of nonlinear systems which are called Hamiltonian systems. This class contains a combination of mechanical systems and electrical systems. We have clarified that the inherent structures such as passivity and adjoint systems, which form a basis of applicability of learning control to this class of systems.As for learning control, major demerits of the existing methods is that they have to use differential of error signals when the precise knowledge of the plants is not available. We have solved this problem in a couple of ways. One is to restrict the input space into a prescribed finite input signals when we adopt an iterative learning control, which turns out to be related to model identification very closely. The other is to use I/O signals of the adjoint systems of Hamiltonian systems in order to calculate its gradient with respect to given cost functions. The point here is that we can achieve this without any model parameters. The usefulness of both methods are demonstrated through experiments using nonlinear manipulators. We also have developed a new method of iterative feedback tuning which is robust against frictions.The development of more effective way of combining learning control with model-set identification is left as a future research work.